Skip to contentSkip to conversation

drm/nvdla: Add driver support for NVDLA

PATCH Superseded Cai Huoqing 7 participants 19 messages Apr 19, 2022 – Apr 27, 2022 Page 1 of 2

Subsystems (3)

v1 of 2 Superseded Latest is v2 →

lore
Share & tools
lore
#0 lore raw

cover drm/nvdla: Add driver support for NVDLA

The NVIDIA Deep Learning Accelerator (NVDLA) is an open source IP which is integrated into NVIDIA Jetson AGX Xavier,
so add driver support for this accelerator.

NVDLA introduce:
http://nvdla.org/primer.html

User mode driver:
https://github.com/caihuoq/nvdla/tree/main/sw/umd

Cai Huoqing (2):
MAINTAINERS: Add the driver info of the NVDLA
drm/nvdla: Add driver support for NVDLA

 MAINTAINERS                             |    7 +
 drivers/gpu/drm/Kconfig                 |    2 +
 drivers/gpu/drm/Makefile                |    1 +
 drivers/gpu/drm/nvdla/Kconfig           |    8 +
 drivers/gpu/drm/nvdla/Makefile          |   19 +
 drivers/gpu/drm/nvdla/nvdla_bdma.c      |  200 +
 drivers/gpu/drm/nvdla/nvdla_cache.c     |  215 +
 drivers/gpu/drm/nvdla/nvdla_cdp.c       |  300 ++
 drivers/gpu/drm/nvdla/nvdla_common.c    |  295 ++
 drivers/gpu/drm/nvdla/nvdla_common.h    |  835 +++
 drivers/gpu/drm/nvdla/nvdla_conv.c      |  683 +++
 drivers/gpu/drm/nvdla/nvdla_drm.c       |  695 +++
 drivers/gpu/drm/nvdla/nvdla_drm.h       |  127 +
 drivers/gpu/drm/nvdla/nvdla_engine.c    |  233 +
 drivers/gpu/drm/nvdla/nvdla_engine.h    |  272 +
 drivers/gpu/drm/nvdla/nvdla_gem.c       |  393 ++
 drivers/gpu/drm/nvdla/nvdla_ioctl.h     |   99 +
 drivers/gpu/drm/nvdla/nvdla_pdp.c       |  446 ++
 drivers/gpu/drm/nvdla/nvdla_reg.h       | 6411 +++++++++++++++++++++++
 drivers/gpu/drm/nvdla/nvdla_rubik.c     |  217 +
 drivers/gpu/drm/nvdla/nvdla_sched.h     |   52 +
 drivers/gpu/drm/nvdla/nvdla_scheduler.c | 1005 ++++
 drivers/gpu/drm/nvdla/nvdla_sdp.c       |  728 +++
 23 files changed, 13243 insertions(+)
 create mode 100644 drivers/gpu/drm/nvdla/Kconfig
 create mode 100644 drivers/gpu/drm/nvdla/Makefile
 create mode 100644 drivers/gpu/drm/nvdla/nvdla_bdma.c
 create mode 100644 drivers/gpu/drm/nvdla/nvdla_cache.c
 create mode 100644 drivers/gpu/drm/nvdla/nvdla_cdp.c
 create mode 100644 drivers/gpu/drm/nvdla/nvdla_common.c
 create mode 100644 drivers/gpu/drm/nvdla/nvdla_common.h
 create mode 100644 drivers/gpu/drm/nvdla/nvdla_conv.c
 create mode 100644 drivers/gpu/drm/nvdla/nvdla_drm.c
 create mode 100644 drivers/gpu/drm/nvdla/nvdla_drm.h
 create mode 100644 drivers/gpu/drm/nvdla/nvdla_engine.c
 create mode 100644 drivers/gpu/drm/nvdla/nvdla_engine.h
 create mode 100644 drivers/gpu/drm/nvdla/nvdla_gem.c
 create mode 100644 drivers/gpu/drm/nvdla/nvdla_ioctl.h
 create mode 100644 drivers/gpu/drm/nvdla/nvdla_pdp.c
 create mode 100644 drivers/gpu/drm/nvdla/nvdla_reg.h
 create mode 100644 drivers/gpu/drm/nvdla/nvdla_rubik.c
 create mode 100644 drivers/gpu/drm/nvdla/nvdla_sched.h
 create mode 100644 drivers/gpu/drm/nvdla/nvdla_scheduler.c
 create mode 100644 drivers/gpu/drm/nvdla/nvdla_sdp.c
Signature · 1 line
2.25.1
#1 lore raw

PATCH 1/2 MAINTAINERS: Add the driver info of the NVDLA

The NVIDIA Deep Learning Accelerator (NVDLA) is an open source IP which is integrated into Jetson AGX Xavier. After adding the driver support for it, I add the driver info of the NVDLA to MAINTAINERS file.

---
 MAINTAINERS | 7 +++++++
 1 file changed, 7 insertions(+)
MAINTAINERS +7 -0
index 452f3662e5ac..0e5464d61a6d 100644
@@ -6485,6 +6485,13 @@ T:	git git://anongit.freedesktop.org/drm/drm-misc
 F:	Documentation/devicetree/bindings/display/panel/samsung,lms380kf01.yaml
 F:	drivers/gpu/drm/panel/panel-widechips-ws2401.c
 
+DRM DRIVER FOR NVDLA
+M:	Cai Huoqing <[email protected]>
+L:	[email protected]
+S:	Maintained
+T:	git git://anongit.freedesktop.org/drm/drm-misc
+F:	drivers/gpu/drm/nvdla/
+
 DRM DRIVERS
 M:	David Airlie <[email protected]>
 M:	Daniel Vetter <[email protected]>
Signature · 1 line
2.25.1
#2 lore raw

PATCH 2/2 drm/nvdla: Add driver support for NVDLA

In reply to Cai Huoqing

The NVIDIA Deep Learning Accelerator (NVDLA) is an open source IP which is integrated into NVIDIA Jetson AGX Xavier,
so add driver support for this accelerator.

---
 drivers/gpu/drm/Kconfig                 |    2 +
 drivers/gpu/drm/Makefile                |    1 +
 drivers/gpu/drm/nvdla/Kconfig           |    8 +
 drivers/gpu/drm/nvdla/Makefile          |   19 +
 drivers/gpu/drm/nvdla/nvdla_bdma.c      |  200 +
 drivers/gpu/drm/nvdla/nvdla_cache.c     |  215 +
 drivers/gpu/drm/nvdla/nvdla_cdp.c       |  300 ++
 drivers/gpu/drm/nvdla/nvdla_common.c    |  295 ++
 drivers/gpu/drm/nvdla/nvdla_common.h    |  835 +++
 drivers/gpu/drm/nvdla/nvdla_conv.c      |  683 +++
 drivers/gpu/drm/nvdla/nvdla_drm.c       |  695 +++
 drivers/gpu/drm/nvdla/nvdla_drm.h       |  127 +
 drivers/gpu/drm/nvdla/nvdla_engine.c    |  233 +
 drivers/gpu/drm/nvdla/nvdla_engine.h    |  272 +
 drivers/gpu/drm/nvdla/nvdla_gem.c       |  393 ++
 drivers/gpu/drm/nvdla/nvdla_ioctl.h     |   99 +
 drivers/gpu/drm/nvdla/nvdla_pdp.c       |  446 ++
 drivers/gpu/drm/nvdla/nvdla_reg.h       | 6411 +++++++++++++++++++++++
 drivers/gpu/drm/nvdla/nvdla_rubik.c     |  217 +
 drivers/gpu/drm/nvdla/nvdla_sched.h     |   52 +
 drivers/gpu/drm/nvdla/nvdla_scheduler.c | 1005 ++++
 drivers/gpu/drm/nvdla/nvdla_sdp.c       |  728 +++
 22 files changed, 13236 insertions(+)
 create mode 100644 drivers/gpu/drm/nvdla/Kconfig
 create mode 100644 drivers/gpu/drm/nvdla/Makefile
 create mode 100644 drivers/gpu/drm/nvdla/nvdla_bdma.c
 create mode 100644 drivers/gpu/drm/nvdla/nvdla_cache.c
 create mode 100644 drivers/gpu/drm/nvdla/nvdla_cdp.c
 create mode 100644 drivers/gpu/drm/nvdla/nvdla_common.c
 create mode 100644 drivers/gpu/drm/nvdla/nvdla_common.h
 create mode 100644 drivers/gpu/drm/nvdla/nvdla_conv.c
 create mode 100644 drivers/gpu/drm/nvdla/nvdla_drm.c
 create mode 100644 drivers/gpu/drm/nvdla/nvdla_drm.h
 create mode 100644 drivers/gpu/drm/nvdla/nvdla_engine.c
 create mode 100644 drivers/gpu/drm/nvdla/nvdla_engine.h
 create mode 100644 drivers/gpu/drm/nvdla/nvdla_gem.c
 create mode 100644 drivers/gpu/drm/nvdla/nvdla_ioctl.h
 create mode 100644 drivers/gpu/drm/nvdla/nvdla_pdp.c
 create mode 100644 drivers/gpu/drm/nvdla/nvdla_reg.h
 create mode 100644 drivers/gpu/drm/nvdla/nvdla_rubik.c
 create mode 100644 drivers/gpu/drm/nvdla/nvdla_sched.h
 create mode 100644 drivers/gpu/drm/nvdla/nvdla_scheduler.c
 create mode 100644 drivers/gpu/drm/nvdla/nvdla_sdp.c
drivers/gpu/drm/Kconfig +2 -0
index 5133c3f028ab..a55cff374abd 100644
@@ -409,6 +409,8 @@ source "drivers/gpu/drm/solomon/Kconfig"
 
 source "drivers/gpu/drm/sprd/Kconfig"
 
+source "drivers/gpu/drm/nvdla/Kconfig"
+
 config DRM_HYPERV
 	tristate "DRM Support for Hyper-V synthetic video device"
 	depends on DRM && PCI && MMU && HYPERV
drivers/gpu/drm/Makefile +1 -0
index c2ef5f9fce54..8fa3537f308a 100644
@@ -134,3 +134,4 @@ obj-y			+= gud/
 obj-$(CONFIG_DRM_HYPERV) += hyperv/
 obj-y			+= solomon/
 obj-$(CONFIG_DRM_SPRD) += sprd/
+obj-$(CONFIG_DRM_NVDLA) += nvdla/
drivers/gpu/drm/nvdla/Kconfig new file +8 -0
new file mode 100644
index 000000000000..11c04f5da877
@@ -0,0 +1,8 @@
+# SPDX-License-Identifier: GPL-2.0-only
+config DRM_NVDLA
+	tristate "NVDLA DRM"
+	depends on DRM
+	select DRM_GEM_CMA_HELPER
+	help
+	  Choose this option for open-source NVIDIA DLA support.
+	  If M is selected the module will be called nvdla-drm.
drivers/gpu/drm/nvdla/Makefile new file +19 -0
new file mode 100644
index 000000000000..74f37d258f8d
@@ -0,0 +1,19 @@
+
+# SPDX-License-Identifier: GPL-2.0
+nvdla-drm-y := \
+	nvdla_drm.o \
+	nvdla_gem.o \
+	nvdla_scheduler.o \
+	nvdla_engine.o \
+	nvdla_bdma.o \
+	nvdla_conv.o \
+	nvdla_sdp.o \
+	nvdla_cdp.o \
+	nvdla_pdp.o \
+	nvdla_rubik.o \
+	nvdla_cache.o \
+	nvdla_common.o \
+	nvdla_engine_data.o \
+	nvdla_engine_debug.o \
+
+obj-$(CONFIG_DRM_NVDLA) += nvdla-drm.o
drivers/gpu/drm/nvdla/nvdla_bdma.c new file +200 -0

200 lines not on this page · raw message

drivers/gpu/drm/nvdla/nvdla_cache.c new file +215 -0

215 lines not on this page · raw message

drivers/gpu/drm/nvdla/nvdla_cdp.c new file +300 -0

300 lines not on this page · raw message

drivers/gpu/drm/nvdla/nvdla_common.c new file +295 -0

295 lines not on this page · raw message

drivers/gpu/drm/nvdla/nvdla_common.h new file +835 -0

835 lines not on this page · raw message

drivers/gpu/drm/nvdla/nvdla_conv.c new file +683 -0

683 lines not on this page · raw message

drivers/gpu/drm/nvdla/nvdla_drm.c new file +695 -0

433 lines not on this page · raw message

drivers/gpu/drm/nvdla/nvdla_drm.h new file +127 -0

127 lines not on this page · raw message

drivers/gpu/drm/nvdla/nvdla_engine.c new file +233 -0

233 lines not on this page · raw message

drivers/gpu/drm/nvdla/nvdla_engine.h new file +272 -0

272 lines not on this page · raw message

drivers/gpu/drm/nvdla/nvdla_gem.c new file +393 -0

393 lines not on this page · raw message

drivers/gpu/drm/nvdla/nvdla_ioctl.h new file +99 -0

99 lines not on this page · raw message

drivers/gpu/drm/nvdla/nvdla_pdp.c new file +446 -0

446 lines not on this page · raw message

drivers/gpu/drm/nvdla/nvdla_reg.h new file +6411 -0

6,411 lines not on this page · raw message

drivers/gpu/drm/nvdla/nvdla_rubik.c new file +217 -0

217 lines not on this page · raw message

drivers/gpu/drm/nvdla/nvdla_sched.h new file +52 -0

52 lines not on this page · raw message

drivers/gpu/drm/nvdla/nvdla_scheduler.c new file +1005 -0

1,005 lines not on this page · raw message

drivers/gpu/drm/nvdla/nvdla_sdp.c new file +728 -0

728 lines not on this page · raw message

The rest of this diff is not shown. View the raw message.

Signature · 1 line
2.25.1

Part of the diff is not shown here. View the raw message.

#3 lore raw

Am 19.04.22 um 15:59 schrieb Cai Huoqing:

Quoted 6 lines

The NVIDIA Deep Learning Accelerator (NVDLA) is an open source IP which is integrated into NVIDIA Jetson AGX Xavier,
so add driver support for this accelerator.

Well doesn't looks so bad on first glance (regarding coding style etc..)

But am I blind or isn't there any UAPI for the driver? I mean adding a DRM driver without any change to include/uapi/drm is really odd.

Regards,
Christian.

#4 lore raw
Quoted 5 lines · Christian König

Am 19.04.22 um 15:59 schrieb Cai Huoqing:

Quoted 4 lines

The NVIDIA Deep Learning Accelerator (NVDLA) is an open source IP which is integrated into NVIDIA Jetson AGX Xavier,
so add driver support for this accelerator.

Quoted 7 lines
Quoted 2 lines

Well doesn't looks so bad on first glance (regarding coding style etc..)

But am I blind or isn't there any UAPI for the driver? I mean adding a DRM driver without any change to include/uapi/drm is really odd.

thanks for your reply, I will rename nvdla_ioctl.h which is UAPI headfile, then put it to include/uapi/drm.

thanks,
Cai

Quoted 3 lines

Regards,
Christian.

#5 lore raw
In reply to Cai Huoqing
I love your patch! Yet something to improve:

Hi Cai,

I love your patch! Yet something to improve:

[auto build test ERROR on drm/drm-next]
[also build test ERROR on drm-intel/for-linux-next drm-tip/drm-tip linus/master v5.18-rc3] [If your patch is applied to the wrong git tree, kindly drop us a note.
And when submitting patch, we suggest to use '--base' as documented in
https://git-scm.com/docs/git-format-patch]

url:    https://github.com/intel-lab-lkp/linux/commits/Cai-Huoqing/drm-nvdla-Add-driver-support-for-NVDLA/20220419-220255
base:   git://anongit.freedesktop.org/drm/drm drm-next
config: h8300-randconfig-r014-20220420 (https://download.01.org/0day-ci/archive/20220420/[email protected]/config)
compiler: h8300-linux-gcc (GCC) 11.2.0

reproduce (this is a W=1 build):
wget https://raw.githubusercontent.com/intel/lkp-tests/master/sbin/make.cross -O ~/bin/make.cross chmod +x ~/bin/make.cross
# https://github.com/intel-lab-lkp/linux/commit/7539e5487eb7d0c6f13c03bba596e51a2238106d
git remote add linux-review https://github.com/intel-lab-lkp/linux
git fetch --no-tags linux-review Cai-Huoqing/drm-nvdla-Add-driver-support-for-NVDLA/20220419-220255 git checkout 7539e5487eb7d0c6f13c03bba596e51a2238106d
# save the config file
mkdir build_dir && cp config build_dir/.config
COMPILER_INSTALL_PATH=$HOME/0day COMPILER=gcc-11.2.0 make.cross W=1 O=build_dir ARCH=h8300 SHELL=/bin/bash drivers/gpu/

If you fix the issue, kindly add following tag as appropriate

All error/warnings (new ones prefixed by >>):

>> make[5]: *** No rule to make target 'drivers/gpu/drm/nvdla/nvdla_engine_data.o', needed by 'drivers/gpu/drm/nvdla/built-in.a'.
>> make[5]: *** No rule to make target 'drivers/gpu/drm/nvdla/nvdla_engine_debug.o', needed by 'drivers/gpu/drm/nvdla/built-in.a'.
   make[5]: Target '__build' not remade because of errors.
--
>> drivers/gpu/drm/nvdla/nvdla_drm.c:45:9: warning: no previous prototype for 'dla_get_time_us' [-Wmissing-prototypes]
      45 | int64_t dla_get_time_us(void)
         |         ^~~~~~~~~~~~~~~
   drivers/gpu/drm/nvdla/nvdla_drm.c: In function 'nvdla_engine_isr':
>> drivers/gpu/drm/nvdla/nvdla_drm.c:75:18: warning: variable 'mask' set but not used [-Wunused-but-set-variable]
      75 |         uint32_t mask;
         |                  ^~~~
--
   drivers/gpu/drm/nvdla/nvdla_gem.c: In function 'nvdla_fill_task_desc':
>> drivers/gpu/drm/nvdla/nvdla_gem.c:39:17: warning: cast to pointer from integer of different size [-Wint-to-pointer-cast]
      39 |                 (void __user *)local_task->address_list,
         |                 ^
--
   drivers/gpu/drm/nvdla/nvdla_scheduler.c: In function 'dla_op_completion':
>> drivers/gpu/drm/nvdla/nvdla_scheduler.c:513:26: warning: variable 'task' set but not used [-Wunused-but-set-variable]
     513 |         struct dla_task *task;
         |                          ^~~~
--
>> drivers/gpu/drm/nvdla/nvdla_scheduler.c:363: warning: This comment starts with '/**', but isn't a kernel-doc comment. Refer Documentation/doc-guide/kernel-doc.rst

* Dequeue next operation of same type from list of operations
drivers/gpu/drm/nvdla/nvdla_scheduler.c:505: warning: This comment starts with '/**', but isn't a kernel-doc comment. Refer Documentation/doc-guide/kernel-doc.rst
* Handle operation completion notification
drivers/gpu/drm/nvdla/nvdla_scheduler.c:610: warning: This comment starts with '/**', but isn't a kernel-doc comment. Refer Documentation/doc-guide/kernel-doc.rst
* Read network configuration from DRAM, network descriptor address
drivers/gpu/drm/nvdla/nvdla_scheduler.c:920: warning: This comment starts with '/**', but isn't a kernel-doc comment. Refer Documentation/doc-guide/kernel-doc.rst
* Execute task selected by task scheduler
--
drivers/gpu/drm/nvdla/nvdla_bdma.c: In function 'dla_bdma_dump_config':

>> drivers/gpu/drm/nvdla/nvdla_bdma.c:157:39: warning: variable 'bdma_surface' set but not used [-Wunused-but-set-variable]
     157 |         struct dla_bdma_surface_desc *bdma_surface;
         |                                       ^~~~~~~~~~~~
>> drivers/gpu/drm/nvdla/nvdla_bdma.c:156:34: warning: variable 'bdma_op' set but not used [-Wunused-but-set-variable]
     156 |         struct dla_bdma_op_desc *bdma_op;
         |                                  ^~~~~~~
--
   drivers/gpu/drm/nvdla/nvdla_conv.c: In function 'dla_conv_dump_config':
>> drivers/gpu/drm/nvdla/nvdla_conv.c:666:39: warning: variable 'conv_surface' set but not used [-Wunused-but-set-variable]
     666 |         struct dla_conv_surface_desc *conv_surface;
         |                                       ^~~~~~~~~~~~
>> drivers/gpu/drm/nvdla/nvdla_conv.c:665:34: warning: variable 'conv_op' set but not used [-Wunused-but-set-variable]
     665 |         struct dla_conv_op_desc *conv_op;
         |                                  ^~~~~~~
--
>> drivers/gpu/drm/nvdla/nvdla_engine.c:67: warning: This comment starts with '/**', but isn't a kernel-doc comment. Refer Documentation/doc-guide/kernel-doc.rst

* Get DMA data cube address
drivers/gpu/drm/nvdla/nvdla_engine.c:88: warning: This comment starts with '/**', but isn't a kernel-doc comment. Refer Documentation/doc-guide/kernel-doc.rst
* Read input buffer address
--

>> drivers/gpu/drm/nvdla/nvdla_bdma.c:56: warning: This comment starts with '/**', but isn't a kernel-doc comment. Refer Documentation/doc-guide/kernel-doc.rst
    * Program BDMA slot for transfer
--
   drivers/gpu/drm/nvdla/nvdla_sdp.c: In function 'processor_sdp_program':
>> drivers/gpu/drm/nvdla/nvdla_sdp.c:190:18: warning: variable 'atom_size' set but not used [-Wunused-but-set-variable]
     190 |         uint32_t atom_size;
         |                  ^~~~~~~~~
   drivers/gpu/drm/nvdla/nvdla_sdp.c: In function 'dla_sdp_dump_config':
>> drivers/gpu/drm/nvdla/nvdla_sdp.c:708:38: warning: variable 'sdp_surface' set but not used [-Wunused-but-set-variable]
     708 |         struct dla_sdp_surface_desc *sdp_surface;
         |                                      ^~~~~~~~~~~
>> drivers/gpu/drm/nvdla/nvdla_sdp.c:707:33: warning: variable 'sdp_op' set but not used [-Wunused-but-set-variable]
     707 |         struct dla_sdp_op_desc *sdp_op;
         |                                 ^~~~~~
   At top level:
   drivers/gpu/drm/nvdla/nvdla_sdp.c:118:22: warning: 'map_perf_nan_inf' defined but not used [-Wunused-const-variable=]
     118 | static const uint8_t map_perf_nan_inf[] = {
         |                      ^~~~~~~~~~~~~~~~
   drivers/gpu/drm/nvdla/nvdla_sdp.c:113:22: warning: 'map_perf_sat' defined but not used [-Wunused-const-variable=]
     113 | static const uint8_t map_perf_sat[] = {
         |                      ^~~~~~~~~~~~
   drivers/gpu/drm/nvdla/nvdla_sdp.c:108:22: warning: 'map_perf_lut' defined but not used [-Wunused-const-variable=]
     108 | static const uint8_t map_perf_lut[] = {
         |                      ^~~~~~~~~~~~
   drivers/gpu/drm/nvdla/nvdla_sdp.c:103:22: warning: 'map_perf_dma' defined but not used [-Wunused-const-variable=]
     103 | static const uint8_t map_perf_dma[] = {
         |                      ^~~~~~~~~~~~
--
   drivers/gpu/drm/nvdla/nvdla_cdp.c: In function 'dla_cdp_dump_config':
>> drivers/gpu/drm/nvdla/nvdla_cdp.c:280:38: warning: variable 'cdp_surface' set but not used [-Wunused-but-set-variable]
     280 |         struct dla_cdp_surface_desc *cdp_surface;
         |                                      ^~~~~~~~~~~
>> drivers/gpu/drm/nvdla/nvdla_cdp.c:279:33: warning: variable 'cdp_op' set but not used [-Wunused-but-set-variable]
     279 |         struct dla_cdp_op_desc *cdp_op;
         |                                 ^~~~~~
   At top level:
   drivers/gpu/drm/nvdla/nvdla_cdp.c:28:22: warning: 'map_perf_lut' defined but not used [-Wunused-const-variable=]
      28 | static const uint8_t map_perf_lut[] = {
         |                      ^~~~~~~~~~~~
   drivers/gpu/drm/nvdla/nvdla_cdp.c:23:22: warning: 'map_perf_dma' defined but not used [-Wunused-const-variable=]
      23 | static const uint8_t map_perf_dma[] = {
         |                      ^~~~~~~~~~~~

..

vim +/dla_get_time_us +45 drivers/gpu/drm/nvdla/nvdla_drm.c

    44	
  > 45	int64_t dla_get_time_us(void)
    46	{
    47		return ktime_get_ns() / NSEC_PER_USEC;
    48	}
    49	
    50	void dla_reg_write(void *driver_context, uint32_t addr, uint32_t reg)
    51	{
    52		struct nvdla_device *nvdla_dev =
    53				(struct nvdla_device *)driver_context;
    54	
    55		if (!nvdla_dev)
    56			return;
    57	
    58		writel(reg, nvdla_dev->base + addr);
    59	}
    60	
    61	uint32_t dla_reg_read(void *driver_context, uint32_t addr)
    62	{
    63		struct nvdla_device *nvdla_dev =
    64				(struct nvdla_device *)driver_context;
    65	
    66		if (!nvdla_dev)
    67			return 0;
    68	
    69		return readl(nvdla_dev->base + addr);
    70	}
    71	
    72	static irqreturn_t nvdla_engine_isr(int32_t irq, void *data)
    73	{
    74		unsigned long flags;
  > 75		uint32_t mask;
    76		uint32_t reg;
    77		struct dla_processor *processor = NULL;
    78		struct dla_processor_group *group;
    79		struct dla_engine *engine;
    80		struct nvdla_device *nvdla_dev = (struct nvdla_device *)data;
    81	
    82		if (!nvdla_dev)
    83			return IRQ_NONE;
    84	
    85		engine = nvdla_dev->engine_context;
    86		spin_lock_irqsave(&nvdla_dev->nvdla_lock, flags);
    87	
    88		mask = glb_reg_read(engine, S_INTR_MASK);
    89		reg = glb_reg_read(engine, S_INTR_STATUS);
    90	
    91		if (reg & MASK(GLB_S_INTR_STATUS_0, CACC_DONE_STATUS0)) {
    92			processor = &engine->processors[DLA_OP_CONV];
    93			group = &processor->groups[0];
    94			group->events |= (1 << DLA_EVENT_OP_COMPLETED);
    95		}
    96		if (reg & MASK(GLB_S_INTR_STATUS_0, CACC_DONE_STATUS1)) {
    97			processor = &engine->processors[DLA_OP_CONV];
    98			group = &processor->groups[1];
    99			group->events |= (1 << DLA_EVENT_OP_COMPLETED);

100 }
101 if (reg & MASK(GLB_S_INTR_STATUS_0, SDP_DONE_STATUS0)) {
102 processor = &engine->processors[DLA_OP_SDP];
103 group = &processor->groups[0];
104 group->events |= (1 << DLA_EVENT_OP_COMPLETED);
105 }
106 if (reg & MASK(GLB_S_INTR_STATUS_0, SDP_DONE_STATUS1)) {
107 processor = &engine->processors[DLA_OP_SDP];
108 group = &processor->groups[1];
109 group->events |= (1 << DLA_EVENT_OP_COMPLETED);
110 }
111 if (reg & MASK(GLB_S_INTR_STATUS_0, CDP_DONE_STATUS0)) {
112 processor = &engine->processors[DLA_OP_CDP];
113 group = &processor->groups[0];
114 group->events |= (1 << DLA_EVENT_OP_COMPLETED);
115 }
116 if (reg & MASK(GLB_S_INTR_STATUS_0, CDP_DONE_STATUS1)) {
117 processor = &engine->processors[DLA_OP_CDP];
118 group = &processor->groups[1];
119 group->events |= (1 << DLA_EVENT_OP_COMPLETED);
120 }
121 if (reg & MASK(GLB_S_INTR_STATUS_0, RUBIK_DONE_STATUS0)) {
122 processor = &engine->processors[DLA_OP_RUBIK];
123 group = &processor->groups[0];
124 group->events |= (1 << DLA_EVENT_OP_COMPLETED);
125 }
126 if (reg & MASK(GLB_S_INTR_STATUS_0, RUBIK_DONE_STATUS1)) {
127 processor = &engine->processors[DLA_OP_RUBIK];
128 group = &processor->groups[1];
129 group->events |= (1 << DLA_EVENT_OP_COMPLETED);
130 }
131 if (reg & MASK(GLB_S_INTR_STATUS_0, PDP_DONE_STATUS0)) {
132 processor = &engine->processors[DLA_OP_PDP];
133 group = &processor->groups[0];
134 group->events |= (1 << DLA_EVENT_OP_COMPLETED);
135 }
136 if (reg & MASK(GLB_S_INTR_STATUS_0, PDP_DONE_STATUS1)) {
137 processor = &engine->processors[DLA_OP_PDP];
138 group = &processor->groups[1];
139 group->events |= (1 << DLA_EVENT_OP_COMPLETED);
140 }
141 if (reg & MASK(GLB_S_INTR_STATUS_0, BDMA_DONE_STATUS0)) {
142 processor = &engine->processors[DLA_OP_BDMA];
143 group = &processor->groups[0];
144 group->events |= (1 << DLA_EVENT_OP_COMPLETED);
145 }
146 if (reg & MASK(GLB_S_INTR_STATUS_0, BDMA_DONE_STATUS1)) {
147 processor = &engine->processors[DLA_OP_BDMA];
148 group = &processor->groups[1];
149 group->events |= (1 << DLA_EVENT_OP_COMPLETED);
150 }
151 if (reg & MASK(GLB_S_INTR_STATUS_0, CDMA_DAT_DONE_STATUS0)) {
152 processor = &engine->processors[DLA_OP_CONV];
153 group = &processor->groups[0];
154 group->events |= (1 << DLA_EVENT_CDMA_DT_DONE);
155 }
156 if (reg & MASK(GLB_S_INTR_STATUS_0, CDMA_DAT_DONE_STATUS1)) {
157 processor = &engine->processors[DLA_OP_CONV];
158 group = &processor->groups[1];
159 group->events |= (1 << DLA_EVENT_CDMA_DT_DONE);
160 }
161 if (reg & MASK(GLB_S_INTR_STATUS_0, CDMA_WT_DONE_STATUS0)) {
162 processor = &engine->processors[DLA_OP_CONV];
163 group = &processor->groups[0];
164 group->events |= (1 << DLA_EVENT_CDMA_WT_DONE);
165 }
166 if (reg & MASK(GLB_S_INTR_STATUS_0, CDMA_WT_DONE_STATUS1)) {
167 processor = &engine->processors[DLA_OP_CONV];
168 group = &processor->groups[1];
169 group->events |= (1 << DLA_EVENT_CDMA_WT_DONE);
170 }
171
172 glb_reg_write(engine, S_INTR_STATUS, reg);
173 mask = glb_reg_read(engine, S_INTR_MASK);
174 reg = glb_reg_read(engine, S_INTR_STATUS);
175
176 complete(&nvdla_dev->event_notifier);
177 spin_unlock_irqrestore(&nvdla_dev->nvdla_lock, flags);
178
179 return IRQ_HANDLED;
180 }
181

Signature · 2 lines
0-DAY CI Kernel Test Service
https://01.org/lkp
#6 lore raw
In reply to Cai Huoqing
I love your patch! Yet something to improve:

Hi Cai,

I love your patch! Yet something to improve:

[auto build test ERROR on drm/drm-next]
[also build test ERROR on drm-intel/for-linux-next drm-tip/drm-tip linus/master v5.18-rc3 next-20220419] [If your patch is applied to the wrong git tree, kindly drop us a note.
And when submitting patch, we suggest to use '--base' as documented in
https://git-scm.com/docs/git-format-patch]

url:    https://github.com/intel-lab-lkp/linux/commits/Cai-Huoqing/drm-nvdla-Add-driver-support-for-NVDLA/20220419-220255
base:   git://anongit.freedesktop.org/drm/drm drm-next
config: m68k-allmodconfig (https://download.01.org/0day-ci/archive/20220420/[email protected]/config)
compiler: m68k-linux-gcc (GCC) 11.2.0

reproduce (this is a W=1 build):
wget https://raw.githubusercontent.com/intel/lkp-tests/master/sbin/make.cross -O ~/bin/make.cross chmod +x ~/bin/make.cross
# https://github.com/intel-lab-lkp/linux/commit/7539e5487eb7d0c6f13c03bba596e51a2238106d
git remote add linux-review https://github.com/intel-lab-lkp/linux
git fetch --no-tags linux-review Cai-Huoqing/drm-nvdla-Add-driver-support-for-NVDLA/20220419-220255 git checkout 7539e5487eb7d0c6f13c03bba596e51a2238106d
# save the config file
mkdir build_dir && cp config build_dir/.config
COMPILER_INSTALL_PATH=$HOME/0day COMPILER=gcc-11.2.0 make.cross W=1 O=build_dir ARCH=m68k SHELL=/bin/bash drivers/gpu/

If you fix the issue, kindly add following tag as appropriate

All errors (new ones prefixed by >>):

>> make[5]: *** No rule to make target 'drivers/gpu/drm/nvdla/nvdla_engine_data.o', needed by 'drivers/gpu/drm/nvdla/nvdla-drm.o'.
>> make[5]: *** No rule to make target 'drivers/gpu/drm/nvdla/nvdla_engine_debug.o', needed by 'drivers/gpu/drm/nvdla/nvdla-drm.o'.

make[5]: Target '__build' not remade because of errors.

Signature · 2 lines
0-DAY CI Kernel Test Service
https://01.org/lkp
#7 lore raw
In reply to Cai Huoqing

Hi,

thanks for your submission. Some general comments:

* some functions are prefixed with dla_, others use nvdla_. It seems
arbitrary to me. Please use nvdla_ consistently throughout the source code.

* For reporting errors, please use drm_err(), drm_warn(), etc. I
suggest to rearrange the error messages to not be located in the innermost functions.

Please find more review comments below. It's not a full review, but at least something to start with.

Best regards
Thomas

Am 19.04.22 um 15:59 schrieb Cai Huoqing:

Quoted 99 lines

The NVIDIA Deep Learning Accelerator (NVDLA) is an open source IP which is integrated into NVIDIA Jetson AGX Xavier,
so add driver support for this accelerator.

---
  drivers/gpu/drm/Kconfig                 |    2 +
  drivers/gpu/drm/Makefile                |    1 +
  drivers/gpu/drm/nvdla/Kconfig           |    8 +
  drivers/gpu/drm/nvdla/Makefile          |   19 +
  drivers/gpu/drm/nvdla/nvdla_bdma.c      |  200 +
  drivers/gpu/drm/nvdla/nvdla_cache.c     |  215 +
  drivers/gpu/drm/nvdla/nvdla_cdp.c       |  300 ++
  drivers/gpu/drm/nvdla/nvdla_common.c    |  295 ++
  drivers/gpu/drm/nvdla/nvdla_common.h    |  835 +++
  drivers/gpu/drm/nvdla/nvdla_conv.c      |  683 +++
  drivers/gpu/drm/nvdla/nvdla_drm.c       |  695 +++
  drivers/gpu/drm/nvdla/nvdla_drm.h       |  127 +
  drivers/gpu/drm/nvdla/nvdla_engine.c    |  233 +
  drivers/gpu/drm/nvdla/nvdla_engine.h    |  272 +
  drivers/gpu/drm/nvdla/nvdla_gem.c       |  393 ++
  drivers/gpu/drm/nvdla/nvdla_ioctl.h     |   99 +
  drivers/gpu/drm/nvdla/nvdla_pdp.c       |  446 ++
  drivers/gpu/drm/nvdla/nvdla_reg.h       | 6411 +++++++++++++++++++++++
  drivers/gpu/drm/nvdla/nvdla_rubik.c     |  217 +
  drivers/gpu/drm/nvdla/nvdla_sched.h     |   52 +
  drivers/gpu/drm/nvdla/nvdla_scheduler.c | 1005 ++++
  drivers/gpu/drm/nvdla/nvdla_sdp.c       |  728 +++
  22 files changed, 13236 insertions(+)

create mode 100644 drivers/gpu/drm/nvdla/Kconfig create mode 100644 drivers/gpu/drm/nvdla/Makefile create mode 100644 drivers/gpu/drm/nvdla/nvdla_bdma.c create mode 100644 drivers/gpu/drm/nvdla/nvdla_cache.c create mode 100644 drivers/gpu/drm/nvdla/nvdla_cdp.c create mode 100644 drivers/gpu/drm/nvdla/nvdla_common.c create mode 100644 drivers/gpu/drm/nvdla/nvdla_common.h create mode 100644 drivers/gpu/drm/nvdla/nvdla_conv.c create mode 100644 drivers/gpu/drm/nvdla/nvdla_drm.c create mode 100644 drivers/gpu/drm/nvdla/nvdla_drm.h create mode 100644 drivers/gpu/drm/nvdla/nvdla_engine.c create mode 100644 drivers/gpu/drm/nvdla/nvdla_engine.h create mode 100644 drivers/gpu/drm/nvdla/nvdla_gem.c create mode 100644 drivers/gpu/drm/nvdla/nvdla_ioctl.h create mode 100644 drivers/gpu/drm/nvdla/nvdla_pdp.c create mode 100644 drivers/gpu/drm/nvdla/nvdla_reg.h create mode 100644 drivers/gpu/drm/nvdla/nvdla_rubik.c create mode 100644 drivers/gpu/drm/nvdla/nvdla_sched.h create mode 100644 drivers/gpu/drm/nvdla/nvdla_scheduler.c create mode 100644 drivers/gpu/drm/nvdla/nvdla_sdp.c

drivers/gpu/drm/Kconfig +2 -0
index 5133c3f028ab..a55cff374abd 100644
@@ -409,6 +409,8 @@ source "drivers/gpu/drm/solomon/Kconfig"
 
  source "drivers/gpu/drm/sprd/Kconfig"
 
+source "drivers/gpu/drm/nvdla/Kconfig"
+
  config DRM_HYPERV
  	tristate "DRM Support for Hyper-V synthetic video device"
  	depends on DRM && PCI && MMU && HYPERV
drivers/gpu/drm/Makefile +1 -0
index c2ef5f9fce54..8fa3537f308a 100644
@@ -134,3 +134,4 @@ obj-y			+= gud/
  obj-$(CONFIG_DRM_HYPERV) += hyperv/
  obj-y			+= solomon/
  obj-$(CONFIG_DRM_SPRD) += sprd/
+obj-$(CONFIG_DRM_NVDLA) += nvdla/
drivers/gpu/drm/nvdla/Kconfig new file +8 -0
new file mode 100644
index 000000000000..11c04f5da877
@@ -0,0 +1,8 @@
+# SPDX-License-Identifier: GPL-2.0-only
+config DRM_NVDLA
+	tristate "NVDLA DRM"
+	depends on DRM
+	select DRM_GEM_CMA_HELPER
+	help
+	  Choose this option for open-source NVIDIA DLA support.
+	  If M is selected the module will be called nvdla-drm.
drivers/gpu/drm/nvdla/Makefile new file +7 -0
new file mode 100644
index 000000000000..74f37d258f8d
@@ -0,0 +1,19 @@
+
+# SPDX-License-Identifier: GPL-2.0
+nvdla-drm-y := \
+	nvdla_drm.o \
+	nvdla_gem.o \
+	nvdla_scheduler.o \
+	nvdla_engine.o \

drivers/gpu/drm/nvdla/Makefile line 17

Quoted 10 lines
drivers/gpu/drm/nvdla/Makefile new file +10 -0
+	nvdla_bdma.o \
+	nvdla_conv.o \
+	nvdla_sdp.o \
+	nvdla_cdp.o \
+	nvdla_pdp.o \
+	nvdla_rubik.o \
+	nvdla_cache.o \
+	nvdla_common.o \
+	nvdla_engine_data.o \
+	nvdla_engine_debug.o \

File names should be sorted alphabetically here.

Quoted 2,824 lines
drivers/gpu/drm/nvdla/Makefile new file +2 -0
+
+obj-$(CONFIG_DRM_NVDLA) += nvdla-drm.o
drivers/gpu/drm/nvdla/nvdla_bdma.c new file +200 -0
new file mode 100644
index 000000000000..225613f27acf
@@ -0,0 +1,200 @@
+// SPDX-License-Identifier: BSD-3-Clause OR GPL-2.0
+/*
+ * Copyright (C) 2017-2018 NVIDIA CORPORATION
+ * Copyright (C) 2022 Cai Huoqing
+ */
+
+#include "nvdla_common.h"
+#include "nvdla_drm.h"
+#include "nvdla_reg.h"
+#include "nvdla_engine.h"
+
+static const uint8_t map_mem[] = {
+	FIELD_ENUM(BDMA_CFG_CMD_0, SRC_RAM_TYPE, MC),
+	FIELD_ENUM(BDMA_CFG_CMD_0, SRC_RAM_TYPE, CVSRAM),
+};
+
+void
+dla_bdma_set_producer(struct dla_engine *engine, int32_t group_id, int32_t rdma_group_id)
+{
+	/**
+	 * There is no producer bit for BDMA operation,
+	 * interrupt pointer decides which outstanding request
+	 * to use for this BDMA operation
+	 */
+}
+
+int
+dla_bdma_enable(struct dla_engine *engine, struct dla_processor_group *group)
+{
+	if (group->surface_desc->bdma_surface.num_transfers == (uint16_t)0) {
+		group->events |= ((uint8_t)1 << DLA_EVENT_OP_COMPLETED);
+		goto exit;
+	}
+
+	/**
+	 * Launch BDMA transfer
+	 */
+	if (group->id == 0)
+		bdma_reg_write(engine, CFG_LAUNCH0, FIELD_ENUM(BDMA_CFG_LAUNCH0_0,
+							GRP0_LAUNCH, YES));
+	else
+		bdma_reg_write(engine, CFG_LAUNCH1, FIELD_ENUM(BDMA_CFG_LAUNCH1_0,
+							GRP1_LAUNCH, YES));
+
+exit:
+	return 0;
+}
+
+void
+dla_bdma_rdma_check(struct dla_processor_group *group)
+{
+	group->is_rdma_needed = 0;
+}
+
+/**
+ * Program BDMA slot for transfer
+ */
+static int32_t
+processor_bdma_program_slot(struct dla_engine *engine,
+							struct dla_bdma_surface_desc *bdma_surface,
+							struct dla_bdma_transfer_desc *transfer)
+{
+	int32_t ret = 0;
+	uint64_t source_addr = 0;
+	uint64_t destination_addr = 0;
+	uint32_t high, low, reg;
+	uint8_t  bdma_free_slots = 0;
+
+	/* make sure there're enough free slots */
+	if (bdma_free_slots <= 0) {
+		do {
+			reg = bdma_reg_read(engine, STATUS);
+			reg = (reg & MASK(BDMA_STATUS_0, FREE_SLOT)) >>
+					SHIFT(BDMA_STATUS_0, FREE_SLOT);
+		} while (reg == 0);
+		bdma_free_slots = (uint8_t)reg;
+	}
+
+	dla_get_dma_address(engine->driver_context, engine->task->task_data,
+						transfer->source_address,
+						(void *)&source_addr,
+						DESTINATION_DMA);
+	dla_get_dma_address(engine->driver_context, engine->task->task_data,
+						transfer->destination_address,
+						(void *)&destination_addr,
+						DESTINATION_DMA);
+
+	ASSERT_GOTO((transfer->line_repeat <= 8192),
+				ret, -EINVAL, exit);
+	ASSERT_GOTO((transfer->surface_repeat <= 8192),
+				ret, -EINVAL, exit);
+	ASSERT_GOTO((transfer->line_size % 32) == 0,
+				ret, -EINVAL, exit);
+	ASSERT_GOTO(transfer->source_line >= transfer->line_size,
+				ret, -EINVAL, exit);
+	ASSERT_GOTO(transfer->destination_line >= transfer->line_size,
+				ret, -EINVAL, exit);
+	ASSERT_GOTO(transfer->source_surface >=
+			(transfer->source_line * transfer->line_repeat),
+				ret, -EINVAL, exit);
+	ASSERT_GOTO(transfer->destination_surface >=
+			(transfer->destination_line * transfer->line_repeat),
+				ret, -EINVAL, exit);
+
+	/* config registers */
+	high = upper_32_bits(source_addr);
+	low = lower_32_bits(source_addr);
+	bdma_reg_write(engine, CFG_SRC_ADDR_LOW, low);
+	bdma_reg_write(engine, CFG_SRC_ADDR_HIGH, high);
+	high = upper_32_bits(destination_addr);
+	low = lower_32_bits(destination_addr);
+	bdma_reg_write(engine, CFG_DST_ADDR_LOW, low);
+	bdma_reg_write(engine, CFG_DST_ADDR_HIGH, high);
+	bdma_reg_write(engine, CFG_LINE, (transfer->line_size >> 5) - 1);
+	reg = (map_mem[bdma_surface->source_type] <<
+				SHIFT(BDMA_CFG_CMD_0, SRC_RAM_TYPE)) |
+		(map_mem[bdma_surface->destination_type] <<
+				SHIFT(BDMA_CFG_CMD_0, DST_RAM_TYPE));
+	bdma_reg_write(engine, CFG_CMD, reg);
+	bdma_reg_write(engine, CFG_LINE_REPEAT, transfer->line_repeat - 1);
+	bdma_reg_write(engine, CFG_SRC_LINE, transfer->source_line);
+	bdma_reg_write(engine, CFG_DST_LINE, transfer->destination_line);
+	bdma_reg_write(engine, CFG_SURF_REPEAT, transfer->surface_repeat - 1);
+	bdma_reg_write(engine, CFG_SRC_SURF, transfer->source_surface);
+	bdma_reg_write(engine, CFG_DST_SURF, transfer->destination_surface);
+	bdma_reg_write(engine, CFG_OP, FIELD_ENUM(BDMA_CFG_OP_0, EN, ENABLE));
+
+exit:
+	return ret;
+}
+
+int
+dla_bdma_is_ready(struct dla_processor *processor,
+				  struct dla_processor_group *group)
+{
+	struct dla_processor_group *next_group;
+
+	next_group = &processor->groups[!group->id];
+
+	/**
+	 * If another group is already programmed but not active then
+	 * do not program this operation as BDMA does not really
+	 * have shadow copies for groups. It will end programming
+	 * same group. Wait for another group to get enabled.
+	 */
+	if ((processor->group_status & (1 << next_group->id)) &&
+						!next_group->active)
+		return 0;
+
+	return 1;
+}
+
+void
+dla_bdma_dump_config(struct dla_processor_group *group)
+{
+	struct dla_bdma_op_desc *bdma_op;
+	struct dla_bdma_surface_desc *bdma_surface;
+
+	bdma_surface = &group->surface_desc->bdma_surface;
+	bdma_op = &group->operation_desc->bdma_op;
+}
+
+int
+dla_bdma_program(struct dla_engine *engine, struct dla_processor_group *group)
+{
+	int32_t i;
+	int32_t ret = 0;
+	struct dla_bdma_surface_desc *bdma_surface;
+
+	if (!engine->config_data->bdma_enable) {
+		pr_err("BDMA is not supported for this configuration\n");
+		ret = -EINVAL;
+		goto exit;
+	}
+
+	bdma_surface = &group->surface_desc->bdma_surface;
+
+	pr_debug("Num of transfers %u\n", bdma_surface->num_transfers);
+	if (bdma_surface->num_transfers == (uint16_t)0)
+		goto exit;
+
+	if (bdma_surface->num_transfers > NUM_MAX_BDMA_OPS) {
+		pr_err("Invalid number of transfers\n");
+		ret = -EINVAL;
+		goto exit;
+	}
+
+	for (i = 0; i < bdma_surface->num_transfers; i++) {
+		ret = processor_bdma_program_slot(engine, bdma_surface,
+					&bdma_surface->transfers[i]);
+		if (ret)
+			goto exit;
+	}
+
+	dla_enable_intr(engine, MASK(GLB_S_INTR_MASK_0, BDMA_DONE_MASK1) |
+			MASK(GLB_S_INTR_MASK_0, BDMA_DONE_MASK0));
+
+exit:
+	return ret;
+}
drivers/gpu/drm/nvdla/nvdla_cache.c new file +215 -0
new file mode 100644
index 000000000000..f8bd7b514aab
@@ -0,0 +1,215 @@
+// SPDX-License-Identifier: BSD-3-Clause OR GPL-2.0
+/*
+ * Copyright (C) 2017-2018 NVIDIA CORPORATION
+ * Copyright (C) 2022 Cai Huoqing
+ */
+
+#include "nvdla_common.h"
+#include "nvdla_drm.h"
+#include "nvdla_reg.h"
+#include "nvdla_engine.h"
+
+#define DLA_OP_CACHE_SIZE (DLA_NUM_GROUPS * ((DLA_OP_NUM + 2) * 2))
+
+static struct dla_common_op_desc desc_cache[DLA_OP_NUM][DLA_OP_CACHE_SIZE];
+static int32_t desc_refcount[DLA_OP_NUM][DLA_OP_CACHE_SIZE];
+
+void
+dla_get_refcount(struct dla_common_op_desc *op_desc)
+{
+	int32_t i;
+	struct dla_common_op_desc *desc = NULL;
+
+	if (op_desc == NULL)
+		return;
+
+	if (op_desc->index == -1)
+		return;
+
+	desc = &desc_cache[op_desc->op_type][0];
+
+	for (i = 0; i < DLA_OP_CACHE_SIZE; i++, desc++) {
+		if (desc->index == op_desc->index &&
+				desc->roi_index == op_desc->roi_index) {
+			desc_refcount[op_desc->op_type][i]++;
+			return;
+		}
+	}
+}
+
+struct dla_common_op_desc *
+dla_get_op_desc(struct dla_engine *engine,
+				struct dla_task *task, int16_t index,
+				uint8_t op_type, uint8_t roi_index)
+{
+	int32_t i;
+	int32_t ret;
+	uint64_t op_base;
+	uint64_t dep_graph_addr;
+	struct dla_common_op_desc *desc = NULL;
+
+	if (index == -1) {
+		pr_debug("no desc get due to index==-1\n");
+		goto exit;
+	}
+
+	dep_graph_addr = (sizeof(struct dla_common_op_desc) *
+				engine->network->num_operations * roi_index);
+
+	desc = &desc_cache[op_type][0];
+
+	for (i = 0; i < DLA_OP_CACHE_SIZE; i++, desc++) {
+		if (desc->index == index && desc->roi_index == roi_index) {
+			if (desc->op_type != op_type) {
+				pr_err("op_cache[op=%u] contains incorrect entry of op[%u]\n",
+					   op_type, desc->op_type);
+				continue;
+			}
+			desc_refcount[op_type][i]++;
+			goto exit;
+		}
+	}
+
+	desc = &desc_cache[op_type][0];
+
+	for (i = 0; i < DLA_OP_CACHE_SIZE; i++, desc++) {
+		if (desc->index == -1) {
+			op_base = dep_graph_addr +
+					(sizeof(struct dla_common_op_desc) *
+					(uint64_t)index);
+			ret = dla_data_read(engine->driver_context,
+					task->task_data,
+					task->dependency_graph_addr,
+					(void *)(desc),
+					sizeof(struct dla_common_op_desc),
+					op_base);
+			if (ret) {
+				desc = NULL;
+				goto exit;
+			}
+
+			if (op_type != desc->op_type) {
+				/*
+				 * op_type of entry read from DRAM should not
+				 * mismatch with given op_type. If they
+				 * mismatches, then wrong entry is fetched, so
+				 * report this issue by throwing error.
+				 */
+				pr_err("Fetched [op_type=%u] from DRAM doesn't match with op_type[%u]\n",
+					   desc->op_type, op_type);
+				desc->op_type = op_type;
+				desc->index = -1;
+				desc->roi_index = -1;
+				desc = NULL;
+				goto exit;
+			}
+
+			desc->index = index;
+			desc->roi_index = roi_index;
+
+			desc_refcount[op_type][i]++;
+			goto exit;
+		}
+	}
+
+exit:
+	return desc;
+}
+
+static void
+dla_free_op_desc(struct dla_engine *engine, struct dla_common_op_desc *op_desc)
+{
+	uint64_t op_base;
+	uint64_t dep_graph_addr;
+	struct dla_task *task;
+
+	pr_debug("Enter: %s op desc index %u ROI %d\n", __func__,
+				op_desc->index, op_desc->roi_index);
+
+	task = engine->task;
+	dep_graph_addr = (sizeof(struct dla_common_op_desc) *
+				engine->network->num_operations *
+				op_desc->roi_index);
+
+	if (op_desc->index == -1)
+		goto exit;
+
+	if (op_desc == NULL)
+		goto exit;
+
+	/**
+	 * TODO: keeping the depth value hardcoded as 0 for now,
+	 * need to replace it once corresponding implementation is done.
+	 */
+	op_base = (dep_graph_addr +
+			(sizeof(struct dla_common_op_desc) *
+			(uint64_t)op_desc->index));
+
+	/**
+	 * Flush descriptor to DRAM
+	 */
+	dla_data_write(engine->driver_context,
+			task->task_data,
+			(void *)op_desc,
+			task->dependency_graph_addr,
+			sizeof(struct dla_common_op_desc),
+			op_base);
+
+	/**
+	 * Release it
+	 */
+	op_desc->index = -1;
+	op_desc->roi_index = -1;
+exit:
+	return;
+}
+
+void
+dla_put_op_desc(struct dla_engine *engine, struct dla_common_op_desc *op_desc)
+{
+	int32_t i;
+	struct dla_common_op_desc *desc;
+
+	if (op_desc == NULL)
+		return;
+
+	if (op_desc->index == -1)
+		return;
+
+	desc = &desc_cache[op_desc->op_type][0];
+
+	for (i = 0; i < DLA_OP_CACHE_SIZE; i++, desc++) {
+		if (desc->index == op_desc->index &&
+				desc->roi_index == op_desc->roi_index) {
+
+			desc_refcount[op_desc->op_type][i]--;
+
+			/**
+			 * Free desc if refcount is 0
+			 */
+			if (desc_refcount[op_desc->op_type][i] == 0)
+				dla_free_op_desc(engine, op_desc);
+
+			return;
+		}
+	}
+}
+
+void
+dla_init_op_cache(struct dla_engine *engine)
+{
+	int32_t i, j;
+	struct dla_common_op_desc *desc = &desc_cache[0][0];
+
+	memset((uint8_t *)&desc_cache[0][0], 0, sizeof(desc_cache));
+	memset((uint8_t *)&desc_refcount[0][0], 0, sizeof(desc_refcount));
+
+	for (i = 0; i < DLA_OP_NUM; i++) {
+		for (j = 0; j < DLA_OP_CACHE_SIZE; j++) {
+			desc->index = -1;
+			desc->roi_index = -1;
+			desc->op_type = (uint8_t)i;
+			desc++;
+		}
+	}
+}
drivers/gpu/drm/nvdla/nvdla_cdp.c new file +300 -0
new file mode 100644
index 000000000000..793161e50b1c
@@ -0,0 +1,300 @@
+// SPDX-License-Identifier: BSD-3-Clause OR GPL-2.0
+/*
+ * Copyright (C) 2017-2018 NVIDIA CORPORATION
+ * Copyright (C) 2022 Cai Huoqing
+ */
+
+#include "nvdla_common.h"
+#include "nvdla_drm.h"
+#include "nvdla_reg.h"
+#include "nvdla_engine.h"
+
+static const uint8_t map_ram[] = {
+	FIELD_ENUM(CDP_RDMA_D_SRC_DMA_CFG_0, SRC_RAM_TYPE, MC),
+	FIELD_ENUM(CDP_RDMA_D_SRC_DMA_CFG_0, SRC_RAM_TYPE, CV),
+};
+
+static const uint8_t map_precision[] = {
+	FIELD_ENUM(CDP_RDMA_D_DATA_FORMAT_0, INPUT_DATA, INT8),
+	FIELD_ENUM(CDP_RDMA_D_DATA_FORMAT_0, INPUT_DATA, INT16),
+	FIELD_ENUM(CDP_RDMA_D_DATA_FORMAT_0, INPUT_DATA, FP16),
+};
+
+static const uint8_t map_perf_dma[] = {
+	FIELD_ENUM(CDP_D_PERF_ENABLE_0, DMA_EN, DISABLE),
+	FIELD_ENUM(CDP_D_PERF_ENABLE_0, DMA_EN, ENABLE),
+};
+
+static const uint8_t map_perf_lut[] = {
+	FIELD_ENUM(CDP_D_PERF_ENABLE_0, LUT_EN, DISABLE),
+	FIELD_ENUM(CDP_D_PERF_ENABLE_0, LUT_EN, ENABLE),
+};
+
+
+static uint32_t
+map_local_size(uint8_t local_size)
+{
+	return ((local_size-1)/2)-1;
+}
+
+void
+dla_cdp_set_producer(struct dla_engine *engine, int32_t group_id, int32_t rdma_group_id)
+{
+	uint32_t reg;
+
+	/**
+	 * set producer pointer for all sub-modules
+	 */
+	reg = group_id << SHIFT(CDP_S_POINTER_0, PRODUCER);
+	cdp_reg_write(engine, S_POINTER, reg);
+	reg = group_id << SHIFT(CDP_RDMA_S_POINTER_0, PRODUCER);
+	cdp_rdma_reg_write(engine, S_POINTER, reg);
+}
+
+int
+dla_cdp_enable(struct dla_engine *engine, struct dla_processor_group *group)
+{
+	uint32_t reg;
+
+	/**
+	 * enable all sub-modules
+	 */
+	reg = FIELD_ENUM(CDP_RDMA_D_OP_ENABLE_0, OP_EN, ENABLE);
+	cdp_rdma_reg_write(engine, D_OP_ENABLE, reg);
+	reg = FIELD_ENUM(CDP_D_OP_ENABLE_0, OP_EN, ENABLE);
+	cdp_reg_write(engine, D_OP_ENABLE, reg);
+
+	return 0;
+}
+
+void
+dla_cdp_rdma_check(struct dla_processor_group *group)
+{
+	group->is_rdma_needed = 1;
+}
+
+static int32_t
+processor_cdp_program(struct dla_engine *engine, struct dla_processor_group *group)
+{
+	int32_t ret = 0;
+	uint32_t reg, high, low;
+	uint64_t input_address = 0;
+	uint64_t output_address = 0;
+	struct dla_lut_param lut;
+	struct dla_cdp_op_desc *cdp_op;
+	struct dla_cdp_surface_desc *cdp_surface;
+
+	cdp_op = &group->operation_desc->cdp_op;
+	cdp_surface = &group->surface_desc->cdp_surface;
+
+	/* Argument check */
+	if (cdp_surface->src_data.type == DLA_MEM_HW) {
+		pr_err("Invalid source memory type\n");
+		ret = -EINVAL;
+		goto exit;
+	}
+	if (cdp_surface->dst_data.type == DLA_MEM_HW) {
+		pr_err("Invalid destination memory type\n");
+		ret = -EINVAL;
+		goto exit;
+	}
+
+	if (cdp_op->in_precision != cdp_op->out_precision) {
+		pr_err("CDP does not support precision conversion\n");
+		ret = -EINVAL;
+		goto exit;
+	}
+
+	/* get the addresses from task descriptor */
+	ret = dla_read_input_address(engine, &cdp_surface->src_data,
+								 &input_address,
+								 group->op_desc->index,
+								 group->roi_index, 1);
+	if (ret)
+		goto exit;
+
+	dla_get_dma_cube_address(engine->driver_context,
+							 engine->task->task_data,
+							 cdp_surface->dst_data.address,
+							 cdp_surface->dst_data.offset,
+							 (void *)&output_address,
+							 DESTINATION_DMA);
+	if (cdp_op->lut_index >= 0) {
+		group->lut_index = cdp_op->lut_index;
+		dla_read_lut(engine, cdp_op->lut_index, (void *)&lut);
+	}
+
+	/* config CDP RDMA registers */
+	reg = ((cdp_surface->src_data.width - 1)
+		<< SHIFT(CDP_RDMA_D_DATA_CUBE_WIDTH_0, WIDTH));
+	cdp_rdma_reg_write(engine, D_DATA_CUBE_WIDTH, reg);
+
+	reg = ((cdp_surface->src_data.height - 1)
+		<< SHIFT(CDP_RDMA_D_DATA_CUBE_HEIGHT_0, HEIGHT));
+	cdp_rdma_reg_write(engine, D_DATA_CUBE_HEIGHT, reg);
+
+	reg = ((cdp_surface->src_data.channel - 1)
+		<< SHIFT(CDP_RDMA_D_DATA_CUBE_CHANNEL_0, CHANNEL));
+	cdp_rdma_reg_write(engine, D_DATA_CUBE_CHANNEL, reg);
+
+	high = upper_32_bits(input_address);
+	low = lower_32_bits(input_address);
+	cdp_rdma_reg_write(engine, D_SRC_BASE_ADDR_LOW, low);
+	cdp_rdma_reg_write(engine, D_SRC_BASE_ADDR_HIGH, high);
+
+	cdp_rdma_reg_write(engine, D_SRC_LINE_STRIDE,
+			cdp_surface->src_data.line_stride);
+	cdp_rdma_reg_write(engine, D_SRC_SURFACE_STRIDE,
+			cdp_surface->src_data.surf_stride);
+
+	reg = (map_ram[cdp_surface->src_data.type]
+		<< SHIFT(CDP_RDMA_D_SRC_DMA_CFG_0, SRC_RAM_TYPE));
+	cdp_rdma_reg_write(engine, D_SRC_DMA_CFG, reg);
+
+	reg = (map_precision[cdp_op->in_precision]
+		<< SHIFT(CDP_RDMA_D_DATA_FORMAT_0, INPUT_DATA));
+	cdp_rdma_reg_write(engine, D_DATA_FORMAT, reg);
+
+	/* config CDP */
+	if (cdp_op->lut_index >= 0)
+		update_lut(engine, CDP_S_LUT_ACCESS_CFG_0, &lut, cdp_op->in_precision);
+
+	high = upper_32_bits(output_address);
+	low = lower_32_bits(output_address);
+	cdp_reg_write(engine, D_DST_BASE_ADDR_LOW, low);
+	cdp_reg_write(engine, D_DST_BASE_ADDR_HIGH, high);
+
+	cdp_reg_write(engine, D_DST_LINE_STRIDE, cdp_surface->dst_data.line_stride);
+	cdp_reg_write(engine, D_DST_SURFACE_STRIDE, cdp_surface->dst_data.surf_stride);
+
+	reg = (map_ram[cdp_surface->dst_data.type]
+		<< SHIFT(CDP_D_DST_DMA_CFG_0, DST_RAM_TYPE));
+	cdp_reg_write(engine, D_DST_DMA_CFG, reg);
+
+	reg = (map_precision[cdp_op->in_precision]
+		<< SHIFT(CDP_D_DATA_FORMAT_0, INPUT_DATA_TYPE));
+	cdp_reg_write(engine, D_DATA_FORMAT, reg);
+
+	reg = (map_local_size(cdp_op->local_size)
+		<< SHIFT(CDP_D_LRN_CFG_0, NORMALZ_LEN));
+	cdp_reg_write(engine, D_LRN_CFG, reg);
+
+	reg = (cdp_op->in_cvt.offset
+		<< SHIFT(CDP_D_DATIN_OFFSET_0, DATIN_OFFSET));
+	cdp_reg_write(engine, D_DATIN_OFFSET, reg);
+
+	reg = (cdp_op->in_cvt.scale
+		<< SHIFT(CDP_D_DATIN_SCALE_0, DATIN_SCALE));
+	cdp_reg_write(engine, D_DATIN_SCALE, reg);
+
+	reg = (cdp_op->in_cvt.truncate
+		<< SHIFT(CDP_D_DATIN_SHIFTER_0, DATIN_SHIFTER));
+	cdp_reg_write(engine, D_DATIN_SHIFTER, reg);
+
+	reg = (cdp_op->out_cvt.offset
+		<< SHIFT(CDP_D_DATOUT_OFFSET_0, DATOUT_OFFSET));
+	cdp_reg_write(engine, D_DATOUT_OFFSET, reg);
+
+	reg = (cdp_op->out_cvt.scale
+		<< SHIFT(CDP_D_DATOUT_SCALE_0, DATOUT_SCALE));
+	cdp_reg_write(engine, D_DATOUT_SCALE, reg);
+
+	reg = (cdp_op->out_cvt.truncate
+		<< SHIFT(CDP_D_DATOUT_SHIFTER_0, DATOUT_SHIFTER));
+	cdp_reg_write(engine, D_DATOUT_SHIFTER, reg);
+
+	reg = ((cdp_op->bypass_sqsum ?
+		FIELD_ENUM(CDP_D_FUNC_BYPASS_0, SQSUM_BYPASS, ENABLE) :
+		FIELD_ENUM(CDP_D_FUNC_BYPASS_0, SQSUM_BYPASS, DISABLE)) <<
+		SHIFT(CDP_D_FUNC_BYPASS_0, SQSUM_BYPASS)) |
+		((cdp_op->bypass_out_mul ?
+		FIELD_ENUM(CDP_D_FUNC_BYPASS_0, MUL_BYPASS, ENABLE) :
+		FIELD_ENUM(CDP_D_FUNC_BYPASS_0, MUL_BYPASS, DISABLE)) <<
+		SHIFT(CDP_D_FUNC_BYPASS_0, MUL_BYPASS));
+	cdp_reg_write(engine, D_FUNC_BYPASS, reg);
+
+exit:
+	return ret;
+}
+
+int
+dla_cdp_is_ready(struct dla_processor *processor,
+				 struct dla_processor_group *group)
+{
+	struct dla_processor_group *next_group;
+	struct dla_cdp_op_desc *cdp_op;
+
+	cdp_op = &group->operation_desc->cdp_op;
+	next_group = &processor->groups[!group->id];
+
+	/**
+	 * Single LUT is shared between two CDP groups, need to make
+	 * sure that usage does not conflict. Also, LUT write
+	 * access is locked when CDP sub-engine is active, so delay
+	 * writing LUT when another group is active.
+	 */
+
+	/**
+	 * if no LUT required for current group then it can be programmed
+	 * without further checks
+	 */
+	if (cdp_op->lut_index == -1)
+		return 1;
+
+	/**
+	 * if same LUT is used for both groups then it can be programmed
+	 * without more checks. Even if another group is active and LUT
+	 * is locked, it would have been programmed by another group.
+	 */
+	if (next_group->lut_index == cdp_op->lut_index)
+		return 1;
+
+	/**
+	 * if LUT index of another group is not -1 means some LUT is programmed,
+	 * then do not program current LUT as we already know current LUT is not
+	 * -1 and neither same as another group.
+	 */
+	if (next_group->lut_index != -1)
+		return 0;
+
+	/**
+	 * if current group needs LUT different than another group and that
+	 * group is not active then program it.
+	 */
+	if (!next_group->active)
+		return 1;
+
+	/**
+	 * if control is here it means current group is using LUT different than
+	 * another group and that group is active. Wait for another group to
+	 * become idle.
+	 */
+
+	return 0;
+}
+
+void
+dla_cdp_dump_config(struct dla_processor_group *group)
+{
+	struct dla_cdp_op_desc *cdp_op;
+	struct dla_cdp_surface_desc *cdp_surface;
+
+	cdp_surface = &group->surface_desc->cdp_surface;
+	cdp_op = &group->operation_desc->cdp_op;
+}
+
+int
+dla_cdp_program(struct dla_engine *engine, struct dla_processor_group *group)
+{
+	int32_t ret;
+
+	dla_enable_intr(engine, MASK(GLB_S_INTR_MASK_0, CDP_DONE_MASK1) |
+					MASK(GLB_S_INTR_MASK_0, CDP_DONE_MASK0));
+
+	ret = processor_cdp_program(engine, group);
+	if (ret)
+		goto exit;
+
+exit:
+	return ret;
+}
drivers/gpu/drm/nvdla/nvdla_common.c new file +295 -0
new file mode 100644
index 000000000000..3abc863d9fce
@@ -0,0 +1,295 @@
+// SPDX-License-Identifier: BSD-3-Clause OR GPL-2.0
+/*
+ * Copyright (C) 2017-2018 NVIDIA CORPORATION
+ * Copyright (C) 2022 Cai Huoqing
+ */
+
+#include "nvdla_reg.h"
+#include "nvdla_common.h"
+#include "nvdla_drm.h"
+#include "nvdla_engine.h"
+
+static const uint8_t map_lut_method[] = {
+	FIELD_ENUM(CDP_S_LUT_CFG_0, LUT_LE_FUNCTION, EXPONENT),
+	FIELD_ENUM(CDP_S_LUT_CFG_0, LUT_LE_FUNCTION, LINEAR)
+};
+static const uint8_t map_lut_out[] = {
+	FIELD_ENUM(CDP_S_LUT_CFG_0, LUT_UFLOW_PRIORITY, LE),
+	FIELD_ENUM(CDP_S_LUT_CFG_0, LUT_UFLOW_PRIORITY, LO)
+};
+
+static const uint16_t access_data_offset[] = {
+	CDP_S_LUT_ACCESS_DATA_0 - CDP_S_LUT_ACCESS_CFG_0,
+	SDP_S_LUT_ACCESS_DATA_0 - SDP_S_LUT_ACCESS_CFG_0,
+};
+static const uint16_t lut_cfg_offset[] = {
+	CDP_S_LUT_CFG_0 - CDP_S_LUT_ACCESS_CFG_0,
+	SDP_S_LUT_CFG_0 - SDP_S_LUT_ACCESS_CFG_0,
+};
+static const uint16_t lut_info_offset[] = {
+	CDP_S_LUT_INFO_0 - CDP_S_LUT_ACCESS_CFG_0,
+	SDP_S_LUT_INFO_0 - SDP_S_LUT_ACCESS_CFG_0,
+};
+static const uint16_t le_start_offset[] = {
+	CDP_S_LUT_LE_START_LOW_0 - CDP_S_LUT_ACCESS_CFG_0,
+	SDP_S_LUT_LE_START_0 - SDP_S_LUT_ACCESS_CFG_0,
+};
+static const uint16_t le_end_offset[] = {
+	CDP_S_LUT_LE_END_LOW_0 - CDP_S_LUT_ACCESS_CFG_0,
+	SDP_S_LUT_LE_END_0 - SDP_S_LUT_ACCESS_CFG_0,
+};
+static const uint16_t lo_start_offset[] = {
+	CDP_S_LUT_LO_START_LOW_0 - CDP_S_LUT_ACCESS_CFG_0,
+	SDP_S_LUT_LO_START_0 - SDP_S_LUT_ACCESS_CFG_0,
+};
+static const uint16_t lo_end_offset[] = {
+	CDP_S_LUT_LO_END_LOW_0 - CDP_S_LUT_ACCESS_CFG_0,
+	SDP_S_LUT_LO_END_0 - SDP_S_LUT_ACCESS_CFG_0,
+};
+static const uint16_t le_slope_scale_offset[] = {
+	CDP_S_LUT_LE_SLOPE_SCALE_0 - CDP_S_LUT_ACCESS_CFG_0,
+	SDP_S_LUT_LE_SLOPE_SCALE_0 - SDP_S_LUT_ACCESS_CFG_0,
+};
+static const uint16_t le_slope_shift_offset[] = {
+	CDP_S_LUT_LE_SLOPE_SHIFT_0 - CDP_S_LUT_ACCESS_CFG_0,
+	SDP_S_LUT_LE_SLOPE_SHIFT_0 - SDP_S_LUT_ACCESS_CFG_0,
+};
+static const uint16_t lo_slope_scale_offset[] = {
+	CDP_S_LUT_LO_SLOPE_SCALE_0 - CDP_S_LUT_ACCESS_CFG_0,
+	SDP_S_LUT_LO_SLOPE_SCALE_0 - SDP_S_LUT_ACCESS_CFG_0,
+};
+static const uint16_t lo_slope_shift_offset[] = {
+	CDP_S_LUT_LO_SLOPE_SHIFT_0 - CDP_S_LUT_ACCESS_CFG_0,
+	SDP_S_LUT_LO_SLOPE_SHIFT_0 - SDP_S_LUT_ACCESS_CFG_0,
+};
+
+void update_lut(struct dla_engine *engine, uint32_t reg_base,
+				struct dla_lut_param *lut, uint8_t precision)
+{
+	int32_t i;
+	uint32_t reg;
+	uint32_t high, low;
+	int32_t is_sdp = reg_base == SDP_S_LUT_ACCESS_CFG_0;
+
+	/* program raw table */
+	reg = (FIELD_ENUM(CDP_S_LUT_ACCESS_CFG_0, LUT_TABLE_ID, LE)
+		<< SHIFT(CDP_S_LUT_ACCESS_CFG_0, LUT_TABLE_ID)) |
+		(FIELD_ENUM(CDP_S_LUT_ACCESS_CFG_0, LUT_ACCESS_TYPE, WRITE)
+		<< SHIFT(CDP_S_LUT_ACCESS_CFG_0, LUT_ACCESS_TYPE));
+	reg_write(engine, reg_base, reg);
+
+	for (i = 0; i < (1<<LUT_LINEAR_EXP_TABLE_ENTRY_LOG2)+1; i++) {
+		dla_reg_write(engine->driver_context,
+				reg_base + access_data_offset[is_sdp],
+				lut->linear_exp_table[i]);
+	}
+
+	/* program density table */
+	reg = (FIELD_ENUM(CDP_S_LUT_ACCESS_CFG_0, LUT_TABLE_ID, LO)
+		<< SHIFT(CDP_S_LUT_ACCESS_CFG_0, LUT_TABLE_ID)) |
+		(FIELD_ENUM(CDP_S_LUT_ACCESS_CFG_0, LUT_ACCESS_TYPE, WRITE)
+		<< SHIFT(CDP_S_LUT_ACCESS_CFG_0, LUT_ACCESS_TYPE));
+	dla_reg_write(engine->driver_context, reg_base, reg);
+
+	for (i = 0; i < (1<<LUT_LINEAR_ONLY_TABLE_ENTRY_LOG2)+1; i++) {
+		dla_reg_write(engine->driver_context,
+				reg_base + access_data_offset[is_sdp],
+				lut->linear_only_table[i]);
+	}
+
+	/* program other configurations */
+	reg = (map_lut_method[lut->method] <<
+		SHIFT(CDP_S_LUT_CFG_0, LUT_LE_FUNCTION)) |
+		(map_lut_out[lut->hybrid_priority] <<
+		SHIFT(CDP_S_LUT_CFG_0, LUT_HYBRID_PRIORITY)) |
+		(map_lut_out[lut->underflow_priority] <<
+		SHIFT(CDP_S_LUT_CFG_0, LUT_UFLOW_PRIORITY)) |
+		(map_lut_out[lut->overflow_priority] <<
+		SHIFT(CDP_S_LUT_CFG_0, LUT_OFLOW_PRIORITY));
+	dla_reg_write(engine->driver_context,
+			reg_base + lut_cfg_offset[is_sdp], reg);
+
+	if (lut->method == FIELD_ENUM(CDP_S_LUT_CFG_0,
+					LUT_LE_FUNCTION, EXPONENT)) {
+		reg = ((((uint32_t)lut->linear_exp_offset.exp_offset) <<
+			SHIFT(CDP_S_LUT_INFO_0, LUT_LE_INDEX_OFFSET))&
+		MASK(CDP_S_LUT_INFO_0, LUT_LE_INDEX_OFFSET)) |
+			((((uint32_t)lut->linear_only_offset.frac_bits) <<
+			SHIFT(CDP_S_LUT_INFO_0, LUT_LO_INDEX_SELECT))&
+		MASK(CDP_S_LUT_INFO_0, LUT_LO_INDEX_SELECT));
+	} else {
+		reg = ((((uint32_t)lut->linear_exp_offset.frac_bits) <<
+			SHIFT(CDP_S_LUT_INFO_0, LUT_LE_INDEX_SELECT))&
+		MASK(CDP_S_LUT_INFO_0, LUT_LE_INDEX_SELECT)) |
+			((((uint32_t)lut->linear_only_offset.frac_bits) <<
+			SHIFT(CDP_S_LUT_INFO_0, LUT_LO_INDEX_SELECT))&
+		MASK(CDP_S_LUT_INFO_0, LUT_LO_INDEX_SELECT));
+	}
+	dla_reg_write(engine->driver_context,
+			reg_base + lut_info_offset[is_sdp], reg);
+	high = upper_32_bits(lut->linear_exp_start);
+	low = lower_32_bits(lut->linear_exp_start);
+	dla_reg_write(engine->driver_context,
+			reg_base + le_start_offset[is_sdp], low);
+	if (!is_sdp)
+		dla_reg_write(engine->driver_context,
+				reg_base + le_start_offset[is_sdp] + 4, high);
+
+	high = upper_32_bits(lut->linear_exp_end);
+	low = lower_32_bits(lut->linear_exp_end);
+	dla_reg_write(engine->driver_context,
+				reg_base + le_end_offset[is_sdp], low);
+	if (!is_sdp)
+		dla_reg_write(engine->driver_context,
+				reg_base + le_end_offset[is_sdp] + 4, high);
+
+	high = upper_32_bits(lut->linear_only_start);
+	low = lower_32_bits(lut->linear_only_start);
+	dla_reg_write(engine->driver_context,
+				reg_base + lo_start_offset[is_sdp], low);
+	if (!is_sdp)
+		dla_reg_write(engine->driver_context,
+				reg_base + lo_start_offset[is_sdp] + 4, high);
+
+	high = upper_32_bits(lut->linear_only_end);
+	low = lower_32_bits(lut->linear_only_end);
+	dla_reg_write(engine->driver_context,
+				reg_base + lo_end_offset[is_sdp], low);
+	if (!is_sdp)
+		dla_reg_write(engine->driver_context,
+				reg_base + lo_end_offset[is_sdp] + 4, high);
+
+	if (precision == PRECISION_FP16) {
+		reg = (lut->linear_exp_underflow_slope.data_f <<
+			SHIFT(CDP_S_LUT_LE_SLOPE_SCALE_0,
+					LUT_LE_SLOPE_UFLOW_SCALE)) |
+			(lut->linear_exp_overflow_slope.data_f <<
+			SHIFT(CDP_S_LUT_LE_SLOPE_SCALE_0,
+					LUT_LE_SLOPE_OFLOW_SCALE));
+		dla_reg_write(engine->driver_context,
+				reg_base + le_slope_scale_offset[is_sdp], reg);
+
+		reg = (lut->linear_only_underflow_slope.data_f <<
+			SHIFT(CDP_S_LUT_LO_SLOPE_SCALE_0,
+					LUT_LO_SLOPE_UFLOW_SCALE)) |
+			(lut->linear_only_overflow_slope.data_f <<
+			SHIFT(CDP_S_LUT_LO_SLOPE_SCALE_0,
+					LUT_LO_SLOPE_OFLOW_SCALE));
+		dla_reg_write(engine->driver_context,
+				reg_base + lo_slope_scale_offset[is_sdp], reg);
+	} else {
+		union dla_slope *oslope;
+		union dla_slope *uslope;
+
+		uslope = &lut->linear_exp_underflow_slope;
+		oslope = &lut->linear_exp_overflow_slope;
+		reg = ((((uint32_t)uslope->data_i.scale)
+			<< SHIFT(CDP_S_LUT_LE_SLOPE_SCALE_0,
+					LUT_LE_SLOPE_UFLOW_SCALE))&
+			MASK(CDP_S_LUT_LE_SLOPE_SCALE_0,
+					LUT_LE_SLOPE_UFLOW_SCALE)) |
+			((((uint32_t)oslope->data_i.scale)
+			<< SHIFT(CDP_S_LUT_LE_SLOPE_SCALE_0,
+					LUT_LE_SLOPE_OFLOW_SCALE))&
+			MASK(CDP_S_LUT_LE_SLOPE_SCALE_0,
+					LUT_LE_SLOPE_OFLOW_SCALE));
+		dla_reg_write(engine->driver_context,
+				reg_base + le_slope_scale_offset[is_sdp], reg);
+
+		reg = ((((uint32_t)uslope->data_i.shifter) <<
+			SHIFT(CDP_S_LUT_LE_SLOPE_SHIFT_0,
+					LUT_LE_SLOPE_UFLOW_SHIFT))&
+			MASK(CDP_S_LUT_LE_SLOPE_SHIFT_0,
+					LUT_LE_SLOPE_UFLOW_SHIFT)) |
+			((((uint32_t)oslope->data_i.shifter) <<
+			SHIFT(CDP_S_LUT_LE_SLOPE_SHIFT_0,
+					LUT_LE_SLOPE_OFLOW_SHIFT))&
+			MASK(CDP_S_LUT_LE_SLOPE_SHIFT_0,
+					LUT_LE_SLOPE_OFLOW_SHIFT));
+		dla_reg_write(engine->driver_context,
+				reg_base + le_slope_shift_offset[is_sdp], reg);
+
+		uslope = &lut->linear_only_underflow_slope;
+		oslope = &lut->linear_only_overflow_slope;
+		reg = ((((uint32_t)uslope->data_i.scale) <<
+			SHIFT(CDP_S_LUT_LO_SLOPE_SCALE_0,
+					LUT_LO_SLOPE_UFLOW_SCALE))&
+			MASK(CDP_S_LUT_LO_SLOPE_SCALE_0,
+					LUT_LO_SLOPE_UFLOW_SCALE)) |
+			((((uint32_t)oslope->data_i.scale) <<
+			SHIFT(CDP_S_LUT_LO_SLOPE_SCALE_0,
+					LUT_LO_SLOPE_OFLOW_SCALE))&
+			MASK(CDP_S_LUT_LO_SLOPE_SCALE_0,
+					LUT_LO_SLOPE_OFLOW_SCALE));
+		dla_reg_write(engine->driver_context,
+				reg_base + lo_slope_scale_offset[is_sdp], reg);
+		reg = ((((uint32_t)uslope->data_i.shifter) <<
+			SHIFT(CDP_S_LUT_LO_SLOPE_SHIFT_0,
+					LUT_LO_SLOPE_UFLOW_SHIFT))&
+			MASK(CDP_S_LUT_LO_SLOPE_SHIFT_0,
+					LUT_LO_SLOPE_UFLOW_SHIFT)) |
+			((((uint32_t)oslope->data_i.shifter) <<
+			SHIFT(CDP_S_LUT_LO_SLOPE_SHIFT_0,
+					LUT_LO_SLOPE_OFLOW_SHIFT))&
+			MASK(CDP_S_LUT_LO_SLOPE_SHIFT_0,
+					LUT_LO_SLOPE_OFLOW_SHIFT));
+		dla_reg_write(engine->driver_context,
+				reg_base + lo_slope_shift_offset[is_sdp], reg);
+	}
+}
+
+int
+validate_data_cube(struct dla_data_cube src_data_cube,
+				   struct dla_data_cube dst_data_cube,
+				   uint8_t mem_type)
+{
+	int32_t ret = 0;
+
+	if ((src_data_cube.width > DCUBE_MAX_WIDTH) ||
+	    (src_data_cube.height > DCUBE_MAX_HEIGHT) ||
+	    (src_data_cube.channel > DCUBE_MAX_CHANNEL)) {
+		pr_err("Invalid SrcInput Cude[W: %u, H: %u, C: %u]",
+				src_data_cube.width, src_data_cube.height,
+				src_data_cube.channel);
+		ret = -EINVAL;
+		goto exit;
+	}
+
+	if ((dst_data_cube.width > DCUBE_MAX_WIDTH) ||
+	    (dst_data_cube.height > DCUBE_MAX_HEIGHT) ||
+	    (dst_data_cube.channel > DCUBE_MAX_CHANNEL)) {
+		pr_err("Invalid DstInput Cude[W: %u, H: %u, C: %u]",
+				dst_data_cube.width, dst_data_cube.height,
+				dst_data_cube.channel);
+		ret = -EINVAL;
+		goto exit;
+	}
+
+	if (src_data_cube.type > mem_type) {
+		pr_err("Invalid src_data.mem_type: %u\n", src_data_cube.type);
+		ret = -EINVAL;
+		goto exit;
+	}
+
+	if (dst_data_cube.type > mem_type) {
+		pr_err("Invalid dst_data.mem_type: %u\n", dst_data_cube.type);
+		ret = -EINVAL;
+		goto exit;
+	}
+
+exit:
+	return ret;
+}
+
+int
+validate_precision(uint8_t precision, uint8_t map_precision)
+{
+	int32_t ret = 0;
+
+	if (precision >= map_precision) {
+		pr_err("Invalid precision: %u\n", precision);
+		ret = -EINVAL;
+	}
+
+	return ret;
+}
drivers/gpu/drm/nvdla/nvdla_common.h new file +835 -0
new file mode 100644
index 000000000000..38cf43246890
@@ -0,0 +1,835 @@
+/* SPDX-License-Identifier: GPL-2.0 OR BSD-3-Clause */
+/*
+ * Copyright (C) 2017-2018 NVIDIA CORPORATION.
+ * Copyright (C) 2022 Cai Huoqing
+ */
+
+#ifndef __NVDLA_COMMON_H_
+#define __NVDLA_COMMON_H_
+
+#include <linux/types.h>
+
+#define DCUBE_MAX_WIDTH		8192
+#define DCUBE_MAX_HEIGHT	8192
+#define DCUBE_MAX_CHANNEL	8192
+
+/**
+ * @ingroup Processors
+ * @name DLA Processors
+ * Processor modules in DLA engine. Each processor has it's
+ * own operation a.k.a. HW layer. Network is formed using
+ * graph of these operations
+ * @{
+ */
+#define DLA_OP_BDMA		0
+#define DLA_OP_CONV		1
+#define DLA_OP_SDP		2
+#define DLA_OP_PDP		3
+#define DLA_OP_CDP		4
+#define DLA_OP_RUBIK	5
+/** @} */
+
+/**
+ * @ingroup Processors
+ * @name Maximum number of processors
+ * @brief DLA ash 6 processors
+ * @{
+ */
+#define DLA_OP_NUM		6
+/** @} */
+
+/**
+ * @ingroup Processors
+ * @name Number of groups
+ * @brief Each processor has 2 groups of registers
+ * @{
+ */
+#define DLA_NUM_GROUPS		2
+/** @} */
+
+/**
+ * Network descriptor
+ *
+ * Contains all information to execute a network
+ *
+ * @op_head: Index of first operation of each type in operations list
+ * @num_rois: Number of ROIs
+ * @num_operations: Number of operations in one list
+ * @num_luts: Number of LUTs
+ */
+struct dla_network_desc {
+	int16_t operation_desc_index;
+	int16_t surface_desc_index;
+
+	int16_t dependency_graph_index;
+	int16_t lut_data_index;
+
+	int16_t roi_array_index;
+	int16_t surface_index;
+
+	int16_t stat_list_index;
+	int16_t reserved1;
+
+	int16_t op_head[DLA_OP_NUM];
+
+	uint16_t num_rois;
+	uint16_t num_operations;
+
+	uint16_t num_luts;
+	uint16_t num_addresses;
+
+	int16_t input_layer;
+	uint8_t dynamic_roi;
+	uint8_t reserved0;
+} __packed __aligned(4);
+
+/**
+ * @name Memory types
+ * @brief DLA engnine can read/write to/from 3 memory types
+ * @{
+ */
+#define DLA_MEM_MC			0 /* External DRAM */
+#define DLA_MEM_CV			1 /* CV-SRAM */
+#define DLA_MEM_HW			2 /* DLA sub-module */
+/** @} */
+
+/**
+ * @ingroup Events
+ * @name Operation events
+ * @brief Different events triggered by an operations
+ * @{
+ */
+#define DLA_EVENT_OP_COMPLETED		1
+#define DLA_EVENT_OP_PROGRAMMED		2
+#define DLA_EVENT_OP_ENABLED		3
+#define DLA_EVENT_CDMA_WT_DONE		4
+#define DLA_EVENT_CDMA_DT_DONE		5
+/** @} */
+
+struct dla_consumer {
+	int16_t index; /* the index of dla_common_op_desc in dep_graph_addr */
+	uint8_t event;
+	uint8_t res;
+} __packed __aligned(4);
+
+struct dla_common_op_desc {
+	int16_t index; /* set by ucode */
+	int8_t roi_index;
+	uint8_t op_type;
+
+	uint8_t dependency_count;
+	uint8_t reserved0[3];
+
+	struct dla_consumer consumers[DLA_OP_NUM];
+	struct dla_consumer fused_parent;
+} __packed __aligned(4);
+
+struct dla_roi_array_desc {
+	uint32_t array_length;
+
+	uint32_t array_reserved;
+} __packed __aligned(4);
+
+struct dla_roi_desc {
+	uint32_t left;
+
+	uint32_t top;
+
+	uint32_t right;
+
+	uint32_t bottom;
+} __packed __aligned(4);
+
+/**
+ * @ingroup BDMA
+ * @name Maximum BDMA transfers
+ * @brief BDMA supports multiple transfers in operation. This indicates
+ *        maximum number of transfers possible in one operation.
+ * @{
+ */
+#define NUM_MAX_BDMA_OPS	20
+/** @} */
+
+struct dla_bdma_transfer_desc {
+	int16_t source_address;
+	int16_t destination_address;
+
+	uint32_t line_size;
+
+	uint32_t line_repeat;
+
+	uint32_t source_line;
+
+	uint32_t destination_line;
+
+	uint32_t surface_repeat;
+
+	uint32_t source_surface;
+
+	uint32_t destination_surface;
+} __packed __aligned(4);
+
+struct dla_bdma_surface_desc {
+	uint8_t source_type;
+	uint8_t destination_type;
+	uint16_t num_transfers;
+
+	struct dla_bdma_transfer_desc transfers[NUM_MAX_BDMA_OPS];
+} __packed __aligned(4);
+
+struct dla_bdma_op_desc {
+	uint16_t num_transfers;
+	uint16_t reserved0;
+} __packed __aligned(4);
+
+struct dla_bdma_stat_desc {
+	uint32_t read_stall;
+	uint32_t write_stall;
+} __packed __aligned(4);
+
+/**
+ * @ingroup Convolution
+ * @name Convolution mode
+ * @brief Convolution modes support by DLA
+ * @{
+ */
+#define CONV_MODE_DIRECT	0
+#define CONV_MODE_WINOGRAD	1
+/** @} */
+
+/**
+ * @ingroup Processors
+ * @name Precision BPE mapping
+ * @brief Precision formats and Bit Per Elements mapping
+ * @{
+ */
+#define BPE_PRECISION_INT8		1
+#define BPE_PRECISION_INT16		2
+#define BPE_PRECISION_FP16		2
+/** @} */
+
+
+/**
+ * @ingroup Processors
+ * @name Precision types
+ * @brief Precision formats supported by DLA engine
+ * @{
+ */
+#define PRECISION_INT8		0
+#define PRECISION_INT16		1
+#define PRECISION_FP16		2
+/** @} */
+
+/**
+ * @ingroup Processors
+ * @name Data formats
+ * @brief Data formats supported by DLA engine
+ * @{
+ */
+#define FORMAT_T_R8			0
+#define FORMAT_T_R10			1
+#define FORMAT_T_R12			2
+#define FORMAT_T_R16			3
+#define FORMAT_T_R16_I			4
+#define FORMAT_T_R16_F			5
+#define FORMAT_T_A16B16G16R16		6
+#define FORMAT_T_X16B16G16R16		7
+#define FORMAT_T_A16B16G16R16_F		8
+#define FORMAT_T_A16Y16U16V16		9
+#define FORMAT_T_V16U16Y16A16		10
+#define FORMAT_T_A16Y16U16V16_F		11
+#define FORMAT_T_A8B8G8R8		12
+#define FORMAT_T_A8R8G8B8		13
+#define FORMAT_T_B8G8R8A8		14
+#define FORMAT_T_R8G8B8A8		15
+#define FORMAT_T_X8B8G8R8		16
+#define FORMAT_T_X8R8G8B8		17
+#define FORMAT_T_B8G8R8X8		18
+#define FORMAT_T_R8G8B8X8		19
+#define FORMAT_T_A2B10G10R10		20
+#define FORMAT_T_A2R10G10B10		21
+#define FORMAT_T_B10G10R10A2		22
+#define FORMAT_T_R10G10B10A2		23
+#define FORMAT_T_A2Y10U10V10		24
+#define FORMAT_T_V10U10Y10A2		25
+#define FORMAT_T_A8Y8U8V8			26
+#define FORMAT_T_V8U8Y8A8			27
+#define FORMAT_T_Y8___U8V8_N444		28
+#define FORMAT_T_Y8___V8U8_N444		29
+#define FORMAT_T_Y10___U10V10_N444	30
+#define FORMAT_T_Y10___V10U10_N444	31
+#define FORMAT_T_Y12___U12V12_N444	32
+#define FORMAT_T_Y12___V12U12_N444	33
+#define FORMAT_T_Y16___U16V16_N444	34
+#define FORMAT_T_Y16___V16U16_N444	35
+#define FORMAT_FEATURE			36
+/** @} */
+
+/**
+ * @ingroup Convolution
+ * @name Pixel mapping
+ * @brief Pixel mapping formats supported for image input in Convolution
+ * @{
+ */
+#define MAP_PITCH_LINEAR		0
+/** @} */
+
+/**
+ * @ingroup Convolution
+ * @name Weight formats
+ * @brief Weight data formats supported in Convolution
+ * @{
+ */
+#define WEIGHT_FORMAT_UNCOMPRESSED	0
+#define WEIGHT_FORMAT_COMPRESSED	1
+/** @} */
+
+/**
+ * @ingroup Convolution
+ * @name Mean data format
+ * @brief Mean data formats supported in Convolution
+ * @{
+ */
+#define MEAN_FORMAT_DISABLE     0
+#define MEAN_FORMAT_ENABLE      1
+/** @} */
+
+struct dla_cvt_param {
+	int16_t  scale;
+	uint8_t  truncate;
+	uint8_t  enable;
+
+	int32_t  offset;
+} __packed __aligned(4);
+
+struct dla_data_cube {
+	uint16_t type; /* dla_mem_type */
+	int16_t address; /* offset to the actual IOVA in task.address_list */
+
+	uint32_t offset; /* offset within address */
+	uint32_t size;
+
+	/* cube dimensions */
+	uint16_t width;
+	uint16_t height;
+
+	uint16_t channel;
+	uint16_t reserved0;
+
+	/* stride information */
+	uint32_t line_stride;
+	uint32_t surf_stride;
+
+	/* For Rubik only */
+	uint32_t plane_stride;
+} __packed __aligned(4);
+
+#define PIXEL_OVERRIDE_UINT 0
+#define PIXEL_OVERRIDE_INT  1
+
+struct dla_conv_surface_desc {
+	/* Data cube */
+	struct dla_data_cube weight_data;
+	struct dla_data_cube wmb_data;
+	struct dla_data_cube wgs_data;
+	struct dla_data_cube src_data;
+	struct dla_data_cube dst_data;
+
+	/**
+	 * u_addr = input_data.source_addr + offset_u
+	 * this field should be set when YUV is not interleave format
+	 *
+	 */
+	int64_t offset_u;
+
+	/* line stride for 2nd plane, must be 32bytes aligned */
+	uint32_t in_line_uv_stride;
+} __packed __aligned(4);
+
+struct dla_conv_op_desc {
+	/* Performance parameters */
+
+	/* dla_conv_mode */
+	uint8_t conv_mode;
+	uint8_t data_reuse;
+	uint8_t weight_reuse;
+	uint8_t skip_data_rls;
+
+	uint8_t skip_weight_rls;
+	uint8_t reserved0;
+	uint16_t entry_per_slice;
+
+	/* dla_data_format */
+	uint8_t data_format;
+	/* dla_pixel_mapping */
+	uint8_t pixel_mapping;
+	/* number of free slices before fetch */
+	uint16_t fetch_grain;
+
+	uint8_t reserved_b[8];
+
+	/* batch_num */
+	uint8_t batch;
+	/* dla_weight_format */
+	uint8_t weight_format;
+	uint8_t data_bank;
+	uint8_t weight_bank;
+
+	/* the offset in bytes of each data cube in a batch */
+	uint32_t batch_stride;
+
+	uint8_t post_extension;
+	uint8_t pixel_override;
+	/* number of slices need to be released */
+	uint16_t release;
+
+	 /* The input cube dimension for CSC */
+	uint16_t input_width_csc;
+	uint16_t input_height_csc;
+
+	uint16_t input_channel_csc;
+	uint16_t kernel_width_csc;
+
+	uint16_t kernel_height_csc;
+	uint16_t kernel_channel_csc;
+
+	/* The input cube dimension for CMAC */
+	uint16_t input_width_cmac;
+	uint16_t input_height_cmac;
+
+	/* actual size in bytes */
+	uint32_t bytes_per_kernel;
+
+	/* Algorithm parameters */
+
+	int16_t mean_ry; /* mean value for red in RGB or Y in YUV */
+	int16_t mean_gu; /* mean value for green in RGB or U in YUV */
+
+	int16_t mean_bv; /* mean value for blue in RGB or V in YUV */
+	int16_t mean_ax;
+
+	uint8_t mean_format; /* dla_mean_format */
+	uint8_t conv_stride_x;
+	uint8_t conv_stride_y;
+	uint8_t pad_x_left;
+
+	uint8_t pad_x_right;
+	uint8_t pad_y_top;
+	uint8_t pad_y_bottom;
+	uint8_t dilation_x;
+
+	uint8_t dilation_y;
+	uint8_t reserved2[2];
+
+	/* Precision parameters */
+	uint8_t pra_truncate;
+
+	uint8_t in_precision;
+	/* The output precision from CONV, it's the MAC processing precison */
+	uint8_t out_precision;
+	int16_t pad_val;
+
+	/* input converter parameters */
+	struct dla_cvt_param in_cvt;
+	/* output converter parameters, support truncate only */
+	struct dla_cvt_param out_cvt;
+
+} __packed __aligned(4);
+
+struct dla_conv_stat_desc {
+	uint32_t data_read_stall;
+	uint32_t weight_read_stall;
+	uint32_t data_read_latency;
+	uint32_t weight_read_latency;
+	uint32_t saturation_count;
+	uint32_t nan_data_num;
+	uint32_t nan_weight_num;
+	uint32_t inf_data_num;
+	uint32_t inf_weight_num;
+} __packed __aligned(4);
+
+/**
+ * @ingroup SDP
+ * @name Activation functions
+ * @brief Activation functions supported in SDP
+ * @{
+ */
+#define ACTIVATION_NONE		0
+#define ACTIVATION_RELU		1
+#define ACTIVATION_LUT		2
+#define ACTIVATION_PRELU	3
+/** @} */
+
+/**
+ * @ingroup LUT
+ * @name LUT size
+ * @brief LUT sizes for linear and exponentila LUT
+ * @{
+ */
+#define LUT_LINEAR_EXP_TABLE_ENTRY_LOG2		6
+#define LUT_LINEAR_ONLY_TABLE_ENTRY_LOG2	8
+/** @} */
+
+/**
+ * @ingroup LUT
+ * @name LUT types
+ * @brief DLA supports two types of LUT, linear and exonential
+ * @{
+ */
+#define LUT_LINEAR_EXP_TABLE		0
+#define LUT_LINEAR_ONLY_TABLE		1
+/** @} */
+
+/**
+ * @ingroup LUT
+ * @name LUT methods
+ * @brief DLA supports two types of LUT, linear and exonential
+ * @{
+ */
+#define LUT_METHOD_EXPONENTIAL		0
+#define LUT_METHOD_LINEAR		1
+/** @} */
+
+/**
+ * @ingroup LUT
+ * @name LUT
+ * @brief DLA supports two types of LUT, linear and exonential
+ * @{
+ */
+#define LUT_PRI_LINEAR_EXP		0
+#define LUT_PRI_LINEAR_ONLY		1
+/** @} */
+
+union dla_lut_offset {
+	/**
+	 * Number should be substracted on log domain before look up
+	 * exponetial table it has the same definition as hardware
+	 * thus input scaling should also take into account when
+	 * set this field.
+	 */
+	int8_t exp_offset;
+	/**
+	 * Number of bits should be right shift before looking
+	 * up linear table
+	 */
+	int8_t frac_bits;
+	uint16_t reserved0;
+};
+
+/**
+ * This struct is used to represent floating point values by INT
+ * suppose we have a float point number fp_x, it will be represented
+ * as:
+ *
+ * fp_x = scale_int_x>>(shifter_x)
+ *
+ * This is very useful for INT pipeline;
+ */
+struct dla_float_data {
+	int16_t scale;
+	int8_t shifter;
+	uint8_t reserved0;
+} __packed __aligned(4);
+
+/**
+ * For INT pipeline, we use the struct above to represent a floating number;
+ * For FP16 pipeline, we should store the FP16 encoded value into a uint16_t
+ * container
+ */
+union dla_slope {
+	struct dla_float_data data_i;
+
+	uint16_t data_f;
+};
+
+struct dla_lut_param {
+	/**
+	 * value of expression ((1<<LUT_LINEAR_EXP_TABLE_ENTRY_LOG2)+1) is 65,
+	 * ((1<<LUT_LINEAR_ONLY_TABLE_ENTRY_LOG2)+1) is 257, and int16_t is of
+	 * 2Byte. And below two statement's combined memory size is 644 Byte.
+	 *
+	 * NOTE: below two declaration combined size should always be multiple
+	 * of 4.
+	 */
+	int16_t linear_exp_table[(1<<LUT_LINEAR_EXP_TABLE_ENTRY_LOG2)+1];
+	int16_t linear_only_table[(1<<LUT_LINEAR_ONLY_TABLE_ENTRY_LOG2)+1];
+
+	union dla_lut_offset linear_exp_offset;
+	union dla_lut_offset linear_only_offset;
+
+	/**
+	 * The start and end point of raw table,
+	 * valid when raw_method=LINEAR only
+	 */
+	uint64_t linear_exp_start;
+	uint64_t linear_exp_end;
+	uint64_t linear_only_start;
+	uint64_t linear_only_end;
+
+	union dla_slope linear_exp_underflow_slope;
+	union dla_slope linear_exp_overflow_slope;
+	union dla_slope linear_only_underflow_slope;
+	union dla_slope linear_only_overflow_slope;
+
+	/**
+	 * dla_lut_priority, when both lut are hit(or one overflow,
+	 * the other underflow), which one should be selected as output
+	 */
+	uint8_t hybrid_priority;
+	uint8_t underflow_priority;
+	uint8_t overflow_priority;
+	uint8_t method; /* dla_lut_method */
+} __packed __aligned(4);
+
+struct dla_sdp_surface_desc {
+	/* Data cube */
+	/* source input cube, available when SDP working on offline mode */
+	struct dla_data_cube src_data;
+
+	/* X1 input cube */
+	struct dla_data_cube x1_data;
+
+	/* X2 input cube */
+	struct dla_data_cube x2_data;
+
+	/* Y input cube */
+	struct dla_data_cube y_data;
+
+	/* Output cube */
+	struct dla_data_cube dst_data;
+} __packed __aligned(4);
+
+#define SDP_OP_NONE		0
+#define SDP_OP_MUL		1
+#define SDP_OP_ADD		2
+#define SDP_OP_BOTH		3
+
+#define SDP_ALU_OP_MAX		0
+#define SDP_ALU_OP_MIN		1
+#define SDP_ALU_OP_SUM		2
+#define SDP_ALU_OP_EQL		3
+
+#define SDP_OP_PER_LAYER	0
+#define SDP_OP_PER_KERNEL	1
+#define SDP_OP_PER_POINT	2
+
+struct dla_sdp_cvt {
+	struct dla_cvt_param alu_cvt;
+	struct dla_cvt_param mul_cvt;
+} __packed __aligned(4);
+
+struct dla_sdp_op {
+	uint8_t enable;
+	uint8_t alu_type; /* dla_sdp_alu_op_type */
+	uint8_t type; /* dla_sdp_op_type */
+	uint8_t mode; /* dla_sdp_op_mode */
+
+	uint8_t act; /* dla_act_type */
+	uint8_t shift_value; /* left shift */
+	uint8_t truncate;
+	uint8_t precision;
+
+	int32_t alu_operand;
+	int32_t mul_operand;
+
+	struct dla_sdp_cvt  cvt;
+} __packed __aligned(4);
+
+struct dla_sdp_op_desc {
+	/* Precision parameters */
+	/* dla_precision */
+	uint8_t src_precision;
+	uint8_t dst_precision;
+	int16_t lut_index;
+
+	struct dla_cvt_param out_cvt;
+
+	/* Performance parameters */
+	/* dla_conv_mode */
+	uint8_t conv_mode;
+	uint8_t batch_num;
+	uint16_t reserved0;
+
+	uint32_t batch_stride;	/* will be used when batch_num > 1 */
+
+	/* Algorithm parameters */
+	struct dla_sdp_op x1_op;
+	struct dla_sdp_op x2_op;
+	struct dla_sdp_op y_op;
+} __packed __aligned(4);
+
+#define POOL_MODE_AVG		0
+#define POOL_MODE_MAX		1
+#define POOL_MODE_MIN		2
+
+#define POOL_SIZE_1		0
+#define POOL_SIZE_2		1
+#define POOL_SIZE_3		2
+#define POOL_SIZE_4		3
+#define POOL_SIZE_5		4
+#define POOL_SIZE_6		5
+#define POOL_SIZE_7		6
+#define POOL_SIZE_8		7
+
+#define PDP_PAD_VAL_NUM	7
+
+struct dla_pdp_surface_desc {
+	/* Data cube */
+	struct dla_data_cube src_data;
+
+	struct dla_data_cube dst_data;
+} __packed __aligned(4);
+
+struct dla_pdp_op_desc {
+	/* Performance parameters */
+	uint16_t  partial_in_width_first;
+	uint16_t  partial_in_width_mid;
+
+	uint16_t  partial_in_width_last;
+	uint16_t  partial_width_first;
+
+	uint16_t  partial_width_mid;
+	uint16_t  partial_width_last;
+
+	uint8_t   split_num;
+
+	/* Algorithm parameters */
+	uint8_t  pool_mode; /* dla_pool_mode */
+	uint8_t  pool_width; /* dla_pool_width */
+	uint8_t  pool_height; /* dla_pool_height */
+
+	uint8_t  stride_x;
+	uint8_t  stride_y;
+
+	/**
+	 * The left/right padding size,
+	 * pad_right might be less than pad_left
+	 */
+	uint8_t  pad_left;
+	uint8_t  pad_right;
+
+	/* The top/bottom padding size */
+	uint8_t  pad_top;
+	uint8_t  pad_bottom;
+
+	/* Precision parameters */
+	uint8_t  precision; /* dla_precision */
+	uint8_t  reserved0;
+	/**
+	 * if input has non-zero "offset", this value should be set
+	 * There'll be 7 different paddding values, the relationship between
+	 * those versions are:
+	 * padding_value[0] = -offset*scaling;
+	 * padding_value[1] = 2*padding_value[0]
+	 * padding_value[2] = 3*padding_value[0]
+	 * ...
+	 * The purpose is to avoid ucode implement FP16
+	 * multiplier(for FP16 mode)
+	 */
+	int32_t  padding_value[PDP_PAD_VAL_NUM];
+} __packed __aligned(4);
+
+struct dla_pdp_stat_desc {
+	uint32_t inf_input_num;
+	uint32_t nan_input_num;
+	uint32_t nan_output_num;
+	uint32_t write_stall;
+	uint32_t runtime;
+} __packed __aligned(4);
+
+struct dla_cdp_surface_desc {
+	/* Data cube */
+	struct dla_data_cube src_data;
+
+	struct dla_data_cube dst_data;
+} __packed __aligned(4);
+
+struct dla_cdp_op_desc {
+	/* Precision parameters */
+
+	/* dla_precision */
+	uint8_t  in_precision;
+	uint8_t  out_precision;
+	int16_t  lut_index;
+
+	struct dla_cvt_param in_cvt;
+	struct dla_cvt_param out_cvt;
+
+	/* Performance parameters */
+
+	/* Algorithm parameters */
+	uint8_t  local_size;
+	uint8_t  bypass_sqsum;
+	uint8_t  bypass_out_mul;
+	uint8_t  reserved0;
+} __packed __aligned(4);
+
+struct dla_cdp_stat_desc {
+	uint32_t nan_input_num;
+	uint32_t inf_input_num;
+	uint32_t nan_output_num;
+	uint32_t write_stall;
+	uint32_t lut_uflow;
+	uint32_t lut_oflow;
+	uint32_t lut_hybrid;
+	uint32_t lut_le_hit;
+	uint32_t lut_lo_hit;
+	uint32_t saturation_count;
+	uint32_t runtime;
+} __packed __aligned(4);
+
+struct dla_rubik_surface_desc {
+	/* Data cube */
+	struct dla_data_cube src_data;
+
+	struct dla_data_cube dst_data;
+} __packed __aligned(4);
+
+/* rubik mode */
+#define RUBIK_MODE_CONTRACT	0
+#define RUBIK_MODE_SPLIT	1
+#define RUBIK_MODE_MERGE	2
+
+struct dla_rubik_op_desc {
+	/* Precision parameters */
+	uint8_t mode;
+	uint8_t precision;
+	uint8_t stride_x;
+	uint8_t stride_y;
+} __packed __aligned(4);
+
+struct dla_rubik_stat_desc {
+	uint32_t read_stall;
+	uint32_t write_stall;
+	uint32_t runtime;
+} __packed __aligned(4);
+
+union dla_surface_container {
+	struct dla_bdma_surface_desc bdma_surface;
+	struct dla_conv_surface_desc conv_surface;
+	struct dla_sdp_surface_desc sdp_surface;
+	struct dla_pdp_surface_desc pdp_surface;
+	struct dla_cdp_surface_desc cdp_surface;
+	struct dla_rubik_surface_desc rubik_surface;
+};
+
+union dla_operation_container {
+	struct dla_bdma_op_desc bdma_op;
+	struct dla_conv_op_desc conv_op;
+	struct dla_sdp_op_desc sdp_op;
+	struct dla_pdp_op_desc pdp_op;
+	struct dla_cdp_op_desc cdp_op;
+	struct dla_rubik_op_desc rubik_op;
+};
+
+struct dla_engine;
+void update_lut(struct dla_engine *engine, uint32_t reg_base,
+		struct dla_lut_param *lut,
+		uint8_t precision);
+int32_t validate_data_cube(struct dla_data_cube src_data_cube,
+			struct dla_data_cube dst_data_cube,
+			uint8_t mem_type);
+int32_t validate_precision(uint8_t precision,
+			uint8_t map_precision);
+
+#endif
drivers/gpu/drm/nvdla/nvdla_conv.c new file +683 -0
new file mode 100644
index 000000000000..bb24909e1afa
@@ -0,0 +1,683 @@
+// SPDX-License-Identifier: BSD-3-Clause OR GPL-2.0
+/*
+ * Copyright (C) 2017-2018 NVIDIA CORPORATION
+ * Copyright (C) 2022 Cai Huoqing
+ */
+
+#include "nvdla_common.h"
+#include "nvdla_drm.h"
+#include "nvdla_reg.h"
+#include "nvdla_common.h"
+#include "nvdla_engine.h"
+
+static const uint8_t map_precision[] = {
+	FIELD_ENUM(CDMA_D_MISC_CFG_0, IN_PRECISION, INT8),
+	FIELD_ENUM(CDMA_D_MISC_CFG_0, IN_PRECISION, INT16),
+	FIELD_ENUM(CDMA_D_MISC_CFG_0, IN_PRECISION, FP16),
+};
+
+static const uint8_t map_conv[] = {
+	FIELD_ENUM(CACC_D_MISC_CFG_0, CONV_MODE, DIRECT),
+	FIELD_ENUM(CACC_D_MISC_CFG_0, CONV_MODE, WINOGRAD),
+};
+
+static const uint8_t map_weight_fmt[] = {
+	FIELD_ENUM(CSC_D_WEIGHT_FORMAT_0, WEIGHT_FORMAT, UNCOMPRESSED),
+	FIELD_ENUM(CSC_D_WEIGHT_FORMAT_0, WEIGHT_FORMAT, COMPRESSED),
+};
+
+static const uint8_t map_img_fmt[][2] = {
+	{FIELD_ENUM(CDMA_D_DATAIN_FORMAT_0,
+			PIXEL_FORMAT, T_R8), 1},
+	{FIELD_ENUM(CDMA_D_DATAIN_FORMAT_0,
+			PIXEL_FORMAT, T_R10), 2},
+	{FIELD_ENUM(CDMA_D_DATAIN_FORMAT_0,
+			PIXEL_FORMAT, T_R12), 2},
+	{FIELD_ENUM(CDMA_D_DATAIN_FORMAT_0,
+			PIXEL_FORMAT, T_R16), 2},
+	{FIELD_ENUM(CDMA_D_DATAIN_FORMAT_0,
+			PIXEL_FORMAT, T_R16_I), 2},
+	{FIELD_ENUM(CDMA_D_DATAIN_FORMAT_0,
+			PIXEL_FORMAT, T_R16_F), 2},
+	{FIELD_ENUM(CDMA_D_DATAIN_FORMAT_0,
+			PIXEL_FORMAT, T_A16B16G16R16), 8},
+	{FIELD_ENUM(CDMA_D_DATAIN_FORMAT_0,
+			PIXEL_FORMAT, T_X16B16G16R16), 8},
+	{FIELD_ENUM(CDMA_D_DATAIN_FORMAT_0,
+			PIXEL_FORMAT, T_A16B16G16R16_F), 8},
+	{FIELD_ENUM(CDMA_D_DATAIN_FORMAT_0,
+			PIXEL_FORMAT, T_A16Y16U16V16), 8},
+	{FIELD_ENUM(CDMA_D_DATAIN_FORMAT_0,
+			PIXEL_FORMAT, T_V16U16Y16A16), 8},
+	{FIELD_ENUM(CDMA_D_DATAIN_FORMAT_0,
+			PIXEL_FORMAT, T_A16Y16U16V16_F), 8},
+	{FIELD_ENUM(CDMA_D_DATAIN_FORMAT_0,
+			PIXEL_FORMAT, T_A8B8G8R8), 4},
+	{FIELD_ENUM(CDMA_D_DATAIN_FORMAT_0,
+			PIXEL_FORMAT, T_A8R8G8B8), 4},
+	{FIELD_ENUM(CDMA_D_DATAIN_FORMAT_0,
+			PIXEL_FORMAT, T_B8G8R8A8), 4},
+	{FIELD_ENUM(CDMA_D_DATAIN_FORMAT_0,
+			PIXEL_FORMAT, T_R8G8B8A8), 4},
+	{FIELD_ENUM(CDMA_D_DATAIN_FORMAT_0,
+			PIXEL_FORMAT, T_X8B8G8R8), 4},
+	{FIELD_ENUM(CDMA_D_DATAIN_FORMAT_0,
+			PIXEL_FORMAT, T_X8R8G8B8), 4},
+	{FIELD_ENUM(CDMA_D_DATAIN_FORMAT_0,
+			PIXEL_FORMAT, T_B8G8R8X8), 4},
+	{FIELD_ENUM(CDMA_D_DATAIN_FORMAT_0,
+			PIXEL_FORMAT, T_R8G8B8X8), 4},
+	{FIELD_ENUM(CDMA_D_DATAIN_FORMAT_0,
+			PIXEL_FORMAT, T_A2B10G10R10), 4},
+	{FIELD_ENUM(CDMA_D_DATAIN_FORMAT_0,
+			PIXEL_FORMAT, T_A2R10G10B10), 4},
+	{FIELD_ENUM(CDMA_D_DATAIN_FORMAT_0,
+			PIXEL_FORMAT, T_B10G10R10A2), 4},
+	{FIELD_ENUM(CDMA_D_DATAIN_FORMAT_0,
+			PIXEL_FORMAT, T_R10G10B10A2), 4},
+	{FIELD_ENUM(CDMA_D_DATAIN_FORMAT_0,
+			PIXEL_FORMAT, T_A2Y10U10V10), 4},
+	{FIELD_ENUM(CDMA_D_DATAIN_FORMAT_0,
+			PIXEL_FORMAT, T_V10U10Y10A2), 4},
+	{FIELD_ENUM(CDMA_D_DATAIN_FORMAT_0,
+			PIXEL_FORMAT, T_A8Y8U8V8), 4},
+	{FIELD_ENUM(CDMA_D_DATAIN_FORMAT_0,
+			PIXEL_FORMAT, T_V8U8Y8A8), 4},
+	{FIELD_ENUM(CDMA_D_DATAIN_FORMAT_0,
+			PIXEL_FORMAT, T_Y8___U8V8_N444), 1},
+	{FIELD_ENUM(CDMA_D_DATAIN_FORMAT_0,
+			PIXEL_FORMAT, T_Y8___V8U8_N444), 1},
+	{FIELD_ENUM(CDMA_D_DATAIN_FORMAT_0,
+			PIXEL_FORMAT, T_Y10___U10V10_N444), 2},
+	{FIELD_ENUM(CDMA_D_DATAIN_FORMAT_0,
+			PIXEL_FORMAT, T_Y10___V10U10_N444), 2},
+	{FIELD_ENUM(CDMA_D_DATAIN_FORMAT_0,
+			PIXEL_FORMAT, T_Y12___U12V12_N444), 2},
+	{FIELD_ENUM(CDMA_D_DATAIN_FORMAT_0,
+			PIXEL_FORMAT, T_Y12___V12U12_N444), 2},
+	{FIELD_ENUM(CDMA_D_DATAIN_FORMAT_0,
+			PIXEL_FORMAT, T_Y16___U16V16_N444), 2},
+	{FIELD_ENUM(CDMA_D_DATAIN_FORMAT_0,
+			PIXEL_FORMAT, T_Y16___V16U16_N444), 2},
+	{FIELD_ENUM(CDMA_D_DATAIN_FORMAT_0,
+			DATAIN_FORMAT, FEATURE), 2},
+	{FIELD_ENUM(CDMA_D_DATAIN_FORMAT_0,
+			DATAIN_FORMAT, PIXEL), 1},
+};
+
+static const uint8_t map_pixel[] = {
+	FIELD_ENUM(CDMA_D_DATAIN_FORMAT_0, PIXEL_MAPPING, PITCH_LINEAR),
+};
+
+static const uint8_t map_ram[] = {
+	FIELD_ENUM(CDMA_D_DAIN_RAM_TYPE_0, DATAIN_RAM_TYPE, MCIF),
+	FIELD_ENUM(CDMA_D_DAIN_RAM_TYPE_0, DATAIN_RAM_TYPE, CVIF),
+};
+
+static const uint8_t map_mean[] = {
+	FIELD_ENUM(CDMA_D_MEAN_FORMAT_0, MEAN_FORMAT, DISABLE),
+	FIELD_ENUM(CDMA_D_MEAN_FORMAT_0, MEAN_FORMAT, ENABLE),
+};
+
+static uint32_t
+get_in_format(uint8_t format)
+{
+	uint32_t in_format = 0;
+
+	if (format >= FORMAT_T_R8 && format < FORMAT_FEATURE) {
+		in_format = FIELD_ENUM(CDMA_D_DATAIN_FORMAT_0,
+						DATAIN_FORMAT, PIXEL);
+	} else if (format == FORMAT_FEATURE) {
+		in_format = FIELD_ENUM(CDMA_D_DATAIN_FORMAT_0,
+						DATAIN_FORMAT, FEATURE);
+	}
+
+	return in_format;
+}
+
+void
+dla_conv_set_producer(struct dla_engine *engine, int32_t group_id, int32_t rdma_group_id)
+{
+	uint32_t reg;
+
+	/* set producer pointer for all sub-modules */
+	reg = group_id << SHIFT(CACC_S_POINTER_0, PRODUCER);
+	cacc_reg_write(engine, S_POINTER, reg);
+	cmac_a_reg_write(engine, S_POINTER, reg);
+	cmac_b_reg_write(engine, S_POINTER, reg);
+	csc_reg_write(engine, S_POINTER, reg);
+	cdma_reg_write(engine, S_POINTER, reg);
+}
+
+int
+dla_conv_enable(struct dla_engine *engine, struct dla_processor_group *group)
+{
+	uint32_t reg;
+
+	do {
+		reg = cdma_reg_read(engine, S_CBUF_FLUSH_STATUS);
+	} while (!(reg & MASK(CDMA_S_CBUF_FLUSH_STATUS_0, FLUSH_DONE)));
+
+	/* enable all sub-modules */
+	reg = FIELD_ENUM(CACC_D_OP_ENABLE_0, OP_EN, ENABLE);
+	cacc_reg_write(engine, D_OP_ENABLE, reg);
+	cmac_a_reg_write(engine, D_OP_ENABLE, reg);
+	cmac_b_reg_write(engine, D_OP_ENABLE, reg);
+	csc_reg_write(engine, D_OP_ENABLE, reg);
+	cdma_reg_write(engine, D_OP_ENABLE, reg);
+
+	return 0;
+}
+
+void
+dla_conv_rdma_check(struct dla_processor_group *group)
+{
+	group->is_rdma_needed = 0;
+}
+
+static int32_t
+processor_conv_program(struct dla_engine *engine, struct dla_processor_group *group)
+{
+	int32_t ret = 0;
+	uint32_t reg, high, low, shift, mask;
+	uint32_t stride_x, stride_y, pad_x, pad_y;
+	uint64_t weight_address = 0;
+	uint64_t wmb_address = 0;
+	uint64_t wgs_address = 0;
+	uint64_t input_address = 0;
+	uint64_t output_address = 0;
+	uint32_t atom_size = 0;
+	bool weight_compress_support = false;
+	struct dla_conv_op_desc *conv_op;
+	struct dla_conv_surface_desc *conv_surface;
+
+	weight_compress_support = engine->config_data->weight_compress_support;
+	atom_size = engine->config_data->atom_size;
+	conv_op = &group->operation_desc->conv_op;
+	conv_surface = &group->surface_desc->conv_surface;
+
+	if (conv_op->weight_format == WEIGHT_FORMAT_COMPRESSED) {
+		ASSERT_GOTO((weight_compress_support), ret, -EINVAL, exit);
+		ASSERT_GOTO((conv_surface->wmb_data.address != -1),
+			ret, -EINVAL, exit);
+		dla_get_dma_cube_address(engine->driver_context,
+								 engine->task->task_data,
+								 conv_surface->wmb_data.address,
+								 conv_surface->wmb_data.offset,
+								 (void *)&wmb_address,
+								 DESTINATION_DMA);
+
+		ASSERT_GOTO((conv_surface->wgs_data.address != -1),
+			ret, -EINVAL, exit);
+		dla_get_dma_cube_address(engine->driver_context,
+								 engine->task->task_data,
+								 conv_surface->wgs_data.address,
+								 conv_surface->wgs_data.offset,
+								 (void *)&wgs_address,
+								 DESTINATION_DMA);
+	}
+
+	if (conv_surface->weight_data.address != -1) {
+		dla_get_dma_cube_address(engine->driver_context,
+								 engine->task->task_data,
+								 conv_surface->weight_data.address,
+								 conv_surface->weight_data.offset,
+								 (void *)&weight_address,
+								 DESTINATION_DMA);
+	}
+
+	if (conv_surface->dst_data.address != -1) {
+		dla_get_dma_cube_address(engine->driver_context,
+								 engine->task->task_data,
+								 conv_surface->dst_data.address,
+								 conv_surface->dst_data.offset,
+								 (void *)&output_address,
+								 DESTINATION_DMA);
+	}
+
+	ret = dla_read_input_address(engine, &conv_surface->src_data, &input_address,
+								 group->op_desc->index,
+								 group->roi_index,
+								 map_img_fmt[conv_op->data_format][1]);
+	if (ret)
+		goto exit;
+
+	ASSERT_GOTO((conv_op->out_cvt.scale  == 1), ret, -EINVAL, exit);
+	ASSERT_GOTO((conv_op->out_cvt.offset == 0), ret, -EINVAL, exit);
+
+	/* check if the register group is idle */
+	reg = cacc_reg_read(engine, S_STATUS);
+	mask = group->id ? MASK(CACC_S_STATUS_0, STATUS_1) :
+		MASK(CACC_S_STATUS_0, STATUS_0);
+	shift = group->id ? SHIFT(CACC_S_STATUS_0, STATUS_1) :
+		SHIFT(CACC_S_STATUS_0, STATUS_0);
+	reg = (reg & mask) >> shift;
+	ASSERT_GOTO((reg == FIELD_ENUM(CACC_S_STATUS_0, STATUS_0, IDLE)),
+				ret, -EINVAL, exit);
+
+	reg = cmac_a_reg_read(engine, S_STATUS);
+	mask = group->id ? MASK(CMAC_A_S_STATUS_0, STATUS_1) :
+		MASK(CMAC_A_S_STATUS_0, STATUS_0);
+	shift = group->id ? SHIFT(CMAC_A_S_STATUS_0, STATUS_1) :
+		SHIFT(CMAC_A_S_STATUS_0, STATUS_0);
+	reg = (reg & mask) >> shift;
+	ASSERT_GOTO((reg == FIELD_ENUM(CMAC_A_S_STATUS_0, STATUS_0, IDLE)),
+				ret, -EINVAL, exit);
+
+	reg = cmac_b_reg_read(engine, S_STATUS);
+	mask = group->id ? MASK(CMAC_B_S_STATUS_0, STATUS_1) :
+		MASK(CMAC_B_S_STATUS_0, STATUS_0);
+	shift = group->id ? SHIFT(CMAC_B_S_STATUS_0, STATUS_1) :
+		SHIFT(CMAC_B_S_STATUS_0, STATUS_0);
+	reg = (reg & mask) >> shift;
+	ASSERT_GOTO((reg == FIELD_ENUM(CMAC_B_S_STATUS_0, STATUS_0, IDLE)),
+				ret, -EINVAL, exit);
+
+	reg = csc_reg_read(engine, S_STATUS);
+	mask = group->id ? MASK(CSC_S_STATUS_0, STATUS_1) :
+		MASK(CSC_S_STATUS_0, STATUS_0);
+	shift = group->id ? SHIFT(CSC_S_STATUS_0, STATUS_1) :
+		SHIFT(CSC_S_STATUS_0, STATUS_0);
+	reg = (reg & mask) >> shift;
+	ASSERT_GOTO((reg == FIELD_ENUM(CSC_S_STATUS_0, STATUS_0, IDLE)),
+				ret, -EINVAL, exit);
+
+	reg = cdma_reg_read(engine, S_STATUS);
+	mask = group->id ? MASK(CDMA_S_STATUS_0, STATUS_1) :
+		MASK(CDMA_S_STATUS_0, STATUS_0);
+	shift = group->id ? SHIFT(CDMA_S_STATUS_0, STATUS_1) :
+		SHIFT(CDMA_S_STATUS_0, STATUS_0);
+	reg = (reg & mask) >> shift;
+	ASSERT_GOTO((reg == FIELD_ENUM(CDMA_S_STATUS_0, STATUS_0, IDLE)),
+		ret, -EINVAL, exit);
+
+	/* reverse config each sub-module in CC */
+
+	/* CACC */
+	reg = (map_conv[conv_op->conv_mode]
+		<< SHIFT(CACC_D_MISC_CFG_0, CONV_MODE)) |
+		(map_precision[conv_op->out_precision]
+		<< SHIFT(CACC_D_MISC_CFG_0, PROC_PRECISION));
+	cacc_reg_write(engine, D_MISC_CFG, reg);
+
+	reg = ((conv_surface->dst_data.width - 1)
+		<< SHIFT(CACC_D_DATAOUT_SIZE_0_0, DATAOUT_WIDTH)) |
+		((conv_surface->dst_data.height - 1)
+		<< SHIFT(CACC_D_DATAOUT_SIZE_0_0, DATAOUT_HEIGHT));
+	cacc_reg_write(engine, D_DATAOUT_SIZE_0, reg);
+
+	reg = ((conv_surface->dst_data.channel - 1)
+		<< SHIFT(CACC_D_DATAOUT_SIZE_1_0, DATAOUT_CHANNEL));
+	cacc_reg_write(engine, D_DATAOUT_SIZE_1, reg);
+
+	low = lower_32_bits(output_address);
+	cacc_reg_write(engine, D_DATAOUT_ADDR, low);
+	cacc_reg_write(engine, D_BATCH_NUMBER, conv_op->batch - 1);
+	cacc_reg_write(engine, D_LINE_STRIDE, conv_surface->dst_data.line_stride);
+	cacc_reg_write(engine, D_SURF_STRIDE, conv_surface->dst_data.surf_stride);
+
+	if (conv_surface->dst_data.width == 1 &&
+				conv_surface->dst_data.height == 1) {
+		ASSERT_GOTO((((uint32_t)conv_surface->dst_data.line_stride ==
+			(uint32_t)(conv_surface->dst_data.width * atom_size))),
+			ret, -EINVAL, exit);
+		reg = (CACC_D_DATAOUT_MAP_0_LINE_PACKED_TRUE <<
+				SHIFT(CACC_D_DATAOUT_MAP_0, LINE_PACKED));
+		reg |= (CACC_D_DATAOUT_MAP_0_SURF_PACKED_TRUE <<
+				SHIFT(CACC_D_DATAOUT_MAP_0, SURF_PACKED));
+	} else {
+		reg = (FIELD_ENUM(CACC_D_DATAOUT_MAP_0, LINE_PACKED, FALSE) <<
+				SHIFT(CACC_D_DATAOUT_MAP_0, LINE_PACKED));
+		reg |= (FIELD_ENUM(CACC_D_DATAOUT_MAP_0, SURF_PACKED, FALSE) <<
+				SHIFT(CACC_D_DATAOUT_MAP_0, SURF_PACKED));
+	}
+	cacc_reg_write(engine, D_DATAOUT_MAP, reg);
+
+	cacc_reg_write(engine, D_CLIP_CFG, conv_op->out_cvt.truncate);
+
+	/* CMAC */
+	reg = (map_conv[conv_op->conv_mode]
+		<< SHIFT(CMAC_A_D_MISC_CFG_0, CONV_MODE)) |
+		(map_precision[conv_op->out_precision]
+		<< SHIFT(CMAC_A_D_MISC_CFG_0, PROC_PRECISION));
+	cmac_a_reg_write(engine, D_MISC_CFG, reg);
+	cmac_b_reg_write(engine, D_MISC_CFG, reg);
+
+	/* CSC */
+	reg = (map_conv[conv_op->conv_mode]
+		<< SHIFT(CSC_D_MISC_CFG_0, CONV_MODE)) |
+		(map_precision[conv_op->out_precision]
+		<< SHIFT(CSC_D_MISC_CFG_0, IN_PRECISION)) |
+		(map_precision[conv_op->out_precision]
+		<< SHIFT(CSC_D_MISC_CFG_0, PROC_PRECISION)) |
+		(conv_op->data_reuse
+		<< SHIFT(CSC_D_MISC_CFG_0, DATA_REUSE)) |
+		(conv_op->weight_reuse
+		<< SHIFT(CSC_D_MISC_CFG_0, WEIGHT_REUSE)) |
+		(conv_op->skip_data_rls
+		<< SHIFT(CSC_D_MISC_CFG_0, SKIP_DATA_RLS)) |
+		(conv_op->skip_weight_rls
+		<< SHIFT(CSC_D_MISC_CFG_0, SKIP_WEIGHT_RLS));
+	csc_reg_write(engine, D_MISC_CFG, reg);
+
+	reg = (get_in_format(conv_op->data_format) <<
+		SHIFT(CSC_D_DATAIN_FORMAT_0, DATAIN_FORMAT));
+	csc_reg_write(engine, D_DATAIN_FORMAT, reg);
+
+	reg = ((conv_op->input_width_csc - 1)
+		<< SHIFT(CSC_D_DATAIN_SIZE_EXT_0_0, DATAIN_WIDTH_EXT)) |
+		((conv_op->input_height_csc - 1)
+		<< SHIFT(CSC_D_DATAIN_SIZE_EXT_0_0, DATAIN_HEIGHT_EXT));
+	csc_reg_write(engine, D_DATAIN_SIZE_EXT_0, reg);
+
+	reg = ((conv_op->input_channel_csc - 1)
+		<< SHIFT(CSC_D_DATAIN_SIZE_EXT_1_0, DATAIN_CHANNEL_EXT));
+	csc_reg_write(engine, D_DATAIN_SIZE_EXT_1, reg);
+
+	reg = ((conv_op->batch - 1)
+		<< SHIFT(CSC_D_BATCH_NUMBER_0, BATCHES));
+	csc_reg_write(engine, D_BATCH_NUMBER, reg);
+	reg = ((conv_op->post_extension)
+		<< SHIFT(CSC_D_POST_Y_EXTENSION_0, Y_EXTENSION));
+	csc_reg_write(engine, D_POST_Y_EXTENSION, reg);
+
+	reg = ((conv_op->entry_per_slice - 1)
+		<< SHIFT(CSC_D_ENTRY_PER_SLICE_0, ENTRIES));
+	csc_reg_write(engine, D_ENTRY_PER_SLICE, reg);
+
+	reg = (map_weight_fmt[conv_op->weight_format]
+		<< SHIFT(CSC_D_WEIGHT_FORMAT_0, WEIGHT_FORMAT));
+	csc_reg_write(engine, D_WEIGHT_FORMAT, reg);
+
+	reg = ((conv_op->kernel_width_csc - 1)
+		<< SHIFT(CSC_D_WEIGHT_SIZE_EXT_0_0, WEIGHT_WIDTH_EXT)) |
+		((conv_op->kernel_height_csc - 1)
+		<< SHIFT(CSC_D_WEIGHT_SIZE_EXT_0_0, WEIGHT_HEIGHT_EXT));
+	csc_reg_write(engine, D_WEIGHT_SIZE_EXT_0, reg);
+
+	reg = ((conv_op->kernel_channel_csc - 1)
+		<< SHIFT(CSC_D_WEIGHT_SIZE_EXT_1_0, WEIGHT_CHANNEL_EXT)) |
+		((conv_surface->dst_data.channel - 1)
+		<< SHIFT(CSC_D_WEIGHT_SIZE_EXT_1_0, WEIGHT_KERNEL));
+	csc_reg_write(engine, D_WEIGHT_SIZE_EXT_1, reg);
+
+	csc_reg_write(engine, D_WEIGHT_BYTES, conv_surface->weight_data.size);
+	csc_reg_write(engine, D_WMB_BYTES, conv_surface->wmb_data.size);
+
+	reg = ((conv_op->input_width_cmac - 1)
+		<< SHIFT(CSC_D_DATAOUT_SIZE_0_0, DATAOUT_WIDTH)) |
+		((conv_op->input_height_cmac - 1)
+		<< SHIFT(CSC_D_DATAOUT_SIZE_0_0, DATAOUT_HEIGHT));
+	csc_reg_write(engine, D_DATAOUT_SIZE_0, reg);
+
+	reg = ((conv_surface->dst_data.channel - 1)
+		<< SHIFT(CSC_D_DATAOUT_SIZE_1_0, DATAOUT_CHANNEL));
+	csc_reg_write(engine, D_DATAOUT_SIZE_1, reg);
+
+	reg = ((conv_surface->dst_data.width *
+				conv_surface->dst_data.height - 1)
+		<< SHIFT(CSC_D_ATOMICS_0, ATOMICS));
+	csc_reg_write(engine, D_ATOMICS, reg);
+	reg = ((conv_op->release - 1)
+		<< SHIFT(CSC_D_RELEASE_0, RLS_SLICES));
+	csc_reg_write(engine, D_RELEASE, reg);
+
+	if (conv_op->conv_mode == CONV_MODE_DIRECT) {
+		stride_x = conv_op->conv_stride_x - 1;
+		stride_y = conv_op->conv_stride_y - 1;
+		pad_x = conv_op->pad_x_left;
+		pad_y = conv_op->pad_y_top;
+	} else {
+		stride_x = 0;
+		stride_y = 0;
+		pad_x = 0;
+		pad_y = 0;
+	}
+
+	reg = (stride_x
+		<< SHIFT(CSC_D_CONV_STRIDE_EXT_0, CONV_X_STRIDE_EXT)) |
+		(stride_y
+		<< SHIFT(CSC_D_CONV_STRIDE_EXT_0, CONV_Y_STRIDE_EXT));
+	csc_reg_write(engine, D_CONV_STRIDE_EXT, reg);
+
+	reg = ((conv_op->dilation_x - 1)
+		<< SHIFT(CSC_D_DILATION_EXT_0, X_DILATION_EXT)) |
+		((conv_op->dilation_y - 1)
+		<< SHIFT(CSC_D_DILATION_EXT_0, Y_DILATION_EXT));
+	csc_reg_write(engine, D_DILATION_EXT, reg);
+
+	reg = (pad_x
+		<< SHIFT(CSC_D_ZERO_PADDING_0, PAD_LEFT)) |
+		(pad_y
+		<< SHIFT(CSC_D_ZERO_PADDING_0, PAD_TOP));
+	csc_reg_write(engine, D_ZERO_PADDING, reg);
+
+	reg = (conv_op->pad_val
+		<< SHIFT(CSC_D_ZERO_PADDING_VALUE_0, PAD_VALUE)) &
+		MASK(CSC_D_ZERO_PADDING_VALUE_0, PAD_VALUE);
+	csc_reg_write(engine, D_ZERO_PADDING_VALUE, reg);
+
+	reg = ((conv_op->data_bank - 1)
+		<< SHIFT(CSC_D_BANK_0, DATA_BANK)) |
+		((conv_op->weight_bank - 1)
+		<< SHIFT(CSC_D_BANK_0, WEIGHT_BANK));
+	csc_reg_write(engine, D_BANK, reg);
+	csc_reg_write(engine, D_PRA_CFG, conv_op->pra_truncate);
+
+	/* CBUF */
+	/* there's no CBUF register */
+
+	/* CDMA */
+	reg = (map_conv[conv_op->conv_mode]
+		<< SHIFT(CDMA_D_MISC_CFG_0, CONV_MODE)) |
+		(map_precision[conv_op->in_precision]
+		<< SHIFT(CDMA_D_MISC_CFG_0, IN_PRECISION)) |
+		(map_precision[conv_op->out_precision]
+		<< SHIFT(CDMA_D_MISC_CFG_0, PROC_PRECISION)) |
+		(conv_op->data_reuse
+		<< SHIFT(CDMA_D_MISC_CFG_0, DATA_REUSE)) |
+		(conv_op->weight_reuse
+		<< SHIFT(CDMA_D_MISC_CFG_0, WEIGHT_REUSE)) |
+		(conv_op->skip_data_rls
+		<< SHIFT(CDMA_D_MISC_CFG_0, SKIP_DATA_RLS)) |
+		(conv_op->skip_weight_rls
+		<< SHIFT(CDMA_D_MISC_CFG_0, SKIP_WEIGHT_RLS));
+	cdma_reg_write(engine, D_MISC_CFG, reg);
+
+	reg = (get_in_format(conv_op->data_format) <<
+		SHIFT(CDMA_D_DATAIN_FORMAT_0, DATAIN_FORMAT)) |
+		(map_img_fmt[conv_op->data_format][0]
+		<< SHIFT(CDMA_D_DATAIN_FORMAT_0, PIXEL_FORMAT)) |
+		(map_pixel[conv_op->pixel_mapping]
+		<< SHIFT(CDMA_D_DATAIN_FORMAT_0, PIXEL_MAPPING)) |
+		(conv_op->pixel_override
+		<< SHIFT(CDMA_D_DATAIN_FORMAT_0, PIXEL_SIGN_OVERRIDE));
+	cdma_reg_write(engine, D_DATAIN_FORMAT, reg);
+
+	reg = ((conv_surface->src_data.width - 1)
+		<< SHIFT(CDMA_D_DATAIN_SIZE_0_0, DATAIN_WIDTH)) |
+		((conv_surface->src_data.height - 1)
+		<< SHIFT(CDMA_D_DATAIN_SIZE_0_0, DATAIN_HEIGHT));
+	cdma_reg_write(engine, D_DATAIN_SIZE_0, reg);
+
+	reg = ((conv_surface->src_data.channel - 1)
+		<< SHIFT(CDMA_D_DATAIN_SIZE_1_0, DATAIN_CHANNEL));
+	cdma_reg_write(engine, D_DATAIN_SIZE_1, reg);
+
+	reg = ((conv_op->input_width_csc - 1)
+		<< SHIFT(CDMA_D_DATAIN_SIZE_EXT_0_0, DATAIN_WIDTH_EXT)) |
+		((conv_op->input_height_csc - 1)
+		<< SHIFT(CDMA_D_DATAIN_SIZE_EXT_0_0, DATAIN_HEIGHT_EXT));
+	cdma_reg_write(engine, D_DATAIN_SIZE_EXT_0, reg);
+
+	reg = (map_ram[conv_surface->src_data.type]
+		<< SHIFT(CDMA_D_DAIN_RAM_TYPE_0, DATAIN_RAM_TYPE));
+	cdma_reg_write(engine, D_DAIN_RAM_TYPE, reg);
+
+	high = upper_32_bits(input_address);
+	low = lower_32_bits(input_address);
+	cdma_reg_write(engine, D_DAIN_ADDR_HIGH_0, high);
+	cdma_reg_write(engine, D_DAIN_ADDR_LOW_0, low);
+
+	high = upper_32_bits((input_address + conv_surface->offset_u));
+	low = lower_32_bits(input_address + conv_surface->offset_u);
+	cdma_reg_write(engine, D_DAIN_ADDR_HIGH_1, high);
+	cdma_reg_write(engine, D_DAIN_ADDR_LOW_1, low);
+
+	cdma_reg_write(engine, D_LINE_STRIDE, conv_surface->src_data.line_stride);
+	cdma_reg_write(engine, D_SURF_STRIDE, conv_surface->src_data.surf_stride);
+	cdma_reg_write(engine, D_LINE_UV_STRIDE, conv_surface->in_line_uv_stride);
+
+	reg = ((conv_surface->src_data.line_stride ==
+			((uint32_t)conv_surface->src_data.width * atom_size))
+		<< SHIFT(CDMA_D_DAIN_MAP_0, LINE_PACKED));
+	reg |= ((conv_surface->src_data.surf_stride ==
+			((uint32_t)(conv_surface->src_data.width *
+			conv_surface->src_data.height) * atom_size))
+		<< SHIFT(CDMA_D_DAIN_MAP_0, SURF_PACKED));
+	cdma_reg_write(engine, D_DAIN_MAP, reg);
+
+	reg = ((conv_op->batch - 1)
+		<< SHIFT(CDMA_D_BATCH_NUMBER_0, BATCHES));
+	cdma_reg_write(engine, D_BATCH_NUMBER, reg);
+
+	cdma_reg_write(engine, D_BATCH_STRIDE, conv_op->batch_stride);
+
+	reg = ((conv_op->entry_per_slice - 1)
+		<< SHIFT(CDMA_D_ENTRY_PER_SLICE_0, ENTRIES));
+	cdma_reg_write(engine, D_ENTRY_PER_SLICE, reg);
+
+	reg = ((conv_op->fetch_grain - 1)
+		<< SHIFT(CDMA_D_FETCH_GRAIN_0, GRAINS));
+	cdma_reg_write(engine, D_FETCH_GRAIN, reg);
+
+	reg = (map_weight_fmt[conv_op->weight_format]
+		<< SHIFT(CDMA_D_WEIGHT_FORMAT_0, WEIGHT_FORMAT));
+	cdma_reg_write(engine, D_WEIGHT_FORMAT, reg);
+
+	reg = ((conv_op->bytes_per_kernel - 1)
+		<< SHIFT(CDMA_D_WEIGHT_SIZE_0_0, BYTE_PER_KERNEL));
+	cdma_reg_write(engine, D_WEIGHT_SIZE_0, reg);
+
+	reg = ((conv_surface->dst_data.channel - 1)
+		<< SHIFT(CDMA_D_WEIGHT_SIZE_1_0, WEIGHT_KERNEL));
+	cdma_reg_write(engine, D_WEIGHT_SIZE_1, reg);
+
+	reg = (map_ram[conv_surface->weight_data.type]
+		<< SHIFT(CDMA_D_WEIGHT_RAM_TYPE_0, WEIGHT_RAM_TYPE));
+	cdma_reg_write(engine, D_WEIGHT_RAM_TYPE, reg);
+
+	high = upper_32_bits(weight_address);
+	low = lower_32_bits(weight_address);
+	cdma_reg_write(engine, D_WEIGHT_ADDR_HIGH, high);
+	cdma_reg_write(engine, D_WEIGHT_ADDR_LOW, low);
+	cdma_reg_write(engine, D_WEIGHT_BYTES, conv_surface->weight_data.size);
+
+	if (conv_op->weight_format == WEIGHT_FORMAT_COMPRESSED) {
+		high = upper_32_bits(wgs_address);
+		low = lower_32_bits(wgs_address);
+		cdma_reg_write(engine, D_WGS_ADDR_HIGH, high);
+		cdma_reg_write(engine, D_WGS_ADDR_LOW, low);
+
+		high = upper_32_bits(wmb_address);
+		low = lower_32_bits(wmb_address);
+		cdma_reg_write(engine, D_WMB_ADDR_HIGH, high);
+		cdma_reg_write(engine, D_WMB_ADDR_LOW, low);
+		cdma_reg_write(engine, D_WMB_BYTES, conv_surface->wmb_data.size);
+	}
+
+	reg = (map_mean[conv_op->mean_format]
+		<< SHIFT(CDMA_D_MEAN_FORMAT_0, MEAN_FORMAT));
+	cdma_reg_write(engine, D_MEAN_FORMAT, reg);
+
+	if (conv_op->mean_format == MEAN_FORMAT_ENABLE) {
+		reg = ((conv_op->mean_ry
+			<< SHIFT(CDMA_D_MEAN_GLOBAL_0_0, MEAN_RY)) &
+			MASK(CDMA_D_MEAN_GLOBAL_0_0, MEAN_RY)) |
+			((conv_op->mean_gu
+			<< SHIFT(CDMA_D_MEAN_GLOBAL_0_0, MEAN_GU)) &
+			MASK(CDMA_D_MEAN_GLOBAL_0_0, MEAN_GU));
+		cdma_reg_write(engine, D_MEAN_GLOBAL_0, reg);
+
+		reg = ((conv_op->mean_bv
+			<< SHIFT(CDMA_D_MEAN_GLOBAL_1_0, MEAN_BV))&
+			MASK(CDMA_D_MEAN_GLOBAL_1_0, MEAN_BV)) |
+			((conv_op->mean_ax
+			<< SHIFT(CDMA_D_MEAN_GLOBAL_1_0, MEAN_AX))&
+			MASK(CDMA_D_MEAN_GLOBAL_1_0, MEAN_AX));
+		cdma_reg_write(engine, D_MEAN_GLOBAL_1, reg);
+	}
+
+	if (conv_op->in_cvt.enable) {
+		reg = ((FIELD_ENUM(CDMA_D_CVT_CFG_0, CVT_EN, ENABLE))
+			<< SHIFT(CDMA_D_CVT_CFG_0, CVT_EN)) |
+			(conv_op->in_cvt.truncate
+			<< SHIFT(CDMA_D_CVT_CFG_0, CVT_TRUNCATE));
+		cdma_reg_write(engine, D_CVT_CFG, reg);
+		cdma_reg_write(engine, D_CVT_OFFSET, conv_op->in_cvt.offset);
+		cdma_reg_write(engine, D_CVT_SCALE, conv_op->in_cvt.scale);
+	} else {
+		reg = ((FIELD_ENUM(CDMA_D_CVT_CFG_0, CVT_EN, DISABLE))
+			<< SHIFT(CDMA_D_CVT_CFG_0, CVT_EN));
+		cdma_reg_write(engine, D_CVT_CFG, reg);
+	}
+
+	reg = ((conv_op->conv_stride_x - 1)
+		<< SHIFT(CDMA_D_CONV_STRIDE_0, CONV_X_STRIDE)) |
+		((conv_op->conv_stride_y - 1)
+		<< SHIFT(CDMA_D_CONV_STRIDE_0, CONV_Y_STRIDE));
+	cdma_reg_write(engine, D_CONV_STRIDE, reg);
+
+	reg = (conv_op->pad_x_left <<
+		SHIFT(CDMA_D_ZERO_PADDING_0, PAD_LEFT)) |
+		(conv_op->pad_x_right
+		<< SHIFT(CDMA_D_ZERO_PADDING_0, PAD_RIGHT)) |
+		(conv_op->pad_y_top
+		<< SHIFT(CDMA_D_ZERO_PADDING_0, PAD_TOP)) |
+		(conv_op->pad_y_bottom
+		<< SHIFT(CDMA_D_ZERO_PADDING_0, PAD_BOTTOM));
+	cdma_reg_write(engine, D_ZERO_PADDING,   reg);
+
+	reg = conv_op->pad_val <<
+		SHIFT(CDMA_D_ZERO_PADDING_VALUE_0, PAD_VALUE) &
+		MASK(CDMA_D_ZERO_PADDING_VALUE_0, PAD_VALUE);
+	cdma_reg_write(engine, D_ZERO_PADDING_VALUE, reg);
+	reg = ((conv_op->weight_bank - 1)
+		<< SHIFT(CDMA_D_BANK_0, WEIGHT_BANK)) |
+		((conv_op->data_bank - 1)
+		<< SHIFT(CDMA_D_BANK_0, DATA_BANK));
+	cdma_reg_write(engine, D_BANK, reg);
+
+exit:
+	return ret;
+}
+
+int
+dla_conv_is_ready(struct dla_processor *processor,
+				  struct dla_processor_group *group)
+{
+	return 1;
+}
+
+void
+dla_conv_dump_config(struct dla_processor_group *group)
+{
+	struct dla_conv_op_desc *conv_op;
+	struct dla_conv_surface_desc *conv_surface;
+
+	conv_surface = &group->surface_desc->conv_surface;
+	conv_op = &group->operation_desc->conv_op;
+}
+
+int
+dla_conv_program(struct dla_engine *engine, struct dla_processor_group *group)
+{
+	int32_t ret;
+
+	ret = processor_conv_program(engine, group);
+	if (ret)
+		goto exit;
+
+exit:
+	return ret;
+}
drivers/gpu/drm/nvdla/nvdla_drm.c new file +252 -0
new file mode 100644
index 000000000000..9217eee1de3b
@@ -0,0 +1,695 @@
+// SPDX-License-Identifier: BSD-3-Clause OR GPL-2.0
+/*
+ * Copyright (C) 2017-2018 NVIDIA CORPORATION
+ * Copyright (C) 2022 Cai Huoqing
+ */
+
+#include <linux/dma-buf.h>
+#include <linux/dma-mapping.h>
+#include <linux/fs.h>
+#include <linux/interrupt.h>
+#include <linux/irq.h>
+#include <linux/irqdomain.h>
+#include <linux/module.h>
+#include <linux/of.h>
+#include <linux/of_device.h>
+#include <linux/of_irq.h>
+#include <linux/of_platform.h>
+#include <linux/platform_device.h>
+#include <linux/printk.h>
+#include <linux/slab.h>
+#include <linux/spinlock.h>
+#include <linux/time.h>
+#include <linux/uaccess.h>
+#include <linux/types.h>
+
+#include "nvdla_drm.h"
+#include "nvdla_ioctl.h"
+#include "nvdla_engine.h"
+
+static struct nvdla_config nvdla_config_os_initial = {
+	.atom_size = 32,
+	.bdma_enable = true,
+	.rubik_enable = true,
+	.weight_compress_support = true,
+};
+
+static struct nvdla_config nvdla_config_small = {
+	//.atom_size = 8,
+	.atom_size = 32,  // nv_large config
+	.bdma_enable = false,
+	.rubik_enable = false,
+	.weight_compress_support = false,
+};
+
+int64_t dla_get_time_us(void)
+{
+	return ktime_get_ns() / NSEC_PER_USEC;
+}
+
+void dla_reg_write(void *driver_context, uint32_t addr, uint32_t reg)
+{
+	struct nvdla_device *nvdla_dev =
+			(struct nvdla_device *)driver_context;
+
+	if (!nvdla_dev)
+		return;
+
+	writel(reg, nvdla_dev->base + addr);
+}
+
+uint32_t dla_reg_read(void *driver_context, uint32_t addr)
+{
+	struct nvdla_device *nvdla_dev =
+			(struct nvdla_device *)driver_context;
+
+	if (!nvdla_dev)
+		return 0;
+
+	return readl(nvdla_dev->base + addr);
+}
+
+static irqreturn_t nvdla_engine_isr(int32_t irq, void *data)
+{
+	unsigned long flags;
+	uint32_t mask;
+	uint32_t reg;
+	struct dla_processor *processor = NULL;
+	struct dla_processor_group *group;
+	struct dla_engine *engine;
+	struct nvdla_device *nvdla_dev = (struct nvdla_device *)data;
+
+	if (!nvdla_dev)
+		return IRQ_NONE;
+
+	engine = nvdla_dev->engine_context;
+	spin_lock_irqsave(&nvdla_dev->nvdla_lock, flags);
+
+	mask = glb_reg_read(engine, S_INTR_MASK);
+	reg = glb_reg_read(engine, S_INTR_STATUS);
+
+	if (reg & MASK(GLB_S_INTR_STATUS_0, CACC_DONE_STATUS0)) {
+		processor = &engine->processors[DLA_OP_CONV];
+		group = &processor->groups[0];
+		group->events |= (1 << DLA_EVENT_OP_COMPLETED);
+	}
+	if (reg & MASK(GLB_S_INTR_STATUS_0, CACC_DONE_STATUS1)) {
+		processor = &engine->processors[DLA_OP_CONV];
+		group = &processor->groups[1];
+		group->events |= (1 << DLA_EVENT_OP_COMPLETED);
+	}
+	if (reg & MASK(GLB_S_INTR_STATUS_0, SDP_DONE_STATUS0)) {
+		processor = &engine->processors[DLA_OP_SDP];
+		group = &processor->groups[0];
+		group->events |= (1 << DLA_EVENT_OP_COMPLETED);
+	}
+	if (reg & MASK(GLB_S_INTR_STATUS_0, SDP_DONE_STATUS1)) {
+		processor = &engine->processors[DLA_OP_SDP];
+		group = &processor->groups[1];
+		group->events |= (1 << DLA_EVENT_OP_COMPLETED);
+	}
+	if (reg & MASK(GLB_S_INTR_STATUS_0, CDP_DONE_STATUS0)) {
+		processor = &engine->processors[DLA_OP_CDP];
+		group = &processor->groups[0];
+		group->events |= (1 << DLA_EVENT_OP_COMPLETED);
+	}
+	if (reg & MASK(GLB_S_INTR_STATUS_0, CDP_DONE_STATUS1)) {
+		processor = &engine->processors[DLA_OP_CDP];
+		group = &processor->groups[1];
+		group->events |= (1 << DLA_EVENT_OP_COMPLETED);
+	}
+	if (reg & MASK(GLB_S_INTR_STATUS_0, RUBIK_DONE_STATUS0)) {
+		processor = &engine->processors[DLA_OP_RUBIK];
+		group = &processor->groups[0];
+		group->events |= (1 << DLA_EVENT_OP_COMPLETED);
+	}
+	if (reg & MASK(GLB_S_INTR_STATUS_0, RUBIK_DONE_STATUS1)) {
+		processor = &engine->processors[DLA_OP_RUBIK];
+		group = &processor->groups[1];
+		group->events |= (1 << DLA_EVENT_OP_COMPLETED);
+	}
+	if (reg & MASK(GLB_S_INTR_STATUS_0, PDP_DONE_STATUS0)) {
+		processor = &engine->processors[DLA_OP_PDP];
+		group = &processor->groups[0];
+		group->events |= (1 << DLA_EVENT_OP_COMPLETED);
+	}
+	if (reg & MASK(GLB_S_INTR_STATUS_0, PDP_DONE_STATUS1)) {
+		processor = &engine->processors[DLA_OP_PDP];
+		group = &processor->groups[1];
+		group->events |= (1 << DLA_EVENT_OP_COMPLETED);
+	}
+	if (reg & MASK(GLB_S_INTR_STATUS_0, BDMA_DONE_STATUS0)) {
+		processor = &engine->processors[DLA_OP_BDMA];
+		group = &processor->groups[0];
+		group->events |= (1 << DLA_EVENT_OP_COMPLETED);
+	}
+	if (reg & MASK(GLB_S_INTR_STATUS_0, BDMA_DONE_STATUS1)) {
+		processor = &engine->processors[DLA_OP_BDMA];
+		group = &processor->groups[1];
+		group->events |= (1 << DLA_EVENT_OP_COMPLETED);
+	}
+	if (reg & MASK(GLB_S_INTR_STATUS_0, CDMA_DAT_DONE_STATUS0)) {
+		processor = &engine->processors[DLA_OP_CONV];
+		group = &processor->groups[0];
+		group->events |= (1 << DLA_EVENT_CDMA_DT_DONE);
+	}
+	if (reg & MASK(GLB_S_INTR_STATUS_0, CDMA_DAT_DONE_STATUS1)) {
+		processor = &engine->processors[DLA_OP_CONV];
+		group = &processor->groups[1];
+		group->events |= (1 << DLA_EVENT_CDMA_DT_DONE);
+	}
+	if (reg & MASK(GLB_S_INTR_STATUS_0, CDMA_WT_DONE_STATUS0)) {
+		processor = &engine->processors[DLA_OP_CONV];
+		group = &processor->groups[0];
+		group->events |= (1 << DLA_EVENT_CDMA_WT_DONE);
+	}
+	if (reg & MASK(GLB_S_INTR_STATUS_0, CDMA_WT_DONE_STATUS1)) {
+		processor = &engine->processors[DLA_OP_CONV];
+		group = &processor->groups[1];
+		group->events |= (1 << DLA_EVENT_CDMA_WT_DONE);
+	}
+
+	glb_reg_write(engine, S_INTR_STATUS, reg);
+	mask = glb_reg_read(engine, S_INTR_MASK);
+	reg = glb_reg_read(engine, S_INTR_STATUS);
+
+	complete(&nvdla_dev->event_notifier);
+	spin_unlock_irqrestore(&nvdla_dev->nvdla_lock, flags);
+
+	return IRQ_HANDLED;
+}
+
+static int32_t dla_read_dma_address(void *driver_context, void *task_data,
+						int16_t index, void *dst)
+{
+	int32_t ret = 0;
+	struct nvdla_mem_handle *handles;
+	dma_addr_t *phys_addr = (dma_addr_t *)(dst);
+	struct nvdla_device *nvdla_dev =
+			(struct nvdla_device *)driver_context;
+	struct nvdla_task *task = (struct nvdla_task *)task_data;
+
+	if (index == -1 || index > task->num_addresses)
+		return -EINVAL;
+
+	handles = (struct nvdla_mem_handle *)task->address_list;
+	ret = nvdla_gem_dma_addr(nvdla_dev->drm, task->file,
+					handles[index].handle,
+					phys_addr);
+
+	/* Add offset to IOVA address */
+	*phys_addr = *phys_addr + handles[index].offset;
+
+	return ret;
+}
+
+static int32_t dla_read_cpu_address(void *driver_context, void *task_data,
+						int16_t index, void *dst)
+{
+	uint64_t *temp = (uint64_t *)dst;
+	struct nvdla_task *task = (struct nvdla_task *)task_data;
+
+	if (index == -1 || index > task->num_addresses)
+		return -EINVAL;
+
+	*temp = (uint64_t)index;
+	return 0;
+}
+
+int32_t dla_get_dma_address(void *driver_context, void *task_data,
+					int16_t index, void *dst_ptr,
+					uint32_t destination)
+{
+	int32_t ret = 0;
+
+	if (destination == DESTINATION_PROCESSOR) {
+		ret = dla_read_cpu_address(driver_context, task_data,
+						index, dst_ptr);
+	} else if (destination == DESTINATION_DMA) {
+		ret = dla_read_dma_address(driver_context, task_data,
+						index, dst_ptr);
+	} else {
+		ret = -EINVAL;
+	}
+
+	return ret;
+}
+
+int32_t dla_data_write(void *driver_context, void *task_data,
+				void *src, uint64_t dst,
+				uint32_t size, uint64_t offset)
+{
+	int32_t ret;
+	void *ptr = NULL;
+	struct dma_buf *buf;
+	struct iosys_map map;
+	struct nvdla_mem_handle *handles;
+	struct nvdla_task *task = (struct nvdla_task *)task_data;
+
+	handles = task->address_list;
+	buf = dma_buf_get(handles[dst].handle);
+	if (IS_ERR(buf)) {
+		pr_err("%s: Failed get dma_buf for handle=%d\n", __func__,

drivers/gpu/drm/nvdla/nvdla_drm.c line 262

Quoted 10 lines
drivers/gpu/drm/nvdla/nvdla_drm.c new file +10 -0
+						handles[dst].handle);
+		return -EFAULT;
+	}
+
+	ret = dma_buf_begin_cpu_access(buf, DMA_BIDIRECTIONAL);
+	if (ret)
+		goto put_dma_buf;
+
+	ret = dma_buf_vmap(buf, &map);
+	ptr = ret ? NULL : map.vaddr;

Never extract the pointer's address without good reason. You don't know if this points to a location in I/O memory.

drivers/gpu/drm/nvdla/nvdla_drm.c line 263

Quoted 1 line
drivers/gpu/drm/nvdla/nvdla_drm.c new file +1 -0
+	if (!ptr) {

Simply test for ret here.

drivers/gpu/drm/nvdla/nvdla_drm.c line 266

Quoted 3 lines
drivers/gpu/drm/nvdla/nvdla_drm.c new file +3 -0
+		pr_err("%s: Failed to vmap dma_buf for handle=%d\n", __func__,
+						handles[dst].handle);
+		ret = -ENOMEM;

You already got an errno code. Don't override it.

drivers/gpu/drm/nvdla/nvdla_drm.c line 271

Quoted 5 lines
drivers/gpu/drm/nvdla/nvdla_drm.c new file +5 -0
+		goto end_cpu_access;
+	}
+
+
+	memcpy((void *)((uint8_t *)ptr + offset), src, size);

Use iosys_map_memcpy_to() here. It does the right thing

drivers/gpu/drm/nvdla/nvdla_drm.c line 273

Quoted 2 lines
drivers/gpu/drm/nvdla/nvdla_drm.c new file +2 -0
+
+	dma_buf_vunmap(buf, ptr);

You have to pass map as the second argument.

Quoted 32 lines
drivers/gpu/drm/nvdla/nvdla_drm.c new file +32 -0
+
+end_cpu_access:
+	dma_buf_end_cpu_access(buf, DMA_BIDIRECTIONAL);
+
+put_dma_buf:
+	dma_buf_put(buf);
+
+	return ret;
+}
+
+int32_t dla_data_read(void *driver_context, void *task_data,
+				uint64_t src, void *dst,
+				uint32_t size, uint64_t offset)
+{
+	int32_t ret;
+	void *ptr = NULL;
+	struct dma_buf *buf;
+	struct iosys_map map;
+	struct nvdla_mem_handle *handles;
+	struct nvdla_task *task = (struct nvdla_task *)task_data;
+
+	handles = task->address_list;
+
+	buf = dma_buf_get(handles[src].handle);
+	if (IS_ERR(buf)) {
+		pr_err("%s: Failed get dma_buf for handle=%d\n", __func__,
+						handles[src].handle);
+		return -EFAULT;
+	}
+
+	ret = dma_buf_begin_cpu_access(buf, DMA_BIDIRECTIONAL);
+	if (ret)

drivers/gpu/drm/nvdla/nvdla_drm.c line 315

Quoted 10 lines
drivers/gpu/drm/nvdla/nvdla_drm.c new file +10 -0
+		goto put_dma_buf;
+
+	ret = dma_buf_vmap(buf, &map);
+	ptr = ret ? NULL : map.vaddr;
+	if (!ptr) {
+		pr_err("%s: Failed to vmap dma_buf for handle=%d\n", __func__,
+						handles[src].handle);
+		ret = -ENOMEM;
+		goto end_cpu_access;
+	}

All the same problems as in dla_data_write().

drivers/gpu/drm/nvdla/nvdla_drm.c line 317

Quoted 2 lines
drivers/gpu/drm/nvdla/nvdla_drm.c new file +2 -0
+
+	memcpy(dst, (void *)(((uint8_t *)ptr) + offset), size);

Use iosys_map_memcpy_from() here.

drivers/gpu/drm/nvdla/nvdla_drm.c line 319

Quoted 2 lines
drivers/gpu/drm/nvdla/nvdla_drm.c new file +2 -0
+
+	dma_buf_vunmap(buf, ptr);

'map' instead of 'ptr'

drivers/gpu/drm/nvdla/nvdla_gem.c

Quoted 1,137 lines
drivers/gpu/drm/nvdla/nvdla_drm.c new file +376 -0
+
+end_cpu_access:
+	dma_buf_end_cpu_access(buf, DMA_BIDIRECTIONAL);
+
+put_dma_buf:
+	dma_buf_put(buf);
+
+	return ret;
+}
+
+int32_t nvdla_task_submit(struct nvdla_device *nvdla_dev, struct nvdla_task *task)
+{
+	int32_t err = 0;
+	uint32_t task_complete = 0;
+
+	//trace_printk("__nvdla_task_submit_entry\n");
+	nvdla_dev->task = task;
+
+	err = dla_execute_task(nvdla_dev->engine_context, (void *)task, nvdla_dev->config_data);
+	if (err) {
+		pr_err("Task execution failed\n");
+		return err;
+	}
+
+	pr_debug("Wait for task complete\n");
+
+	while (1) {
+		unsigned long flags;
+
+		wait_for_completion(&nvdla_dev->event_notifier);
+
+		spin_lock_irqsave(&nvdla_dev->nvdla_lock, flags);
+
+		err = dla_process_events(nvdla_dev->engine_context, &task_complete);
+
+		spin_unlock_irqrestore(&nvdla_dev->nvdla_lock, flags);
+
+		if (err || task_complete)
+			break;
+	}
+
+	pr_debug("Task complete\n");
+	dla_clear_task(nvdla_dev->engine_context);
+	//trace_printk("__nvdla_task_submit_exit\n");
+
+	return err;
+}
+
+static union dla_operation_container operation_desc[DLA_OP_NUM][DLA_NUM_GROUPS];
+static union dla_surface_container surface_desc[DLA_OP_NUM][DLA_NUM_GROUPS];
+
+static struct dla_task global_task;
+
+static struct dla_engine engine = {
+	.processors[DLA_OP_BDMA] = {
+		.name = "BDMA",
+		.op_type = DLA_OP_BDMA,
+		.program = dla_bdma_program,
+		.enable = dla_bdma_enable,
+		.set_producer = dla_bdma_set_producer,
+		.is_ready = dla_bdma_is_ready,
+		.dump_config = dla_bdma_dump_config,
+		.rdma_check = dla_bdma_rdma_check,
+		.consumer_ptr = 0,
+		.roi_index = 0,
+		.group_status = 0,
+		.rdma_status = 0,
+		.last_group = 1,
+		.groups[0] = {
+			.id = 0,
+			.rdma_id = 0,
+			.active = 0,
+			.events = 0,
+			.roi_index = 0,
+			.is_rdma_needed = 0,
+			.lut_index = -1,
+			.operation_desc = &operation_desc[DLA_OP_BDMA][0],
+			.surface_desc = &surface_desc[DLA_OP_BDMA][0],
+		},
+		.groups[1] = {
+			.id = 1,
+			.rdma_id = 0,
+			.active = 0,
+			.events = 0,
+			.roi_index = 0,
+			.is_rdma_needed = 0,
+			.lut_index = -1,
+			.operation_desc = &operation_desc[DLA_OP_BDMA][1],
+			.surface_desc = &surface_desc[DLA_OP_BDMA][1],
+		},
+	},
+	.processors[DLA_OP_CONV] = {
+		.name = "Convolution",
+		.op_type = DLA_OP_CONV,
+		.program = dla_conv_program,
+		.enable = dla_conv_enable,
+		.set_producer = dla_conv_set_producer,
+		.is_ready = dla_conv_is_ready,
+		.dump_config = dla_conv_dump_config,
+		.rdma_check = dla_conv_rdma_check,
+		.consumer_ptr = 0,
+		.roi_index = 0,
+		.group_status = 0,
+		.rdma_status = 0,
+		.last_group = 1,
+		.groups[0] = {
+			.id = 0,
+			.rdma_id = 0,
+			.active = 0,
+			.events = 0,
+			.roi_index = 0,
+			.is_rdma_needed = 0,
+			.lut_index = -1,
+			.operation_desc = &operation_desc[DLA_OP_CONV][0],
[… more lines in the raw message]

Lines of this file are not shown. View the raw message.

drivers/gpu/drm/nvdla/nvdla_drm.h new file +127 -0
new file mode 100644
index 000000000000..356fc399accb

Lines of this file are not shown. View the raw message.

drivers/gpu/drm/nvdla/nvdla_engine.c new file +233 -0
new file mode 100644
index 000000000000..26ebd6c362b0

Lines of this file are not shown. View the raw message.

drivers/gpu/drm/nvdla/nvdla_engine.h new file +272 -0
new file mode 100644
index 000000000000..1976f5919543

Lines of this file are not shown. View the raw message.

drivers/gpu/drm/nvdla/nvdla_gem.c new file +105 -0
new file mode 100644
index 000000000000..cccf6d01a564

Lines of this file are not shown. View the raw message.

The rest of this diff is not shown. View the raw message.

Store an iosys-map address in nobj and initialize it with iosys_map_set_vaddr(); or iosys_map_set_vaddr_iomem() if you're working with I/O memory.

drivers/gpu/drm/nvdla/nvdla_gem.c

Quoted 85 lines
drivers/gpu/drm/nvdla/nvdla_gem.c new file +85 -0

Lines of this file are not shown. View the raw message.

The rest of this diff is not shown. View the raw message.

Instead of kvaddr, store the pointer as struct iosys_map. Then simply copy it here, as in

  *map = nobj->map;

drivers/gpu/drm/nvdla/nvdla_gem.c

Quoted 18 lines
drivers/gpu/drm/nvdla/nvdla_gem.c new file +18 -0

Lines of this file are not shown. View the raw message.

The rest of this diff is not shown. View the raw message.

It's cleaner to do this as

vma->vm_pgoff -= drm_vma_node_start(&obj->vma_node)

drivers/gpu/drm/nvdla/nvdla_gem.c

Quoted 84 lines
drivers/gpu/drm/nvdla/nvdla_gem.c new file +84 -0

Lines of this file are not shown. View the raw message.

The rest of this diff is not shown. View the raw message.

I don't understand these two lines. This is part of what drm_gem_mmap() does. It shouldn't be necessary here.

drivers/gpu/drm/nvdla/nvdla_gem.c

Quoted 42 lines
drivers/gpu/drm/nvdla/nvdla_gem.c new file +42 -0

Lines of this file are not shown. View the raw message.

The rest of this diff is not shown. View the raw message.

It should be fine to use drm_gem_mmap here. Then you should use DEFINE_DRM_GEM_FOPS() to define nvdla_drm_fops.

drivers/gpu/drm/nvdla/nvdla_gem.c

Quoted 22 lines
drivers/gpu/drm/nvdla/nvdla_gem.c new file +22 -0

Lines of this file are not shown. View the raw message.

The rest of this diff is not shown. View the raw message.

Use drm_gem_prime_mmap() here.

Some context: the situation with these mmap functions has been confusing and inconsistent among DMR drivers. But we cleaned it up so that you only have to provide a minimal implementation of struct drm_gem_object_funcs.mmap. All other mmap callbacks can then be filled with standard DRM helpers.

Quoted 9,037 lines
drivers/gpu/drm/nvdla/nvdla_gem.c new file +37 -0

Lines of this file are not shown. View the raw message.

drivers/gpu/drm/nvdla/nvdla_ioctl.h new file +99 -0
new file mode 100644
index 000000000000..e21d4007b7f4

Lines of this file are not shown. View the raw message.

drivers/gpu/drm/nvdla/nvdla_pdp.c new file +446 -0
new file mode 100644
index 000000000000..2ed0ac0f021a

Lines of this file are not shown. View the raw message.

drivers/gpu/drm/nvdla/nvdla_reg.h new file +6411 -0
new file mode 100644
index 000000000000..5ca2897405bc

Lines of this file are not shown. View the raw message.

drivers/gpu/drm/nvdla/nvdla_rubik.c new file +217 -0
new file mode 100644
index 000000000000..3bfb3b0cb786

Lines of this file are not shown. View the raw message.

drivers/gpu/drm/nvdla/nvdla_sched.h new file +52 -0
new file mode 100644
index 000000000000..d3a198127006

Lines of this file are not shown. View the raw message.

drivers/gpu/drm/nvdla/nvdla_scheduler.c new file +1005 -0
new file mode 100644
index 000000000000..b814077478c6

Lines of this file are not shown. View the raw message.

drivers/gpu/drm/nvdla/nvdla_sdp.c new file +728 -0
new file mode 100644
index 000000000000..655ec7195ee9

Lines of this file are not shown. View the raw message.

The rest of this diff is not shown. View the raw message.

Signature · 6 lines
Thomas Zimmermann
Graphics Driver Developer
SUSE Software Solutions Germany GmbH
Maxfeldstr. 5, 90409 Nürnberg, Germany
(HRB 36809, AG Nürnberg)
Geschäftsführer: Ivo Totev

Part of the diff is not shown here. View the raw message.

#8 lore raw
In reply to Cai Huoqing
Quoted 5 lines · Cai Huoqing

The NVIDIA Deep Learning Accelerator (NVDLA) is an open source IP which is integrated into NVIDIA Jetson AGX Xavier,
so add driver support for this accelerator.

Hi, all
requests for more comments

many thanks
CAI

Quoted 13,423 lines
---
 drivers/gpu/drm/Kconfig                 |    2 +
 drivers/gpu/drm/Makefile                |    1 +
 drivers/gpu/drm/nvdla/Kconfig           |    8 +
 drivers/gpu/drm/nvdla/Makefile          |   19 +
 drivers/gpu/drm/nvdla/nvdla_bdma.c      |  200 +
 drivers/gpu/drm/nvdla/nvdla_cache.c     |  215 +
 drivers/gpu/drm/nvdla/nvdla_cdp.c       |  300 ++
 drivers/gpu/drm/nvdla/nvdla_common.c    |  295 ++
 drivers/gpu/drm/nvdla/nvdla_common.h    |  835 +++
 drivers/gpu/drm/nvdla/nvdla_conv.c      |  683 +++
 drivers/gpu/drm/nvdla/nvdla_drm.c       |  695 +++
 drivers/gpu/drm/nvdla/nvdla_drm.h       |  127 +
 drivers/gpu/drm/nvdla/nvdla_engine.c    |  233 +
 drivers/gpu/drm/nvdla/nvdla_engine.h    |  272 +
 drivers/gpu/drm/nvdla/nvdla_gem.c       |  393 ++
 drivers/gpu/drm/nvdla/nvdla_ioctl.h     |   99 +
 drivers/gpu/drm/nvdla/nvdla_pdp.c       |  446 ++
 drivers/gpu/drm/nvdla/nvdla_reg.h       | 6411 +++++++++++++++++++++++
 drivers/gpu/drm/nvdla/nvdla_rubik.c     |  217 +
 drivers/gpu/drm/nvdla/nvdla_sched.h     |   52 +
 drivers/gpu/drm/nvdla/nvdla_scheduler.c | 1005 ++++
 drivers/gpu/drm/nvdla/nvdla_sdp.c       |  728 +++
 22 files changed, 13236 insertions(+)
 create mode 100644 drivers/gpu/drm/nvdla/Kconfig
 create mode 100644 drivers/gpu/drm/nvdla/Makefile
 create mode 100644 drivers/gpu/drm/nvdla/nvdla_bdma.c
 create mode 100644 drivers/gpu/drm/nvdla/nvdla_cache.c
 create mode 100644 drivers/gpu/drm/nvdla/nvdla_cdp.c
 create mode 100644 drivers/gpu/drm/nvdla/nvdla_common.c
 create mode 100644 drivers/gpu/drm/nvdla/nvdla_common.h
 create mode 100644 drivers/gpu/drm/nvdla/nvdla_conv.c
 create mode 100644 drivers/gpu/drm/nvdla/nvdla_drm.c
 create mode 100644 drivers/gpu/drm/nvdla/nvdla_drm.h
 create mode 100644 drivers/gpu/drm/nvdla/nvdla_engine.c
 create mode 100644 drivers/gpu/drm/nvdla/nvdla_engine.h
 create mode 100644 drivers/gpu/drm/nvdla/nvdla_gem.c
 create mode 100644 drivers/gpu/drm/nvdla/nvdla_ioctl.h
 create mode 100644 drivers/gpu/drm/nvdla/nvdla_pdp.c
 create mode 100644 drivers/gpu/drm/nvdla/nvdla_reg.h
 create mode 100644 drivers/gpu/drm/nvdla/nvdla_rubik.c
 create mode 100644 drivers/gpu/drm/nvdla/nvdla_sched.h
 create mode 100644 drivers/gpu/drm/nvdla/nvdla_scheduler.c
 create mode 100644 drivers/gpu/drm/nvdla/nvdla_sdp.c
drivers/gpu/drm/Kconfig +2 -0
index 5133c3f028ab..a55cff374abd 100644
@@ -409,6 +409,8 @@ source "drivers/gpu/drm/solomon/Kconfig"
 
 source "drivers/gpu/drm/sprd/Kconfig"
 
+source "drivers/gpu/drm/nvdla/Kconfig"
+
 config DRM_HYPERV
 	tristate "DRM Support for Hyper-V synthetic video device"
 	depends on DRM && PCI && MMU && HYPERV
drivers/gpu/drm/Makefile +1 -0
index c2ef5f9fce54..8fa3537f308a 100644
@@ -134,3 +134,4 @@ obj-y			+= gud/
 obj-$(CONFIG_DRM_HYPERV) += hyperv/
 obj-y			+= solomon/
 obj-$(CONFIG_DRM_SPRD) += sprd/
+obj-$(CONFIG_DRM_NVDLA) += nvdla/
drivers/gpu/drm/nvdla/Kconfig new file +8 -0
new file mode 100644
index 000000000000..11c04f5da877
@@ -0,0 +1,8 @@
+# SPDX-License-Identifier: GPL-2.0-only
+config DRM_NVDLA
+	tristate "NVDLA DRM"
+	depends on DRM
+	select DRM_GEM_CMA_HELPER
+	help
+	  Choose this option for open-source NVIDIA DLA support.
+	  If M is selected the module will be called nvdla-drm.
drivers/gpu/drm/nvdla/Makefile new file +19 -0
new file mode 100644
index 000000000000..74f37d258f8d
@@ -0,0 +1,19 @@
+
+# SPDX-License-Identifier: GPL-2.0
+nvdla-drm-y := \
+	nvdla_drm.o \
+	nvdla_gem.o \
+	nvdla_scheduler.o \
+	nvdla_engine.o \
+	nvdla_bdma.o \
+	nvdla_conv.o \
+	nvdla_sdp.o \
+	nvdla_cdp.o \
+	nvdla_pdp.o \
+	nvdla_rubik.o \
+	nvdla_cache.o \
+	nvdla_common.o \
+	nvdla_engine_data.o \
+	nvdla_engine_debug.o \
+
+obj-$(CONFIG_DRM_NVDLA) += nvdla-drm.o
drivers/gpu/drm/nvdla/nvdla_bdma.c new file +200 -0
new file mode 100644
index 000000000000..225613f27acf
@@ -0,0 +1,200 @@
+// SPDX-License-Identifier: BSD-3-Clause OR GPL-2.0
+/*
+ * Copyright (C) 2017-2018 NVIDIA CORPORATION
+ * Copyright (C) 2022 Cai Huoqing
+ */
+
+#include "nvdla_common.h"
+#include "nvdla_drm.h"
+#include "nvdla_reg.h"
+#include "nvdla_engine.h"
+
+static const uint8_t map_mem[] = {
+	FIELD_ENUM(BDMA_CFG_CMD_0, SRC_RAM_TYPE, MC),
+	FIELD_ENUM(BDMA_CFG_CMD_0, SRC_RAM_TYPE, CVSRAM),
+};
+
+void
+dla_bdma_set_producer(struct dla_engine *engine, int32_t group_id, int32_t rdma_group_id)
+{
+	/**
+	 * There is no producer bit for BDMA operation,
+	 * interrupt pointer decides which outstanding request
+	 * to use for this BDMA operation
+	 */
+}
+
+int
+dla_bdma_enable(struct dla_engine *engine, struct dla_processor_group *group)
+{
+	if (group->surface_desc->bdma_surface.num_transfers == (uint16_t)0) {
+		group->events |= ((uint8_t)1 << DLA_EVENT_OP_COMPLETED);
+		goto exit;
+	}
+
+	/**
+	 * Launch BDMA transfer
+	 */
+	if (group->id == 0)
+		bdma_reg_write(engine, CFG_LAUNCH0, FIELD_ENUM(BDMA_CFG_LAUNCH0_0,
+							GRP0_LAUNCH, YES));
+	else
+		bdma_reg_write(engine, CFG_LAUNCH1, FIELD_ENUM(BDMA_CFG_LAUNCH1_0,
+							GRP1_LAUNCH, YES));
+
+exit:
+	return 0;
+}
+
+void
+dla_bdma_rdma_check(struct dla_processor_group *group)
+{
+	group->is_rdma_needed = 0;
+}
+
+/**
+ * Program BDMA slot for transfer
+ */
+static int32_t
+processor_bdma_program_slot(struct dla_engine *engine,
+							struct dla_bdma_surface_desc *bdma_surface,
+							struct dla_bdma_transfer_desc *transfer)
+{
+	int32_t ret = 0;
+	uint64_t source_addr = 0;
+	uint64_t destination_addr = 0;
+	uint32_t high, low, reg;
+	uint8_t  bdma_free_slots = 0;
+
+	/* make sure there're enough free slots */
+	if (bdma_free_slots <= 0) {
+		do {
+			reg = bdma_reg_read(engine, STATUS);
+			reg = (reg & MASK(BDMA_STATUS_0, FREE_SLOT)) >>
+					SHIFT(BDMA_STATUS_0, FREE_SLOT);
+		} while (reg == 0);
+		bdma_free_slots = (uint8_t)reg;
+	}
+
+	dla_get_dma_address(engine->driver_context, engine->task->task_data,
+						transfer->source_address,
+						(void *)&source_addr,
+						DESTINATION_DMA);
+	dla_get_dma_address(engine->driver_context, engine->task->task_data,
+						transfer->destination_address,
+						(void *)&destination_addr,
+						DESTINATION_DMA);
+
+	ASSERT_GOTO((transfer->line_repeat <= 8192),
+				ret, -EINVAL, exit);
+	ASSERT_GOTO((transfer->surface_repeat <= 8192),
+				ret, -EINVAL, exit);
+	ASSERT_GOTO((transfer->line_size % 32) == 0,
+				ret, -EINVAL, exit);
+	ASSERT_GOTO(transfer->source_line >= transfer->line_size,
+				ret, -EINVAL, exit);
+	ASSERT_GOTO(transfer->destination_line >= transfer->line_size,
+				ret, -EINVAL, exit);
+	ASSERT_GOTO(transfer->source_surface >=
+			(transfer->source_line * transfer->line_repeat),
+				ret, -EINVAL, exit);
+	ASSERT_GOTO(transfer->destination_surface >=
+			(transfer->destination_line * transfer->line_repeat),
+				ret, -EINVAL, exit);
+
+	/* config registers */
+	high = upper_32_bits(source_addr);
+	low = lower_32_bits(source_addr);
+	bdma_reg_write(engine, CFG_SRC_ADDR_LOW, low);
+	bdma_reg_write(engine, CFG_SRC_ADDR_HIGH, high);
+	high = upper_32_bits(destination_addr);
+	low = lower_32_bits(destination_addr);
+	bdma_reg_write(engine, CFG_DST_ADDR_LOW, low);
+	bdma_reg_write(engine, CFG_DST_ADDR_HIGH, high);
+	bdma_reg_write(engine, CFG_LINE, (transfer->line_size >> 5) - 1);
+	reg = (map_mem[bdma_surface->source_type] <<
+				SHIFT(BDMA_CFG_CMD_0, SRC_RAM_TYPE)) |
+		(map_mem[bdma_surface->destination_type] <<
+				SHIFT(BDMA_CFG_CMD_0, DST_RAM_TYPE));
+	bdma_reg_write(engine, CFG_CMD, reg);
+	bdma_reg_write(engine, CFG_LINE_REPEAT, transfer->line_repeat - 1);
+	bdma_reg_write(engine, CFG_SRC_LINE, transfer->source_line);
+	bdma_reg_write(engine, CFG_DST_LINE, transfer->destination_line);
+	bdma_reg_write(engine, CFG_SURF_REPEAT, transfer->surface_repeat - 1);
+	bdma_reg_write(engine, CFG_SRC_SURF, transfer->source_surface);
+	bdma_reg_write(engine, CFG_DST_SURF, transfer->destination_surface);
+	bdma_reg_write(engine, CFG_OP, FIELD_ENUM(BDMA_CFG_OP_0, EN, ENABLE));
+
+exit:
+	return ret;
+}
+
+int
+dla_bdma_is_ready(struct dla_processor *processor,
+				  struct dla_processor_group *group)
+{
+	struct dla_processor_group *next_group;
+
+	next_group = &processor->groups[!group->id];
+
+	/**
+	 * If another group is already programmed but not active then
+	 * do not program this operation as BDMA does not really
+	 * have shadow copies for groups. It will end programming
+	 * same group. Wait for another group to get enabled.
+	 */
+	if ((processor->group_status & (1 << next_group->id)) &&
+						!next_group->active)
+		return 0;
+
+	return 1;
+}
+
+void
+dla_bdma_dump_config(struct dla_processor_group *group)
+{
+	struct dla_bdma_op_desc *bdma_op;
+	struct dla_bdma_surface_desc *bdma_surface;
+
+	bdma_surface = &group->surface_desc->bdma_surface;
+	bdma_op = &group->operation_desc->bdma_op;
+}
+
+int
+dla_bdma_program(struct dla_engine *engine, struct dla_processor_group *group)
+{
+	int32_t i;
+	int32_t ret = 0;
+	struct dla_bdma_surface_desc *bdma_surface;
+
+	if (!engine->config_data->bdma_enable) {
+		pr_err("BDMA is not supported for this configuration\n");
+		ret = -EINVAL;
+		goto exit;
+	}
+
+	bdma_surface = &group->surface_desc->bdma_surface;
+
+	pr_debug("Num of transfers %u\n", bdma_surface->num_transfers);
+	if (bdma_surface->num_transfers == (uint16_t)0)
+		goto exit;
+
+	if (bdma_surface->num_transfers > NUM_MAX_BDMA_OPS) {
+		pr_err("Invalid number of transfers\n");
+		ret = -EINVAL;
+		goto exit;
+	}
+
+	for (i = 0; i < bdma_surface->num_transfers; i++) {
+		ret = processor_bdma_program_slot(engine, bdma_surface,
+					&bdma_surface->transfers[i]);
+		if (ret)
+			goto exit;
+	}
+
+	dla_enable_intr(engine, MASK(GLB_S_INTR_MASK_0, BDMA_DONE_MASK1) |
+			MASK(GLB_S_INTR_MASK_0, BDMA_DONE_MASK0));
+
+exit:
+	return ret;
+}
drivers/gpu/drm/nvdla/nvdla_cache.c new file +215 -0
new file mode 100644
index 000000000000..f8bd7b514aab
@@ -0,0 +1,215 @@
+// SPDX-License-Identifier: BSD-3-Clause OR GPL-2.0
+/*
+ * Copyright (C) 2017-2018 NVIDIA CORPORATION
+ * Copyright (C) 2022 Cai Huoqing
+ */
+
+#include "nvdla_common.h"
+#include "nvdla_drm.h"
+#include "nvdla_reg.h"
+#include "nvdla_engine.h"
+
+#define DLA_OP_CACHE_SIZE (DLA_NUM_GROUPS * ((DLA_OP_NUM + 2) * 2))
+
+static struct dla_common_op_desc desc_cache[DLA_OP_NUM][DLA_OP_CACHE_SIZE];
+static int32_t desc_refcount[DLA_OP_NUM][DLA_OP_CACHE_SIZE];
+
+void
+dla_get_refcount(struct dla_common_op_desc *op_desc)
+{
+	int32_t i;
+	struct dla_common_op_desc *desc = NULL;
+
+	if (op_desc == NULL)
+		return;
+
+	if (op_desc->index == -1)
+		return;
+
+	desc = &desc_cache[op_desc->op_type][0];
+
+	for (i = 0; i < DLA_OP_CACHE_SIZE; i++, desc++) {
+		if (desc->index == op_desc->index &&
+				desc->roi_index == op_desc->roi_index) {
+			desc_refcount[op_desc->op_type][i]++;
+			return;
+		}
+	}
+}
+
+struct dla_common_op_desc *
+dla_get_op_desc(struct dla_engine *engine,
+				struct dla_task *task, int16_t index,
+				uint8_t op_type, uint8_t roi_index)
+{
+	int32_t i;
+	int32_t ret;
+	uint64_t op_base;
+	uint64_t dep_graph_addr;
+	struct dla_common_op_desc *desc = NULL;
+
+	if (index == -1) {
+		pr_debug("no desc get due to index==-1\n");
+		goto exit;
+	}
+
+	dep_graph_addr = (sizeof(struct dla_common_op_desc) *
+				engine->network->num_operations * roi_index);
+
+	desc = &desc_cache[op_type][0];
+
+	for (i = 0; i < DLA_OP_CACHE_SIZE; i++, desc++) {
+		if (desc->index == index && desc->roi_index == roi_index) {
+			if (desc->op_type != op_type) {
+				pr_err("op_cache[op=%u] contains incorrect entry of op[%u]\n",
+					   op_type, desc->op_type);
+				continue;
+			}
+			desc_refcount[op_type][i]++;
+			goto exit;
+		}
+	}
+
+	desc = &desc_cache[op_type][0];
+
+	for (i = 0; i < DLA_OP_CACHE_SIZE; i++, desc++) {
+		if (desc->index == -1) {
+			op_base = dep_graph_addr +
+					(sizeof(struct dla_common_op_desc) *
+					(uint64_t)index);
+			ret = dla_data_read(engine->driver_context,
+					task->task_data,
+					task->dependency_graph_addr,
+					(void *)(desc),
+					sizeof(struct dla_common_op_desc),
+					op_base);
+			if (ret) {
+				desc = NULL;
+				goto exit;
+			}
+
+			if (op_type != desc->op_type) {
+				/*
+				 * op_type of entry read from DRAM should not
+				 * mismatch with given op_type. If they
+				 * mismatches, then wrong entry is fetched, so
+				 * report this issue by throwing error.
+				 */
+				pr_err("Fetched [op_type=%u] from DRAM doesn't match with op_type[%u]\n",
+					   desc->op_type, op_type);
+				desc->op_type = op_type;
+				desc->index = -1;
+				desc->roi_index = -1;
+				desc = NULL;
+				goto exit;
+			}
+
+			desc->index = index;
+			desc->roi_index = roi_index;
+
+			desc_refcount[op_type][i]++;
+			goto exit;
+		}
+	}
+
+exit:
+	return desc;
+}
+
+static void
+dla_free_op_desc(struct dla_engine *engine, struct dla_common_op_desc *op_desc)
+{
+	uint64_t op_base;
+	uint64_t dep_graph_addr;
+	struct dla_task *task;
+
+	pr_debug("Enter: %s op desc index %u ROI %d\n", __func__,
+				op_desc->index, op_desc->roi_index);
+
+	task = engine->task;
+	dep_graph_addr = (sizeof(struct dla_common_op_desc) *
+				engine->network->num_operations *
+				op_desc->roi_index);
+
+	if (op_desc->index == -1)
+		goto exit;
+
+	if (op_desc == NULL)
+		goto exit;
+
+	/**
+	 * TODO: keeping the depth value hardcoded as 0 for now,
+	 * need to replace it once corresponding implementation is done.
+	 */
+	op_base = (dep_graph_addr +
+			(sizeof(struct dla_common_op_desc) *
+			(uint64_t)op_desc->index));
+
+	/**
+	 * Flush descriptor to DRAM
+	 */
+	dla_data_write(engine->driver_context,
+			task->task_data,
+			(void *)op_desc,
+			task->dependency_graph_addr,
+			sizeof(struct dla_common_op_desc),
+			op_base);
+
+	/**
+	 * Release it
+	 */
+	op_desc->index = -1;
+	op_desc->roi_index = -1;
+exit:
+	return;
+}
+
+void
+dla_put_op_desc(struct dla_engine *engine, struct dla_common_op_desc *op_desc)
+{
+	int32_t i;
+	struct dla_common_op_desc *desc;
+
+	if (op_desc == NULL)
+		return;
+
+	if (op_desc->index == -1)
+		return;
+
+	desc = &desc_cache[op_desc->op_type][0];
+
+	for (i = 0; i < DLA_OP_CACHE_SIZE; i++, desc++) {
+		if (desc->index == op_desc->index &&
+				desc->roi_index == op_desc->roi_index) {
+
+			desc_refcount[op_desc->op_type][i]--;
+
+			/**
+			 * Free desc if refcount is 0
+			 */
+			if (desc_refcount[op_desc->op_type][i] == 0)
+				dla_free_op_desc(engine, op_desc);
+
+			return;
+		}
+	}
+}
+
+void
+dla_init_op_cache(struct dla_engine *engine)
+{
+	int32_t i, j;
+	struct dla_common_op_desc *desc = &desc_cache[0][0];
+
+	memset((uint8_t *)&desc_cache[0][0], 0, sizeof(desc_cache));
+	memset((uint8_t *)&desc_refcount[0][0], 0, sizeof(desc_refcount));
+
+	for (i = 0; i < DLA_OP_NUM; i++) {
+		for (j = 0; j < DLA_OP_CACHE_SIZE; j++) {
+			desc->index = -1;
+			desc->roi_index = -1;
+			desc->op_type = (uint8_t)i;
+			desc++;
+		}
+	}
+}
drivers/gpu/drm/nvdla/nvdla_cdp.c new file +300 -0
new file mode 100644
index 000000000000..793161e50b1c
@@ -0,0 +1,300 @@
+// SPDX-License-Identifier: BSD-3-Clause OR GPL-2.0
+/*
+ * Copyright (C) 2017-2018 NVIDIA CORPORATION
+ * Copyright (C) 2022 Cai Huoqing
+ */
+
+#include "nvdla_common.h"
+#include "nvdla_drm.h"
+#include "nvdla_reg.h"
+#include "nvdla_engine.h"
+
+static const uint8_t map_ram[] = {
+	FIELD_ENUM(CDP_RDMA_D_SRC_DMA_CFG_0, SRC_RAM_TYPE, MC),
+	FIELD_ENUM(CDP_RDMA_D_SRC_DMA_CFG_0, SRC_RAM_TYPE, CV),
+};
+
+static const uint8_t map_precision[] = {
+	FIELD_ENUM(CDP_RDMA_D_DATA_FORMAT_0, INPUT_DATA, INT8),
+	FIELD_ENUM(CDP_RDMA_D_DATA_FORMAT_0, INPUT_DATA, INT16),
+	FIELD_ENUM(CDP_RDMA_D_DATA_FORMAT_0, INPUT_DATA, FP16),
+};
+
+static const uint8_t map_perf_dma[] = {
+	FIELD_ENUM(CDP_D_PERF_ENABLE_0, DMA_EN, DISABLE),
+	FIELD_ENUM(CDP_D_PERF_ENABLE_0, DMA_EN, ENABLE),
+};
+
+static const uint8_t map_perf_lut[] = {
+	FIELD_ENUM(CDP_D_PERF_ENABLE_0, LUT_EN, DISABLE),
+	FIELD_ENUM(CDP_D_PERF_ENABLE_0, LUT_EN, ENABLE),
+};
+
+
+static uint32_t
+map_local_size(uint8_t local_size)
+{
+	return ((local_size-1)/2)-1;
+}
+
+void
+dla_cdp_set_producer(struct dla_engine *engine, int32_t group_id, int32_t rdma_group_id)
+{
+	uint32_t reg;
+
+	/**
+	 * set producer pointer for all sub-modules
+	 */
+	reg = group_id << SHIFT(CDP_S_POINTER_0, PRODUCER);
+	cdp_reg_write(engine, S_POINTER, reg);
+	reg = group_id << SHIFT(CDP_RDMA_S_POINTER_0, PRODUCER);
+	cdp_rdma_reg_write(engine, S_POINTER, reg);
+}
+
+int
+dla_cdp_enable(struct dla_engine *engine, struct dla_processor_group *group)
+{
+	uint32_t reg;
+
+	/**
+	 * enable all sub-modules
+	 */
+	reg = FIELD_ENUM(CDP_RDMA_D_OP_ENABLE_0, OP_EN, ENABLE);
+	cdp_rdma_reg_write(engine, D_OP_ENABLE, reg);
+	reg = FIELD_ENUM(CDP_D_OP_ENABLE_0, OP_EN, ENABLE);
+	cdp_reg_write(engine, D_OP_ENABLE, reg);
+
+	return 0;
+}
+
+void
+dla_cdp_rdma_check(struct dla_processor_group *group)
+{
+	group->is_rdma_needed = 1;
+}
+
+static int32_t
+processor_cdp_program(struct dla_engine *engine, struct dla_processor_group *group)
+{
+	int32_t ret = 0;
+	uint32_t reg, high, low;
+	uint64_t input_address = 0;
+	uint64_t output_address = 0;
+	struct dla_lut_param lut;
+	struct dla_cdp_op_desc *cdp_op;
+	struct dla_cdp_surface_desc *cdp_surface;
+
+	cdp_op = &group->operation_desc->cdp_op;
+	cdp_surface = &group->surface_desc->cdp_surface;
+
+	/* Argument check */
+	if (cdp_surface->src_data.type == DLA_MEM_HW) {
+		pr_err("Invalid source memory type\n");
+		ret = -EINVAL;
+		goto exit;
+	}
+	if (cdp_surface->dst_data.type == DLA_MEM_HW) {
+		pr_err("Invalid destination memory type\n");
+		ret = -EINVAL;
+		goto exit;
+	}
+
+	if (cdp_op->in_precision != cdp_op->out_precision) {
+		pr_err("CDP does not support precision conversion\n");
+		ret = -EINVAL;
+		goto exit;
+	}
+
+	/* get the addresses from task descriptor */
+	ret = dla_read_input_address(engine, &cdp_surface->src_data,
+								 &input_address,
+								 group->op_desc->index,
+								 group->roi_index, 1);
+	if (ret)
+		goto exit;
+
+	dla_get_dma_cube_address(engine->driver_context,
+							 engine->task->task_data,
+							 cdp_surface->dst_data.address,
+							 cdp_surface->dst_data.offset,
+							 (void *)&output_address,
+							 DESTINATION_DMA);
+	if (cdp_op->lut_index >= 0) {
+		group->lut_index = cdp_op->lut_index;
+		dla_read_lut(engine, cdp_op->lut_index, (void *)&lut);
+	}
+
+	/* config CDP RDMA registers */
+	reg = ((cdp_surface->src_data.width - 1)
+		<< SHIFT(CDP_RDMA_D_DATA_CUBE_WIDTH_0, WIDTH));
+	cdp_rdma_reg_write(engine, D_DATA_CUBE_WIDTH, reg);
+
+	reg = ((cdp_surface->src_data.height - 1)
+		<< SHIFT(CDP_RDMA_D_DATA_CUBE_HEIGHT_0, HEIGHT));
+	cdp_rdma_reg_write(engine, D_DATA_CUBE_HEIGHT, reg);
+
+	reg = ((cdp_surface->src_data.channel - 1)
+		<< SHIFT(CDP_RDMA_D_DATA_CUBE_CHANNEL_0, CHANNEL));
+	cdp_rdma_reg_write(engine, D_DATA_CUBE_CHANNEL, reg);
+
+	high = upper_32_bits(input_address);
+	low = lower_32_bits(input_address);
+	cdp_rdma_reg_write(engine, D_SRC_BASE_ADDR_LOW, low);
+	cdp_rdma_reg_write(engine, D_SRC_BASE_ADDR_HIGH, high);
+
+	cdp_rdma_reg_write(engine, D_SRC_LINE_STRIDE,
+			cdp_surface->src_data.line_stride);
+	cdp_rdma_reg_write(engine, D_SRC_SURFACE_STRIDE,
+			cdp_surface->src_data.surf_stride);
+
+	reg = (map_ram[cdp_surface->src_data.type]
+		<< SHIFT(CDP_RDMA_D_SRC_DMA_CFG_0, SRC_RAM_TYPE));
+	cdp_rdma_reg_write(engine, D_SRC_DMA_CFG, reg);
+
+	reg = (map_precision[cdp_op->in_precision]
+		<< SHIFT(CDP_RDMA_D_DATA_FORMAT_0, INPUT_DATA));
+	cdp_rdma_reg_write(engine, D_DATA_FORMAT, reg);
+
+	/* config CDP */
+	if (cdp_op->lut_index >= 0)
+		update_lut(engine, CDP_S_LUT_ACCESS_CFG_0, &lut, cdp_op->in_precision);
+
+	high = upper_32_bits(output_address);
+	low = lower_32_bits(output_address);
+	cdp_reg_write(engine, D_DST_BASE_ADDR_LOW, low);
+	cdp_reg_write(engine, D_DST_BASE_ADDR_HIGH, high);
+
+	cdp_reg_write(engine, D_DST_LINE_STRIDE, cdp_surface->dst_data.line_stride);
+	cdp_reg_write(engine, D_DST_SURFACE_STRIDE, cdp_surface->dst_data.surf_stride);
+
+	reg = (map_ram[cdp_surface->dst_data.type]
+		<< SHIFT(CDP_D_DST_DMA_CFG_0, DST_RAM_TYPE));
+	cdp_reg_write(engine, D_DST_DMA_CFG, reg);
+
+	reg = (map_precision[cdp_op->in_precision]
+		<< SHIFT(CDP_D_DATA_FORMAT_0, INPUT_DATA_TYPE));
+	cdp_reg_write(engine, D_DATA_FORMAT, reg);
+
+	reg = (map_local_size(cdp_op->local_size)
+		<< SHIFT(CDP_D_LRN_CFG_0, NORMALZ_LEN));
+	cdp_reg_write(engine, D_LRN_CFG, reg);
+
+	reg = (cdp_op->in_cvt.offset
+		<< SHIFT(CDP_D_DATIN_OFFSET_0, DATIN_OFFSET));
+	cdp_reg_write(engine, D_DATIN_OFFSET, reg);
+
+	reg = (cdp_op->in_cvt.scale
+		<< SHIFT(CDP_D_DATIN_SCALE_0, DATIN_SCALE));
+	cdp_reg_write(engine, D_DATIN_SCALE, reg);
+
+	reg = (cdp_op->in_cvt.truncate
+		<< SHIFT(CDP_D_DATIN_SHIFTER_0, DATIN_SHIFTER));
+	cdp_reg_write(engine, D_DATIN_SHIFTER, reg);
+
+	reg = (cdp_op->out_cvt.offset
+		<< SHIFT(CDP_D_DATOUT_OFFSET_0, DATOUT_OFFSET));
+	cdp_reg_write(engine, D_DATOUT_OFFSET, reg);
+
+	reg = (cdp_op->out_cvt.scale
+		<< SHIFT(CDP_D_DATOUT_SCALE_0, DATOUT_SCALE));
+	cdp_reg_write(engine, D_DATOUT_SCALE, reg);
+
+	reg = (cdp_op->out_cvt.truncate
+		<< SHIFT(CDP_D_DATOUT_SHIFTER_0, DATOUT_SHIFTER));
+	cdp_reg_write(engine, D_DATOUT_SHIFTER, reg);
+
+	reg = ((cdp_op->bypass_sqsum ?
+		FIELD_ENUM(CDP_D_FUNC_BYPASS_0, SQSUM_BYPASS, ENABLE) :
+		FIELD_ENUM(CDP_D_FUNC_BYPASS_0, SQSUM_BYPASS, DISABLE)) <<
+		SHIFT(CDP_D_FUNC_BYPASS_0, SQSUM_BYPASS)) |
+		((cdp_op->bypass_out_mul ?
+		FIELD_ENUM(CDP_D_FUNC_BYPASS_0, MUL_BYPASS, ENABLE) :
+		FIELD_ENUM(CDP_D_FUNC_BYPASS_0, MUL_BYPASS, DISABLE)) <<
+		SHIFT(CDP_D_FUNC_BYPASS_0, MUL_BYPASS));
+	cdp_reg_write(engine, D_FUNC_BYPASS, reg);
+
+exit:
+	return ret;
+}
+
+int
+dla_cdp_is_ready(struct dla_processor *processor,
+				 struct dla_processor_group *group)
+{
+	struct dla_processor_group *next_group;
+	struct dla_cdp_op_desc *cdp_op;
+
+	cdp_op = &group->operation_desc->cdp_op;
+	next_group = &processor->groups[!group->id];
+
+	/**
+	 * Single LUT is shared between two CDP groups, need to make
+	 * sure that usage does not conflict. Also, LUT write
+	 * access is locked when CDP sub-engine is active, so delay
+	 * writing LUT when another group is active.
+	 */
+
+	/**
+	 * if no LUT required for current group then it can be programmed
+	 * without further checks
+	 */
+	if (cdp_op->lut_index == -1)
+		return 1;
+
+	/**
+	 * if same LUT is used for both groups then it can be programmed
+	 * without more checks. Even if another group is active and LUT
+	 * is locked, it would have been programmed by another group.
+	 */
+	if (next_group->lut_index == cdp_op->lut_index)
+		return 1;
+
+	/**
+	 * if LUT index of another group is not -1 means some LUT is programmed,
+	 * then do not program current LUT as we already know current LUT is not
+	 * -1 and neither same as another group.
+	 */
+	if (next_group->lut_index != -1)
+		return 0;
+
+	/**
+	 * if current group needs LUT different than another group and that
+	 * group is not active then program it.
+	 */
+	if (!next_group->active)
+		return 1;
+
+	/**
+	 * if control is here it means current group is using LUT different than
+	 * another group and that group is active. Wait for another group to
+	 * become idle.
+	 */
+
+	return 0;
+}
+
+void
+dla_cdp_dump_config(struct dla_processor_group *group)
+{
+	struct dla_cdp_op_desc *cdp_op;
+	struct dla_cdp_surface_desc *cdp_surface;
+
+	cdp_surface = &group->surface_desc->cdp_surface;
+	cdp_op = &group->operation_desc->cdp_op;
+}
+
+int
+dla_cdp_program(struct dla_engine *engine, struct dla_processor_group *group)
+{
+	int32_t ret;
+
+	dla_enable_intr(engine, MASK(GLB_S_INTR_MASK_0, CDP_DONE_MASK1) |
+					MASK(GLB_S_INTR_MASK_0, CDP_DONE_MASK0));
+
+	ret = processor_cdp_program(engine, group);
+	if (ret)
+		goto exit;
+
+exit:
+	return ret;
+}
drivers/gpu/drm/nvdla/nvdla_common.c new file +295 -0
new file mode 100644
index 000000000000..3abc863d9fce
@@ -0,0 +1,295 @@
+// SPDX-License-Identifier: BSD-3-Clause OR GPL-2.0
+/*
+ * Copyright (C) 2017-2018 NVIDIA CORPORATION
+ * Copyright (C) 2022 Cai Huoqing
+ */
+
+#include "nvdla_reg.h"
+#include "nvdla_common.h"
+#include "nvdla_drm.h"
+#include "nvdla_engine.h"
+
+static const uint8_t map_lut_method[] = {
+	FIELD_ENUM(CDP_S_LUT_CFG_0, LUT_LE_FUNCTION, EXPONENT),
+	FIELD_ENUM(CDP_S_LUT_CFG_0, LUT_LE_FUNCTION, LINEAR)
+};
+static const uint8_t map_lut_out[] = {
+	FIELD_ENUM(CDP_S_LUT_CFG_0, LUT_UFLOW_PRIORITY, LE),
+	FIELD_ENUM(CDP_S_LUT_CFG_0, LUT_UFLOW_PRIORITY, LO)
+};
+
+static const uint16_t access_data_offset[] = {
+	CDP_S_LUT_ACCESS_DATA_0 - CDP_S_LUT_ACCESS_CFG_0,
+	SDP_S_LUT_ACCESS_DATA_0 - SDP_S_LUT_ACCESS_CFG_0,
+};
+static const uint16_t lut_cfg_offset[] = {
+	CDP_S_LUT_CFG_0 - CDP_S_LUT_ACCESS_CFG_0,
+	SDP_S_LUT_CFG_0 - SDP_S_LUT_ACCESS_CFG_0,
+};
+static const uint16_t lut_info_offset[] = {
+	CDP_S_LUT_INFO_0 - CDP_S_LUT_ACCESS_CFG_0,
+	SDP_S_LUT_INFO_0 - SDP_S_LUT_ACCESS_CFG_0,
+};
+static const uint16_t le_start_offset[] = {
+	CDP_S_LUT_LE_START_LOW_0 - CDP_S_LUT_ACCESS_CFG_0,
+	SDP_S_LUT_LE_START_0 - SDP_S_LUT_ACCESS_CFG_0,
+};
+static const uint16_t le_end_offset[] = {
+	CDP_S_LUT_LE_END_LOW_0 - CDP_S_LUT_ACCESS_CFG_0,
+	SDP_S_LUT_LE_END_0 - SDP_S_LUT_ACCESS_CFG_0,
+};
+static const uint16_t lo_start_offset[] = {
+	CDP_S_LUT_LO_START_LOW_0 - CDP_S_LUT_ACCESS_CFG_0,
+	SDP_S_LUT_LO_START_0 - SDP_S_LUT_ACCESS_CFG_0,
+};
+static const uint16_t lo_end_offset[] = {
+	CDP_S_LUT_LO_END_LOW_0 - CDP_S_LUT_ACCESS_CFG_0,
+	SDP_S_LUT_LO_END_0 - SDP_S_LUT_ACCESS_CFG_0,
+};
+static const uint16_t le_slope_scale_offset[] = {
+	CDP_S_LUT_LE_SLOPE_SCALE_0 - CDP_S_LUT_ACCESS_CFG_0,
+	SDP_S_LUT_LE_SLOPE_SCALE_0 - SDP_S_LUT_ACCESS_CFG_0,
+};
+static const uint16_t le_slope_shift_offset[] = {
+	CDP_S_LUT_LE_SLOPE_SHIFT_0 - CDP_S_LUT_ACCESS_CFG_0,
+	SDP_S_LUT_LE_SLOPE_SHIFT_0 - SDP_S_LUT_ACCESS_CFG_0,
+};
+static const uint16_t lo_slope_scale_offset[] = {
+	CDP_S_LUT_LO_SLOPE_SCALE_0 - CDP_S_LUT_ACCESS_CFG_0,
+	SDP_S_LUT_LO_SLOPE_SCALE_0 - SDP_S_LUT_ACCESS_CFG_0,
+};
+static const uint16_t lo_slope_shift_offset[] = {
+	CDP_S_LUT_LO_SLOPE_SHIFT_0 - CDP_S_LUT_ACCESS_CFG_0,
+	SDP_S_LUT_LO_SLOPE_SHIFT_0 - SDP_S_LUT_ACCESS_CFG_0,
+};
+
+void update_lut(struct dla_engine *engine, uint32_t reg_base,
+				struct dla_lut_param *lut, uint8_t precision)
+{
+	int32_t i;
+	uint32_t reg;
+	uint32_t high, low;
+	int32_t is_sdp = reg_base == SDP_S_LUT_ACCESS_CFG_0;
+
+	/* program raw table */
+	reg = (FIELD_ENUM(CDP_S_LUT_ACCESS_CFG_0, LUT_TABLE_ID, LE)
+		<< SHIFT(CDP_S_LUT_ACCESS_CFG_0, LUT_TABLE_ID)) |
+		(FIELD_ENUM(CDP_S_LUT_ACCESS_CFG_0, LUT_ACCESS_TYPE, WRITE)
+		<< SHIFT(CDP_S_LUT_ACCESS_CFG_0, LUT_ACCESS_TYPE));
+	reg_write(engine, reg_base, reg);
+
+	for (i = 0; i < (1<<LUT_LINEAR_EXP_TABLE_ENTRY_LOG2)+1; i++) {
+		dla_reg_write(engine->driver_context,
+				reg_base + access_data_offset[is_sdp],
+				lut->linear_exp_table[i]);
+	}
+
+	/* program density table */
+	reg = (FIELD_ENUM(CDP_S_LUT_ACCESS_CFG_0, LUT_TABLE_ID, LO)
+		<< SHIFT(CDP_S_LUT_ACCESS_CFG_0, LUT_TABLE_ID)) |
+		(FIELD_ENUM(CDP_S_LUT_ACCESS_CFG_0, LUT_ACCESS_TYPE, WRITE)
+		<< SHIFT(CDP_S_LUT_ACCESS_CFG_0, LUT_ACCESS_TYPE));
+	dla_reg_write(engine->driver_context, reg_base, reg);
+
+	for (i = 0; i < (1<<LUT_LINEAR_ONLY_TABLE_ENTRY_LOG2)+1; i++) {
+		dla_reg_write(engine->driver_context,
+				reg_base + access_data_offset[is_sdp],
+				lut->linear_only_table[i]);
+	}
+
+	/* program other configurations */
+	reg = (map_lut_method[lut->method] <<
+		SHIFT(CDP_S_LUT_CFG_0, LUT_LE_FUNCTION)) |
+		(map_lut_out[lut->hybrid_priority] <<
+		SHIFT(CDP_S_LUT_CFG_0, LUT_HYBRID_PRIORITY)) |
+		(map_lut_out[lut->underflow_priority] <<
+		SHIFT(CDP_S_LUT_CFG_0, LUT_UFLOW_PRIORITY)) |
+		(map_lut_out[lut->overflow_priority] <<
+		SHIFT(CDP_S_LUT_CFG_0, LUT_OFLOW_PRIORITY));
+	dla_reg_write(engine->driver_context,
+			reg_base + lut_cfg_offset[is_sdp], reg);
+
+	if (lut->method == FIELD_ENUM(CDP_S_LUT_CFG_0,
+					LUT_LE_FUNCTION, EXPONENT)) {
+		reg = ((((uint32_t)lut->linear_exp_offset.exp_offset) <<
+			SHIFT(CDP_S_LUT_INFO_0, LUT_LE_INDEX_OFFSET))&
+		MASK(CDP_S_LUT_INFO_0, LUT_LE_INDEX_OFFSET)) |
+			((((uint32_t)lut->linear_only_offset.frac_bits) <<
+			SHIFT(CDP_S_LUT_INFO_0, LUT_LO_INDEX_SELECT))&
+		MASK(CDP_S_LUT_INFO_0, LUT_LO_INDEX_SELECT));
+	} else {
+		reg = ((((uint32_t)lut->linear_exp_offset.frac_bits) <<
+			SHIFT(CDP_S_LUT_INFO_0, LUT_LE_INDEX_SELECT))&
+		MASK(CDP_S_LUT_INFO_0, LUT_LE_INDEX_SELECT)) |
+			((((uint32_t)lut->linear_only_offset.frac_bits) <<
+			SHIFT(CDP_S_LUT_INFO_0, LUT_LO_INDEX_SELECT))&
+		MASK(CDP_S_LUT_INFO_0, LUT_LO_INDEX_SELECT));
+	}
+	dla_reg_write(engine->driver_context,
+			reg_base + lut_info_offset[is_sdp], reg);
+	high = upper_32_bits(lut->linear_exp_start);
+	low = lower_32_bits(lut->linear_exp_start);
+	dla_reg_write(engine->driver_context,
+			reg_base + le_start_offset[is_sdp], low);
+	if (!is_sdp)
+		dla_reg_write(engine->driver_context,
+				reg_base + le_start_offset[is_sdp] + 4, high);
+
+	high = upper_32_bits(lut->linear_exp_end);
+	low = lower_32_bits(lut->linear_exp_end);
+	dla_reg_write(engine->driver_context,
+				reg_base + le_end_offset[is_sdp], low);
+	if (!is_sdp)
+		dla_reg_write(engine->driver_context,
+				reg_base + le_end_offset[is_sdp] + 4, high);
+
+	high = upper_32_bits(lut->linear_only_start);
+	low = lower_32_bits(lut->linear_only_start);
+	dla_reg_write(engine->driver_context,
+				reg_base + lo_start_offset[is_sdp], low);
+	if (!is_sdp)
+		dla_reg_write(engine->driver_context,
+				reg_base + lo_start_offset[is_sdp] + 4, high);
+
+	high = upper_32_bits(lut->linear_only_end);
+	low = lower_32_bits(lut->linear_only_end);
+	dla_reg_write(engine->driver_context,
+				reg_base + lo_end_offset[is_sdp], low);
+	if (!is_sdp)
+		dla_reg_write(engine->driver_context,
+				reg_base + lo_end_offset[is_sdp] + 4, high);
+
+	if (precision == PRECISION_FP16) {
+		reg = (lut->linear_exp_underflow_slope.data_f <<
+			SHIFT(CDP_S_LUT_LE_SLOPE_SCALE_0,
+					LUT_LE_SLOPE_UFLOW_SCALE)) |
+			(lut->linear_exp_overflow_slope.data_f <<
+			SHIFT(CDP_S_LUT_LE_SLOPE_SCALE_0,
+					LUT_LE_SLOPE_OFLOW_SCALE));
+		dla_reg_write(engine->driver_context,
+				reg_base + le_slope_scale_offset[is_sdp], reg);
+
+		reg = (lut->linear_only_underflow_slope.data_f <<
+			SHIFT(CDP_S_LUT_LO_SLOPE_SCALE_0,
+					LUT_LO_SLOPE_UFLOW_SCALE)) |
+			(lut->linear_only_overflow_slope.data_f <<
+			SHIFT(CDP_S_LUT_LO_SLOPE_SCALE_0,
+					LUT_LO_SLOPE_OFLOW_SCALE));
+		dla_reg_write(engine->driver_context,
+				reg_base + lo_slope_scale_offset[is_sdp], reg);
+	} else {
+		union dla_slope *oslope;
+		union dla_slope *uslope;
+
+		uslope = &lut->linear_exp_underflow_slope;
+		oslope = &lut->linear_exp_overflow_slope;
+		reg = ((((uint32_t)uslope->data_i.scale)
+			<< SHIFT(CDP_S_LUT_LE_SLOPE_SCALE_0,
+					LUT_LE_SLOPE_UFLOW_SCALE))&
+			MASK(CDP_S_LUT_LE_SLOPE_SCALE_0,
+					LUT_LE_SLOPE_UFLOW_SCALE)) |
+			((((uint32_t)oslope->data_i.scale)
+			<< SHIFT(CDP_S_LUT_LE_SLOPE_SCALE_0,
+					LUT_LE_SLOPE_OFLOW_SCALE))&
+			MASK(CDP_S_LUT_LE_SLOPE_SCALE_0,
+					LUT_LE_SLOPE_OFLOW_SCALE));
+		dla_reg_write(engine->driver_context,
+				reg_base + le_slope_scale_offset[is_sdp], reg);
+
+		reg = ((((uint32_t)uslope->data_i.shifter) <<
+			SHIFT(CDP_S_LUT_LE_SLOPE_SHIFT_0,
+					LUT_LE_SLOPE_UFLOW_SHIFT))&
+			MASK(CDP_S_LUT_LE_SLOPE_SHIFT_0,
+					LUT_LE_SLOPE_UFLOW_SHIFT)) |
+			((((uint32_t)oslope->data_i.shifter) <<
+			SHIFT(CDP_S_LUT_LE_SLOPE_SHIFT_0,
+					LUT_LE_SLOPE_OFLOW_SHIFT))&
+			MASK(CDP_S_LUT_LE_SLOPE_SHIFT_0,
+					LUT_LE_SLOPE_OFLOW_SHIFT));
+		dla_reg_write(engine->driver_context,
+				reg_base + le_slope_shift_offset[is_sdp], reg);
+
+		uslope = &lut->linear_only_underflow_slope;
+		oslope = &lut->linear_only_overflow_slope;
+		reg = ((((uint32_t)uslope->data_i.scale) <<
+			SHIFT(CDP_S_LUT_LO_SLOPE_SCALE_0,
+					LUT_LO_SLOPE_UFLOW_SCALE))&
+			MASK(CDP_S_LUT_LO_SLOPE_SCALE_0,
+					LUT_LO_SLOPE_UFLOW_SCALE)) |
+			((((uint32_t)oslope->data_i.scale) <<
+			SHIFT(CDP_S_LUT_LO_SLOPE_SCALE_0,
+					LUT_LO_SLOPE_OFLOW_SCALE))&
+			MASK(CDP_S_LUT_LO_SLOPE_SCALE_0,
+					LUT_LO_SLOPE_OFLOW_SCALE));
+		dla_reg_write(engine->driver_context,
+				reg_base + lo_slope_scale_offset[is_sdp], reg);
+		reg = ((((uint32_t)uslope->data_i.shifter) <<
+			SHIFT(CDP_S_LUT_LO_SLOPE_SHIFT_0,
+					LUT_LO_SLOPE_UFLOW_SHIFT))&
+			MASK(CDP_S_LUT_LO_SLOPE_SHIFT_0,
+					LUT_LO_SLOPE_UFLOW_SHIFT)) |
+			((((uint32_t)oslope->data_i.shifter) <<
+			SHIFT(CDP_S_LUT_LO_SLOPE_SHIFT_0,
+					LUT_LO_SLOPE_OFLOW_SHIFT))&
+			MASK(CDP_S_LUT_LO_SLOPE_SHIFT_0,
+					LUT_LO_SLOPE_OFLOW_SHIFT));
+		dla_reg_write(engine->driver_context,
+				reg_base + lo_slope_shift_offset[is_sdp], reg);
+	}
+}
+
+int
+validate_data_cube(struct dla_data_cube src_data_cube,
+				   struct dla_data_cube dst_data_cube,
+				   uint8_t mem_type)
+{
+	int32_t ret = 0;
+
+	if ((src_data_cube.width > DCUBE_MAX_WIDTH) ||
+	    (src_data_cube.height > DCUBE_MAX_HEIGHT) ||
+	    (src_data_cube.channel > DCUBE_MAX_CHANNEL)) {
+		pr_err("Invalid SrcInput Cude[W: %u, H: %u, C: %u]",
+				src_data_cube.width, src_data_cube.height,
+				src_data_cube.channel);
+		ret = -EINVAL;
+		goto exit;
+	}
+
+	if ((dst_data_cube.width > DCUBE_MAX_WIDTH) ||
+	    (dst_data_cube.height > DCUBE_MAX_HEIGHT) ||
+	    (dst_data_cube.channel > DCUBE_MAX_CHANNEL)) {
+		pr_err("Invalid DstInput Cude[W: %u, H: %u, C: %u]",
+				dst_data_cube.width, dst_data_cube.height,
+				dst_data_cube.channel);
+		ret = -EINVAL;
+		goto exit;
+	}
+
+	if (src_data_cube.type > mem_type) {
+		pr_err("Invalid src_data.mem_type: %u\n", src_data_cube.type);
+		ret = -EINVAL;
+		goto exit;
+	}
+
+	if (dst_data_cube.type > mem_type) {
+		pr_err("Invalid dst_data.mem_type: %u\n", dst_data_cube.type);
+		ret = -EINVAL;
+		goto exit;
+	}
+
+exit:
+	return ret;
+}
+
+int
+validate_precision(uint8_t precision, uint8_t map_precision)
+{
+	int32_t ret = 0;
+
+	if (precision >= map_precision) {
+		pr_err("Invalid precision: %u\n", precision);
+		ret = -EINVAL;
+	}
+
+	return ret;
+}
drivers/gpu/drm/nvdla/nvdla_common.h new file +835 -0
new file mode 100644
index 000000000000..38cf43246890
@@ -0,0 +1,835 @@
+/* SPDX-License-Identifier: GPL-2.0 OR BSD-3-Clause */
+/*
+ * Copyright (C) 2017-2018 NVIDIA CORPORATION.
+ * Copyright (C) 2022 Cai Huoqing
+ */
+
+#ifndef __NVDLA_COMMON_H_
+#define __NVDLA_COMMON_H_
+
+#include <linux/types.h>
+
+#define DCUBE_MAX_WIDTH		8192
+#define DCUBE_MAX_HEIGHT	8192
+#define DCUBE_MAX_CHANNEL	8192
+
+/**
+ * @ingroup Processors
+ * @name DLA Processors
+ * Processor modules in DLA engine. Each processor has it's
+ * own operation a.k.a. HW layer. Network is formed using
+ * graph of these operations
+ * @{
+ */
+#define DLA_OP_BDMA		0
+#define DLA_OP_CONV		1
+#define DLA_OP_SDP		2
+#define DLA_OP_PDP		3
+#define DLA_OP_CDP		4
+#define DLA_OP_RUBIK	5
+/** @} */
+
+/**
+ * @ingroup Processors
+ * @name Maximum number of processors
+ * @brief DLA ash 6 processors
+ * @{
+ */
+#define DLA_OP_NUM		6
+/** @} */
+
+/**
+ * @ingroup Processors
+ * @name Number of groups
+ * @brief Each processor has 2 groups of registers
+ * @{
+ */
+#define DLA_NUM_GROUPS		2
+/** @} */
+
+/**
+ * Network descriptor
+ *
+ * Contains all information to execute a network
+ *
+ * @op_head: Index of first operation of each type in operations list
+ * @num_rois: Number of ROIs
+ * @num_operations: Number of operations in one list
+ * @num_luts: Number of LUTs
+ */
+struct dla_network_desc {
+	int16_t operation_desc_index;
+	int16_t surface_desc_index;
+
+	int16_t dependency_graph_index;
+	int16_t lut_data_index;
+
+	int16_t roi_array_index;
+	int16_t surface_index;
+
+	int16_t stat_list_index;
+	int16_t reserved1;
+
+	int16_t op_head[DLA_OP_NUM];
+
+	uint16_t num_rois;
+	uint16_t num_operations;
+
+	uint16_t num_luts;
+	uint16_t num_addresses;
+
+	int16_t input_layer;
+	uint8_t dynamic_roi;
+	uint8_t reserved0;
+} __packed __aligned(4);
+
+/**
+ * @name Memory types
+ * @brief DLA engnine can read/write to/from 3 memory types
+ * @{
+ */
+#define DLA_MEM_MC			0 /* External DRAM */
+#define DLA_MEM_CV			1 /* CV-SRAM */
+#define DLA_MEM_HW			2 /* DLA sub-module */
+/** @} */
+
+/**
+ * @ingroup Events
+ * @name Operation events
+ * @brief Different events triggered by an operations
+ * @{
+ */
+#define DLA_EVENT_OP_COMPLETED		1
+#define DLA_EVENT_OP_PROGRAMMED		2
+#define DLA_EVENT_OP_ENABLED		3
+#define DLA_EVENT_CDMA_WT_DONE		4
+#define DLA_EVENT_CDMA_DT_DONE		5
+/** @} */
+
+struct dla_consumer {
+	int16_t index; /* the index of dla_common_op_desc in dep_graph_addr */
+	uint8_t event;
+	uint8_t res;
+} __packed __aligned(4);
+
+struct dla_common_op_desc {
+	int16_t index; /* set by ucode */
+	int8_t roi_index;
+	uint8_t op_type;
+
+	uint8_t dependency_count;
+	uint8_t reserved0[3];
+
+	struct dla_consumer consumers[DLA_OP_NUM];
+	struct dla_consumer fused_parent;
+} __packed __aligned(4);
+
+struct dla_roi_array_desc {
+	uint32_t array_length;
+
+	uint32_t array_reserved;
+} __packed __aligned(4);
+
+struct dla_roi_desc {
+	uint32_t left;
+
+	uint32_t top;
+
+	uint32_t right;
+
+	uint32_t bottom;
+} __packed __aligned(4);
+
+/**
+ * @ingroup BDMA
+ * @name Maximum BDMA transfers
+ * @brief BDMA supports multiple transfers in operation. This indicates
+ *        maximum number of transfers possible in one operation.
+ * @{
+ */
+#define NUM_MAX_BDMA_OPS	20
+/** @} */
+
+struct dla_bdma_transfer_desc {
+	int16_t source_address;
+	int16_t destination_address;
+
+	uint32_t line_size;
+
+	uint32_t line_repeat;
+
+	uint32_t source_line;
+
+	uint32_t destination_line;
+
+	uint32_t surface_repeat;
+
+	uint32_t source_surface;
+
+	uint32_t destination_surface;
+} __packed __aligned(4);
+
+struct dla_bdma_surface_desc {
+	uint8_t source_type;
+	uint8_t destination_type;
+	uint16_t num_transfers;
+
+	struct dla_bdma_transfer_desc transfers[NUM_MAX_BDMA_OPS];
+} __packed __aligned(4);
+
+struct dla_bdma_op_desc {
+	uint16_t num_transfers;
+	uint16_t reserved0;
+} __packed __aligned(4);
+
+struct dla_bdma_stat_desc {
+	uint32_t read_stall;
+	uint32_t write_stall;
+} __packed __aligned(4);
+
+/**
+ * @ingroup Convolution
+ * @name Convolution mode
+ * @brief Convolution modes support by DLA
+ * @{
+ */
+#define CONV_MODE_DIRECT	0
+#define CONV_MODE_WINOGRAD	1
+/** @} */
+
+/**
+ * @ingroup Processors
+ * @name Precision BPE mapping
+ * @brief Precision formats and Bit Per Elements mapping
+ * @{
+ */
+#define BPE_PRECISION_INT8		1
+#define BPE_PRECISION_INT16		2
+#define BPE_PRECISION_FP16		2
+/** @} */
+
+
+/**
+ * @ingroup Processors
+ * @name Precision types
+ * @brief Precision formats supported by DLA engine
+ * @{
+ */
+#define PRECISION_INT8		0
+#define PRECISION_INT16		1
+#define PRECISION_FP16		2
+/** @} */
+
+/**
+ * @ingroup Processors
+ * @name Data formats
+ * @brief Data formats supported by DLA engine
+ * @{
+ */
+#define FORMAT_T_R8			0
+#define FORMAT_T_R10			1
+#define FORMAT_T_R12			2
+#define FORMAT_T_R16			3
+#define FORMAT_T_R16_I			4
+#define FORMAT_T_R16_F			5
+#define FORMAT_T_A16B16G16R16		6
+#define FORMAT_T_X16B16G16R16		7
+#define FORMAT_T_A16B16G16R16_F		8
+#define FORMAT_T_A16Y16U16V16		9
+#define FORMAT_T_V16U16Y16A16		10
+#define FORMAT_T_A16Y16U16V16_F		11
+#define FORMAT_T_A8B8G8R8		12
+#define FORMAT_T_A8R8G8B8		13
+#define FORMAT_T_B8G8R8A8		14
+#define FORMAT_T_R8G8B8A8		15
+#define FORMAT_T_X8B8G8R8		16
+#define FORMAT_T_X8R8G8B8		17
+#define FORMAT_T_B8G8R8X8		18
+#define FORMAT_T_R8G8B8X8		19
+#define FORMAT_T_A2B10G10R10		20
+#define FORMAT_T_A2R10G10B10		21
+#define FORMAT_T_B10G10R10A2		22
+#define FORMAT_T_R10G10B10A2		23
+#define FORMAT_T_A2Y10U10V10		24
+#define FORMAT_T_V10U10Y10A2		25
+#define FORMAT_T_A8Y8U8V8			26
+#define FORMAT_T_V8U8Y8A8			27
+#define FORMAT_T_Y8___U8V8_N444		28
+#define FORMAT_T_Y8___V8U8_N444		29
+#define FORMAT_T_Y10___U10V10_N444	30
+#define FORMAT_T_Y10___V10U10_N444	31
+#define FORMAT_T_Y12___U12V12_N444	32
+#define FORMAT_T_Y12___V12U12_N444	33
+#define FORMAT_T_Y16___U16V16_N444	34
+#define FORMAT_T_Y16___V16U16_N444	35
+#define FORMAT_FEATURE			36
+/** @} */
+
+/**
+ * @ingroup Convolution
+ * @name Pixel mapping
+ * @brief Pixel mapping formats supported for image input in Convolution
+ * @{
+ */
+#define MAP_PITCH_LINEAR		0
+/** @} */
+
+/**
+ * @ingroup Convolution
+ * @name Weight formats
+ * @brief Weight data formats supported in Convolution
+ * @{
+ */
+#define WEIGHT_FORMAT_UNCOMPRESSED	0
+#define WEIGHT_FORMAT_COMPRESSED	1
+/** @} */
+
+/**
+ * @ingroup Convolution
+ * @name Mean data format
+ * @brief Mean data formats supported in Convolution
+ * @{
+ */
+#define MEAN_FORMAT_DISABLE     0
+#define MEAN_FORMAT_ENABLE      1
+/** @} */
+
+struct dla_cvt_param {
+	int16_t  scale;
+	uint8_t  truncate;
+	uint8_t  enable;
+
+	int32_t  offset;
+} __packed __aligned(4);
+
+struct dla_data_cube {
+	uint16_t type; /* dla_mem_type */
+	int16_t address; /* offset to the actual IOVA in task.address_list */
+
+	uint32_t offset; /* offset within address */
+	uint32_t size;
+
+	/* cube dimensions */
+	uint16_t width;
+	uint16_t height;
+
+	uint16_t channel;
+	uint16_t reserved0;
+
+	/* stride information */
+	uint32_t line_stride;
+	uint32_t surf_stride;
+
+	/* For Rubik only */
+	uint32_t plane_stride;
+} __packed __aligned(4);
+
+#define PIXEL_OVERRIDE_UINT 0
+#define PIXEL_OVERRIDE_INT  1
+
+struct dla_conv_surface_desc {
+	/* Data cube */
+	struct dla_data_cube weight_data;
+	struct dla_data_cube wmb_data;
+	struct dla_data_cube wgs_data;
+	struct dla_data_cube src_data;
+	struct dla_data_cube dst_data;
+
+	/**
+	 * u_addr = input_data.source_addr + offset_u
+	 * this field should be set when YUV is not interleave format
+	 *
+	 */
+	int64_t offset_u;
+
+	/* line stride for 2nd plane, must be 32bytes aligned */
+	uint32_t in_line_uv_stride;
+} __packed __aligned(4);
+
+struct dla_conv_op_desc {
+	/* Performance parameters */
+
+	/* dla_conv_mode */
+	uint8_t conv_mode;
+	uint8_t data_reuse;
+	uint8_t weight_reuse;
+	uint8_t skip_data_rls;
+
+	uint8_t skip_weight_rls;
+	uint8_t reserved0;
+	uint16_t entry_per_slice;
+
+	/* dla_data_format */
+	uint8_t data_format;
+	/* dla_pixel_mapping */
+	uint8_t pixel_mapping;
+	/* number of free slices before fetch */
+	uint16_t fetch_grain;
+
+	uint8_t reserved_b[8];
+
+	/* batch_num */
+	uint8_t batch;
+	/* dla_weight_format */
+	uint8_t weight_format;
+	uint8_t data_bank;
+	uint8_t weight_bank;
+
+	/* the offset in bytes of each data cube in a batch */
+	uint32_t batch_stride;
+
+	uint8_t post_extension;
+	uint8_t pixel_override;
+	/* number of slices need to be released */
+	uint16_t release;
+
+	 /* The input cube dimension for CSC */
+	uint16_t input_width_csc;
+	uint16_t input_height_csc;
+
+	uint16_t input_channel_csc;
+	uint16_t kernel_width_csc;
+
+	uint16_t kernel_height_csc;
+	uint16_t kernel_channel_csc;
+
+	/* The input cube dimension for CMAC */
+	uint16_t input_width_cmac;
+	uint16_t input_height_cmac;
+
+	/* actual size in bytes */
+	uint32_t bytes_per_kernel;
+
+	/* Algorithm parameters */
+
+	int16_t mean_ry; /* mean value for red in RGB or Y in YUV */
+	int16_t mean_gu; /* mean value for green in RGB or U in YUV */
+
+	int16_t mean_bv; /* mean value for blue in RGB or V in YUV */
+	int16_t mean_ax;
+
+	uint8_t mean_format; /* dla_mean_format */
+	uint8_t conv_stride_x;
+	uint8_t conv_stride_y;
+	uint8_t pad_x_left;
+
+	uint8_t pad_x_right;
+	uint8_t pad_y_top;
+	uint8_t pad_y_bottom;
+	uint8_t dilation_x;
+
+	uint8_t dilation_y;
+	uint8_t reserved2[2];
+
+	/* Precision parameters */
+	uint8_t pra_truncate;
+
+	uint8_t in_precision;
+	/* The output precision from CONV, it's the MAC processing precison */
+	uint8_t out_precision;
+	int16_t pad_val;
+
+	/* input converter parameters */
+	struct dla_cvt_param in_cvt;
+	/* output converter parameters, support truncate only */
+	struct dla_cvt_param out_cvt;
+
+} __packed __aligned(4);
+
+struct dla_conv_stat_desc {
+	uint32_t data_read_stall;
+	uint32_t weight_read_stall;
+	uint32_t data_read_latency;
+	uint32_t weight_read_latency;
+	uint32_t saturation_count;
+	uint32_t nan_data_num;
+	uint32_t nan_weight_num;
+	uint32_t inf_data_num;
+	uint32_t inf_weight_num;
+} __packed __aligned(4);
+
+/**
+ * @ingroup SDP
+ * @name Activation functions
+ * @brief Activation functions supported in SDP
+ * @{
+ */
+#define ACTIVATION_NONE		0
+#define ACTIVATION_RELU		1
+#define ACTIVATION_LUT		2
+#define ACTIVATION_PRELU	3
+/** @} */
+
+/**
+ * @ingroup LUT
+ * @name LUT size
+ * @brief LUT sizes for linear and exponentila LUT
+ * @{
+ */
+#define LUT_LINEAR_EXP_TABLE_ENTRY_LOG2		6
+#define LUT_LINEAR_ONLY_TABLE_ENTRY_LOG2	8
+/** @} */
+
+/**
+ * @ingroup LUT
+ * @name LUT types
+ * @brief DLA supports two types of LUT, linear and exonential
+ * @{
+ */
+#define LUT_LINEAR_EXP_TABLE		0
+#define LUT_LINEAR_ONLY_TABLE		1
+/** @} */
+
+/**
+ * @ingroup LUT
+ * @name LUT methods
+ * @brief DLA supports two types of LUT, linear and exonential
+ * @{
+ */
+#define LUT_METHOD_EXPONENTIAL		0
+#define LUT_METHOD_LINEAR		1
+/** @} */
+
+/**
+ * @ingroup LUT
+ * @name LUT
+ * @brief DLA supports two types of LUT, linear and exonential
+ * @{
+ */
+#define LUT_PRI_LINEAR_EXP		0
+#define LUT_PRI_LINEAR_ONLY		1
+/** @} */
+
+union dla_lut_offset {
+	/**
+	 * Number should be substracted on log domain before look up
+	 * exponetial table it has the same definition as hardware
+	 * thus input scaling should also take into account when
+	 * set this field.
+	 */
+	int8_t exp_offset;
+	/**
+	 * Number of bits should be right shift before looking
+	 * up linear table
+	 */
+	int8_t frac_bits;
+	uint16_t reserved0;
+};
+
+/**
+ * This struct is used to represent floating point values by INT
+ * suppose we have a float point number fp_x, it will be represented
+ * as:
+ *
+ * fp_x = scale_int_x>>(shifter_x)
+ *
+ * This is very useful for INT pipeline;
+ */
+struct dla_float_data {
+	int16_t scale;
+	int8_t shifter;
+	uint8_t reserved0;
+} __packed __aligned(4);
+
+/**
+ * For INT pipeline, we use the struct above to represent a floating number;
+ * For FP16 pipeline, we should store the FP16 encoded value into a uint16_t
+ * container
+ */
+union dla_slope {
+	struct dla_float_data data_i;
+
+	uint16_t data_f;
+};
+
+struct dla_lut_param {
+	/**
+	 * value of expression ((1<<LUT_LINEAR_EXP_TABLE_ENTRY_LOG2)+1) is 65,
+	 * ((1<<LUT_LINEAR_ONLY_TABLE_ENTRY_LOG2)+1) is 257, and int16_t is of
+	 * 2Byte. And below two statement's combined memory size is 644 Byte.
+	 *
+	 * NOTE: below two declaration combined size should always be multiple
+	 * of 4.
+	 */
+	int16_t linear_exp_table[(1<<LUT_LINEAR_EXP_TABLE_ENTRY_LOG2)+1];
+	int16_t linear_only_table[(1<<LUT_LINEAR_ONLY_TABLE_ENTRY_LOG2)+1];
+
+	union dla_lut_offset linear_exp_offset;
+	union dla_lut_offset linear_only_offset;
+
+	/**
+	 * The start and end point of raw table,
+	 * valid when raw_method=LINEAR only
+	 */
+	uint64_t linear_exp_start;
+	uint64_t linear_exp_end;
+	uint64_t linear_only_start;
+	uint64_t linear_only_end;
+
+	union dla_slope linear_exp_underflow_slope;
+	union dla_slope linear_exp_overflow_slope;
+	union dla_slope linear_only_underflow_slope;
+	union dla_slope linear_only_overflow_slope;
+
+	/**
+	 * dla_lut_priority, when both lut are hit(or one overflow,
+	 * the other underflow), which one should be selected as output
+	 */
+	uint8_t hybrid_priority;
+	uint8_t underflow_priority;
+	uint8_t overflow_priority;
+	uint8_t method; /* dla_lut_method */
+} __packed __aligned(4);
+
+struct dla_sdp_surface_desc {
+	/* Data cube */
+	/* source input cube, available when SDP working on offline mode */
+	struct dla_data_cube src_data;
+
+	/* X1 input cube */
+	struct dla_data_cube x1_data;
+
+	/* X2 input cube */
+	struct dla_data_cube x2_data;
+
+	/* Y input cube */
+	struct dla_data_cube y_data;
+
+	/* Output cube */
+	struct dla_data_cube dst_data;
+} __packed __aligned(4);
+
+#define SDP_OP_NONE		0
+#define SDP_OP_MUL		1
+#define SDP_OP_ADD		2
+#define SDP_OP_BOTH		3
+
+#define SDP_ALU_OP_MAX		0
+#define SDP_ALU_OP_MIN		1
+#define SDP_ALU_OP_SUM		2
+#define SDP_ALU_OP_EQL		3
+
+#define SDP_OP_PER_LAYER	0
+#define SDP_OP_PER_KERNEL	1
+#define SDP_OP_PER_POINT	2
+
+struct dla_sdp_cvt {
+	struct dla_cvt_param alu_cvt;
+	struct dla_cvt_param mul_cvt;
+} __packed __aligned(4);
+
+struct dla_sdp_op {
+	uint8_t enable;
+	uint8_t alu_type; /* dla_sdp_alu_op_type */
+	uint8_t type; /* dla_sdp_op_type */
+	uint8_t mode; /* dla_sdp_op_mode */
+
+	uint8_t act; /* dla_act_type */
+	uint8_t shift_value; /* left shift */
+	uint8_t truncate;
+	uint8_t precision;
+
+	int32_t alu_operand;
+	int32_t mul_operand;
+
+	struct dla_sdp_cvt  cvt;
+} __packed __aligned(4);
+
+struct dla_sdp_op_desc {
+	/* Precision parameters */
+	/* dla_precision */
+	uint8_t src_precision;
+	uint8_t dst_precision;
+	int16_t lut_index;
+
+	struct dla_cvt_param out_cvt;
+
+	/* Performance parameters */
+	/* dla_conv_mode */
+	uint8_t conv_mode;
+	uint8_t batch_num;
+	uint16_t reserved0;
+
+	uint32_t batch_stride;	/* will be used when batch_num > 1 */
+
+	/* Algorithm parameters */
+	struct dla_sdp_op x1_op;
+	struct dla_sdp_op x2_op;
+	struct dla_sdp_op y_op;
+} __packed __aligned(4);
+
+#define POOL_MODE_AVG		0
+#define POOL_MODE_MAX		1
+#define POOL_MODE_MIN		2
+
+#define POOL_SIZE_1		0
+#define POOL_SIZE_2		1
+#define POOL_SIZE_3		2
+#define POOL_SIZE_4		3
+#define POOL_SIZE_5		4
+#define POOL_SIZE_6		5
+#define POOL_SIZE_7		6
+#define POOL_SIZE_8		7
+
+#define PDP_PAD_VAL_NUM	7
+
+struct dla_pdp_surface_desc {
+	/* Data cube */
+	struct dla_data_cube src_data;
+
+	struct dla_data_cube dst_data;
+} __packed __aligned(4);
+
+struct dla_pdp_op_desc {
+	/* Performance parameters */
+	uint16_t  partial_in_width_first;
+	uint16_t  partial_in_width_mid;
+
+	uint16_t  partial_in_width_last;
+	uint16_t  partial_width_first;
+
+	uint16_t  partial_width_mid;
+	uint16_t  partial_width_last;
+
+	uint8_t   split_num;
+
+	/* Algorithm parameters */
+	uint8_t  pool_mode; /* dla_pool_mode */
+	uint8_t  pool_width; /* dla_pool_width */
+	uint8_t  pool_height; /* dla_pool_height */
+
+	uint8_t  stride_x;
+	uint8_t  stride_y;
+
+	/**
+	 * The left/right padding size,
+	 * pad_right might be less than pad_left
+	 */
+	uint8_t  pad_left;
+	uint8_t  pad_right;
+
+	/* The top/bottom padding size */
+	uint8_t  pad_top;
+	uint8_t  pad_bottom;
+
+	/* Precision parameters */
+	uint8_t  precision; /* dla_precision */
+	uint8_t  reserved0;
+	/**
+	 * if input has non-zero "offset", this value should be set
+	 * There'll be 7 different paddding values, the relationship between
+	 * those versions are:
+	 * padding_value[0] = -offset*scaling;
+	 * padding_value[1] = 2*padding_value[0]
+	 * padding_value[2] = 3*padding_value[0]
+	 * ...
+	 * The purpose is to avoid ucode implement FP16
+	 * multiplier(for FP16 mode)
+	 */
+	int32_t  padding_value[PDP_PAD_VAL_NUM];
+} __packed __aligned(4);
+
+struct dla_pdp_stat_desc {
+	uint32_t inf_input_num;
+	uint32_t nan_input_num;
+	uint32_t nan_output_num;
+	uint32_t write_stall;
+	uint32_t runtime;
+} __packed __aligned(4);
+
+struct dla_cdp_surface_desc {
+	/* Data cube */
+	struct dla_data_cube src_data;
+
+	struct dla_data_cube dst_data;
+} __packed __aligned(4);
+
+struct dla_cdp_op_desc {
+	/* Precision parameters */
+
+	/* dla_precision */
+	uint8_t  in_precision;
+	uint8_t  out_precision;
+	int16_t  lut_index;
+
+	struct dla_cvt_param in_cvt;
+	struct dla_cvt_param out_cvt;
+
+	/* Performance parameters */
+
+	/* Algorithm parameters */
+	uint8_t  local_size;
+	uint8_t  bypass_sqsum;
+	uint8_t  bypass_out_mul;
+	uint8_t  reserved0;
+} __packed __aligned(4);
+
+struct dla_cdp_stat_desc {
+	uint32_t nan_input_num;
+	uint32_t inf_input_num;
+	uint32_t nan_output_num;
+	uint32_t write_stall;
+	uint32_t lut_uflow;
+	uint32_t lut_oflow;
+	uint32_t lut_hybrid;
+	uint32_t lut_le_hit;
+	uint32_t lut_lo_hit;
+	uint32_t saturation_count;
+	uint32_t runtime;
+} __packed __aligned(4);
+
+struct dla_rubik_surface_desc {
+	/* Data cube */
+	struct dla_data_cube src_data;
+
+	struct dla_data_cube dst_data;
+} __packed __aligned(4);
+
+/* rubik mode */
+#define RUBIK_MODE_CONTRACT	0
+#define RUBIK_MODE_SPLIT	1
+#define RUBIK_MODE_MERGE	2
+
+struct dla_rubik_op_desc {
+	/* Precision parameters */
+	uint8_t mode;
+	uint8_t precision;
+	uint8_t stride_x;
+	uint8_t stride_y;
+} __packed __aligned(4);
+
+struct dla_rubik_stat_desc {
+	uint32_t read_stall;
+	uint32_t write_stall;
+	uint32_t runtime;
+} __packed __aligned(4);
+
+union dla_surface_container {
+	struct dla_bdma_surface_desc bdma_surface;
+	struct dla_conv_surface_desc conv_surface;
+	struct dla_sdp_surface_desc sdp_surface;
+	struct dla_pdp_surface_desc pdp_surface;
+	struct dla_cdp_surface_desc cdp_surface;
+	struct dla_rubik_surface_desc rubik_surface;
+};
+
+union dla_operation_container {
+	struct dla_bdma_op_desc bdma_op;
+	struct dla_conv_op_desc conv_op;
+	struct dla_sdp_op_desc sdp_op;
+	struct dla_pdp_op_desc pdp_op;
+	struct dla_cdp_op_desc cdp_op;
+	struct dla_rubik_op_desc rubik_op;
+};
+
+struct dla_engine;
+void update_lut(struct dla_engine *engine, uint32_t reg_base,
+		struct dla_lut_param *lut,
+		uint8_t precision);
+int32_t validate_data_cube(struct dla_data_cube src_data_cube,
+			struct dla_data_cube dst_data_cube,
+			uint8_t mem_type);
+int32_t validate_precision(uint8_t precision,
+			uint8_t map_precision);
+
+#endif
drivers/gpu/drm/nvdla/nvdla_conv.c new file +683 -0
new file mode 100644
index 000000000000..bb24909e1afa
@@ -0,0 +1,683 @@
+// SPDX-License-Identifier: BSD-3-Clause OR GPL-2.0
+/*
+ * Copyright (C) 2017-2018 NVIDIA CORPORATION
+ * Copyright (C) 2022 Cai Huoqing
+ */
+
+#include "nvdla_common.h"
+#include "nvdla_drm.h"
+#include "nvdla_reg.h"
+#include "nvdla_common.h"
+#include "nvdla_engine.h"
+
+static const uint8_t map_precision[] = {
+	FIELD_ENUM(CDMA_D_MISC_CFG_0, IN_PRECISION, INT8),
+	FIELD_ENUM(CDMA_D_MISC_CFG_0, IN_PRECISION, INT16),
+	FIELD_ENUM(CDMA_D_MISC_CFG_0, IN_PRECISION, FP16),
+};
+
+static const uint8_t map_conv[] = {
+	FIELD_ENUM(CACC_D_MISC_CFG_0, CONV_MODE, DIRECT),
+	FIELD_ENUM(CACC_D_MISC_CFG_0, CONV_MODE, WINOGRAD),
+};
+
+static const uint8_t map_weight_fmt[] = {
+	FIELD_ENUM(CSC_D_WEIGHT_FORMAT_0, WEIGHT_FORMAT, UNCOMPRESSED),
+	FIELD_ENUM(CSC_D_WEIGHT_FORMAT_0, WEIGHT_FORMAT, COMPRESSED),
+};
+
+static const uint8_t map_img_fmt[][2] = {
+	{FIELD_ENUM(CDMA_D_DATAIN_FORMAT_0,
+			PIXEL_FORMAT, T_R8), 1},
+	{FIELD_ENUM(CDMA_D_DATAIN_FORMAT_0,
+			PIXEL_FORMAT, T_R10), 2},
+	{FIELD_ENUM(CDMA_D_DATAIN_FORMAT_0,
+			PIXEL_FORMAT, T_R12), 2},
+	{FIELD_ENUM(CDMA_D_DATAIN_FORMAT_0,
+			PIXEL_FORMAT, T_R16), 2},
+	{FIELD_ENUM(CDMA_D_DATAIN_FORMAT_0,
+			PIXEL_FORMAT, T_R16_I), 2},
+	{FIELD_ENUM(CDMA_D_DATAIN_FORMAT_0,
+			PIXEL_FORMAT, T_R16_F), 2},
+	{FIELD_ENUM(CDMA_D_DATAIN_FORMAT_0,
+			PIXEL_FORMAT, T_A16B16G16R16), 8},
+	{FIELD_ENUM(CDMA_D_DATAIN_FORMAT_0,
+			PIXEL_FORMAT, T_X16B16G16R16), 8},
+	{FIELD_ENUM(CDMA_D_DATAIN_FORMAT_0,
+			PIXEL_FORMAT, T_A16B16G16R16_F), 8},
+	{FIELD_ENUM(CDMA_D_DATAIN_FORMAT_0,
+			PIXEL_FORMAT, T_A16Y16U16V16), 8},
+	{FIELD_ENUM(CDMA_D_DATAIN_FORMAT_0,
+			PIXEL_FORMAT, T_V16U16Y16A16), 8},
+	{FIELD_ENUM(CDMA_D_DATAIN_FORMAT_0,
+			PIXEL_FORMAT, T_A16Y16U16V16_F), 8},
+	{FIELD_ENUM(CDMA_D_DATAIN_FORMAT_0,
+			PIXEL_FORMAT, T_A8B8G8R8), 4},
+	{FIELD_ENUM(CDMA_D_DATAIN_FORMAT_0,
+			PIXEL_FORMAT, T_A8R8G8B8), 4},
+	{FIELD_ENUM(CDMA_D_DATAIN_FORMAT_0,
+			PIXEL_FORMAT, T_B8G8R8A8), 4},
+	{FIELD_ENUM(CDMA_D_DATAIN_FORMAT_0,
+			PIXEL_FORMAT, T_R8G8B8A8), 4},
+	{FIELD_ENUM(CDMA_D_DATAIN_FORMAT_0,
+			PIXEL_FORMAT, T_X8B8G8R8), 4},
+	{FIELD_ENUM(CDMA_D_DATAIN_FORMAT_0,
+			PIXEL_FORMAT, T_X8R8G8B8), 4},
+	{FIELD_ENUM(CDMA_D_DATAIN_FORMAT_0,
+			PIXEL_FORMAT, T_B8G8R8X8), 4},
+	{FIELD_ENUM(CDMA_D_DATAIN_FORMAT_0,
+			PIXEL_FORMAT, T_R8G8B8X8), 4},
+	{FIELD_ENUM(CDMA_D_DATAIN_FORMAT_0,
+			PIXEL_FORMAT, T_A2B10G10R10), 4},
+	{FIELD_ENUM(CDMA_D_DATAIN_FORMAT_0,
+			PIXEL_FORMAT, T_A2R10G10B10), 4},
+	{FIELD_ENUM(CDMA_D_DATAIN_FORMAT_0,
+			PIXEL_FORMAT, T_B10G10R10A2), 4},
+	{FIELD_ENUM(CDMA_D_DATAIN_FORMAT_0,
+			PIXEL_FORMAT, T_R10G10B10A2), 4},
+	{FIELD_ENUM(CDMA_D_DATAIN_FORMAT_0,
+			PIXEL_FORMAT, T_A2Y10U10V10), 4},
+	{FIELD_ENUM(CDMA_D_DATAIN_FORMAT_0,
+			PIXEL_FORMAT, T_V10U10Y10A2), 4},
+	{FIELD_ENUM(CDMA_D_DATAIN_FORMAT_0,
+			PIXEL_FORMAT, T_A8Y8U8V8), 4},
+	{FIELD_ENUM(CDMA_D_DATAIN_FORMAT_0,
+			PIXEL_FORMAT, T_V8U8Y8A8), 4},
+	{FIELD_ENUM(CDMA_D_DATAIN_FORMAT_0,
+			PIXEL_FORMAT, T_Y8___U8V8_N444), 1},
+	{FIELD_ENUM(CDMA_D_DATAIN_FORMAT_0,
+			PIXEL_FORMAT, T_Y8___V8U8_N444), 1},
+	{FIELD_ENUM(CDMA_D_DATAIN_FORMAT_0,
+			PIXEL_FORMAT, T_Y10___U10V10_N444), 2},
+	{FIELD_ENUM(CDMA_D_DATAIN_FORMAT_0,
+			PIXEL_FORMAT, T_Y10___V10U10_N444), 2},
+	{FIELD_ENUM(CDMA_D_DATAIN_FORMAT_0,
+			PIXEL_FORMAT, T_Y12___U12V12_N444), 2},
+	{FIELD_ENUM(CDMA_D_DATAIN_FORMAT_0,
+			PIXEL_FORMAT, T_Y12___V12U12_N444), 2},
+	{FIELD_ENUM(CDMA_D_DATAIN_FORMAT_0,
+			PIXEL_FORMAT, T_Y16___U16V16_N444), 2},
+	{FIELD_ENUM(CDMA_D_DATAIN_FORMAT_0,
+			PIXEL_FORMAT, T_Y16___V16U16_N444), 2},
+	{FIELD_ENUM(CDMA_D_DATAIN_FORMAT_0,
+			DATAIN_FORMAT, FEATURE), 2},
+	{FIELD_ENUM(CDMA_D_DATAIN_FORMAT_0,
+			DATAIN_FORMAT, PIXEL), 1},
+};
+
+static const uint8_t map_pixel[] = {
+	FIELD_ENUM(CDMA_D_DATAIN_FORMAT_0, PIXEL_MAPPING, PITCH_LINEAR),
+};
+
+static const uint8_t map_ram[] = {
+	FIELD_ENUM(CDMA_D_DAIN_RAM_TYPE_0, DATAIN_RAM_TYPE, MCIF),
+	FIELD_ENUM(CDMA_D_DAIN_RAM_TYPE_0, DATAIN_RAM_TYPE, CVIF),
+};
+
+static const uint8_t map_mean[] = {
+	FIELD_ENUM(CDMA_D_MEAN_FORMAT_0, MEAN_FORMAT, DISABLE),
+	FIELD_ENUM(CDMA_D_MEAN_FORMAT_0, MEAN_FORMAT, ENABLE),
+};
+
+static uint32_t
+get_in_format(uint8_t format)
+{
+	uint32_t in_format = 0;
+
+	if (format >= FORMAT_T_R8 && format < FORMAT_FEATURE) {
+		in_format = FIELD_ENUM(CDMA_D_DATAIN_FORMAT_0,
+						DATAIN_FORMAT, PIXEL);
+	} else if (format == FORMAT_FEATURE) {
+		in_format = FIELD_ENUM(CDMA_D_DATAIN_FORMAT_0,
+						DATAIN_FORMAT, FEATURE);
+	}
+
+	return in_format;
+}
+
+void
+dla_conv_set_producer(struct dla_engine *engine, int32_t group_id, int32_t rdma_group_id)
+{
+	uint32_t reg;
+
+	/* set producer pointer for all sub-modules */
+	reg = group_id << SHIFT(CACC_S_POINTER_0, PRODUCER);
+	cacc_reg_write(engine, S_POINTER, reg);
+	cmac_a_reg_write(engine, S_POINTER, reg);
+	cmac_b_reg_write(engine, S_POINTER, reg);
+	csc_reg_write(engine, S_POINTER, reg);
+	cdma_reg_write(engine, S_POINTER, reg);
+}
+
+int
+dla_conv_enable(struct dla_engine *engine, struct dla_processor_group *group)
+{
+	uint32_t reg;
+
+	do {
+		reg = cdma_reg_read(engine, S_CBUF_FLUSH_STATUS);
+	} while (!(reg & MASK(CDMA_S_CBUF_FLUSH_STATUS_0, FLUSH_DONE)));
+
+	/* enable all sub-modules */
+	reg = FIELD_ENUM(CACC_D_OP_ENABLE_0, OP_EN, ENABLE);
+	cacc_reg_write(engine, D_OP_ENABLE, reg);
+	cmac_a_reg_write(engine, D_OP_ENABLE, reg);
+	cmac_b_reg_write(engine, D_OP_ENABLE, reg);
+	csc_reg_write(engine, D_OP_ENABLE, reg);
+	cdma_reg_write(engine, D_OP_ENABLE, reg);
+
+	return 0;
+}
+
+void
+dla_conv_rdma_check(struct dla_processor_group *group)
+{
+	group->is_rdma_needed = 0;
+}
+
+static int32_t
+processor_conv_program(struct dla_engine *engine, struct dla_processor_group *group)
+{
+	int32_t ret = 0;
+	uint32_t reg, high, low, shift, mask;
+	uint32_t stride_x, stride_y, pad_x, pad_y;
+	uint64_t weight_address = 0;
+	uint64_t wmb_address = 0;
+	uint64_t wgs_address = 0;
+	uint64_t input_address = 0;
+	uint64_t output_address = 0;
+	uint32_t atom_size = 0;
+	bool weight_compress_support = false;
+	struct dla_conv_op_desc *conv_op;
+	struct dla_conv_surface_desc *conv_surface;
+
+	weight_compress_support = engine->config_data->weight_compress_support;
+	atom_size = engine->config_data->atom_size;
+	conv_op = &group->operation_desc->conv_op;
+	conv_surface = &group->surface_desc->conv_surface;
+
+	if (conv_op->weight_format == WEIGHT_FORMAT_COMPRESSED) {
+		ASSERT_GOTO((weight_compress_support), ret, -EINVAL, exit);
+		ASSERT_GOTO((conv_surface->wmb_data.address != -1),
+			ret, -EINVAL, exit);
+		dla_get_dma_cube_address(engine->driver_context,
+								 engine->task->task_data,
+								 conv_surface->wmb_data.address,
+								 conv_surface->wmb_data.offset,
+								 (void *)&wmb_address,
+								 DESTINATION_DMA);
+
+		ASSERT_GOTO((conv_surface->wgs_data.address != -1),
+			ret, -EINVAL, exit);
+		dla_get_dma_cube_address(engine->driver_context,
+								 engine->task->task_data,
+								 conv_surface->wgs_data.address,
+								 conv_surface->wgs_data.offset,
+								 (void *)&wgs_address,
+								 DESTINATION_DMA);
+	}
+
+	if (conv_surface->weight_data.address != -1) {
+		dla_get_dma_cube_address(engine->driver_context,
+								 engine->task->task_data,
+								 conv_surface->weight_data.address,
+								 conv_surface->weight_data.offset,
+								 (void *)&weight_address,
+								 DESTINATION_DMA);
+	}
+
+	if (conv_surface->dst_data.address != -1) {
+		dla_get_dma_cube_address(engine->driver_context,
+								 engine->task->task_data,
+								 conv_surface->dst_data.address,
+								 conv_surface->dst_data.offset,
+								 (void *)&output_address,
+								 DESTINATION_DMA);
+	}
+
+	ret = dla_read_input_address(engine, &conv_surface->src_data, &input_address,
+								 group->op_desc->index,
+								 group->roi_index,
+								 map_img_fmt[conv_op->data_format][1]);
+	if (ret)
+		goto exit;
+
+	ASSERT_GOTO((conv_op->out_cvt.scale  == 1), ret, -EINVAL, exit);
+	ASSERT_GOTO((conv_op->out_cvt.offset == 0), ret, -EINVAL, exit);
+
+	/* check if the register group is idle */
+	reg = cacc_reg_read(engine, S_STATUS);
+	mask = group->id ? MASK(CACC_S_STATUS_0, STATUS_1) :
+		MASK(CACC_S_STATUS_0, STATUS_0);
+	shift = group->id ? SHIFT(CACC_S_STATUS_0, STATUS_1) :
+		SHIFT(CACC_S_STATUS_0, STATUS_0);
+	reg = (reg & mask) >> shift;
+	ASSERT_GOTO((reg == FIELD_ENUM(CACC_S_STATUS_0, STATUS_0, IDLE)),
+				ret, -EINVAL, exit);
+
+	reg = cmac_a_reg_read(engine, S_STATUS);
+	mask = group->id ? MASK(CMAC_A_S_STATUS_0, STATUS_1) :
+		MASK(CMAC_A_S_STATUS_0, STATUS_0);
+	shift = group->id ? SHIFT(CMAC_A_S_STATUS_0, STATUS_1) :
+		SHIFT(CMAC_A_S_STATUS_0, STATUS_0);
+	reg = (reg & mask) >> shift;
+	ASSERT_GOTO((reg == FIELD_ENUM(CMAC_A_S_STATUS_0, STATUS_0, IDLE)),
+				ret, -EINVAL, exit);
+
+	reg = cmac_b_reg_read(engine, S_STATUS);
+	mask = group->id ? MASK(CMAC_B_S_STATUS_0, STATUS_1) :
+		MASK(CMAC_B_S_STATUS_0, STATUS_0);
+	shift = group->id ? SHIFT(CMAC_B_S_STATUS_0, STATUS_1) :
+		SHIFT(CMAC_B_S_STATUS_0, STATUS_0);
+	reg = (reg & mask) >> shift;
+	ASSERT_GOTO((reg == FIELD_ENUM(CMAC_B_S_STATUS_0, STATUS_0, IDLE)),
+				ret, -EINVAL, exit);
+
+	reg = csc_reg_read(engine, S_STATUS);
+	mask = group->id ? MASK(CSC_S_STATUS_0, STATUS_1) :
+		MASK(CSC_S_STATUS_0, STATUS_0);
+	shift = group->id ? SHIFT(CSC_S_STATUS_0, STATUS_1) :
+		SHIFT(CSC_S_STATUS_0, STATUS_0);
+	reg = (reg & mask) >> shift;
+	ASSERT_GOTO((reg == FIELD_ENUM(CSC_S_STATUS_0, STATUS_0, IDLE)),
+				ret, -EINVAL, exit);
+
+	reg = cdma_reg_read(engine, S_STATUS);
+	mask = group->id ? MASK(CDMA_S_STATUS_0, STATUS_1) :
+		MASK(CDMA_S_STATUS_0, STATUS_0);
+	shift = group->id ? SHIFT(CDMA_S_STATUS_0, STATUS_1) :
+		SHIFT(CDMA_S_STATUS_0, STATUS_0);
+	reg = (reg & mask) >> shift;
+	ASSERT_GOTO((reg == FIELD_ENUM(CDMA_S_STATUS_0, STATUS_0, IDLE)),
+		ret, -EINVAL, exit);
+
+	/* reverse config each sub-module in CC */
+
+	/* CACC */
+	reg = (map_conv[conv_op->conv_mode]
+		<< SHIFT(CACC_D_MISC_CFG_0, CONV_MODE)) |
+		(map_precision[conv_op->out_precision]
+		<< SHIFT(CACC_D_MISC_CFG_0, PROC_PRECISION));
+	cacc_reg_write(engine, D_MISC_CFG, reg);
+
+	reg = ((conv_surface->dst_data.width - 1)
+		<< SHIFT(CACC_D_DATAOUT_SIZE_0_0, DATAOUT_WIDTH)) |
+		((conv_surface->dst_data.height - 1)
+		<< SHIFT(CACC_D_DATAOUT_SIZE_0_0, DATAOUT_HEIGHT));
+	cacc_reg_write(engine, D_DATAOUT_SIZE_0, reg);
+
+	reg = ((conv_surface->dst_data.channel - 1)
+		<< SHIFT(CACC_D_DATAOUT_SIZE_1_0, DATAOUT_CHANNEL));
+	cacc_reg_write(engine, D_DATAOUT_SIZE_1, reg);
+
+	low = lower_32_bits(output_address);
+	cacc_reg_write(engine, D_DATAOUT_ADDR, low);
+	cacc_reg_write(engine, D_BATCH_NUMBER, conv_op->batch - 1);
+	cacc_reg_write(engine, D_LINE_STRIDE, conv_surface->dst_data.line_stride);
+	cacc_reg_write(engine, D_SURF_STRIDE, conv_surface->dst_data.surf_stride);
+
+	if (conv_surface->dst_data.width == 1 &&
+				conv_surface->dst_data.height == 1) {
+		ASSERT_GOTO((((uint32_t)conv_surface->dst_data.line_stride ==
+			(uint32_t)(conv_surface->dst_data.width * atom_size))),
+			ret, -EINVAL, exit);
+		reg = (CACC_D_DATAOUT_MAP_0_LINE_PACKED_TRUE <<
+				SHIFT(CACC_D_DATAOUT_MAP_0, LINE_PACKED));
+		reg |= (CACC_D_DATAOUT_MAP_0_SURF_PACKED_TRUE <<
+				SHIFT(CACC_D_DATAOUT_MAP_0, SURF_PACKED));
+	} else {
+		reg = (FIELD_ENUM(CACC_D_DATAOUT_MAP_0, LINE_PACKED, FALSE) <<
+				SHIFT(CACC_D_DATAOUT_MAP_0, LINE_PACKED));
+		reg |= (FIELD_ENUM(CACC_D_DATAOUT_MAP_0, SURF_PACKED, FALSE) <<
+				SHIFT(CACC_D_DATAOUT_MAP_0, SURF_PACKED));
+	}
+	cacc_reg_write(engine, D_DATAOUT_MAP, reg);
+
+	cacc_reg_write(engine, D_CLIP_CFG, conv_op->out_cvt.truncate);
+
+	/* CMAC */
+	reg = (map_conv[conv_op->conv_mode]
+		<< SHIFT(CMAC_A_D_MISC_CFG_0, CONV_MODE)) |
+		(map_precision[conv_op->out_precision]
+		<< SHIFT(CMAC_A_D_MISC_CFG_0, PROC_PRECISION));
+	cmac_a_reg_write(engine, D_MISC_CFG, reg);
+	cmac_b_reg_write(engine, D_MISC_CFG, reg);
+
+	/* CSC */
+	reg = (map_conv[conv_op->conv_mode]
+		<< SHIFT(CSC_D_MISC_CFG_0, CONV_MODE)) |
+		(map_precision[conv_op->out_precision]
+		<< SHIFT(CSC_D_MISC_CFG_0, IN_PRECISION)) |
+		(map_precision[conv_op->out_precision]
+		<< SHIFT(CSC_D_MISC_CFG_0, PROC_PRECISION)) |
+		(conv_op->data_reuse
+		<< SHIFT(CSC_D_MISC_CFG_0, DATA_REUSE)) |
+		(conv_op->weight_reuse
+		<< SHIFT(CSC_D_MISC_CFG_0, WEIGHT_REUSE)) |
+		(conv_op->skip_data_rls
+		<< SHIFT(CSC_D_MISC_CFG_0, SKIP_DATA_RLS)) |
+		(conv_op->skip_weight_rls
+		<< SHIFT(CSC_D_MISC_CFG_0, SKIP_WEIGHT_RLS));
+	csc_reg_write(engine, D_MISC_CFG, reg);
+
+	reg = (get_in_format(conv_op->data_format) <<
+		SHIFT(CSC_D_DATAIN_FORMAT_0, DATAIN_FORMAT));
+	csc_reg_write(engine, D_DATAIN_FORMAT, reg);
+
+	reg = ((conv_op->input_width_csc - 1)
+		<< SHIFT(CSC_D_DATAIN_SIZE_EXT_0_0, DATAIN_WIDTH_EXT)) |
+		((conv_op->input_height_csc - 1)
+		<< SHIFT(CSC_D_DATAIN_SIZE_EXT_0_0, DATAIN_HEIGHT_EXT));
+	csc_reg_write(engine, D_DATAIN_SIZE_EXT_0, reg);
+
+	reg = ((conv_op->input_channel_csc - 1)
+		<< SHIFT(CSC_D_DATAIN_SIZE_EXT_1_0, DATAIN_CHANNEL_EXT));
+	csc_reg_write(engine, D_DATAIN_SIZE_EXT_1, reg);
+
+	reg = ((conv_op->batch - 1)
+		<< SHIFT(CSC_D_BATCH_NUMBER_0, BATCHES));
+	csc_reg_write(engine, D_BATCH_NUMBER, reg);
+	reg = ((conv_op->post_extension)
+		<< SHIFT(CSC_D_POST_Y_EXTENSION_0, Y_EXTENSION));
+	csc_reg_write(engine, D_POST_Y_EXTENSION, reg);
+
+	reg = ((conv_op->entry_per_slice - 1)
+		<< SHIFT(CSC_D_ENTRY_PER_SLICE_0, ENTRIES));
+	csc_reg_write(engine, D_ENTRY_PER_SLICE, reg);
+
+	reg = (map_weight_fmt[conv_op->weight_format]
+		<< SHIFT(CSC_D_WEIGHT_FORMAT_0, WEIGHT_FORMAT));
+	csc_reg_write(engine, D_WEIGHT_FORMAT, reg);
+
+	reg = ((conv_op->kernel_width_csc - 1)
+		<< SHIFT(CSC_D_WEIGHT_SIZE_EXT_0_0, WEIGHT_WIDTH_EXT)) |
+		((conv_op->kernel_height_csc - 1)
+		<< SHIFT(CSC_D_WEIGHT_SIZE_EXT_0_0, WEIGHT_HEIGHT_EXT));
+	csc_reg_write(engine, D_WEIGHT_SIZE_EXT_0, reg);
+
+	reg = ((conv_op->kernel_channel_csc - 1)
+		<< SHIFT(CSC_D_WEIGHT_SIZE_EXT_1_0, WEIGHT_CHANNEL_EXT)) |
+		((conv_surface->dst_data.channel - 1)
+		<< SHIFT(CSC_D_WEIGHT_SIZE_EXT_1_0, WEIGHT_KERNEL));
+	csc_reg_write(engine, D_WEIGHT_SIZE_EXT_1, reg);
+
+	csc_reg_write(engine, D_WEIGHT_BYTES, conv_surface->weight_data.size);
+	csc_reg_write(engine, D_WMB_BYTES, conv_surface->wmb_data.size);
+
+	reg = ((conv_op->input_width_cmac - 1)
+		<< SHIFT(CSC_D_DATAOUT_SIZE_0_0, DATAOUT_WIDTH)) |
+		((conv_op->input_height_cmac - 1)
+		<< SHIFT(CSC_D_DATAOUT_SIZE_0_0, DATAOUT_HEIGHT));
+	csc_reg_write(engine, D_DATAOUT_SIZE_0, reg);
+
+	reg = ((conv_surface->dst_data.channel - 1)
+		<< SHIFT(CSC_D_DATAOUT_SIZE_1_0, DATAOUT_CHANNEL));
+	csc_reg_write(engine, D_DATAOUT_SIZE_1, reg);
+
+	reg = ((conv_surface->dst_data.width *
+				conv_surface->dst_data.height - 1)
+		<< SHIFT(CSC_D_ATOMICS_0, ATOMICS));
+	csc_reg_write(engine, D_ATOMICS, reg);
+	reg = ((conv_op->release - 1)
+		<< SHIFT(CSC_D_RELEASE_0, RLS_SLICES));
+	csc_reg_write(engine, D_RELEASE, reg);
+
+	if (conv_op->conv_mode == CONV_MODE_DIRECT) {
+		stride_x = conv_op->conv_stride_x - 1;
+		stride_y = conv_op->conv_stride_y - 1;
+		pad_x = conv_op->pad_x_left;
+		pad_y = conv_op->pad_y_top;
+	} else {
+		stride_x = 0;
+		stride_y = 0;
+		pad_x = 0;
+		pad_y = 0;
+	}
+
+	reg = (stride_x
+		<< SHIFT(CSC_D_CONV_STRIDE_EXT_0, CONV_X_STRIDE_EXT)) |
+		(stride_y
+		<< SHIFT(CSC_D_CONV_STRIDE_EXT_0, CONV_Y_STRIDE_EXT));
+	csc_reg_write(engine, D_CONV_STRIDE_EXT, reg);
+
+	reg = ((conv_op->dilation_x - 1)
+		<< SHIFT(CSC_D_DILATION_EXT_0, X_DILATION_EXT)) |
+		((conv_op->dilation_y - 1)
+		<< SHIFT(CSC_D_DILATION_EXT_0, Y_DILATION_EXT));
+	csc_reg_write(engine, D_DILATION_EXT, reg);
+
+	reg = (pad_x
+		<< SHIFT(CSC_D_ZERO_PADDING_0, PAD_LEFT)) |
+		(pad_y
+		<< SHIFT(CSC_D_ZERO_PADDING_0, PAD_TOP));
+	csc_reg_write(engine, D_ZERO_PADDING, reg);
+
+	reg = (conv_op->pad_val
+		<< SHIFT(CSC_D_ZERO_PADDING_VALUE_0, PAD_VALUE)) &
+		MASK(CSC_D_ZERO_PADDING_VALUE_0, PAD_VALUE);
+	csc_reg_write(engine, D_ZERO_PADDING_VALUE, reg);
+
+	reg = ((conv_op->data_bank - 1)
+		<< SHIFT(CSC_D_BANK_0, DATA_BANK)) |
+		((conv_op->weight_bank - 1)
+		<< SHIFT(CSC_D_BANK_0, WEIGHT_BANK));
+	csc_reg_write(engine, D_BANK, reg);
+	csc_reg_write(engine, D_PRA_CFG, conv_op->pra_truncate);
+
+	/* CBUF */
+	/* there's no CBUF register */
+
+	/* CDMA */
+	reg = (map_conv[conv_op->conv_mode]
+		<< SHIFT(CDMA_D_MISC_CFG_0, CONV_MODE)) |
+		(map_precision[conv_op->in_precision]
+		<< SHIFT(CDMA_D_MISC_CFG_0, IN_PRECISION)) |
+		(map_precision[conv_op->out_precision]
+		<< SHIFT(CDMA_D_MISC_CFG_0, PROC_PRECISION)) |
+		(conv_op->data_reuse
+		<< SHIFT(CDMA_D_MISC_CFG_0, DATA_REUSE)) |
+		(conv_op->weight_reuse
+		<< SHIFT(CDMA_D_MISC_CFG_0, WEIGHT_REUSE)) |
+		(conv_op->skip_data_rls
+		<< SHIFT(CDMA_D_MISC_CFG_0, SKIP_DATA_RLS)) |
+		(conv_op->skip_weight_rls
+		<< SHIFT(CDMA_D_MISC_CFG_0, SKIP_WEIGHT_RLS));
+	cdma_reg_write(engine, D_MISC_CFG, reg);
+
+	reg = (get_in_format(conv_op->data_format) <<
+		SHIFT(CDMA_D_DATAIN_FORMAT_0, DATAIN_FORMAT)) |
+		(map_img_fmt[conv_op->data_format][0]
+		<< SHIFT(CDMA_D_DATAIN_FORMAT_0, PIXEL_FORMAT)) |
+		(map_pixel[conv_op->pixel_mapping]
+		<< SHIFT(CDMA_D_DATAIN_FORMAT_0, PIXEL_MAPPING)) |
+		(conv_op->pixel_override
+		<< SHIFT(CDMA_D_DATAIN_FORMAT_0, PIXEL_SIGN_OVERRIDE));
+	cdma_reg_write(engine, D_DATAIN_FORMAT, reg);
+
+	reg = ((conv_surface->src_data.width - 1)
+		<< SHIFT(CDMA_D_DATAIN_SIZE_0_0, DATAIN_WIDTH)) |
+		((conv_surface->src_data.height - 1)
+		<< SHIFT(CDMA_D_DATAIN_SIZE_0_0, DATAIN_HEIGHT));
+	cdma_reg_write(engine, D_DATAIN_SIZE_0, reg);
+
+	reg = ((conv_surface->src_data.channel - 1)
+		<< SHIFT(CDMA_D_DATAIN_SIZE_1_0, DATAIN_CHANNEL));
+	cdma_reg_write(engine, D_DATAIN_SIZE_1, reg);
+
+	reg = ((conv_op->input_width_csc - 1)
+		<< SHIFT(CDMA_D_DATAIN_SIZE_EXT_0_0, DATAIN_WIDTH_EXT)) |
+		((conv_op->input_height_csc - 1)
+		<< SHIFT(CDMA_D_DATAIN_SIZE_EXT_0_0, DATAIN_HEIGHT_EXT));
+	cdma_reg_write(engine, D_DATAIN_SIZE_EXT_0, reg);
+
+	reg = (map_ram[conv_surface->src_data.type]
+		<< SHIFT(CDMA_D_DAIN_RAM_TYPE_0, DATAIN_RAM_TYPE));
+	cdma_reg_write(engine, D_DAIN_RAM_TYPE, reg);
+
+	high = upper_32_bits(input_address);
+	low = lower_32_bits(input_address);
+	cdma_reg_write(engine, D_DAIN_ADDR_HIGH_0, high);
+	cdma_reg_write(engine, D_DAIN_ADDR_LOW_0, low);
+
+	high = upper_32_bits((input_address + conv_surface->offset_u));
+	low = lower_32_bits(input_address + conv_surface->offset_u);
+	cdma_reg_write(engine, D_DAIN_ADDR_HIGH_1, high);
+	cdma_reg_write(engine, D_DAIN_ADDR_LOW_1, low);
+
+	cdma_reg_write(engine, D_LINE_STRIDE, conv_surface->src_data.line_stride);
+	cdma_reg_write(engine, D_SURF_STRIDE, conv_surface->src_data.surf_stride);
+	cdma_reg_write(engine, D_LINE_UV_STRIDE, conv_surface->in_line_uv_stride);
+
+	reg = ((conv_surface->src_data.line_stride ==
+			((uint32_t)conv_surface->src_data.width * atom_size))
+		<< SHIFT(CDMA_D_DAIN_MAP_0, LINE_PACKED));
+	reg |= ((conv_surface->src_data.surf_stride ==
+			((uint32_t)(conv_surface->src_data.width *
+			conv_surface->src_data.height) * atom_size))
+		<< SHIFT(CDMA_D_DAIN_MAP_0, SURF_PACKED));
+	cdma_reg_write(engine, D_DAIN_MAP, reg);
+
+	reg = ((conv_op->batch - 1)
+		<< SHIFT(CDMA_D_BATCH_NUMBER_0, BATCHES));
+	cdma_reg_write(engine, D_BATCH_NUMBER, reg);
+
+	cdma_reg_write(engine, D_BATCH_STRIDE, conv_op->batch_stride);
+
+	reg = ((conv_op->entry_per_slice - 1)
+		<< SHIFT(CDMA_D_ENTRY_PER_SLICE_0, ENTRIES));
+	cdma_reg_write(engine, D_ENTRY_PER_SLICE, reg);
+
+	reg = ((conv_op->fetch_grain - 1)
+		<< SHIFT(CDMA_D_FETCH_GRAIN_0, GRAINS));
+	cdma_reg_write(engine, D_FETCH_GRAIN, reg);
+
+	reg = (map_weight_fmt[conv_op->weight_format]
+		<< SHIFT(CDMA_D_WEIGHT_FORMAT_0, WEIGHT_FORMAT));
+	cdma_reg_write(engine, D_WEIGHT_FORMAT, reg);
+
+	reg = ((conv_op->bytes_per_kernel - 1)
+		<< SHIFT(CDMA_D_WEIGHT_SIZE_0_0, BYTE_PER_KERNEL));
+	cdma_reg_write(engine, D_WEIGHT_SIZE_0, reg);
+
+	reg = ((conv_surface->dst_data.channel - 1)
+		<< SHIFT(CDMA_D_WEIGHT_SIZE_1_0, WEIGHT_KERNEL));
+	cdma_reg_write(engine, D_WEIGHT_SIZE_1, reg);
+
+	reg = (map_ram[conv_surface->weight_data.type]
+		<< SHIFT(CDMA_D_WEIGHT_RAM_TYPE_0, WEIGHT_RAM_TYPE));
+	cdma_reg_write(engine, D_WEIGHT_RAM_TYPE, reg);
+
+	high = upper_32_bits(weight_address);
+	low = lower_32_bits(weight_address);
+	cdma_reg_write(engine, D_WEIGHT_ADDR_HIGH, high);
+	cdma_reg_write(engine, D_WEIGHT_ADDR_LOW, low);
+	cdma_reg_write(engine, D_WEIGHT_BYTES, conv_surface->weight_data.size);
+
+	if (conv_op->weight_format == WEIGHT_FORMAT_COMPRESSED) {
+		high = upper_32_bits(wgs_address);
+		low = lower_32_bits(wgs_address);
+		cdma_reg_write(engine, D_WGS_ADDR_HIGH, high);
+		cdma_reg_write(engine, D_WGS_ADDR_LOW, low);
+
+		high = upper_32_bits(wmb_address);
+		low = lower_32_bits(wmb_address);
+		cdma_reg_write(engine, D_WMB_ADDR_HIGH, high);
+		cdma_reg_write(engine, D_WMB_ADDR_LOW, low);
+		cdma_reg_write(engine, D_WMB_BYTES, conv_surface->wmb_data.size);
+	}
+
+	reg = (map_mean[conv_op->mean_format]
+		<< SHIFT(CDMA_D_MEAN_FORMAT_0, MEAN_FORMAT));
+	cdma_reg_write(engine, D_MEAN_FORMAT, reg);
+
+	if (conv_op->mean_format == MEAN_FORMAT_ENABLE) {
+		reg = ((conv_op->mean_ry
+			<< SHIFT(CDMA_D_MEAN_GLOBAL_0_0, MEAN_RY)) &
+			MASK(CDMA_D_MEAN_GLOBAL_0_0, MEAN_RY)) |
+			((conv_op->mean_gu
+			<< SHIFT(CDMA_D_MEAN_GLOBAL_0_0, MEAN_GU)) &
+			MASK(CDMA_D_MEAN_GLOBAL_0_0, MEAN_GU));
+		cdma_reg_write(engine, D_MEAN_GLOBAL_0, reg);
+
+		reg = ((conv_op->mean_bv
+			<< SHIFT(CDMA_D_MEAN_GLOBAL_1_0, MEAN_BV))&
+			MASK(CDMA_D_MEAN_GLOBAL_1_0, MEAN_BV)) |
+			((conv_op->mean_ax
+			<< SHIFT(CDMA_D_MEAN_GLOBAL_1_0, MEAN_AX))&
+			MASK(CDMA_D_MEAN_GLOBAL_1_0, MEAN_AX));
+		cdma_reg_write(engine, D_MEAN_GLOBAL_1, reg);
+	}
+
+	if (conv_op->in_cvt.enable) {
+		reg = ((FIELD_ENUM(CDMA_D_CVT_CFG_0, CVT_EN, ENABLE))
+			<< SHIFT(CDMA_D_CVT_CFG_0, CVT_EN)) |
+			(conv_op->in_cvt.truncate
+			<< SHIFT(CDMA_D_CVT_CFG_0, CVT_TRUNCATE));
+		cdma_reg_write(engine, D_CVT_CFG, reg);
+		cdma_reg_write(engine, D_CVT_OFFSET, conv_op->in_cvt.offset);
+		cdma_reg_write(engine, D_CVT_SCALE, conv_op->in_cvt.scale);
+	} else {
+		reg = ((FIELD_ENUM(CDMA_D_CVT_CFG_0, CVT_EN, DISABLE))
+			<< SHIFT(CDMA_D_CVT_CFG_0, CVT_EN));
+		cdma_reg_write(engine, D_CVT_CFG, reg);
+	}
+
+	reg = ((conv_op->conv_stride_x - 1)
+		<< SHIFT(CDMA_D_CONV_STRIDE_0, CONV_X_STRIDE)) |
+		((conv_op->conv_stride_y - 1)
+		<< SHIFT(CDMA_D_CONV_STRIDE_0, CONV_Y_STRIDE));
+	cdma_reg_write(engine, D_CONV_STRIDE, reg);
+
+	reg = (conv_op->pad_x_left <<
+		SHIFT(CDMA_D_ZERO_PADDING_0, PAD_LEFT)) |
+		(conv_op->pad_x_right
+		<< SHIFT(CDMA_D_ZERO_PADDING_0, PAD_RIGHT)) |
+		(conv_op->pad_y_top
+		<< SHIFT(CDMA_D_ZERO_PADDING_0, PAD_TOP)) |
+		(conv_op->pad_y_bottom
+		<< SHIFT(CDMA_D_ZERO_PADDING_0, PAD_BOTTOM));
+	cdma_reg_write(engine, D_ZERO_PADDING,   reg);
+
+	reg = conv_op->pad_val <<
+		SHIFT(CDMA_D_ZERO_PADDING_VALUE_0, PAD_VALUE) &
+		MASK(CDMA_D_ZERO_PADDING_VALUE_0, PAD_VALUE);
+	cdma_reg_write(engine, D_ZERO_PADDING_VALUE, reg);
+	reg = ((conv_op->weight_bank - 1)
+		<< SHIFT(CDMA_D_BANK_0, WEIGHT_BANK)) |
+		((conv_op->data_bank - 1)
+		<< SHIFT(CDMA_D_BANK_0, DATA_BANK));
+	cdma_reg_write(engine, D_BANK, reg);
+
+exit:
+	return ret;
+}
+
+int
+dla_conv_is_ready(struct dla_processor *processor,
+				  struct dla_processor_group *group)
+{
+	return 1;
+}
+
+void
+dla_conv_dump_config(struct dla_processor_group *group)
+{
+	struct dla_conv_op_desc *conv_op;
+	struct dla_conv_surface_desc *conv_surface;
+
+	conv_surface = &group->surface_desc->conv_surface;
+	conv_op = &group->operation_desc->conv_op;
+}
+
+int
+dla_conv_program(struct dla_engine *engine, struct dla_processor_group *group)
+{
+	int32_t ret;
+
+	ret = processor_conv_program(engine, group);
+	if (ret)
+		goto exit;
+
+exit:
+	return ret;
+}
drivers/gpu/drm/nvdla/nvdla_drm.c new file +695 -0
new file mode 100644
index 000000000000..9217eee1de3b
@@ -0,0 +1,695 @@
+// SPDX-License-Identifier: BSD-3-Clause OR GPL-2.0
+/*
+ * Copyright (C) 2017-2018 NVIDIA CORPORATION
+ * Copyright (C) 2022 Cai Huoqing
+ */
+
+#include <linux/dma-buf.h>
+#include <linux/dma-mapping.h>
+#include <linux/fs.h>
+#include <linux/interrupt.h>
+#include <linux/irq.h>
+#include <linux/irqdomain.h>
+#include <linux/module.h>
+#include <linux/of.h>
+#include <linux/of_device.h>
+#include <linux/of_irq.h>
+#include <linux/of_platform.h>
+#include <linux/platform_device.h>
+#include <linux/printk.h>
+#include <linux/slab.h>
+#include <linux/spinlock.h>
+#include <linux/time.h>
+#include <linux/uaccess.h>
+#include <linux/types.h>
+
+#include "nvdla_drm.h"
+#include "nvdla_ioctl.h"
+#include "nvdla_engine.h"
+
+static struct nvdla_config nvdla_config_os_initial = {
+	.atom_size = 32,
+	.bdma_enable = true,
+	.rubik_enable = true,
+	.weight_compress_support = true,
+};
+
+static struct nvdla_config nvdla_config_small = {
+	//.atom_size = 8,
+	.atom_size = 32,  // nv_large config
+	.bdma_enable = false,
+	.rubik_enable = false,
+	.weight_compress_support = false,
+};
+
+int64_t dla_get_time_us(void)
+{
+	return ktime_get_ns() / NSEC_PER_USEC;
+}
+
+void dla_reg_write(void *driver_context, uint32_t addr, uint32_t reg)
+{
+	struct nvdla_device *nvdla_dev =
+			(struct nvdla_device *)driver_context;
+
+	if (!nvdla_dev)
+		return;
+
+	writel(reg, nvdla_dev->base + addr);
+}
+
+uint32_t dla_reg_read(void *driver_context, uint32_t addr)
+{
+	struct nvdla_device *nvdla_dev =
+			(struct nvdla_device *)driver_context;
+
+	if (!nvdla_dev)
+		return 0;
+
+	return readl(nvdla_dev->base + addr);
+}
+
+static irqreturn_t nvdla_engine_isr(int32_t irq, void *data)
+{
+	unsigned long flags;
+	uint32_t mask;
+	uint32_t reg;
+	struct dla_processor *processor = NULL;
+	struct dla_processor_group *group;
+	struct dla_engine *engine;
+	struct nvdla_device *nvdla_dev = (struct nvdla_device *)data;
+
+	if (!nvdla_dev)
+		return IRQ_NONE;
+
+	engine = nvdla_dev->engine_context;
+	spin_lock_irqsave(&nvdla_dev->nvdla_lock, flags);
+
+	mask = glb_reg_read(engine, S_INTR_MASK);
+	reg = glb_reg_read(engine, S_INTR_STATUS);
+
+	if (reg & MASK(GLB_S_INTR_STATUS_0, CACC_DONE_STATUS0)) {
+		processor = &engine->processors[DLA_OP_CONV];
+		group = &processor->groups[0];
+		group->events |= (1 << DLA_EVENT_OP_COMPLETED);
+	}
+	if (reg & MASK(GLB_S_INTR_STATUS_0, CACC_DONE_STATUS1)) {
+		processor = &engine->processors[DLA_OP_CONV];
+		group = &processor->groups[1];
+		group->events |= (1 << DLA_EVENT_OP_COMPLETED);
+	}
+	if (reg & MASK(GLB_S_INTR_STATUS_0, SDP_DONE_STATUS0)) {
+		processor = &engine->processors[DLA_OP_SDP];
+		group = &processor->groups[0];
+		group->events |= (1 << DLA_EVENT_OP_COMPLETED);
+	}
+	if (reg & MASK(GLB_S_INTR_STATUS_0, SDP_DONE_STATUS1)) {
+		processor = &engine->processors[DLA_OP_SDP];
+		group = &processor->groups[1];
+		group->events |= (1 << DLA_EVENT_OP_COMPLETED);
+	}
+	if (reg & MASK(GLB_S_INTR_STATUS_0, CDP_DONE_STATUS0)) {
+		processor = &engine->processors[DLA_OP_CDP];
+		group = &processor->groups[0];
+		group->events |= (1 << DLA_EVENT_OP_COMPLETED);
+	}
+	if (reg & MASK(GLB_S_INTR_STATUS_0, CDP_DONE_STATUS1)) {
+		processor = &engine->processors[DLA_OP_CDP];
+		group = &processor->groups[1];
+		group->events |= (1 << DLA_EVENT_OP_COMPLETED);
+	}
+	if (reg & MASK(GLB_S_INTR_STATUS_0, RUBIK_DONE_STATUS0)) {
+		processor = &engine->processors[DLA_OP_RUBIK];
+		group = &processor->groups[0];
+		group->events |= (1 << DLA_EVENT_OP_COMPLETED);
+	}
+	if (reg & MASK(GLB_S_INTR_STATUS_0, RUBIK_DONE_STATUS1)) {
+		processor = &engine->processors[DLA_OP_RUBIK];
+		group = &processor->groups[1];
+		group->events |= (1 << DLA_EVENT_OP_COMPLETED);
+	}
+	if (reg & MASK(GLB_S_INTR_STATUS_0, PDP_DONE_STATUS0)) {
+		processor = &engine->processors[DLA_OP_PDP];
+		group = &processor->groups[0];
+		group->events |= (1 << DLA_EVENT_OP_COMPLETED);
+	}
+	if (reg & MASK(GLB_S_INTR_STATUS_0, PDP_DONE_STATUS1)) {
+		processor = &engine->processors[DLA_OP_PDP];
+		group = &processor->groups[1];
+		group->events |= (1 << DLA_EVENT_OP_COMPLETED);
+	}
+	if (reg & MASK(GLB_S_INTR_STATUS_0, BDMA_DONE_STATUS0)) {
+		processor = &engine->processors[DLA_OP_BDMA];
+		group = &processor->groups[0];
+		group->events |= (1 << DLA_EVENT_OP_COMPLETED);
+	}
+	if (reg & MASK(GLB_S_INTR_STATUS_0, BDMA_DONE_STATUS1)) {
+		processor = &engine->processors[DLA_OP_BDMA];
+		group = &processor->groups[1];
+		group->events |= (1 << DLA_EVENT_OP_COMPLETED);
+	}
+	if (reg & MASK(GLB_S_INTR_STATUS_0, CDMA_DAT_DONE_STATUS0)) {
+		processor = &engine->processors[DLA_OP_CONV];
+		group = &processor->groups[0];
+		group->events |= (1 << DLA_EVENT_CDMA_DT_DONE);
+	}
+	if (reg & MASK(GLB_S_INTR_STATUS_0, CDMA_DAT_DONE_STATUS1)) {
+		processor = &engine->processors[DLA_OP_CONV];
+		group = &processor->groups[1];
+		group->events |= (1 << DLA_EVENT_CDMA_DT_DONE);
+	}
+	if (reg & MASK(GLB_S_INTR_STATUS_0, CDMA_WT_DONE_STATUS0)) {
+		processor = &engine->processors[DLA_OP_CONV];
+		group = &processor->groups[0];
+		group->events |= (1 << DLA_EVENT_CDMA_WT_DONE);
+	}
+	if (reg & MASK(GLB_S_INTR_STATUS_0, CDMA_WT_DONE_STATUS1)) {
+		processor = &engine->processors[DLA_OP_CONV];
+		group = &processor->groups[1];
+		group->events |= (1 << DLA_EVENT_CDMA_WT_DONE);
+	}
+
+	glb_reg_write(engine, S_INTR_STATUS, reg);
+	mask = glb_reg_read(engine, S_INTR_MASK);
+	reg = glb_reg_read(engine, S_INTR_STATUS);
+
+	complete(&nvdla_dev->event_notifier);
+	spin_unlock_irqrestore(&nvdla_dev->nvdla_lock, flags);
+
+	return IRQ_HANDLED;
+}
+
+static int32_t dla_read_dma_address(void *driver_context, void *task_data,
+						int16_t index, void *dst)
+{
+	int32_t ret = 0;
+	struct nvdla_mem_handle *handles;
+	dma_addr_t *phys_addr = (dma_addr_t *)(dst);
+	struct nvdla_device *nvdla_dev =
+			(struct nvdla_device *)driver_context;
+	struct nvdla_task *task = (struct nvdla_task *)task_data;
+
+	if (index == -1 || index > task->num_addresses)
+		return -EINVAL;
+
+	handles = (struct nvdla_mem_handle *)task->address_list;
+	ret = nvdla_gem_dma_addr(nvdla_dev->drm, task->file,
+					handles[index].handle,
+					phys_addr);
+
+	/* Add offset to IOVA address */
+	*phys_addr = *phys_addr + handles[index].offset;
+
+	return ret;
+}
+
+static int32_t dla_read_cpu_address(void *driver_context, void *task_data,
+						int16_t index, void *dst)
+{
+	uint64_t *temp = (uint64_t *)dst;
+	struct nvdla_task *task = (struct nvdla_task *)task_data;
+
+	if (index == -1 || index > task->num_addresses)
+		return -EINVAL;
+
+	*temp = (uint64_t)index;
+	return 0;
+}
+
+int32_t dla_get_dma_address(void *driver_context, void *task_data,
+					int16_t index, void *dst_ptr,
+					uint32_t destination)
+{
+	int32_t ret = 0;
+
+	if (destination == DESTINATION_PROCESSOR) {
+		ret = dla_read_cpu_address(driver_context, task_data,
+						index, dst_ptr);
+	} else if (destination == DESTINATION_DMA) {
+		ret = dla_read_dma_address(driver_context, task_data,
+						index, dst_ptr);
+	} else {
+		ret = -EINVAL;
+	}
+
+	return ret;
+}
+
+int32_t dla_data_write(void *driver_context, void *task_data,
+				void *src, uint64_t dst,
+				uint32_t size, uint64_t offset)
+{
+	int32_t ret;
+	void *ptr = NULL;
+	struct dma_buf *buf;
+	struct iosys_map map;
+	struct nvdla_mem_handle *handles;
+	struct nvdla_task *task = (struct nvdla_task *)task_data;
+
+	handles = task->address_list;
+	buf = dma_buf_get(handles[dst].handle);
+	if (IS_ERR(buf)) {
+		pr_err("%s: Failed get dma_buf for handle=%d\n", __func__,
+						handles[dst].handle);
+		return -EFAULT;
+	}
+
+	ret = dma_buf_begin_cpu_access(buf, DMA_BIDIRECTIONAL);
+	if (ret)
+		goto put_dma_buf;
+
+	ret = dma_buf_vmap(buf, &map);
+	ptr = ret ? NULL : map.vaddr;
+	if (!ptr) {
+		pr_err("%s: Failed to vmap dma_buf for handle=%d\n", __func__,
+						handles[dst].handle);
+		ret = -ENOMEM;
+		goto end_cpu_access;
+	}
+
+
+	memcpy((void *)((uint8_t *)ptr + offset), src, size);
+
+	dma_buf_vunmap(buf, ptr);
+
+end_cpu_access:
+	dma_buf_end_cpu_access(buf, DMA_BIDIRECTIONAL);
+
+put_dma_buf:
+	dma_buf_put(buf);
+
+	return ret;
+}
+
+int32_t dla_data_read(void *driver_context, void *task_data,
+				uint64_t src, void *dst,
+				uint32_t size, uint64_t offset)
+{
+	int32_t ret;
+	void *ptr = NULL;
+	struct dma_buf *buf;
+	struct iosys_map map;
+	struct nvdla_mem_handle *handles;
+	struct nvdla_task *task = (struct nvdla_task *)task_data;
+
+	handles = task->address_list;
+
+	buf = dma_buf_get(handles[src].handle);
+	if (IS_ERR(buf)) {
+		pr_err("%s: Failed get dma_buf for handle=%d\n", __func__,
+						handles[src].handle);
+		return -EFAULT;
+	}
+
+	ret = dma_buf_begin_cpu_access(buf, DMA_BIDIRECTIONAL);
+	if (ret)
+		goto put_dma_buf;
+
+	ret = dma_buf_vmap(buf, &map);
+	ptr = ret ? NULL : map.vaddr;
+	if (!ptr) {
+		pr_err("%s: Failed to vmap dma_buf for handle=%d\n", __func__,
+						handles[src].handle);
+		ret = -ENOMEM;
+		goto end_cpu_access;
+	}
+
+	memcpy(dst, (void *)(((uint8_t *)ptr) + offset), size);
+
+	dma_buf_vunmap(buf, ptr);
+
+end_cpu_access:
+	dma_buf_end_cpu_access(buf, DMA_BIDIRECTIONAL);
+
+put_dma_buf:
+	dma_buf_put(buf);
+
+	return ret;
+}
+
+int32_t nvdla_task_submit(struct nvdla_device *nvdla_dev, struct nvdla_task *task)
+{
+	int32_t err = 0;
+	uint32_t task_complete = 0;
+
+	//trace_printk("__nvdla_task_submit_entry\n");
+	nvdla_dev->task = task;
+
+	err = dla_execute_task(nvdla_dev->engine_context, (void *)task, nvdla_dev->config_data);
+	if (err) {
+		pr_err("Task execution failed\n");
+		return err;
+	}
+
+	pr_debug("Wait for task complete\n");
+
+	while (1) {
+		unsigned long flags;
+
+		wait_for_completion(&nvdla_dev->event_notifier);
+
+		spin_lock_irqsave(&nvdla_dev->nvdla_lock, flags);
+
+		err = dla_process_events(nvdla_dev->engine_context, &task_complete);
+
+		spin_unlock_irqrestore(&nvdla_dev->nvdla_lock, flags);
+
+		if (err || task_complete)
+			break;
+	}
+
+	pr_debug("Task complete\n");
+	dla_clear_task(nvdla_dev->engine_context);
+	//trace_printk("__nvdla_task_submit_exit\n");
+
+	return err;
+}
+
+static union dla_operation_container operation_desc[DLA_OP_NUM][DLA_NUM_GROUPS];
+static union dla_surface_container surface_desc[DLA_OP_NUM][DLA_NUM_GROUPS];
+
+static struct dla_task global_task;
+
+static struct dla_engine engine = {
+	.processors[DLA_OP_BDMA] = {
+		.name = "BDMA",
+		.op_type = DLA_OP_BDMA,
+		.program = dla_bdma_program,
+		.enable = dla_bdma_enable,
+		.set_producer = dla_bdma_set_producer,
+		.is_ready = dla_bdma_is_ready,
+		.dump_config = dla_bdma_dump_config,
+		.rdma_check = dla_bdma_rdma_check,
+		.consumer_ptr = 0,
+		.roi_index = 0,
+		.group_status = 0,
+		.rdma_status = 0,
+		.last_group = 1,
+		.groups[0] = {
+			.id = 0,
+			.rdma_id = 0,
+			.active = 0,
+			.events = 0,
+			.roi_index = 0,
+			.is_rdma_needed = 0,
+			.lut_index = -1,
+			.operation_desc = &operation_desc[DLA_OP_BDMA][0],
+			.surface_desc = &surface_desc[DLA_OP_BDMA][0],
+		},
+		.groups[1] = {
+			.id = 1,
+			.rdma_id = 0,
+			.active = 0,
+			.events = 0,
+			.roi_index = 0,
+			.is_rdma_needed = 0,
+			.lut_index = -1,
+			.operation_desc = &operation_desc[DLA_OP_BDMA][1],
+			.surface_desc = &surface_desc[DLA_OP_BDMA][1],
+		},
+	},
+	.processors[DLA_OP_CONV] = {
+		.name = "Convolution",
+		.op_type = DLA_OP_CONV,
+		.program = dla_conv_program,
+		.enable = dla_conv_enable,
+		.set_producer = dla_conv_set_producer,
+		.is_ready = dla_conv_is_ready,
+		.dump_config = dla_conv_dump_config,
+		.rdma_check = dla_conv_rdma_check,
+		.consumer_ptr = 0,
+		.roi_index = 0,
+		.group_status = 0,
+		.rdma_status = 0,
+		.last_group = 1,
+		.groups[0] = {
+			.id = 0,
+			.rdma_id = 0,
+			.active = 0,
+			.events = 0,
+			.roi_index = 0,
+			.is_rdma_needed = 0,
+			.lut_index = -1,
+			.operation_desc = &operation_desc[DLA_OP_CONV][0],
[… more lines in the raw message]

Lines of this file are not shown. View the raw message.

drivers/gpu/drm/nvdla/nvdla_drm.h new file +127 -0
new file mode 100644
index 000000000000..356fc399accb

Lines of this file are not shown. View the raw message.

drivers/gpu/drm/nvdla/nvdla_engine.c new file +233 -0
new file mode 100644
index 000000000000..26ebd6c362b0

Lines of this file are not shown. View the raw message.

drivers/gpu/drm/nvdla/nvdla_engine.h new file +272 -0
new file mode 100644
index 000000000000..1976f5919543

Lines of this file are not shown. View the raw message.

drivers/gpu/drm/nvdla/nvdla_gem.c new file +393 -0
new file mode 100644
index 000000000000..cccf6d01a564

Lines of this file are not shown. View the raw message.

drivers/gpu/drm/nvdla/nvdla_ioctl.h new file +99 -0
new file mode 100644
index 000000000000..e21d4007b7f4

Lines of this file are not shown. View the raw message.

drivers/gpu/drm/nvdla/nvdla_pdp.c new file +446 -0
new file mode 100644
index 000000000000..2ed0ac0f021a

Lines of this file are not shown. View the raw message.

drivers/gpu/drm/nvdla/nvdla_reg.h new file +6411 -0
new file mode 100644
index 000000000000..5ca2897405bc

Lines of this file are not shown. View the raw message.

drivers/gpu/drm/nvdla/nvdla_rubik.c new file +217 -0
new file mode 100644
index 000000000000..3bfb3b0cb786

Lines of this file are not shown. View the raw message.

drivers/gpu/drm/nvdla/nvdla_sched.h new file +52 -0
new file mode 100644
index 000000000000..d3a198127006

Lines of this file are not shown. View the raw message.

drivers/gpu/drm/nvdla/nvdla_scheduler.c new file +1005 -0
new file mode 100644
index 000000000000..b814077478c6

Lines of this file are not shown. View the raw message.

drivers/gpu/drm/nvdla/nvdla_sdp.c new file +728 -0
new file mode 100644
index 000000000000..655ec7195ee9

Lines of this file are not shown. View the raw message.

The rest of this diff is not shown. View the raw message.

Signature · 1 line
2.25.1

Part of the diff is not shown here. View the raw message.

#9 lore raw
In reply to Cai Huoqing

(Resending, as some MLs didn't like the size of the origninal mail.)

Hi,

thanks for your submission. Some general comments:

* some functions are prefixed with dla_, others use nvdla_. It seems
arbitrary to me. Please use nvdla_ consistently throughout the source code.

* For reporting errors, please use drm_err(), drm_warn(), etc. I
suggest to rearrange the error messages to not be located in the innermost functions.

* Could you please split this patch into smaller pieces? It currently
hits size limits of some mailing lists. Maybe add the register constants separately.

Please find more review comments below. It's not a full review, but at least something to start with.

Best regards
Thomas

Am 19.04.22 um 15:59 schrieb Cai Huoqing:

Quoted 99 lines

The NVIDIA Deep Learning Accelerator (NVDLA) is an open source IP which is integrated into NVIDIA Jetson AGX Xavier,
so add driver support for this accelerator.

---
  drivers/gpu/drm/Kconfig                 |    2 +
  drivers/gpu/drm/Makefile                |    1 +
  drivers/gpu/drm/nvdla/Kconfig           |    8 +
  drivers/gpu/drm/nvdla/Makefile          |   19 +
  drivers/gpu/drm/nvdla/nvdla_bdma.c      |  200 +
  drivers/gpu/drm/nvdla/nvdla_cache.c     |  215 +
  drivers/gpu/drm/nvdla/nvdla_cdp.c       |  300 ++
  drivers/gpu/drm/nvdla/nvdla_common.c    |  295 ++
  drivers/gpu/drm/nvdla/nvdla_common.h    |  835 +++
  drivers/gpu/drm/nvdla/nvdla_conv.c      |  683 +++
  drivers/gpu/drm/nvdla/nvdla_drm.c       |  695 +++
  drivers/gpu/drm/nvdla/nvdla_drm.h       |  127 +
  drivers/gpu/drm/nvdla/nvdla_engine.c    |  233 +
  drivers/gpu/drm/nvdla/nvdla_engine.h    |  272 +
  drivers/gpu/drm/nvdla/nvdla_gem.c       |  393 ++
  drivers/gpu/drm/nvdla/nvdla_ioctl.h     |   99 +
  drivers/gpu/drm/nvdla/nvdla_pdp.c       |  446 ++
  drivers/gpu/drm/nvdla/nvdla_reg.h       | 6411 +++++++++++++++++++++++
  drivers/gpu/drm/nvdla/nvdla_rubik.c     |  217 +
  drivers/gpu/drm/nvdla/nvdla_sched.h     |   52 +
  drivers/gpu/drm/nvdla/nvdla_scheduler.c | 1005 ++++
  drivers/gpu/drm/nvdla/nvdla_sdp.c       |  728 +++
  22 files changed, 13236 insertions(+)

create mode 100644 drivers/gpu/drm/nvdla/Kconfig create mode 100644 drivers/gpu/drm/nvdla/Makefile create mode 100644 drivers/gpu/drm/nvdla/nvdla_bdma.c create mode 100644 drivers/gpu/drm/nvdla/nvdla_cache.c create mode 100644 drivers/gpu/drm/nvdla/nvdla_cdp.c create mode 100644 drivers/gpu/drm/nvdla/nvdla_common.c create mode 100644 drivers/gpu/drm/nvdla/nvdla_common.h create mode 100644 drivers/gpu/drm/nvdla/nvdla_conv.c create mode 100644 drivers/gpu/drm/nvdla/nvdla_drm.c create mode 100644 drivers/gpu/drm/nvdla/nvdla_drm.h create mode 100644 drivers/gpu/drm/nvdla/nvdla_engine.c create mode 100644 drivers/gpu/drm/nvdla/nvdla_engine.h create mode 100644 drivers/gpu/drm/nvdla/nvdla_gem.c create mode 100644 drivers/gpu/drm/nvdla/nvdla_ioctl.h create mode 100644 drivers/gpu/drm/nvdla/nvdla_pdp.c create mode 100644 drivers/gpu/drm/nvdla/nvdla_reg.h create mode 100644 drivers/gpu/drm/nvdla/nvdla_rubik.c create mode 100644 drivers/gpu/drm/nvdla/nvdla_sched.h create mode 100644 drivers/gpu/drm/nvdla/nvdla_scheduler.c create mode 100644 drivers/gpu/drm/nvdla/nvdla_sdp.c

drivers/gpu/drm/Kconfig +2 -0
index 5133c3f028ab..a55cff374abd 100644
@@ -409,6 +409,8 @@ source "drivers/gpu/drm/solomon/Kconfig"
 
  source "drivers/gpu/drm/sprd/Kconfig"
 
+source "drivers/gpu/drm/nvdla/Kconfig"
+
  config DRM_HYPERV
  	tristate "DRM Support for Hyper-V synthetic video device"
  	depends on DRM && PCI && MMU && HYPERV
drivers/gpu/drm/Makefile +1 -0
index c2ef5f9fce54..8fa3537f308a 100644
@@ -134,3 +134,4 @@ obj-y			+= gud/
  obj-$(CONFIG_DRM_HYPERV) += hyperv/
  obj-y			+= solomon/
  obj-$(CONFIG_DRM_SPRD) += sprd/
+obj-$(CONFIG_DRM_NVDLA) += nvdla/
drivers/gpu/drm/nvdla/Kconfig new file +8 -0
new file mode 100644
index 000000000000..11c04f5da877
@@ -0,0 +1,8 @@
+# SPDX-License-Identifier: GPL-2.0-only
+config DRM_NVDLA
+	tristate "NVDLA DRM"
+	depends on DRM
+	select DRM_GEM_CMA_HELPER
+	help
+	  Choose this option for open-source NVIDIA DLA support.
+	  If M is selected the module will be called nvdla-drm.
drivers/gpu/drm/nvdla/Makefile new file +7 -0
new file mode 100644
index 000000000000..74f37d258f8d
@@ -0,0 +1,19 @@
+
+# SPDX-License-Identifier: GPL-2.0
+nvdla-drm-y := \
+	nvdla_drm.o \
+	nvdla_gem.o \
+	nvdla_scheduler.o \
+	nvdla_engine.o \

drivers/gpu/drm/nvdla/Makefile line 17

Quoted 10 lines
drivers/gpu/drm/nvdla/Makefile new file +10 -0
+	nvdla_bdma.o \
+	nvdla_conv.o \
+	nvdla_sdp.o \
+	nvdla_cdp.o \
+	nvdla_pdp.o \
+	nvdla_rubik.o \
+	nvdla_cache.o \
+	nvdla_common.o \
+	nvdla_engine_data.o \
+	nvdla_engine_debug.o \

File names should be sorted alphabetically here.

drivers/gpu/drm/nvdla/nvdla_bdma.c

Quoted 6 lines
drivers/gpu/drm/nvdla/Makefile new file +2 -0
+
+obj-$(CONFIG_DRM_NVDLA) += nvdla-drm.o
drivers/gpu/drm/nvdla/nvdla_bdma.c new file +0 -1
new file mode 100644
index 000000000000..225613f27acf
--- /dev/null

}

Quoted 258 lines
drivers/gpu/drm/nvdla/nvdla_drm.c new file +252 -0
new file mode 100644
index 000000000000..9217eee1de3b
@@ -0,0 +1,695 @@
+// SPDX-License-Identifier: BSD-3-Clause OR GPL-2.0
+/*
+ * Copyright (C) 2017-2018 NVIDIA CORPORATION
+ * Copyright (C) 2022 Cai Huoqing
+ */
+
+#include <linux/dma-buf.h>
+#include <linux/dma-mapping.h>
+#include <linux/fs.h>
+#include <linux/interrupt.h>
+#include <linux/irq.h>
+#include <linux/irqdomain.h>
+#include <linux/module.h>
+#include <linux/of.h>
+#include <linux/of_device.h>
+#include <linux/of_irq.h>
+#include <linux/of_platform.h>
+#include <linux/platform_device.h>
+#include <linux/printk.h>
+#include <linux/slab.h>
+#include <linux/spinlock.h>
+#include <linux/time.h>
+#include <linux/uaccess.h>
+#include <linux/types.h>
+
+#include "nvdla_drm.h"
+#include "nvdla_ioctl.h"
+#include "nvdla_engine.h"
+
+static struct nvdla_config nvdla_config_os_initial = {
+	.atom_size = 32,
+	.bdma_enable = true,
+	.rubik_enable = true,
+	.weight_compress_support = true,
+};
+
+static struct nvdla_config nvdla_config_small = {
+	//.atom_size = 8,
+	.atom_size = 32,  // nv_large config
+	.bdma_enable = false,
+	.rubik_enable = false,
+	.weight_compress_support = false,
+};
+
+int64_t dla_get_time_us(void)
+{
+	return ktime_get_ns() / NSEC_PER_USEC;
+}
+
+void dla_reg_write(void *driver_context, uint32_t addr, uint32_t reg)
+{
+	struct nvdla_device *nvdla_dev =
+			(struct nvdla_device *)driver_context;
+
+	if (!nvdla_dev)
+		return;
+
+	writel(reg, nvdla_dev->base + addr);
+}
+
+uint32_t dla_reg_read(void *driver_context, uint32_t addr)
+{
+	struct nvdla_device *nvdla_dev =
+			(struct nvdla_device *)driver_context;
+
+	if (!nvdla_dev)
+		return 0;
+
+	return readl(nvdla_dev->base + addr);
+}
+
+static irqreturn_t nvdla_engine_isr(int32_t irq, void *data)
+{
+	unsigned long flags;
+	uint32_t mask;
+	uint32_t reg;
+	struct dla_processor *processor = NULL;
+	struct dla_processor_group *group;
+	struct dla_engine *engine;
+	struct nvdla_device *nvdla_dev = (struct nvdla_device *)data;
+
+	if (!nvdla_dev)
+		return IRQ_NONE;
+
+	engine = nvdla_dev->engine_context;
+	spin_lock_irqsave(&nvdla_dev->nvdla_lock, flags);
+
+	mask = glb_reg_read(engine, S_INTR_MASK);
+	reg = glb_reg_read(engine, S_INTR_STATUS);
+
+	if (reg & MASK(GLB_S_INTR_STATUS_0, CACC_DONE_STATUS0)) {
+		processor = &engine->processors[DLA_OP_CONV];
+		group = &processor->groups[0];
+		group->events |= (1 << DLA_EVENT_OP_COMPLETED);
+	}
+	if (reg & MASK(GLB_S_INTR_STATUS_0, CACC_DONE_STATUS1)) {
+		processor = &engine->processors[DLA_OP_CONV];
+		group = &processor->groups[1];
+		group->events |= (1 << DLA_EVENT_OP_COMPLETED);
+	}
+	if (reg & MASK(GLB_S_INTR_STATUS_0, SDP_DONE_STATUS0)) {
+		processor = &engine->processors[DLA_OP_SDP];
+		group = &processor->groups[0];
+		group->events |= (1 << DLA_EVENT_OP_COMPLETED);
+	}
+	if (reg & MASK(GLB_S_INTR_STATUS_0, SDP_DONE_STATUS1)) {
+		processor = &engine->processors[DLA_OP_SDP];
+		group = &processor->groups[1];
+		group->events |= (1 << DLA_EVENT_OP_COMPLETED);
+	}
+	if (reg & MASK(GLB_S_INTR_STATUS_0, CDP_DONE_STATUS0)) {
+		processor = &engine->processors[DLA_OP_CDP];
+		group = &processor->groups[0];
+		group->events |= (1 << DLA_EVENT_OP_COMPLETED);
+	}
+	if (reg & MASK(GLB_S_INTR_STATUS_0, CDP_DONE_STATUS1)) {
+		processor = &engine->processors[DLA_OP_CDP];
+		group = &processor->groups[1];
+		group->events |= (1 << DLA_EVENT_OP_COMPLETED);
+	}
+	if (reg & MASK(GLB_S_INTR_STATUS_0, RUBIK_DONE_STATUS0)) {
+		processor = &engine->processors[DLA_OP_RUBIK];
+		group = &processor->groups[0];
+		group->events |= (1 << DLA_EVENT_OP_COMPLETED);
+	}
+	if (reg & MASK(GLB_S_INTR_STATUS_0, RUBIK_DONE_STATUS1)) {
+		processor = &engine->processors[DLA_OP_RUBIK];
+		group = &processor->groups[1];
+		group->events |= (1 << DLA_EVENT_OP_COMPLETED);
+	}
+	if (reg & MASK(GLB_S_INTR_STATUS_0, PDP_DONE_STATUS0)) {
+		processor = &engine->processors[DLA_OP_PDP];
+		group = &processor->groups[0];
+		group->events |= (1 << DLA_EVENT_OP_COMPLETED);
+	}
+	if (reg & MASK(GLB_S_INTR_STATUS_0, PDP_DONE_STATUS1)) {
+		processor = &engine->processors[DLA_OP_PDP];
+		group = &processor->groups[1];
+		group->events |= (1 << DLA_EVENT_OP_COMPLETED);
+	}
+	if (reg & MASK(GLB_S_INTR_STATUS_0, BDMA_DONE_STATUS0)) {
+		processor = &engine->processors[DLA_OP_BDMA];
+		group = &processor->groups[0];
+		group->events |= (1 << DLA_EVENT_OP_COMPLETED);
+	}
+	if (reg & MASK(GLB_S_INTR_STATUS_0, BDMA_DONE_STATUS1)) {
+		processor = &engine->processors[DLA_OP_BDMA];
+		group = &processor->groups[1];
+		group->events |= (1 << DLA_EVENT_OP_COMPLETED);
+	}
+	if (reg & MASK(GLB_S_INTR_STATUS_0, CDMA_DAT_DONE_STATUS0)) {
+		processor = &engine->processors[DLA_OP_CONV];
+		group = &processor->groups[0];
+		group->events |= (1 << DLA_EVENT_CDMA_DT_DONE);
+	}
+	if (reg & MASK(GLB_S_INTR_STATUS_0, CDMA_DAT_DONE_STATUS1)) {
+		processor = &engine->processors[DLA_OP_CONV];
+		group = &processor->groups[1];
+		group->events |= (1 << DLA_EVENT_CDMA_DT_DONE);
+	}
+	if (reg & MASK(GLB_S_INTR_STATUS_0, CDMA_WT_DONE_STATUS0)) {
+		processor = &engine->processors[DLA_OP_CONV];
+		group = &processor->groups[0];
+		group->events |= (1 << DLA_EVENT_CDMA_WT_DONE);
+	}
+	if (reg & MASK(GLB_S_INTR_STATUS_0, CDMA_WT_DONE_STATUS1)) {
+		processor = &engine->processors[DLA_OP_CONV];
+		group = &processor->groups[1];
+		group->events |= (1 << DLA_EVENT_CDMA_WT_DONE);
+	}
+
+	glb_reg_write(engine, S_INTR_STATUS, reg);
+	mask = glb_reg_read(engine, S_INTR_MASK);
+	reg = glb_reg_read(engine, S_INTR_STATUS);
+
+	complete(&nvdla_dev->event_notifier);
+	spin_unlock_irqrestore(&nvdla_dev->nvdla_lock, flags);
+
+	return IRQ_HANDLED;
+}
+
+static int32_t dla_read_dma_address(void *driver_context, void *task_data,
+						int16_t index, void *dst)
+{
+	int32_t ret = 0;
+	struct nvdla_mem_handle *handles;
+	dma_addr_t *phys_addr = (dma_addr_t *)(dst);
+	struct nvdla_device *nvdla_dev =
+			(struct nvdla_device *)driver_context;
+	struct nvdla_task *task = (struct nvdla_task *)task_data;
+
+	if (index == -1 || index > task->num_addresses)
+		return -EINVAL;
+
+	handles = (struct nvdla_mem_handle *)task->address_list;
+	ret = nvdla_gem_dma_addr(nvdla_dev->drm, task->file,
+					handles[index].handle,
+					phys_addr);
+
+	/* Add offset to IOVA address */
+	*phys_addr = *phys_addr + handles[index].offset;
+
+	return ret;
+}
+
+static int32_t dla_read_cpu_address(void *driver_context, void *task_data,
+						int16_t index, void *dst)
+{
+	uint64_t *temp = (uint64_t *)dst;
+	struct nvdla_task *task = (struct nvdla_task *)task_data;
+
+	if (index == -1 || index > task->num_addresses)
+		return -EINVAL;
+
+	*temp = (uint64_t)index;
+	return 0;
+}
+
+int32_t dla_get_dma_address(void *driver_context, void *task_data,
+					int16_t index, void *dst_ptr,
+					uint32_t destination)
+{
+	int32_t ret = 0;
+
+	if (destination == DESTINATION_PROCESSOR) {
+		ret = dla_read_cpu_address(driver_context, task_data,
+						index, dst_ptr);
+	} else if (destination == DESTINATION_DMA) {
+		ret = dla_read_dma_address(driver_context, task_data,
+						index, dst_ptr);
+	} else {
+		ret = -EINVAL;
+	}
+
+	return ret;
+}
+
+int32_t dla_data_write(void *driver_context, void *task_data,
+				void *src, uint64_t dst,
+				uint32_t size, uint64_t offset)
+{
+	int32_t ret;
+	void *ptr = NULL;
+	struct dma_buf *buf;
+	struct iosys_map map;
+	struct nvdla_mem_handle *handles;
+	struct nvdla_task *task = (struct nvdla_task *)task_data;
+
+	handles = task->address_list;
+	buf = dma_buf_get(handles[dst].handle);
+	if (IS_ERR(buf)) {
+		pr_err("%s: Failed get dma_buf for handle=%d\n", __func__,

drivers/gpu/drm/nvdla/nvdla_drm.c line 262

Quoted 10 lines
drivers/gpu/drm/nvdla/nvdla_drm.c new file +10 -0
+						handles[dst].handle);
+		return -EFAULT;
+	}
+
+	ret = dma_buf_begin_cpu_access(buf, DMA_BIDIRECTIONAL);
+	if (ret)
+		goto put_dma_buf;
+
+	ret = dma_buf_vmap(buf, &map);
+	ptr = ret ? NULL : map.vaddr;

Never extract the pointer's address without good reason. You don't know if this points to a location in I/O memory.

drivers/gpu/drm/nvdla/nvdla_drm.c line 263

Quoted 1 line
drivers/gpu/drm/nvdla/nvdla_drm.c new file +1 -0
+	if (!ptr) {

Simply test for ret here.

drivers/gpu/drm/nvdla/nvdla_drm.c line 266

Quoted 3 lines
drivers/gpu/drm/nvdla/nvdla_drm.c new file +3 -0
+		pr_err("%s: Failed to vmap dma_buf for handle=%d\n", __func__,
+						handles[dst].handle);
+		ret = -ENOMEM;

You already got an errno code. Don't override it.

drivers/gpu/drm/nvdla/nvdla_drm.c line 271

Quoted 5 lines
drivers/gpu/drm/nvdla/nvdla_drm.c new file +5 -0
+		goto end_cpu_access;
+	}
+
+
+	memcpy((void *)((uint8_t *)ptr + offset), src, size);

Use iosys_map_memcpy_to() here. It does the right thing

drivers/gpu/drm/nvdla/nvdla_drm.c line 273

Quoted 2 lines
drivers/gpu/drm/nvdla/nvdla_drm.c new file +2 -0
+
+	dma_buf_vunmap(buf, ptr);

You have to pass map as the second argument.

Quoted 32 lines
drivers/gpu/drm/nvdla/nvdla_drm.c new file +32 -0
+
+end_cpu_access:
+	dma_buf_end_cpu_access(buf, DMA_BIDIRECTIONAL);
+
+put_dma_buf:
+	dma_buf_put(buf);
+
+	return ret;
+}
+
+int32_t dla_data_read(void *driver_context, void *task_data,
+				uint64_t src, void *dst,
+				uint32_t size, uint64_t offset)
+{
+	int32_t ret;
+	void *ptr = NULL;
+	struct dma_buf *buf;
+	struct iosys_map map;
+	struct nvdla_mem_handle *handles;
+	struct nvdla_task *task = (struct nvdla_task *)task_data;
+
+	handles = task->address_list;
+
+	buf = dma_buf_get(handles[src].handle);
+	if (IS_ERR(buf)) {
+		pr_err("%s: Failed get dma_buf for handle=%d\n", __func__,
+						handles[src].handle);
+		return -EFAULT;
+	}
+
+	ret = dma_buf_begin_cpu_access(buf, DMA_BIDIRECTIONAL);
+	if (ret)

drivers/gpu/drm/nvdla/nvdla_drm.c line 315

Quoted 10 lines
drivers/gpu/drm/nvdla/nvdla_drm.c new file +10 -0
+		goto put_dma_buf;
+
+	ret = dma_buf_vmap(buf, &map);
+	ptr = ret ? NULL : map.vaddr;
+	if (!ptr) {
+		pr_err("%s: Failed to vmap dma_buf for handle=%d\n", __func__,
+						handles[src].handle);
+		ret = -ENOMEM;
+		goto end_cpu_access;
+	}

All the same problems as in dla_data_write().

drivers/gpu/drm/nvdla/nvdla_drm.c line 317

Quoted 2 lines
drivers/gpu/drm/nvdla/nvdla_drm.c new file +2 -0
+
+	memcpy(dst, (void *)(((uint8_t *)ptr) + offset), size);

Use iosys_map_memcpy_from() here.

drivers/gpu/drm/nvdla/nvdla_drm.c line 319

Quoted 2 lines
drivers/gpu/drm/nvdla/nvdla_drm.c new file +2 -0
+
+	dma_buf_vunmap(buf, ptr);

'map' instead of 'ptr'

Quoted 56 lines
drivers/gpu/drm/nvdla/nvdla_drm.c new file +55 -0
+
+end_cpu_access:
+	dma_buf_end_cpu_access(buf, DMA_BIDIRECTIONAL);
+
+put_dma_buf:
+	dma_buf_put(buf);
+
+	return ret;
+}
+
 
+
+static int32_t nvdla_submit(struct drm_device *drm, void *arg,
+					struct drm_file *file)
+{
+	int32_t err = 0;
+	struct nvdla_task *task;
+	struct nvdla_ioctl_submit_task local_task;
+	struct nvdla_ioctl_submit_task __user *user_task;
+	struct nvdla_device *nvdla_dev = dev_get_drvdata(drm->dev);
+	struct nvdla_submit_args *args =
+			(struct nvdla_submit_args *)arg;
+
+	user_task = (struct nvdla_ioctl_submit_task __user *)
+			(uintptr_t)args->tasks;
+	if (!user_task)
+		return -EINVAL;
+
+	/* IOCTL copy descriptors */
+	if (copy_from_user(&local_task, (void __user *)user_task,
+			(sizeof(*user_task))))
+		return -EFAULT;
+
+	task = kzalloc(sizeof(*task), GFP_KERNEL);
+	if (task == NULL)
+		return -EFAULT;
+
+	nvdla_dev->task = task;
+	kref_init(&task->ref);
+	task->nvdla_dev = nvdla_dev;
+	task->file = file;
+
+	/* update task desc fields */
+	err = nvdla_fill_task_desc(&local_task, task);
+	if (err)
+		goto free_task_desc;
+
+	err = nvdla_task_submit(nvdla_dev, task);
+
+	kfree(task->address_list);
+
+free_task_desc:
+	kfree(task);
+	return err;
+}
+

drivers/gpu/drm/nvdla/nvdla_drm.c

Quoted 10 lines
drivers/gpu/drm/nvdla/nvdla_drm.c new file +10 -0
+static int32_t nvdla_gem_alloc(struct nvdla_gem_object *nobj)
+{
+	struct drm_gem_object *dobj = &nobj->object;
+	struct drm_device *drm = dobj->dev;
+
+	nobj->dma_attrs = DMA_ATTR_WRITE_COMBINE;
+
+	nobj->kvaddr = dma_alloc_attrs(drm->dev, dobj->size, &nobj->dma_addr,
+						GFP_KERNEL, nobj->dma_attrs);
+

Store an iosys-map address in nobj and initialize it with iosys_map_set_vaddr(); or iosys_map_set_vaddr_iomem() if you're working with I/O memory.

Quoted 75 lines
drivers/gpu/drm/nvdla/nvdla_drm.c new file +75 -0
+	if (!nobj->kvaddr)
+		return -ENOMEM;
+
+	return 0;
+}
+
+static void nvdla_gem_free(struct nvdla_gem_object *nobj)
+{
+	struct drm_gem_object *dobj = &nobj->object;
+	struct drm_device *drm = dobj->dev;
+
+	dma_free_attrs(drm->dev, dobj->size, nobj->kvaddr, nobj->dma_addr,
+				nobj->dma_attrs);
+}
+
+static void nvdla_gem_free_object(struct drm_gem_object *dobj)
+{
+	struct nvdla_gem_object *nobj;
+
+	drm_gem_free_mmap_offset(dobj);
+
+	nobj = to_nvdla_obj(dobj);
+
+	nvdla_gem_free(nobj);
+
+	kfree(nobj);
+}
+
+static struct nvdla_gem_object *
+nvdla_gem_create_object(struct drm_device *drm, uint32_t size)
+{
+	int32_t ret;
+	struct drm_gem_object *dobj;
+	struct nvdla_gem_object *nobj;
+
+	size = round_up(size, PAGE_SIZE);
+
+	nobj = kzalloc(sizeof(*nobj), GFP_KERNEL);
+	if (!nobj)
+		return ERR_PTR(-ENOMEM);
+
+	dobj = &nobj->object;
+
+	drm_gem_private_object_init(drm, dobj, size);
+
+	ret = nvdla_gem_alloc(nobj);
+	if (ret)
+		goto free_nvdla_obj;
+
+	return nobj;
+
+free_nvdla_obj:
+	kfree(nobj);
+	return ERR_PTR(ret);
+}
+
+static struct sg_table*
+nvdla_drm_gem_prime_get_sg_table(struct drm_gem_object *dobj)
+{
+	int32_t ret;
+	struct sg_table *sgt;
+	struct drm_device *drm = dobj->dev;
+	struct nvdla_gem_object *nobj = to_nvdla_obj(dobj);
+
+	sgt = kzalloc(sizeof(*sgt), GFP_KERNEL);
+	if (!sgt)
+		return ERR_PTR(-ENOMEM);
+
+	ret = dma_get_sgtable_attrs(drm->dev, sgt, nobj->kvaddr,
+				    nobj->dma_addr, dobj->size,
+				    nobj->dma_attrs);
+	if (ret) {
+		DRM_ERROR("failed to allocate sgt, %d\n", ret);
+		kfree(sgt);
+		return ERR_PTR(ret);

drivers/gpu/drm/nvdla/nvdla_drm.c

Quoted 10 lines
drivers/gpu/drm/nvdla/nvdla_drm.c new file +10 -0
+	}
+
+	return sgt;
+}
+
+static int nvdla_drm_gem_prime_vmap(struct drm_gem_object *obj, struct iosys_map *map)
+{
+	struct nvdla_gem_object *nobj = to_nvdla_obj(obj);
+
+	map->vaddr = nobj->kvaddr;

Instead of kvaddr, store the pointer as struct iosys_map. Then simply copy it here, as in

   *map = nobj->map;
Quoted 8 lines
drivers/gpu/drm/nvdla/nvdla_drm.c new file +8 -0
+
+	return 0;
+}
+
+static void nvdla_drm_gem_prime_vunmap(struct drm_gem_object *obj, struct iosys_map *map)
+{
+	/* Nothing to do */
+}

drivers/gpu/drm/nvdla/nvdla_drm.c

Quoted 10 lines
drivers/gpu/drm/nvdla/nvdla_drm.c new file +10 -0
+
+static int32_t nvdla_drm_gem_object_mmap(struct drm_gem_object *dobj,
+					struct vm_area_struct *vma)
+{
+	int32_t ret;
+	struct nvdla_gem_object *nobj = to_nvdla_obj(dobj);
+	struct drm_device *drm = dobj->dev;
+
+	vma->vm_flags &= ~VM_PFNMAP;
+	vma->vm_pgoff = 0;

It's cleaner to do this as

vma->vm_pgoff -= drm_vma_node_start(&obj->vma_node)

Quoted 74 lines
drivers/gpu/drm/nvdla/nvdla_drm.c new file +74 -0
+
+	ret = dma_mmap_attrs(drm->dev, vma, nobj->kvaddr, nobj->dma_addr,
+			     dobj->size, nobj->dma_attrs);
+	if (ret)
+		drm_gem_vm_close(vma);
+
+	return ret;
+}
+
+static const struct drm_gem_object_funcs nvdla_gem_object_funcs = {
+	.free = nvdla_gem_free_object,
+	.get_sg_table = nvdla_drm_gem_prime_get_sg_table,
+	.vmap = nvdla_drm_gem_prime_vmap,
+	.vunmap = nvdla_drm_gem_prime_vunmap,
+	.mmap = nvdla_drm_gem_object_mmap,
+};
+
+static struct nvdla_gem_object*
+nvdla_gem_create_with_handle(struct drm_file *file_priv,
+							 struct drm_device *drm, uint32_t size,
+							 uint32_t *handle)
+{
+	int32_t ret;
+	struct drm_gem_object *dobj;
+	struct nvdla_gem_object *nobj;
+
+	nobj = nvdla_gem_create_object(drm, size);
+	if (IS_ERR(nobj))
+		return ERR_CAST(nobj);
+
+	dobj = &nobj->object;
+	dobj->funcs = &nvdla_gem_object_funcs;
+	ret = drm_gem_handle_create(file_priv, dobj, handle);
+	if (ret)
+		goto free_drm_object;
+
+	drm_gem_object_put(dobj);
+
+	return nobj;
+
+free_drm_object:
+	nvdla_gem_free_object(dobj);
+
+	return ERR_PTR(ret);
+}
+
+static int32_t nvdla_gem_create(struct drm_device *drm, void *data,
+								struct drm_file *file)
+{
+	struct nvdla_gem_object *nobj;
+	struct nvdla_gem_create_args *args = data;
+
+	nobj = nvdla_gem_create_with_handle(file, drm, args->size,
+					 &args->handle);
+	if (IS_ERR(nobj))
+		return PTR_ERR(nobj);
+
+	return 0;
+}
+
+static int32_t nvdla_drm_gem_mmap_buf(struct drm_gem_object *obj,
+									  struct vm_area_struct *vma)
+{
+	int32_t ret;
+
+	ret = drm_gem_mmap_obj(obj, obj->size, vma);
+	if (ret)
+		return ret;
+
+	return nvdla_drm_gem_object_mmap(obj, vma);
+}
+
+static int32_t nvdla_drm_gem_mmap(struct file *filp, struct vm_area_struct *vma)
+{

drivers/gpu/drm/nvdla/nvdla_drm.c

Quoted 10 lines
drivers/gpu/drm/nvdla/nvdla_drm.c new file +10 -0
+	int32_t ret;
+	struct drm_gem_object *obj;
+
+	ret = drm_gem_mmap(filp, vma);
+	if (ret)
+		return ret;
+
+	obj = vma->vm_private_data;
+
+	return nvdla_drm_gem_object_mmap(obj, vma);

I don't understand these two lines. This is part of what drm_gem_mmap() does. It shouldn't be necessary here.

Quoted 32 lines
drivers/gpu/drm/nvdla/nvdla_drm.c new file +32 -0
+}
+
+int32_t nvdla_gem_dma_addr(struct drm_device *dev, struct drm_file *file,
+						   uint32_t fd, dma_addr_t *addr)
+{
+	int32_t ret;
+	uint32_t handle;
+	struct nvdla_gem_object *nobj;
+	struct drm_gem_object *dobj;
+
+	ret = drm_gem_prime_fd_to_handle(dev, file, fd, &handle);
+	if (ret)
+		return ret;
+
+	dobj = drm_gem_object_lookup(file, handle);
+	if (!dobj)
+		return -EINVAL;
+
+	nobj = to_nvdla_obj(dobj);
+
+	*addr = nobj->dma_addr;
+
+	drm_gem_object_put(dobj);
+
+	return 0;
+}
+
+static int32_t nvdla_gem_map_offset(struct drm_device *drm, void *data,
+									struct drm_file *file)
+{
+	struct nvdla_gem_map_offset_args *args = data;
+

drivers/gpu/drm/nvdla/nvdla_drm.c

Quoted 10 lines
drivers/gpu/drm/nvdla/nvdla_drm.c new file +10 -0
+	return drm_gem_dumb_map_offset(file, drm, args->handle,
+								   &args->offset);
+}
+
+static const struct file_operations nvdla_drm_fops = {
+	.owner = THIS_MODULE,
+	.open = drm_open,
+	.release = drm_release,
+	.unlocked_ioctl = drm_ioctl,
+	.mmap = nvdla_drm_gem_mmap,

It should be fine to use drm_gem_mmap here. Then you should use DEFINE_DRM_GEM_FOPS() to define nvdla_drm_fops.

Quoted 12 lines
drivers/gpu/drm/nvdla/nvdla_drm.c new file +12 -0
+	.poll = drm_poll,
+	.read = drm_read,
+#ifdef CONFIG_COMPAT
+	.compat_ioctl = drm_compat_ioctl,
+#endif
+	.llseek = noop_llseek,
+};
+
+static const struct drm_ioctl_desc nvdla_drm_ioctls[] = {
+	DRM_IOCTL_DEF_DRV(NVDLA_SUBMIT, nvdla_submit, DRM_RENDER_ALLOW),
+	DRM_IOCTL_DEF_DRV(NVDLA_GEM_CREATE, nvdla_gem_create, DRM_RENDER_ALLOW),
+	DRM_IOCTL_DEF_DRV(NVDLA_GEM_MMAP, nvdla_gem_map_offset, DRM_RENDER_ALLOW),

drivers/gpu/drm/nvdla/nvdla_drm.c

Quoted 10 lines
drivers/gpu/drm/nvdla/nvdla_drm.c new file +10 -0
+	/* use DRM_IOCTL_MODE_DESTROY_DUMB to destory */
+};
+
+static struct drm_driver nvdla_drm_driver = {
+	.driver_features = DRIVER_GEM | DRIVER_RENDER,
+
+	.ioctls = nvdla_drm_ioctls,
+	.num_ioctls = ARRAY_SIZE(nvdla_drm_ioctls),
+	.fops = &nvdla_drm_fops,
+	.gem_prime_mmap		= nvdla_drm_gem_mmap_buf,

Use drm_gem_prime_mmap() here.

Some context: the situation with these mmap functions has been confusing and inconsistent among DRM drivers. But we cleaned it up so that you only have to provide a minimal implementation of struct drm_gem_object_funcs.mmap. All other mmap callbacks can then be filled with standard DRM helpers.

Quoted 37 lines
drivers/gpu/drm/nvdla/nvdla_drm.c new file +37 -0
+
+	.name = "nvdla",
+	.desc = "NVDLA driver",
+	.date = "20171017",
+	.major = 0,
+	.minor = 0,
+	.patchlevel = 0,
+};
+
+int32_t nvdla_drm_probe(struct nvdla_device *nvdla_dev)
+{
+	int32_t err;
+	struct drm_device *drm;
+	struct drm_driver *driver = &nvdla_drm_driver;
+
+	drm = drm_dev_alloc(driver, &nvdla_dev->pdev->dev);
+	if (IS_ERR(drm))
+		return PTR_ERR(drm);
+
+	nvdla_dev->drm = drm;
+
+	err = drm_dev_register(drm, 0);
+	if (err < 0)
+		goto unref;
+
+	return 0;
+
+unref:
+	drm_dev_put(drm);
+	return err;
+}
+
+void nvdla_drm_remove(struct nvdla_device *nvdla_dev)
+{
+	drm_dev_unregister(nvdla_dev->drm);
+	drm_dev_put(nvdla_dev->drm);
+}
Signature · 6 lines
Thomas Zimmermann
Graphics Driver Developer
SUSE Software Solutions Germany GmbH
Maxfeldstr. 5, 90409 Nürnberg, Germany
(HRB 36809, AG Nürnberg)
Geschäftsführer: Ivo Totev
#10 lore raw

Am 21.04.22 um 10:30 schrieb Thomas Zimmermann:

Quoted 9 lines

(Resending, as some MLs didn't like the size of the origninal mail.)

Hi,

thanks for your submission. Some general comments:

  * some functions are prefixed with dla_, others use nvdla_. It seems arbitrary to me. Please use nvdla_ consistently throughout the source code.

  * For reporting errors, please use drm_err(), drm_warn(), etc. I suggest to rearrange the error messages to not be located in the innermost functions.

If you plan to have multiple instances of the driver loaded at the same time, using drm_dev_err(), drm_dev_warn() etc.. would be even better.

BTW: I'm still absolutely not keen to enforcing drm_* log functions. So if you prefer to stick with pr_err() and dev_err() we could discuss that once more.

Regards,
Christian.

Quoted 7 lines

  * Could you please split this patch into smaller pieces? It currently hits size limits of some mailing lists. Maybe add the register constants separately.

Please find more review comments below. It's not a full review, but at least something to start with.

Best regards
Thomas

#11 lore raw

Hi

Am 21.04.22 um 10:34 schrieb Christian König:

Quoted 5 lines

Am 21.04.22 um 10:30 schrieb Thomas Zimmermann:

Quoted 4 lines

(Resending, as some MLs didn't like the size of the origninal mail.)

Hi,

Quoted 7 lines
Quoted 5 lines

thanks for your submission. Some general comments:

  * some functions are prefixed with dla_, others use nvdla_. It seems arbitrary to me. Please use nvdla_ consistently throughout the source code.

  * For reporting errors, please use drm_err(), drm_warn(), etc. I suggest to rearrange the error messages to not be located in the innermost functions.

If you plan to have multiple instances of the driver loaded at the same time, using drm_dev_err(), drm_dev_warn() etc.. would be even better.

I think that's what I mean. Thanks for pointing this out.

Best regards
Thomas

Quoted 14 lines

BTW: I'm still absolutely not keen to enforcing drm_* log functions. So if you prefer to stick with pr_err() and dev_err() we could discuss that once more.

Regards,
Christian.

Quoted 7 lines

  * Could you please split this patch into smaller pieces? It currently hits size limits of some mailing lists. Maybe add the register constants separately.

Please find more review comments below. It's not a full review, but at least something to start with.

Best regards
Thomas

Signature · 6 lines
Thomas Zimmermann
Graphics Driver Developer
SUSE Software Solutions Germany GmbH
Maxfeldstr. 5, 90409 Nürnberg, Germany
(HRB 36809, AG Nürnberg)
Geschäftsführer: Ivo Totev
#12 lore raw
In reply to Christian König

Am 21.04.22 um 10:34 schrieb Christian König:

Quoted 7 lines

Am 21.04.22 um 10:30 schrieb Thomas Zimmermann:

Quoted 6 lines

(Resending, as some MLs didn't like the size of the origninal mail.)

Hi,

thanks for your submission. Some general comments:

Quoted 7 lines
Quoted 3 lines

  * some functions are prefixed with dla_, others use nvdla_. It seems arbitrary to me. Please use nvdla_ consistently throughout the source code.

  * For reporting errors, please use drm_err(), drm_warn(), etc. I suggest to rearrange the error messages to not be located in the innermost functions.

If you plan to have multiple instances of the driver loaded at the same time, using drm_dev_err(), drm_dev_warn() etc.. would be even better.

BTW: I'm still absolutely not keen to enforcing drm_* log functions. So if you prefer to stick with pr_err() and dev_err() we could discuss that once more.

I often do 'dmesg | grep drm' to quickly look for errors. Not using drm logging helpers makes this less useful.

Best regards
Thomas

Quoted 12 lines

Regards,
Christian.

Quoted 7 lines

  * Could you please split this patch into smaller pieces? It currently hits size limits of some mailing lists. Maybe add the register constants separately.

Please find more review comments below. It's not a full review, but at least something to start with.

Best regards
Thomas

Signature · 6 lines
Thomas Zimmermann
Graphics Driver Developer
SUSE Software Solutions Germany GmbH
Maxfeldstr. 5, 90409 Nürnberg, Germany
(HRB 36809, AG Nürnberg)
Geschäftsführer: Ivo Totev
#13 lore raw
In reply to Christian König

Hi

Am 21.04.22 um 10:34 schrieb Christian König:

Quoted 5 lines

Am 21.04.22 um 10:30 schrieb Thomas Zimmermann:

Quoted 4 lines

(Resending, as some MLs didn't like the size of the origninal mail.)

Hi,

Quoted 7 lines
Quoted 5 lines

thanks for your submission. Some general comments:

  * some functions are prefixed with dla_, others use nvdla_. It seems arbitrary to me. Please use nvdla_ consistently throughout the source code.

  * For reporting errors, please use drm_err(), drm_warn(), etc. I suggest to rearrange the error messages to not be located in the innermost functions.

If you plan to have multiple instances of the driver loaded at the same time, using drm_dev_err(), drm_dev_warn() etc.. would be even better.

I thought that these functions exist, but looking for them now I cannot find them. The macros DRM_DEV_ERR(), etc are deprecated.

Quoted 14 lines

BTW: I'm still absolutely not keen to enforcing drm_* log functions. So if you prefer to stick with pr_err() and dev_err() we could discuss that once more.

Regards,
Christian.

Quoted 7 lines

  * Could you please split this patch into smaller pieces? It currently hits size limits of some mailing lists. Maybe add the register constants separately.

Please find more review comments below. It's not a full review, but at least something to start with.

Best regards
Thomas

Signature · 6 lines
Thomas Zimmermann
Graphics Driver Developer
SUSE Software Solutions Germany GmbH
Maxfeldstr. 5, 90409 Nürnberg, Germany
(HRB 36809, AG Nürnberg)
Geschäftsführer: Ivo Totev
#14 lore raw

Am 21.04.22 um 11:13 schrieb Thomas Zimmermann:

Quoted 10 lines

Hi

Am 21.04.22 um 10:34 schrieb Christian König:

Quoted 7 lines

Am 21.04.22 um 10:30 schrieb Thomas Zimmermann:

Quoted 6 lines

(Resending, as some MLs didn't like the size of the origninal mail.)

Hi,

thanks for your submission. Some general comments:

Quoted 7 lines
Quoted 5 lines
Quoted 3 lines

  * some functions are prefixed with dla_, others use nvdla_. It seems arbitrary to me. Please use nvdla_ consistently throughout the source code.

  * For reporting errors, please use drm_err(), drm_warn(), etc. I suggest to rearrange the error messages to not be located in the innermost functions.

If you plan to have multiple instances of the driver loaded at the same time, using drm_dev_err(), drm_dev_warn() etc.. would be even better.

I thought that these functions exist, but looking for them now I cannot find them. The macros DRM_DEV_ERR(), etc are deprecated.

That's what I meant with the comment below.

I wasn't 100%, but dev_err() etc.. seems to now be the preferred form since that allows filtering for log messages of a certain device.

Regards,
Christian.

Quoted 16 lines
Quoted 15 lines

BTW: I'm still absolutely not keen to enforcing drm_* log functions. So if you prefer to stick with pr_err() and dev_err() we could discuss that once more.

Regards,
Christian.

Quoted 7 lines

  * Could you please split this patch into smaller pieces? It currently hits size limits of some mailing lists. Maybe add the register constants separately.

Please find more review comments below. It's not a full review, but at least something to start with.

Best regards
Thomas