Skip to contentSkip to conversation

imagination: Update allocation flags for buffers in submit job

PATCH Applied Brajesh Gupta 2 participants 4 messages Nov 16 Library code

lore
Share & tools
lore
#0 lore raw

v4 cover imagination: Update allocation flags for buffers in submit job

Update BO allocation flags to match expectation:
- Remove the CPU_ACCESS flag from BOs that are not host-accessed.

---
Changes in v4:
- Drop patch 3 for dropping DRM_PVR_BO_PM_FW_PROTECT for MList from v3.
- Link to v3: https://patch.msgid.link/[email protected]

Changes in v3:
- Fixed a formatting in Patch 1 as suggested by Luigi.
- Added RB tags for patch 1 and patch 2.
- Link to v2: https://patch.msgid.link/[email protected]

Changes in v2:
- Update as per Kamil's feedback.
- Updated description to map commits are specific to Imagination.
- Spilt lib/igt change to two: Interface update and flags update.
- Link to v1: https://patch.msgid.link/[email protected]

---
Brajesh Gupta (2):
lib/igt_pvr: Update general heap allocation interface to pass flags
imagination: Skip CPU access flag for a few Render related allocations

 lib/igt_pvr.c                       | 29 +++++++++++++++++------------
 lib/igt_pvr.h                       |  3 ++-
 tests/imagination/pvr_submit_jobs.c | 13 ++++++++++---
 3 files changed, 29 insertions(+), 16 deletions(-)

---
base-commit: fc75d6d7ffcb79e8a3316e80c7085460568a7670
change-id: 20261029-b4-render_update-79d0cf70eb0e

Best regards,

Signature · 1 line
Brajesh Gupta <[email protected]>
#1 lore raw

PATCH v4 1/2 lib/igt_pvr: Update general heap allocation interface to pass flags

Update igt_pvr_allocate_general() interface to pass flags for allocation.

---
 lib/igt_pvr.c                       | 36 ++++++++++++++++++++++++------------
 lib/igt_pvr.h                       |  3 ++-
 tests/imagination/pvr_submit_jobs.c | 13 ++++++++++---
 3 files changed, 36 insertions(+), 16 deletions(-)
lib/igt_pvr.c +24 -12
index 799852da5..e771d0660 100644
@@ -399,7 +399,9 @@ uint32_t igt_pvr_ioctl_create_context(int fd, enum drm_pvr_ctx_type type,
 		create_context_args.callstack_addr =
 			igt_pvr_get_gpu_addr(igt_pvr_allocate_general(fd,
 								      vm_ctx_handle,
-								      POWERVR_GPU_PAGE_SIZE));
+								      POWERVR_GPU_PAGE_SIZE,
+								      DRM_PVR_BO_ALLOW_CPU_USERSPACE_ACCESS |
+								      DRM_PVR_BO_BYPASS_DEVICE_CACHE));
 		create_context_args.static_context_state = to_user_pointer(init_render_context_stream);
 		create_context_args.static_context_state_len =
 			init_render_context_state_stream(init_render_context_stream);
@@ -673,17 +675,15 @@ struct igt_pvr_allocation *igt_pvr_allocate(int fd, uint32_t vm_ctx, size_t size
  * @fd: The file descriptor of the DRM device.
  * @vm_ctx: The virtual memory context.
  * @size: The size of the allocation.
+ * @flags: Allocation flags.
  *
  * Function to allocate and map GPU memory from the general heap.
  *
  * Returns: Pointer to the allocation structure.
  */
-struct igt_pvr_allocation *igt_pvr_allocate_general(int fd, uint32_t vm_ctx, size_t size)
+struct igt_pvr_allocation *igt_pvr_allocate_general(int fd, uint32_t vm_ctx, size_t size, uint64_t flags)
 {
-	return igt_pvr_allocate(fd, vm_ctx, size,
-				DRM_PVR_BO_ALLOW_CPU_USERSPACE_ACCESS |
-				DRM_PVR_BO_BYPASS_DEVICE_CACHE,
-				DRM_PVR_HEAP_GENERAL);
+	return igt_pvr_allocate(fd, vm_ctx, size, flags, DRM_PVR_HEAP_GENERAL);
 }
 
 /**
@@ -817,11 +817,17 @@ static void igt_pvr_create_hwrt_geom_data_args(int fd, uint32_t vm_ctx,
 					       *geom_data_args)
 {
 	geom_data_args->rtc_dev_addr =
-		igt_pvr_get_gpu_addr(igt_pvr_allocate_general(fd, vm_ctx, POWERVR_GPU_PAGE_SIZE));
+		igt_pvr_get_gpu_addr(igt_pvr_allocate_general(fd, vm_ctx, POWERVR_GPU_PAGE_SIZE,
+							      DRM_PVR_BO_ALLOW_CPU_USERSPACE_ACCESS |
+							      DRM_PVR_BO_BYPASS_DEVICE_CACHE));
 	geom_data_args->tpc_dev_addr =
-		igt_pvr_get_gpu_addr(igt_pvr_allocate_general(fd, vm_ctx, POWERVR_GPU_PAGE_SIZE));
+		igt_pvr_get_gpu_addr(igt_pvr_allocate_general(fd, vm_ctx, POWERVR_GPU_PAGE_SIZE,
+							      DRM_PVR_BO_ALLOW_CPU_USERSPACE_ACCESS |
+							      DRM_PVR_BO_BYPASS_DEVICE_CACHE));
 	geom_data_args->vheap_table_dev_addr =
-		igt_pvr_get_gpu_addr(igt_pvr_allocate_general(fd, vm_ctx, POWERVR_GPU_PAGE_SIZE));
+		igt_pvr_get_gpu_addr(igt_pvr_allocate_general(fd, vm_ctx, POWERVR_GPU_PAGE_SIZE,
+							      DRM_PVR_BO_ALLOW_CPU_USERSPACE_ACCESS |
+							      DRM_PVR_BO_BYPASS_DEVICE_CACHE));
 	geom_data_args->tpc_size = POWERVR_GPU_PAGE_SIZE;
 	geom_data_args->tpc_stride = 0x1;
 }
@@ -839,11 +845,17 @@ static void igt_pvr_create_hwrt_rt_data_args(int fd, uint32_t vm_ctx,
 					     struct drm_pvr_create_hwrt_rt_data_args *rt_data_args)
 {
 	rt_data_args->pm_mlist_dev_addr =
-		igt_pvr_get_gpu_addr(igt_pvr_allocate_general(fd, vm_ctx, POWERVR_GPU_PAGE_SIZE));
+		igt_pvr_get_gpu_addr(igt_pvr_allocate_general(fd, vm_ctx, POWERVR_GPU_PAGE_SIZE,
+							      DRM_PVR_BO_ALLOW_CPU_USERSPACE_ACCESS |
+							      DRM_PVR_BO_BYPASS_DEVICE_CACHE));
 	rt_data_args->macrotile_array_dev_addr =
-		igt_pvr_get_gpu_addr(igt_pvr_allocate_general(fd, vm_ctx, POWERVR_GPU_PAGE_SIZE));
+		igt_pvr_get_gpu_addr(igt_pvr_allocate_general(fd, vm_ctx, POWERVR_GPU_PAGE_SIZE,
+							      DRM_PVR_BO_ALLOW_CPU_USERSPACE_ACCESS |
+							      DRM_PVR_BO_BYPASS_DEVICE_CACHE));
 	rt_data_args->region_header_dev_addr =
-		igt_pvr_get_gpu_addr(igt_pvr_allocate_general(fd, vm_ctx, POWERVR_GPU_PAGE_SIZE));
+		igt_pvr_get_gpu_addr(igt_pvr_allocate_general(fd, vm_ctx, POWERVR_GPU_PAGE_SIZE,
+							      DRM_PVR_BO_ALLOW_CPU_USERSPACE_ACCESS |
+							      DRM_PVR_BO_BYPASS_DEVICE_CACHE));
 }
 
 /**
lib/igt_pvr.h +2 -1
index b6e3dd676..bed02f968 100644
@@ -67,7 +67,8 @@ struct igt_pvr_allocation *igt_pvr_allocate_addr(int fd, uint32_t vm_ctx, size_t
 						 uint64_t flags, uint64_t gpu_addr);
 struct igt_pvr_allocation *igt_pvr_allocate(int fd, uint32_t vm_ctx, size_t size, uint64_t flags,
 					    uint32_t heap_index);
-struct igt_pvr_allocation *igt_pvr_allocate_general(int fd, uint32_t vm_ctx, size_t size);
+struct igt_pvr_allocation *igt_pvr_allocate_general(int fd, uint32_t vm_ctx, size_t size,
+					    uint64_t flags);
 void *igt_pvr_get_cpu_addr(int fd, struct igt_pvr_allocation *alloc);
 uint64_t igt_pvr_get_gpu_addr(struct igt_pvr_allocation *alloc);
 size_t igt_pvr_get_size(struct igt_pvr_allocation *alloc);
tests/imagination/pvr_submit_jobs.c +10 -3
index 566f9d2b2..21ccb03fd 100644
@@ -52,7 +52,10 @@ static uint32_t init_compute_cmd_stream(int fd, uint32_t vm_ctx,
 	void *cdm_stream_map;
 	uint32_t stream_length;
 
-	struct igt_pvr_allocation *cdm_stream_alloc = igt_pvr_allocate_general(fd, vm_ctx, 0x1000);
+	struct igt_pvr_allocation *cdm_stream_alloc =
+		igt_pvr_allocate_general(fd, vm_ctx, 0x1000,
+				         DRM_PVR_BO_ALLOW_CPU_USERSPACE_ACCESS |
+				         DRM_PVR_BO_BYPASS_DEVICE_CACHE);
 
 	cdm_stream_gpu_addr = igt_pvr_get_gpu_addr(cdm_stream_alloc);
 
@@ -212,14 +215,18 @@ int igt_main()
 		pds_buffer_base[0x20] = pds_inst_encode_halt();
 
 		vdm_alloc =
-			igt_pvr_allocate_general(fd, render_vm_ctx_handle, size);
+			igt_pvr_allocate_general(fd, render_vm_ctx_handle, size,
+						 DRM_PVR_BO_ALLOW_CPU_USERSPACE_ACCESS |
+						 DRM_PVR_BO_BYPASS_DEVICE_CACHE);
 
 		vdm_buffer = igt_pvr_get_cpu_addr(fd, vdm_alloc);
 
 		size = PVR_BO_SIZE;
 
 		ppp_alloc =
-			igt_pvr_allocate_general(fd, render_vm_ctx_handle, size);
+			igt_pvr_allocate_general(fd, render_vm_ctx_handle, size,
+						 DRM_PVR_BO_ALLOW_CPU_USERSPACE_ACCESS |
+						 DRM_PVR_BO_BYPASS_DEVICE_CACHE);
 
 		ppp_buffer_base = igt_pvr_get_cpu_addr(fd, ppp_alloc);
 
Signature · 1 line
2.43.0
#2 lore raw

Re: [PATCH i-g-t v4 1/2] lib/igt_pvr: Update general heap allocation interface to pass flags

Hi Brajesh,

Quoted 4 lines · Brajesh Gupta

Update igt_pvr_allocate_general() interface to pass flags for allocation.

Applied, thank you all!

Regards,
Kamil

PS. when resending after a rebase on new top, please also update base-commit in cover letter, or just do not use it. Stale commit hash could broke my automated workflow...

Quoted 6 lines
---
 lib/igt_pvr.c                       | 36 ++++++++++++++++++++++++------------
 lib/igt_pvr.h                       |  3 ++-
 tests/imagination/pvr_submit_jobs.c | 13 ++++++++++---
 3 files changed, 36 insertions(+), 16 deletions(-)

[cut]

#3 lore raw

PATCH v4 2/2 imagination: Skip CPU access flag for a few Render related allocations

In reply to Brajesh Gupta

Skip passing CPU access flag for BO's that don't need host access.

---
 lib/igt_pvr.c | 7 -------
 1 file changed, 7 deletions(-)
lib/igt_pvr.c +0 -7
index e771d0660..9e88b2e41 100644
@@ -400,7 +400,6 @@ uint32_t igt_pvr_ioctl_create_context(int fd, enum drm_pvr_ctx_type type,
 			igt_pvr_get_gpu_addr(igt_pvr_allocate_general(fd,
 								      vm_ctx_handle,
 								      POWERVR_GPU_PAGE_SIZE,
-								      DRM_PVR_BO_ALLOW_CPU_USERSPACE_ACCESS |
 								      DRM_PVR_BO_BYPASS_DEVICE_CACHE));
 		create_context_args.static_context_state = to_user_pointer(init_render_context_stream);
 		create_context_args.static_context_state_len =
@@ -818,15 +817,12 @@ static void igt_pvr_create_hwrt_geom_data_args(int fd, uint32_t vm_ctx,
 {
 	geom_data_args->rtc_dev_addr =
 		igt_pvr_get_gpu_addr(igt_pvr_allocate_general(fd, vm_ctx, POWERVR_GPU_PAGE_SIZE,
-							      DRM_PVR_BO_ALLOW_CPU_USERSPACE_ACCESS |
 							      DRM_PVR_BO_BYPASS_DEVICE_CACHE));
 	geom_data_args->tpc_dev_addr =
 		igt_pvr_get_gpu_addr(igt_pvr_allocate_general(fd, vm_ctx, POWERVR_GPU_PAGE_SIZE,
-							      DRM_PVR_BO_ALLOW_CPU_USERSPACE_ACCESS |
 							      DRM_PVR_BO_BYPASS_DEVICE_CACHE));
 	geom_data_args->vheap_table_dev_addr =
 		igt_pvr_get_gpu_addr(igt_pvr_allocate_general(fd, vm_ctx, POWERVR_GPU_PAGE_SIZE,
-							      DRM_PVR_BO_ALLOW_CPU_USERSPACE_ACCESS |
 							      DRM_PVR_BO_BYPASS_DEVICE_CACHE));
 	geom_data_args->tpc_size = POWERVR_GPU_PAGE_SIZE;
 	geom_data_args->tpc_stride = 0x1;
@@ -846,15 +842,12 @@ static void igt_pvr_create_hwrt_rt_data_args(int fd, uint32_t vm_ctx,
 {
 	rt_data_args->pm_mlist_dev_addr =
 		igt_pvr_get_gpu_addr(igt_pvr_allocate_general(fd, vm_ctx, POWERVR_GPU_PAGE_SIZE,
-							      DRM_PVR_BO_ALLOW_CPU_USERSPACE_ACCESS |
 							      DRM_PVR_BO_BYPASS_DEVICE_CACHE));
 	rt_data_args->macrotile_array_dev_addr =
 		igt_pvr_get_gpu_addr(igt_pvr_allocate_general(fd, vm_ctx, POWERVR_GPU_PAGE_SIZE,
-							      DRM_PVR_BO_ALLOW_CPU_USERSPACE_ACCESS |
 							      DRM_PVR_BO_BYPASS_DEVICE_CACHE));
 	rt_data_args->region_header_dev_addr =
 		igt_pvr_get_gpu_addr(igt_pvr_allocate_general(fd, vm_ctx, POWERVR_GPU_PAGE_SIZE,
-							      DRM_PVR_BO_ALLOW_CPU_USERSPACE_ACCESS |
 							      DRM_PVR_BO_BYPASS_DEVICE_CACHE));
 }
 
Signature · 1 line
2.43.0