Skip to contentSkip to conversation

Compression support for Lunarlake

PATCH Superseded Akshata Jahagirdar 5 participants 11 messages Nov 16, 2023 Library code

v1 of 8 Superseded Latest is v8 →

lore
Share & tools
lore
#0 lore raw

cover Compression support for Lunarlake

Series enables the compression feature for Lunarlake and address various changes of the feature from Gen12. The Main-to-CCS ratio has been changed to 512:1. This changes the calculations and value for fields such as CCS Copy size for blitter command APIs.
This patch series updates tests xe_ccs and gem_ccs for XE2.

These changes are based on top of the "vm_bind pat_index" patch here:
https://patchwork.freedesktop.org/series/124667/

v2: Lots of improvements/tweaks (Kamil)

v3: Addressed review comments (Matthew, Karolina)

#1 lore raw

PATCH 1/5 lib: add blt command properties for lunarlake

Add blt_cmd_info structs to describe properties of XY_BLOCK_COPY and XY_FAST_COPY blitter commands for XE2 platform. Updated the definitions for Lunarlake.

---
 lib/intel_cmds_info.c   | 24 ++++++++++++++++++++++++
 lib/intel_cmds_info.h   |  1 +
 lib/intel_device_info.c |  2 +-
 3 files changed, 26 insertions(+), 1 deletion(-)
lib/intel_cmds_info.c +24 -0
index 2e51ec081..fc3554271 100644
@@ -67,6 +67,22 @@ static const struct blt_cmd_info
 						 BLT_CMD_EXTENDED |
 						 BLT_CMD_SUPPORTS_COMPRESSION);
 
+static const struct blt_cmd_info				
+		xe2_xy_block_copy = BLT_INFO_EXT(XY_BLOCK_COPY,
+						 BIT(T_LINEAR) |
+						 BIT(T_XMAJOR) |
+						 BIT(T_TILE4)  |
+						 BIT(T_TILE64),
+						 BLT_CMD_EXTENDED |
+						 BLT_CMD_SUPPORTS_COMPRESSION);
+
+static const struct blt_cmd_info
+		xe2_xy_fast_copy = BLT_INFO(XY_FAST_COPY,
+					    BIT(T_LINEAR) |
+					    BIT(T_XMAJOR) |
+					    BIT(T_TILE4)  |
+					    BIT(T_TILE64));
+
 static const struct blt_cmd_info
 		mtl_xy_block_copy = BLT_INFO_EXT(XY_BLOCK_COPY,
 						 BIT(T_LINEAR) |
@@ -169,6 +185,14 @@ const struct intel_cmds_info gen12_pvc_cmds_info = {
 	}
 };
 
+
+const struct intel_cmds_info xe2_cmds_info  = {
+	.blt_cmds = {
+		[XY_FAST_COPY] = &xe2_xy_fast_copy,
+		[XY_BLOCK_COPY] = &xe2_xy_block_copy,
+	}
+};
+
 const struct blt_cmd_info *blt_get_cmd_info(const struct intel_cmds_info *cmds_info,
 					    enum blt_cmd_type cmd)
 {
lib/intel_cmds_info.h +1 -0
index f9e3932d1..0a83b6a44 100644
@@ -55,6 +55,7 @@ extern const struct intel_cmds_info gen12_cmds_info;
 extern const struct intel_cmds_info gen12_dg2_cmds_info;
 extern const struct intel_cmds_info gen12_mtl_cmds_info;
 extern const struct intel_cmds_info gen12_pvc_cmds_info;
+extern const struct intel_cmds_info xe2_cmds_info;
 
 #define for_each_tiling(__tiling) \
 	for (__tiling = T_LINEAR; __tiling < __BLT_MAX_TILING; __tiling++)
lib/intel_device_info.c +1 -1
index 34817f7b6..a669797c3 100644
@@ -511,7 +511,7 @@ static const struct intel_device_info intel_lunarlake_info = {
 	.has_4tile = true,
 	.is_lunarlake = true,
 	.codename = "lunarlake",
-	.cmds_info = &gen12_pvc_cmds_info,
+	.cmds_info = &xe2_cmds_info,
 };
 
 static const struct pci_id_match intel_device_match[] = {
Signature · 1 line
2.34.1
#2 lore raw

Re: [igt-dev] [PATCH i-g-t 1/5] lib: add blt command properties for lunarlake

Quoted 1 line · Akshata Jahagirdar

Add blt_cmd_info structs to describe properties of XY_BLOCK_COPY and XY_FAST_COPY blitter commands for XE2 platform. Updated the definitions for Lunarlake.

Hmm, I'm slightly confused. The patch looks the same as the previous one, and doesn't address my comments (like the one with reusing command definition). Could you fix that in the new version?

Also, nit for the patch subject. It should be "lib: Add" (delete extra space and start with a capital letter).

All the best,
Karolina

Quoted 75 lines
---
  lib/intel_cmds_info.c   | 24 ++++++++++++++++++++++++
  lib/intel_cmds_info.h   |  1 +
  lib/intel_device_info.c |  2 +-
  3 files changed, 26 insertions(+), 1 deletion(-)
lib/intel_cmds_info.c +24 -0
index 2e51ec081..fc3554271 100644
@@ -67,6 +67,22 @@ static const struct blt_cmd_info
  						 BLT_CMD_EXTENDED |
  						 BLT_CMD_SUPPORTS_COMPRESSION);
 
+static const struct blt_cmd_info				
+		xe2_xy_block_copy = BLT_INFO_EXT(XY_BLOCK_COPY,
+						 BIT(T_LINEAR) |
+						 BIT(T_XMAJOR) |
+						 BIT(T_TILE4)  |
+						 BIT(T_TILE64),
+						 BLT_CMD_EXTENDED |
+						 BLT_CMD_SUPPORTS_COMPRESSION);
+
+static const struct blt_cmd_info
+		xe2_xy_fast_copy = BLT_INFO(XY_FAST_COPY,
+					    BIT(T_LINEAR) |
+					    BIT(T_XMAJOR) |
+					    BIT(T_TILE4)  |
+					    BIT(T_TILE64));
+
  static const struct blt_cmd_info
  		mtl_xy_block_copy = BLT_INFO_EXT(XY_BLOCK_COPY,
  						 BIT(T_LINEAR) |
@@ -169,6 +185,14 @@ const struct intel_cmds_info gen12_pvc_cmds_info = {
  	}
  };
 
+
+const struct intel_cmds_info xe2_cmds_info  = {
+	.blt_cmds = {
+		[XY_FAST_COPY] = &xe2_xy_fast_copy,
+		[XY_BLOCK_COPY] = &xe2_xy_block_copy,
+	}
+};
+
  const struct blt_cmd_info *blt_get_cmd_info(const struct intel_cmds_info *cmds_info,
  					    enum blt_cmd_type cmd)
  {
lib/intel_cmds_info.h +1 -0
index f9e3932d1..0a83b6a44 100644
@@ -55,6 +55,7 @@ extern const struct intel_cmds_info gen12_cmds_info;
  extern const struct intel_cmds_info gen12_dg2_cmds_info;
  extern const struct intel_cmds_info gen12_mtl_cmds_info;
  extern const struct intel_cmds_info gen12_pvc_cmds_info;
+extern const struct intel_cmds_info xe2_cmds_info;
 
  #define for_each_tiling(__tiling) \
  	for (__tiling = T_LINEAR; __tiling < __BLT_MAX_TILING; __tiling++)
lib/intel_device_info.c +1 -1
index 34817f7b6..a669797c3 100644
@@ -511,7 +511,7 @@ static const struct intel_device_info intel_lunarlake_info = {
  	.has_4tile = true,
  	.is_lunarlake = true,
  	.codename = "lunarlake",
-	.cmds_info = &gen12_pvc_cmds_info,
+	.cmds_info = &xe2_cmds_info,
  };
 
  static const struct pci_id_match intel_device_match[] = {
#3 lore raw

PATCH 2/5 lib/intel_blt: Update calculation of ccs_size and size_of_ctrl_copy

In reply to Akshata Jahagirdar

The Main-to-CCS Ratio for XE2 has been changed to 512:1.
Update the CCS_RATIO macro to select relevant ratio based on platform.
Since the PAGE_SIZE of sysmem is 4K, update the size of ctrl_copy to reflect this change. The pat-index and caching mode for compression need to change to xe2_uc_comp and WC in case of compression, else they just take the default value of pat_index and WB.

---
 lib/intel_blt.c | 23 +++++++++++++++--------
 lib/intel_blt.h |  2 +-
 2 files changed, 16 insertions(+), 9 deletions(-)
lib/intel_blt.c +15 -8
index c0593930c..764b27213 100644
@@ -948,15 +948,16 @@ int blt_block_copy(int fd,
 	return ret;
 }
 
-static uint16_t __ccs_size(const struct blt_ctrl_surf_copy_data *surf)
+static uint16_t __ccs_size(int fd, const struct blt_ctrl_surf_copy_data *surf)
 {
 	uint32_t src_size, dst_size;
+	uint16_t ccsratio = CCS_RATIO(fd);
 
 	src_size = surf->src.access_type == DIRECT_ACCESS ?
-				surf->src.size : surf->src.size / CCS_RATIO;
+				surf->src.size : surf->src.size / ccsratio;
 
 	dst_size = surf->dst.access_type == DIRECT_ACCESS ?
-				surf->dst.size : surf->dst.size / CCS_RATIO;
+				surf->dst.size : surf->dst.size / ccsratio;
 
 	igt_assert_f(src_size <= dst_size, "dst size must be >= src size for CCS copy\n");
 
@@ -1118,6 +1119,7 @@ uint64_t emit_blt_ctrl_surf_copy(int fd,
 	uint64_t dst_offset, src_offset, bb_offset, alignment;
 	uint32_t bbe = MI_BATCH_BUFFER_END;
 	uint32_t *bb;
+	uint16_t num_ccs_blocks = xe_get_default_alignment(fd) / (CCS_RATIO(fd));
 
 	igt_assert_f(ahnd, "ctrl-surf-copy supports softpin only\n");
 	igt_assert_f(surf, "ctrl-surf-copy requires data to do ctrl-surf-copy blit\n");
@@ -1136,7 +1138,7 @@ uint64_t emit_blt_ctrl_surf_copy(int fd,
 		data.xe2.dw00.dst_access_type = surf->dst.access_type;
 
 		/* Ensure dst has size capable to keep src ccs aux */
-		data.xe2.dw00.size_of_ctrl_copy = __ccs_size(surf) / CCS_RATIO - 1;
+		data.xe2.dw00.size_of_ctrl_copy = __ccs_size(fd, surf) / num_ccs_blocks - 1;
 		data.xe2.dw00.length = 0x3;
 
 		data.xe2.dw01.src_address_lo = src_offset;
@@ -1155,7 +1157,7 @@ uint64_t emit_blt_ctrl_surf_copy(int fd,
 		data.gen12.dw00.dst_access_type = surf->dst.access_type;
 
 		/* Ensure dst has size capable to keep src ccs aux */
-		data.gen12.dw00.size_of_ctrl_copy = __ccs_size(surf) / CCS_RATIO - 1;
+		data.gen12.dw00.size_of_ctrl_copy = __ccs_size(fd,surf) / num_ccs_blocks - 1;
 		data.gen12.dw00.length = 0x3;
 
 		data.gen12.dw01.src_address_lo = src_offset;
@@ -1794,7 +1796,8 @@ blt_create_object(const struct blt_copy_data *blt, uint32_t region,
 	uint64_t size = width * height * bpp / 8;
 	uint32_t stride = tiling == T_LINEAR ? width * 4 : width;
 	uint32_t handle;
-
+	uint8_t pat_index = intel_get_pat_idx_uc(blt->fd);
+	uint16_t cpu_caching = DRM_XE_GEM_CPU_CACHING_WB;
 	igt_assert_f(blt->driver, "Driver isn't set, have you called blt_copy_init()?\n");
 
 	obj = calloc(1, sizeof(*obj));
@@ -1804,17 +1807,21 @@ blt_create_object(const struct blt_copy_data *blt, uint32_t region,
 	if (blt->driver == INTEL_DRIVER_XE) {
 		uint64_t flags = region;
 
+		if(AT_LEAST_GEN(intel_get_drm_devid(xe), 20) && mid->compression){
+			pat_index = xe2_get_pat_idx_uc_comp(blt->fd);
+			cpu_caching = DRM_XE_GEM_CPU_CACHING_WC;
+		}
 		if (create_mapping && region != system_memory(blt->fd))
 			flags |= XE_GEM_CREATE_FLAG_NEEDS_VISIBLE_VRAM;
 
 		size = ALIGN(size, xe_get_default_alignment(blt->fd));
-		handle = xe_bo_create_flags(blt->fd, 0, size, flags);
+		handle = xe_bo_create_caching(blt->fd, 0, size, region, cpu_caching);
 	} else {
 		igt_assert(__gem_create_in_memory_regions(blt->fd, &handle,
 							  &size, region) == 0);
 	}
 
-	blt_set_object(obj, handle, size, region, mocs_index, DEFAULT_PAT_INDEX, tiling,
+	blt_set_object(obj, handle, size, region, mocs_index, pat_index, tiling,
 		       compression, compression_type);
 	blt_set_geom(obj, stride, 0, 0, width, height, 0, 0);
 
lib/intel_blt.h +1 -1
index 5934ccd67..a7aeab595 100644
@@ -52,7 +52,7 @@
 #include "igt.h"
 #include "intel_cmds_info.h"
 
-#define CCS_RATIO 256
+#define CCS_RATIO(xe) AT_LEAST_GEN(intel_get_drm_devid(xe), 20) ? 512 : 256
 
 enum blt_color_depth {
 	CD_8bit,
Signature · 1 line
2.34.1
#4 lore raw

Re: [igt-dev] [PATCH i-g-t 2/5] lib/intel_blt: Update calculation of ccs_size and size_of_ctrl_copy

Hi Akshata,
On 2023-11-16 at 07:38:31 -0800, Akshata Jahagirdar wrote:

your patch do not apply, please first make sure it will merge with upstream igt repo, then send new version. Also add version number like:

[PATCH i-g-t v3 2/5] lib/intel_blt: Update calculation of ...

Quoted 5 lines

The Main-to-CCS Ratio for XE2 has been changed to 512:1.
Update the CCS_RATIO macro to select relevant ratio based on platform.
Since the PAGE_SIZE of sysmem is 4K, update the size of ctrl_copy to reflect this change. The pat-index and caching mode for compression need to change to xe2_uc_comp and WC in case of compression, else they just take the default value of pat_index and WB.

Could you split this patch into two, first with calculations and next one with pat index?

lib/intel_blt.c blt_create_object(const struct blt_copy_data *blt, uint32_t region,

Quoted 68 lines
---
 lib/intel_blt.c | 23 +++++++++++++++--------
 lib/intel_blt.h |  2 +-
 2 files changed, 16 insertions(+), 9 deletions(-)
lib/intel_blt.c +9 -6
index c0593930c..764b27213 100644
@@ -948,15 +948,16 @@ int blt_block_copy(int fd,
 	return ret;
 }
 
-static uint16_t __ccs_size(const struct blt_ctrl_surf_copy_data *surf)
+static uint16_t __ccs_size(int fd, const struct blt_ctrl_surf_copy_data *surf)
 {
 	uint32_t src_size, dst_size;
+	uint16_t ccsratio = CCS_RATIO(fd);
 
 	src_size = surf->src.access_type == DIRECT_ACCESS ?
-				surf->src.size : surf->src.size / CCS_RATIO;
+				surf->src.size : surf->src.size / ccsratio;
 
 	dst_size = surf->dst.access_type == DIRECT_ACCESS ?
-				surf->dst.size : surf->dst.size / CCS_RATIO;
+				surf->dst.size : surf->dst.size / ccsratio;
 
 	igt_assert_f(src_size <= dst_size, "dst size must be >= src size for CCS copy\n");
 
@@ -1118,6 +1119,7 @@ uint64_t emit_blt_ctrl_surf_copy(int fd,
 	uint64_t dst_offset, src_offset, bb_offset, alignment;
 	uint32_t bbe = MI_BATCH_BUFFER_END;
 	uint32_t *bb;
+	uint16_t num_ccs_blocks = xe_get_default_alignment(fd) / (CCS_RATIO(fd));
 
 	igt_assert_f(ahnd, "ctrl-surf-copy supports softpin only\n");
 	igt_assert_f(surf, "ctrl-surf-copy requires data to do ctrl-surf-copy blit\n");
@@ -1136,7 +1138,7 @@ uint64_t emit_blt_ctrl_surf_copy(int fd,
 		data.xe2.dw00.dst_access_type = surf->dst.access_type;
 
 		/* Ensure dst has size capable to keep src ccs aux */
-		data.xe2.dw00.size_of_ctrl_copy = __ccs_size(surf) / CCS_RATIO - 1;
+		data.xe2.dw00.size_of_ctrl_copy = __ccs_size(fd, surf) / num_ccs_blocks - 1;
 		data.xe2.dw00.length = 0x3;
 
 		data.xe2.dw01.src_address_lo = src_offset;
@@ -1155,7 +1157,7 @@ uint64_t emit_blt_ctrl_surf_copy(int fd,
 		data.gen12.dw00.dst_access_type = surf->dst.access_type;
 
 		/* Ensure dst has size capable to keep src ccs aux */
-		data.gen12.dw00.size_of_ctrl_copy = __ccs_size(surf) / CCS_RATIO - 1;
+		data.gen12.dw00.size_of_ctrl_copy = __ccs_size(fd,surf) / num_ccs_blocks - 1;
 		data.gen12.dw00.length = 0x3;
 
 		data.gen12.dw01.src_address_lo = src_offset;
@@ -1794,7 +1796,8 @@ blt_create_object(const struct blt_copy_data *blt, uint32_t region,
 	uint64_t size = width * height * bpp / 8;
 	uint32_t stride = tiling == T_LINEAR ? width * 4 : width;
 	uint32_t handle;
-
+	uint8_t pat_index = intel_get_pat_idx_uc(blt->fd);
+	uint16_t cpu_caching = DRM_XE_GEM_CPU_CACHING_WB;
 	igt_assert_f(blt->driver, "Driver isn't set, have you called blt_copy_init()?\n");
 
 	obj = calloc(1, sizeof(*obj));
@@ -1804,17 +1807,21 @@ blt_create_object(const struct blt_copy_data *blt, uint32_t region,

This hunk do not apply.

Regards,
Kamil

Quoted 39 lines
lib/intel_blt.c +6 -2
 	if (blt->driver == INTEL_DRIVER_XE) {
 		uint64_t flags = region;
 
+		if(AT_LEAST_GEN(intel_get_drm_devid(xe), 20) && mid->compression){
+			pat_index = xe2_get_pat_idx_uc_comp(blt->fd);
+			cpu_caching = DRM_XE_GEM_CPU_CACHING_WC;
+		}
 		if (create_mapping && region != system_memory(blt->fd))
 			flags |= XE_GEM_CREATE_FLAG_NEEDS_VISIBLE_VRAM;
 
 		size = ALIGN(size, xe_get_default_alignment(blt->fd));
-		handle = xe_bo_create_flags(blt->fd, 0, size, flags);
+		handle = xe_bo_create_caching(blt->fd, 0, size, region, cpu_caching);
 	} else {
 		igt_assert(__gem_create_in_memory_regions(blt->fd, &handle,
 							  &size, region) == 0);
 	}
 
-	blt_set_object(obj, handle, size, region, mocs_index, DEFAULT_PAT_INDEX, tiling,
+	blt_set_object(obj, handle, size, region, mocs_index, pat_index, tiling,
 		       compression, compression_type);
 	blt_set_geom(obj, stride, 0, 0, width, height, 0, 0);
 
lib/intel_blt.h +1 -1
index 5934ccd67..a7aeab595 100644
@@ -52,7 +52,7 @@
 #include "igt.h"
 #include "intel_cmds_info.h"
 
-#define CCS_RATIO 256
+#define CCS_RATIO(xe) AT_LEAST_GEN(intel_get_drm_devid(xe), 20) ? 512 : 256
 
 enum blt_color_depth {
 	CD_8bit,
Signature · 1 line
2.34.1
#5 lore raw

Re: [igt-dev] [PATCH i-g-t 2/5] lib/intel_blt: Update calculation of ccs_size and size_of_ctrl_copy

Quoted 13 lines

-----Original Message-----
From: igt-dev <[email protected]> On Behalf Of Kamil Konieczny
Sent: Thursday, November 16, 2023 7:09 AM
To: [email protected]
Cc: Auld, Matthew <[email protected]>; Siddiqui, Ayaz A <[email protected]>; Jahagirdar, Akshata <[email protected]>
Subject: Re: [igt-dev] [PATCH i-g-t 2/5] lib/intel_blt: Update calculation of ccs_size and size_of_ctrl_copy

Quoted 4 lines

Hi Akshata,
On 2023-11-16 at 07:38:31 -0800, Akshata Jahagirdar wrote:

your patch do not apply, please first make sure it will merge with upstream igt

These changes are based on https://patchwork.freedesktop.org/series/124667/ which is not merged yet. Hence patch does not apply.

Thanks,
Pallavi

Quoted 147 lines

repo, then send new version. Also add version number like:

[PATCH i-g-t v3 2/5] lib/intel_blt: Update calculation of ...

Quoted 3 lines

The Main-to-CCS Ratio for XE2 has been changed to 512:1.
Update the CCS_RATIO macro to select relevant ratio based on platform. Since the PAGE_SIZE of sysmem is 4K, update the size of ctrl_copy to reflect

this change.

Quoted 2 lines

The pat-index and caching mode for compression need to change to xe2_uc_comp and WC in case of compression, else they just take the default

value of pat_index and WB.

Could you split this patch into two, first with calculations and next one with pat index?

Quoted 42 lines
---
 lib/intel_blt.c | 23 +++++++++++++++--------  lib/intel_blt.h |  2 +-
 2 files changed, 16 insertions(+), 9 deletions(-)
lib/intel_blt.c index renamed from lib/intel_blt.c +0 -0

c0593930c..764b27213 100644

lib/intel_blt.c +5 -3
@@ -948,15 +948,16 @@ int blt_block_copy(int fd,
 	return ret;
 }
 
-static uint16_t __ccs_size(const struct blt_ctrl_surf_copy_data
 *surf)
+static uint16_t __ccs_size(int fd, const struct
+blt_ctrl_surf_copy_data *surf)
 {
 	uint32_t src_size, dst_size;
+	uint16_t ccsratio = CCS_RATIO(fd);
 
 	src_size = surf->src.access_type == DIRECT_ACCESS ?
-				surf->src.size : surf->src.size / CCS_RATIO;
+				surf->src.size : surf->src.size / ccsratio;
 
 	dst_size = surf->dst.access_type == DIRECT_ACCESS ?
-				surf->dst.size : surf->dst.size / CCS_RATIO;
+				surf->dst.size : surf->dst.size / ccsratio;
 
 	igt_assert_f(src_size <= dst_size, "dst size must be >= src size for

CCS copy\n");

lib/intel_blt.c +2 -0
@@ -1118,6 +1119,7 @@ uint64_t emit_blt_ctrl_surf_copy(int fd,
 	uint64_t dst_offset, src_offset, bb_offset, alignment;
 	uint32_t bbe = MI_BATCH_BUFFER_END;
 	uint32_t *bb;
+	uint16_t num_ccs_blocks = xe_get_default_alignment(fd) /
+(CCS_RATIO(fd));
 
 	igt_assert_f(ahnd, "ctrl-surf-copy supports softpin only\n");
 	igt_assert_f(surf, "ctrl-surf-copy requires data to do
Quoted 1 line

ctrl-surf-copy blit\n"); @@ -1136,7 +1138,7 @@ uint64_t

emit_blt_ctrl_surf_copy(int fd,

lib/intel_blt.c uint64_t emit_blt_ctrl_surf_copy(int fd,

Quoted 4 lines
lib/intel_blt.c +0 -1
 		data.xe2.dw00.dst_access_type = surf->dst.access_type;
 
 		/* Ensure dst has size capable to keep src ccs aux */
-		data.xe2.dw00.size_of_ctrl_copy = __ccs_size(surf) /

CCS_RATIO - 1;

lib/intel_blt.c

Quoted 5 lines
lib/intel_blt.c +2 -0
+		data.xe2.dw00.size_of_ctrl_copy = __ccs_size(fd, surf) /
+num_ccs_blocks - 1;
 		data.xe2.dw00.length = 0x3;
 
 		data.xe2.dw01.src_address_lo = src_offset; @@ -1155,7

+1157,7 @@

lib/intel_blt.c uint64_t emit_blt_ctrl_surf_copy(int fd,

Quoted 5 lines

uint64_t emit_blt_ctrl_surf_copy(int fd,

lib/intel_blt.c +0 -1
 		data.gen12.dw00.dst_access_type = surf->dst.access_type;
 
 		/* Ensure dst has size capable to keep src ccs aux */
-		data.gen12.dw00.size_of_ctrl_copy = __ccs_size(surf) /

CCS_RATIO - 1;

lib/intel_blt.c

Quoted 5 lines
lib/intel_blt.c +2 -0
+		data.gen12.dw00.size_of_ctrl_copy = __ccs_size(fd,surf) /
+num_ccs_blocks - 1;
 		data.gen12.dw00.length = 0x3;
 
 		data.gen12.dw01.src_address_lo = src_offset; @@ -1794,7

+1796,8 @@

lib/intel_blt.c blt_create_object(const struct blt_copy_data · line 1807

Quoted 13 lines

blt_create_object(const struct blt_copy_data *blt, uint32_t region,

lib/intel_blt.c +2 -1
 	uint64_t size = width * height * bpp / 8;
 	uint32_t stride = tiling == T_LINEAR ? width * 4 : width;
 	uint32_t handle;
-
+	uint8_t pat_index = intel_get_pat_idx_uc(blt->fd);
+	uint16_t cpu_caching = DRM_XE_GEM_CPU_CACHING_WB;
 	igt_assert_f(blt->driver, "Driver isn't set, have you called

blt_copy_init()?\n");

lib/intel_blt.c +0 -0
 	obj = calloc(1, sizeof(*obj));
@@ -1804,17 +1807,21 @@ blt_create_object(const struct blt_copy_data
 *blt, uint32_t region,

This hunk do not apply.

Regards,
Kamil

lib/intel_blt.c blt_create_object(const struct blt_copy_data *blt, uint32_t region,

Quoted 10 lines
lib/intel_blt.c +4 -0
 	if (blt->driver == INTEL_DRIVER_XE) {
 		uint64_t flags = region;
 
+		if(AT_LEAST_GEN(intel_get_drm_devid(xe), 20) && mid-
 compression){
+			pat_index = xe2_get_pat_idx_uc_comp(blt->fd);
+			cpu_caching = DRM_XE_GEM_CPU_CACHING_WC;
+		}
 		if (create_mapping && region != system_memory(blt->fd))
 			flags |=

XE_GEM_CREATE_FLAG_NEEDS_VISIBLE_VRAM;

lib/intel_blt.c blt_create_object(const struct blt_copy_data *blt, uint32_t region,

Quoted 7 lines
lib/intel_blt.c +2 -1
 		size = ALIGN(size, xe_get_default_alignment(blt->fd));
-		handle = xe_bo_create_flags(blt->fd, 0, size, flags);
+		handle = xe_bo_create_caching(blt->fd, 0, size, region,
+cpu_caching);
 	} else {
 		igt_assert(__gem_create_in_memory_regions(blt->fd,

&handle,

lib/intel_blt.c blt_create_object(const struct blt_copy_data *blt, uint32_t region,

Quoted 4 lines
lib/intel_blt.c +0 -1
 							  &size, region) == 0);
 	}
 
-	blt_set_object(obj, handle, size, region, mocs_index,

DEFAULT_PAT_INDEX, tiling,

Quoted 22 lines
lib/intel_blt.c +2 -0
+	blt_set_object(obj, handle, size, region, mocs_index, pat_index,
+tiling,
 		       compression, compression_type);
 	blt_set_geom(obj, stride, 0, 0, width, height, 0, 0);
lib/intel_blt.h index renamed from lib/intel_blt.h +0 -0

5934ccd67..a7aeab595 100644

lib/intel_blt.h +2 -2
@@ -52,7 +52,7 @@
 #include "igt.h"
 #include "intel_cmds_info.h"
 
-#define CCS_RATIO 256
+#define CCS_RATIO(xe) AT_LEAST_GEN(intel_get_drm_devid(xe), 20) ? 512
+: 256
 
 enum blt_color_depth {
 	CD_8bit,
--

2.34.1

#6 lore raw

PATCH 3/5 lib/intel_pat : Add uc_comp pat_index

In reply to Akshata Jahagirdar

Compression in XE2 is programmed through pat-index attribute. Add a dedicated pat-index for compression for XE2 and later platforms. The caller to this helper function ensures GFX version is correct, else it will call the intel_uc_pat_index() function..

---
 lib/intel_pat.c | 12 +++++++++++-
 lib/intel_pat.h |  2 ++
 2 files changed, 13 insertions(+), 1 deletion(-)
lib/intel_pat.c +11 -1
index 2b892ee52..3c8043f59 100644
@@ -11,7 +11,7 @@ struct intel_pat_cache {
 	uint8_t uc; /* UC + COH_NONE */
 	uint8_t wt; /* WT + COH_NONE */
 	uint8_t wb; /* WB + COH_AT_LEAST_1WAY */
-
+	uint8_t xe2_uc_comp; /* UC + COH_NONE + COMPRESSION*/
 	uint8_t max_index;
 };
 
@@ -23,6 +23,7 @@ static void intel_get_pat_idx(int fd, struct intel_pat_cache *pat)
 		pat->uc = 3;
 		pat->wt = 15; /* Compressed + WB-transient */
 		pat->wb = 2;
+		pat->xe2_uc_comp = 12; /* Compressed + UC */
 		pat->max_index = 31;
 	} else if (IS_METEORLAKE(dev_id)) {
 		pat->uc = 2;
@@ -60,6 +61,15 @@ uint8_t intel_get_pat_idx_uc(int fd)
 	return pat.uc;
 }
 
+uint8_t xe2_get_pat_idx_uc_comp(int fd)
+{
+	struct intel_pat_cache pat = {};
+	uint16_t dev_id = intel_get_drm_devid(fd);
+	
+	intel_get_pat_idx(fd, &pat);
+	return pat.xe2_uc_comp;
+}
+
 uint8_t intel_get_pat_idx_wt(int fd)
 {
 	struct intel_pat_cache pat = {};
lib/intel_pat.h +2 -0
index c24dbc275..de929e16a 100644
@@ -16,4 +16,6 @@ uint8_t intel_get_pat_idx_uc(int fd);
 uint8_t intel_get_pat_idx_wt(int fd);
 uint8_t intel_get_pat_idx_wb(int fd);
 
+uint8_t xe2_get_pat_idx_uc_comp(int fd);
+
 #endif /* INTEL_PAT_H */
Signature · 1 line
2.34.1
#7 lore raw

Re: [igt-dev] [PATCH i-g-t 3/5] lib/intel_pat : Add uc_comp pat_index

Hi Akshata,

regarding subject, remove space before ':' [PATCH i-g-t 3/5] lib/intel_pat : Add uc_comp pat_index

should be:

[PATCH i-g-t 3/5] lib/intel_pat: Add uc_comp pat_index

Also when sending new version, add version number, for example for version 2:

[PATCH i-g-t v2 3/5] lib/intel_pat: Add uc_comp pat_index

Quoted 4 lines · Akshata Jahagirdar

Compression in XE2 is programmed through pat-index attribute. Add a dedicated pat-index for compression for XE2 and later platforms. The caller to this helper function ensures GFX version is correct, else it will call the intel_uc_pat_index() function..

Two dots at end, remove one: s/\.\././

Quoted 12 lines
---
 lib/intel_pat.c | 12 +++++++++++-
 lib/intel_pat.h |  2 ++
 2 files changed, 13 insertions(+), 1 deletion(-)
lib/intel_pat.c +0 -0
index 2b892ee52..3c8043f59 100644

lib/intel_pat.c struct intel_pat_cache { · line 14

Quoted 5 lines
lib/intel_pat.c +1 -1
@@ -11,7 +11,7 @@ struct intel_pat_cache {
 	uint8_t uc; /* UC + COH_NONE */
 	uint8_t wt; /* WT + COH_NONE */
 	uint8_t wb; /* WB + COH_AT_LEAST_1WAY */
-
+	uint8_t xe2_uc_comp; /* UC + COH_NONE + COMPRESSION*/

----------- ^^^
As you stated in description, xe2 or later, imho this should be "uc_comp" and if you want you may add note about xe2 in comment, like:

uint8_t uc_comp; /* UC + COH_NONE + COMPRESSION, xe2 and later */

lib/intel_pat.c static void intel_get_pat_idx(int fd, struct intel_pat_cache *pat) · line 26

Quoted 8 lines
lib/intel_pat.c +1 -0
 	uint8_t max_index;
 };
 
@@ -23,6 +23,7 @@ static void intel_get_pat_idx(int fd, struct intel_pat_cache *pat)
 		pat->uc = 3;
 		pat->wt = 15; /* Compressed + WB-transient */
 		pat->wb = 2;
+		pat->xe2_uc_comp = 12; /* Compressed + UC */
		pat->uc_comp = 12; /* Compressed + UC */

lib/intel_pat.c uint8_t intel_get_pat_idx_uc(int fd) · line 64

Quoted 8 lines
lib/intel_pat.c +1 -0
 		pat->max_index = 31;
 	} else if (IS_METEORLAKE(dev_id)) {
 		pat->uc = 2;
@@ -60,6 +61,15 @@ uint8_t intel_get_pat_idx_uc(int fd)
 	return pat.uc;
 }
 
+uint8_t xe2_get_pat_idx_uc_comp(int fd)

---------- ^^^
imho should start with intel_, see note at header.

lib/intel_pat.c uint8_t intel_get_pat_idx_uc(int fd) · line 67

Quoted 3 lines
lib/intel_pat.c +3 -0
+{
+	struct intel_pat_cache pat = {};
+	uint16_t dev_id = intel_get_drm_devid(fd);

------------ ^
Unused, did you want use it for checking xe2?

lib/intel_pat.c uint8_t intel_get_pat_idx_uc(int fd) · line 69

Quoted 2 lines
lib/intel_pat.c +2 -0
+	
+	intel_get_pat_idx(fd, &pat);

Put newline before return.

lib/intel_pat.c uint8_t intel_get_pat_idx_uc(int fd) · line 70

Quoted 1 line
lib/intel_pat.c +1 -0
+	return pat.xe2_uc_comp;

from description imho something like: IS_XE2_OR_ABOVE(dev_id) ? pat.uc_comp : pat.uc or assert if < Xe2

Quoted 10 lines
lib/intel_pat.c +2 -0
+}
+
 uint8_t intel_get_pat_idx_wt(int fd)
 {
 	struct intel_pat_cache pat = {};
lib/intel_pat.h +0 -0
index c24dbc275..de929e16a 100644

lib/intel_pat.h uint8_t intel_get_pat_idx_uc(int fd); · line 19

Quoted 4 lines
lib/intel_pat.h +1 -0
@@ -16,4 +16,6 @@ uint8_t intel_get_pat_idx_uc(int fd);
 uint8_t intel_get_pat_idx_wt(int fd);
 uint8_t intel_get_pat_idx_wb(int fd);
 
+uint8_t xe2_get_pat_idx_uc_comp(int fd);

---------- ^^^^
Inconsistent, imho should start with intel_, so:

uint8_t intel_get_pat_idx_uc_comp(int fd);

Regards,
Kamil

Quoted 5 lines
lib/intel_pat.h +1 -0
+
 #endif /* INTEL_PAT_H */
Signature · 1 line
2.34.1
#8 lore raw

PATCH 4/5 tests/intel/gem_ccs: Add compression support for Lunarlake

In reply to Akshata Jahagirdar

Update the ccs_ratio call to get value according to platform.

---
 tests/intel/gem_ccs.c | 2 +-
 1 file changed, 1 insertion(+), 1 deletion(-)
tests/intel/gem_ccs.c +1 -1
index 0a691778d..5a17be4cb 100644
@@ -99,7 +99,7 @@ static void surf_copy(int i915,
 	struct blt_block_copy_data_ext ext = {};
 	struct blt_ctrl_surf_copy_data surf = {};
 	uint32_t bb1, bb2, ccs, ccs2, *ccsmap, *ccsmap2;
-	uint64_t bb_size, ccssize = mid->size / CCS_RATIO;
+	uint64_t bb_size, ccssize = mid->size / (CCS_RATIO(i915));
 	uint32_t *ccscopy;
 	uint8_t uc_mocs = intel_get_uc_mocs_index(i915);
 	int result;
Signature · 1 line
2.34.1
#9 lore raw

PATCH 5/5 tests/intel/xe_ccs: Add compression support for Lunarlake

In reply to Akshata Jahagirdar

In XE2 IGFX platform, sysmem also participates in compression.
So, create all blt objects in sysmem itself, and update the pat-index to reflect the compression status. Since we need to align the buffer object size with page size and also have the src size and dst size of CCS copy to be equal, change the default width and height to 1024.

---
 tests/intel/xe_ccs.c | 44 +++++++++++++++++++++++++-------------------
 1 file changed, 25 insertions(+), 19 deletions(-)
tests/intel/xe_ccs.c +25 -19
index 647a6bd2e..bcf6cbd29 100644
@@ -63,8 +63,8 @@ static struct param {
 	.write_png = false,
 	.print_bb = false,
 	.print_surface_info = false,
-	.width = 512,
-	.height = 512,
+	.width = 1024,
+	.height = 1024,
 };
 
 struct test_config {
@@ -95,21 +95,27 @@ static void surf_copy(int xe,
 	struct blt_block_copy_data_ext ext = {};
 	struct blt_ctrl_surf_copy_data surf = {};
 	uint32_t bb1, bb2, ccs, ccs2, *ccsmap, *ccsmap2;
-	uint64_t bb_size, ccssize = mid->size / CCS_RATIO;
+	uint64_t bb_size, ccssize = mid->size / (CCS_RATIO(xe));
 	uint32_t *ccscopy;
 	uint8_t uc_mocs = intel_get_uc_mocs_index(xe);
+	uint8_t comp_pat_index = DEFAULT_PAT_INDEX;
+	uint16_t cpu_caching = DRM_XE_GEM_CPU_CACHING_WB;
 	uint32_t sysmem = system_memory(xe);
 	int result;
 
 	igt_assert(mid->compression);
+	if(AT_LEAST_GEN(intel_get_drm_devid(xe), 20) && mid->compression){
+		comp_pat_index = xe2_get_pat_idx_uc_comp(xe);
+		cpu_caching = DRM_XE_GEM_CPU_CACHING_WC;
+	}
 	ccscopy = (uint32_t *) malloc(ccssize);
-	ccs = xe_bo_create_flags(xe, 0, ccssize, sysmem);
-	ccs2 = xe_bo_create_flags(xe, 0, ccssize, sysmem);
+	ccs = xe_bo_create_caching(xe, 0, ccssize, sysmem, cpu_caching);
+	ccs2 = xe_bo_create_caching(xe, 0, ccssize, sysmem, cpu_caching);
 
 	blt_ctrl_surf_copy_init(xe, &surf);
 	surf.print_bb = param.print_bb;
 	blt_set_ctrl_surf_object(&surf.src, mid->handle, mid->region, mid->size,
-				 uc_mocs, DEFAULT_PAT_INDEX, BLT_INDIRECT_ACCESS);
+				 uc_mocs, comp_pat_index, BLT_INDIRECT_ACCESS);
 	blt_set_ctrl_surf_object(&surf.dst, ccs, sysmem, ccssize, uc_mocs,
 				 DEFAULT_PAT_INDEX, DIRECT_ACCESS);
 	bb_size = xe_get_default_alignment(xe);
@@ -157,7 +163,7 @@ static void surf_copy(int xe,
 	blt_set_ctrl_surf_object(&surf.src, ccs, sysmem, ccssize,
 				 uc_mocs, DEFAULT_PAT_INDEX, DIRECT_ACCESS);
 	blt_set_ctrl_surf_object(&surf.dst, mid->handle, mid->region, mid->size,
-				 uc_mocs, DEFAULT_PAT_INDEX, INDIRECT_ACCESS);
+				 uc_mocs, comp_pat_index, INDIRECT_ACCESS);
 	blt_ctrl_surf_copy(xe, ctx, NULL, ahnd, &surf);
 	intel_ctx_xe_sync(ctx, true);
 
@@ -234,10 +240,10 @@ static int blt_block_copy3(int xe,
 	igt_assert_f(blt3, "block-copy3 requires data to do blit\n");
 
 	alignment = xe_get_default_alignment(xe);
-	get_offset(ahnd, blt3->src.handle, blt3->src.size, alignment);
-	get_offset(ahnd, blt3->mid.handle, blt3->mid.size, alignment);
-	get_offset(ahnd, blt3->dst.handle, blt3->dst.size, alignment);
-	get_offset(ahnd, blt3->final.handle, blt3->final.size, alignment);
+	get_offset_pat_index(ahnd, blt3->src.handle, blt3->src.size, alignment, blt3->src.pat_index);
+	get_offset_pat_index(ahnd, blt3->mid.handle, blt3->mid.size, alignment, blt3->mid.pat_index);
+	get_offset_pat_index(ahnd, blt3->dst.handle, blt3->dst.size, alignment, blt3->dst.pat_index);
+	get_offset_pat_index(ahnd, blt3->final.handle, blt3->final.size, alignment, blt3->final.pat_index);
 	bb_offset = get_offset(ahnd, blt3->bb.handle, blt3->bb.size, alignment);
 
 	/* First blit src -> mid */
@@ -291,8 +297,8 @@ static void block_copy(int xe,
 	uint64_t bb_size = xe_get_default_alignment(xe);
 	uint64_t ahnd = intel_allocator_open(xe, ctx->vm, INTEL_ALLOCATOR_RELOC);
 	uint32_t run_id = mid_tiling;
-	uint32_t mid_region = region2, bb;
-	uint32_t width = param.width, height = param.height;
+	uint32_t mid_region = (AT_LEAST_GEN(intel_get_drm_devid(xe), 20) & !xe_has_vram(xe)) ? region1 : region2;
+	uint32_t width = param.width, height = param.height, bb;
 	enum blt_compression mid_compression = config->compression;
 	int mid_compression_format = param.compression_format;
 	enum blt_compression_type comp_type = COMPRESSION_TYPE_3D;
@@ -413,8 +419,8 @@ static void block_multicopy(int xe,
 	uint64_t bb_size = xe_get_default_alignment(xe);
 	uint64_t ahnd = intel_allocator_open(xe, ctx->vm, INTEL_ALLOCATOR_RELOC);
 	uint32_t run_id = mid_tiling;
-	uint32_t mid_region = region2, bb;
-	uint32_t width = param.width, height = param.height;
+	uint32_t mid_region = (AT_LEAST_GEN(intel_get_drm_devid(xe), 20) & !xe_has_vram(xe)) ? region1 : region2;
+	uint32_t width = param.width, height = param.height, bb;
 	enum blt_compression mid_compression = config->compression;
 	int mid_compression_format = param.compression_format;
 	enum blt_compression_type comp_type = COMPRESSION_TYPE_3D;
@@ -539,8 +545,8 @@ static void block_copy_test(int xe,
 			region1 = igt_collection_get_value(regions, 0);
 			region2 = igt_collection_get_value(regions, 1);
 
-			/* Compressed surface must be in device memory */
-			if (config->compression && !XE_IS_VRAM_MEMORY_REGION(xe, region2))
+			/* if not XE2, then Compressed surface must be in device memory */
+			if (config->compression && !(AT_LEAST_GEN((intel_get_drm_devid(xe)), 20)) && !XE_IS_VRAM_MEMORY_REGION(xe, region2))
 				continue;
 
 			regtxt = xe_memregion_dynamic_subtest_name(xe, regions);
@@ -621,8 +627,8 @@ const char *help_str =
 	"  -p\tWrite PNG\n"
 	"  -s\tPrint surface info\n"
 	"  -t\tTiling format (0 - linear, 1 - XMAJOR, 2 - YMAJOR, 3 - TILE4, 4 - TILE64)\n"
-	"  -W\tWidth (default 512)\n"
-	"  -H\tHeight (default 512)"
+	"  -W\tWidth (default 1024)\n"
+	"  -H\tHeight (default 1024)"
 	;
 
 igt_main_args("bf:pst:W:H:", NULL, help_str, opt_handler, NULL)
Signature · 1 line
2.34.1
#10 lore raw

[igt-dev] ✗ Fi.CI.BUILD: failure for Compression support for Lunarlake (rev2)

In reply to Akshata Jahagirdar
== Series Details ==

== Series Details ==

Series: Compression support for Lunarlake (rev2)
URL : https://patchwork.freedesktop.org/series/126254/
State : failure

== Summary ==

Applying: lib: add blt command properties for lunarlake
Applying: lib/intel_blt: Update calculation of ccs_size and size_of_ctrl_copy Using index info to reconstruct a base tree...
M lib/intel_blt.c
M lib/intel_blt.h
Falling back to patching base and 3-way merge...
Auto-merging lib/intel_blt.h
Auto-merging lib/intel_blt.c
CONFLICT (content): Merge conflict in lib/intel_blt.c
Patch failed at 0002 lib/intel_blt: Update calculation of ccs_size and size_of_ctrl_copy When you have resolved this problem, run "git am --continue".
If you prefer to skip this patch, run "git am --skip" instead.
To restore the original branch and stop patching, run "git am --abort".