Compression support for Lunarlake
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)
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
@@ -67,6 +67,22 @@ static const struct blt_cmd_infoBLT_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
@@ -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
@@ -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
PATCH 2/5 lib/intel_blt: Update calculation of ccs_size and size_of_ctrl_copy
In reply to Akshata JahagirdarThe 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.
--- lib/intel_blt.c | 19 +++++++++++-------- lib/intel_blt.h | 2 +- 2 files changed, 12 insertions(+), 9 deletions(-)
lib/intel_blt.c +11 -8
@@ -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 = compression ?intel_get_pat_idx_uc_comp (blt->fd) :intel_get_pat_idx_uc (blt->fd); +uint16_t cpu_caching =DRM_XE_GEM_CPU_CACHING_WC ;igt_assert_f (blt->driver,"Driver isn't set, have you called blt_copy_init()?\n" ); obj =calloc (1 ,sizeof (*obj));
@@ -1808,13 +1811,13 @@ blt_create_object(const struct blt_copy_data *blt, uint32_t region, 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
@@ -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 : 256enum blt_color_depth { CD_8bit,
Signature · 1 line
2.34.1
PATCH 3/5 lib/intel_pat : Add uc_comp pat_index
In reply to Akshata JahagirdarCompression in XE2 is programmed through pat-index attribute. Add a dedicated pat-index for compression for XE2 and later platforms. The helper function returns uc_pat_index for previous platforms.
--- lib/intel_pat.c | 12 +++++++++++- lib/intel_pat.h | 1 + 2 files changed, 12 insertions(+), 1 deletion(-)
lib/intel_pat.c +11 -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 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->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 intel_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 (intel_get_device_info (dev_id)->graphics_ver >=20 ) ? pat.uc_comp : pat.uc; +} +uint8_t intel_get_pat_idx_wt (int fd) {struct intel_pat_cache pat = {};
lib/intel_pat.h +1 -0
@@ -13,6 +13,7 @@uint8_t intel_get_max_pat_index (int fd);uint8_t intel_get_pat_idx_uc (int fd); +uint8_t intel_get_pat_idx_uc_comp (int fd);uint8_t intel_get_pat_idx_wt (int fd);uint8_t intel_get_pat_idx_wb (int fd);
Signature · 1 line
2.34.1
PATCH 4/5 tests/intel/gem_ccs: Add compression support for Lunarlake
In reply to Akshata JahagirdarUpdate 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
@@ -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
PATCH 5/5 tests/intel/xe_ccs: Add compression support for Lunarlake
In reply to Akshata JahagirdarIn 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 | 38 +++++++++++++++++++------------------- 1 file changed, 19 insertions(+), 19 deletions(-)
tests/intel/xe_ccs.c +19 -19
@@ -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 {
@@ -95,7 +95,7 @@ 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);uint32_t sysmem =system_memory (xe);
@@ -103,13 +103,13 @@ static void surf_copy(int xe,igt_assert (mid->compression); 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,DRM_XE_GEM_CPU_CACHING_WC ); + ccs2 =xe_bo_create_caching (xe,0 , ccssize, sysmem,DRM_XE_GEM_CPU_CACHING_WC );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,intel_get_pat_idx_uc_comp (xe),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 +157,7 @@ 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,intel_get_pat_idx_uc_comp (xe),INDIRECT_ACCESS );blt_ctrl_surf_copy (xe, ctx,NULL , ahnd, &surf);intel_ctx_xe_sync (ctx, true);
@@ -234,10 +234,10 @@ 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 +291,8 @@ 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 +413,8 @@ 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 +539,8 @@ 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 +621,8 @@ 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
Patchwork
CI[igt-dev] ✗ Fi.CI.BUILD: failure for Compression support for Lunarlake
In reply to Akshata Jahagirdar== Series Details ==
== Series Details ==
Series: Compression support for Lunarlake
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".
Re: [igt-dev] [PATCH i-g-t 2/5] lib/intel_blt: Update calculation of ccs_size and size_of_ctrl_copy
In reply to Akshata JahagirdarQuoted 2 lines · Akshata Jahagirdar
The Main-to-CCS Ratio for XE2 has been changed to 512:1. Update the
CCS_RATIOmacro to select relevant ratio based on platform. Since thePAGE_SIZEof sysmem is 4K, update the size ofctrl_copyto reflect this change.
Nit: very long line.
Quoted 58 lines
Signed-off-by Akshata Jahagirdar <[email protected]>--- lib/intel_blt.c | 19 +++++++++++-------- lib/intel_blt.h | 2 +- 2 files changed, 12 insertions(+), 9 deletions(-)lib/intel_blt.c +7
-5@@ -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;
lib/intel_blt.c blt_create_object(const struct blt_copy_data *blt, uint32_t region, · line 1799
Quoted 5 lines
lib/intel_blt.c +1
-1@@ -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 = compression ?intel_get_pat_idx_uc_comp (blt->fd) :intel_get_pat_idx_uc (blt->fd);
intel_get_pat_idx_uc_comp() has not been defined yet in the series?
lib/intel_blt.c blt_create_object(const struct blt_copy_data *blt, uint32_t region, · line 1800
Quoted 1 line
lib/intel_blt.c +1
-0+uint16_t cpu_caching =DRM_XE_GEM_CPU_CACHING_WC ;
Maybe make this conditional on xe2+ and compression = true? I figure we still want wb (if possible) for non-compressed stuff.
Also commit doesn't really explain this change in behaviour? Should this change be separated?
Quoted 32 lines
lib/intel_blt.c +2
-2igt_assert_f (blt->driver,"Driver isn't set, have you called blt_copy_init()?\n" ); obj =calloc (1 ,sizeof (*obj));@@ -1808,13 +1811,13 @@ blt_create_object(const struct blt_copy_data *blt, uint32_t region, 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@@ -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 : 256enum blt_color_depth { CD_8bit,
Re: [igt-dev] [PATCH i-g-t 3/5] lib/intel_pat : Add uc_comp pat_index
In reply to Akshata JahagirdarQuoted 13 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 helper function returns
uc_pat_indexfor previous platforms.Signed-off-by Akshata Jahagirdar <[email protected]>--- lib/intel_pat.c | 12 +++++++++++- lib/intel_pat.h | 1 + 2 files changed, 12 insertions(+), 1 deletion(-)lib/intel_pat.c +0
-0
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 uc_comp;/* UC + COH_NONE + COMPRESSION*/
Maybe prefix this with xe2? i.e xe2_uc_comp.
Quoted 12 lines
lib/intel_pat.c +1
-0uint8_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->uc_comp =12 ;/* Compressed + UC */ pat->max_index =31 ; }else if (IS_METEORLAKE (dev_id)) { pat->uc =2 ;
lib/intel_pat.c uint8_t intel_get_pat_idx_uc(int fd) · line 70
Quoted 10 lines
lib/intel_pat.c +7
-0@@ -60,6 +61,15 @@ uint8_t intel_get_pat_idx_uc(int fd)return pat.uc; } +uint8_t intel_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 (intel_get_device_info (dev_id)->graphics_ver >=20 ) ? pat.uc_comp : pat.uc;
Maybe make just make this xe2_get_pat_idx_uc_comp()? Or if we really do want to return plain uc on pre-xe2 then maybe make that clearer somehow in the api? intel_get_pat_idx_uc_comp_if_possible()? But perhaps caller should just handle this.
Quoted 17 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 +1
-0@@ -13,6 +13,7 @@uint8_t intel_get_max_pat_index (int fd);uint8_t intel_get_pat_idx_uc (int fd); +uint8_t intel_get_pat_idx_uc_comp (int fd);uint8_t intel_get_pat_idx_wt (int fd);uint8_t intel_get_pat_idx_wb (int fd);
Re: [igt-dev] [PATCH i-g-t 1/5] lib: add blt command properties for lunarlake
In reply to Akshata JahagirdarQuoted 1 line · Akshata Jahagirdar
Add
blt_cmd_infostructs to describe properties ofXY_BLOCK_COPYandXY_FAST_COPYblitter commands for XE2 platform. Updated the definitions for Lunarlake.
Nit: break the commit message at 80 column
Quoted 14 lines
Signed-off-by Akshata Jahagirdar <[email protected]>--- 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 +0
-0@@ -67,6 +67,22 @@ static const struct blt_cmd_info BLT_CMD_EXTENDED |
lib/intel_cmds_info.c static const struct blt_cmd_info · line 77
Quoted 10 lines
lib/intel_cmds_info.c +8
-0BLT_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 );
I know this might be too much work, but could we confirm that XMAJOR is actually supported with block copy on this platform? This could be as easy as running xe_ccs test with -p parameter, and then checking the mid image and the tiling pattern. But I won't argue it's essential for r-b, it's more of a sanity check.
lib/intel_cmds_info.c static const struct blt_cmd_info · line 84
Quoted 7 lines
lib/intel_cmds_info.c +7
-0+ +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 ));
This looks a lot like dg2_xy_fast_copy definition. Why not reuse it when defining xe2_cmds_info?
With these comments addressed:
Quoted 44 lines
lib/intel_cmds_info.c +9
-0+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@@ -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@@ -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[] = {
Re: [igt-dev] [PATCH i-g-t 2/5] lib/intel_blt: Update calculation of ccs_size and size_of_ctrl_copy
In reply to Matthew AuldHi,
Comments inline.
Best,
Akshata
-----Original Message-----
From: Auld, Matthew <[email protected]>
Sent: Friday, November 10, 2023 3:00 AM
To: Jahagirdar, Akshata <[email protected]>
Cc: [email protected]; Siddiqui, Ayaz A <[email protected]>; Stolarek, Karolina <[email protected]>; Kempczynski, Zbigniew <[email protected]>
Subject: Re: [PATCH i-g-t 2/5] lib/intel_blt: Update calculation of ccs_size and size_of_ctrl_copy
Quoted 2 lines · Akshata Jahagirdar
The Main-to-CCS Ratio for XE2 has been changed to 512:1. Update the
CCS_RATIOmacro to select relevant ratio based on platform. Since thePAGE_SIZEof sysmem is 4K, update the size ofctrl_copyto reflect this change.
Nit: very long line.
Will fix this.
Quoted 67 lines
Signed-off-by Akshata Jahagirdar <[email protected]>--- lib/intel_blt.c | 19 +++++++++++-------- lib/intel_blt.h | 2 +- 2 files changed, 12 insertions(+), 9 deletions(-)lib/intel_blt.c index renamed from lib/intel_blt.c +0
-0
c0593930c..8168ce7b5100644lib/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 sizefor 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 doctrl-surf-copy blit\n"); @@ -1136,7 +1138,7 @@
uint64_temit_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 */
lib/intel_blt.c +2
-1- 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_temit_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 */
lib/intel_blt.c +2
-1- 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_tregion,
uint64_tsize = width * height * bpp / 8;uint32_tstride = tiling ==T_LINEAR? width * 4 : width;uint32_thandle;
lib/intel_blt.c
Quoted 3 lines
lib/intel_blt.c +2
-1- +uint8_t pat_index = compression ?intel_get_pat_idx_uc_comp (blt->fd) +:intel_get_pat_idx_uc (blt->fd);
intel_get_pat_idx_uc_comp() has not been defined yet in the series?
The patch order seems to be messed up, will fix this.
lib/intel_blt.c blt_create_object(const struct blt_copy_data *blt, uint32_t region, · line 1800
Quoted 1 line
lib/intel_blt.c +1
-0+uint16_t cpu_caching =DRM_XE_GEM_CPU_CACHING_WC ;
Maybe make this conditional on xe2+ and compression = true? I figure we still want wb (if possible) for non-compressed stuff.
Also commit doesn't really explain this change in behaviour? Should this change be separated?
Yes, adding the condition sounds good. Will address this change is commit message too.
Quoted 36 lines
igt_assert_f(blt->driver, "Driver isn't set, have you called
blt_copy_init()?\n");obj = calloc(1, sizeof(*obj));lib/intel_blt.c +4
-2@@ -1808,13 +1811,13 @@ blt_create_object(const struct blt_copy_data *blt, uint32_t region, 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 index renamed from lib/intel_blt.h +0
-0
5934ccd67..a7aeab595100644lib/intel_blt.h +2
-1@@ -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 +: 256enum blt_color_depth { CD_8bit,
Re: [igt-dev] [PATCH i-g-t 3/5] lib/intel_pat : Add uc_comp pat_index
In reply to Matthew AuldHI,
Comments inline.
Best,
Akshata
-----Original Message-----
From: Auld, Matthew <[email protected]>
Sent: Friday, November 10, 2023 3:06 AM
To: Jahagirdar, Akshata <[email protected]>
Cc: [email protected]; Siddiqui, Ayaz A <[email protected]>; Stolarek, Karolina <[email protected]>; Kempczynski, Zbigniew <[email protected]>
Subject: Re: [PATCH i-g-t 3/5] lib/intel_pat : Add uc_comp pat_index
Quoted 13 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 helper function returns
uc_pat_indexfor previous platforms.Signed-off-by Akshata Jahagirdar <[email protected]>--- lib/intel_pat.c | 12 +++++++++++- lib/intel_pat.h | 1 + 2 files changed, 12 insertions(+), 1 deletion(-)lib/intel_pat.c index renamed from lib/intel_pat.c +0
-0
2b892ee52..36f13642f100644
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 uc_comp;/* UC + COH_NONE + COMPRESSION*/
Maybe prefix this with xe2? i.e xe2_uc_comp.
AJ: yes, will fix this.
Quoted 12 lines
lib/intel_pat.c +1
-0uint8_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->uc_comp =12 ;/* Compressed + UC */ pat->max_index =31 ; }else if (IS_METEORLAKE (dev_id)) { pat->uc =2 ;
lib/intel_pat.c uint8_t intel_get_pat_idx_uc(int fd) · line 70
Quoted 10 lines
lib/intel_pat.c +7
-0@@ -60,6 +61,15 @@ uint8_t intel_get_pat_idx_uc(int fd)return pat.uc; } +uint8_t intel_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 (intel_get_device_info (dev_id)->graphics_ver >=20 ) ? +pat.uc_comp : pat.uc;
Maybe make just make this xe2_get_pat_idx_uc_comp()? Or if we really do want to return plain uc on pre-xe2 then maybe make that clearer somehow in the api? intel_get_pat_idx_uc_comp_if_possible()? But perhaps caller should just handle this.
AJ: yeah, I think we can make it xe2_get_pat_idx_uc_comp() and not return pre-xe2, instead handle it from caller.
Quoted 17 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 index renamed from lib/intel_pat.h +0
-0
c24dbc275..6d68b9fc2100644lib/intel_pat.h +1
-0@@ -13,6 +13,7 @@uint8_t intel_get_max_pat_index (int fd);uint8_t intel_get_pat_idx_uc (int fd); +uint8_t intel_get_pat_idx_uc_comp (int fd);uint8_t intel_get_pat_idx_wt (int fd);uint8_t intel_get_pat_idx_wb (int fd);
Jahagirdar, Akshata
Reviewed-by Karolina StolarekRe: [igt-dev] [PATCH i-g-t 1/5] lib: add blt command properties for lunarlake
In reply to Karolina StolarekHI,
Comments inline.
-----Original Message-----
From: Stolarek, Karolina <[email protected]>
Sent: Monday, November 13, 2023 12:42 AM
To: Jahagirdar, Akshata <[email protected]>
Cc: [email protected]; Siddiqui, Ayaz A <[email protected]>; Kempczynski, Zbigniew <[email protected]>; Auld, Matthew <[email protected]>
Subject: Re: [PATCH i-g-t 1/5] lib: add blt command properties for lunarlake
Quoted 1 line · Akshata Jahagirdar
Add
blt_cmd_infostructs to describe properties ofXY_BLOCK_COPYandXY_FAST_COPYblitter commands for XE2 platform. Updated the definitions for Lunarlake.
Nit: break the commit message at 80 column
AJ: Will fix this.
Quoted 14 lines
Signed-off-by Akshata Jahagirdar <[email protected]>--- 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 index renamed from lib/intel_cmds_info.c +0
-0
2e51ec081..fc3554271100644lib/intel_cmds_info.c +0
-0@@ -67,6 +67,22 @@ static const struct blt_cmd_info BLT_CMD_EXTENDED |
lib/intel_cmds_info.c static const struct blt_cmd_info · line 77
Quoted 10 lines
lib/intel_cmds_info.c +8
-0BLT_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 );
I know this might be too much work, but could we confirm that XMAJOR is actually supported with block copy on this platform? This could be as easy as running xe_ccs test with -p parameter, and then checking the mid image and the tiling pattern. But I won't argue it's essential for r-b, it's more of a sanity check.
AJ: I have verified this through bspec, as well as tests are passing on the platform.
lib/intel_cmds_info.c static const struct blt_cmd_info · line 84
Quoted 7 lines
lib/intel_cmds_info.c +7
-0+ +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 ));
This looks a lot like dg2_xy_fast_copy definition. Why not reuse it when defining xe2_cmds_info?
AJ: Yes, I thought of keeping it separate to address any other differences.
With these comments addressed:
Quoted 44 lines
lib/intel_cmds_info.c +9
-0+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 index renamed from lib/intel_cmds_info.h +0
-0
f9e3932d1..0a83b6a44100644lib/intel_cmds_info.h +1
-0@@ -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 index renamed from lib/intel_device_info.c +0
-0
34817f7b6..a669797c3100644lib/intel_device_info.c +1
-1@@ -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[] = {