diff --git a/script/cmake-ck-dev.sh b/script/cmake-ck-dev.sh index 086359a79f..6220009b03 100755 --- a/script/cmake-ck-dev.sh +++ b/script/cmake-ck-dev.sh @@ -20,7 +20,7 @@ fi GPU_TARGETS="gfx908;gfx90a;gfx942" if [ $# -ge 1 ]; then - case "$1" in + case "$1" in gfx*) GPU_TARGETS=$1 shift 1 @@ -38,7 +38,7 @@ fi cmake \ -D CMAKE_PREFIX_PATH=/opt/rocm/ \ -D CMAKE_CXX_COMPILER=/opt/rocm/llvm/bin/clang++ \ --D CMAKE_CXX_FLAGS="-ftemplate-backtrace-limit=0 -fPIE -Wno-gnu-line-marker" \ +-D CMAKE_CXX_FLAGS="-ftemplate-backtrace-limit=0 -fPIE -Wno-gnu-line-marker -fbracket-depth=512" \ -D CMAKE_BUILD_TYPE=Release \ -D BUILD_DEV=ON \ -D GPU_TARGETS=$GPU_TARGETS \ diff --git a/tile_engine/ops/gemm/CMakeLists.txt b/tile_engine/ops/gemm/CMakeLists.txt index d52351af2d..a486802343 100644 --- a/tile_engine/ops/gemm/CMakeLists.txt +++ b/tile_engine/ops/gemm/CMakeLists.txt @@ -13,38 +13,38 @@ function(create_individual_gemm_target datatype layout trait tile_config config_ message(WARNING "Skipping individual GEMM target ${datatype}_${layout}_${trait}_${tile_config}: No supported GPU targets") return() endif() - + # Parse tile configuration: format is tile_mxtile_nxtile_k_warp_mxwarp_nxwarp_k_warp_tile_mxwarp_tile_nxwarp_tile_k # First split by underscore to get three groups string(REPLACE "_" ";" config_groups ${tile_config}) list(GET config_groups 0 tile_dims) # e.g., 256x256x32 list(GET config_groups 1 warp_dims) # e.g., 4x1x1 list(GET config_groups 2 warp_tile_dims) # e.g., 16x16x16 - + # Parse tile dimensions string(REPLACE "x" ";" tile_parts ${tile_dims}) list(GET tile_parts 0 tile_m) list(GET tile_parts 1 tile_n) list(GET tile_parts 2 tile_k) - + # Parse warp dimensions string(REPLACE "x" ";" warp_parts ${warp_dims}) list(GET warp_parts 0 warp_m) list(GET warp_parts 1 warp_n) list(GET warp_parts 2 warp_k) - + # Parse warp tile dimensions string(REPLACE "x" ";" warp_tile_parts ${warp_tile_dims}) list(GET warp_tile_parts 0 warp_tile_m) list(GET warp_tile_parts 1 warp_tile_n) list(GET warp_tile_parts 2 warp_tile_k) - + set(target_name "benchmark_gemm_${datatype}_${layout}_${trait}_${tile_config}") set(working_path "${CMAKE_CURRENT_BINARY_DIR}/${datatype}/${layout}") - + # Generate the single instance header for this kernel set(instance_header "${working_path}/gemm_single_${datatype}_${layout}_${trait}_${tile_config}.hpp") - + # Add custom command to generate the header file at build time add_custom_command( OUTPUT ${instance_header} @@ -60,27 +60,27 @@ function(create_individual_gemm_target datatype layout trait tile_config config_ DEPENDS ${GEMM_SOURCE_DIR}/gemm_instance_builder.py ${config_json} COMMENT "Generating ${instance_header}" ) - + # Create the executable - add_executable(${target_name} + add_executable(${target_name} ${GEMM_SOURCE_DIR}/benchmark_gemm_single.cpp ${instance_header} ) - + # Set GPU architectures set_property(TARGET ${target_name} PROPERTY HIP_ARCHITECTURES ${GEMM_GPU_TARGETS_INDIVIDUAL}) - + # Set compile definitions target_compile_definitions(${target_name} PRIVATE GEMM_SINGLE_INSTANCE_HPP="${instance_header}" ) - + # Include directories target_include_directories(${target_name} PRIVATE ${GEMM_SOURCE_DIR} ${working_path} ) - + # Compile options target_compile_options(${target_name} PRIVATE -Wno-undefined-func-template @@ -88,19 +88,19 @@ function(create_individual_gemm_target datatype layout trait tile_config config_ --offload-compress -include ${instance_header} ) - + # Add to collection targets add_dependencies(benchmark_gemm_all ${target_name}) add_dependencies(benchmark_gemm_${datatype} ${target_name}) add_dependencies(benchmark_gemm_${layout} ${target_name}) add_dependencies(benchmark_gemm_${datatype}_${layout} ${target_name}) - + # Add to trait-specific targets string(REPLACE "_" ";" trait_parts ${trait}) list(GET trait_parts 0 pipeline) list(GET trait_parts 1 epilogue) list(GET trait_parts 2 scheduler) - + add_dependencies(benchmark_gemm_${pipeline} ${target_name}) add_dependencies(benchmark_gemm_${epilogue} ${target_name}) add_dependencies(benchmark_gemm_${scheduler} ${target_name}) @@ -109,13 +109,13 @@ endfunction() # Function to build individual GEMM targets function(build_individual_gemm_targets datatype layout) set(working_path "${CMAKE_CURRENT_BINARY_DIR}/${datatype}/${layout}") - + # Choose config file # Priority order: # 1. Environment variable GEMM_CONFIG_FILE - # 2. CMake variable GEMM_CONFIG_FILE + # 2. CMake variable GEMM_CONFIG_FILE # 3. Default based on layout - + # Check environment variable first if(DEFINED ENV{GEMM_CONFIG_FILE} AND NOT "$ENV{GEMM_CONFIG_FILE}" STREQUAL "") set(config_filename "$ENV{GEMM_CONFIG_FILE}") @@ -130,12 +130,12 @@ function(build_individual_gemm_targets datatype layout) set(json_blob "${CMAKE_CURRENT_LIST_DIR}/configs/default_config.json") message(STATUS " Using default config for layout ${layout}") endif() - + # Check if config file exists if(NOT EXISTS ${json_blob}) message(FATAL_ERROR "Config file not found: ${json_blob}") endif() - + # Determine number of workers for parallel generation if(DEFINED ENV{CMAKE_BUILD_PARALLEL_LEVEL}) set(num_workers $ENV{CMAKE_BUILD_PARALLEL_LEVEL}) @@ -147,17 +147,17 @@ function(build_individual_gemm_targets datatype layout) set(num_workers 8) endif() endif() - + # Generate individual kernel files using parallel version message(STATUS "Generating individual kernels for ${datatype} ${layout} using ${num_workers} workers...") message(STATUS " Working path: ${working_path}") message(STATUS " Config file: ${json_blob}") message(STATUS " Python executable: ${Python3_EXECUTABLE}") message(STATUS " Script path: ${CMAKE_CURRENT_LIST_DIR}/gemm_instance_builder.py") - + # Create working directory first file(MAKE_DIRECTORY ${working_path}) - + # First, just list the kernels (fast operation) message(STATUS " Listing kernel configurations...") execute_process( @@ -172,11 +172,11 @@ function(build_individual_gemm_targets datatype layout) OUTPUT_VARIABLE list_output ERROR_VARIABLE list_error ) - + if(NOT ret EQUAL 0) message(FATAL_ERROR "Failed to list kernels for ${datatype} ${layout}: ${list_error}") endif() - + # Read kernel count if(EXISTS ${working_path}/gemm_kernel_count.txt) file(READ ${working_path}/gemm_kernel_count.txt kernel_count) @@ -185,7 +185,7 @@ function(build_individual_gemm_targets datatype layout) else() message(FATAL_ERROR "Kernel count file not found") endif() - + # Read kernel list and create targets if(EXISTS ${working_path}/gemm_kernel_list.txt) file(STRINGS ${working_path}/gemm_kernel_list.txt kernel_lines) @@ -195,7 +195,7 @@ function(build_individual_gemm_targets datatype layout) list(GET parts 0 kernel_name) list(GET parts 1 tile_config) list(GET parts 2 trait_combo) - + # Create individual target create_individual_gemm_target("${datatype}" "${layout}" "${trait_combo}" "${tile_config}" "${json_blob}") endforeach() @@ -210,9 +210,9 @@ message(STATUS "GEMM_DATATYPE: ${GEMM_DATATYPE}") message(STATUS "GEMM_LAYOUT: ${GEMM_LAYOUT}") message(STATUS "SUPPORTED_GPU_TARGETS: ${SUPPORTED_GPU_TARGETS}") -# Filter GPU targets to only gfx90a, gfx942, and gfx950 +# Filter GPU targets to only gfx90a, gfx942, gfx950, gfx1201 set(GEMM_GPU_TARGETS_INDIVIDUAL "") -set(DESIRED_TARGETS "gfx90a;gfx942;gfx950") +set(DESIRED_TARGETS "gfx90a;gfx942;gfx950;gfx1201") foreach(target IN LISTS SUPPORTED_GPU_TARGETS) if(target IN_LIST DESIRED_TARGETS) @@ -223,13 +223,13 @@ endforeach() # Skip build if no matching targets found if(NOT GEMM_GPU_TARGETS_INDIVIDUAL) - message(WARNING "Skipping Tile Engine GEMM build: No supported GPU targets (gfx90a, gfx942, gfx950) found in SUPPORTED_GPU_TARGETS: ${SUPPORTED_GPU_TARGETS}") + message(WARNING "Skipping Tile Engine GEMM build: No supported GPU targets (gfx90a, gfx942, gfx950, gfx1201) found in SUPPORTED_GPU_TARGETS: ${SUPPORTED_GPU_TARGETS}") else() message(STATUS "Building individual GEMM targets for GPU targets: ${GEMM_GPU_TARGETS_INDIVIDUAL}") # Enable parallel compilation optimizations # Set up job pools for better parallel compilation control - set_property(GLOBAL PROPERTY JOB_POOLS + set_property(GLOBAL PROPERTY JOB_POOLS compile_heavy=4 # Limit heavy compilations to prevent OOM compile_normal=16 # Allow more parallel normal compilations ) diff --git a/tile_engine/ops/gemm/codegen_utils.py b/tile_engine/ops/gemm/codegen_utils.py index 6a87193043..4f991fa34b 100644 --- a/tile_engine/ops/gemm/codegen_utils.py +++ b/tile_engine/ops/gemm/codegen_utils.py @@ -179,9 +179,14 @@ warp_tile_supported_combinations = { [32, 32, 64], ], }, + "gfx1201": { + "fp16_fp16_fp16": [ + [16, 16, 16], + ], + }, } -# To Do: remove some unsupported combinations +# To Do: remove some unsupp£orted combinations trait_unsupported_combinations = { ("compv3", "cshuffle", "interwave"), ("compv3", "default", "interwave"), diff --git a/tile_engine/ops/gemm/configs/gfx120x_config.json b/tile_engine/ops/gemm/configs/gfx120x_config.json new file mode 100644 index 0000000000..8965063c69 --- /dev/null +++ b/tile_engine/ops/gemm/configs/gfx120x_config.json @@ -0,0 +1,97 @@ +{ + "problem": { + }, + "tile_config": { + "tile_m": { + "max": 256, + "min": 64, + "step": 64 + }, + "tile_n": { + "max": 256, + "min": 64, + "step": 64 + }, + "tile_k": { + "max": 256, + "min": 64, + "step": 64 + }, + "warp_m": { + "values": [ + 4, + 2, + 1 + ] + }, + "warp_n": { + "values": [ + 4, + 2, + 1 + ] + }, + "warp_k": { + "values": [ + 1 + ] + }, + "warp_tile_m": { + "values": [ + 16 + ] + }, + "warp_tile_n": { + "values": [ + 16 + ] + }, + "warp_tile_k": { + "values": [ + 16 + ] + } + }, + "trait_config": { + "pipeline": { + "values": [ + "compv3", + "compv4", + "mem" + ] + }, + "scheduler": { + "values": [ + "intrawave", + "interwave" + ] + }, + "epilogue": { + "values": [ + "cshuffle", + "default" + ] + }, + "pad_m": { + "values": [ + false + ] + }, + "pad_n": { + "values": [ + false + ] + }, + "pad_k": { + "values": [ + false + ] + }, + "persistent": { + "values": [ + false, + true + ] + } + } +}