Merge branch 'develop' into lwpck-976

library/src/tensor_operation_instance/gpu/CMakeLists.txt

@@ -19,7 +19,7 @@ function(add_instance_library INSTANCE_NAME)
                     set(type1 "_i8")
                 endif()
                 #make an exception for reduction kernels
-                if("${source}" MATCHES "${type}" OR "${source}" MATCHES "${type1}" OR "${source}" MATCHES "device_reduce_instance")
+                if("${source}" MATCHES "${type}" OR "${source}" MATCHES "${type1}" OR "${source}" MATCHES "device_reduce_instance" OR ${source} MATCHES "device_image_to_column")
                     #if filename matches any selected type, exit type loop and do no exclude the file from the list
                     set(test 0)
                     break()
@@ -49,8 +49,10 @@ function(add_instance_library INSTANCE_NAME)
         set_target_properties(${INSTANCE_NAME} PROPERTIES POSITION_INDEPENDENT_CODE ON)
         clang_tidy_check(${INSTANCE_NAME})
         set(result 0)
+        message("add_instance_library ${INSTANCE_NAME}")
+    else()
+        message("skip_instance_libary ${INSTANCE_NAME}")
     endif()
-    #message("add_instance_library returns ${result}")
     set(result ${result} PARENT_SCOPE)
 endfunction(add_instance_library INSTANCE_NAME)
 
@@ -58,8 +60,8 @@ endfunction(add_instance_library INSTANCE_NAME)
 file(GLOB dir_list LIST_DIRECTORIES true *)
 set(CK_DEVICE_INSTANCES)
 FOREACH(subdir_path ${dir_list})
-set(target_dir)
-IF(IS_DIRECTORY "${subdir_path}")
+    set(target_dir)
+    IF(IS_DIRECTORY "${subdir_path}")
         set(cmake_instance)
         file(READ "${subdir_path}/CMakeLists.txt" cmake_instance)
         set(add_inst 0)
@@ -87,36 +89,41 @@ IF(IS_DIRECTORY "${subdir_path}")
             message("int8 instance found!")
             set(add_inst 1)
         endif()
-    if(NOT "${cmake_instance}" MATCHES "_fp8" OR
-                   NOT "${cmake_instance}" MATCHES "_f8" OR
-                   NOT "${cmake_instance}" MATCHES "_fp16" OR
-                   NOT "${cmake_instance}" MATCHES "_f16" OR
-                   NOT "${cmake_instance}" MATCHES "_fp32" OR
-                   NOT "${cmake_instance}" MATCHES "_f32" OR
-                   NOT "${cmake_instance}" MATCHES "_fp64" OR
-                   NOT "${cmake_instance}" MATCHES "_f64" OR
-                   NOT "${cmake_instance}" MATCHES "_bf16" OR
-                   NOT "${cmake_instance}" MATCHES "_int8" OR
-                   NOT "${cmake_instance}" MATCHES "_i8" OR
-                   NOT "${cmake_instance}" MATCHES "_int4" OR
-                   NOT DEFINED DTYPES)
+        if(NOT ("${cmake_instance}" MATCHES "_fp8" OR
+                "${cmake_instance}" MATCHES "_f8" OR
+                "${cmake_instance}" MATCHES "_fp16" OR
+                "${cmake_instance}" MATCHES "_f16" OR
+                "${cmake_instance}" MATCHES "_fp32" OR
+                "${cmake_instance}" MATCHES "_f32" OR
+                "${cmake_instance}" MATCHES "_fp64" OR
+                "${cmake_instance}" MATCHES "_f64" OR
+                "${cmake_instance}" MATCHES "_bf16" OR
+                "${cmake_instance}" MATCHES "_int8" OR
+                "${cmake_instance}" MATCHES "_i8" OR
+                "${cmake_instance}" MATCHES "_int4"))
             message("instance should be built for all types!")
             set(add_inst 1)
         endif()
-    if("${cmake_instance}" MATCHES "quantization" AND DEFINED DTYPES AND NOT DTYPES MATCHES "int8")
+        if(NOT DEFINED DTYPES)
+            set(add_inst 1)
+        endif()
+        if(("${cmake_instance}" MATCHES "quantization") AND (DEFINED DTYPES) AND (NOT DTYPES MATCHES "int8"))
             message("quantization instances will not be built!")
             set(add_inst 0)
         endif()
-    if("${cmake_instance}" MATCHES "ONLY DL_KERNELS" AND NOT DEFINED DL_KERNELS)
+        if(("${cmake_instance}" MATCHES "ONLY DL_KERNELS") AND (NOT DEFINED DL_KERNELS))
             message("Found only dl instances, but DL_KERNELS is not set. Skipping.")
             set(add_inst 0)
         endif()
-    if(add_inst EQUAL 1)
+        if((add_inst EQUAL 1))
             get_filename_component(target_dir ${subdir_path} NAME)
             add_subdirectory(${target_dir})
             list(APPEND CK_DEVICE_INSTANCES $<TARGET_OBJECTS:device_${target_dir}_instance>)
+            message("add_instance_directory ${subdir_path}")
+        else()
+            message("skip_instance_directory ${subdir_path}")
         endif()
-ENDIF()
+    ENDIF()
 ENDFOREACH()
 
 add_library(device_operations STATIC ${CK_DEVICE_INSTANCES})

library/src/tensor_operation_instance/gpu/gemm/CMakeLists.txt

@@ -96,13 +96,9 @@ list(APPEND GEMM_INSTANCES device_gemm_xdl_c_shuffle_fp8_fp8_fp8_mk_kn_mn_instan
 
 add_instance_library(device_gemm_instance ${GEMM_INSTANCES})
 
-set(ENABLE_PIPELINE_V2_OPT OFF)
+set(ENABLE_PIPELINE_V2_OPT)
 
 if (ENABLE_PIPELINE_V2_OPT)
-    set(MAX_ILP_OPTS
-       -mllvm
-       -amdgpu-enable-max-ilp-scheduling-strategy
-    )
     set(WAVES_PER_EU_DEFS
        CK_USE_WAVES_PER_EU=1
        CK_MIN_WAVES_PER_EU=1
@@ -118,7 +114,7 @@ if (ENABLE_PIPELINE_V2_OPT)
        COMPILE_DEFINITIONS "${WAVES_PER_EU_DEFS};${IGLP_OPT_DEFS}")
     # layout=NN
     set_source_files_properties(device_gemm_xdl_f16_f16_f16/km_nk_mn_default_pipeline_v2_opt_instance.cpp PROPERTIES
-       COMPILE_OPTIONS "${MAX_ILP_OPTS}"
+       COMPILE_OPTIONS ";;"
        COMPILE_DEFINITIONS "${WAVES_PER_EU_DEFS};${IGLP_OPT_DEFS}")
     # layout=TT
     set_source_files_properties(device_gemm_xdl_f16_f16_f16/mk_kn_mn_default_pipeline_v2_opt_instance.cpp PROPERTIES
@@ -126,7 +122,7 @@ if (ENABLE_PIPELINE_V2_OPT)
        COMPILE_DEFINITIONS "${WAVES_PER_EU_DEFS};${IGLP_OPT_DEFS}")
     # layout=TN
     set_source_files_properties(device_gemm_xdl_f16_f16_f16/mk_nk_mn_default_pipeline_v2_opt_instance.cpp PROPERTIES
-       COMPILE_OPTIONS "${MAX_ILP_OPTS}"
+       COMPILE_OPTIONS ";;"
        COMPILE_DEFINITIONS "${WAVES_PER_EU_DEFS};${IGLP_OPT_DEFS}")
 endif(ENABLE_PIPELINE_V2_OPT)
 

library/src/tensor_operation_instance/gpu/grouped_conv1d_bwd_weight/CMakeLists.txt

 set(GROUPED_CONV1D_BWD_WEIGHT
-    device_grouped_conv1d_bwd_weight_xdl_gnwc_gkxc_gnwk_f16_instance.cpp
-    device_grouped_conv1d_bwd_weight_xdl_gnwc_gkxc_gnwk_f32_instance.cpp
-    device_grouped_conv1d_bwd_weight_xdl_gnwc_gkxc_gnwk_bf16_instance.cpp)
+    xdl/device_grouped_conv1d_bwd_weight_xdl_gnwc_gkxc_gnwk_f16_instance.cpp
+    xdl/device_grouped_conv1d_bwd_weight_xdl_gnwc_gkxc_gnwk_f32_instance.cpp
+    xdl/device_grouped_conv1d_bwd_weight_xdl_gnwc_gkxc_gnwk_bf16_instance.cpp)
 
 if(DL_KERNELS)
     list(APPEND GROUPED_CONV1D_BWD_WEIGHT
-        device_grouped_conv1d_bwd_weight_dl_gnwc_gkxc_gnwk_f16_instance.cpp
-        device_grouped_conv1d_bwd_weight_dl_gnwc_gkxc_gnwk_f32_instance.cpp
-        device_grouped_conv1d_bwd_weight_dl_gnwc_gkxc_gnwk_bf16_instance.cpp
-        device_grouped_conv1d_bwd_weight_dl_nwgc_gkxc_nwgk_f16_instance.cpp
-        device_grouped_conv1d_bwd_weight_dl_nwgc_gkxc_nwgk_f32_instance.cpp
-        device_grouped_conv1d_bwd_weight_dl_nwgc_gkxc_nwgk_bf16_instance.cpp)
+        dl/device_grouped_conv1d_bwd_weight_dl_gnwc_gkxc_gnwk_f16_instance.cpp
+        dl/device_grouped_conv1d_bwd_weight_dl_gnwc_gkxc_gnwk_f32_instance.cpp
+        dl/device_grouped_conv1d_bwd_weight_dl_gnwc_gkxc_gnwk_bf16_instance.cpp
+        dl/device_grouped_conv1d_bwd_weight_dl_nwgc_gkxc_nwgk_f16_instance.cpp
+        dl/device_grouped_conv1d_bwd_weight_dl_nwgc_gkxc_nwgk_f32_instance.cpp
+        dl/device_grouped_conv1d_bwd_weight_dl_nwgc_gkxc_nwgk_bf16_instance.cpp)
 endif()
 
 add_instance_library(device_grouped_conv1d_bwd_weight_instance ${GROUPED_CONV1D_BWD_WEIGHT})

library/src/tensor_operation_instance/gpu/grouped_conv2d_bwd_weight/CMakeLists.txt

 set(GROUPED_CONV2D_BWD_WEIGHT
-    device_grouped_conv2d_bwd_weight_xdl_gnhwc_gkyxc_gnhwk_f16_instance.cpp
-    device_grouped_conv2d_bwd_weight_xdl_gnhwc_gkyxc_gnhwk_f32_instance.cpp
-    device_grouped_conv2d_bwd_weight_xdl_gnhwc_gkyxc_gnhwk_bf16_instance.cpp
-    device_grouped_conv2d_bwd_weight_xdl_nhwgc_gkyxc_nhwgk_f16_instance.cpp
-    device_grouped_conv2d_bwd_weight_xdl_nhwgc_gkyxc_nhwgk_f32_instance.cpp
-    device_grouped_conv2d_bwd_weight_xdl_nhwgc_gkyxc_nhwgk_bf16_instance.cpp)
+    xdl/device_grouped_conv2d_bwd_weight_xdl_gnhwc_gkyxc_gnhwk_f16_instance.cpp
+    xdl/device_grouped_conv2d_bwd_weight_xdl_gnhwc_gkyxc_gnhwk_f32_instance.cpp
+    xdl/device_grouped_conv2d_bwd_weight_xdl_gnhwc_gkyxc_gnhwk_bf16_instance.cpp
+    xdl/device_grouped_conv2d_bwd_weight_xdl_nhwgc_gkyxc_nhwgk_f16_instance.cpp
+    xdl/device_grouped_conv2d_bwd_weight_xdl_nhwgc_gkyxc_nhwgk_f32_instance.cpp
+    xdl/device_grouped_conv2d_bwd_weight_xdl_nhwgc_gkyxc_nhwgk_bf16_instance.cpp)
 
 if(DL_KERNELS)
     list(APPEND GROUPED_CONV2D_BWD_WEIGHT
-        device_grouped_conv2d_bwd_weight_dl_gnhwc_gkyxc_gnhwk_f16_instance.cpp
-        device_grouped_conv2d_bwd_weight_dl_gnhwc_gkyxc_gnhwk_f32_instance.cpp
-        device_grouped_conv2d_bwd_weight_dl_gnhwc_gkyxc_gnhwk_bf16_instance.cpp
-        device_grouped_conv2d_bwd_weight_dl_nhwgc_gkyxc_nhwgk_f16_instance.cpp
-        device_grouped_conv2d_bwd_weight_dl_nhwgc_gkyxc_nhwgk_f32_instance.cpp
-        device_grouped_conv2d_bwd_weight_dl_nhwgc_gkyxc_nhwgk_bf16_instance.cpp)
+        dl/device_grouped_conv2d_bwd_weight_dl_gnhwc_gkyxc_gnhwk_f16_instance.cpp
+        dl/device_grouped_conv2d_bwd_weight_dl_gnhwc_gkyxc_gnhwk_f32_instance.cpp
+        dl/device_grouped_conv2d_bwd_weight_dl_gnhwc_gkyxc_gnhwk_bf16_instance.cpp
+        dl/device_grouped_conv2d_bwd_weight_dl_nhwgc_gkyxc_nhwgk_f16_instance.cpp
+        dl/device_grouped_conv2d_bwd_weight_dl_nhwgc_gkyxc_nhwgk_f32_instance.cpp
+        dl/device_grouped_conv2d_bwd_weight_dl_nhwgc_gkyxc_nhwgk_bf16_instance.cpp)
 endif()
 
 add_instance_library(device_grouped_conv2d_bwd_weight_instance ${GROUPED_CONV2D_BWD_WEIGHT})