merge from develop and revisison for pr#881

29448ffd · Harisankar Sadasivan · 9223a5e2 · 8f84a012 · 29448ffd · 29448ffd
Commit 29448ffd authored Sep 08, 2023 by Harisankar Sadasivan
20 changed files
--- a/example/21_gemm_layernorm/gemm_bias_relu_add_layernorm_xdl_welford_fp16.cpp
+++ b/example/21_gemm_layernorm/gemm_bias_relu_add_layernorm_xdl_welford_fp16.cpp
--- a/example/21_gemm_layernorm/gemm_layernorm_xdl_naive_fp16.cpp
+++ b/example/21_gemm_layernorm/gemm_layernorm_xdl_naive_fp16.cpp
--- a/example/21_gemm_layernorm/gemm_xdl_layernorm_naive_single_kernel_fp16.cpp
+++ b/example/21_gemm_layernorm/gemm_xdl_layernorm_naive_single_kernel_fp16.cpp
--- a/example/22_cgemm/CMakeLists.txt
+++ b/example/22_cgemm/CMakeLists.txt
 add_custom_target(example_cgemm_xdl)
-add_example_executable(example_cgemm_xdl_bf16 cgemm_xdl_bf16.cpp)
+if(DTYPES MATCHES "bf16" OR NOT DEFINED DTYPES)
-add_example_executable(example_cgemm_xdl_fp16 cgemm_xdl_fp16.cpp)
+  add_example_executable(example_cgemm_xdl_bf16 cgemm_xdl_bf16.cpp)
+  add_dependencies(example_cgemm_xdl example_cgemm_xdl_bf16)
+endif()
+if(DTYPES MATCHES "fp16" OR NOT DEFINED DTYPES)
+  add_example_executable(example_cgemm_xdl_fp16 cgemm_xdl_fp16.cpp)
+  add_dependencies(example_cgemm_xdl example_cgemm_xdl_fp16)
+endif()
+if(DTYPES MATCHES "fp32" OR NOT DEFINED DTYPES)
 add_example_executable(example_cgemm_xdl_fp32 cgemm_xdl_fp32.cpp)
-add_example_executable(example_cgemm_xdl_int8 cgemm_xdl_int8.cpp)
+add_dependencies(example_cgemm_xdl example_cgemm_xdl_fp32)
+endif()
-add_dependencies(example_cgemm_xdl 
+if(DTYPES MATCHES "int8" OR NOT DEFINED DTYPES)
-                 example_cgemm_xdl_bf16
+  add_example_executable(example_cgemm_xdl_int8 cgemm_xdl_int8.cpp)
-                 example_cgemm_xdl_fp16
+  add_dependencies(example_cgemm_xdl example_cgemm_xdl_int8)
-                 example_cgemm_xdl_fp32
+endif()
-                 example_cgemm_xdl_int8)
 if(USE_BITINT_EXTENSION_INT4)
  add_example_executable(example_cgemm_xdl_int4 cgemm_xdl_int4.cpp)
  add_dependencies(example_cgemm_xdl example_cgemm_xdl_int4)

--- a/example/22_cgemm/cgemm_xdl_bf16.cpp
+++ b/example/22_cgemm/cgemm_xdl_bf16.cpp
--- a/example/22_cgemm/cgemm_xdl_common.hpp
+++ b/example/22_cgemm/cgemm_xdl_common.hpp
--- a/example/22_cgemm/cgemm_xdl_fp16.cpp
+++ b/example/22_cgemm/cgemm_xdl_fp16.cpp
--- a/example/22_cgemm/cgemm_xdl_fp32.cpp
+++ b/example/22_cgemm/cgemm_xdl_fp32.cpp
--- a/example/22_cgemm/cgemm_xdl_int4.cpp
+++ b/example/22_cgemm/cgemm_xdl_int4.cpp
--- a/example/22_cgemm/cgemm_xdl_int8.cpp
+++ b/example/22_cgemm/cgemm_xdl_int8.cpp
--- a/example/23_softmax/CMakeLists.txt
+++ b/example/23_softmax/CMakeLists.txt
--- a/example/23_softmax/README.md
+++ b/example/23_softmax/README.md
--- a/example/23_softmax/softmax_blockwise.cpp
+++ b/example/23_softmax/softmax_blockwise.cpp
--- a/example/24_batched_gemm/CMakeLists.txt
+++ b/example/24_batched_gemm/CMakeLists.txt
 add_custom_target(example_batched_gemm_xdl)
+if(DTYPES MATCHES "fp32" OR NOT DEFINED DTYPES)
-add_example_executable(example_batched_gemm_xdl_fp32 batched_gemm_xdl_fp32.cpp)
+  add_example_executable(example_batched_gemm_xdl_fp32 batched_gemm_xdl_fp32.cpp)
-add_example_executable(example_batched_gemm_xdl_fp16 batched_gemm_xdl_fp16.cpp)
+  add_dependencies(example_batched_gemm_xdl example_batched_gemm_xdl_fp32)
-add_example_executable(example_batched_gemm_xdl_bfp16 batched_gemm_xdl_bfp16.cpp)
+endif()
-add_example_executable(example_batched_gemm_xdl_int8 batched_gemm_xdl_int8.cpp)
+if(DTYPES MATCHES "fp16" OR NOT DEFINED DTYPES)
+  add_example_executable(example_batched_gemm_xdl_fp16 batched_gemm_xdl_fp16.cpp)
-add_dependencies(example_batched_gemm_xdl
+  add_dependencies(example_batched_gemm_xdl example_batched_gemm_xdl_fp16)
-                 example_batched_gemm_xdl_fp32
+endif()
-                 example_batched_gemm_xdl_fp16
+if(DTYPES MATCHES "bf16" OR NOT DEFINED DTYPES)
-                 example_batched_gemm_xdl_bfp16
+  add_example_executable(example_batched_gemm_xdl_bfp16 batched_gemm_xdl_bfp16.cpp)
-                 example_batched_gemm_xdl_int8)
+  add_dependencies(example_batched_gemm_xdl example_batched_gemm_xdl_bfp16)
+endif()
+if(DTYPES MATCHES "int8" OR NOT DEFINED DTYPES)
+  add_example_executable(example_batched_gemm_xdl_int8 batched_gemm_xdl_int8.cpp)
+  add_dependencies(example_batched_gemm_xdl example_batched_gemm_xdl_int8)
+endif()
 if(USE_BITINT_EXTENSION_INT4)
  add_example_executable(example_batched_gemm_xdl_int4 batched_gemm_xdl_int4.cpp)
  add_dependencies(example_batched_gemm_xdl example_batched_gemm_xdl_int4)

--- a/example/24_batched_gemm/batched_gemm_xdl_bfp16.cpp
+++ b/example/24_batched_gemm/batched_gemm_xdl_bfp16.cpp
--- a/example/24_batched_gemm/batched_gemm_xdl_fp16.cpp
+++ b/example/24_batched_gemm/batched_gemm_xdl_fp16.cpp
--- a/example/24_batched_gemm/batched_gemm_xdl_fp32.cpp
+++ b/example/24_batched_gemm/batched_gemm_xdl_fp32.cpp
--- a/example/24_batched_gemm/batched_gemm_xdl_int4.cpp
+++ b/example/24_batched_gemm/batched_gemm_xdl_int4.cpp
--- a/example/24_batched_gemm/batched_gemm_xdl_int8.cpp
+++ b/example/24_batched_gemm/batched_gemm_xdl_int8.cpp
--- a/example/24_batched_gemm/run_batched_gemm_example.inc
+++ b/example/24_batched_gemm/run_batched_gemm_example.inc