Merge branch 'develop' into wavelet_model

b89a88b5 · Adam Osewski · 41d5fca7 · 43c898f6 · b89a88b5 · b89a88b5
Commit b89a88b5 authored Sep 19, 2022 by Adam Osewski
20 changed files
--- a/example/16_gemm_multi_d_multi_reduces/gemm_max_xdl_fp32.cpp
+++ b/example/16_gemm_multi_d_multi_reduces/gemm_max_xdl_fp32.cpp
--- a/example/16_gemm_multi_d_multi_reduces/gemm_max_xdl_int4.cpp
+++ b/example/16_gemm_multi_d_multi_reduces/gemm_max_xdl_int4.cpp
--- a/example/16_gemm_multi_d_multi_reduces/gemm_max_xdl_int8.cpp
+++ b/example/16_gemm_multi_d_multi_reduces/gemm_max_xdl_int8.cpp
--- a/example/16_gemm_multi_d_multi_reduces/gemm_mean_meansquare_xdl_bf16.cpp
+++ b/example/16_gemm_multi_d_multi_reduces/gemm_mean_meansquare_xdl_bf16.cpp
--- a/example/16_gemm_multi_d_multi_reduces/gemm_mean_meansquare_xdl_fp16.cpp
+++ b/example/16_gemm_multi_d_multi_reduces/gemm_mean_meansquare_xdl_fp16.cpp
--- a/example/16_gemm_multi_d_multi_reduces/gemm_mean_meansquare_xdl_fp32.cpp
+++ b/example/16_gemm_multi_d_multi_reduces/gemm_mean_meansquare_xdl_fp32.cpp
--- a/example/16_gemm_multi_d_multi_reduces/gemm_reduce_xdl_common.hpp
+++ b/example/16_gemm_multi_d_multi_reduces/gemm_reduce_xdl_common.hpp
--- a/example/19_binary_elementwise/broadcast_add_2d_amn_bn.cpp
+++ b/example/19_binary_elementwise/broadcast_add_2d_amn_bn.cpp
--- a/example/19_binary_elementwise/broadcast_add_3d_am_bmnk.cpp
+++ b/example/19_binary_elementwise/broadcast_add_3d_am_bmnk.cpp
--- a/example/19_binary_elementwise/elementwise_add_1d.cpp
+++ b/example/19_binary_elementwise/elementwise_add_1d.cpp
--- a/example/19_binary_elementwise/elementwise_add_4d.cpp
+++ b/example/19_binary_elementwise/elementwise_add_4d.cpp
--- a/example/21_gemm_layernorm/gemm_bias_relu_add_layernorm_xdl_fp16.cpp
+++ b/example/21_gemm_layernorm/gemm_bias_relu_add_layernorm_xdl_fp16.cpp
--- a/example/21_gemm_layernorm/gemm_layernorm_xdl_fp16.cpp
+++ b/example/21_gemm_layernorm/gemm_layernorm_xdl_fp16.cpp
--- a/example/22_cgemm/CMakeLists.txt
+++ b/example/22_cgemm/CMakeLists.txt
@@ -5,7 +5,13 @@ add_example_executable(example_cgemm_xdl_fp16 cgemm_xdl_fp16.cpp)
 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_bf16)
+add_dependencies(example_cgemm_xdl 
-add_dependencies(example_cgemm_xdl example_cgemm_xdl_fp16)
+                 example_cgemm_xdl_bf16
-add_dependencies(example_cgemm_xdl example_cgemm_xdl_fp32)
+                 example_cgemm_xdl_fp16
-add_dependencies(example_cgemm_xdl example_cgemm_xdl_int8)
+                 example_cgemm_xdl_fp32
+                 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)
+endif()
--- 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