Merge pull request #3 from ROCmSoftwarePlatform/downstream-internal

Downstream from public to internal

Merge pull request #3 from ROCmSoftwarePlatform/downstream-internal
Downstream from public to internal
52f64967 · Illia Silin · GitHub · dda3a0a1 · 246ceee4 · 52f64967
Unverified Commit 52f64967 authored Feb 23, 2023 by Illia Silin Committed by GitHub Feb 23, 2023
20 changed files
--- a/example/01_gemm/gemm_xdl_skip_b_lds_fp16.cpp
+++ b/example/01_gemm/gemm_xdl_skip_b_lds_fp16.cpp
--- a/example/01_gemm/gemm_xdl_wavelet_fp16.cpp
+++ b/example/01_gemm/gemm_xdl_wavelet_fp16.cpp
--- a/example/01_gemm/run_gemm_example.inc
+++ b/example/01_gemm/run_gemm_example.inc
--- a/example/02_gemm_bilinear/CMakeLists.txt
+++ b/example/02_gemm_bilinear/CMakeLists.txt
 add_example_executable(example_gemm_bilinear_xdl_fp16 gemm_bilinear_xdl_fp16.cpp)
+if(GPU_TARGETS MATCHES "gfx1100")
+    add_example_executable(example_gemm_bilinear_wmma_fp16 gemm_bilinear_wmma_fp16.cpp)
+endif()
--- a/example/02_gemm_bilinear/gemm_bilinear_wmma_fp16.cpp
+++ b/example/02_gemm_bilinear/gemm_bilinear_wmma_fp16.cpp
--- a/example/02_gemm_bilinear/gemm_bilinear_xdl_fp16.cpp
+++ b/example/02_gemm_bilinear/gemm_bilinear_xdl_fp16.cpp
--- a/example/03_gemm_bias_relu/gemm_bias_relu_xdl_fp16.cpp
+++ b/example/03_gemm_bias_relu/gemm_bias_relu_xdl_fp16.cpp
--- a/example/04_gemm_add_add_fastgelu/run_gemm_add_add_fastgelu_example.inc
+++ b/example/04_gemm_add_add_fastgelu/run_gemm_add_add_fastgelu_example.inc
--- a/example/09_convnd_fwd/CMakeLists.txt
+++ b/example/09_convnd_fwd/CMakeLists.txt
--- a/example/09_convnd_fwd/convnd_fwd_common.hpp
+++ b/example/09_convnd_fwd/convnd_fwd_common.hpp
--- a/example/09_convnd_fwd/convnd_fwd_dl_common.hpp
+++ b/example/09_convnd_fwd/convnd_fwd_dl_common.hpp
--- a/example/09_convnd_fwd/convnd_fwd_dl_fp16.cpp
+++ b/example/09_convnd_fwd/convnd_fwd_dl_fp16.cpp
--- a/example/09_convnd_fwd/convnd_fwd_dl_fp32.cpp
+++ b/example/09_convnd_fwd/convnd_fwd_dl_fp32.cpp
--- a/example/09_convnd_fwd/convnd_fwd_dl_int8.cpp
+++ b/example/09_convnd_fwd/convnd_fwd_dl_int8.cpp
--- a/example/09_convnd_fwd/run_convnd_fwd_dl_example.inc
+++ b/example/09_convnd_fwd/run_convnd_fwd_dl_example.inc
--- a/example/10_convnd_fwd_multiple_d_multiple_reduce/common.hpp
+++ b/example/10_convnd_fwd_multiple_d_multiple_reduce/common.hpp
--- a/example/10_convnd_fwd_multiple_d_multiple_reduce/run_convnd_fwd_max_example.inc
+++ b/example/10_convnd_fwd_multiple_d_multiple_reduce/run_convnd_fwd_max_example.inc
--- a/example/12_reduce/reduce_blockwise.cpp
+++ b/example/12_reduce/reduce_blockwise.cpp
--- a/example/12_reduce/reduce_blockwise_impl.hpp
+++ b/example/12_reduce/reduce_blockwise_impl.hpp
--- a/example/12_reduce/reduce_blockwise_two_call.cpp
+++ b/example/12_reduce/reduce_blockwise_two_call.cpp