Ported xdlops kernels to debug bwdwrw fp32/fp16/bfp16 issue. Verified atleast fwd data fp32 works.

32850b93 · Wen-Heng (Jack) Chung · 583755a7 · 32850b93 · 32850b93 · 32850b93
Commit 32850b93 authored Oct 09, 2019 by Wen-Heng (Jack) Chung
17 changed files
--- a/composable_kernel/include/utility/common_header.hpp
+++ b/composable_kernel/include/utility/common_header.hpp
 #ifndef CK_COMMON_HEADER_HPP
 #define CK_COMMON_HEADER_HPP
-#define MIOPEN_USE_FP16 1
+#define MIOPEN_USE_FP16 0
 #define MIOPEN_USE_BFP16 0
-#define MIOPEN_USE_FP32 0
+#define MIOPEN_USE_FP32 1
 #define __HIP_PLATFORM_HCC__ 1

--- a/composable_kernel/include/utility/config.hpp.bkup
+++ b/composable_kernel/include/utility/config.hpp.bkup
--- a/composable_kernel/include/utility/config_amd.hpp.in
+++ b/composable_kernel/include/utility/config_amd.hpp.in
--- a/composable_kernel/include/utility/config_nvidia.hpp.in
+++ b/composable_kernel/include/utility/config_nvidia.hpp.in
--- a/composable_kernel/include/utility/functional2.hpp
+++ b/composable_kernel/include/utility/functional2.hpp
--- a/composable_kernel/include/utility/functional3.hpp
+++ b/composable_kernel/include/utility/functional3.hpp
--- a/composable_kernel/include/utility/integral_constant.hpp
+++ b/composable_kernel/include/utility/integral_constant.hpp
--- a/composable_kernel/include/utility/math.hpp
+++ b/composable_kernel/include/utility/math.hpp
--- a/composable_kernel/include/utility/vector_type.hpp
+++ b/composable_kernel/include/utility/vector_type.hpp
--- a/composable_kernel/src/kernel_wrapper/gridwise_convolution_implicit_gemm_v4_nchw_kc1x1_nkhw_lds_double_buffer.cpp
+++ b/composable_kernel/src/kernel_wrapper/gridwise_convolution_implicit_gemm_v4_nchw_kc1x1_nkhw_lds_double_buffer.cpp
--- a/composable_kernel/src/kernel_wrapper/gridwise_convolution_implicit_gemm_v4_nchw_kcyx_nkhw_lds_double_buffer.cpp
+++ b/composable_kernel/src/kernel_wrapper/gridwise_convolution_implicit_gemm_v4_nchw_kcyx_nkhw_lds_double_buffer.cpp
--- a/composable_kernel/src/kernel_wrapper/gridwise_convolution_implicit_gemm_v4r4_xdlops_nchw_kc1x1_nkhw_lds_double_buffer.cpp
+++ b/composable_kernel/src/kernel_wrapper/gridwise_convolution_implicit_gemm_v4r4_xdlops_nchw_kc1x1_nkhw_lds_double_buffer.cpp
--- a/composable_kernel/src/kernel_wrapper/gridwise_convolution_implicit_gemm_v4r4_xdlops_nchw_kcyx_nkhw_lds_double_buffer.cpp
+++ b/composable_kernel/src/kernel_wrapper/gridwise_convolution_implicit_gemm_v4r4_xdlops_nchw_kcyx_nkhw_lds_double_buffer.cpp
--- a/driver/include/conv_common.hpp
+++ b/driver/include/conv_common.hpp
@@ -32,6 +32,7 @@ constexpr auto get_convolution_output_default_4d_tensor_descriptor(InDesc, WeiDe
    constexpr auto HO = HI + 1 - Y;
    constexpr auto WO = WI + 1 - X;
+    printf("H0=%d, W0=%d\n", HO, WO);
    return make_ConstantTensorDescriptor_packed(Sequence<N, K, HO, WO>{});
 }

--- a/driver/include/device_convolution_implicit_gemm_v4_nchw_kcyx_nkhw.hpp
+++ b/driver/include/device_convolution_implicit_gemm_v4_nchw_kcyx_nkhw.hpp
--- a/driver/include/device_convolution_implicit_gemm_v5_nchw_kcyx_nkhw.hpp
+++ b/driver/include/device_convolution_implicit_gemm_v5_nchw_kcyx_nkhw.hpp
--- a/driver/src/driver.cpp
+++ b/driver/src/driver.cpp