Skip to content

GitLab

Unverified Commit 52c3fe05 authored by Chao Liu's avatar Chao Liu Committed by GitHub
Browse files

Refactor for MIOpen integration (#4) 

Refactor, so can bring multi-index transformation and padding support into MIOpen
parent 9aaeacc8
Expand all Hide whitespace changes
Inline Side-by-side
Showing with 901 additions and 781 deletions
CMakeLists.txt
View file @ 52c3fe05
......@@ -47,14 +47,17 @@ include_directories(BEFORE
${PROJECT_SOURCE_DIR}/composable_kernel/include/tensor_description
${PROJECT_SOURCE_DIR}/composable_kernel/include/tensor_operation
${PROJECT_SOURCE_DIR}/composable_kernel/include/kernel_algorithm
${PROJECT_SOURCE_DIR}/external/include
${PROJECT_SOURCE_DIR}/driver/include
${PROJECT_BINARY_DIR}/composable_kernel/include/utility
)
if(DEVICE_BACKEND STREQUAL "AMD")
configure_file("${PROJECT_SOURCE_DIR}/composable_kernel/include/utility/config_amd.hpp.in" "${PROJECT_BINARY_DIR}/composable_kernel/include/utility/config.hpp")
configure_file("${PROJECT_SOURCE_DIR}/composable_kernel/include/utility/config.amd.hpp.in" "${PROJECT_BINARY_DIR}/composable_kernel/include/utility/config.hpp")
configure_file("${PROJECT_SOURCE_DIR}/composable_kernel/include/utility/float_type.amd.hpp.in" "${PROJECT_BINARY_DIR}/composable_kernel/include/utility/float_type.hpp")
elseif(DEVICE_BACKEND STREQUAL "NVIDIA")
configure_file("${PROJECT_SOURCE_DIR}/composable_kernel/include/utility/config_nvidia.hpp.in" "${PROJECT_BINARY_DIR}/composable_kernel/include/utility/config.hpp")
configure_file("${PROJECT_SOURCE_DIR}/composable_kernel/include/utility/config.nvidia.hpp.in" "${PROJECT_BINARY_DIR}/composable_kernel/include/utility/config.hpp")
configure_file("${PROJECT_SOURCE_DIR}/composable_kernel/include/utility/float_type.nvidia.hpp.in" "${PROJECT_BINARY_DIR}/composable_kernel/include/utility/float_type.hpp")
endif()
add_subdirectory(driver)
composable_kernel/include/kernel_algorithm/convolution_common.hpp 0 → 100644
View file @ 52c3fe05
#ifndef CK_CONVOLUTION_COMMON_HPP
#define CK_CONVOLUTION_COMMON_HPP
namespace ck {
enum ConvolutionDirection
{
Forward,
BackwardData,
BackwardWeight
};
} // namespace ck
#endif
composable_kernel/include/kernel_algorithm/gridwise_convolution_direct_v2_nchw_kcyx_nkhw.hpp
View file @ 52c3fe05
......@@ -2,7 +2,7 @@
#define CK_GRIDWISE_CONVOLUTION_DIRECT_V2_NCHW_KCYX_NKHW
#include "common_header.hpp"
#include "ConstantTensorDescriptor.hpp"
#include "ConstantTensorDescriptor_deprecated.hpp"
#include "blockwise_2d_tensor_op.hpp"
#include "blockwise_4d_tensor_op.hpp"
#include "threadwise_tensor_slice_copy.hpp"
......
composable_kernel/include/kernel_algorithm/gridwise_convolution_implicit_gemm_v1r1_chwn_cyxk_khwn.hpp
View file @ 52c3fe05
......@@ -2,7 +2,7 @@
#define CK_GRIDWISE_CONVOLUTION_IMPLICIT_GEMM_V1R1_CHWN_CYXK_KHWN
#include "common_header.hpp"
#include "ConstantTensorDescriptor.hpp"
#include "ConstantTensorDescriptor_deprecated.hpp"
#include "ConstantMatrixDescriptor.hpp"
#include "blockwise_4d_tensor_op.hpp"
#include "blockwise_2d_tensor_op.hpp"
......
composable_kernel/include/kernel_algorithm/gridwise_convolution_implicit_gemm_v1r2_chwn_cyxk_khwn.hpp
View file @ 52c3fe05
......@@ -2,7 +2,7 @@
#define CK_GRIDWISE_CONVOLUTION_IMPLICIT_GEMM_V1R2_CHWN_CYXK_KHWN
#include "common_header.hpp"
#include "ConstantTensorDescriptor.hpp"
#include "ConstantTensorDescriptor_deprecated.hpp"
#include "ConstantMatrixDescriptor.hpp"
#include "blockwise_2d_tensor_op.hpp"
#include "blockwise_3d_tensor_op.hpp"
......
composable_kernel/include/kernel_algorithm/gridwise_convolution_implicit_gemm_v1r3_chwn_cyxk_khwn.hpp
View file @ 52c3fe05
......@@ -2,7 +2,7 @@
#define CK_GRIDWISE_CONVOLUTION_IMPLICIT_GEMM_V1R3_CHWN_CYXK_KHWN_HPP
#include "common_header.hpp"
#include "ConstantTensorDescriptor.hpp"
#include "ConstantTensorDescriptor_deprecated.hpp"
#include "ConstantMatrixDescriptor.hpp"
#include "blockwise_generic_tensor_slice_copy.hpp"
#include "threadwise_generic_tensor_slice_copy.hpp"
......@@ -125,38 +125,38 @@ struct GridwiseConvolutionImplicitGemm_v1r3_chwn_cyxk_khwn
// blockwise copy
// input: format is [C, Hi, Wi, N]
auto blockwise_in_copy =
BlockwiseGenericTensorSliceCopy_v1<BlockSize,
decltype(in_c_h_w_n_global_desc),
decltype(in_c_h_w_n_block_desc),
decltype(in_c_h_w_n_block_desc.GetLengths()),
InBlockCopySubLengths_CHWN,
InBlockCopyClusterLengths_CHWN,
Sequence<0, 1, 2, 3>,
Sequence<0, 1, 2, 3>,
Sequence<0, 1, 2, 3>,
3,
3,
InBlockCopyDataPerAccess_N,
InBlockCopyDataPerAccess_N>({0, 0, 0, 0},
{0, 0, 0, 0});
auto blockwise_in_copy = BlockwiseGenericTensorSliceCopy_v1_deprecated<
BlockSize,
decltype(in_c_h_w_n_global_desc),
decltype(in_c_h_w_n_block_desc),
decltype(in_c_h_w_n_block_desc.GetLengths()),
InBlockCopySubLengths_CHWN,
InBlockCopyClusterLengths_CHWN,
Sequence<0, 1, 2, 3>,
Sequence<0, 1, 2, 3>,
Sequence<0, 1, 2, 3>,
3,
3,
InBlockCopyDataPerAccess_N,
InBlockCopyDataPerAccess_N>({0, 0, 0, 0}, {0, 0, 0, 0});
// blockwise wei copy
// format is [CPerBlock, X * KPerBlock]
const auto blockwise_wei_copy =
BlockwiseGenericTensorSliceCopy_v1<BlockSize,
decltype(wei_c_k_global_desc),
decltype(wei_c_k_block_desc),
decltype(wei_c_k_block_desc.GetLengths()),
WeiBlockCopySubLengths_CK,
WeiBlockCopyClusterLengths_CK,
Sequence<0, 1>,
Sequence<0, 1>,
Sequence<0, 1>,
1,
1,
WeiBlockCopyDataPerAccess_K,
WeiBlockCopyDataPerAccess_K>({0, 0}, {0, 0});
BlockwiseGenericTensorSliceCopy_v1_deprecated<BlockSize,
decltype(wei_c_k_global_desc),
decltype(wei_c_k_block_desc),
decltype(wei_c_k_block_desc.GetLengths()),
WeiBlockCopySubLengths_CK,
WeiBlockCopyClusterLengths_CK,
Sequence<0, 1>,
Sequence<0, 1>,
Sequence<0, 1>,
1,
1,
WeiBlockCopyDataPerAccess_K,
WeiBlockCopyDataPerAccess_K>({0, 0},
{0, 0});
// a series of blockwise batched GEMM
// C_matrix += transpose(A_matrix) * B_matrix
......@@ -318,14 +318,15 @@ struct GridwiseConvolutionImplicitGemm_v1r3_chwn_cyxk_khwn
n_block_data_begin + n_thread_data_begin);
#if 1
ThreadwiseGenericTensorSliceCopy_v1r2<decltype(out_10d_thread_desc),
decltype(out_10d_global_desc),
decltype(out_10d_thread_desc.GetLengths()),
arithmetic_sequence_gen<0, 10, 1>::type,
9,
OutThreadCopyDataPerAccess_N,
OutThreadCopyDataPerAccess_N>(
make_zero_array<index_t, 10>(), make_zero_array<index_t, 10>())
ThreadwiseGenericTensorSliceCopy_v1r2_deprecated<
decltype(out_10d_thread_desc),
decltype(out_10d_global_desc),
decltype(out_10d_thread_desc.GetLengths()),
arithmetic_sequence_gen<0, 10, 1>::type,
9,
OutThreadCopyDataPerAccess_N,
OutThreadCopyDataPerAccess_N>(make_zero_array<index_t, 10>(),
make_zero_array<index_t, 10>())
.Run(p_out_thread, p_out_thread_on_global);
#elif 0
ThreadwiseGenericTensorSliceCopy_v1r1<decltype(out_10d_thread_desc),
......@@ -388,14 +389,15 @@ struct GridwiseConvolutionImplicitGemm_v1r3_chwn_cyxk_khwn
n_block_data_begin + n_thread_data_begin);
#if 1
ThreadwiseGenericTensorSliceCopy_v1r2<decltype(out_10d_thread_desc),
decltype(out_10d_global_desc),
decltype(out_10d_thread_desc.GetLengths()),
arithmetic_sequence_gen<0, 10, 1>::type,
9,
OutThreadCopyDataPerAccess_N,
OutThreadCopyDataPerAccess_N>(
make_zero_array<index_t, 10>(), make_zero_array<index_t, 10>())
ThreadwiseGenericTensorSliceCopy_v1r2_deprecated<
decltype(out_10d_thread_desc),
decltype(out_10d_global_desc),
decltype(out_10d_thread_desc.GetLengths()),
arithmetic_sequence_gen<0, 10, 1>::type,
9,
OutThreadCopyDataPerAccess_N,
OutThreadCopyDataPerAccess_N>(make_zero_array<index_t, 10>(),
make_zero_array<index_t, 10>())
.Run(p_out_thread, p_out_thread_on_global);
#elif 0
ThreadwiseGenericTensorSliceCopy_v1r1<decltype(out_10d_thread_desc),
......
composable_kernel/include/kernel_algorithm/gridwise_convolution_implicit_gemm_v1r3_chwn_cyxk_khwn_lds_double_buffer.hpp
View file @ 52c3fe05
......@@ -2,7 +2,7 @@
#define CK_GRIDWISE_CONVOLUTION_IMPLICIT_GEMM_V1R3_CHWN_CYXK_KHWN_LDS_DOUBLE_BUFFER_HPP
#include "common_header.hpp"
#include "ConstantTensorDescriptor.hpp"
#include "ConstantTensorDescriptor_deprecated.hpp"
#include "ConstantMatrixDescriptor.hpp"
#include "blockwise_generic_tensor_slice_copy.hpp"
#include "threadwise_generic_tensor_slice_copy.hpp"
......@@ -127,9 +127,9 @@ struct GridwiseConvolutionImplicitGemm_v1r3_chwn_cyxk_khwn_lds_double_buffer
// input: format is [C, Hi, Wi, N]
auto blockwise_in_copy =
#if 0
BlockwiseGenericTensorSliceCopy_v1
BlockwiseGenericTensorSliceCopy_v1_deprecated
#else
BlockwiseGenericTensorSliceCopy_v2
BlockwiseGenericTensorSliceCopy_v2_deprecated
#endif
<BlockSize,
decltype(in_c_h_w_n_global_desc),
......@@ -149,9 +149,9 @@ struct GridwiseConvolutionImplicitGemm_v1r3_chwn_cyxk_khwn_lds_double_buffer
// format is [CPerBlock, X * KPerBlock]
const auto blockwise_wei_copy =
#if 0
BlockwiseGenericTensorSliceCopy_v1
BlockwiseGenericTensorSliceCopy_v1_deprecated
#else
BlockwiseGenericTensorSliceCopy_v2
BlockwiseGenericTensorSliceCopy_v2_deprecated
#endif
<BlockSize,
decltype(wei_c_k_global_desc),
......@@ -406,14 +406,15 @@ struct GridwiseConvolutionImplicitGemm_v1r3_chwn_cyxk_khwn_lds_double_buffer
n_block_data_begin + n_thread_data_begin);
#if 1
ThreadwiseGenericTensorSliceCopy_v1r2<decltype(out_10d_thread_desc),
decltype(out_10d_global_desc),
decltype(out_10d_thread_desc.GetLengths()),
arithmetic_sequence_gen<0, 10, 1>::type,
9,
OutThreadCopyDataPerAccess_N,
OutThreadCopyDataPerAccess_N>(
make_zero_array<index_t, 10>(), make_zero_array<index_t, 10>())
ThreadwiseGenericTensorSliceCopy_v1r2_deprecated<
decltype(out_10d_thread_desc),
decltype(out_10d_global_desc),
decltype(out_10d_thread_desc.GetLengths()),
arithmetic_sequence_gen<0, 10, 1>::type,
9,
OutThreadCopyDataPerAccess_N,
OutThreadCopyDataPerAccess_N>(make_zero_array<index_t, 10>(),
make_zero_array<index_t, 10>())
.Run(p_out_thread, p_out_thread_on_global);
#elif 0
ThreadwiseGenericTensorSliceCopy_v1r1<decltype(out_10d_thread_desc),
......@@ -476,14 +477,15 @@ struct GridwiseConvolutionImplicitGemm_v1r3_chwn_cyxk_khwn_lds_double_buffer
n_block_data_begin + n_thread_data_begin);
#if 1
ThreadwiseGenericTensorSliceCopy_v1r2<decltype(out_10d_thread_desc),
decltype(out_10d_global_desc),
decltype(out_10d_thread_desc.GetLengths()),
arithmetic_sequence_gen<0, 10, 1>::type,
9,
OutThreadCopyDataPerAccess_N,
OutThreadCopyDataPerAccess_N>(
make_zero_array<index_t, 10>(), make_zero_array<index_t, 10>())
ThreadwiseGenericTensorSliceCopy_v1r2_deprecated<
decltype(out_10d_thread_desc),
decltype(out_10d_global_desc),
decltype(out_10d_thread_desc.GetLengths()),
arithmetic_sequence_gen<0, 10, 1>::type,
9,
OutThreadCopyDataPerAccess_N,
OutThreadCopyDataPerAccess_N>(make_zero_array<index_t, 10>(),
make_zero_array<index_t, 10>())
.Run(p_out_thread, p_out_thread_on_global);
#elif 0
ThreadwiseGenericTensorSliceCopy_v1r1<decltype(out_10d_thread_desc),
......
composable_kernel/include/kernel_algorithm/gridwise_convolution_implicit_gemm_v1r3_chwn_cyxk_khwn_padded.hpp
View file @ 52c3fe05
......@@ -2,7 +2,7 @@
#define CK_GRIDWISE_CONVOLUTION_IMPLICIT_GEMM_V1R3_CHWN_CYXK_KHWN_PADDED_HPP
#include "common_header.hpp"
#include "ConstantTensorDescriptor.hpp"
#include "ConstantTensorDescriptor_deprecated.hpp"
#include "ConstantMatrixDescriptor.hpp"
#include "tensor_descriptor.hpp"
#include "tensor_descriptor_helper.hpp"
......
composable_kernel/include/kernel_algorithm/gridwise_convolution_implicit_gemm_v1r3_nchw_cyxk_nkhw.hpp
View file @ 52c3fe05
......@@ -2,7 +2,7 @@
#define CK_GRIDWISE_CONVOLUTION_IMPLICIT_GEMM_V1R3_NCHW_CYXK_NKHW
#include "common_header.hpp"
#include "ConstantTensorDescriptor.hpp"
#include "ConstantTensorDescriptor_deprecated.hpp"
#include "ConstantMatrixDescriptor.hpp"
#include "blockwise_2d_tensor_op.hpp"
#include "blockwise_tensor_slice_copy.hpp"
......
composable_kernel/include/kernel_algorithm/gridwise_convolution_implicit_gemm_v1r3_nchw_cyxk_nkhw_lds_double_buffer.hpp
View file @ 52c3fe05
......@@ -2,7 +2,7 @@
#define CK_GRIDWISE_CONVOLUTION_IMPLICIT_GEMM_V1R3_NCHW_CYXK_NKHW_LDS_DOUBLE_BUFFER
#include "common_header.hpp"
#include "ConstantTensorDescriptor.hpp"
#include "ConstantTensorDescriptor_deprecated.hpp"
#include "ConstantMatrixDescriptor.hpp"
#include "blockwise_2d_tensor_op.hpp"
#include "blockwise_tensor_slice_copy.hpp"
......
composable_kernel/include/kernel_algorithm/gridwise_convolution_implicit_gemm_v2_chwn_cyxk_khwn.hpp
View file @ 52c3fe05
......@@ -2,7 +2,7 @@
#define CK_GRIDWISE_CONVOLUTION_IMPLICIT_GEMM_V2_CHWN_CYXK_KHWN
#include "common_header.hpp"
#include "ConstantTensorDescriptor.hpp"
#include "ConstantTensorDescriptor_deprecated.hpp"
#include "ConstantMatrixDescriptor.hpp"
#include "blockwise_4d_tensor_op.hpp"
#include "blockwise_2d_tensor_op.hpp"
......
composable_kernel/include/kernel_algorithm/gridwise_convolution_implicit_gemm_v2_chwn_cyxk_khwn_lds_double_buffer.hpp
View file @ 52c3fe05
......@@ -2,7 +2,7 @@
#define CK_GRIDWISE_CONVOLUTION_IMPLICIT_GEMM_V2_CHWN_CYXK_KHWN_LDS_DOUBLE_BUFFER
#include "common_header.hpp"
#include "ConstantTensorDescriptor.hpp"
#include "ConstantTensorDescriptor_deprecated.hpp"
#include "ConstantMatrixDescriptor.hpp"
#include "blockwise_4d_tensor_op.hpp"
#include "blockwise_2d_tensor_op.hpp"
......
composable_kernel/include/kernel_algorithm/gridwise_convolution_implicit_gemm_v3_nchw_cyxk_nkhw.hpp
View file @ 52c3fe05
......@@ -2,8 +2,8 @@
#define CK_GRIDWISE_CONVOLUTION_IMPLICIT_GEMM_V3_NCHW_CYXK_NKHW
#include "common_header.hpp"
#include "ConstantTensorDescriptor.hpp"
#include "ConstantMergedTensorDescriptor.hpp"
#include "ConstantTensorDescriptor_deprecated.hpp"
#include "ConstantMergedTensorDescriptor_deprecated.hpp"
#include "ConstantMatrixDescriptor.hpp"
#include "blockwise_generic_tensor_slice_copy.hpp"
#include "blockwise_gemm.hpp"
......@@ -128,7 +128,7 @@ struct GridwiseConvolutionImplicitGemm_v3_nchw_cyxk_nkhw
// input blockwise copy
// slice a merged tensor, reorder and copy to a normal tensor
// this copy operator already has blockwise offset built-in
auto blockwise_in_copy = BlockwiseGenericTensorSliceCopy_v1<
auto blockwise_in_copy = BlockwiseGenericTensorSliceCopy_v1_deprecated<
BlockSize,
Float,
decltype(in_c_n1_b_n2_global_merged_desc),
......@@ -155,20 +155,19 @@ struct GridwiseConvolutionImplicitGemm_v3_nchw_cyxk_nkhw
// operator for blockwise copy of weight into LDS
// slice a tensor, and copy it into another tensor
// this copy operator already have blockwise offset built-in
auto blockwise_wei_copy =
BlockwiseGenericTensorSliceCopy_v1<BlockSize,
Float,
decltype(wei_c_k_global_desc),
decltype(wei_c_k_block_desc),
decltype(wei_c_k_block_desc.GetLengths()),
WeiBlockCopySubLengths_C_K,
WeiBlockCopyClusterLengths_C_K,
Sequence<0, 1>, // thread_arrange_order [C, K]
Sequence<0, 1>, // src_access_order [C, K]
Sequence<0, 1>, // dst_access_order [C, K]
WeiBlockCopyDataPerAccess_K,
WeiBlockCopyDataPerAccess_K>(
{0, k_block_data_on_global}, {0, 0});
auto blockwise_wei_copy = BlockwiseGenericTensorSliceCopy_v1_deprecated<
BlockSize,
Float,
decltype(wei_c_k_global_desc),
decltype(wei_c_k_block_desc),
decltype(wei_c_k_block_desc.GetLengths()),
WeiBlockCopySubLengths_C_K,
WeiBlockCopyClusterLengths_C_K,
Sequence<0, 1>, // thread_arrange_order [C, K]
Sequence<0, 1>, // src_access_order [C, K]
Sequence<0, 1>, // dst_access_order [C, K]
WeiBlockCopyDataPerAccess_K,
WeiBlockCopyDataPerAccess_K>({0, k_block_data_on_global}, {0, 0});
// GEMM definition
// c_mtx += transpose(a_mtx) * b_mtx
......
composable_kernel/include/kernel_algorithm/gridwise_convolution_implicit_gemm_v3_nchw_cyxk_nkhw_lds_double_buffer.hpp
View file @ 52c3fe05
......@@ -2,8 +2,8 @@
#define CK_GRIDWISE_CONVOLUTION_IMPLICIT_GEMM_V3_NCHW_CYXK_NKHW_LDS_DOUBLE_BUFFER
#include "common_header.hpp"
#include "ConstantTensorDescriptor.hpp"
#include "ConstantMergedTensorDescriptor.hpp"
#include "ConstantTensorDescriptor_deprecated.hpp"
#include "ConstantMergedTensorDescriptor_deprecated.hpp"
#include "ConstantMatrixDescriptor.hpp"
#include "blockwise_generic_tensor_slice_copy.hpp"
#include "blockwise_gemm.hpp"
......@@ -125,7 +125,7 @@ struct GridwiseConvolutionImplicitGemm_v3_nchw_cyxk_nkhw_lds_double_buffer
// input blockwise copy
// slice a merged tensor, reorder and copy to a normal tensor
// this copy operator already has blockwise offset built-in
const auto blockwise_in_copy = BlockwiseGenericTensorSliceCopy_v1<
const auto blockwise_in_copy = BlockwiseGenericTensorSliceCopy_v1_deprecated<
BlockSize,
Float,
decltype(in_c_n1_b_n2_global_merged_desc),
......@@ -152,20 +152,19 @@ struct GridwiseConvolutionImplicitGemm_v3_nchw_cyxk_nkhw_lds_double_buffer
// operator for blockwise copy of weight into LDS
// slice a tensor, and copy it into another tensor
// this copy operator already have blockwise offset built-in
const auto blockwise_wei_copy =
BlockwiseGenericTensorSliceCopy_v1<BlockSize,
Float,
decltype(wei_c_k_global_desc),
decltype(wei_c_k_block_desc),
decltype(wei_c_k_block_desc.GetLengths()),
WeiBlockCopySubLengths_C_K,
WeiBlockCopyClusterLengths_C_K,
Sequence<0, 1>, // thread_arrange_order [C, K]
Sequence<0, 1>, // src_access_order [C, K]
Sequence<0, 1>, // dst_access_order [C, K]
WeiBlockCopyDataPerAccess_K,
WeiBlockCopyDataPerAccess_K>(
{0, k_block_data_on_global}, {0, 0});
const auto blockwise_wei_copy = BlockwiseGenericTensorSliceCopy_v1_deprecated<
BlockSize,
Float,
decltype(wei_c_k_global_desc),
decltype(wei_c_k_block_desc),
decltype(wei_c_k_block_desc.GetLengths()),
WeiBlockCopySubLengths_C_K,
WeiBlockCopyClusterLengths_C_K,
Sequence<0, 1>, // thread_arrange_order [C, K]
Sequence<0, 1>, // src_access_order [C, K]
Sequence<0, 1>, // dst_access_order [C, K]
WeiBlockCopyDataPerAccess_K,
WeiBlockCopyDataPerAccess_K>({0, k_block_data_on_global}, {0, 0});
// GEMM definition
// c_mtx += transpose(a_mtx) * b_mtx
......
composable_kernel/include/kernel_algorithm/gridwise_convolution_implicit_gemm_v4r2_nchw_kcyx_nkhw_lds_double_buffer.hpp
View file @ 52c3fe05
......@@ -2,8 +2,8 @@
#define CK_GRIDWISE_CONVOLUTION_IMPLICIT_GEMM_V4R2_NCHW_KCYX_NKHW_LDS_DOUBLE_BUFFER
#include "common_header.hpp"
#include "ConstantTensorDescriptor.hpp"
#include "ConstantMergedTensorDescriptor.hpp"
#include "ConstantTensorDescriptor_deprecated.hpp"
#include "ConstantMergedTensorDescriptor_deprecated.hpp"
#include "ConstantMatrixDescriptor.hpp"
#include "blockwise_generic_tensor_slice_copy.hpp"
#include "blockwise_gemm.hpp"
......@@ -166,7 +166,7 @@ struct GridwiseConvolutionImplicitGemm_v4r2_nchw_kcyx_nkhw_lds_double_buffer
// input blockwise copy
// slice a merged tensor, reorder and copy to a normal tensor
// this copy operator already has blockwise offset built-in
auto blockwise_in_copy = BlockwiseGenericTensorSliceCopy_v1<
auto blockwise_in_copy = BlockwiseGenericTensorSliceCopy_v1_deprecated<
BlockSize,
Float,
decltype(in_e_n0_ho0_wo0_b_n2_ho2_wo2_global_merged_desc),
......@@ -196,18 +196,18 @@ struct GridwiseConvolutionImplicitGemm_v4r2_nchw_kcyx_nkhw_lds_double_buffer
// slice a tensor, and copy it into another tensor
// this copy operator already have blockwise offset built-in
auto blockwise_wei_copy =
BlockwiseGenericTensorSliceCopy_v1<BlockSize,
Float,
decltype(wei_e_k_global_desc),
decltype(wei_e_k_block_desc),
decltype(wei_e_k_block_desc.GetLengths()),
WeiBlockCopySubLengths_E_K,
WeiBlockCopyClusterLengths_E_K,
WeiBlockCopyThreadClusterArrangeOrder,
WeiBlockCopySrcAccessOrder,
WeiBlockCopyDstAccessOrder,
WeiBlockCopySrcDataPerRead_E,
WeiBlockCopyDstDataPerWrite_K>(
BlockwiseGenericTensorSliceCopy_v1_deprecated<BlockSize,
Float,
decltype(wei_e_k_global_desc),
decltype(wei_e_k_block_desc),
decltype(wei_e_k_block_desc.GetLengths()),
WeiBlockCopySubLengths_E_K,
WeiBlockCopyClusterLengths_E_K,
WeiBlockCopyThreadClusterArrangeOrder,
WeiBlockCopySrcAccessOrder,
WeiBlockCopyDstAccessOrder,
WeiBlockCopySrcDataPerRead_E,
WeiBlockCopyDstDataPerWrite_K>(
{0, k_block_data_on_global}, {0, 0});
// GEMM definition
......
composable_kernel/include/kernel_algorithm/gridwise_convolution_implicit_gemm_v4r3_nchw_kcyx_nkhw_lds_double_buffer.hpp
View file @ 52c3fe05
......@@ -2,8 +2,8 @@
#define CK_GRIDWISE_CONVOLUTION_IMPLICIT_GEMM_V4R3_NCHW_KCYX_NKHW_LDS_DOUBLE_BUFFER
#include "common_header.hpp"
#include "ConstantTensorDescriptor.hpp"
#include "ConstantMergedTensorDescriptor.hpp"
#include "ConstantTensorDescriptor_deprecated.hpp"
#include "ConstantMergedTensorDescriptor_deprecated.hpp"
#include "ConstantMatrixDescriptor.hpp"
#include "blockwise_generic_tensor_slice_copy.hpp"
#include "blockwise_gemm.hpp"
......@@ -165,7 +165,7 @@ struct GridwiseConvolutionImplicitGemm_v4r3_nchw_kcyx_nkhw_lds_double_buffer
// input blockwise copy
// slice a merged tensor, reorder and copy to a normal tensor
// this copy operator already has blockwise offset built-in
auto blockwise_in_copy = BlockwiseGenericTensorSliceCopy_v1<
auto blockwise_in_copy = BlockwiseGenericTensorSliceCopy_v1_deprecated<
BlockSize,
Float,
decltype(in_e_n1_ho1_wo1_b_n2_ho2_wo2_global_merged_desc),
......@@ -195,18 +195,18 @@ struct GridwiseConvolutionImplicitGemm_v4r3_nchw_kcyx_nkhw_lds_double_buffer
// slice a tensor, and copy it into another tensor
// this copy operator already have blockwise offset built-in
auto blockwise_wei_copy =
BlockwiseGenericTensorSliceCopy_v1<BlockSize,
Float,
decltype(wei_e_k_global_desc),
decltype(wei_e_k_block_desc),
decltype(wei_e_k_block_desc.GetLengths()),
WeiBlockCopySubLengths_E_K,
WeiBlockCopyClusterLengths_E_K,
WeiBlockCopyThreadClusterArrangeOrder,
WeiBlockCopySrcAccessOrder,
WeiBlockCopyDstAccessOrder,
WeiBlockCopySrcDataPerRead_E,
WeiBlockCopyDstDataPerWrite_K>(
BlockwiseGenericTensorSliceCopy_v1_deprecated<BlockSize,
Float,
decltype(wei_e_k_global_desc),
decltype(wei_e_k_block_desc),
decltype(wei_e_k_block_desc.GetLengths()),
WeiBlockCopySubLengths_E_K,
WeiBlockCopyClusterLengths_E_K,
WeiBlockCopyThreadClusterArrangeOrder,
WeiBlockCopySrcAccessOrder,
WeiBlockCopyDstAccessOrder,
WeiBlockCopySrcDataPerRead_E,
WeiBlockCopyDstDataPerWrite_K>(
{0, k_block_data_on_global}, {0, 0});
#if 0
......
Markdown is supported
0% or .
You are about to add 0 people to the discussion. Proceed with caution.
Finish editing this message first!
Please register or to comment