Skip to content

GitLab

"git@developer.sourcefind.cn:OpenDAS/opencompass.git" did not exist on "e03526535215f732809e3beb7832104e20b8c9f5"
Commit ed068043 authored by Jing Zhang's avatar Jing Zhang
Browse files

merged develop

parents 41852668 e823d518
Expand all Hide whitespace changes
Inline Side-by-side
Showing with 1774 additions and 375 deletions
CMakeLists.txt
View file @ ed068043
...@@ -40,6 +40,7 @@ message(STATUS "Build with HIP ${hip_VERSION}") ...@@ -40,6 +40,7 @@ message(STATUS "Build with HIP ${hip_VERSION}")
## half ## half
#find_path(HALF_INCLUDE_DIR half.hpp) #find_path(HALF_INCLUDE_DIR half.hpp)
set(HALF_INCLUDE_DIR "${PROJECT_SOURCE_DIR}/external/half/include")
message("HALF_INCLUDE_DIR: ${HALF_INCLUDE_DIR}") message("HALF_INCLUDE_DIR: ${HALF_INCLUDE_DIR}")
# CMAKE_CXX_FLAGS # CMAKE_CXX_FLAGS
...@@ -185,6 +186,7 @@ enable_cppcheck( ...@@ -185,6 +186,7 @@ enable_cppcheck(
composable_kernel/src/kernel_wrapper composable_kernel/src/kernel_wrapper
INCLUDE INCLUDE
host/host_tensor/include host/host_tensor/include
host/device/include
host/solver/include host/solver/include
host/driver_offline/include host/driver_offline/include
composable_kernel/include/* composable_kernel/include/*
...@@ -196,3 +198,5 @@ enable_cppcheck( ...@@ -196,3 +198,5 @@ enable_cppcheck(
) )
add_subdirectory(host) add_subdirectory(host)
add_subdirectory(example)
add_subdirectory(profiler)
README.md
View file @ ed068043
# How to build and run
# Docker
```
docker run \
-it \
--rm \
--privileged \
--group-add sudo \
-w /root/workspace \
-v ${PATH_TO_LOCAL_WORKSPACE}:/root/workspace \
rocm/tensorflow:rocm4.2-tf2.4-dev \
/bin/bash
```
# Install Boost for online compilation
https://www.boost.org/doc/libs/1_66_0/more/getting_started/unix-variants.html#easy-build-and-install
# Build
Add path of Boost
```
export LD_LIBRARY_PATH=/usr/local/lib:$LD_LIBRARY_PATH
```
```
mkdir build && cd build
```
cmake cmd. Need to Specify target ID, example below is gfx908
```
cmake \
-D CMAKE_BUILD_TYPE=Release \
-D CMAKE_CXX_FLAGS="-DCK_AMD_GPU_GFX908 -O3 --amdgpu-target=gfx908 -mllvm --amdgpu-spill-vgpr-to-agpr=0 -gline-tables-only -save-temps=$PWD" \
-D HIP_ONLINE_COMPILER_FLAGS="-DCK_AMD_GPU_GFX908" \
-D CMAKE_CXX_COMPILER=/opt/rocm/bin/hipcc \
-D CMAKE_PREFIX_PATH=/opt/rocm \
-D CMAKE_VERBOSE_MAKEFILE:BOOL=ON \
..
```
Build drivers: \
``conv_fwd_driver_offline`` is (offline compilation) driver for forward convolution, \
``conv_bwd_driver_offline`` is (offline compilation) driver for backward-data convolution \
``conv_fwd_driver_online`` is (online compilation) driver for forward convolution
```
make -j conv_fwd_driver_offline
make -j conv_bwd_driver_offline
make -j conv_fwd_driver_online
```
# Run
* layout: 0 = NCHW; 1 = NHWC
* algo: algorithm
* verify: 0 = no verification; 1 = do verification
* init: 0 ~ 5. initialization method
* log: 0 = no log; 1 = do log
* repeat: number of time kernel being launched
```
######################################################## layout algo verify init log repeat N__ K___ C___ Y X Hi_ Wi__ Strides Dilations LeftPads RightPads
./host/driver_offline/conv_fwd_driver_offline 0 4 0 0 0 1 128 256 192 3 3 71 71 2 2 1 1 1 1 1 1
./host/driver_offline/conv_fwd_driver_offline 0 4 0 0 0 1 256 1024 256 3 3 14 14 1 1 1 1 1 1 1 1
./host/driver_offline/conv_fwd_driver_offline 1 5 0 0 0 1 128 256 192 3 3 71 71 2 2 1 1 1 1 1 1
./host/driver_offline/conv_fwd_driver_offline 1 5 0 0 0 1 256 1024 256 3 3 14 14 1 1 1 1 1 1 1 1
./host/driver_offline/conv_bwd_driver_offline 1 5 0 0 0 1 256 256 1024 3 3 14 14 1 1 1 1 1 1 1 1
```
# Result
Forward convoltuion, FP16, NCHW
```
./host/driver_offline/conv_fwd_driver_offline 0 4 0 0 0 1 128 256 192 3 3 71 71 2 2 1 1 1 1 1 1
layout: 0
in: dim 4, lengths {128, 192, 71, 71}, strides {967872, 5041, 71, 1}
wei: dim 4, lengths {256, 192, 3, 3}, strides {1728, 9, 3, 1}
out: dim 4, lengths {128, 256, 36, 36}, strides {331776, 1296, 36, 1}
InLeftPads size 2, {1, 1, }
InRightPads size 2, {1, 1, }
ConvStrides size 2, {2, 2, }
ConvDilations size 2, {1, 1, }
device_convolution_forward_implicit_gemm_v4r4r2_xdlops_nchw_kcyx_nkhw
a_k0_m_k1_grid_desc{216, 256, 8}
b_k0_n_k1_grid_desc{216, 165888, 8}
c_m_n_grid_desc{ 256, 165888}
launch_and_time_kernel: grid_dim {1296, 1, 1}, block_dim {256, 1, 1}
Warm up
Start running 1 times...
Average time : 1.4155 ms, 103.686 TFlop/s
```
Forward convoltuion, FP16, NCHW
```
./host/driver_offline/conv_fwd_driver_offline 0 4 0 0 0 1 256 1024 256 3 3 14 14 1 1 1 1 1 1 1 1
layout: 0
in: dim 4, lengths {256, 256, 14, 14}, strides {50176, 196, 14, 1}
wei: dim 4, lengths {1024, 256, 3, 3}, strides {2304, 9, 3, 1}
out: dim 4, lengths {256, 1024, 14, 14}, strides {200704, 196, 14, 1}
InLeftPads size 2, {1, 1, }
InRightPads size 2, {1, 1, }
ConvStrides size 2, {1, 1, }
ConvDilations size 2, {1, 1, }
device_convolution_forward_implicit_gemm_v4r4r2_xdlops_nchw_kcyx_nkhw
a_k0_m_k1_grid_desc{288, 1024, 8}
b_k0_n_k1_grid_desc{288, 50176, 8}
c_m_n_grid_desc{ 1024, 50176}
launch_and_time_kernel: grid_dim {1568, 1, 1}, block_dim {256, 1, 1}
Warm up
Start running 1 times...
Average time : 2.21357 ms, 106.959 TFlop/s
```
Forward convolution, FP16, NHWC
```
./host/driver_offline/conv_fwd_driver_offline 1 5 0 0 0 1 128 256 192 3 3 71 71 2 2 1 1 1 1 1 1
layout: 1
in: dim 4, lengths {128, 71, 71, 192}, strides {967872, 13632, 192, 1}
wei: dim 4, lengths {256, 3, 3, 192}, strides {1728, 576, 192, 1}
out: dim 4, lengths {128, 36, 36, 256}, strides {331776, 9216, 256, 1}
InLeftPads size 2, {1, 1, }
InRightPads size 2, {1, 1, }
ConvStrides size 2, {2, 2, }
ConvDilations size 2, {1, 1, }
device_convolution_forward_implicit_gemm_v4r4r4_xdlops_nhwc_kyxc_nhwk
a_k0_m_k1_grid_desc{216, 165888, 8}
b_k0_n_k1_grid_desc{216, 256, 8}
c_m_n_grid_desc{ 165888, 256}
launch_and_time_kernel: grid_dim {1296, 1, 1}, block_dim {256, 1, 1}
Warm up
Start running 1 times...
Average time : 1.12014 ms, 131.025 TFlop/s
```
Forward convolution, FP16, NHWC
```
./host/driver_offline/conv_fwd_driver_offline 1 5 0 0 0 1 256 1024 256 3 3 14 14 1 1 1 1 1 1 1 1
layout: 1
in: dim 4, lengths {256, 14, 14, 256}, strides {50176, 3584, 256, 1}
wei: dim 4, lengths {1024, 3, 3, 256}, strides {2304, 768, 256, 1}
out: dim 4, lengths {256, 14, 14, 1024}, strides {200704, 14336, 1024, 1}
InLeftPads size 2, {1, 1, }
InRightPads size 2, {1, 1, }
ConvStrides size 2, {1, 1, }
ConvDilations size 2, {1, 1, }
device_convolution_forward_implicit_gemm_v4r4r4_xdlops_nhwc_kyxc_nhwk
a_k0_m_k1_grid_desc{288, 50176, 8}
b_k0_n_k1_grid_desc{288, 1024, 8}
c_m_n_grid_desc{ 50176, 1024}
launch_and_time_kernel: grid_dim {1568, 1, 1}, block_dim {256, 1, 1}
Warm up
Start running 1 times...
Average time : 1.86877 ms, 126.693 TFlop/s
```
Backward data convolution, FP16, NHWC
```
./host/driver_offline/conv_bwd_driver_offline 1 1 0 3 0 1 256 256 1024 3 3 14 14 1 1 1 1 1 1 1 1
layout: 1
in: dim 4, lengths {256, 14, 14, 1024}, strides {200704, 14336, 1024, 1}
wei: dim 4, lengths {256, 3, 3, 1024}, strides {9216, 3072, 1024, 1}
out: dim 4, lengths {256, 14, 14, 256}, strides {50176, 3584, 256, 1}
InLeftPads size 2, {1, 1, }
InRightPads size 2, {1, 1, }
ConvStrides size 2, {1, 1, }
ConvDilations size 2, {1, 1, }
device_convolution_backward_data_implicit_gemm_v4r1r2_xdlops_nhwc_kyxc_nhwk
a_k0_m_k1_grid_desc{288, 50176, 8}
b_k0_n_k1_grid_desc{288, 1024, 8}
c_m_n_grid_desc{ 50176, 1024}
launch_and_time_kernel: grid_dim {1568, 1, 1}, block_dim {256, 1, 1}
Warm up
Start running 1 times...
Average time : 2.22461 ms, 106.428 TFlop/s
```
composable_kernel/include/problem_transform/transform_backward_weight_convolution_into_gemm_v4r4r2_atomic_nchw_kcyx_nkhw.hpp 0 → 100644
View file @ ed068043
#ifndef CK_TRANSFORM_BACKWARD_WEIGHT_CONVOLUTION_INTO_GEMM_V4R4R2_ATOMIC_NCHW_KCYX_NKHW_HPP
#define CK_TRANSFORM_BACKWARD_WEIGHT_CONVOLUTION_INTO_GEMM_V4R4R2_ATOMIC_NCHW_KCYX_NKHW_HPP
#include "common_header.hpp"
#include "tensor_descriptor.hpp"
#include "tensor_descriptor_helper.hpp"
namespace ck {
// GemmM = K
// GemmK = N * Ho * Wo
// GemmN = C * Y * X
template <typename... Wei,
typename... In,
typename... Out,
typename ConvStrides,
typename ConvDilations,
typename InLeftPads,
typename InRightPads,
index_t GemmK1Value,
typename GemmKBatchType,
typename GemmKPadType>
__host__ __device__ constexpr auto
transform_backward_weight_convolution_into_gemm_v4r4r2_atomic_nchw_kcyx_nkhw_pad(
const TensorDescriptor<Wei...>& wei_k_c_y_x_grid_desc,
const TensorDescriptor<In...>& in_n_c_hi_wi_grid_desc,
const TensorDescriptor<Out...>& out_n_k_ho_wo_grid_desc,
const ConvStrides& conv_strides,
const ConvDilations& conv_dilations,
const InLeftPads& in_left_pads,
const InRightPads& in_right_pads,
Number<GemmK1Value>,
GemmKBatchType GemmKBatch,
GemmKPadType GemmKPad)
{
constexpr auto I0 = Number<0>{};
constexpr auto I1 = Number<1>{};
constexpr auto I2 = Number<2>{};
constexpr auto I3 = Number<3>{};
constexpr auto GemmK1 = Number<GemmK1Value>{};
const auto N = in_n_c_hi_wi_grid_desc.GetLength(I0);
const auto C = in_n_c_hi_wi_grid_desc.GetLength(I1);
const auto K = out_n_k_ho_wo_grid_desc.GetLength(I1);
const auto Hi = in_n_c_hi_wi_grid_desc.GetLength(I2);
const auto Wi = in_n_c_hi_wi_grid_desc.GetLength(I3);
const auto Ho = out_n_k_ho_wo_grid_desc.GetLength(I2);
const auto Wo = out_n_k_ho_wo_grid_desc.GetLength(I3);
const auto Y = wei_k_c_y_x_grid_desc.GetLength(I2);
const auto X = wei_k_c_y_x_grid_desc.GetLength(I3);
const auto ConvStrideH = conv_strides[I0];
const auto ConvStrideW = conv_strides[I1];
const auto ConvDilationH = conv_dilations[I0];
const auto ConvDilationW = conv_dilations[I1];
const auto InLeftPadH = in_left_pads[I0];
const auto InLeftPadW = in_left_pads[I1];
const auto InRightPadH = in_right_pads[I0];
const auto InRightPadW = in_right_pads[I1];
const auto GemmM = K;
const auto GemmN = C * Y * X;
const auto GemmKTotal = N * Ho * Wo;
const index_t GemmK0 = GemmKPad / (GemmKBatch * GemmK1);
// A: output tensor
const auto out_gemmktotal_gemmm_grid_desc = transform_tensor_descriptor(
make_naive_tensor_descriptor_packed(make_tuple(N, K, Ho * Wo)),
make_tuple(make_pass_through_transform(K), make_merge_transform(make_tuple(N, Ho * Wo))),
make_tuple(Sequence<1>{}, Sequence<0, 2>{}),
make_tuple(Sequence<1>{}, Sequence<0>{}));
const auto out_gemmkpad_gemmm_grid_desc = transform_tensor_descriptor(
out_gemmktotal_gemmm_grid_desc,
make_tuple(make_right_pad_transform(GemmKTotal, GemmKPad - GemmKTotal),
make_pass_through_transform(GemmM)),
make_tuple(Sequence<0>{}, Sequence<1>{}),
make_tuple(Sequence<0>{}, Sequence<1>{}));
const auto out_gemmkbatch_gemmk0_gemmm_gemmk1_grid_desc = transform_tensor_descriptor(
out_gemmkpad_gemmm_grid_desc,
make_tuple(make_unmerge_transform(make_tuple(GemmKBatch, GemmK0, GemmK1)),
make_pass_through_transform(GemmM)),
make_tuple(Sequence<0>{}, Sequence<1>{}),
make_tuple(Sequence<0, 1, 3>{}, Sequence<2>{}));
// B: input tensor
const auto in_n_c_hip_wip_grid_desc = transform_tensor_descriptor(
in_n_c_hi_wi_grid_desc,
make_tuple(make_pass_through_transform(N),
make_pass_through_transform(C),
make_pad_transform(Hi, InLeftPadH, InRightPadH),
make_pad_transform(Wi, InLeftPadW, InRightPadW)),
make_tuple(Sequence<0>{}, Sequence<1>{}, Sequence<2>{}, Sequence<3>{}),
make_tuple(Sequence<0>{}, Sequence<1>{}, Sequence<2>{}, Sequence<3>{}));
const auto in_n_c_y_ho_x_wo_grid_desc = transform_tensor_descriptor(
in_n_c_hip_wip_grid_desc,
make_tuple(make_pass_through_transform(N),
make_pass_through_transform(C),
make_embed_transform(make_tuple(Y, Ho), make_tuple(ConvDilationH, ConvStrideH)),
make_embed_transform(make_tuple(X, Wo), make_tuple(ConvDilationW, ConvStrideW))),
make_tuple(Sequence<0>{}, Sequence<1>{}, Sequence<2>{}, Sequence<3>{}),
make_tuple(Sequence<0>{}, Sequence<1>{}, Sequence<2, 3>{}, Sequence<4, 5>{}));
const auto in_gemmktotal_gemmn_grid_desc =
transform_tensor_descriptor(in_n_c_y_ho_x_wo_grid_desc,
make_tuple(make_merge_transform(make_tuple(C, Y, X)),
make_merge_transform(make_tuple(N, Ho, Wo))),
make_tuple(Sequence<1, 2, 4>{}, Sequence<0, 3, 5>{}),
make_tuple(Sequence<1>{}, Sequence<0>{}));
const auto in_gemmkpad_gemmn_grid_desc = transform_tensor_descriptor(
in_gemmktotal_gemmn_grid_desc,
make_tuple(make_right_pad_transform(GemmKTotal, GemmKPad - GemmKTotal),
make_pass_through_transform(GemmN)),
make_tuple(Sequence<0>{}, Sequence<1>{}),
make_tuple(Sequence<0>{}, Sequence<1>{}));
const auto in_gemmkbatch_gemmk0_gemmn_gemmk1_grid_desc = transform_tensor_descriptor(
in_gemmkpad_gemmn_grid_desc,
make_tuple(make_unmerge_transform(make_tuple(GemmKBatch, GemmK0, GemmK1)),
make_pass_through_transform(GemmN)),
make_tuple(Sequence<0>{}, Sequence<1>{}),
make_tuple(Sequence<0, 1, 3>{}, Sequence<2>{}));
// C: weight tensor
const auto wei_gemmm_gemmn_grid_desc = transform_tensor_descriptor(
make_naive_tensor_descriptor_packed(make_tuple(K, C * Y * X)),
make_tuple(make_pass_through_transform(K), make_pass_through_transform(C * Y * X)),
make_tuple(Sequence<0>{}, Sequence<1>{}),
make_tuple(Sequence<0>{}, Sequence<1>{}));
return make_tuple(out_gemmkbatch_gemmk0_gemmm_gemmk1_grid_desc,
in_gemmkbatch_gemmk0_gemmn_gemmk1_grid_desc,
wei_gemmm_gemmn_grid_desc);
}
} // namespace ck
#endif
composable_kernel/include/problem_transform/transform_backward_weight_convolution_into_gemm_v4r4r4_atomic_nhwc_kyxc_nhwk.hpp 0 → 100644
View file @ ed068043
#ifndef CK_TRANSFORM_BACKWARD_WEIGHT_CONVOLUTION_INTO_GEMM_V4R4R4_ATOMIC_NHWC_KYXC_NHWK_HPP
#define CK_TRANSFORM_BACKWARD_WEIGHT_CONVOLUTION_INTO_GEMM_V4R4R4_ATOMIC_NHWC_KYXC_NHWK_HPP
#include "common_header.hpp"
#include "tensor_descriptor.hpp"
#include "tensor_descriptor_helper.hpp"
namespace ck {
// A: in
// B: wei
// C: out
// GemmM = N * Ho * Wo
// GemmN = K
// GemmK = Y * X * C
template <typename... In,
typename... Wei,
typename... Out,
typename ConvStrides,
typename ConvDilations,
typename InLeftPads,
typename InRightPads,
index_t GemmK1Value,
typename GemmKBatchType,
typename GemmKPadType>
__host__ __device__ constexpr auto
transform_backward_weight_convolution_into_gemm_v4r4r4_atomic_nhwc_kyxc_nhwk_pad(
const TensorDescriptor<In...>& in_n_hi_wi_c_grid_desc,
const TensorDescriptor<Wei...>& wei_k_y_x_c_grid_desc,
const TensorDescriptor<Out...>& out_n_ho_wo_k_grid_desc,
const ConvStrides& conv_strides,
const ConvDilations& conv_dilations,
const InLeftPads& in_left_pads,
const InRightPads& in_right_pads,
Number<GemmK1Value>,
GemmKBatchType GemmKBatch,
GemmKPadType GemmKPad)
{
constexpr auto I0 = Number<0>{};
constexpr auto I1 = Number<1>{};
constexpr auto I2 = Number<2>{};
constexpr auto I3 = Number<3>{};
constexpr auto GemmK1 = Number<GemmK1Value>{};
const auto N = in_n_hi_wi_c_grid_desc.GetLength(I0);
const auto C = in_n_hi_wi_c_grid_desc.GetLength(I3);
const auto K = out_n_ho_wo_k_grid_desc.GetLength(I3);
const auto Hi = in_n_hi_wi_c_grid_desc.GetLength(I1);
const auto Wi = in_n_hi_wi_c_grid_desc.GetLength(I2);
const auto Ho = out_n_ho_wo_k_grid_desc.GetLength(I1);
const auto Wo = out_n_ho_wo_k_grid_desc.GetLength(I2);
const auto Y = wei_k_y_x_c_grid_desc.GetLength(I1);
const auto X = wei_k_y_x_c_grid_desc.GetLength(I2);
const auto ConvStrideH = conv_strides[I0];
const auto ConvStrideW = conv_strides[I1];
const auto ConvDilationH = conv_dilations[I0];
const auto ConvDilationW = conv_dilations[I1];
const auto InLeftPadH = in_left_pads[I0];
const auto InLeftPadW = in_left_pads[I1];
const auto InRightPadH = in_right_pads[I0];
const auto InRightPadW = in_right_pads[I1];
const auto GemmM = Y * X * C;
const auto GemmN = K;
const auto GemmKTotal = N * Ho * Wo;
const index_t GemmK0 = GemmKPad / (GemmKBatch * GemmK1);
// A: input tensor
const auto in_n_hip_wip_c_grid_desc = transform_tensor_descriptor(
in_n_hi_wi_c_grid_desc,
make_tuple(make_pass_through_transform(N),
make_pad_transform(Hi, InLeftPadH, InRightPadH),
make_pad_transform(Wi, InLeftPadW, InRightPadW),
make_pass_through_transform(C)),
make_tuple(Sequence<0>{}, Sequence<1>{}, Sequence<2>{}, Sequence<3>{}),
make_tuple(Sequence<0>{}, Sequence<1>{}, Sequence<2>{}, Sequence<3>{}));
const auto in_n_y_ho_x_wo_c_grid_desc = transform_tensor_descriptor(
in_n_hip_wip_c_grid_desc,
make_tuple(make_pass_through_transform(N),
make_embed_transform(make_tuple(Y, Ho), make_tuple(ConvDilationH, ConvStrideH)),
make_embed_transform(make_tuple(X, Wo), make_tuple(ConvDilationW, ConvStrideW)),
make_pass_through_transform(C)),
make_tuple(Sequence<0>{}, Sequence<1>{}, Sequence<2>{}, Sequence<3>{}),
make_tuple(Sequence<0>{}, Sequence<1, 2>{}, Sequence<3, 4>{}, Sequence<5>{}));
const auto in_gemmktotal_gemmm_grid_desc =
transform_tensor_descriptor(in_n_y_ho_x_wo_c_grid_desc,
make_tuple(make_merge_transform(make_tuple(Y, X, C)),
make_merge_transform(make_tuple(N, Ho, Wo))),
make_tuple(Sequence<1, 3, 5>{}, Sequence<0, 2, 4>{}),
make_tuple(Sequence<1>{}, Sequence<0>{}));
const auto in_gemmkpad_gemmm_grid_desc = transform_tensor_descriptor(
in_gemmktotal_gemmm_grid_desc,
make_tuple(make_right_pad_transform(GemmKTotal, GemmKPad - GemmKTotal),
make_pass_through_transform(GemmM)),
make_tuple(Sequence<0>{}, Sequence<1>{}),
make_tuple(Sequence<0>{}, Sequence<1>{}));
const auto in_gemmkbatch_gemmk0_gemmm_gemmk1_grid_desc = transform_tensor_descriptor(
in_gemmkpad_gemmm_grid_desc,
make_tuple(make_unmerge_transform(make_tuple(GemmKBatch, GemmK0, GemmK1)),
make_pass_through_transform(GemmM)),
make_tuple(Sequence<0>{}, Sequence<1>{}),
make_tuple(Sequence<0, 1, 3>{}, Sequence<2>{}));
// B: output tensor
const auto out_gemmktotal_gemmn_grid_desc =
make_naive_tensor_descriptor_packed(make_tuple(N * Ho * Wo, K));
const auto out_gemmkpad_gemmn_grid_desc = transform_tensor_descriptor(
out_gemmktotal_gemmn_grid_desc,
make_tuple(make_right_pad_transform(GemmKTotal, GemmKPad - GemmKTotal),
make_pass_through_transform(GemmN)),
make_tuple(Sequence<0>{}, Sequence<1>{}),
make_tuple(Sequence<0>{}, Sequence<1>{}));
const auto out_gemmkbatch_gemmk0_gemmn_gemmk1_grid_desc = transform_tensor_descriptor(
out_gemmkpad_gemmn_grid_desc,
make_tuple(make_unmerge_transform(make_tuple(GemmKBatch, GemmK0, GemmK1)),
make_pass_through_transform(GemmN)),
make_tuple(Sequence<0>{}, Sequence<1>{}),
make_tuple(Sequence<0, 1, 3>{}, Sequence<2>{}));
// C: weight tensor
const auto wei_gemmm_gemmn_grid_desc = transform_tensor_descriptor(
make_naive_tensor_descriptor_packed(make_tuple(K, Y * X * C)),
make_tuple(make_pass_through_transform(K), make_pass_through_transform(Y * X * C)),
make_tuple(Sequence<0>{}, Sequence<1>{}),
make_tuple(Sequence<1>{}, Sequence<0>{}));
return make_tuple(in_gemmkbatch_gemmk0_gemmm_gemmk1_grid_desc,
out_gemmkbatch_gemmk0_gemmn_gemmk1_grid_desc,
wei_gemmm_gemmn_grid_desc);
}
} // namespace ck
#endif
composable_kernel/include/problem_transform/transform_backward_weight_convolution_into_gemm_v4r4r4_nhwc_kyxc_nhwk.hpp 0 → 100644
View file @ ed068043
#ifndef CK_TRANSFORM_BACKWARD_WEIGHT_CONVOLUTION_INTO_GEMM_V4R4R4_NHWC_KYXC_NHWK_HPP
#define CK_TRANSFORM_BACKWARD_WEIGHT_CONVOLUTION_INTO_GEMM_V4R4R4_NHWC_KYXC_NHWK_HPP
#include "common_header.hpp"
#include "tensor_descriptor.hpp"
#include "tensor_descriptor_helper.hpp"
namespace ck {
// A: in
// B: wei
// C: out
// GemmM = N * Ho * Wo
// GemmN = K
// GemmK = Y * X * C
template <typename... In,
typename... Wei,
typename... Out,
typename ConvStrides,
typename ConvDilations,
typename InLeftPads,
typename InRightPads,
index_t GemmK1Value>
__host__ __device__ constexpr auto
transform_backward_weight_convolution_into_gemm_v4r4r4_nhwc_kyxc_nhwk_pad(
const TensorDescriptor<In...>& in_n_hi_wi_c_grid_desc,
const TensorDescriptor<Wei...>& wei_k_y_x_c_grid_desc,
const TensorDescriptor<Out...>& out_n_ho_wo_k_grid_desc,
const ConvStrides& conv_strides,
const ConvDilations& conv_dilations,
const InLeftPads& in_left_pads,
const InRightPads& in_right_pads,
Number<GemmK1Value>)
{
constexpr auto I0 = Number<0>{};
constexpr auto I1 = Number<1>{};
constexpr auto I2 = Number<2>{};
constexpr auto I3 = Number<3>{};
constexpr auto GemmK1 = Number<GemmK1Value>{};
const auto N = in_n_hi_wi_c_grid_desc.GetLength(I0);
const auto C = in_n_hi_wi_c_grid_desc.GetLength(I3);
const auto K = out_n_ho_wo_k_grid_desc.GetLength(I3);
const auto Hi = in_n_hi_wi_c_grid_desc.GetLength(I1);
const auto Wi = in_n_hi_wi_c_grid_desc.GetLength(I2);
const auto Ho = out_n_ho_wo_k_grid_desc.GetLength(I1);
const auto Wo = out_n_ho_wo_k_grid_desc.GetLength(I2);
const auto Y = wei_k_y_x_c_grid_desc.GetLength(I1);
const auto X = wei_k_y_x_c_grid_desc.GetLength(I2);
const auto ConvStrideH = conv_strides[I0];
const auto ConvStrideW = conv_strides[I1];
const auto ConvDilationH = conv_dilations[I0];
const auto ConvDilationW = conv_dilations[I1];
const auto InLeftPadH = in_left_pads[I0];
const auto InLeftPadW = in_left_pads[I1];
const auto InRightPadH = in_right_pads[I0];
const auto InRightPadW = in_right_pads[I1];
const auto GemmM = Y * X * C;
const auto GemmN = K;
const auto GemmK = N * Ho * Wo;
const auto GemmK0 = GemmK / GemmK1;
// A: input tensor
const auto in_n_hip_wip_c_grid_desc = transform_tensor_descriptor(
in_n_hi_wi_c_grid_desc,
make_tuple(make_pass_through_transform(N),
make_pad_transform(Hi, InLeftPadH, InRightPadH),
make_pad_transform(Wi, InLeftPadW, InRightPadW),
make_pass_through_transform(C)),
make_tuple(Sequence<0>{}, Sequence<1>{}, Sequence<2>{}, Sequence<3>{}),
make_tuple(Sequence<0>{}, Sequence<1>{}, Sequence<2>{}, Sequence<3>{}));
const auto in_n_y_ho_x_wo_c_grid_desc = transform_tensor_descriptor(
in_n_hip_wip_c_grid_desc,
make_tuple(make_pass_through_transform(N),
make_embed_transform(make_tuple(Y, Ho), make_tuple(ConvDilationH, ConvStrideH)),
make_embed_transform(make_tuple(X, Wo), make_tuple(ConvDilationW, ConvStrideW)),
make_pass_through_transform(C)),
make_tuple(Sequence<0>{}, Sequence<1>{}, Sequence<2>{}, Sequence<3>{}),
make_tuple(Sequence<0>{}, Sequence<1, 2>{}, Sequence<3, 4>{}, Sequence<5>{}));
const auto in_gemmk_gemmm_grid_desc =
transform_tensor_descriptor(in_n_y_ho_x_wo_c_grid_desc,
make_tuple(make_merge_transform(make_tuple(Y, X, C)),
make_merge_transform(make_tuple(N, Ho, Wo))),
make_tuple(Sequence<1, 3, 5>{}, Sequence<0, 2, 4>{}),
make_tuple(Sequence<1>{}, Sequence<0>{}));
const auto in_gemmk0_gemmm_gemmk1_grid_desc =
transform_tensor_descriptor(in_gemmk_gemmm_grid_desc,
make_tuple(make_unmerge_transform(make_tuple(GemmK0, GemmK1)),
make_pass_through_transform(GemmM)),
make_tuple(Sequence<0>{}, Sequence<1>{}),
make_tuple(Sequence<0, 2>{}, Sequence<1>{}));
// B: output tensor
const auto out_gemmk_gemmn_grid_desc = transform_tensor_descriptor(
make_naive_tensor_descriptor_packed(make_tuple(N * Ho * Wo, K)),
make_tuple(make_pass_through_transform(N * Ho * Wo), make_pass_through_transform(K)),
make_tuple(Sequence<0>{}, Sequence<1>{}),
make_tuple(Sequence<0>{}, Sequence<1>{}));
const auto out_gemmk0_gemmn_gemmk1_grid_desc =
transform_tensor_descriptor(out_gemmk_gemmn_grid_desc,
make_tuple(make_unmerge_transform(make_tuple(GemmK0, GemmK1)),
make_pass_through_transform(GemmN)),
make_tuple(Sequence<0>{}, Sequence<1>{}),
make_tuple(Sequence<0, 2>{}, Sequence<1>{}));
// C: weight tensor
const auto wei_gemmm_gemmn_grid_desc = transform_tensor_descriptor(
make_naive_tensor_descriptor_packed(make_tuple(K, Y * X * C)),
make_tuple(make_pass_through_transform(K), make_pass_through_transform(Y * X * C)),
make_tuple(Sequence<0>{}, Sequence<1>{}),
make_tuple(Sequence<1>{}, Sequence<0>{}));
return make_tuple(in_gemmk0_gemmm_gemmk1_grid_desc,
out_gemmk0_gemmn_gemmk1_grid_desc,
wei_gemmm_gemmn_grid_desc);
}
} // namespace ck
#endif
composable_kernel/include/problem_transform/transform_backward_weight_convolution_into_gemm_v4r4r5_nhwc_kyxc_nhwk.hpp 0 → 100644
View file @ ed068043
#ifndef CK_TRANSFORM_BACKWARD_WEIGHT_CONVOLUTION_INTO_GEMM_V4R4R5_NHWC_KYXC_NHWK_HPP
#define CK_TRANSFORM_BACKWARD_WEIGHT_CONVOLUTION_INTO_GEMM_V4R4R5_NHWC_KYXC_NHWK_HPP
#include "common_header.hpp"
#include "tensor_descriptor.hpp"
#include "tensor_descriptor_helper.hpp"
namespace ck {
// A: out
// B: in
// C: wei
// GemmM = K
// GemmN = Y * X * C
// GemmKTotal = N * Ho * Wo
template <typename... In,
typename... Wei,
typename... Out,
typename ConvStrides,
typename ConvDilations,
typename InLeftPads,
typename InRightPads,
index_t GemmK1Value,
typename GemmKBatchType,
typename GemmKPadType>
__host__ __device__ constexpr auto
transform_backward_weight_convolution_into_gemm_v4r4r5_nhwc_kyxc_nhwk_pad(
const TensorDescriptor<In...>& in_n_hi_wi_c_grid_desc,
const TensorDescriptor<Wei...>& wei_k_y_x_c_grid_desc,
const TensorDescriptor<Out...>& out_n_ho_wo_k_grid_desc,
const ConvStrides& conv_strides,
const ConvDilations& conv_dilations,
const InLeftPads& in_left_pads,
const InRightPads& in_right_pads,
Number<GemmK1Value>,
GemmKBatchType GemmKBatch,
GemmKPadType GemmKPad)
{
constexpr auto I0 = Number<0>{};
constexpr auto I1 = Number<1>{};
constexpr auto I2 = Number<2>{};
constexpr auto I3 = Number<3>{};
constexpr auto GemmK1 = Number<GemmK1Value>{};
const auto N = in_n_hi_wi_c_grid_desc.GetLength(I0);
const auto C = in_n_hi_wi_c_grid_desc.GetLength(I3);
const auto K = out_n_ho_wo_k_grid_desc.GetLength(I3);
const auto Hi = in_n_hi_wi_c_grid_desc.GetLength(I1);
const auto Wi = in_n_hi_wi_c_grid_desc.GetLength(I2);
const auto Ho = out_n_ho_wo_k_grid_desc.GetLength(I1);
const auto Wo = out_n_ho_wo_k_grid_desc.GetLength(I2);
const auto Y = wei_k_y_x_c_grid_desc.GetLength(I1);
const auto X = wei_k_y_x_c_grid_desc.GetLength(I2);
const auto ConvStrideH = conv_strides[I0];
const auto ConvStrideW = conv_strides[I1];
const auto ConvDilationH = conv_dilations[I0];
const auto ConvDilationW = conv_dilations[I1];
const auto InLeftPadH = in_left_pads[I0];
const auto InLeftPadW = in_left_pads[I1];
const auto InRightPadH = in_right_pads[I0];
const auto InRightPadW = in_right_pads[I1];
const auto GemmM = K;
const auto GemmN = Y * X * C;
const auto GemmKTotal = N * Ho * Wo;
const index_t GemmK0 = GemmKPad / (GemmKBatch * GemmK1);
// A: output tensor
const auto out_gemmktotal_gemmm_grid_desc =
make_naive_tensor_descriptor_packed(make_tuple(N * Ho * Wo, K));
const auto out_gemmkpad_gemmm_grid_desc = transform_tensor_descriptor(
out_gemmktotal_gemmm_grid_desc,
make_tuple(make_right_pad_transform(GemmKTotal, GemmKPad - GemmKTotal),
make_pass_through_transform(GemmM)),
make_tuple(Sequence<0>{}, Sequence<1>{}),
make_tuple(Sequence<0>{}, Sequence<1>{}));
const auto out_gemmkbatch_gemmk0_gemmm_gemmk1_grid_desc = transform_tensor_descriptor(
out_gemmkpad_gemmm_grid_desc,
make_tuple(make_unmerge_transform(make_tuple(GemmKBatch, GemmK0, GemmK1)),
make_pass_through_transform(GemmM)),
make_tuple(Sequence<0>{}, Sequence<1>{}),
make_tuple(Sequence<0, 1, 3>{}, Sequence<2>{}));
// B: input tensor
const auto in_n_hip_wip_c_grid_desc = transform_tensor_descriptor(
in_n_hi_wi_c_grid_desc,
make_tuple(make_pass_through_transform(N),
make_pad_transform(Hi, InLeftPadH, InRightPadH),
make_pad_transform(Wi, InLeftPadW, InRightPadW),
make_pass_through_transform(C)),
make_tuple(Sequence<0>{}, Sequence<1>{}, Sequence<2>{}, Sequence<3>{}),
make_tuple(Sequence<0>{}, Sequence<1>{}, Sequence<2>{}, Sequence<3>{}));
const auto in_n_y_ho_x_wo_c_grid_desc = transform_tensor_descriptor(
in_n_hip_wip_c_grid_desc,
make_tuple(make_pass_through_transform(N),
make_embed_transform(make_tuple(Y, Ho), make_tuple(ConvDilationH, ConvStrideH)),
make_embed_transform(make_tuple(X, Wo), make_tuple(ConvDilationW, ConvStrideW)),
make_pass_through_transform(C)),
make_tuple(Sequence<0>{}, Sequence<1>{}, Sequence<2>{}, Sequence<3>{}),
make_tuple(Sequence<0>{}, Sequence<1, 2>{}, Sequence<3, 4>{}, Sequence<5>{}));
const auto in_gemmktotal_gemmn_grid_desc =
transform_tensor_descriptor(in_n_y_ho_x_wo_c_grid_desc,
make_tuple(make_merge_transform(make_tuple(Y, X, C)),
make_merge_transform(make_tuple(N, Ho, Wo))),
make_tuple(Sequence<1, 3, 5>{}, Sequence<0, 2, 4>{}),
make_tuple(Sequence<1>{}, Sequence<0>{}));
const auto in_gemmkpad_gemmn_grid_desc = transform_tensor_descriptor(
in_gemmktotal_gemmn_grid_desc,
make_tuple(make_right_pad_transform(GemmKTotal, GemmKPad - GemmKTotal),
make_pass_through_transform(GemmN)),
make_tuple(Sequence<0>{}, Sequence<1>{}),
make_tuple(Sequence<0>{}, Sequence<1>{}));
const auto in_gemmkbatch_gemmk0_gemmn_gemmk1_grid_desc = transform_tensor_descriptor(
in_gemmkpad_gemmn_grid_desc,
make_tuple(make_unmerge_transform(make_tuple(GemmKBatch, GemmK0, GemmK1)),
make_pass_through_transform(GemmN)),
make_tuple(Sequence<0>{}, Sequence<1>{}),
make_tuple(Sequence<0, 1, 3>{}, Sequence<2>{}));
// C: weight tensor
const auto wei_gemmm_gemmn_grid_desc =
make_naive_tensor_descriptor_packed(make_tuple(K, Y * X * C));
return make_tuple(out_gemmkbatch_gemmk0_gemmm_gemmk1_grid_desc,
in_gemmkbatch_gemmk0_gemmn_gemmk1_grid_desc,
wei_gemmm_gemmn_grid_desc);
}
} // namespace ck
#endif
composable_kernel/include/problem_transform/transform_forward_convolution_into_gemm_v4r4r4_nhwc_kyxc_nhwk.hpp
View file @ ed068043
...@@ -21,8 +21,7 @@ template <typename... In, ...@@ -21,8 +21,7 @@ template <typename... In,
typename InLeftPads, typename InLeftPads,
typename InRightPads, typename InRightPads,
index_t GemmK1Value> index_t GemmK1Value>
__host__ __device__ constexpr auto __host__ __device__ constexpr auto transform_forward_convolution_into_gemm_v4r4r4_nhwc_kyxc_nhwk(
transform_forward_convolution_into_gemm_v4r4r4_nhwc_kyxc_nhwk_pad(
const TensorDescriptor<In...>& in_n_hi_wi_c_grid_desc, const TensorDescriptor<In...>& in_n_hi_wi_c_grid_desc,
const TensorDescriptor<Wei...>& wei_k_y_x_c_grid_desc, const TensorDescriptor<Wei...>& wei_k_y_x_c_grid_desc,
const TensorDescriptor<Out...>& out_n_ho_wo_k_grid_desc, const TensorDescriptor<Out...>& out_n_ho_wo_k_grid_desc,
......
composable_kernel/include/tensor_operation/blockwise_gemm_xdlops.hpp
View file @ ed068043
...@@ -124,7 +124,7 @@ struct BlockwiseGemmXdlops_k0mk1_k0nk1_m0n0m1n1m2m3m4n2_v1 ...@@ -124,7 +124,7 @@ struct BlockwiseGemmXdlops_k0mk1_k0nk1_m0n0m1n1m2m3m4n2_v1
"wrong!"); "wrong!");
} }
__host__ __device__ static constexpr auto GetCM0N0M1N1M2M3M4N2ThreadDescriptor() __host__ __device__ static constexpr auto GetCThreadDescriptor_M0_N0_M1_N1_M2_M3_M4_N2()
{ {
constexpr auto c_m0_m1_m2_n_tblk_lens = xdlops_gemm.GetCM0M1M2NThreadBlkLengths(); constexpr auto c_m0_m1_m2_n_tblk_lens = xdlops_gemm.GetCM0M1M2NThreadBlkLengths();
...@@ -136,9 +136,9 @@ struct BlockwiseGemmXdlops_k0mk1_k0nk1_m0n0m1n1m2m3m4n2_v1 ...@@ -136,9 +136,9 @@ struct BlockwiseGemmXdlops_k0mk1_k0nk1_m0n0m1n1m2m3m4n2_v1
return make_naive_tensor_descriptor_packed(make_tuple(I1, I1, I1, I1, M0, M1, M2, N)); return make_naive_tensor_descriptor_packed(make_tuple(I1, I1, I1, I1, M0, M1, M2, N));
} }
__host__ __device__ static constexpr auto GetCM0N0M1N1M2M3M4N2BlockDescriptor() __host__ __device__ static constexpr auto GetCBlockDescriptor_M0_N0_M1_N1_M2_M3_M4_N2()
{ {
constexpr auto c_m0_n0_m1_n1_m2_n2_block_desc = constexpr auto c_block_desc_m0_n0_m1_n1_m2_n2 =
make_naive_tensor_descriptor_packed(make_tuple(Number<MRepeat>{}, make_naive_tensor_descriptor_packed(make_tuple(Number<MRepeat>{},
Number<NRepeat>{}, Number<NRepeat>{},
Number<MWaves>{}, Number<MWaves>{},
...@@ -146,24 +146,24 @@ struct BlockwiseGemmXdlops_k0mk1_k0nk1_m0n0m1n1m2m3m4n2_v1 ...@@ -146,24 +146,24 @@ struct BlockwiseGemmXdlops_k0mk1_k0nk1_m0n0m1n1m2m3m4n2_v1
Number<MPerXDL>{}, Number<MPerXDL>{},
Number<NPerXDL>{})); Number<NPerXDL>{}));
return xdlops_gemm.MakeCM0N0M1N1M2M3M4N2Descriptor(c_m0_n0_m1_n1_m2_n2_block_desc); return xdlops_gemm.MakeCDescriptor_M0_N0_M1_N1_M2_M3_M4_N2(c_block_desc_m0_n0_m1_n1_m2_n2);
} }
template <typename CMNGridDesc> template <typename CGridDesc_M_N>
__host__ __device__ static constexpr auto __host__ __device__ static constexpr auto
MakeCM0N0M1N1M2M3M4N2GridDescriptor(const CMNGridDesc& c_m_n_grid_desc) MakeCGridDescriptor_M0_N0_M1_N1_M2_M3_M4_N2(const CGridDesc_M_N& c_grid_desc_m_n)
{ {
const auto c_m0_n0_m1_n1_m2_n2_grid_desc = transform_tensor_descriptor( const auto c_grid_desc_m0_n0_m1_n1_m2_n2 = transform_tensor_descriptor(
c_m_n_grid_desc, c_grid_desc_m_n,
make_tuple(make_unmerge_transform(make_tuple(MRepeat, MWaves, MPerXDL)), make_tuple(make_unmerge_transform(make_tuple(MRepeat, MWaves, MPerXDL)),
make_unmerge_transform(make_tuple(NRepeat, NWaves, NPerXDL))), make_unmerge_transform(make_tuple(NRepeat, NWaves, NPerXDL))),
make_tuple(Sequence<0>{}, Sequence<1>{}), make_tuple(Sequence<0>{}, Sequence<1>{}),
make_tuple(Sequence<0, 2, 4>{}, Sequence<1, 3, 5>{})); make_tuple(Sequence<0, 2, 4>{}, Sequence<1, 3, 5>{}));
return xdlops_gemm.MakeCM0N0M1N1M2M3M4N2Descriptor(c_m0_n0_m1_n1_m2_n2_grid_desc); return xdlops_gemm.MakeCDescriptor_M0_N0_M1_N1_M2_M3_M4_N2(c_grid_desc_m0_n0_m1_n1_m2_n2);
} }
__host__ __device__ static constexpr auto MakeAK0M0M1M2K1BlockDescriptor() __host__ __device__ static constexpr auto MakeABlockDescriptor_K0_M0_M1_M2_K1()
{ {
return transform_tensor_descriptor( return transform_tensor_descriptor(
AK0MK1BlockDesc{}, AK0MK1BlockDesc{},
...@@ -175,7 +175,7 @@ struct BlockwiseGemmXdlops_k0mk1_k0nk1_m0n0m1n1m2m3m4n2_v1 ...@@ -175,7 +175,7 @@ struct BlockwiseGemmXdlops_k0mk1_k0nk1_m0n0m1n1m2m3m4n2_v1
make_tuple(Sequence<0>{}, Sequence<1, 2, 3>{}, Sequence<4>{})); make_tuple(Sequence<0>{}, Sequence<1, 2, 3>{}, Sequence<4>{}));
} }
__host__ __device__ static constexpr auto MakeBK0N0N1N2K1BlockDescriptor() __host__ __device__ static constexpr auto MakeBBlockDescriptor_K0_N0_N1_N2_K1()
{ {
return transform_tensor_descriptor( return transform_tensor_descriptor(
BK0NK1BlockDesc{}, BK0NK1BlockDesc{},
...@@ -187,8 +187,8 @@ struct BlockwiseGemmXdlops_k0mk1_k0nk1_m0n0m1n1m2m3m4n2_v1 ...@@ -187,8 +187,8 @@ struct BlockwiseGemmXdlops_k0mk1_k0nk1_m0n0m1n1m2m3m4n2_v1
make_tuple(Sequence<0>{}, Sequence<1, 2, 3>{}, Sequence<4>{})); make_tuple(Sequence<0>{}, Sequence<1, 2, 3>{}, Sequence<4>{}));
} }
static constexpr auto a_k0_m0_m1_m2_k1_block_desc = MakeAK0M0M1M2K1BlockDescriptor(); static constexpr auto a_block_desc_k0_m0_m1_m2_k1 = MakeABlockDescriptor_K0_M0_M1_M2_K1();
static constexpr auto b_k0_n0_n1_n2_k1_block_desc = MakeBK0N0N1N2K1BlockDescriptor(); static constexpr auto b_block_desc_k0_n0_n1_n2_k1 = MakeBBlockDescriptor_K0_N0_N1_N2_K1();
template <typename ABlockBuffer, typename BBlockBuffer, typename CThreadBuffer> template <typename ABlockBuffer, typename BBlockBuffer, typename CThreadBuffer>
__device__ void Run(const ABlockBuffer& a_block_buf, __device__ void Run(const ABlockBuffer& a_block_buf,
...@@ -202,7 +202,7 @@ struct BlockwiseGemmXdlops_k0mk1_k0nk1_m0n0m1n1m2m3m4n2_v1 ...@@ -202,7 +202,7 @@ struct BlockwiseGemmXdlops_k0mk1_k0nk1_m0n0m1n1m2m3m4n2_v1
static_for<0, MRepeat, 1>{}([&](auto m0) { static_for<0, MRepeat, 1>{}([&](auto m0) {
// read A // read A
a_thread_copy_.Run(a_k0_m0_m1_m2_k1_block_desc, a_thread_copy_.Run(a_block_desc_k0_m0_m1_m2_k1,
make_tuple(I0, m0, I0, I0, I0), make_tuple(I0, m0, I0, I0, I0),
a_block_buf, a_block_buf,
a_thread_desc_, a_thread_desc_,
...@@ -211,7 +211,7 @@ struct BlockwiseGemmXdlops_k0mk1_k0nk1_m0n0m1n1m2m3m4n2_v1 ...@@ -211,7 +211,7 @@ struct BlockwiseGemmXdlops_k0mk1_k0nk1_m0n0m1n1m2m3m4n2_v1
static_for<0, NRepeat, 1>{}([&](auto n0) { static_for<0, NRepeat, 1>{}([&](auto n0) {
// read B // read B
b_thread_copy_.Run(b_k0_n0_n1_n2_k1_block_desc, b_thread_copy_.Run(b_block_desc_k0_n0_n1_n2_k1,
make_tuple(I0, n0, I0, I0, I0), make_tuple(I0, n0, I0, I0, I0),
b_block_buf, b_block_buf,
b_thread_desc_, b_thread_desc_,
...@@ -256,7 +256,7 @@ struct BlockwiseGemmXdlops_k0mk1_k0nk1_m0n0m1n1m2m3m4n2_v1 ...@@ -256,7 +256,7 @@ struct BlockwiseGemmXdlops_k0mk1_k0nk1_m0n0m1n1m2m3m4n2_v1
using AThreadCopy = ThreadwiseTensorSliceTransfer_v4<FloatAB, using AThreadCopy = ThreadwiseTensorSliceTransfer_v4<FloatAB,
FloatAB, FloatAB,
decltype(a_k0_m0_m1_m2_k1_block_desc), decltype(a_block_desc_k0_m0_m1_m2_k1),
decltype(a_thread_desc_), decltype(a_thread_desc_),
Sequence<K0, 1, 1, 1, K1>, Sequence<K0, 1, 1, 1, K1>,
Sequence<0, 1, 2, 3, 4>, Sequence<0, 1, 2, 3, 4>,
...@@ -266,7 +266,7 @@ struct BlockwiseGemmXdlops_k0mk1_k0nk1_m0n0m1n1m2m3m4n2_v1 ...@@ -266,7 +266,7 @@ struct BlockwiseGemmXdlops_k0mk1_k0nk1_m0n0m1n1m2m3m4n2_v1
using BThreadCopy = ThreadwiseTensorSliceTransfer_v4<FloatAB, using BThreadCopy = ThreadwiseTensorSliceTransfer_v4<FloatAB,
FloatAB, FloatAB,
decltype(b_k0_n0_n1_n2_k1_block_desc), decltype(b_block_desc_k0_n0_n1_n2_k1),
decltype(b_thread_desc_), decltype(b_thread_desc_),
Sequence<K0, 1, 1, 1, K1>, Sequence<K0, 1, 1, 1, K1>,
Sequence<0, 1, 2, 3, 4>, Sequence<0, 1, 2, 3, 4>,
......
composable_kernel/include/tensor_operation/threadwise_tensor_slice_transfer.hpp
View file @ ed068043
...@@ -413,7 +413,7 @@ struct ThreadwiseTensorSliceTransfer_v2 ...@@ -413,7 +413,7 @@ struct ThreadwiseTensorSliceTransfer_v2
"wrong! SrcDesc need to known at compile-time"); "wrong! SrcDesc need to known at compile-time");
} }
__device__ void SetDstSliceOrigin(const SrcDesc& src_desc, const Index& src_slice_origin_idx) __device__ void SetSrcSliceOrigin(const SrcDesc& src_desc, const Index& src_slice_origin_idx)
{ {
src_coord_ = make_tensor_coordinate(src_desc, src_slice_origin_idx); src_coord_ = make_tensor_coordinate(src_desc, src_slice_origin_idx);
} }
......
composable_kernel/include/tensor_operation/xdlops_gemm.hpp
View file @ ed068043
...@@ -644,17 +644,17 @@ struct XdlopsGemm ...@@ -644,17 +644,17 @@ struct XdlopsGemm
static_assert(KPack % mfma_instr.k_per_blk == 0, "KPack cannot be divided by k_per_blk"); static_assert(KPack % mfma_instr.k_per_blk == 0, "KPack cannot be divided by k_per_blk");
} }
template <typename CM0N0M1N1M2N2Desc> template <typename CDesc_M0_N0_M1_N1_M2_N2>
__host__ __device__ static constexpr auto __host__ __device__ static constexpr auto
MakeCM0N0M1N1M2M3M4N2Descriptor(const CM0N0M1N1M2N2Desc& c_m0_n0_m1_n1_m2_n2_desc) MakeCDescriptor_M0_N0_M1_N1_M2_M3_M4_N2(const CDesc_M0_N0_M1_N1_M2_N2& c_desc_m0_n0_m1_n1_m2_n2)
{ {
const auto M0 = c_m0_n0_m1_n1_m2_n2_desc.GetLength(I0); const auto M0 = c_desc_m0_n0_m1_n1_m2_n2.GetLength(I0);
const auto N0 = c_m0_n0_m1_n1_m2_n2_desc.GetLength(I1); const auto N0 = c_desc_m0_n0_m1_n1_m2_n2.GetLength(I1);
const auto M1 = c_m0_n0_m1_n1_m2_n2_desc.GetLength(I2); const auto M1 = c_desc_m0_n0_m1_n1_m2_n2.GetLength(I2);
const auto N1 = c_m0_n0_m1_n1_m2_n2_desc.GetLength(I3); const auto N1 = c_desc_m0_n0_m1_n1_m2_n2.GetLength(I3);
return transform_tensor_descriptor( return transform_tensor_descriptor(
c_m0_n0_m1_n1_m2_n2_desc, c_desc_m0_n0_m1_n1_m2_n2,
make_tuple(make_pass_through_transform(M0), make_tuple(make_pass_through_transform(M0),
make_pass_through_transform(N0), make_pass_through_transform(N0),
make_pass_through_transform(M1), make_pass_through_transform(M1),
......
composable_kernel/include/utility/config.hpp
View file @ ed068043
...@@ -95,7 +95,7 @@ ...@@ -95,7 +95,7 @@
#define CK_EXPERIMENTAL_STATIC_TENSOR_DESCRIPTOR 1 #define CK_EXPERIMENTAL_STATIC_TENSOR_DESCRIPTOR 1
// merge transformation use magic number division // merge transformation use magic number division
#define CK_EXPERIMENTAL_MERGE_USE_MAGIC_DIVISION 0 #define CK_EXPERIMENTAL_MERGE_USE_MAGIC_DIVISION 1
// hack: have underlying assumption that need to be satsified, otherwise it's a bug // hack: have underlying assumption that need to be satsified, otherwise it's a bug
// hack for forcing register to keep idx_diff_low_const in SGPR. idx_diff_low_const must be // hack for forcing register to keep idx_diff_low_const in SGPR. idx_diff_low_const must be
......
composable_kernel/include/utility/type.hpp
View file @ ed068043
...@@ -16,6 +16,9 @@ struct is_same<X, X> : public integral_constant<bool, true> ...@@ -16,6 +16,9 @@ struct is_same<X, X> : public integral_constant<bool, true>
{ {
}; };
template <typename X, typename Y>
inline constexpr bool is_same_v = is_same<X, Y>::value;
template <typename T> template <typename T>
using remove_reference_t = typename std::remove_reference<T>::type; using remove_reference_t = typename std::remove_reference<T>::type;
......
composable_kernel/src/kernel_wrapper/convolution_forward_implicit_gemm_v4r4_xdlops_nhwc_kyxc_nhwk.cpp
View file @ ed068043
...@@ -92,7 +92,7 @@ extern "C" __global__ void convolution_forward_implicit_gemm_v4r4_xdlops_nhwc_ky ...@@ -92,7 +92,7 @@ extern "C" __global__ void convolution_forward_implicit_gemm_v4r4_xdlops_nhwc_ky
const auto wei_k_y_x_c_desc = make_naive_tensor_descriptor_packed(make_tuple(k, y, x, c)); const auto wei_k_y_x_c_desc = make_naive_tensor_descriptor_packed(make_tuple(k, y, x, c));
const auto out_n_ho_wo_k_desc = make_naive_tensor_descriptor_packed(make_tuple(n, ho, wo, k)); const auto out_n_ho_wo_k_desc = make_naive_tensor_descriptor_packed(make_tuple(n, ho, wo, k));
const auto descs = transform_forward_convolution_into_gemm_v4r4r4_nhwc_kyxc_nhwk_pad( const auto descs = transform_forward_convolution_into_gemm_v4r4r4_nhwc_kyxc_nhwk(
in_n_hi_wi_c_desc, in_n_hi_wi_c_desc,
wei_k_y_x_c_desc, wei_k_y_x_c_desc,
out_n_ho_wo_k_desc, out_n_ho_wo_k_desc,
...@@ -230,14 +230,14 @@ extern "C" __global__ void ...@@ -230,14 +230,14 @@ extern "C" __global__ void
make_naive_tensor_descriptor_packed(make_tuple(256, 28, 28, 256)); make_naive_tensor_descriptor_packed(make_tuple(256, 28, 28, 256));
constexpr auto descs = constexpr auto descs =
transform_forward_convolution_into_gemm_v4r4r4_nhwc_kyxc_nhwk_pad(in_n_hi_wi_c_desc, transform_forward_convolution_into_gemm_v4r4r4_nhwc_kyxc_nhwk(in_n_hi_wi_c_desc,
wei_k_y_x_c_desc, wei_k_y_x_c_desc,
out_n_ho_wo_k_desc, out_n_ho_wo_k_desc,
make_tuple(1, 1), make_tuple(1, 1),
make_tuple(1, 1), make_tuple(1, 1),
make_tuple(1, 1), make_tuple(1, 1),
make_tuple(1, 1), make_tuple(1, 1),
Number<K1>{}); Number<K1>{});
constexpr auto a_k0_m_k1_grid_desc_tmp = descs[I0]; constexpr auto a_k0_m_k1_grid_desc_tmp = descs[I0];
constexpr auto b_k0_n_k1_grid_desc_tmp = descs[I1]; constexpr auto b_k0_n_k1_grid_desc_tmp = descs[I1];
......
device_operation/device_conv_xdl_instance_f16_f16_f16_nhwc_kyxc_nhwk.cpp 0 → 100644
View file @ ed068043
#include <stdlib.h>
#include "config.hpp"
#include "device_conv_fwd_xdl_nhwc_kyxc_nhwk.hpp"
#include "device_conv_instance.hpp"
namespace ck {
namespace tensor_operation {
namespace device {
namespace device_conv_instance {
using F16 = ck::half_t;
using F32 = float;
using NHWC = ck::tensor_layout::convolution::NHWC;
using KYXC = ck::tensor_layout::convolution::KYXC;
using NHWK = ck::tensor_layout::convolution::NHWK;
template <ck::index_t... Is>
using S = ck::Sequence<Is...>;
// Compilation parameters for in[n, hi, wi, c] * wei[k, y, x, c] = out[n, ho, wo, k]
using device_conv_fwd_xdl_instances_f16_f16_f16_nhwc_kyxc_nhwk = std::tuple<
// clang-format off
//##############| NDim| InData| WeiData| OutData| AccData| In| Wei| Out| Block| MPer| NPer| K0Per| K1| MPer| NPer| MXdl| NXdl| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| BBlockTransfer| BBlockTransfer| BBlockTransfer| BBlockTransfer| BlockTransfer| BBlockTransfer| BBlockTransfer| CThreadTransfer| CThreadTransfer| ABlockLds| BBlockLds|
//##############| Spatial| Type| Type| Type| Type| Layout| Layout| Layout| Size| Block| Block| Block| | XDL| XDL| Per| Per| ThreadSlice| ThreadCluster| ThreadCluster| SrcAccessOrder| SrcVectorDim| SrcScalar| DstScalar| ThreadSlice| ThreadCluster| ThreadCluster| SrcAccessOrder| SrcVectorDim| SrcScalar| DstScalar| SrcDstVectorDim| DstScalar| AddExtraM| AddExtraN|
//##############| | | | | | | | | | | | | | | | Wave| Wave| Lengths_K0_N_K1| Lengths_K0_M_K1| ArrangeOrder| | | PerVector| PerVector_K1| Lengths_K0_N_K1| Lengths_K0_N_K1| ArrangeOrder| | | PerVector| PerVector_K1| | PerVector| | |
//##############| | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | |
DeviceConvFwdXdl< 2, F16, F16, F16, F32, NHWC, KYXC, NHWK, 256, 256, 128, 4, 8, 32, 32, 4, 2, S<1, 4, 8>, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, S<1, 2, 8>, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 7, 1, true, true>,
DeviceConvFwdXdl< 2, F16, F16, F16, F32, NHWC, KYXC, NHWK, 256, 128, 256, 4, 8, 32, 32, 2, 4, S<1, 2, 8>, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, S<1, 4, 8>, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 7, 1, true, true>,
DeviceConvFwdXdl< 2, F16, F16, F16, F32, NHWC, KYXC, NHWK, 128, 128, 128, 4, 8, 32, 32, 4, 2, S<1, 4, 8>, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, S<1, 4, 8>, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 7, 1, true, true>,
DeviceConvFwdXdl< 2, F16, F16, F16, F32, NHWC, KYXC, NHWK, 256, 128, 128, 4, 8, 32, 32, 2, 2, S<1, 2, 8>, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, S<1, 2, 8>, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 7, 1, true, true>,
DeviceConvFwdXdl< 2, F16, F16, F16, F32, NHWC, KYXC, NHWK, 128, 128, 64, 4, 8, 32, 32, 2, 2, S<1, 4, 8>, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, S<1, 2, 8>, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 7, 1, true, true>,
DeviceConvFwdXdl< 2, F16, F16, F16, F32, NHWC, KYXC, NHWK, 128, 64, 128, 4, 8, 32, 32, 2, 2, S<1, 2, 8>, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, S<1, 4, 8>, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 7, 1, true, true>,
DeviceConvFwdXdl< 2, F16, F16, F16, F32, NHWC, KYXC, NHWK, 64, 64, 64, 4, 8, 32, 32, 2, 2, S<1, 4, 8>, S<4, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, S<1, 4, 8>, S<4, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 7, 1, true, true>,
DeviceConvFwdXdl< 2, F16, F16, F16, F32, NHWC, KYXC, NHWK, 256, 128, 64, 4, 8, 32, 32, 2, 1, S<1, 2, 8>, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, S<1, 1, 8>, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 7, 1, true, true>,
DeviceConvFwdXdl< 2, F16, F16, F16, F32, NHWC, KYXC, NHWK, 256, 64, 128, 4, 8, 32, 32, 1, 2, S<1, 1, 8>, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, S<1, 2, 8>, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 7, 1, true, true>,
DeviceConvFwdXdl< 2, F16, F16, F16, F32, NHWC, KYXC, NHWK, 128, 128, 32, 4, 8, 32, 32, 2, 1, S<1, 4, 8>, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, S<1, 1, 8>, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 7, 1, true, true>,
DeviceConvFwdXdl< 2, F16, F16, F16, F32, NHWC, KYXC, NHWK, 128, 32, 128, 4, 8, 32, 32, 1, 2, S<1, 1, 8>, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, S<1, 4, 8>, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 7, 1, true, true>,
DeviceConvFwdXdl< 2, F16, F16, F16, F32, NHWC, KYXC, NHWK, 64, 64, 32, 4, 8, 32, 32, 2, 1, S<1, 4, 8>, S<4, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, S<1, 2, 8>, S<4, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 7, 1, true, true>,
DeviceConvFwdXdl< 2, F16, F16, F16, F32, NHWC, KYXC, NHWK, 64, 32, 64, 4, 8, 32, 32, 1, 2, S<1, 2, 8>, S<4, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, S<1, 4, 8>, S<4, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 7, 1, true, true>
// clang-format on
>;
template <>
void add_device_conv_fwd_instance<2, F16, F16, F16, NHWC, KYXC, NHWK>(
std::vector<DeviceConvFwdPtr>& device_conv_instances)
{
using DeviceConvs = device_conv_fwd_xdl_instances_f16_f16_f16_nhwc_kyxc_nhwk;
const auto device_convs = DeviceConvs{};
ck::static_for<0, std::tuple_size_v<DeviceConvs>, 1>{}([&](auto i) {
using Conv = remove_cvref_t<decltype(std::get<i>(device_convs))>;
auto conv = Conv{};
device_conv_instances.push_back(std::make_unique<Conv>(conv));
});
}
} // namespace device_conv_instance
} // namespace device
} // namespace tensor_operation
} // namespace ck
device_operation/device_conv_xdl_instance_f32_f32_f32_nhwc_kyxc_nhwk.cpp 0 → 100644
View file @ ed068043
#include <stdlib.h>
#include "config.hpp"
#include "device_conv_fwd_xdl_nhwc_kyxc_nhwk.hpp"
#include "device_conv_instance.hpp"
namespace ck {
namespace tensor_operation {
namespace device {
namespace device_conv_instance {
using F16 = ck::half_t;
using F32 = float;
using NHWC = ck::tensor_layout::convolution::NHWC;
using KYXC = ck::tensor_layout::convolution::KYXC;
using NHWK = ck::tensor_layout::convolution::NHWK;
template <ck::index_t... Is>
using S = ck::Sequence<Is...>;
// Compilation parameters for in[n, hi, wi, c] * wei[k, y, x, c] = out[n, ho, wo, k]
using device_conv_fwd_xdl_instances_f32_f32_f32_nhwc_kyxc_nhwk = std::tuple<
// clang-format off
//##############| NDim| InData| WeiData| OutData| AccData| In| Wei| Out| Block| MPer| NPer| K0Per| K1| MPer| NPer| MXdl| NXdl| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| BBlockTransfer| BBlockTransfer| BBlockTransfer| BBlockTransfer| BlockTransfer| BBlockTransfer| BBlockTransfer| CThreadTransfer| CThreadTransfer| ABlockLds| BBlockLds|
//##############| Spatial| Type| Type| Type| Type| Layout| Layout| Layout| Size| Block| Block| Block| | XDL| XDL| Per| Per| ThreadSlice| ThreadCluster| ThreadCluster| SrcAccessOrder| SrcVectorDim| SrcScalar| DstScalar| ThreadSlice| ThreadCluster| ThreadCluster| SrcAccessOrder| SrcVectorDim| SrcScalar| DstScalar| SrcDstVectorDim| DstScalar| AddExtraM| AddExtraN|
//##############| | | | | | | | | | | | | | | | Wave| Wave| Lengths_K0_N_K1| Lengths_K0_M_K1| ArrangeOrder| | | PerVector| PerVector_K1| Lengths_K0_N_K1| Lengths_K0_N_K1| ArrangeOrder| | | PerVector| PerVector_K1| | PerVector| | |
//##############| | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | |
DeviceConvFwdXdl< 2, F32, F32, F32, F32, NHWC, KYXC, NHWK, 256, 256, 128, 4, 4, 32, 32, 4, 2, S<1, 4, 4>, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 4, 4, S<1, 2, 4>, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 4, 4, 7, 1, true, true>,
DeviceConvFwdXdl< 2, F32, F32, F32, F32, NHWC, KYXC, NHWK, 256, 128, 256, 4, 4, 32, 32, 2, 4, S<1, 2, 4>, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 4, 4, S<1, 4, 4>, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 4, 4, 7, 1, true, true>,
DeviceConvFwdXdl< 2, F32, F32, F32, F32, NHWC, KYXC, NHWK, 128, 128, 128, 4, 4, 32, 32, 4, 2, S<1, 4, 4>, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 4, 4, S<1, 4, 4>, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 4, 4, 7, 1, true, true>,
DeviceConvFwdXdl< 2, F32, F32, F32, F32, NHWC, KYXC, NHWK, 256, 128, 128, 4, 4, 32, 32, 2, 2, S<1, 2, 4>, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 4, 4, S<1, 2, 4>, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 4, 4, 7, 1, true, true>,
DeviceConvFwdXdl< 2, F32, F32, F32, F32, NHWC, KYXC, NHWK, 128, 128, 64, 4, 4, 32, 32, 2, 2, S<1, 4, 4>, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 4, 4, S<1, 2, 4>, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 4, 4, 7, 1, true, true>,
DeviceConvFwdXdl< 2, F32, F32, F32, F32, NHWC, KYXC, NHWK, 128, 64, 128, 4, 4, 32, 32, 2, 2, S<1, 2, 4>, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 4, 4, S<1, 4, 4>, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 4, 4, 7, 1, true, true>,
DeviceConvFwdXdl< 2, F32, F32, F32, F32, NHWC, KYXC, NHWK, 64, 64, 64, 4, 4, 32, 32, 2, 2, S<1, 4, 4>, S<4, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 4, 4, S<1, 4, 4>, S<4, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 4, 4, 7, 1, true, true>,
DeviceConvFwdXdl< 2, F32, F32, F32, F32, NHWC, KYXC, NHWK, 256, 128, 64, 4, 4, 32, 32, 2, 1, S<1, 2, 4>, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 4, 4, S<1, 1, 4>, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 4, 4, 7, 1, true, true>,
DeviceConvFwdXdl< 2, F32, F32, F32, F32, NHWC, KYXC, NHWK, 256, 64, 128, 4, 4, 32, 32, 1, 2, S<1, 1, 4>, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 4, 4, S<1, 2, 4>, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 4, 4, 7, 1, true, true>,
DeviceConvFwdXdl< 2, F32, F32, F32, F32, NHWC, KYXC, NHWK, 128, 128, 32, 4, 4, 32, 32, 2, 1, S<1, 4, 4>, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 4, 4, S<1, 1, 4>, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 4, 4, 7, 1, true, true>,
DeviceConvFwdXdl< 2, F32, F32, F32, F32, NHWC, KYXC, NHWK, 128, 32, 128, 4, 4, 32, 32, 1, 2, S<1, 1, 4>, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 4, 4, S<1, 4, 4>, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 4, 4, 7, 1, true, true>,
DeviceConvFwdXdl< 2, F32, F32, F32, F32, NHWC, KYXC, NHWK, 64, 64, 32, 4, 4, 32, 32, 2, 1, S<1, 4, 4>, S<4, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 4, 4, S<1, 2, 4>, S<4, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 4, 4, 7, 1, true, true>,
DeviceConvFwdXdl< 2, F32, F32, F32, F32, NHWC, KYXC, NHWK, 64, 32, 64, 4, 4, 32, 32, 1, 2, S<1, 2, 4>, S<4, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 4, 4, S<1, 4, 4>, S<4, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 4, 4, 7, 1, true, true>
// clang-format on
>;
template <>
void add_device_conv_fwd_instance<2, F32, F32, F32, NHWC, KYXC, NHWK>(
std::vector<DeviceConvFwdPtr>& device_conv_instances)
{
using DeviceConvs = device_conv_fwd_xdl_instances_f32_f32_f32_nhwc_kyxc_nhwk;
const auto device_convs = DeviceConvs{};
ck::static_for<0, std::tuple_size_v<DeviceConvs>, 1>{}([&](auto i) {
using Conv = remove_cvref_t<decltype(std::get<i>(device_convs))>;
auto conv = Conv{};
device_conv_instances.push_back(std::make_unique<Conv>(conv));
});
}
} // namespace device_conv_instance
} // namespace device
} // namespace tensor_operation
} // namespace ck
device_operation/device_gemm_xdl_instance_f16_f16_f16_km_kn_mn.cpp 0 → 100644
View file @ ed068043
#include <stdlib.h>
#include "config.hpp"
#include "device_gemm_xdl.hpp"
#include "device_gemm_instance.hpp"
namespace ck {
namespace tensor_operation {
namespace device {
namespace device_gemm_instance {
using F16 = ck::half_t;
using F32 = float;
using Row = ck::tensor_layout::gemm::RowMajor;
using Col = ck::tensor_layout::gemm::ColumnMajor;
template <ck::index_t... Is>
using S = ck::Sequence<Is...>;
// Compilation parameters for a[k, m] * b[k, n] = c[m, n]
using device_gemm_xdl_instance_f16_f16_f16_km_kn_mn = std::tuple<
// clang-format off
//##########| AData| BData| CData| AccData| ALayout| BLayout| CLayout| Block| MPer| NPer| K0Per| K1| MPer| NPer| MXdl| NXdl| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| BBlockTransfer| BBlockTransfer| BBlockTransfer| BBlockTransfer| BlockTransfer| BBlockTransfer| BBlockTransfer| CThreadTransfer| CThreadTransfer| ABlockLds| BBlockLds|
//##########| Type| Type| Type| Type| | | | Size| Block| Block| Block| | XDL| XDL| Per| Per| ThreadSlice| ThreadCluster| ThreadCluster| SrcAccessOrder| SrcVectorDim| SrcScalar| DstScalar| ThreadSlice| ThreadCluster| ThreadCluster| SrcAccessOrder| SrcVectorDim| SrcScalar| DstScalar| SrcDstVectorDim| DstScalar| AddExtraM| AddExtraN|
//##########| | | | | | | | | | | | | | | Wave| Wave| Lengths_K0_N_K1| Lengths_K0_M_K1| ArrangeOrder| | | PerVector| PerVector_K1| Lengths_K0_N_K1| Lengths_K0_N_K1| ArrangeOrder| | | PerVector| PerVector_K1| | PerVector| | |
//##########| | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | |
DeviceGemmXdl< F16, F16, F16, F32, Col, Row, Row, 256, 256, 128, 4, 8, 32, 32, 4, 2, S<1, 4, 8>, S<4, 64, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 8, S<1, 2, 8>, S<4, 64, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 2, 8, 7, 1, true, true>,
DeviceGemmXdl< F16, F16, F16, F32, Col, Row, Row, 256, 128, 256, 4, 8, 32, 32, 2, 4, S<1, 2, 8>, S<4, 64, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 2, 8, S<1, 4, 8>, S<4, 64, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 8, 7, 1, true, true>,
DeviceGemmXdl< F16, F16, F16, F32, Col, Row, Row, 128, 128, 128, 4, 8, 32, 32, 4, 2, S<1, 4, 8>, S<4, 32, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 8, S<1, 4, 8>, S<4, 32, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 8, 7, 1, true, true>,
DeviceGemmXdl< F16, F16, F16, F32, Col, Row, Row, 256, 128, 128, 4, 8, 32, 32, 2, 2, S<1, 2, 8>, S<4, 64, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 2, 8, S<1, 2, 8>, S<4, 64, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 2, 8, 7, 1, true, true>,
DeviceGemmXdl< F16, F16, F16, F32, Col, Row, Row, 128, 128, 64, 4, 8, 32, 32, 2, 2, S<1, 4, 8>, S<4, 32, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 8, S<1, 2, 8>, S<4, 32, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 2, 8, 7, 1, true, true>,
DeviceGemmXdl< F16, F16, F16, F32, Col, Row, Row, 128, 64, 128, 4, 8, 32, 32, 2, 2, S<1, 2, 8>, S<4, 32, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 2, 8, S<1, 4, 8>, S<4, 32, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 8, 7, 1, true, true>,
DeviceGemmXdl< F16, F16, F16, F32, Col, Row, Row, 256, 128, 64, 4, 8, 32, 32, 2, 1, S<1, 2, 8>, S<4, 64, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 2, 8, S<1, 1, 8>, S<4, 64, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 1, 8, 7, 1, true, true>,
DeviceGemmXdl< F16, F16, F16, F32, Col, Row, Row, 256, 64, 128, 4, 8, 32, 32, 1, 2, S<1, 1, 8>, S<4, 64, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 1, 8, S<1, 2, 8>, S<4, 64, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 2, 8, 7, 1, true, true>
// clang-format on
>;
template <>
void add_device_gemm_instance<F16, F16, F16, Col, Row, Row>(
std::vector<DeviceGemmPtr>& device_op_instances)
{
using DeviceGemms = device_gemm_instance::device_gemm_xdl_instance_f16_f16_f16_km_kn_mn;
const auto device_gemms = DeviceGemms{};
ck::static_for<0, std::tuple_size_v<DeviceGemms>, 1>{}([&](auto i) {
using Gemm = remove_cvref_t<decltype(std::get<i>(device_gemms))>;
auto gemm = Gemm{};
device_op_instances.push_back(std::make_unique<Gemm>(gemm));
});
}
} // namespace device_gemm_instance
} // namespace device
} // namespace tensor_operation
} // namespace ck
device_operation/device_gemm_xdl_instance_f16_f16_f16_km_nk_mn.cpp 0 → 100644
View file @ ed068043
#include <stdlib.h>
#include "config.hpp"
#include "device_gemm_xdl.hpp"
#include "device_gemm_instance.hpp"
namespace ck {
namespace tensor_operation {
namespace device {
namespace device_gemm_instance {
using F16 = ck::half_t;
using F32 = float;
using Row = ck::tensor_layout::gemm::RowMajor;
using Col = ck::tensor_layout::gemm::ColumnMajor;
template <ck::index_t... Is>
using S = ck::Sequence<Is...>;
// Compilation parameters for a[k, m] * b[n, k] = c[m, n]
using device_gemm_xdl_instance_f16_f16_f16_km_nk_mn = std::tuple<
// clang-format off
//##########| AData| BData| CData| AccData| ALayout| BLayout| CLayout| Block| MPer| NPer| K0Per| K1| MPer| NPer| MXdl| NXdl| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| BBlockTransfer| BBlockTransfer| BBlockTransfer| BBlockTransfer| BlockTransfer| BBlockTransfer| BBlockTransfer| CThreadTransfer| CThreadTransfer| ABlockLds| BBlockLds|
//##########| Type| Type| Type| Type| | | | Size| Block| Block| Block| | XDL| XDL| Per| Per| ThreadSlice| ThreadCluster| ThreadCluster| SrcAccessOrder| SrcVectorDim| SrcScalar| DstScalar| ThreadSlice| ThreadCluster| ThreadCluster| SrcAccessOrder| SrcVectorDim| SrcScalar| DstScalar| SrcDstVectorDim| DstScalar| AddExtraM| AddExtraN|
//##########| | | | | | | | | | | | | | | Wave| Wave| Lengths_K0_N_K1| Lengths_K0_M_K1| ArrangeOrder| | | PerVector| PerVector_K1| Lengths_K0_N_K1| Lengths_K0_N_K1| ArrangeOrder| | | PerVector| PerVector_K1| | PerVector| | |
//##########| | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | |
DeviceGemmXdl< F16, F16, F16, F32, Col, Col, Row, 256, 256, 128, 4, 8, 32, 32, 4, 2, S<1, 4, 8>, S<4, 64, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 8, S<1, 2, 8>, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 7, 1, true, true>,
DeviceGemmXdl< F16, F16, F16, F32, Col, Col, Row, 256, 128, 256, 4, 8, 32, 32, 2, 4, S<1, 2, 8>, S<4, 64, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 2, 8, S<1, 4, 8>, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 7, 1, true, true>,
DeviceGemmXdl< F16, F16, F16, F32, Col, Col, Row, 128, 128, 128, 4, 8, 32, 32, 4, 2, S<1, 4, 8>, S<4, 32, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 8, S<1, 4, 8>, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 7, 1, true, true>,
DeviceGemmXdl< F16, F16, F16, F32, Col, Col, Row, 256, 128, 128, 4, 8, 32, 32, 2, 2, S<1, 2, 8>, S<4, 64, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 2, 8, S<1, 2, 8>, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 7, 1, true, true>,
DeviceGemmXdl< F16, F16, F16, F32, Col, Col, Row, 128, 128, 64, 4, 8, 32, 32, 2, 2, S<1, 4, 8>, S<4, 32, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 8, S<1, 2, 8>, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 7, 1, true, true>,
DeviceGemmXdl< F16, F16, F16, F32, Col, Col, Row, 128, 64, 128, 4, 8, 32, 32, 2, 2, S<1, 2, 8>, S<4, 32, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 2, 8, S<1, 4, 8>, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 7, 1, true, true>,
DeviceGemmXdl< F16, F16, F16, F32, Col, Col, Row, 256, 128, 64, 4, 8, 32, 32, 2, 1, S<1, 2, 8>, S<4, 64, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 2, 8, S<1, 1, 8>, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 7, 1, true, true>,
DeviceGemmXdl< F16, F16, F16, F32, Col, Col, Row, 256, 64, 128, 4, 8, 32, 32, 1, 2, S<1, 1, 8>, S<4, 64, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 1, 8, S<1, 2, 8>, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 7, 1, true, true>
// clang-format on
>;
template <>
void add_device_gemm_instance<F16, F16, F16, Col, Col, Row>(
std::vector<DeviceGemmPtr>& device_op_instances)
{
using DeviceGemms = device_gemm_instance::device_gemm_xdl_instance_f16_f16_f16_km_nk_mn;
const auto device_gemms = DeviceGemms{};
ck::static_for<0, std::tuple_size_v<DeviceGemms>, 1>{}([&](auto i) {
using Gemm = remove_cvref_t<decltype(std::get<i>(device_gemms))>;
auto gemm = Gemm{};
device_op_instances.push_back(std::make_unique<Gemm>(gemm));
});
}
} // namespace device_gemm_instance
} // namespace device
} // namespace tensor_operation
} // namespace ck
device_operation/device_gemm_xdl_instance_f16_f16_f16_mk_kn_mn.cpp 0 → 100644
View file @ ed068043
#include <stdlib.h>
#include "config.hpp"
#include "device_gemm_xdl.hpp"
#include "device_gemm_instance.hpp"
namespace ck {
namespace tensor_operation {
namespace device {
namespace device_gemm_instance {
using F16 = ck::half_t;
using F32 = float;
using Row = ck::tensor_layout::gemm::RowMajor;
using Col = ck::tensor_layout::gemm::ColumnMajor;
template <ck::index_t... Is>
using S = ck::Sequence<Is...>;
// Compilation parameters for a[m, k] * b[k, n] = c[m, n]
using device_gemm_xdl_instance_f16_f16_f16_mk_kn_mn = std::tuple<
// clang-format off
//##########| AData| BData| CData| AccData| ALayout| BLayout| CLayout| Block| MPer| NPer| K0Per| K1| MPer| NPer| MXdl| NXdl| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| BBlockTransfer| BBlockTransfer| BBlockTransfer| BBlockTransfer| BlockTransfer| BBlockTransfer| BBlockTransfer| CThreadTransfer| CThreadTransfer| ABlockLds| BBlockLds|
//##########| Type| Type| Type| Type| | | | Size| Block| Block| Block| | XDL| XDL| Per| Per| ThreadSlice| ThreadCluster| ThreadCluster| SrcAccessOrder| SrcVectorDim| SrcScalar| DstScalar| ThreadSlice| ThreadCluster| ThreadCluster| SrcAccessOrder| SrcVectorDim| SrcScalar| DstScalar| SrcDstVectorDim| DstScalar| AddExtraM| AddExtraN|
//##########| | | | | | | | | | | | | | | Wave| Wave| Lengths_K0_N_K1| Lengths_K0_M_K1| ArrangeOrder| | | PerVector| PerVector_K1| Lengths_K0_N_K1| Lengths_K0_N_K1| ArrangeOrder| | | PerVector| PerVector_K1| | PerVector| | |
//##########| | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | |
DeviceGemmXdl< F16, F16, F16, F32, Row, Row, Row, 256, 256, 128, 4, 8, 32, 32, 4, 2, S<1, 4, 8>, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, S<1, 2, 8>, S<4, 64, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 2, 8, 7, 1, true, true>,
DeviceGemmXdl< F16, F16, F16, F32, Row, Row, Row, 256, 128, 256, 4, 8, 32, 32, 2, 4, S<1, 2, 8>, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, S<1, 4, 8>, S<4, 64, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 8, 7, 1, true, true>,
DeviceGemmXdl< F16, F16, F16, F32, Row, Row, Row, 128, 128, 128, 4, 8, 32, 32, 4, 2, S<1, 4, 8>, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, S<1, 4, 8>, S<4, 32, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 8, 7, 1, true, true>,
DeviceGemmXdl< F16, F16, F16, F32, Row, Row, Row, 256, 128, 128, 4, 8, 32, 32, 2, 2, S<1, 2, 8>, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, S<1, 2, 8>, S<4, 64, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 2, 8, 7, 1, true, true>,
DeviceGemmXdl< F16, F16, F16, F32, Row, Row, Row, 128, 128, 64, 4, 8, 32, 32, 2, 2, S<1, 4, 8>, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, S<1, 2, 8>, S<4, 32, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 2, 8, 7, 1, true, true>,
DeviceGemmXdl< F16, F16, F16, F32, Row, Row, Row, 128, 64, 128, 4, 8, 32, 32, 2, 2, S<1, 2, 8>, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, S<1, 4, 8>, S<4, 32, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 8, 7, 1, true, true>,
DeviceGemmXdl< F16, F16, F16, F32, Row, Row, Row, 256, 128, 64, 4, 8, 32, 32, 2, 1, S<1, 2, 8>, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, S<1, 1, 8>, S<4, 64, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 1, 8, 7, 1, true, true>,
DeviceGemmXdl< F16, F16, F16, F32, Row, Row, Row, 256, 64, 128, 4, 8, 32, 32, 1, 2, S<1, 1, 8>, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, S<1, 2, 8>, S<4, 64, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 2, 8, 7, 1, true, true>
// clang-format on
>;
template <>
void add_device_gemm_instance<F16, F16, F16, Row, Row, Row>(
std::vector<DeviceGemmPtr>& device_op_instances)
{
using DeviceGemms = device_gemm_instance::device_gemm_xdl_instance_f16_f16_f16_mk_kn_mn;
const auto device_gemms = DeviceGemms{};
ck::static_for<0, std::tuple_size_v<DeviceGemms>, 1>{}([&](auto i) {
using Gemm = remove_cvref_t<decltype(std::get<i>(device_gemms))>;
auto gemm = Gemm{};
device_op_instances.push_back(std::make_unique<Gemm>(gemm));
});
}
} // namespace device_gemm_instance
} // namespace device
} // namespace tensor_operation
} // namespace ck
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