Skip to content

GitLab

"vscode:/vscode.git/clone" did not exist on "5aa31bd6741a3469415bd95c91b0b47b98ba87a8"
Commit 9fb64dae authored by aska-0096's avatar aska-0096
Browse files

Merge branch 'develop' of... 

Merge branch 'develop' of https://github.com/ROCmSoftwarePlatform/composable_kernel into e2e_kernellib
parents e330961d fe96e8fb
Hide whitespace changes
Inline Side-by-side
Showing with 824 additions and 78 deletions
CHANGELOG.md
View file @ 9fb64dae
...@@ -2,7 +2,7 @@ ...@@ -2,7 +2,7 @@
Full documentation for Composable Kernel is not yet available. Full documentation for Composable Kernel is not yet available.
## CK 0.1.1 for ROCm 5.5.0 ## CK 0.2.0 for ROCm 5.5.0
### Fixed ### Fixed
- Fixed a bug in 6-dimensional kernels (#555). - Fixed a bug in 6-dimensional kernels (#555).
...@@ -12,6 +12,7 @@ Full documentation for Composable Kernel is not yet available. ...@@ -12,6 +12,7 @@ Full documentation for Composable Kernel is not yet available.
- Improve proformance of normalization kernel - Improve proformance of normalization kernel
### Added ### Added
- Added support on NAVI3x.
- Added user tutorial (#563). - Added user tutorial (#563).
- Added more instances for irregular GEMM sizes (#560). - Added more instances for irregular GEMM sizes (#560).
- Added inter-wave consumer-producer programming model for GEMM kernels (#310). - Added inter-wave consumer-producer programming model for GEMM kernels (#310).
......
Jenkinsfile
View file @ 9fb64dae
...@@ -684,8 +684,8 @@ pipeline { ...@@ -684,8 +684,8 @@ pipeline {
} }
agent{ label rocmnode("navi21") } agent{ label rocmnode("navi21") }
environment{ environment{
setup_args = """ -DCMAKE_INSTALL_PREFIX=../install -DGPU_TARGETS="gfx1030" """ setup_args = """ -DCMAKE_INSTALL_PREFIX=../install """
execute_args = """ cd ../client_example && rm -rf build && mkdir build && cd build && cmake -D CMAKE_PREFIX_PATH="${env.WORKSPACE}/install;/opt/rocm" -DGPU_TARGETS="gfx1030" -D CMAKE_CXX_COMPILER="${build_compiler()}" .. && make -j """ execute_args = """ cd ../client_example && rm -rf build && mkdir build && cd build && cmake -D CMAKE_PREFIX_PATH="${env.WORKSPACE}/install;/opt/rocm" -DGPU_TARGETS="gfx1030;gfx1100;gfx1101;gfx1102" -D CMAKE_CXX_COMPILER="${build_compiler()}" .. && make -j """
} }
steps{ steps{
......
client_example/09_quantization/CMakeLists.txt
View file @ 9fb64dae
...@@ -9,3 +9,6 @@ target_link_libraries(client_conv2d_fwd_perchannel_quantization PRIVATE composab ...@@ -9,3 +9,6 @@ target_link_libraries(client_conv2d_fwd_perchannel_quantization PRIVATE composab
add_executable(client_conv2d_fwd_perlayer_quantization conv2d_fwd_perlayer_quantization.cpp) add_executable(client_conv2d_fwd_perlayer_quantization conv2d_fwd_perlayer_quantization.cpp)
target_link_libraries(client_conv2d_fwd_perlayer_quantization PRIVATE composable_kernel::device_operations) target_link_libraries(client_conv2d_fwd_perlayer_quantization PRIVATE composable_kernel::device_operations)
add_executable(client_gemm_quantization gemm_quantization.cpp)
target_link_libraries(client_gemm_quantization PRIVATE composable_kernel::device_operations)
client_example/09_quantization/conv2d_fwd_bias_relu_perchannel_quantization.cpp
View file @ 9fb64dae
...@@ -28,16 +28,15 @@ using OutElementOp = ck::tensor_operation::element_wise::Add_Activation_Mul2_Cla ...@@ -28,16 +28,15 @@ using OutElementOp = ck::tensor_operation::element_wise::Add_Activation_Mul2_Cla
static constexpr ck::index_t NumDimSpatial = 2; static constexpr ck::index_t NumDimSpatial = 2;
static constexpr ck::index_t G = 1; static constexpr ck::index_t G = 1;
static constexpr ck::index_t N = 4; static constexpr ck::index_t N = 4; // batch size
static constexpr ck::index_t K = 64; static constexpr ck::index_t K = 64; // output channel
static constexpr ck::index_t C = 32; static constexpr ck::index_t C = 192; // input channel
static constexpr ck::index_t Y = 3; static constexpr ck::index_t Y = 3; // filter H
static constexpr ck::index_t X = 3; static constexpr ck::index_t X = 3; // filter W
static constexpr ck::index_t Hi = 71; static constexpr ck::index_t Hi = 71; // input H
static constexpr ck::index_t Wi = 71; static constexpr ck::index_t Wi = 71; // input W
static constexpr ck::index_t Ho = 36; static constexpr ck::index_t Ho = 36; // output H
static constexpr ck::index_t Wo = 36; static constexpr ck::index_t Wo = 36; // output W
struct SimpleDeviceMem struct SimpleDeviceMem
{ {
SimpleDeviceMem() = delete; SimpleDeviceMem() = delete;
...@@ -64,8 +63,8 @@ int main(int argc, char* argv[]) ...@@ -64,8 +63,8 @@ int main(int argc, char* argv[])
std::array<ck::index_t, 5> bias_strides{K, 0, 1, 0, 0}; std::array<ck::index_t, 5> bias_strides{K, 0, 1, 0, 0};
std::array<ck::index_t, 5> requant_scale_lengths{G, N, K, Ho, Wo}; std::array<ck::index_t, 5> requant_scale_lengths{G, N, K, Ho, Wo};
std::array<ck::index_t, 5> requant_scale_strides{K, 0, 1, 0, 0}; std::array<ck::index_t, 5> requant_scale_strides{K, 0, 1, 0, 0};
std::array<ck::index_t, 5> out_lengths{G, N, C, Ho, Wo}; std::array<ck::index_t, 5> out_lengths{G, N, K, Ho, Wo};
std::array<ck::index_t, 5> out_strides{N * Ho * Wo * C, Ho * Wo * C, 1, Wo * C, C}; std::array<ck::index_t, 5> out_strides{N * Ho * Wo * K, Ho * Wo * K, 1, Wo * K, K};
std::array<ck::index_t, 2> in_left_pad{1, 1}; std::array<ck::index_t, 2> in_left_pad{1, 1};
std::array<ck::index_t, 2> in_right_pad{1, 1}; std::array<ck::index_t, 2> in_right_pad{1, 1};
std::array<ck::index_t, 2> conv_strides{2, 2}; std::array<ck::index_t, 2> conv_strides{2, 2};
...@@ -136,10 +135,11 @@ int main(int argc, char* argv[]) ...@@ -136,10 +135,11 @@ int main(int argc, char* argv[])
{ {
float avg_time = invoker_ptr->Run(argument_ptr.get(), StreamConfig{nullptr, true}); float avg_time = invoker_ptr->Run(argument_ptr.get(), StreamConfig{nullptr, true});
std::size_t flop = G * 2 * N * K * C * Ho * Wo * Y * X; std::size_t flop = G * 2 * N * K * C * Ho * Wo * Y * X;
std::size_t num_bytes = G * sizeof(InDataType) * N * Hi * Wi * C + std::size_t num_bytes =
G * sizeof(WeiDataType) * K * Y * X * C + G * sizeof(InDataType) * N * Hi * Wi * C + G * sizeof(WeiDataType) * K * Y * X * C +
G * sizeof(OutDataType) * N * Ho * Wo * K; G * sizeof(BiasDataType) * K + G * sizeof(RequantScaleDataType) * K +
G * sizeof(OutDataType) * N * Ho * Wo * K;
float tflops = static_cast<float>(flop) / 1.E9 / avg_time; float tflops = static_cast<float>(flop) / 1.E9 / avg_time;
float gb_per_sec = num_bytes / 1.E6 / avg_time; float gb_per_sec = num_bytes / 1.E6 / avg_time;
...@@ -162,11 +162,12 @@ int main(int argc, char* argv[]) ...@@ -162,11 +162,12 @@ int main(int argc, char* argv[])
} }
} }
std::cout << "Best Perf: " << std::setw(10) << best_avg_time << " ms, " << best_tflops
<< " TFlops, " << best_gb_per_sec << " GB/s, " << best_op_name << std::endl;
// run the best intance // run the best intance
if(best_op_id != -1)
{ {
std::cout << "Best Perf: " << std::setw(10) << best_avg_time << " ms, " << best_tflops
<< " TFlops, " << best_gb_per_sec << " GB/s, " << best_op_name << std::endl;
auto& op_ptr = op_ptrs[best_op_id]; auto& op_ptr = op_ptrs[best_op_id];
std::cout << "Run the best instance without timing: " << op_ptr->GetTypeString() std::cout << "Run the best instance without timing: " << op_ptr->GetTypeString()
<< std::endl; << std::endl;
......
client_example/09_quantization/conv2d_fwd_bias_relu_perlayer_quantization.cpp
View file @ 9fb64dae
...@@ -26,15 +26,15 @@ using OutElementOp = ck::tensor_operation::element_wise::Add_Activation_Mul_Clam ...@@ -26,15 +26,15 @@ using OutElementOp = ck::tensor_operation::element_wise::Add_Activation_Mul_Clam
static constexpr ck::index_t NumDimSpatial = 2; static constexpr ck::index_t NumDimSpatial = 2;
static constexpr ck::index_t G = 1; static constexpr ck::index_t G = 1;
static constexpr ck::index_t N = 4; static constexpr ck::index_t N = 4; // batch size
static constexpr ck::index_t K = 64; static constexpr ck::index_t K = 64; // output channel
static constexpr ck::index_t C = 32; static constexpr ck::index_t C = 192; // input channel
static constexpr ck::index_t Y = 3; static constexpr ck::index_t Y = 3; // filter H
static constexpr ck::index_t X = 3; static constexpr ck::index_t X = 3; // filter W
static constexpr ck::index_t Hi = 71; static constexpr ck::index_t Hi = 71; // input H
static constexpr ck::index_t Wi = 71; static constexpr ck::index_t Wi = 71; // input W
static constexpr ck::index_t Ho = 36; static constexpr ck::index_t Ho = 36; // output H
static constexpr ck::index_t Wo = 36; static constexpr ck::index_t Wo = 36; // output W
struct SimpleDeviceMem struct SimpleDeviceMem
{ {
...@@ -60,8 +60,8 @@ int main(int argc, char* argv[]) ...@@ -60,8 +60,8 @@ int main(int argc, char* argv[])
std::array<ck::index_t, 5> weight_strides{K * Y * X * C, Y * X * C, 1, X * C, C}; std::array<ck::index_t, 5> weight_strides{K * Y * X * C, Y * X * C, 1, X * C, C};
std::array<ck::index_t, 5> bias_lengths{G, N, K, Ho, Wo}; std::array<ck::index_t, 5> bias_lengths{G, N, K, Ho, Wo};
std::array<ck::index_t, 5> bias_strides{K, 0, 1, 0, 0}; std::array<ck::index_t, 5> bias_strides{K, 0, 1, 0, 0};
std::array<ck::index_t, 5> out_lengths{G, N, C, Ho, Wo}; std::array<ck::index_t, 5> out_lengths{G, N, K, Ho, Wo};
std::array<ck::index_t, 5> out_strides{N * Ho * Wo * C, Ho * Wo * C, 1, Wo * C, C}; std::array<ck::index_t, 5> out_strides{N * Ho * Wo * K, Ho * Wo * K, 1, Wo * K, K};
std::array<ck::index_t, 2> in_left_pad{1, 1}; std::array<ck::index_t, 2> in_left_pad{1, 1};
std::array<ck::index_t, 2> in_right_pad{1, 1}; std::array<ck::index_t, 2> in_right_pad{1, 1};
std::array<ck::index_t, 2> conv_strides{2, 2}; std::array<ck::index_t, 2> conv_strides{2, 2};
...@@ -130,10 +130,10 @@ int main(int argc, char* argv[]) ...@@ -130,10 +130,10 @@ int main(int argc, char* argv[])
{ {
float avg_time = invoker_ptr->Run(argument_ptr.get(), StreamConfig{nullptr, true}); float avg_time = invoker_ptr->Run(argument_ptr.get(), StreamConfig{nullptr, true});
std::size_t flop = G * 2 * N * K * C * Ho * Wo * Y * X; std::size_t flop = G * 2 * N * K * C * Ho * Wo * Y * X;
std::size_t num_bytes = G * sizeof(InDataType) * N * Hi * Wi * C + std::size_t num_bytes =
G * sizeof(WeiDataType) * K * Y * X * C + G * sizeof(InDataType) * N * Hi * Wi * C + G * sizeof(WeiDataType) * K * Y * X * C +
G * sizeof(OutDataType) * N * Ho * Wo * K; G * sizeof(BiasDataType) * K + G * sizeof(OutDataType) * N * Ho * Wo * K;
float tflops = static_cast<float>(flop) / 1.E9 / avg_time; float tflops = static_cast<float>(flop) / 1.E9 / avg_time;
float gb_per_sec = num_bytes / 1.E6 / avg_time; float gb_per_sec = num_bytes / 1.E6 / avg_time;
...@@ -156,11 +156,12 @@ int main(int argc, char* argv[]) ...@@ -156,11 +156,12 @@ int main(int argc, char* argv[])
} }
} }
std::cout << "Best Perf: " << std::setw(10) << best_avg_time << " ms, " << best_tflops
<< " TFlops, " << best_gb_per_sec << " GB/s, " << best_op_name << std::endl;
// run the best intance // run the best intance
if(best_op_id != -1)
{ {
std::cout << "Best Perf: " << std::setw(10) << best_avg_time << " ms, " << best_tflops
<< " TFlops, " << best_gb_per_sec << " GB/s, " << best_op_name << std::endl;
auto& op_ptr = op_ptrs[best_op_id]; auto& op_ptr = op_ptrs[best_op_id];
std::cout << "Run the best instance without timing: " << op_ptr->GetTypeString() std::cout << "Run the best instance without timing: " << op_ptr->GetTypeString()
<< std::endl; << std::endl;
......
client_example/09_quantization/conv2d_fwd_perchannel_quantization.cpp
View file @ 9fb64dae
...@@ -26,15 +26,15 @@ using OutElementOp = ck::tensor_operation::element_wise::Activation_Mul2_C ...@@ -26,15 +26,15 @@ using OutElementOp = ck::tensor_operation::element_wise::Activation_Mul2_C
static constexpr ck::index_t NumDimSpatial = 2; static constexpr ck::index_t NumDimSpatial = 2;
static constexpr ck::index_t G = 1; static constexpr ck::index_t G = 1;
static constexpr ck::index_t N = 4; static constexpr ck::index_t N = 4; // batch size
static constexpr ck::index_t K = 64; static constexpr ck::index_t K = 64; // output channel
static constexpr ck::index_t C = 32; static constexpr ck::index_t C = 192; // input channel
static constexpr ck::index_t Y = 3; static constexpr ck::index_t Y = 3; // filter H
static constexpr ck::index_t X = 3; static constexpr ck::index_t X = 3; // filter W
static constexpr ck::index_t Hi = 71; static constexpr ck::index_t Hi = 71; // input H
static constexpr ck::index_t Wi = 71; static constexpr ck::index_t Wi = 71; // input W
static constexpr ck::index_t Ho = 36; static constexpr ck::index_t Ho = 36; // output H
static constexpr ck::index_t Wo = 36; static constexpr ck::index_t Wo = 36; // output W
struct SimpleDeviceMem struct SimpleDeviceMem
{ {
...@@ -60,8 +60,8 @@ int main(int argc, char* argv[]) ...@@ -60,8 +60,8 @@ int main(int argc, char* argv[])
std::array<ck::index_t, 5> weight_strides{K * Y * X * C, Y * X * C, 1, X * C, C}; std::array<ck::index_t, 5> weight_strides{K * Y * X * C, Y * X * C, 1, X * C, C};
std::array<ck::index_t, 5> requant_scale_lengths{G, N, K, Ho, Wo}; std::array<ck::index_t, 5> requant_scale_lengths{G, N, K, Ho, Wo};
std::array<ck::index_t, 5> requant_scale_strides{K, 0, 1, 0, 0}; std::array<ck::index_t, 5> requant_scale_strides{K, 0, 1, 0, 0};
std::array<ck::index_t, 5> out_lengths{G, N, C, Ho, Wo}; std::array<ck::index_t, 5> out_lengths{G, N, K, Ho, Wo};
std::array<ck::index_t, 5> out_strides{N * Ho * Wo * C, Ho * Wo * C, 1, Wo * C, C}; std::array<ck::index_t, 5> out_strides{N * Ho * Wo * K, Ho * Wo * K, 1, Wo * K, K};
std::array<ck::index_t, 2> in_left_pad{1, 1}; std::array<ck::index_t, 2> in_left_pad{1, 1};
std::array<ck::index_t, 2> in_right_pad{1, 1}; std::array<ck::index_t, 2> in_right_pad{1, 1};
std::array<ck::index_t, 2> conv_strides{2, 2}; std::array<ck::index_t, 2> conv_strides{2, 2};
...@@ -130,10 +130,10 @@ int main(int argc, char* argv[]) ...@@ -130,10 +130,10 @@ int main(int argc, char* argv[])
{ {
float avg_time = invoker_ptr->Run(argument_ptr.get(), StreamConfig{nullptr, true}); float avg_time = invoker_ptr->Run(argument_ptr.get(), StreamConfig{nullptr, true});
std::size_t flop = G * 2 * N * K * C * Ho * Wo * Y * X; std::size_t flop = G * 2 * N * K * C * Ho * Wo * Y * X;
std::size_t num_bytes = G * sizeof(InDataType) * N * Hi * Wi * C + std::size_t num_bytes =
G * sizeof(WeiDataType) * K * Y * X * C + G * sizeof(InDataType) * N * Hi * Wi * C + G * sizeof(WeiDataType) * K * Y * X * C +
G * sizeof(OutDataType) * N * Ho * Wo * K; G * sizeof(RequantScaleDataType) * K + G * sizeof(OutDataType) * N * Ho * Wo * K;
float tflops = static_cast<float>(flop) / 1.E9 / avg_time; float tflops = static_cast<float>(flop) / 1.E9 / avg_time;
float gb_per_sec = num_bytes / 1.E6 / avg_time; float gb_per_sec = num_bytes / 1.E6 / avg_time;
...@@ -156,11 +156,12 @@ int main(int argc, char* argv[]) ...@@ -156,11 +156,12 @@ int main(int argc, char* argv[])
} }
} }
std::cout << "Best Perf: " << std::setw(10) << best_avg_time << " ms, " << best_tflops
<< " TFlops, " << best_gb_per_sec << " GB/s, " << best_op_name << std::endl;
// run the best intance // run the best intance
if(best_op_id != -1)
{ {
std::cout << "Best Perf: " << std::setw(10) << best_avg_time << " ms, " << best_tflops
<< " TFlops, " << best_gb_per_sec << " GB/s, " << best_op_name << std::endl;
auto& op_ptr = op_ptrs[best_op_id]; auto& op_ptr = op_ptrs[best_op_id];
std::cout << "Run the best instance without timing: " << op_ptr->GetTypeString() std::cout << "Run the best instance without timing: " << op_ptr->GetTypeString()
<< std::endl; << std::endl;
......
client_example/09_quantization/conv2d_fwd_perlayer_quantization.cpp
View file @ 9fb64dae
...@@ -24,15 +24,15 @@ using OutElementOp = ck::tensor_operation::element_wise::Activation_Mul_Clamp<Ac ...@@ -24,15 +24,15 @@ using OutElementOp = ck::tensor_operation::element_wise::Activation_Mul_Clamp<Ac
static constexpr ck::index_t NumDimSpatial = 2; static constexpr ck::index_t NumDimSpatial = 2;
static constexpr ck::index_t G = 1; static constexpr ck::index_t G = 1;
static constexpr ck::index_t N = 4; static constexpr ck::index_t N = 4; // batch size
static constexpr ck::index_t K = 64; static constexpr ck::index_t K = 64; // output channel
static constexpr ck::index_t C = 32; static constexpr ck::index_t C = 192; // input channel
static constexpr ck::index_t Y = 3; static constexpr ck::index_t Y = 3; // filter H
static constexpr ck::index_t X = 3; static constexpr ck::index_t X = 3; // filter W
static constexpr ck::index_t Hi = 71; static constexpr ck::index_t Hi = 71; // input H
static constexpr ck::index_t Wi = 71; static constexpr ck::index_t Wi = 71; // input W
static constexpr ck::index_t Ho = 36; static constexpr ck::index_t Ho = 36; // output H
static constexpr ck::index_t Wo = 36; static constexpr ck::index_t Wo = 36; // output W
struct SimpleDeviceMem struct SimpleDeviceMem
{ {
...@@ -56,8 +56,8 @@ int main(int argc, char* argv[]) ...@@ -56,8 +56,8 @@ int main(int argc, char* argv[])
std::array<ck::index_t, 5> in_strides{N * Hi * Wi * C, Hi * Wi * C, 1, Wi * C, C}; std::array<ck::index_t, 5> in_strides{N * Hi * Wi * C, Hi * Wi * C, 1, Wi * C, C};
std::array<ck::index_t, 5> weight_lengths{G, K, C, Y, X}; std::array<ck::index_t, 5> weight_lengths{G, K, C, Y, X};
std::array<ck::index_t, 5> weight_strides{K * Y * X * C, Y * X * C, 1, X * C, C}; std::array<ck::index_t, 5> weight_strides{K * Y * X * C, Y * X * C, 1, X * C, C};
std::array<ck::index_t, 5> out_lengths{G, N, C, Ho, Wo}; std::array<ck::index_t, 5> out_lengths{G, N, K, Ho, Wo};
std::array<ck::index_t, 5> out_strides{N * Ho * Wo * C, Ho * Wo * C, 1, Wo * C, C}; std::array<ck::index_t, 5> out_strides{N * Ho * Wo * K, Ho * Wo * K, 1, Wo * K, K};
std::array<ck::index_t, 2> in_left_pad{1, 1}; std::array<ck::index_t, 2> in_left_pad{1, 1};
std::array<ck::index_t, 2> in_right_pad{1, 1}; std::array<ck::index_t, 2> in_right_pad{1, 1};
std::array<ck::index_t, 2> conv_strides{2, 2}; std::array<ck::index_t, 2> conv_strides{2, 2};
...@@ -150,11 +150,11 @@ int main(int argc, char* argv[]) ...@@ -150,11 +150,11 @@ int main(int argc, char* argv[])
} }
} }
std::cout << "Best Perf: " << std::setw(10) << best_avg_time << " ms, " << best_tflops if(best_op_id != -1)
<< " TFlops, " << best_gb_per_sec << " GB/s, " << best_op_name << std::endl;
// run the best intance
{ {
std::cout << "Best Perf: " << std::setw(10) << best_avg_time << " ms, " << best_tflops
<< " TFlops, " << best_gb_per_sec << " GB/s, " << best_op_name << std::endl;
auto& op_ptr = op_ptrs[best_op_id]; auto& op_ptr = op_ptrs[best_op_id];
std::cout << "Run the best instance without timing: " << op_ptr->GetTypeString() std::cout << "Run the best instance without timing: " << op_ptr->GetTypeString()
<< std::endl; << std::endl;
......
client_example/09_quantization/gemm_quantization.cpp 0 → 100644
View file @ 9fb64dae
// SPDX-License-Identifier: MIT
// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
#include <iomanip>
#include <iostream>
#include <vector>
#include "ck/ck.hpp"
#include "ck/tensor_operation/gpu/device/tensor_layout.hpp"
#include "ck/tensor_operation/gpu/device/device_gemm_multiple_d.hpp"
#include "ck/tensor_operation/gpu/element/element_wise_operation.hpp"
#include "ck/library/tensor_operation_instance/gpu/quantization/gemm_quantization.hpp"
using Row = ck::tensor_layout::gemm::RowMajor;
using Col = ck::tensor_layout::gemm::ColumnMajor;
using PassThrough = ck::tensor_operation::element_wise::PassThrough;
using AElementOp = PassThrough;
using BElementOp = PassThrough;
using ActivationOp = PassThrough;
using CDEElementOp = ck::tensor_operation::element_wise::Activation_Mul_Clamp<ActivationOp>;
using ADataType = int8_t;
using BDataType = int8_t;
using EDataType = int8_t;
using ALayout = Row;
using BLayout = Col;
using ELayout = Row;
struct SimpleDeviceMem
{
SimpleDeviceMem() = delete;
SimpleDeviceMem(std::size_t mem_size) : p_mem_{}
{
(void)hipMalloc(static_cast<void**>(&p_mem_), mem_size);
}
void* GetDeviceBuffer() { return p_mem_; }
~SimpleDeviceMem() { (void)hipFree(p_mem_); }
void* p_mem_;
};
int main(int argc, char* argv[])
{
ck::index_t M = 1024;
ck::index_t N = 1024;
ck::index_t K = 1024;
ck::index_t StrideA = 1024;
ck::index_t StrideB = 1024;
ck::index_t StrideE = 1024;
float requant_scale = 0.03;
auto f_matrix_space_size =
[](std::size_t nRow, std::size_t nCol, std::size_t stride, auto layout) {
using Layout = decltype(layout);
if constexpr(std::is_same<Layout, ck::tensor_layout::gemm::RowMajor>::value)
{
return (nRow - 1) * stride + nCol;
}
else
{
return (nCol - 1) * stride + nRow;
}
};
SimpleDeviceMem a_device_buf(sizeof(ADataType) * f_matrix_space_size(M, K, StrideA, ALayout{}));
SimpleDeviceMem b_device_buf(sizeof(BDataType) * f_matrix_space_size(K, N, StrideB, BLayout{}));
SimpleDeviceMem e_device_buf(sizeof(EDataType) * f_matrix_space_size(M, N, StrideE, ELayout{}));
using DeviceOp = ck::tensor_operation::device::DeviceGemmMultipleD<ALayout,
BLayout,
ck::Tuple<>,
ELayout,
ADataType,
BDataType,
ck::Tuple<>,
EDataType,
AElementOp,
BElementOp,
CDEElementOp>;
// get device op instances
const auto op_ptrs = ck::tensor_operation::device::instance::DeviceOperationInstanceFactory<
DeviceOp>::GetInstances();
std::cout << "found " << op_ptrs.size() << " instances" << std::endl;
const auto a_element_op = AElementOp{};
const auto b_element_op = BElementOp{};
const auto cde_element_op = CDEElementOp{requant_scale, ActivationOp{}};
std::string best_op_name;
int best_op_id = -1;
float best_avg_time = std::numeric_limits<float>::max();
float best_gb_per_sec = 0;
float best_tflops = 0;
// profile device operation instances
std::cout << "Run all instances and do timing" << std::endl;
for(int i = 0; i < op_ptrs.size(); ++i)
{
auto& op_ptr = op_ptrs[i];
auto argument_ptr = op_ptr->MakeArgumentPointer(a_device_buf.GetDeviceBuffer(),
b_device_buf.GetDeviceBuffer(),
{},
e_device_buf.GetDeviceBuffer(),
M,
N,
K,
StrideA,
StrideB,
{},
StrideE,
a_element_op,
b_element_op,
cde_element_op);
auto invoker_ptr = op_ptr->MakeInvokerPointer();
std::string op_name = op_ptr->GetTypeString();
if(op_ptr->IsSupportedArgument(argument_ptr.get()))
{
float avg_time = invoker_ptr->Run(argument_ptr.get(), StreamConfig{nullptr, true});
std::size_t flop = std::size_t(2) * M * N * K;
std::size_t num_bytes =
sizeof(ADataType) * M * K + sizeof(BDataType) * K * N + sizeof(EDataType) * M * N;
float tflops = static_cast<float>(flop) / 1.E9 / avg_time;
float gb_per_sec = num_bytes / 1.E6 / avg_time;
std::cout << "Perf: " << std::setw(10) << avg_time << " ms, " << tflops << " TFlops, "
<< gb_per_sec << " GB/s, " << op_name << std::endl;
if(tflops > best_tflops)
{
best_op_id = i;
best_op_name = op_name;
best_avg_time = avg_time;
best_gb_per_sec = gb_per_sec;
best_tflops = tflops;
}
}
else
{
std::cout << op_name << " does not support this problem" << std::endl;
}
}
if(best_op_id != -1)
{
std::cout << "Best Perf: " << std::setw(10) << best_avg_time << " ms, " << best_tflops
<< " TFlops, " << best_gb_per_sec << " GB/s, " << best_op_name << std::endl;
auto& op_ptr = op_ptrs[best_op_id];
std::cout << "Run the best instance without timing: " << op_ptr->GetTypeString()
<< std::endl;
auto argument_ptr = op_ptr->MakeArgumentPointer(a_device_buf.GetDeviceBuffer(),
b_device_buf.GetDeviceBuffer(),
{},
e_device_buf.GetDeviceBuffer(),
M,
N,
K,
StrideA,
StrideB,
{},
StrideE,
a_element_op,
b_element_op,
cde_element_op);
auto invoker_ptr = op_ptr->MakeInvokerPointer();
if(op_ptr->IsSupportedArgument(argument_ptr.get()))
{
invoker_ptr->Run(argument_ptr.get(), StreamConfig{nullptr, false});
}
std::cout << "Done" << std::endl;
}
return 0;
}
\ No newline at end of file
client_example/17_grouped_gemm_fastgelu/CMakeLists.txt 0 → 100644
View file @ 9fb64dae
add_executable(client_grouped_gemm_fastgelu grouped_gemm_fastgelu.cpp)
target_link_libraries(client_grouped_gemm_fastgelu PRIVATE composable_kernel::device_operations)
\ No newline at end of file
client_example/17_grouped_gemm_fastgelu/grouped_gemm_fastgelu.cpp 0 → 100644
View file @ 9fb64dae
// SPDX-License-Identifier: MIT
// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
#include <iomanip>
#include <iostream>
#include <vector>
#include <random>
#include "ck/ck.hpp"
#include "ck/tensor_operation/gpu/device/tensor_layout.hpp"
#include "ck/tensor_operation/gpu/device/device_grouped_gemm.hpp"
#include "ck/tensor_operation/gpu/element/element_wise_operation.hpp"
#include "ck/library/tensor_operation_instance/gpu/grouped_gemm_fastgelu.hpp"
using F16 = ck::half_t;
using F32 = float;
using Row = ck::tensor_layout::gemm::RowMajor;
using Col = ck::tensor_layout::gemm::ColumnMajor;
using PassThrough = ck::tensor_operation::element_wise::PassThrough;
using FastGelu = ck::tensor_operation::element_wise::FastGelu;
using ADataType = F16;
using BDataType = F16;
using DsDataType = ck::Tuple<>;
using EDataType = F16;
using ALayout = Row;
using BLayout = Col;
using DsLayout = ck::Tuple<>;
using ELayout = Row;
using AElementOp = PassThrough;
using BElementOp = PassThrough;
using CDEElementOp = FastGelu;
struct SimpleDeviceMem
{
SimpleDeviceMem() = delete;
SimpleDeviceMem(std::size_t mem_size) : p_mem_{}
{
(void)hipMalloc(static_cast<void**>(&p_mem_), mem_size);
}
void* GetDeviceBuffer() { return p_mem_; }
~SimpleDeviceMem() { (void)hipFree(p_mem_); }
void* p_mem_;
};
int main()
{
std::mt19937 gen(19391);
std::uniform_int_distribution<> distrib(1, 10);
int group_count = distrib(gen);
std::vector<int> Ms, Ns, Ks, StrideAs, StrideBs, StrideEs;
for(int i = 0; i < group_count; ++i)
{
Ms.push_back(256 + 256 * distrib(gen));
Ns.push_back(256 + 256 * distrib(gen));
Ks.push_back(128 + 128 * distrib(gen));
StrideAs.push_back(std::is_same<Row, ALayout>::value ? Ks[i] : Ms[i]);
StrideBs.push_back(std::is_same<Row, BLayout>::value ? Ns[i] : Ks[i]);
StrideEs.push_back(std::is_same<Row, ELayout>::value ? Ns[i] : Ms[i]);
}
auto f_matrix_space_size =
[](std::size_t nRow, std::size_t nCol, std::size_t stride, auto layout) {
using Layout = decltype(layout);
if constexpr(std::is_same<Layout, ck::tensor_layout::gemm::RowMajor>::value)
{
return (nRow - 1) * stride + nCol;
}
else
{
return (nCol - 1) * stride + nRow;
}
};
std::vector<SimpleDeviceMem> a_dev_bufs, b_dev_bufs, e_dev_bufs;
a_dev_bufs.reserve(group_count);
b_dev_bufs.reserve(group_count);
e_dev_bufs.reserve(group_count);
std::vector<const void*> p_a, p_b;
std::vector<void*> p_e;
p_a.reserve(group_count);
p_b.reserve(group_count);
p_e.reserve(group_count);
std::vector<ck::tensor_operation::device::GemmDesc> gemm_descs;
gemm_descs.reserve(group_count);
for(int i = 0; i < group_count; ++i)
{
a_dev_bufs.emplace_back(sizeof(ADataType) *
f_matrix_space_size(Ms[i], Ks[i], StrideAs[i], ALayout{}));
b_dev_bufs.emplace_back(sizeof(BDataType) *
f_matrix_space_size(Ks[i], Ns[i], StrideBs[i], BLayout{}));
e_dev_bufs.emplace_back(sizeof(EDataType) *
f_matrix_space_size(Ms[i], Ns[i], StrideEs[i], ELayout{}));
gemm_descs.push_back({Ms[i], Ns[i], Ks[i], StrideAs[i], StrideBs[i], StrideEs[i], {}});
p_a.push_back(a_dev_bufs[i].GetDeviceBuffer());
p_b.push_back(b_dev_bufs[i].GetDeviceBuffer());
p_e.push_back(e_dev_bufs[i].GetDeviceBuffer());
}
using DeviceOp = ck::tensor_operation::device::DeviceGroupedGemm<ALayout,
BLayout,
DsLayout,
ELayout,
ADataType,
BDataType,
DsDataType,
EDataType,
AElementOp,
BElementOp,
CDEElementOp>;
// get device op instances
const auto op_ptrs = ck::tensor_operation::device::instance::DeviceOperationInstanceFactory<
DeviceOp>::GetInstances();
std::cout << "found " << op_ptrs.size() << " instances" << std::endl;
const auto a_element_op = AElementOp{};
const auto b_element_op = BElementOp{};
const auto cde_element_op = CDEElementOp{};
std::string best_op_name;
bool found = false;
int best_op_id = -1;
float best_ave_time = 0;
float best_tflops = 0;
float best_gb_per_sec = 0;
auto p_ds = std::vector<std::array<const void*, 0>>{};
// profile device operation instances
std::cout << "Run all instances and do timing" << std::endl;
for(int i = 0; i < op_ptrs.size(); ++i)
{
auto& op_ptr = op_ptrs[i];
auto argument_ptr = op_ptr->MakeArgumentPointer(
p_a, p_b, p_ds, p_e, gemm_descs, a_element_op, b_element_op, cde_element_op);
auto invoker_ptr = op_ptr->MakeInvokerPointer();
SimpleDeviceMem gemm_desc_workspace(op_ptr->GetWorkSpaceSize(argument_ptr.get()));
op_ptr->SetWorkSpacePointer(argument_ptr.get(), gemm_desc_workspace.GetDeviceBuffer());
std::string op_name = op_ptr->GetTypeString();
if(op_ptr->IsSupportedArgument(argument_ptr.get()))
{
float ave_time = invoker_ptr->Run(argument_ptr.get(), StreamConfig{nullptr, true});
std::size_t flop = 0, num_btype = 0;
for(std::size_t j = 0; j < gemm_descs.size(); ++j)
{
flop += std::size_t(2) * Ms[j] * Ns[j] * Ks[j];
num_btype += sizeof(ADataType) * Ms[j] * Ks[j] + sizeof(BDataType) * Ks[j] * Ns[j] +
sizeof(EDataType) * Ms[j] * Ns[j];
}
float tflops = static_cast<float>(flop) / 1.E9 / ave_time;
float gb_per_sec = num_btype / 1.E6 / ave_time;
std::cout << "Perf: " << std::setw(10) << ave_time << " ms, " << tflops << " TFlops, "
<< gb_per_sec << " GB/s, " << op_name << std::endl;
if(tflops > best_tflops)
{
found = true;
best_op_id = i;
best_op_name = op_name;
best_tflops = tflops;
best_ave_time = ave_time;
best_gb_per_sec = gb_per_sec;
}
}
else
{
std::cout << op_name << " does not support this problem" << std::endl;
}
}
std::cout << "Best Perf: " << best_ave_time << " ms, " << best_tflops << " TFlops, "
<< best_gb_per_sec << " GB/s, " << best_op_name << std::endl;
// run the best intance
if(found)
{
auto& op_ptr = op_ptrs[best_op_id];
std::cout << "Run the best instance without timing: " << op_ptr->GetTypeString()
<< std::endl;
auto argument_ptr = op_ptr->MakeArgumentPointer(
p_a, p_b, p_ds, p_e, gemm_descs, a_element_op, b_element_op, cde_element_op);
auto invoker_ptr = op_ptr->MakeInvokerPointer();
SimpleDeviceMem gemm_desc_workspace(op_ptr->GetWorkSpaceSize(argument_ptr.get()));
op_ptr->SetWorkSpacePointer(argument_ptr.get(), gemm_desc_workspace.GetDeviceBuffer());
if(op_ptr->IsSupportedArgument(argument_ptr.get()))
{
invoker_ptr->Run(argument_ptr.get(), StreamConfig{nullptr, false});
}
std::cout << "Done" << std::endl;
}
return 0;
}
example/01_gemm/CMakeLists.txt
View file @ 9fb64dae
...@@ -38,7 +38,7 @@ add_example_executable_no_testing(example_gemm_xdl_fp64 gemm_xdl_fp64.cpp) ...@@ -38,7 +38,7 @@ add_example_executable_no_testing(example_gemm_xdl_fp64 gemm_xdl_fp64.cpp)
add_dependencies(example_gemm_xdl example_gemm_xdl_skip_b_lds_fp16) add_dependencies(example_gemm_xdl example_gemm_xdl_skip_b_lds_fp16)
add_dependencies(example_gemm_xdl example_gemm_xdl_fp64) add_dependencies(example_gemm_xdl example_gemm_xdl_fp64)
if(GPU_TARGETS MATCHES "gfx1100") if(GPU_TARGETS MATCHES "gfx1100" OR GPU_TARGETS MATCHES "gfx1101" OR GPU_TARGETS MATCHES "gfx1102")
add_custom_target(example_gemm_wmma) add_custom_target(example_gemm_wmma)
add_example_executable(example_gemm_wmma_fp16 gemm_wmma_fp16.cpp) add_example_executable(example_gemm_wmma_fp16 gemm_wmma_fp16.cpp)
add_dependencies(example_gemm_wmma example_gemm_wmma_fp16) add_dependencies(example_gemm_wmma example_gemm_wmma_fp16)
......
example/02_gemm_bilinear/CMakeLists.txt
View file @ 9fb64dae
add_example_executable(example_gemm_bilinear_xdl_fp16 gemm_bilinear_xdl_fp16.cpp) add_example_executable(example_gemm_bilinear_xdl_fp16 gemm_bilinear_xdl_fp16.cpp)
if(GPU_TARGETS MATCHES "gfx1100") if(GPU_TARGETS MATCHES "gfx1100" OR GPU_TARGETS MATCHES "gfx1101" OR GPU_TARGETS MATCHES "gfx1102")
add_example_executable(example_gemm_bilinear_wmma_fp16 gemm_bilinear_wmma_fp16.cpp) add_example_executable(example_gemm_bilinear_wmma_fp16 gemm_bilinear_wmma_fp16.cpp)
endif() endif()
example/14_gemm_quantization/CMakeLists.txt
View file @ 9fb64dae
# dlops
add_example_executable(example_gemm_dl_quantization_int8 gemm_dl_quantization_int8.cpp)
# xdlops
add_example_executable(example_gemm_xdl_bias_relu_quantization_int8 gemm_xdl_bias_relu_quantization_int8.cpp) add_example_executable(example_gemm_xdl_bias_relu_quantization_int8 gemm_xdl_bias_relu_quantization_int8.cpp)
add_example_executable(example_gemm_xdl_quantization_int8 gemm_xdl_quantization_int8.cpp) add_example_executable(example_gemm_xdl_quantization_int8 gemm_xdl_quantization_int8.cpp)
\ No newline at end of file
example/14_gemm_quantization/gemm_dl_quantization_int8.cpp 0 → 100644
View file @ 9fb64dae
// SPDX-License-Identifier: MIT
// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
#include <iostream>
#include <numeric>
#include <initializer_list>
#include <cstdlib>
#include "ck/ck.hpp"
#include "ck/tensor_operation/gpu/device/tensor_layout.hpp"
#include "ck/tensor_operation/gpu/device/gemm_specialization.hpp"
#include "ck/tensor_operation/gpu/device/impl/device_gemm_multiple_d_dl.hpp"
#include "ck/tensor_operation/gpu/element/element_wise_operation.hpp"
#include "ck/library/utility/device_memory.hpp"
#include "ck/library/utility/host_tensor.hpp"
#include "ck/library/utility/host_tensor_generator.hpp"
#include "ck/library/utility/literals.hpp"
#include "ck/library/reference_tensor_operation/cpu/reference_gemm.hpp"
#include "ck/library/utility/check_err.hpp"
template <ck::index_t... Is>
using S = ck::Sequence<Is...>;
using I8 = int8_t;
using I32 = int32_t;
using Row = ck::tensor_layout::gemm::RowMajor;
using Col = ck::tensor_layout::gemm::ColumnMajor;
using PassThrough = ck::tensor_operation::element_wise::PassThrough;
using AElementOp = PassThrough;
using BElementOp = PassThrough;
using ActivationOp = PassThrough;
using CDEElementOp = ck::tensor_operation::element_wise::Activation_Mul_Clamp<ActivationOp>;
using ADataType = I8;
using BDataType = I8;
using AccDataType = I32;
using CShuffleDataType = I32;
using DsDataType = ck::Tuple<>;
using EDataType = I8;
using ALayout = Row;
using BLayout = Col;
using DsLayout = ck::Tuple<>;
using ELayout = Row;
static constexpr auto GemmDefault = ck::tensor_operation::device::GemmSpecialization::Default;
using DeviceGemmInstance = ck::tensor_operation::device::DeviceGemmMultipleD_Dl<
ALayout,
BLayout,
DsLayout,
ELayout,
ADataType,
BDataType,
AccDataType,
DsDataType,
EDataType,
AElementOp,
BElementOp,
CDEElementOp,
GemmDefault,
256, // BlockSize
128, // MPerBlock
128, // NPerBlock
16, // K0PerBlock
4, // K1
4, // M1PerThread
4, // N1PerThread
1, // KPerThread
S<8, 2>, // M1N1ThreadClusterM1Xs
S<8, 2>, // M1N1ThreadClusterN1Xs
S<8, 1, 1, 4>, // ABlockTransferThreadSliceLengths_K0_M0_M1_K1
S<2, 1, 128, 1>, // ABlockTransferThreadClusterLengths_K0_M0_M1_K1
S<1, 2, 0, 3>, // ABlockTransferThreadClusterArrangeOrder
S<1, 2, 0, 3>, // ABlockTransferSrcAccessOrder
S<4, 1, 1, 4>, // ABlockTransferSrcVectorTensorLengths_K0_M0_M1_K1
S<1, 2, 0, 3>, // ABlockTransferSrcVectorTensorContiguousDimOrder
S<1, 1, 1, 4>, // ABlockTransferDstVectorTensorLengths_K0_M0_M1_K1
S<8, 1, 1, 4>, // BBlockTransferThreadSliceLengths_K0_N0_N1_K1
S<2, 1, 128, 1>, // BBlockTransferThreadClusterLengths_K0_N0_N1_K1
S<1, 2, 0, 3>, // BBlockTransferThreadClusterArrangeOrder
S<1, 2, 0, 3>, // BBlockTransferSrcAccessOrder
S<4, 1, 1, 4>, // BBlockTransferSrcVectorTensorLengths_K0_N0_N1_K1
S<1, 2, 0, 3>, // BBlockTransferSrcVectorTensorContiguousDimOrder
S<1, 1, 1, 4>, // BBlockTransferDstVectorTensorLengths_K0_N0_N1_K1
S<0, 1, 2, 3, 4, 5>, // CThreadTransferSrcDstAccessOrder
5, // CThreadTransferSrcDstVectorDim
4>; // CThreadTransferDstScalarPerVector
using ReferenceGemmInstance = ck::tensor_operation::host::
ReferenceGemm<ADataType, BDataType, EDataType, float, PassThrough, PassThrough, CDEElementOp>;
int main()
{
bool do_verification = true;
bool time_kernel = false;
// GEMM shape
ck::index_t M = 1024;
ck::index_t N = 1024;
ck::index_t K = 1024;
ck::index_t StrideA = 1024;
ck::index_t StrideB = 1024;
ck::index_t StrideE = 1024;
float requant_scale = 0.03;
auto f_host_tensor_descriptor =
[](std::size_t row, std::size_t col, std::size_t stride, auto layout) {
using namespace ck::literals;
if(std::is_same<decltype(layout), ck::tensor_layout::gemm::RowMajor>::value)
{
return HostTensorDescriptor(std::vector<std::size_t>({row, col}),
std::vector<std::size_t>({stride, 1_uz}));
}
else
{
return HostTensorDescriptor(std::vector<std::size_t>({row, col}),
std::vector<std::size_t>({1_uz, stride}));
}
};
Tensor<ADataType> a_m_k(f_host_tensor_descriptor(M, K, StrideA, ALayout{}));
Tensor<BDataType> b_k_n(f_host_tensor_descriptor(K, N, StrideB, BLayout{}));
Tensor<EDataType> e_m_n_host_result(f_host_tensor_descriptor(M, N, StrideE, ELayout{}));
Tensor<EDataType> e_m_n_device_result(f_host_tensor_descriptor(M, N, StrideE, ELayout{}));
std::cout << "a_m_k: " << a_m_k.mDesc << std::endl;
std::cout << "b_k_n: " << b_k_n.mDesc << std::endl;
std::cout << "e_m_n: " << e_m_n_host_result.mDesc << std::endl;
a_m_k.GenerateTensorValue(GeneratorTensor_2<ADataType>{-5, 5});
b_k_n.GenerateTensorValue(GeneratorTensor_2<BDataType>{-5, 5});
DeviceMem a_device_buf(sizeof(ADataType) * a_m_k.mDesc.GetElementSpaceSize());
DeviceMem b_device_buf(sizeof(BDataType) * b_k_n.mDesc.GetElementSpaceSize());
DeviceMem e_device_buf(sizeof(EDataType) * e_m_n_device_result.mDesc.GetElementSpaceSize());
a_device_buf.ToDevice(a_m_k.mData.data());
b_device_buf.ToDevice(b_k_n.mData.data());
auto a_element_op = AElementOp{};
auto b_element_op = BElementOp{};
auto cde_element_op = CDEElementOp{requant_scale, ActivationOp{}};
// do GEMM
auto gemm = DeviceGemmInstance{};
auto invoker = gemm.MakeInvoker();
auto argument = gemm.MakeArgument(a_device_buf.GetDeviceBuffer(),
b_device_buf.GetDeviceBuffer(),
{},
e_device_buf.GetDeviceBuffer(),
M,
N,
K,
StrideA,
StrideB,
{},
StrideE,
a_element_op,
b_element_op,
cde_element_op);
if(!gemm.IsSupportedArgument(argument))
{
throw std::runtime_error(
"wrong! device_gemm with the specified compilation parameters does "
"not support this GEMM problem");
}
float ave_time = invoker.Run(argument, StreamConfig{nullptr, time_kernel});
std::size_t flop = std::size_t(2) * M * N * K;
std::size_t num_btype =
sizeof(ADataType) * M * K + sizeof(BDataType) * K * N + sizeof(EDataType) * M * N;
float tflops = static_cast<float>(flop) / 1.E9 / ave_time;
float gb_per_sec = num_btype / 1.E6 / ave_time;
std::cout << "Perf: " << ave_time << " ms, " << tflops << " TFlops, " << gb_per_sec << " GB/s, "
<< gemm.GetTypeString() << std::endl;
e_device_buf.FromDevice(e_m_n_device_result.mData.data());
if(do_verification)
{
auto ref_gemm = ReferenceGemmInstance{};
auto ref_invoker = ref_gemm.MakeInvoker();
auto ref_argument = ref_gemm.MakeArgument(
a_m_k, b_k_n, e_m_n_host_result, a_element_op, b_element_op, cde_element_op);
ref_invoker.Run(ref_argument);
return ck::utils::check_err(e_m_n_device_result, e_m_n_host_result) ? 0 : 1;
}
return 0;
}
example/15_grouped_gemm/CMakeLists.txt
View file @ 9fb64dae
...@@ -4,12 +4,15 @@ add_example_executable(example_grouped_gemm_xdl_fp32 grouped_gemm_xdl_fp32.cpp) ...@@ -4,12 +4,15 @@ add_example_executable(example_grouped_gemm_xdl_fp32 grouped_gemm_xdl_fp32.cpp)
add_example_executable(example_grouped_gemm_xdl_fp16 grouped_gemm_xdl_fp16.cpp) add_example_executable(example_grouped_gemm_xdl_fp16 grouped_gemm_xdl_fp16.cpp)
add_example_executable(example_grouped_gemm_xdl_bfp16 grouped_gemm_xdl_bfp16.cpp) add_example_executable(example_grouped_gemm_xdl_bfp16 grouped_gemm_xdl_bfp16.cpp)
add_example_executable(example_grouped_gemm_xdl_int8 grouped_gemm_xdl_int8.cpp) add_example_executable(example_grouped_gemm_xdl_int8 grouped_gemm_xdl_int8.cpp)
add_example_executable(example_grouped_gemm_multiple_d_dl_fp16 grouped_gemm_multiple_d_dl_fp16.cpp)
add_dependencies(example_grouped_gemm_xdl add_dependencies(example_grouped_gemm_xdl
example_grouped_gemm_xdl_fp32 example_grouped_gemm_xdl_fp32
example_grouped_gemm_xdl_fp16 example_grouped_gemm_xdl_fp16
example_grouped_gemm_xdl_bfp16 example_grouped_gemm_xdl_bfp16
example_grouped_gemm_xdl_int8) example_grouped_gemm_xdl_int8
example_grouped_gemm_multiple_d_dl_fp16)
if(USE_BITINT_EXTENSION_INT4) if(USE_BITINT_EXTENSION_INT4)
add_example_executable(example_grouped_gemm_xdl_int4 grouped_gemm_xdl_int4.cpp) add_example_executable(example_grouped_gemm_xdl_int4 grouped_gemm_xdl_int4.cpp)
......
example/15_grouped_gemm/grouped_gemm_multiple_d_dl_fp16.cpp 0 → 100644
View file @ 9fb64dae
// SPDX-License-Identifier: MIT
// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
#include <algorithm>
#include <cstddef>
#include <initializer_list>
#include <iostream>
#include <numeric>
#include <stdexcept>
#include <string>
#include "ck/ck.hpp"
#include "ck/tensor_operation/gpu/device/tensor_layout.hpp"
#include "ck/tensor_operation/gpu/device/gemm_specialization.hpp"
#include "ck/tensor_operation/gpu/device/impl/device_grouped_gemm_multiple_d_dl.hpp"
#include "ck/tensor_operation/gpu/element/element_wise_operation.hpp"
#include "ck/library/utility/check_err.hpp"
#include "ck/library/utility/device_memory.hpp"
#include "ck/library/utility/host_tensor.hpp"
#include "ck/library/utility/host_tensor_generator.hpp"
#include "ck/library/utility/literals.hpp"
#include "ck/library/reference_tensor_operation/cpu/reference_gemm.hpp"
#include "ck/utility/data_type.hpp"
#include "ck/utility/tuple.hpp"
#include "ck/utility/sequence.hpp"
template <ck::index_t... Is>
using S = ck::Sequence<Is...>;
using F16 = ck::half_t;
using F32 = float;
using Row = ck::tensor_layout::gemm::RowMajor;
using Col = ck::tensor_layout::gemm::ColumnMajor;
using PassThrough = ck::tensor_operation::element_wise::PassThrough;
using ADataType = F16;
using BDataType = F16;
using AccDataType = F32;
using DsDataType = ck::Tuple<>;
using EDataType = F16;
using ALayout = Row;
using BLayout = Row;
using DsLayout = ck::Tuple<>;
using ELayout = Row;
using AElementOp = PassThrough;
using BElementOp = PassThrough;
using CDEElementOp = PassThrough;
static constexpr auto GemmSpec = ck::tensor_operation::device::GemmSpecialization::MNPadding;
// clang-format off
using DeviceGemmInstance = ck::tensor_operation::device::
// ##################| ALayout| BLayout| DsLayout| ELayout| AData| BData| AccData| DsData| EData| A| B| CDE| GEMM| Block| MPer| NPer| K0Per| K1| M1Per| N1Per| KPer| M11N11Thread| M11N11Thread| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| BBlockTransfer| BBlockTransfer| BBlockTransfer| BBlockTransfer| BBlockTransfer| BBlockTransfer| BBlockTransfer| CThreadTransfer| CThreadTransfer| CThreadTransfer|
// ##################| | | | | Type| Type| Type| Type| Type| Elementwise| Elementwise| Elementwise| Specialization| Size| Block| Block| Block| | ThreadM111| ThreadN111| Thread| ClusterM110Xs| ClusterN110Xs| ThreadSliceLengths| ThreadClusterLengths| ThreadCluster| SrcAccess| SrcVectorTensor| SrcVectorTensor| DstVectorTensor| ThreadSliceLengths| ThreadClusterLengths| ThreadCluster| SrcAccess| SrcVectorTensor| SrcVectorTensor| DstVectorTensor| SrcDstAccess| SrcDstVectorDim| DstScalarPerVector|
// ##################| | | | | | | | | | Operation| Operation| Operation| | | | | | | | | | | | K0_M0_M1_K1| K0_M0_M1_K1| ArrangeOrder| Order| Lengths_K0_M0_M1_K1| ContiguousDimOrder| Lengths_K0_M0_M1_K1| K0_N0_N1_K1| K0_N0_N1_K1| ArrangeOrder| Order| Lengths_K0_N0_N1_K1| ContiguousDimOrder| Lengths_K0_N0_N1_K1| Order| | |
// ##################| | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | |
DeviceGroupedGemmMultipleD_Dl< ALayout, BLayout, DsLayout, ELayout, ADataType, BDataType, AccDataType, DsDataType, EDataType, AElementOp, BElementOp, CDEElementOp, GemmSpec, 256, 128, 128, 16, 2, 4, 4, 1, S<8, 2>, S<8, 2>, S<8, 1, 1, 2>, S<2, 1, 128, 1>, S<1, 2, 0, 3>, S<1, 2, 0, 3>, S<4, 1, 1, 2>, S<1, 2, 0, 3>, S<1, 1, 1, 2>, S<2, 1, 4, 2>, S<8, 1, 32, 1>, S<0, 3, 1, 2>, S<0, 3, 1, 2>, S<1, 1, 4, 1>, S<0, 3, 1, 2>, S<1, 1, 4, 2>, S<0, 1, 2, 3, 4, 5>, 5, 4>;
// clang-format on
#include "run_grouped_gemm_example.inc"
int main(int argc, char* argv[]) { return !run_grouped_gemm_example(argc, argv); }
example/29_batched_gemm_bias_e_permute/CMakeLists.txt
View file @ 9fb64dae
add_example_executable(example_batched_gemm_bias_e_permute_xdl_fp16 batched_gemm_bias_e_permute_xdl_fp16.cpp) add_example_executable(example_batched_gemm_bias_e_permute_xdl_fp16 batched_gemm_bias_e_permute_xdl_fp16.cpp)
if(GPU_TARGETS MATCHES "gfx1100") if(GPU_TARGETS MATCHES "gfx1100" OR GPU_TARGETS MATCHES "gfx1101" OR GPU_TARGETS MATCHES "gfx1102")
add_example_executable(example_batched_gemm_bias_e_permute_wmma_fp16 batched_gemm_bias_e_permute_wmma_fp16.cpp) add_example_executable(example_batched_gemm_bias_e_permute_wmma_fp16 batched_gemm_bias_e_permute_wmma_fp16.cpp)
endif() endif()
example/30_grouped_conv_fwd_multiple_d/CMakeLists.txt
View file @ 9fb64dae
...@@ -16,7 +16,7 @@ if(USE_BITINT_EXTENSION_INT4) ...@@ -16,7 +16,7 @@ if(USE_BITINT_EXTENSION_INT4)
add_dependencies(example_grouped_conv_fwd_multiple_d example_grouped_conv_fwd_bias_relu_add_xdl_int4) add_dependencies(example_grouped_conv_fwd_multiple_d example_grouped_conv_fwd_bias_relu_add_xdl_int4)
endif() # USE_BITINT_EXTENSION_INT4 endif() # USE_BITINT_EXTENSION_INT4
if(GPU_TARGETS MATCHES "gfx1100") if(GPU_TARGETS MATCHES "gfx1100" OR GPU_TARGETS MATCHES "gfx1101" OR GPU_TARGETS MATCHES "gfx1102")
add_example_executable(example_grouped_conv_fwd_bias_relu_add_wmma_fp16 grouped_conv_fwd_bias_relu_add_wmma_fp16.cpp) add_example_executable(example_grouped_conv_fwd_bias_relu_add_wmma_fp16 grouped_conv_fwd_bias_relu_add_wmma_fp16.cpp)
endif() endif()
......
example/40_conv2d_fwd_quantization/CMakeLists.txt 0 → 100644
View file @ 9fb64dae
# Conv perlayer quantization
add_example_executable(example_conv2d_fwd_dl_perlayer_quantization_int8 conv2d_fwd_dl_perlayer_quantization_int8.cpp)
add_example_executable(example_conv2d_fwd_xdl_perlayer_quantization_int8 conv2d_fwd_xdl_perlayer_quantization_int8.cpp)
# Conv perchannel quantization
add_example_executable(example_conv2d_fwd_dl_perchannel_quantization_int8 conv2d_fwd_dl_perchannel_quantization_int8.cpp)
add_example_executable(example_conv2d_fwd_xdl_perchannel_quantization_int8 conv2d_fwd_xdl_perchannel_quantization_int8.cpp)
# Conv + bias + relu perlayer quantization
add_example_executable(example_conv2d_fwd_dl_bias_relu_perlayer_quantization_int8 conv2d_fwd_dl_bias_relu_perlayer_quantization_int8.cpp)
add_example_executable(example_conv2d_fwd_xdl_bias_relu_perlayer_quantization_int8 conv2d_fwd_xdl_bias_relu_perlayer_quantization_int8.cpp)
# Conv + bias + relu perchannel quantization
add_example_executable(example_conv2d_fwd_dl_bias_relu_perchannel_quantization_int8 conv2d_fwd_dl_bias_relu_perchannel_quantization_int8.cpp)
add_example_executable(example_conv2d_fwd_xdl_bias_relu_perchannel_quantization_int8 conv2d_fwd_xdl_bias_relu_perchannel_quantization_int8.cpp)
example/40_conv2d_fwd_quantization/common.hpp 0 → 100644
View file @ 9fb64dae
// SPDX-License-Identifier: MIT
// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
#pragma once
#include "ck/ck.hpp"
#include "ck/tensor_operation/gpu/device/tensor_layout.hpp"
#include "ck/tensor_operation/gpu/element/element_wise_operation.hpp"
#include "ck/library/utility/algorithm.hpp"
#include "ck/library/utility/check_err.hpp"
#include "ck/library/utility/device_memory.hpp"
#include "ck/library/utility/host_tensor.hpp"
#include "ck/library/utility/host_tensor_generator.hpp"
#include "ck/library/utility/literals.hpp"
#include "ck/library/utility/convolution_parameter.hpp"
#include "ck/library/utility/convolution_host_tensor_descriptor_helper.hpp"
#include "ck/library/reference_tensor_operation/cpu/reference_conv_fwd.hpp"
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