Skip to content

GitLab

"...resnet50_tensorflow.git" did not exist on "c303344e38b4639f24d8ff3de02090e329f0f0df"
Unverified Commit 1724f7c8 authored by zjing14's avatar zjing14 Committed by GitHub
Browse files

Merge branch 'develop' into lwpck-586

parents ddbeddb4 fa998675
Hide whitespace changes
Inline Side-by-side
Showing with 1224 additions and 71 deletions
client_example/09_quantization/CMakeLists.txt
View file @ 1724f7c8
...@@ -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 @ 1724f7c8
...@@ -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 @ 1724f7c8
...@@ -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 @ 1724f7c8
...@@ -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 @ 1724f7c8
...@@ -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 @ 1724f7c8
// 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
example/14_gemm_quantization/CMakeLists.txt
View file @ 1724f7c8
# 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 @ 1724f7c8
// 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 @ 1724f7c8
...@@ -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 @ 1724f7c8
// 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/40_conv2d_fwd_quantization/CMakeLists.txt 0 → 100644
View file @ 1724f7c8
# 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 @ 1724f7c8
// 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"
example/40_conv2d_fwd_quantization/conv2d_fwd_dl_bias_relu_perchannel_quantization_int8.cpp 0 → 100644
View file @ 1724f7c8
// SPDX-License-Identifier: MIT
// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
#include "common.hpp"
#include "ck/tensor_operation/gpu/device/device_grouped_conv_fwd_dl_multiple_d_nhwc_kyxc_nhwk.hpp"
using InDataType = int8_t;
using WeiDataType = int8_t;
using BiasDataType = int32_t;
using RequantScaleDataType = float;
using AccDataType = int32_t;
using OutDataType = int8_t;
template <ck::index_t... Is>
using S = ck::Sequence<Is...>;
using PassThrough = ck::tensor_operation::element_wise::PassThrough;
using InElementOp = PassThrough;
using WeiElementOp = PassThrough;
using ActivationOp = ck::tensor_operation::element_wise::Relu;
using OutElementOp = ck::tensor_operation::element_wise::Add_Activation_Mul2_Clamp<ActivationOp>;
static constexpr auto ConvSpec =
ck::tensor_operation::device::ConvolutionForwardSpecialization::Default;
static constexpr auto GemmSpec = ck::tensor_operation::device::GemmSpecialization::MNKPadding;
template <ck::index_t NDimSpatial,
typename InLayout,
typename WeiLayout,
typename BiasLayout,
typename RequantScaleLayout,
typename OutLayout>
using DeviceGroupedConvNDFwdInstance =
ck::tensor_operation::device::DeviceGroupedConvFwdDlMultipleD_NHWC_KYXC_NHWK<
NDimSpatial,
InDataType,
WeiDataType,
ck::Tuple<BiasDataType, RequantScaleDataType>,
OutDataType,
AccDataType,
InLayout,
WeiLayout,
ck::Tuple<BiasLayout, RequantScaleLayout>,
OutLayout,
InElementOp,
WeiElementOp,
OutElementOp,
ConvSpec, // ConvForwardSpecialization
GemmSpec, // GemmSpecialization
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
#include "run_conv2d_fwd_bias_relu_perchannel_quantization_example.inc"
int main() { run_conv2d_fwd_bias_relu_perchannel_quantization_example(); };
example/40_conv2d_fwd_quantization/conv2d_fwd_dl_bias_relu_perlayer_quantization_int8.cpp 0 → 100644
View file @ 1724f7c8
// SPDX-License-Identifier: MIT
// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
#include "common.hpp"
#include "ck/tensor_operation/gpu/device/device_grouped_conv_fwd_dl_multiple_d_nhwc_kyxc_nhwk.hpp"
using InDataType = int8_t;
using WeiDataType = int8_t;
using BiasDataType = int32_t;
using AccDataType = int32_t;
using OutDataType = int8_t;
template <ck::index_t... Is>
using S = ck::Sequence<Is...>;
using PassThrough = ck::tensor_operation::element_wise::PassThrough;
using InElementOp = PassThrough;
using WeiElementOp = PassThrough;
using ActivationOp = ck::tensor_operation::element_wise::Relu;
using OutElementOp = ck::tensor_operation::element_wise::Add_Activation_Mul_Clamp<ActivationOp>;
static constexpr auto ConvSpec =
ck::tensor_operation::device::ConvolutionForwardSpecialization::Default;
static constexpr auto GemmSpec = ck::tensor_operation::device::GemmSpecialization::MNKPadding;
template <ck::index_t NDimSpatial,
typename InLayout,
typename WeiLayout,
typename BiasLayout,
typename OutLayout>
using DeviceGroupedConvNDFwdInstance =
ck::tensor_operation::device::DeviceGroupedConvFwdDlMultipleD_NHWC_KYXC_NHWK<
NDimSpatial,
InDataType,
WeiDataType,
ck::Tuple<BiasDataType>,
OutDataType,
AccDataType,
InLayout,
WeiLayout,
ck::Tuple<BiasLayout>,
OutLayout,
InElementOp,
WeiElementOp,
OutElementOp,
ConvSpec, // ConvForwardSpecialization
GemmSpec, // GemmSpecialization
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
#include "run_conv2d_fwd_bias_relu_perlayer_quantization_example.inc"
int main() { run_conv2d_fwd_bias_relu_perlayer_quantization_example(); }
example/40_conv2d_fwd_quantization/conv2d_fwd_dl_perchannel_quantization_int8.cpp 0 → 100644
View file @ 1724f7c8
// SPDX-License-Identifier: MIT
// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
#include "common.hpp"
#include "ck/tensor_operation/gpu/device/device_grouped_conv_fwd_dl_multiple_d_nhwc_kyxc_nhwk.hpp"
using InDataType = int8_t;
using WeiDataType = int8_t;
using RequantScaleDataType = float;
using AccDataType = int32_t;
using OutDataType = int8_t;
template <ck::index_t... Is>
using S = ck::Sequence<Is...>;
using PassThrough = ck::tensor_operation::element_wise::PassThrough;
using InElementOp = PassThrough;
using WeiElementOp = PassThrough;
using ActivationOp = ck::tensor_operation::element_wise::Relu;
using OutElementOp = ck::tensor_operation::element_wise::Activation_Mul2_Clamp<ActivationOp>;
static constexpr auto ConvSpec =
ck::tensor_operation::device::ConvolutionForwardSpecialization::Default;
static constexpr auto GemmSpec = ck::tensor_operation::device::GemmSpecialization::MNKPadding;
template <ck::index_t NDimSpatial,
typename InLayout,
typename WeiLayout,
typename RequantScaleLayout,
typename OutLayout>
using DeviceGroupedConvNDFwdInstance =
ck::tensor_operation::device::DeviceGroupedConvFwdDlMultipleD_NHWC_KYXC_NHWK<
NDimSpatial,
InDataType,
WeiDataType,
ck::Tuple<RequantScaleDataType>,
OutDataType,
AccDataType,
InLayout,
WeiLayout,
ck::Tuple<RequantScaleLayout>,
OutLayout,
InElementOp,
WeiElementOp,
OutElementOp,
ConvSpec, // ConvForwardSpecialization
GemmSpec, // GemmSpecialization
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
#include "run_conv2d_fwd_perchannel_quantization_example.inc"
int main() { run_conv2d_fwd_perchannel_quantization_example(); }
example/40_conv2d_fwd_quantization/conv2d_fwd_dl_perlayer_quantization_int8.cpp 0 → 100644
View file @ 1724f7c8
// SPDX-License-Identifier: MIT
// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
#include "common.hpp"
#include "ck/tensor_operation/gpu/device/device_grouped_conv_fwd_dl_multiple_d_nhwc_kyxc_nhwk.hpp"
using InDataType = int8_t;
using WeiDataType = int8_t;
using AccDataType = int32_t;
using OutDataType = int8_t;
template <ck::index_t... Is>
using S = ck::Sequence<Is...>;
using PassThrough = ck::tensor_operation::element_wise::PassThrough;
using InElementOp = PassThrough;
using WeiElementOp = PassThrough;
using ActivationOp = PassThrough;
using OutElementOp = ck::tensor_operation::element_wise::Activation_Mul_Clamp<ActivationOp>;
static constexpr auto ConvSpec =
ck::tensor_operation::device::ConvolutionForwardSpecialization::Default;
static constexpr auto GemmSpec = ck::tensor_operation::device::GemmSpecialization::MNKPadding;
template <ck::index_t NDimSpatial, typename InLayout, typename WeiLayout, typename OutLayout>
using DeviceGroupedConvNDFwdInstance =
ck::tensor_operation::device::DeviceGroupedConvFwdDlMultipleD_NHWC_KYXC_NHWK<
NDimSpatial,
InDataType,
WeiDataType,
ck::Tuple<>,
OutDataType,
AccDataType,
InLayout,
WeiLayout,
ck::Tuple<>,
OutLayout,
InElementOp,
WeiElementOp,
OutElementOp,
ConvSpec, // ConvForwardSpecialization
GemmSpec, // GemmSpecialization
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
#include "run_conv2d_fwd_perlayer_quantization_example.inc"
int main() { run_conv2d_fwd_perlayer_quantization_example(); }
example/40_conv2d_fwd_quantization/conv2d_fwd_xdl_bias_relu_perchannel_quantization_int8.cpp 0 → 100644
View file @ 1724f7c8
// SPDX-License-Identifier: MIT
// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
#include "common.hpp"
#include "ck/tensor_operation/gpu/device/impl/device_grouped_conv_fwd_multiple_d_xdl_cshuffle.hpp"
using InDataType = int8_t;
using WeiDataType = int8_t;
using BiasDataType = int32_t;
using RequantScaleDataType = float;
using AccDataType = int32_t;
using CShuffleDataType = AccDataType;
using OutDataType = int8_t;
template <ck::index_t... Is>
using S = ck::Sequence<Is...>;
using PassThrough = ck::tensor_operation::element_wise::PassThrough;
using InElementOp = PassThrough;
using WeiElementOp = PassThrough;
using ActivationOp = ck::tensor_operation::element_wise::Relu;
using OutElementOp = ck::tensor_operation::element_wise::Add_Activation_Mul2_Clamp<ActivationOp>;
static constexpr auto ConvSpec =
ck::tensor_operation::device::ConvolutionForwardSpecialization::Default;
static constexpr auto GemmSpec = ck::tensor_operation::device::GemmSpecialization::MNKPadding;
template <ck::index_t NDimSpatial,
typename InLayout,
typename WeiLayout,
typename BiasLayout,
typename RequantScaleLayout,
typename OutLayout>
using DeviceGroupedConvNDFwdInstance =
ck::tensor_operation::device::DeviceGroupedConvFwdMultipleD_Xdl_CShuffle<
NDimSpatial,
InLayout,
WeiLayout,
ck::Tuple<BiasLayout, RequantScaleLayout>,
OutLayout,
InDataType,
WeiDataType,
AccDataType,
CShuffleDataType,
ck::Tuple<BiasDataType, RequantScaleDataType>,
OutDataType,
InElementOp,
WeiElementOp,
OutElementOp,
ConvSpec, // ConvForwardSpecialization
GemmSpec, // GemmSpecialization
1, //
256, // BlockSize
128, // MPerBlock
256, // NPerBlock
64, // KPerBlock
16, // AK1
16, // BK1
32, // MPerXdl
32, // NPerXdl
2, // MXdlPerWave
4, // NXdlPerWave
S<4, 64, 1>, // ABlockTransferThreadClusterLengths_AK0_M_AK1
S<1, 0, 2>, // ABlockTransferThreadClusterArrangeOrder
S<1, 0, 2>, // ABlockTransferSrcAccessOrder
2, // ABlockTransferSrcVectorDim
16, // ABlockTransferSrcScalarPerVector
16, // ABlockTransferDstScalarPerVector_AK1
1, // ABlockLdsExtraM
S<4, 64, 1>, // BBlockTransferThreadClusterLengths_BK0_N_BK1
S<1, 0, 2>, // BBlockTransferThreadClusterArrangeOrder
S<1, 0, 2>, // BBlockTransferSrcAccessOrder
2, // BBlockTransferSrcVectorDim
16, // BBlockTransferSrcScalarPerVector
16, // BBlockTransferDstScalarPerVector_BK1
1, // BBlockLdsExtraN
1,
1,
S<1, 64, 1, 4>,
8>;
#include "run_conv2d_fwd_bias_relu_perchannel_quantization_example.inc"
int main() { run_conv2d_fwd_bias_relu_perchannel_quantization_example(); };
example/40_conv2d_fwd_quantization/conv2d_fwd_xdl_bias_relu_perlayer_quantization_int8.cpp 0 → 100644
View file @ 1724f7c8
// SPDX-License-Identifier: MIT
// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
#include "common.hpp"
#include "ck/tensor_operation/gpu/device/impl/device_grouped_conv_fwd_multiple_d_xdl_cshuffle.hpp"
using InDataType = int8_t;
using WeiDataType = int8_t;
using BiasDataType = int32_t;
using AccDataType = int32_t;
using CShuffleDataType = AccDataType;
using OutDataType = int8_t;
template <ck::index_t... Is>
using S = ck::Sequence<Is...>;
using PassThrough = ck::tensor_operation::element_wise::PassThrough;
using InElementOp = PassThrough;
using WeiElementOp = PassThrough;
using ActivationOp = ck::tensor_operation::element_wise::Relu;
using OutElementOp = ck::tensor_operation::element_wise::Add_Activation_Mul_Clamp<ActivationOp>;
static constexpr auto ConvSpec =
ck::tensor_operation::device::ConvolutionForwardSpecialization::Default;
static constexpr auto GemmSpec = ck::tensor_operation::device::GemmSpecialization::MNKPadding;
template <ck::index_t NDimSpatial,
typename InLayout,
typename WeiLayout,
typename BiasLayout,
typename OutLayout>
using DeviceGroupedConvNDFwdInstance =
ck::tensor_operation::device::DeviceGroupedConvFwdMultipleD_Xdl_CShuffle<
NDimSpatial,
InLayout,
WeiLayout,
ck::Tuple<BiasLayout>,
OutLayout,
InDataType,
WeiDataType,
AccDataType,
CShuffleDataType,
ck::Tuple<BiasDataType>,
OutDataType,
InElementOp,
WeiElementOp,
OutElementOp,
ConvSpec, // ConvForwardSpecialization
GemmSpec, // GemmSpecialization
1, //
256, // BlockSize
128, // MPerBlock
256, // NPerBlock
64, // KPerBlock
16, // AK1
16, // BK1
32, // MPerXdl
32, // NPerXdl
2, // MXdlPerWave
4, // NXdlPerWave
S<4, 64, 1>, // ABlockTransferThreadClusterLengths_AK0_M_AK1
S<1, 0, 2>, // ABlockTransferThreadClusterArrangeOrder
S<1, 0, 2>, // ABlockTransferSrcAccessOrder
2, // ABlockTransferSrcVectorDim
16, // ABlockTransferSrcScalarPerVector
16, // ABlockTransferDstScalarPerVector_AK1
1, // ABlockLdsExtraM
S<4, 64, 1>, // BBlockTransferThreadClusterLengths_BK0_N_BK1
S<1, 0, 2>, // BBlockTransferThreadClusterArrangeOrder
S<1, 0, 2>, // BBlockTransferSrcAccessOrder
2, // BBlockTransferSrcVectorDim
16, // BBlockTransferSrcScalarPerVector
16, // BBlockTransferDstScalarPerVector_BK1
1, // BBlockLdsExtraN
1,
1,
S<1, 64, 1, 4>,
8>;
#include "run_conv2d_fwd_bias_relu_perlayer_quantization_example.inc"
int main() { run_conv2d_fwd_bias_relu_perlayer_quantization_example(); }
example/40_conv2d_fwd_quantization/conv2d_fwd_xdl_perchannel_quantization_int8.cpp 0 → 100644
View file @ 1724f7c8
// SPDX-License-Identifier: MIT
// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
#include "common.hpp"
#include "ck/tensor_operation/gpu/device/impl/device_grouped_conv_fwd_multiple_d_xdl_cshuffle.hpp"
using InDataType = int8_t;
using WeiDataType = int8_t;
using RequantScaleDataType = float;
using AccDataType = int32_t;
using CShuffleDataType = AccDataType;
using OutDataType = int8_t;
template <ck::index_t... Is>
using S = ck::Sequence<Is...>;
using PassThrough = ck::tensor_operation::element_wise::PassThrough;
using InElementOp = PassThrough;
using WeiElementOp = PassThrough;
using ActivationOp = ck::tensor_operation::element_wise::Relu;
using OutElementOp = ck::tensor_operation::element_wise::Activation_Mul2_Clamp<ActivationOp>;
static constexpr auto ConvSpec =
ck::tensor_operation::device::ConvolutionForwardSpecialization::Default;
static constexpr auto GemmSpec = ck::tensor_operation::device::GemmSpecialization::MNKPadding;
template <ck::index_t NDimSpatial,
typename InLayout,
typename WeiLayout,
typename RequantScaleLayout,
typename OutLayout>
using DeviceGroupedConvNDFwdInstance =
ck::tensor_operation::device::DeviceGroupedConvFwdMultipleD_Xdl_CShuffle<
NDimSpatial,
InLayout,
WeiLayout,
ck::Tuple<RequantScaleLayout>,
OutLayout,
InDataType,
WeiDataType,
AccDataType,
CShuffleDataType,
ck::Tuple<RequantScaleDataType>,
OutDataType,
InElementOp,
WeiElementOp,
OutElementOp,
ConvSpec, // ConvForwardSpecialization
GemmSpec, // GemmSpecialization
1, //
256, // BlockSize
128, // MPerBlock
256, // NPerBlock
64, // KPerBlock
16, // AK1
16, // BK1
32, // MPerXdl
32, // NPerXdl
2, // MXdlPerWave
4, // NXdlPerWave
S<4, 64, 1>, // ABlockTransferThreadClusterLengths_AK0_M_AK1
S<1, 0, 2>, // ABlockTransferThreadClusterArrangeOrder
S<1, 0, 2>, // ABlockTransferSrcAccessOrder
2, // ABlockTransferSrcVectorDim
16, // ABlockTransferSrcScalarPerVector
16, // ABlockTransferDstScalarPerVector_AK1
1, // ABlockLdsExtraM
S<4, 64, 1>, // BBlockTransferThreadClusterLengths_BK0_N_BK1
S<1, 0, 2>, // BBlockTransferThreadClusterArrangeOrder
S<1, 0, 2>, // BBlockTransferSrcAccessOrder
2, // BBlockTransferSrcVectorDim
16, // BBlockTransferSrcScalarPerVector
16, // BBlockTransferDstScalarPerVector_BK1
1, // BBlockLdsExtraN
1,
1,
S<1, 64, 1, 4>,
8>;
#include "run_conv2d_fwd_perchannel_quantization_example.inc"
int main() { run_conv2d_fwd_perchannel_quantization_example(); }
example/40_conv2d_fwd_quantization/conv2d_fwd_xdl_perlayer_quantization_int8.cpp 0 → 100644
View file @ 1724f7c8
// SPDX-License-Identifier: MIT
// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
#include "common.hpp"
#include "ck/tensor_operation/gpu/device/impl/device_grouped_conv_fwd_multiple_d_xdl_cshuffle.hpp"
using InDataType = int8_t;
using WeiDataType = int8_t;
using AccDataType = int32_t;
using CShuffleDataType = AccDataType;
using OutDataType = int8_t;
template <ck::index_t... Is>
using S = ck::Sequence<Is...>;
using PassThrough = ck::tensor_operation::element_wise::PassThrough;
using InElementOp = PassThrough;
using WeiElementOp = PassThrough;
using ActivationOp = PassThrough;
using OutElementOp = ck::tensor_operation::element_wise::Activation_Mul_Clamp<ActivationOp>;
static constexpr auto ConvSpec =
ck::tensor_operation::device::ConvolutionForwardSpecialization::Default;
static constexpr auto GemmSpec = ck::tensor_operation::device::GemmSpecialization::MNKPadding;
template <ck::index_t NDimSpatial, typename InLayout, typename WeiLayout, typename OutLayout>
using DeviceGroupedConvNDFwdInstance =
ck::tensor_operation::device::DeviceGroupedConvFwdMultipleD_Xdl_CShuffle<
NDimSpatial,
InLayout,
WeiLayout,
ck::Tuple<>,
OutLayout,
InDataType,
WeiDataType,
AccDataType,
CShuffleDataType,
ck::Tuple<>,
OutDataType,
InElementOp,
WeiElementOp,
OutElementOp,
ConvSpec, // ConvForwardSpecialization
GemmSpec, // GemmSpecialization
1, //
256, // BlockSize
128, // MPerBlock
256, // NPerBlock
64, // KPerBlock
16, // AK1
16, // BK1
32, // MPerXdl
32, // NPerXdl
2, // MXdlPerWave
4, // NXdlPerWave
S<4, 64, 1>, // ABlockTransferThreadClusterLengths_AK0_M_AK1
S<1, 0, 2>, // ABlockTransferThreadClusterArrangeOrder
S<1, 0, 2>, // ABlockTransferSrcAccessOrder
2, // ABlockTransferSrcVectorDim
16, // ABlockTransferSrcScalarPerVector
16, // ABlockTransferDstScalarPerVector_AK1
1, // ABlockLdsExtraM
S<4, 64, 1>, // BBlockTransferThreadClusterLengths_BK0_N_BK1
S<1, 0, 2>, // BBlockTransferThreadClusterArrangeOrder
S<1, 0, 2>, // BBlockTransferSrcAccessOrder
2, // BBlockTransferSrcVectorDim
16, // BBlockTransferSrcScalarPerVector
16, // BBlockTransferDstScalarPerVector_BK1
1, // BBlockLdsExtraN
1,
1,
S<1, 64, 1, 4>,
16>;
#include "run_conv2d_fwd_perlayer_quantization_example.inc"
int main() { run_conv2d_fwd_perlayer_quantization_example(); }
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