Skip to content

GitLab

"...composable_kernel_rocm.git" did not exist on "27a05c7ee65d94bf44a2e45afce9e438cbd9cd51"
Commit 10cabc2d authored by Adam Osewski's avatar Adam Osewski
Browse files

Merge remote-tracking branch 'origin/develop' into aosewski/ggemm_splitk

parents baf68688 091570f5
Hide whitespace changes
Inline Side-by-side
Showing with 767 additions and 110 deletions
CMakeLists.txt
View file @ 10cabc2d
......@@ -22,6 +22,7 @@ include(TargetFlags)
list(APPEND CMAKE_PREFIX_PATH ${CMAKE_INSTALL_PREFIX} ${CMAKE_INSTALL_PREFIX}/llvm ${CMAKE_INSTALL_PREFIX}/hip /opt/rocm /opt/rocm/llvm /opt/rocm/hip)
option(USE_BITINT_EXTENSION_INT4, "Whether to enable clang's BitInt extension to provide int4 data type." OFF)
option(USE_OPT_NAVI3X, "Whether to enable LDS cumode and Wavefront32 mode for NAVI3X silicons." OFF)
if(USE_BITINT_EXTENSION_INT4)
add_compile_definitions(CK_EXPERIMENTAL_BIT_INT_EXTENSION_INT4)
......@@ -29,6 +30,12 @@ if(USE_BITINT_EXTENSION_INT4)
message("CK compiled with USE_BITINT_EXTENSION_INT4 set to ${USE_BITINT_EXTENSION_INT4}")
endif()
if(USE_OPT_NAVI3X)
add_compile_options(-mcumode)
add_compile_options(-mno-wavefrontsize64)
message("CK compiled with USE_OPT_NAVI3X set to ${USE_OPT_NAVI3X}")
endif()
## Threads
set(THREADS_PREFER_PTHREAD_FLAG ON)
find_package(Threads REQUIRED)
......
client_example/09_quantization/CMakeLists.txt
View file @ 10cabc2d
add_executable(client_conv2d_fwd_bias_tanh_perchannel_quantization conv2d_fwd_bias_tanh_perchannel_quantization.cpp)
target_link_libraries(client_conv2d_fwd_bias_tanh_perchannel_quantization PRIVATE composable_kernel::device_operations)
add_executable(client_conv2d_fwd_bias_relu_perchannel_quantization conv2d_fwd_bias_relu_perchannel_quantization.cpp)
target_link_libraries(client_conv2d_fwd_bias_relu_perchannel_quantization PRIVATE composable_kernel::device_operations)
add_executable(client_conv2d_fwd_bias_tanh_perlayer_quantization conv2d_fwd_bias_tanh_perlayer_quantization.cpp)
target_link_libraries(client_conv2d_fwd_bias_tanh_perlayer_quantization PRIVATE composable_kernel::device_operations)
add_executable(client_conv2d_fwd_bias_relu_perlayer_quantization conv2d_fwd_bias_relu_perlayer_quantization.cpp)
target_link_libraries(client_conv2d_fwd_bias_relu_perlayer_quantization PRIVATE composable_kernel::device_operations)
......
client_example/09_quantization/conv2d_fwd_bias_relu_perlayer_quantization.cpp
View file @ 10cabc2d
......@@ -26,15 +26,16 @@ using OutElementOp = ck::tensor_operation::element_wise::Add_Activation_Mul_Clam
static constexpr ck::index_t NumDimSpatial = 2;
static constexpr ck::index_t G = 1;
static constexpr ck::index_t N = 4; // batch size
static constexpr ck::index_t K = 64; // output channel
static constexpr ck::index_t C = 192; // input channel
static constexpr ck::index_t Y = 3; // filter H
static constexpr ck::index_t X = 3; // filter W
static constexpr ck::index_t Hi = 71; // input H
static constexpr ck::index_t Wi = 71; // input W
static constexpr ck::index_t Ho = 36; // output H
static constexpr ck::index_t Wo = 36; // output W
static constexpr ck::index_t N = 4; // batch size
static constexpr ck::index_t K = 64; // output channel
static constexpr ck::index_t C = 192; // input channel
static constexpr ck::index_t Y = 3; // filter H
static constexpr ck::index_t X = 3; // filter W
static constexpr ck::index_t Hi = 71; // input H
static constexpr ck::index_t Wi = 71; // input W
static constexpr ck::index_t Ho = 36; // output H
static constexpr ck::index_t Wo = 36; // output W
static constexpr float requant_scale = 0.5f; // requantize qAcc to qz
struct SimpleDeviceMem
{
......@@ -102,26 +103,27 @@ int main(int argc, char* argv[])
for(int i = 0; i < op_ptrs.size(); ++i)
{
auto& op_ptr = op_ptrs[i];
auto argument_ptr = op_ptr->MakeArgumentPointer(in.GetDeviceBuffer(),
wei.GetDeviceBuffer(),
{bias.GetDeviceBuffer()},
out.GetDeviceBuffer(),
in_lengths,
in_strides,
weight_lengths,
weight_strides,
{bias_lengths},
{bias_strides},
out_lengths,
out_strides,
conv_strides,
conv_dilations,
in_left_pad,
in_right_pad,
PassThrough{},
PassThrough{},
OutElementOp{0.5f, ActivationOp{}});
auto& op_ptr = op_ptrs[i];
auto argument_ptr =
op_ptr->MakeArgumentPointer(in.GetDeviceBuffer(),
wei.GetDeviceBuffer(),
{bias.GetDeviceBuffer()},
out.GetDeviceBuffer(),
in_lengths,
in_strides,
weight_lengths,
weight_strides,
{bias_lengths},
{bias_strides},
out_lengths,
out_strides,
conv_strides,
conv_dilations,
in_left_pad,
in_right_pad,
PassThrough{},
PassThrough{},
OutElementOp{requant_scale, ActivationOp{}});
auto invoker_ptr = op_ptr->MakeInvokerPointer();
std::string op_name = op_ptr->GetTypeString();
......@@ -165,25 +167,26 @@ int main(int argc, char* argv[])
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(in.GetDeviceBuffer(),
wei.GetDeviceBuffer(),
{bias.GetDeviceBuffer()},
out.GetDeviceBuffer(),
in_lengths,
in_strides,
weight_lengths,
weight_strides,
{bias_lengths},
{bias_strides},
out_lengths,
out_strides,
conv_strides,
conv_dilations,
in_left_pad,
in_right_pad,
PassThrough{},
PassThrough{},
OutElementOp{0.5f, ActivationOp{}});
auto argument_ptr =
op_ptr->MakeArgumentPointer(in.GetDeviceBuffer(),
wei.GetDeviceBuffer(),
{bias.GetDeviceBuffer()},
out.GetDeviceBuffer(),
in_lengths,
in_strides,
weight_lengths,
weight_strides,
{bias_lengths},
{bias_strides},
out_lengths,
out_strides,
conv_strides,
conv_dilations,
in_left_pad,
in_right_pad,
PassThrough{},
PassThrough{},
OutElementOp{requant_scale, ActivationOp{}});
auto invoker_ptr = op_ptr->MakeInvokerPointer();
......
client_example/09_quantization/conv2d_fwd_bias_tanh_perchannel_quantization.cpp 0 → 100644
View file @ 10cabc2d
// 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/library/tensor_operation_instance/gpu/quantization/grouped_convolution_bias_forward_perchannel_quantization.hpp"
#include "ck/tensor_operation/gpu/device/tensor_layout.hpp"
#include "ck/tensor_operation/gpu/device/device_conv_fwd.hpp"
#include "ck/tensor_operation/gpu/element/element_wise_operation.hpp"
using InDataType = int8_t;
using WeiDataType = int8_t;
using BiasDataType = int32_t;
using RequantScaleDataType = float;
using OutDataType = int8_t;
using InLayout = ck::tensor_layout::convolution::GNHWC;
using WeiLayout = ck::tensor_layout::convolution::GKYXC;
using BiasLayout = ck::tensor_layout::convolution::G_K;
using RequantScaleLayout = ck::tensor_layout::convolution::G_K;
using OutLayout = ck::tensor_layout::convolution::GNHWK;
using PassThrough = ck::tensor_operation::element_wise::PassThrough;
using ActivationOp = ck::tensor_operation::element_wise::TanH;
using OutElementOp =
ck::tensor_operation::element_wise::Add_Mul2_Activation_Mul_Clamp<ActivationOp>;
static constexpr ck::index_t NumDimSpatial = 2;
static constexpr ck::index_t G = 1;
static constexpr ck::index_t N = 4; // batch size
static constexpr ck::index_t K = 64; // output channel
static constexpr ck::index_t C = 192; // input channel
static constexpr ck::index_t Y = 3; // filter H
static constexpr ck::index_t X = 3; // filter W
static constexpr ck::index_t Hi = 71; // input H
static constexpr ck::index_t Wi = 71; // input W
static constexpr ck::index_t Ho = 36; // output H
static constexpr ck::index_t Wo = 36; // output W
static constexpr float sz_inv = 0.5f; // inverse of scale_z
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[])
{
std::array<ck::index_t, 5> in_lengths{G, N, C, Hi, Wi};
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_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_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_strides{K, 0, 1, 0, 0};
std::array<ck::index_t, 5> out_lengths{G, N, K, Ho, Wo};
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_right_pad{1, 1};
std::array<ck::index_t, 2> conv_strides{2, 2};
std::array<ck::index_t, 2> conv_dilations{1, 1};
SimpleDeviceMem in(sizeof(InDataType) * N * Hi * Wi * C);
SimpleDeviceMem wei(sizeof(WeiDataType) * K * Y * X * C);
SimpleDeviceMem bias(sizeof(BiasDataType) * K * Y * X * C);
SimpleDeviceMem requant_scale(sizeof(RequantScaleDataType) * K * Y * X * C);
SimpleDeviceMem out(sizeof(OutDataType) * N * Ho * Wo * K);
using DeviceOp = ck::tensor_operation::device::DeviceGroupedConvFwdMultipleD<
NumDimSpatial,
InLayout,
WeiLayout,
ck::Tuple<BiasLayout, RequantScaleLayout>,
OutLayout,
InDataType,
WeiDataType,
ck::Tuple<BiasDataType, RequantScaleDataType>,
OutDataType,
PassThrough,
PassThrough,
OutElementOp>;
// 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;
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(in.GetDeviceBuffer(),
wei.GetDeviceBuffer(),
{bias.GetDeviceBuffer(), requant_scale.GetDeviceBuffer()},
out.GetDeviceBuffer(),
in_lengths,
in_strides,
weight_lengths,
weight_strides,
{bias_lengths, requant_scale_lengths},
{bias_strides, requant_scale_strides},
out_lengths,
out_strides,
conv_strides,
conv_dilations,
in_left_pad,
in_right_pad,
PassThrough{},
PassThrough{},
OutElementOp{sz_inv, ActivationOp{}});
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 = G * 2 * N * K * C * Ho * Wo * Y * X;
std::size_t num_bytes =
G * sizeof(InDataType) * N * Hi * Wi * C + G * sizeof(WeiDataType) * K * Y * X * C +
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 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;
}
}
// 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];
std::cout << "Run the best instance without timing: " << op_ptr->GetTypeString()
<< std::endl;
auto argument_ptr =
op_ptr->MakeArgumentPointer(in.GetDeviceBuffer(),
wei.GetDeviceBuffer(),
{bias.GetDeviceBuffer(), requant_scale.GetDeviceBuffer()},
out.GetDeviceBuffer(),
in_lengths,
in_strides,
weight_lengths,
weight_strides,
{bias_lengths, requant_scale_lengths},
{bias_strides, requant_scale_strides},
out_lengths,
out_strides,
conv_strides,
conv_dilations,
in_left_pad,
in_right_pad,
PassThrough{},
PassThrough{},
OutElementOp{sz_inv, ActivationOp{}});
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/09_quantization/conv2d_fwd_bias_tanh_perlayer_quantization.cpp 0 → 100644
View file @ 10cabc2d
// 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/library/tensor_operation_instance/gpu/quantization/grouped_convolution_bias_forward_perlayer_quantization.hpp"
#include "ck/tensor_operation/gpu/device/tensor_layout.hpp"
#include "ck/tensor_operation/gpu/device/device_conv_fwd.hpp"
#include "ck/tensor_operation/gpu/element/element_wise_operation.hpp"
using InDataType = int8_t;
using WeiDataType = int8_t;
using BiasDataType = int32_t;
using OutDataType = int8_t;
using InLayout = ck::tensor_layout::convolution::GNHWC;
using WeiLayout = ck::tensor_layout::convolution::GKYXC;
using BiasLayout = ck::tensor_layout::convolution::G_K;
using OutLayout = ck::tensor_layout::convolution::GNHWK;
using PassThrough = ck::tensor_operation::element_wise::PassThrough;
using ActivationOp = ck::tensor_operation::element_wise::TanH;
using OutElementOp = ck::tensor_operation::element_wise::Add_Mul_Activation_Mul_Clamp<ActivationOp>;
static constexpr ck::index_t NumDimSpatial = 2;
static constexpr ck::index_t G = 1;
static constexpr ck::index_t N = 4; // batch size
static constexpr ck::index_t K = 64; // output channel
static constexpr ck::index_t C = 192; // input channel
static constexpr ck::index_t Y = 3; // filter H
static constexpr ck::index_t X = 3; // filter W
static constexpr ck::index_t Hi = 71; // input H
static constexpr ck::index_t Wi = 71; // input W
static constexpr ck::index_t Ho = 36; // output H
static constexpr ck::index_t Wo = 36; // output W
static constexpr float sacc = 0.5f; // scale of acc
static constexpr float sz_inv = 0.5f; // inverse of scale_z
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[])
{
std::array<ck::index_t, 5> in_lengths{G, N, C, Hi, Wi};
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_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_strides{K, 0, 1, 0, 0};
std::array<ck::index_t, 5> out_lengths{G, N, K, Ho, Wo};
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_right_pad{1, 1};
std::array<ck::index_t, 2> conv_strides{2, 2};
std::array<ck::index_t, 2> conv_dilations{1, 1};
SimpleDeviceMem in(sizeof(InDataType) * N * Hi * Wi * C);
SimpleDeviceMem wei(sizeof(WeiDataType) * K * Y * X * C);
SimpleDeviceMem bias(sizeof(BiasDataType) * K * Y * X * C);
SimpleDeviceMem out(sizeof(OutDataType) * N * Ho * Wo * K);
using DeviceOp =
ck::tensor_operation::device::DeviceGroupedConvFwdMultipleD<NumDimSpatial,
InLayout,
WeiLayout,
ck::Tuple<BiasLayout>,
OutLayout,
InDataType,
WeiDataType,
ck::Tuple<BiasDataType>,
OutDataType,
PassThrough,
PassThrough,
OutElementOp>;
// 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;
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(in.GetDeviceBuffer(),
wei.GetDeviceBuffer(),
{bias.GetDeviceBuffer()},
out.GetDeviceBuffer(),
in_lengths,
in_strides,
weight_lengths,
weight_strides,
{bias_lengths},
{bias_strides},
out_lengths,
out_strides,
conv_strides,
conv_dilations,
in_left_pad,
in_right_pad,
PassThrough{},
PassThrough{},
OutElementOp{sacc, sz_inv, ActivationOp{}});
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 = G * 2 * N * K * C * Ho * Wo * Y * X;
std::size_t num_bytes =
G * sizeof(InDataType) * N * Hi * Wi * C + G * sizeof(WeiDataType) * K * Y * X * C +
G * sizeof(BiasDataType) * K + G * sizeof(OutDataType) * N * Ho * Wo * K;
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;
}
}
// 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];
std::cout << "Run the best instance without timing: " << op_ptr->GetTypeString()
<< std::endl;
auto argument_ptr = op_ptr->MakeArgumentPointer(in.GetDeviceBuffer(),
wei.GetDeviceBuffer(),
{bias.GetDeviceBuffer()},
out.GetDeviceBuffer(),
in_lengths,
in_strides,
weight_lengths,
weight_strides,
{bias_lengths},
{bias_strides},
out_lengths,
out_strides,
conv_strides,
conv_dilations,
in_left_pad,
in_right_pad,
PassThrough{},
PassThrough{},
OutElementOp{sacc, sz_inv, ActivationOp{}});
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/09_quantization/conv2d_fwd_perlayer_quantization.cpp
View file @ 10cabc2d
......@@ -24,15 +24,16 @@ using OutElementOp = ck::tensor_operation::element_wise::Activation_Mul_Clamp<Ac
static constexpr ck::index_t NumDimSpatial = 2;
static constexpr ck::index_t G = 1;
static constexpr ck::index_t N = 4; // batch size
static constexpr ck::index_t K = 64; // output channel
static constexpr ck::index_t C = 192; // input channel
static constexpr ck::index_t Y = 3; // filter H
static constexpr ck::index_t X = 3; // filter W
static constexpr ck::index_t Hi = 71; // input H
static constexpr ck::index_t Wi = 71; // input W
static constexpr ck::index_t Ho = 36; // output H
static constexpr ck::index_t Wo = 36; // output W
static constexpr ck::index_t N = 4; // batch size
static constexpr ck::index_t K = 64; // output channel
static constexpr ck::index_t C = 192; // input channel
static constexpr ck::index_t Y = 3; // filter H
static constexpr ck::index_t X = 3; // filter W
static constexpr ck::index_t Hi = 71; // input H
static constexpr ck::index_t Wi = 71; // input W
static constexpr ck::index_t Ho = 36; // output H
static constexpr ck::index_t Wo = 36; // output W
static constexpr float requant_scale = 0.5f; // requantize qAcc to qY
struct SimpleDeviceMem
{
......@@ -96,26 +97,27 @@ int main(int argc, char* argv[])
for(int i = 0; i < op_ptrs.size(); ++i)
{
auto& op_ptr = op_ptrs[i];
auto argument_ptr = op_ptr->MakeArgumentPointer(in.GetDeviceBuffer(),
wei.GetDeviceBuffer(),
{},
out.GetDeviceBuffer(),
in_lengths,
in_strides,
weight_lengths,
weight_strides,
{},
{},
out_lengths,
out_strides,
conv_strides,
conv_dilations,
in_left_pad,
in_right_pad,
PassThrough{},
PassThrough{},
OutElementOp{0.5f, ActivationOp{}});
auto& op_ptr = op_ptrs[i];
auto argument_ptr =
op_ptr->MakeArgumentPointer(in.GetDeviceBuffer(),
wei.GetDeviceBuffer(),
{},
out.GetDeviceBuffer(),
in_lengths,
in_strides,
weight_lengths,
weight_strides,
{},
{},
out_lengths,
out_strides,
conv_strides,
conv_dilations,
in_left_pad,
in_right_pad,
PassThrough{},
PassThrough{},
OutElementOp{requant_scale, ActivationOp{}});
auto invoker_ptr = op_ptr->MakeInvokerPointer();
std::string op_name = op_ptr->GetTypeString();
......@@ -158,25 +160,26 @@ int main(int argc, char* argv[])
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(in.GetDeviceBuffer(),
wei.GetDeviceBuffer(),
{},
out.GetDeviceBuffer(),
in_lengths,
in_strides,
weight_lengths,
weight_strides,
{},
{},
out_lengths,
out_strides,
conv_strides,
conv_dilations,
in_left_pad,
in_right_pad,
PassThrough{},
PassThrough{},
OutElementOp{0.5f, ActivationOp{}});
auto argument_ptr =
op_ptr->MakeArgumentPointer(in.GetDeviceBuffer(),
wei.GetDeviceBuffer(),
{},
out.GetDeviceBuffer(),
in_lengths,
in_strides,
weight_lengths,
weight_strides,
{},
{},
out_lengths,
out_strides,
conv_strides,
conv_dilations,
in_left_pad,
in_right_pad,
PassThrough{},
PassThrough{},
OutElementOp{requant_scale, ActivationOp{}});
auto invoker_ptr = op_ptr->MakeInvokerPointer();
......
docs/.sphinx/requirements.in
View file @ 10cabc2d
git+https://github.com/RadeonOpenCompute/rocm-docs-core.git
sphinxcontrib-bibtex==2.5.0
docs/.sphinx/requirements.txt
View file @ 10cabc2d
......@@ -46,9 +46,11 @@ docutils==0.16
# via
# breathe
# myst-parser
# pybtex-docutils
# pydata-sphinx-theme
# rocm-docs-core
# sphinx
# sphinxcontrib-bibtex
executing==1.2.0
# via stack-data
fastjsonschema==2.16.3
......@@ -94,6 +96,8 @@ jupyter-core==5.3.0
# ipykernel
# jupyter-client
# nbformat
latexcodec==2.0.1
# via pybtex
linkify-it-py==1.0.3
# via myst-parser
markdown-it-py==2.2.0
......@@ -150,6 +154,12 @@ ptyprocess==0.7.0
# via pexpect
pure-eval==0.2.2
# via stack-data
pybtex==0.24.0
# via
# pybtex-docutils
# sphinxcontrib-bibtex
pybtex-docutils==1.0.2
# via sphinxcontrib-bibtex
pycparser==2.21
# via cffi
pydata-sphinx-theme==0.13.1
......@@ -175,6 +185,7 @@ pyyaml==6.0
# jupyter-cache
# myst-nb
# myst-parser
# pybtex
# sphinx-external-toc
pyzmq==25.0.2
# via
......@@ -189,6 +200,8 @@ rocm-docs-core @ git+https://github.com/RadeonOpenCompute/rocm-docs-core.git
six==1.16.0
# via
# asttokens
# latexcodec
# pybtex
# python-dateutil
smmap==5.0.0
# via gitdb
......@@ -208,6 +221,7 @@ sphinx==4.3.1
# sphinx-design
# sphinx-external-toc
# sphinx-notfound-page
# sphinxcontrib-bibtex
sphinx-book-theme==1.0.0rc2
# via rocm-docs-core
sphinx-copybutton==0.5.1
......@@ -220,6 +234,10 @@ sphinx-notfound-page==0.8.3
# via rocm-docs-core
sphinxcontrib-applehelp==1.0.4
# via sphinx
sphinxcontrib-bibtex==2.5.0
# via
# -r requirements.in
# rocm-docs-core
sphinxcontrib-devhelp==1.0.2
# via sphinx
sphinxcontrib-htmlhelp==2.0.1
......
docs/conf.py
View file @ 10cabc2d
......@@ -18,7 +18,8 @@ mathjax3_config = {
}
}
bibtex_bibfiles = ['refs.bib']
for sphinx_var in ROCmDocs.SPHINX_VARS:
globals()[sphinx_var] = getattr(docs_core, sphinx_var)
extensions += ['sphinxcontrib.bibtex']
bibtex_bibfiles = ['refs.bib']
example/30_grouped_conv_fwd_multiple_d/grouped_conv_fwd_xdl_fp16.cpp
View file @ 10cabc2d
// SPDX-License-Identifier: MIT
// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
#include "common.hpp"
......
example/40_conv2d_fwd_quantization/CMakeLists.txt
View file @ 10cabc2d
......@@ -14,3 +14,8 @@ add_example_executable(example_conv2d_fwd_xdl_bias_relu_perlayer_quantization_in
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)
# Conv + bias + tanh perlayer quantization
add_example_executable(example_conv2d_fwd_dl_bias_tanh_perlayer_quantization_int8 conv2d_fwd_dl_bias_tanh_perlayer_quantization_int8.cpp)
# Conv + bias + tanh perchannel quantization
add_example_executable(example_conv2d_fwd_dl_bias_tanh_perchannel_quantization_int8 conv2d_fwd_dl_bias_tanh_perchannel_quantization_int8.cpp)
example/40_conv2d_fwd_quantization/conv2d_fwd_dl_bias_relu_perchannel_quantization_int8.cpp
View file @ 10cabc2d
......@@ -76,6 +76,10 @@ using DeviceGroupedConvNDFwdInstance =
5, // CThreadTransferSrcDstVectorDim
4>; // CThreadTransferDstScalarPerVector
#include "run_conv2d_fwd_bias_relu_perchannel_quantization_example.inc"
#include "run_conv2d_fwd_bias_perchannel_quantization_example.inc"
int main() { run_conv2d_fwd_bias_relu_perchannel_quantization_example(); };
int main()
{
const auto out_element_op = OutElementOp{ActivationOp{}};
run_conv2d_fwd_bias_perchannel_quantization_example(out_element_op);
};
example/40_conv2d_fwd_quantization/conv2d_fwd_dl_bias_relu_perlayer_quantization_int8.cpp
View file @ 10cabc2d
......@@ -74,6 +74,11 @@ using DeviceGroupedConvNDFwdInstance =
5, // CThreadTransferSrcDstVectorDim
4>; // CThreadTransferDstScalarPerVector
#include "run_conv2d_fwd_bias_relu_perlayer_quantization_example.inc"
#include "run_conv2d_fwd_bias_perlayer_quantization_example.inc"
int main() { run_conv2d_fwd_bias_relu_perlayer_quantization_example(); }
int main()
{
float requant_scale = 0.5f;
const auto out_element_op = OutElementOp{requant_scale, ActivationOp{}};
run_conv2d_fwd_bias_perlayer_quantization_example(out_element_op);
}
example/40_conv2d_fwd_quantization/conv2d_fwd_dl_bias_tanh_perchannel_quantization_int8.cpp 0 → 100644
View file @ 10cabc2d
// 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::TanH;
using OutElementOp =
ck::tensor_operation::element_wise::Add_Mul2_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 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_perchannel_quantization_example.inc"
int main()
{
float scale_z_inv = 0.5f;
const auto out_element_op = OutElementOp{scale_z_inv, ActivationOp{}};
run_conv2d_fwd_bias_perchannel_quantization_example(out_element_op);
};
example/40_conv2d_fwd_quantization/conv2d_fwd_dl_bias_tanh_perlayer_quantization_int8.cpp 0 → 100644
View file @ 10cabc2d
// 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::TanH;
using OutElementOp = ck::tensor_operation::element_wise::Add_Mul_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_perlayer_quantization_example.inc"
int main()
{
float scale_acc = 0.5f;
float scale_z_inv = 0.5f;
const auto out_element_op = OutElementOp{scale_z_inv, scale_acc, ActivationOp{}};
run_conv2d_fwd_bias_perlayer_quantization_example(out_element_op);
}
example/40_conv2d_fwd_quantization/conv2d_fwd_dl_perchannel_quantization_int8.cpp
View file @ 10cabc2d
......@@ -76,4 +76,8 @@ using DeviceGroupedConvNDFwdInstance =
#include "run_conv2d_fwd_perchannel_quantization_example.inc"
int main() { run_conv2d_fwd_perchannel_quantization_example(); }
int main()
{
const auto out_element_op = OutElementOp{ActivationOp{}};
run_conv2d_fwd_perchannel_quantization_example(out_element_op);
}
example/40_conv2d_fwd_quantization/conv2d_fwd_dl_perlayer_quantization_int8.cpp
View file @ 10cabc2d
......@@ -71,4 +71,9 @@ using DeviceGroupedConvNDFwdInstance =
#include "run_conv2d_fwd_perlayer_quantization_example.inc"
int main() { run_conv2d_fwd_perlayer_quantization_example(); }
int main()
{
float requant_scale = 0.5f;
const auto out_element_op = OutElementOp{requant_scale, ActivationOp{}};
run_conv2d_fwd_perlayer_quantization_example(out_element_op);
}
example/40_conv2d_fwd_quantization/conv2d_fwd_xdl_bias_relu_perchannel_quantization_int8.cpp
View file @ 10cabc2d
......@@ -80,6 +80,10 @@ using DeviceGroupedConvNDFwdInstance =
S<1, 64, 1, 4>,
8>;
#include "run_conv2d_fwd_bias_relu_perchannel_quantization_example.inc"
#include "run_conv2d_fwd_bias_perchannel_quantization_example.inc"
int main() { run_conv2d_fwd_bias_relu_perchannel_quantization_example(); };
int main()
{
const auto out_element_op = OutElementOp{ActivationOp{}};
run_conv2d_fwd_bias_perchannel_quantization_example(out_element_op);
};
example/40_conv2d_fwd_quantization/conv2d_fwd_xdl_bias_relu_perlayer_quantization_int8.cpp
View file @ 10cabc2d
......@@ -78,6 +78,11 @@ using DeviceGroupedConvNDFwdInstance =
S<1, 64, 1, 4>,
8>;
#include "run_conv2d_fwd_bias_relu_perlayer_quantization_example.inc"
#include "run_conv2d_fwd_bias_perlayer_quantization_example.inc"
int main() { run_conv2d_fwd_bias_relu_perlayer_quantization_example(); }
int main()
{
float requant_scale = 0.5f;
const auto out_element_op = OutElementOp{requant_scale, ActivationOp{}};
run_conv2d_fwd_bias_perlayer_quantization_example(out_element_op);
}
example/40_conv2d_fwd_quantization/conv2d_fwd_xdl_perchannel_quantization_int8.cpp
View file @ 10cabc2d
......@@ -80,4 +80,8 @@ using DeviceGroupedConvNDFwdInstance =
#include "run_conv2d_fwd_perchannel_quantization_example.inc"
int main() { run_conv2d_fwd_perchannel_quantization_example(); }
int main()
{
const auto out_element_op = OutElementOp{ActivationOp{}};
run_conv2d_fwd_perchannel_quantization_example(out_element_op);
}
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