Skip to content

GitLab

Unverified Commit bf5fe7b3 authored by rocking5566's avatar rocking5566 Committed by GitHub
Browse files

Merge branch 'develop' into conv_dlops/quantization

parents 86c32464 9096b1c7
Show whitespace changes
Inline Side-by-side
Showing with 740 additions and 37 deletions
client_example/17_grouped_gemm_fastgelu/CMakeLists.txt 0 → 100644
View file @ bf5fe7b3
add_executable(client_grouped_gemm_fastgelu grouped_gemm_fastgelu.cpp)
target_link_libraries(client_grouped_gemm_fastgelu PRIVATE composable_kernel::device_operations)
\ No newline at end of file
client_example/17_grouped_gemm_fastgelu/grouped_gemm_fastgelu.cpp 0 → 100644
View file @ bf5fe7b3
// SPDX-License-Identifier: MIT
// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
#include <iomanip>
#include <iostream>
#include <vector>
#include <random>
#include "ck/ck.hpp"
#include "ck/tensor_operation/gpu/device/tensor_layout.hpp"
#include "ck/tensor_operation/gpu/device/device_grouped_gemm.hpp"
#include "ck/tensor_operation/gpu/element/element_wise_operation.hpp"
#include "ck/library/tensor_operation_instance/gpu/grouped_gemm_fastgelu.hpp"
using F16 = ck::half_t;
using F32 = float;
using Row = ck::tensor_layout::gemm::RowMajor;
using Col = ck::tensor_layout::gemm::ColumnMajor;
using PassThrough = ck::tensor_operation::element_wise::PassThrough;
using FastGelu = ck::tensor_operation::element_wise::FastGelu;
using ADataType = F16;
using BDataType = F16;
using DsDataType = ck::Tuple<>;
using EDataType = F16;
using ALayout = Row;
using BLayout = Col;
using DsLayout = ck::Tuple<>;
using ELayout = Row;
using AElementOp = PassThrough;
using BElementOp = PassThrough;
using CDEElementOp = FastGelu;
struct SimpleDeviceMem
{
SimpleDeviceMem() = delete;
SimpleDeviceMem(std::size_t mem_size) : p_mem_{}
{
(void)hipMalloc(static_cast<void**>(&p_mem_), mem_size);
}
void* GetDeviceBuffer() { return p_mem_; }
~SimpleDeviceMem() { (void)hipFree(p_mem_); }
void* p_mem_;
};
int main()
{
std::mt19937 gen(19391);
std::uniform_int_distribution<> distrib(1, 10);
int group_count = distrib(gen);
std::vector<int> Ms, Ns, Ks, StrideAs, StrideBs, StrideEs;
for(int i = 0; i < group_count; ++i)
{
Ms.push_back(256 + 256 * distrib(gen));
Ns.push_back(256 + 256 * distrib(gen));
Ks.push_back(128 + 128 * distrib(gen));
StrideAs.push_back(std::is_same<Row, ALayout>::value ? Ks[i] : Ms[i]);
StrideBs.push_back(std::is_same<Row, BLayout>::value ? Ns[i] : Ks[i]);
StrideEs.push_back(std::is_same<Row, ELayout>::value ? Ns[i] : Ms[i]);
}
auto f_matrix_space_size =
[](std::size_t nRow, std::size_t nCol, std::size_t stride, auto layout) {
using Layout = decltype(layout);
if constexpr(std::is_same<Layout, ck::tensor_layout::gemm::RowMajor>::value)
{
return (nRow - 1) * stride + nCol;
}
else
{
return (nCol - 1) * stride + nRow;
}
};
std::vector<SimpleDeviceMem> a_dev_bufs, b_dev_bufs, e_dev_bufs;
a_dev_bufs.reserve(group_count);
b_dev_bufs.reserve(group_count);
e_dev_bufs.reserve(group_count);
std::vector<const void*> p_a, p_b;
std::vector<void*> p_e;
p_a.reserve(group_count);
p_b.reserve(group_count);
p_e.reserve(group_count);
std::vector<ck::tensor_operation::device::GemmDesc> gemm_descs;
gemm_descs.reserve(group_count);
for(int i = 0; i < group_count; ++i)
{
a_dev_bufs.emplace_back(sizeof(ADataType) *
f_matrix_space_size(Ms[i], Ks[i], StrideAs[i], ALayout{}));
b_dev_bufs.emplace_back(sizeof(BDataType) *
f_matrix_space_size(Ks[i], Ns[i], StrideBs[i], BLayout{}));
e_dev_bufs.emplace_back(sizeof(EDataType) *
f_matrix_space_size(Ms[i], Ns[i], StrideEs[i], ELayout{}));
gemm_descs.push_back({Ms[i], Ns[i], Ks[i], StrideAs[i], StrideBs[i], StrideEs[i], {}});
p_a.push_back(a_dev_bufs[i].GetDeviceBuffer());
p_b.push_back(b_dev_bufs[i].GetDeviceBuffer());
p_e.push_back(e_dev_bufs[i].GetDeviceBuffer());
}
using DeviceOp = ck::tensor_operation::device::DeviceGroupedGemm<ALayout,
BLayout,
DsLayout,
ELayout,
ADataType,
BDataType,
DsDataType,
EDataType,
AElementOp,
BElementOp,
CDEElementOp>;
// get device op instances
const auto op_ptrs = ck::tensor_operation::device::instance::DeviceOperationInstanceFactory<
DeviceOp>::GetInstances();
std::cout << "found " << op_ptrs.size() << " instances" << std::endl;
const auto a_element_op = AElementOp{};
const auto b_element_op = BElementOp{};
const auto cde_element_op = CDEElementOp{};
std::string best_op_name;
bool found = false;
int best_op_id = -1;
float best_ave_time = 0;
float best_tflops = 0;
float best_gb_per_sec = 0;
auto p_ds = std::vector<std::array<const void*, 0>>{};
// profile device operation instances
std::cout << "Run all instances and do timing" << std::endl;
for(int i = 0; i < op_ptrs.size(); ++i)
{
auto& op_ptr = op_ptrs[i];
auto argument_ptr = op_ptr->MakeArgumentPointer(
p_a, p_b, p_ds, p_e, gemm_descs, a_element_op, b_element_op, cde_element_op);
auto invoker_ptr = op_ptr->MakeInvokerPointer();
SimpleDeviceMem gemm_desc_workspace(op_ptr->GetWorkSpaceSize(argument_ptr.get()));
op_ptr->SetWorkSpacePointer(argument_ptr.get(), gemm_desc_workspace.GetDeviceBuffer());
std::string op_name = op_ptr->GetTypeString();
if(op_ptr->IsSupportedArgument(argument_ptr.get()))
{
float ave_time = invoker_ptr->Run(argument_ptr.get(), StreamConfig{nullptr, true});
std::size_t flop = 0, num_btype = 0;
for(std::size_t j = 0; j < gemm_descs.size(); ++j)
{
flop += std::size_t(2) * Ms[j] * Ns[j] * Ks[j];
num_btype += sizeof(ADataType) * Ms[j] * Ks[j] + sizeof(BDataType) * Ks[j] * Ns[j] +
sizeof(EDataType) * Ms[j] * Ns[j];
}
float tflops = static_cast<float>(flop) / 1.E9 / ave_time;
float gb_per_sec = num_btype / 1.E6 / ave_time;
std::cout << "Perf: " << std::setw(10) << ave_time << " ms, " << tflops << " TFlops, "
<< gb_per_sec << " GB/s, " << op_name << std::endl;
if(tflops > best_tflops)
{
found = true;
best_op_id = i;
best_op_name = op_name;
best_tflops = tflops;
best_ave_time = ave_time;
best_gb_per_sec = gb_per_sec;
}
}
else
{
std::cout << op_name << " does not support this problem" << std::endl;
}
}
std::cout << "Best Perf: " << best_ave_time << " ms, " << best_tflops << " TFlops, "
<< best_gb_per_sec << " GB/s, " << best_op_name << std::endl;
// run the best intance
if(found)
{
auto& op_ptr = op_ptrs[best_op_id];
std::cout << "Run the best instance without timing: " << op_ptr->GetTypeString()
<< std::endl;
auto argument_ptr = op_ptr->MakeArgumentPointer(
p_a, p_b, p_ds, p_e, gemm_descs, a_element_op, b_element_op, cde_element_op);
auto invoker_ptr = op_ptr->MakeInvokerPointer();
SimpleDeviceMem gemm_desc_workspace(op_ptr->GetWorkSpaceSize(argument_ptr.get()));
op_ptr->SetWorkSpacePointer(argument_ptr.get(), gemm_desc_workspace.GetDeviceBuffer());
if(op_ptr->IsSupportedArgument(argument_ptr.get()))
{
invoker_ptr->Run(argument_ptr.get(), StreamConfig{nullptr, false});
}
std::cout << "Done" << std::endl;
}
return 0;
}
docs/Doxyfile
View file @ bf5fe7b3
...@@ -775,8 +775,10 @@ WARN_LOGFILE = ...@@ -775,8 +775,10 @@ WARN_LOGFILE =
# spaces. See also FILE_PATTERNS and EXTENSION_MAPPING # spaces. See also FILE_PATTERNS and EXTENSION_MAPPING
# Note: If this tag is empty the current directory is searched. # Note: If this tag is empty the current directory is searched.
INPUT = ../library/include \ INPUT = ../include/ck/tensor_operation/gpu/grid \
../library/include/internal ../include/ck/tensor_operation/gpu/block \
../include/ck/tensor_operation/gpu/thread \
../library/include/ck/library/utility
# This tag can be used to specify the character encoding of the source files # This tag can be used to specify the character encoding of the source files
# that doxygen parses. Internally doxygen uses the UTF-8 encoding. Doxygen uses # that doxygen parses. Internally doxygen uses the UTF-8 encoding. Doxygen uses
...@@ -845,7 +847,7 @@ FILE_PATTERNS = *.c \ ...@@ -845,7 +847,7 @@ FILE_PATTERNS = *.c \
# be searched for input files as well. # be searched for input files as well.
# The default value is: NO. # The default value is: NO.
RECURSIVE = NO RECURSIVE = YES
# The EXCLUDE tag can be used to specify files and/or directories that should be # The EXCLUDE tag can be used to specify files and/or directories that should be
# excluded from the INPUT source files. This way you can easily exclude a # excluded from the INPUT source files. This way you can easily exclude a
......
docs/source/API_Reference_Guide.rst
View file @ bf5fe7b3
=================== *******************
API Reference Guide API Reference Guide
=================== *******************
------------ =================
Introduction Introduction
------------ =================
This document contains details of the APIs for the Composable Kernel (CK) library and introduces some of the key design This document contains details of the APIs for the Composable Kernel (CK) library and introduces some of the key design
principles that are used to write new classes that extend CK functionality. principles that are used to write new classes that extend CK functionality.
...@@ -16,8 +16,37 @@ Using CK API ...@@ -16,8 +16,37 @@ Using CK API
This section describes how to use the CK library API. This section describes how to use the CK library API.
----------------- =================
CK Datatypes CK Datatypes
=================
-----------------
DeviceMem
----------------- -----------------
[TODO] .. doxygenstruct:: DeviceMem
\ No newline at end of file
---------------------------
Kernels For Flashattention
---------------------------
The Flashattention algorithm is defined in :cite:t:`dao2022flashattention`. This sections lists the classes that are
used in the CK GPU implementation of Flashattention.
**Gridwise classes**
.. doxygenstruct:: ck::GridwiseBatchedGemmSoftmaxGemm_Xdl_CShuffle
**Blockwise classes**
.. doxygenstruct:: ck::ThreadGroupTensorSliceTransfer_v4r1
.. doxygenstruct:: ck::BlockwiseGemmXdlops_v2
.. doxygenstruct:: ck::BlockwiseSoftmax
**Threadwise classes**
.. doxygenstruct:: ck::ThreadwiseTensorSliceTransfer_StaticToStatic
.. bibliography::
\ No newline at end of file
docs/source/conf.py
View file @ bf5fe7b3
...@@ -59,10 +59,13 @@ if read_the_docs_build: ...@@ -59,10 +59,13 @@ if read_the_docs_build:
# Add any Sphinx extension module names here, as strings. They can be # Add any Sphinx extension module names here, as strings. They can be
# extensions coming with Sphinx (named 'sphinx.ext.*') or your custom # extensions coming with Sphinx (named 'sphinx.ext.*') or your custom
# ones. # ones.
extensions = ['sphinx.ext.mathjax', 'breathe'] extensions = ['sphinx.ext.mathjax', 'breathe', 'sphinxcontrib.bibtex']
breathe_projects = { "CK": "../docBin/xml" } breathe_projects = { "CK": "../docBin/xml" }
breathe_default_project = "CK" breathe_default_project = "CK"
bibtex_bibfiles = ['refs.bib']
# Add any paths that contain templates here, relative to this directory. # Add any paths that contain templates here, relative to this directory.
templates_path = ['_templates'] templates_path = ['_templates']
......
docs/source/refs.bib 0 → 100644
View file @ bf5fe7b3
@article{dao2022flashattention,
title={Flashattention: Fast and memory-efficient exact attention with io-awareness},
author={Dao, Tri and Fu, Daniel Y and Ermon, Stefano and Rudra, Atri and R{\'e}, Christopher},
journal={arXiv preprint arXiv:2205.14135},
year={2022}
}
include/ck/tensor_operation/gpu/block/blockwise_gemm_xdlops.hpp
View file @ bf5fe7b3
...@@ -622,11 +622,16 @@ constexpr auto BlockwiseGemmXdlops_k0mk1_k0nk1_m0n0m1n1m2m3m4n2_Selector() ...@@ -622,11 +622,16 @@ constexpr auto BlockwiseGemmXdlops_k0mk1_k0nk1_m0n0m1n1m2m3m4n2_Selector()
} }
}; };
// Blockwise gemm supporting /**
// 1. regular XDL output M2_M3_M4_M2 and transposed XDL output M2_N2_N3_N4 * @brief Blockwise gemm
// 2. decoupled input tile descriptor and mma tile descriptor in order to support both vgpr and LDS *
// source buffer * Supports
// 3. configurable k index starting position and step size after each FMA/XDL instruction * 1. regular XDL output M2_M3_M4_M2 and transposed XDL output M2_N2_N3_N4
* 2. decoupled input tile descriptor and mma tile descriptor in order to support both vgpr and LDS
* source buffer
* 3. configurable k index starting position and step size after each FMA/XDL instruction
*/
template <index_t BlockSize, template <index_t BlockSize,
typename FloatAB, typename FloatAB,
typename FloatAcc, typename FloatAcc,
......
include/ck/tensor_operation/gpu/block/blockwise_softmax.hpp
View file @ bf5fe7b3
...@@ -12,6 +12,16 @@ ...@@ -12,6 +12,16 @@
namespace ck { namespace ck {
/**
* @brief Blockwise softmax
*
* @tparam BlockSize Block size
* @tparam AccDataType Accumulator data type
* @tparam ThreadMap_M_K Thread id to m_k
* @tparam ThreadClusterDesc_M_K Threadwise cluster descriptor
* @tparam ThreadSliceDesc_M_K Threadwise slices descriptor
* @tparam IgnoreNaN Flag to ignore NaN, false by default
*/
template <index_t BlockSize, template <index_t BlockSize,
typename AccDataType, typename AccDataType,
typename ThreadMap_M_K, // thread_id to m_k typename ThreadMap_M_K, // thread_id to m_k
......
include/ck/tensor_operation/gpu/block/thread_group_tensor_slice_transfer_v4r1.hpp
View file @ bf5fe7b3
...@@ -11,10 +11,15 @@ ...@@ -11,10 +11,15 @@
namespace ck { namespace ck {
// this version does following things to avoid scratch memory issue /**
// 1. Use StaticallyIndexedArray instead of C array for thread buffer * @brief Blockwise data transfer
// 2. ThreadwiseTensorSliceTransfer_v3 does not keep reference to tensor descriptor *
// 3. ThreadwiseTensorSliceTransfer_v3::Run() does not construct new tensor coordinate * This version does following things to avoid scratch memory issue
* 1. Use StaticallyIndexedArray instead of C array for thread buffer
* 2. ThreadwiseTensorSliceTransfer_v3 does not keep reference to tensor descriptor
* 3. ThreadwiseTensorSliceTransfer_v3::Run() does not construct new tensor coordinate
*
*/
template <typename ThreadGroup, template <typename ThreadGroup,
typename SrcElementwiseOperation, typename SrcElementwiseOperation,
typename DstElementwiseOperation, typename DstElementwiseOperation,
......
include/ck/tensor_operation/gpu/device/impl/device_grouped_gemm_xdl.hpp
View file @ bf5fe7b3
...@@ -381,6 +381,9 @@ struct DeviceGroupedGemm_Xdl : public DeviceGroupedGemm<ALayout, ...@@ -381,6 +381,9 @@ struct DeviceGroupedGemm_Xdl : public DeviceGroupedGemm<ALayout,
const index_t N = gemm_descs[i].N_; const index_t N = gemm_descs[i].N_;
const index_t K = gemm_descs[i].K_; const index_t K = gemm_descs[i].K_;
a_mtx_mraw_kraw_.emplace_back(M, K);
b_mtx_nraw_kraw_.emplace_back(N, K);
if(M == 0) if(M == 0)
{ {
skipped_group_count_++; skipped_group_count_++;
...@@ -485,6 +488,8 @@ struct DeviceGroupedGemm_Xdl : public DeviceGroupedGemm<ALayout, ...@@ -485,6 +488,8 @@ struct DeviceGroupedGemm_Xdl : public DeviceGroupedGemm<ALayout,
CDEElementwiseOperation c_element_op_; CDEElementwiseOperation c_element_op_;
std::vector<GemmBiasTransKernelArg> gemm_desc_kernel_arg_; std::vector<GemmBiasTransKernelArg> gemm_desc_kernel_arg_;
std::vector<Tuple<index_t, index_t>> a_mtx_mraw_kraw_;
std::vector<Tuple<index_t, index_t>> b_mtx_nraw_kraw_;
index_t grid_size_; index_t grid_size_;
}; };
...@@ -599,7 +604,28 @@ struct DeviceGroupedGemm_Xdl : public DeviceGroupedGemm<ALayout, ...@@ -599,7 +604,28 @@ struct DeviceGroupedGemm_Xdl : public DeviceGroupedGemm<ALayout,
return false; return false;
} }
return true; bool supported = true;
// If we use padding we do not support vector loads for dimensions not divisible by vector
// load size.
if constexpr(GemmSpec != GemmSpecialization::Default)
{
// [A|B]BlockTransferSrcVectorDim value define dimension in the block {K0,M,K1} layout,
// thus we have to adapt it to the {M,K} or {N,K} layout.
const auto a_raw_vector_dim = ABlockTransferSrcVectorDim != 1 ? 1 : 0;
const auto b_raw_vector_dim = BBlockTransferSrcVectorDim != 1 ? 1 : 0;
for(index_t i = 0; i < arg.group_count_; ++i)
{
const auto a_vector_dim = arg.a_mtx_mraw_kraw_[i].At(Number<a_raw_vector_dim>{});
const auto b_vector_dim = arg.b_mtx_nraw_kraw_[i].At(Number<b_raw_vector_dim>{});
supported = supported & (a_vector_dim % ABlockTransferSrcScalarPerVector == 0);
supported = supported & (b_vector_dim % BBlockTransferSrcScalarPerVector == 0);
}
}
return supported;
} }
// polymorphic // polymorphic
...@@ -661,7 +687,8 @@ struct DeviceGroupedGemm_Xdl : public DeviceGroupedGemm<ALayout, ...@@ -661,7 +687,8 @@ struct DeviceGroupedGemm_Xdl : public DeviceGroupedGemm<ALayout,
<< MPerXDL << ", " << MPerXDL << ", "
<< NPerXDL << ", " << NPerXDL << ", "
<< MXdlPerWave << ", " << MXdlPerWave << ", "
<< NXdlPerWave << NXdlPerWave << ", "
<< getGemmSpecializationString(GemmSpec)
<< ">"; << ">";
// clang-format on // clang-format on
......
include/ck/tensor_operation/gpu/grid/gridwise_batched_gemm_softmax_gemm_xdl_cshuffle_v1.hpp
View file @ bf5fe7b3
...@@ -18,6 +18,10 @@ ...@@ -18,6 +18,10 @@
namespace ck { namespace ck {
/**
* @brief Gridwise gemm + softmax + gemm fusion
*
*/
template <typename FloatAB, template <typename FloatAB,
typename FloatGemmAcc, typename FloatGemmAcc,
typename FloatCShuffle, typename FloatCShuffle,
......
include/ck/tensor_operation/gpu/thread/threadwise_tensor_slice_transfer.hpp
View file @ bf5fe7b3
...@@ -1201,7 +1201,12 @@ struct ThreadwiseTensorSliceTransfer_v4 ...@@ -1201,7 +1201,12 @@ struct ThreadwiseTensorSliceTransfer_v4
SrcCoord src_ref_coord_; SrcCoord src_ref_coord_;
}; };
// Do NOT involve any tensor coordinates with StaticBuffer /**
* @brief Threadwise data transfer
*
* Do NOT involve any tensor coordinates with StaticBuffer
*
*/
template <typename SrcData, template <typename SrcData,
typename DstData, typename DstData,
typename SrcDesc, typename SrcDesc,
......
library/include/ck/library/tensor_operation_instance/device_operation_instance_factory.hpp
View file @ bf5fe7b3
...@@ -93,6 +93,7 @@ using AddReluAdd = ck::tensor_operation::element_wise::AddReluAdd; ...@@ -93,6 +93,7 @@ using AddReluAdd = ck::tensor_operation::element_wise::AddReluAdd;
using FastGelu = ck::tensor_operation::element_wise::FastGelu; using FastGelu = ck::tensor_operation::element_wise::FastGelu;
using AddMultiply = ck::tensor_operation::element_wise::AddMultiply; using AddMultiply = ck::tensor_operation::element_wise::AddMultiply;
using ScaleAdd = ck::tensor_operation::element_wise::ScaleAdd; using ScaleAdd = ck::tensor_operation::element_wise::ScaleAdd;
using Gelu = ck::tensor_operation::element_wise::Gelu;
template <typename Activation> template <typename Activation>
using Activation_Mul_Clamp = ck::tensor_operation::element_wise::Activation_Mul_Clamp<Activation>; using Activation_Mul_Clamp = ck::tensor_operation::element_wise::Activation_Mul_Clamp<Activation>;
......
library/include/ck/library/tensor_operation_instance/gpu/grouped_gemm.hpp
View file @ bf5fe7b3
...@@ -74,8 +74,7 @@ template <typename ALayout, ...@@ -74,8 +74,7 @@ template <typename ALayout,
typename ADataType, typename ADataType,
typename BDataType, typename BDataType,
typename EDataType> typename EDataType>
struct DeviceOperationInstanceFactory<ck::tensor_operation::device::DeviceGroupedGemm< struct DeviceOperationInstanceFactory<ck::tensor_operation::device::DeviceGroupedGemm<ALayout,
ALayout,
BLayout, BLayout,
Empty_Tuple, Empty_Tuple,
ELayout, ELayout,
...@@ -83,9 +82,9 @@ struct DeviceOperationInstanceFactory<ck::tensor_operation::device::DeviceGroupe ...@@ -83,9 +82,9 @@ struct DeviceOperationInstanceFactory<ck::tensor_operation::device::DeviceGroupe
BDataType, BDataType,
Empty_Tuple, Empty_Tuple,
EDataType, EDataType,
ck::tensor_operation::element_wise::PassThrough, PassThrough,
ck::tensor_operation::element_wise::PassThrough, PassThrough,
ck::tensor_operation::element_wise::PassThrough>> PassThrough>>
{ {
using DeviceOp = DeviceGroupedGemm<ALayout, using DeviceOp = DeviceGroupedGemm<ALayout,
BLayout, BLayout,
...@@ -95,9 +94,9 @@ struct DeviceOperationInstanceFactory<ck::tensor_operation::device::DeviceGroupe ...@@ -95,9 +94,9 @@ struct DeviceOperationInstanceFactory<ck::tensor_operation::device::DeviceGroupe
BDataType, BDataType,
Empty_Tuple, Empty_Tuple,
EDataType, EDataType,
ck::tensor_operation::element_wise::PassThrough, PassThrough,
ck::tensor_operation::element_wise::PassThrough, PassThrough,
ck::tensor_operation::element_wise::PassThrough>; PassThrough>;
static auto GetInstances() static auto GetInstances()
{ {
......
library/include/ck/library/tensor_operation_instance/gpu/grouped_gemm_fastgelu.hpp 0 → 100644
View file @ bf5fe7b3
// SPDX-License-Identifier: MIT
// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
#pragma once
#include <memory>
#include <vector>
#include "ck/ck.hpp"
#include "ck/library/tensor_operation_instance/device_operation_instance_factory.hpp"
#include "ck/tensor_operation/gpu/device/device_grouped_gemm.hpp"
namespace ck {
namespace tensor_operation {
namespace device {
namespace instance {
void add_device_grouped_gemm_fastgelu_xdl_f16_f16_f16_mk_kn_mn_instances(
std::vector<std::unique_ptr<DeviceGroupedGemm<Row,
Row,
Empty_Tuple,
Row,
F16,
F16,
Empty_Tuple,
F16,
PassThrough,
PassThrough,
FastGelu>>>& instances);
void add_device_grouped_gemm_fastgelu_xdl_f16_f16_f16_mk_nk_mn_instances(
std::vector<std::unique_ptr<DeviceGroupedGemm<Row,
Col,
Empty_Tuple,
Row,
F16,
F16,
Empty_Tuple,
F16,
PassThrough,
PassThrough,
FastGelu>>>& instances);
void add_device_grouped_gemm_fastgelu_xdl_f16_f16_f16_km_kn_mn_instances(
std::vector<std::unique_ptr<DeviceGroupedGemm<Col,
Row,
Empty_Tuple,
Row,
F16,
F16,
Empty_Tuple,
F16,
PassThrough,
PassThrough,
FastGelu>>>& instances);
void add_device_grouped_gemm_fastgelu_xdl_f16_f16_f16_km_nk_mn_instances(
std::vector<std::unique_ptr<DeviceGroupedGemm<Col,
Col,
Empty_Tuple,
Row,
F16,
F16,
Empty_Tuple,
F16,
PassThrough,
PassThrough,
FastGelu>>>& instances);
// GroupedGEMM + GELU
template <typename ALayout,
typename BLayout,
typename ELayout,
typename ADataType,
typename BDataType,
typename EDataType>
struct DeviceOperationInstanceFactory<ck::tensor_operation::device::DeviceGroupedGemm<ALayout,
BLayout,
Empty_Tuple,
ELayout,
ADataType,
BDataType,
Empty_Tuple,
EDataType,
PassThrough,
PassThrough,
FastGelu>>
{
using DeviceOp = DeviceGroupedGemm<ALayout,
BLayout,
Empty_Tuple,
ELayout,
ADataType,
BDataType,
Empty_Tuple,
EDataType,
PassThrough,
PassThrough,
FastGelu>;
static auto GetInstances()
{
std::vector<std::unique_ptr<DeviceOp>> op_ptrs;
if constexpr(is_same_v<ADataType, half_t> && is_same_v<BDataType, half_t> &&
is_same_v<EDataType, half_t>)
{
if constexpr(is_same_v<ALayout, Row> && is_same_v<BLayout, Row> &&
is_same_v<ELayout, Row>)
{
add_device_grouped_gemm_fastgelu_xdl_f16_f16_f16_mk_kn_mn_instances(op_ptrs);
}
else if constexpr(is_same_v<ALayout, Row> && is_same_v<BLayout, Col> &&
is_same_v<ELayout, Row>)
{
add_device_grouped_gemm_fastgelu_xdl_f16_f16_f16_mk_nk_mn_instances(op_ptrs);
}
else if constexpr(is_same_v<ALayout, Col> && is_same_v<BLayout, Row> &&
is_same_v<ELayout, Row>)
{
add_device_grouped_gemm_fastgelu_xdl_f16_f16_f16_km_kn_mn_instances(op_ptrs);
}
else if constexpr(is_same_v<ALayout, Col> && is_same_v<BLayout, Col> &&
is_same_v<ELayout, Row>)
{
add_device_grouped_gemm_fastgelu_xdl_f16_f16_f16_km_nk_mn_instances(op_ptrs);
}
}
return op_ptrs;
}
};
} // namespace instance
} // namespace device
} // namespace tensor_operation
} // namespace ck
library/include/ck/library/utility/device_memory.hpp
View file @ bf5fe7b3
...@@ -14,6 +14,10 @@ __global__ void set_buffer_value(T* p, T x, uint64_t buffer_element_size) ...@@ -14,6 +14,10 @@ __global__ void set_buffer_value(T* p, T x, uint64_t buffer_element_size)
} }
} }
/**
* @brief Container for storing data in GPU device memory
*
*/
struct DeviceMem struct DeviceMem
{ {
DeviceMem() = delete; DeviceMem() = delete;
......
library/include/ck/library/utility/fill.hpp
View file @ bf5fe7b3
...@@ -100,6 +100,15 @@ struct FillMonotonicSeq ...@@ -100,6 +100,15 @@ struct FillMonotonicSeq
return tmp; return tmp;
}); });
} }
template <typename ForwardRange>
auto operator()(ForwardRange&& range) const -> std::void_t<decltype(
std::declval<const FillMonotonicSeq&>()(std::begin(std::forward<ForwardRange>(range)),
std::end(std::forward<ForwardRange>(range))))>
{
(*this)(std::begin(std::forward<ForwardRange>(range)),
std::end(std::forward<ForwardRange>(range)));
}
}; };
template <typename T> template <typename T>
...@@ -112,6 +121,15 @@ struct FillConstant ...@@ -112,6 +121,15 @@ struct FillConstant
{ {
std::fill(first, last, value_); std::fill(first, last, value_);
} }
template <typename ForwardRange>
auto operator()(ForwardRange&& range) const -> std::void_t<
decltype(std::declval<const FillConstant&>()(std::begin(std::forward<ForwardRange>(range)),
std::end(std::forward<ForwardRange>(range))))>
{
(*this)(std::begin(std::forward<ForwardRange>(range)),
std::end(std::forward<ForwardRange>(range)));
}
}; };
} // namespace utils } // namespace utils
......
library/src/tensor_operation_instance/gpu/grouped_gemm_fastgelu/CMakeLists.txt 0 → 100644
View file @ bf5fe7b3
add_instance_library(device_grouped_gemm_fastgelu_instance
device_grouped_gemm_fastgelu_xdl_f16_f16_f16_mk_kn_mn_instance.cpp
device_grouped_gemm_fastgelu_xdl_f16_f16_f16_mk_nk_mn_instance.cpp
device_grouped_gemm_fastgelu_xdl_f16_f16_f16_km_kn_mn_instance.cpp
device_grouped_gemm_fastgelu_xdl_f16_f16_f16_km_nk_mn_instance.cpp
)
library/src/tensor_operation_instance/gpu/grouped_gemm_fastgelu/device_grouped_gemm_fastgelu_xdl_f16_f16_f16_km_kn_mn_instance.cpp 0 → 100644
View file @ bf5fe7b3
// SPDX-License-Identifier: MIT
// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
#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_grouped_gemm_xdl.hpp"
#include "ck/library/tensor_operation_instance/add_device_operation_instance.hpp"
namespace ck {
namespace tensor_operation {
namespace device {
namespace instance {
using F16 = ck::half_t;
using F32 = float;
using Row = ck::tensor_layout::gemm::RowMajor;
using Col = ck::tensor_layout::gemm::ColumnMajor;
template <ck::index_t... Is>
using S = ck::Sequence<Is...>;
using Empty_Tuple = ck::Tuple<>;
using PassThrough = ck::tensor_operation::element_wise::PassThrough;
using FastGelu = ck::tensor_operation::element_wise::FastGelu;
static constexpr auto GemmDefault = ck::tensor_operation::device::GemmSpecialization::Default;
// a[k, m] * b[k, n] = e[m, n]
using device_grouped_gemm_fastgelu_xdl_f16_f16_f16_km_kn_mn_instances = std::tuple<
// clang-format off
//###################| A| B| Ds| E| AData| BData| AccData| CShuffle| DsData| EData| A| B| C| GEMM| NumGemmK| Block| MPer| NPer| KPer| AK1| BK1| MPer| NPer| MXdl| NXdl| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockLds| BBlockTransfer| BBlockTransfer| BBlockTransfer| BlockTransfer| BBlockTransfer| BBlockTransfer| BBlockLds| CShuffle| CShuffle| CBlockTransferClusterLengths| CBlockTransfer|
//###################| Layout| Layout| Layout| Layout| Type| Type| Type| DataType| Type| Type| Elementwise| Elementwise| Elementwise| Spacialization| Prefetch| Size| Block| Block| Block| | | XDL| XDL| Per| Per| ThreadCluster| ThreadCluster| SrcAccessOrder| SrcVectorDim| SrcScalar| DstScalar| AddExtraM| ThreadCluster| ThreadCluster| SrcAccessOrder| SrcVectorDim| SrcScalar| DstScalar| AddExtraN| MXdlPerWave| NXdlPerWave| _MBlock_MWaveMPerXdl| ScalarPerVector|
//###################| | | | | | | | | | | Operation| Operation| Operation| | Stage| | | | | | | | | Wave| Wave| Lengths_K0_M_K1| ArrangeOrder| | | PerVector| PerVector_K1| | Lengths_K0_N_K1| ArrangeOrder| | | PerVector| PerVector_K1| | PerShuffle| PerShuffle| _NBlock_NWaveNPerXdl| _NWaveNPerXdl|
//###################| | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | |
DeviceGroupedGemm_Xdl< Col, Row, Empty_Tuple, Row, F16, F16, F32, F16, Empty_Tuple, F16, PassThrough, PassThrough, FastGelu, GemmDefault, 1, 256, 256, 128, 32, 2, 2, 32, 32, 4, 2, S<4, 64, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 2, 0, S<8, 32, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 2, 0, 1, 1, S<1, 32, 1, 8>, 8>,
DeviceGroupedGemm_Xdl< Col, Row, Empty_Tuple, Row, F16, F16, F32, F16, Empty_Tuple, F16, PassThrough, PassThrough, FastGelu, GemmDefault, 1, 256, 256, 128, 32, 8, 8, 32, 32, 4, 2, S<4, 64, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 8, 1, S<4, 64, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 2, 8, 1, 1, 1, S<1, 32, 1, 8>, 8>,
DeviceGroupedGemm_Xdl< Col, Row, Empty_Tuple, Row, F16, F16, F32, F16, Empty_Tuple, F16, PassThrough, PassThrough, FastGelu, GemmDefault, 1, 256, 128, 256, 32, 2, 2, 32, 32, 2, 4, S<8, 32, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 2, 0, S<4, 64, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 2, 0, 1, 1, S<1, 32, 1, 8>, 8>,
DeviceGroupedGemm_Xdl< Col, Row, Empty_Tuple, Row, F16, F16, F32, F16, Empty_Tuple, F16, PassThrough, PassThrough, FastGelu, GemmDefault, 1, 256, 128, 256, 32, 8, 8, 32, 32, 2, 4, S<4, 64, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 2, 8, 1, S<4, 64, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 8, 1, 1, 1, S<1, 32, 1, 8>, 8>,
DeviceGroupedGemm_Xdl< Col, Row, Empty_Tuple, Row, F16, F16, F32, F16, Empty_Tuple, F16, PassThrough, PassThrough, FastGelu, GemmDefault, 1, 128, 128, 128, 32, 2, 2, 32, 32, 4, 2, S<4, 32, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 2, 0, S<4, 32, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 2, 0, 1, 1, S<1, 16, 1, 8>, 8>,
DeviceGroupedGemm_Xdl< Col, Row, Empty_Tuple, Row, F16, F16, F32, F16, Empty_Tuple, F16, PassThrough, PassThrough, FastGelu, GemmDefault, 1, 128, 128, 128, 32, 8, 8, 32, 32, 4, 2, S<4, 32, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 8, 1, S<4, 32, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 8, 1, 1, 1, S<1, 16, 1, 8>, 8>,
DeviceGroupedGemm_Xdl< Col, Row, Empty_Tuple, Row, F16, F16, F32, F16, Empty_Tuple, F16, PassThrough, PassThrough, FastGelu, GemmDefault, 1, 256, 128, 128, 32, 2, 2, 32, 32, 2, 2, S<8, 32, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 2, 0, S<8, 32, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 2, 0, 1, 1, S<1, 32, 1, 8>, 8>,
DeviceGroupedGemm_Xdl< Col, Row, Empty_Tuple, Row, F16, F16, F32, F16, Empty_Tuple, F16, PassThrough, PassThrough, FastGelu, GemmDefault, 1, 256, 128, 128, 32, 8, 8, 32, 32, 2, 2, S<4, 64, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 2, 8, 1, S<4, 64, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 2, 8, 1, 1, 1, S<1, 32, 1, 8>, 8>,
DeviceGroupedGemm_Xdl< Col, Row, Empty_Tuple, Row, F16, F16, F32, F16, Empty_Tuple, F16, PassThrough, PassThrough, FastGelu, GemmDefault, 1, 128, 128, 64, 32, 2, 2, 32, 32, 2, 2, S<4, 32, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 2, 0, S<4, 16, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 2, 0, 1, 1, S<1, 32, 1, 4>, 8>,
DeviceGroupedGemm_Xdl< Col, Row, Empty_Tuple, Row, F16, F16, F32, F16, Empty_Tuple, F16, PassThrough, PassThrough, FastGelu, GemmDefault, 1, 128, 128, 64, 32, 8, 8, 32, 32, 2, 2, S<4, 32, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 8, 1, S<4, 32, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 2, 8, 1, 1, 1, S<1, 32, 1, 4>, 8>,
DeviceGroupedGemm_Xdl< Col, Row, Empty_Tuple, Row, F16, F16, F32, F16, Empty_Tuple, F16, PassThrough, PassThrough, FastGelu, GemmDefault, 1, 128, 64, 128, 32, 2, 2, 32, 32, 2, 2, S<8, 16, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 2, 0, S<4, 32, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 2, 0, 1, 1, S<1, 16, 1, 8>, 8>,
DeviceGroupedGemm_Xdl< Col, Row, Empty_Tuple, Row, F16, F16, F32, F16, Empty_Tuple, F16, PassThrough, PassThrough, FastGelu, GemmDefault, 1, 128, 64, 128, 32, 8, 8, 32, 32, 2, 2, S<4, 32, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 2, 8, 1, S<4, 32, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 8, 1, 1, 1, S<1, 16, 1, 8>, 8>,
DeviceGroupedGemm_Xdl< Col, Row, Empty_Tuple, Row, F16, F16, F32, F16, Empty_Tuple, F16, PassThrough, PassThrough, FastGelu, GemmDefault, 1, 256, 128, 64, 32, 2, 2, 32, 32, 2, 1, S<8, 32, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 2, 0, S<16,16, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 2, 0, 1, 1, S<1, 32, 1, 8>, 8>,
DeviceGroupedGemm_Xdl< Col, Row, Empty_Tuple, Row, F16, F16, F32, F16, Empty_Tuple, F16, PassThrough, PassThrough, FastGelu, GemmDefault, 1, 256, 128, 64, 32, 8, 8, 32, 32, 2, 1, S<4, 64, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 2, 8, 1, S<4, 64, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 1, 8, 1, 1, 1, S<1, 32, 1, 8>, 8>,
DeviceGroupedGemm_Xdl< Col, Row, Empty_Tuple, Row, F16, F16, F32, F16, Empty_Tuple, F16, PassThrough, PassThrough, FastGelu, GemmDefault, 1, 256, 64, 128, 32, 2, 2, 32, 32, 1, 2, S<16,16, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 2, 0, S<8, 32, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 2, 0, 1, 1, S<1, 32, 1, 8>, 8>,
DeviceGroupedGemm_Xdl< Col, Row, Empty_Tuple, Row, F16, F16, F32, F16, Empty_Tuple, F16, PassThrough, PassThrough, FastGelu, GemmDefault, 1, 256, 64, 128, 32, 8, 8, 32, 32, 1, 2, S<4, 64, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 1, 8, 1, S<4, 64, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 2, 8, 1, 1, 1, S<1, 32, 1, 8>, 8>
// clang-format on
>;
static constexpr auto GemmMNKPadding = ck::tensor_operation::device::GemmSpecialization::MNKPadding;
using device_grouped_gemm_fastgelu_xdl_f16_f16_f16_km_kn_mn_irregular_tile_instances = std::tuple<
// clang-format off
//###################| A| B| Ds| E| AData| BData| AccData| CShuffle| DsData| EData| A| B| C| GEMM| NumGemmK| Block| MPer| NPer| KPer| AK1| BK1| MPer| NPer| MXdl| NXdl| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockLds| BBlockTransfer| BBlockTransfer| BBlockTransfer| BlockTransfer| BBlockTransfer| BBlockTransfer| BBlockLds| CShuffle| CShuffle| CBlockTransferClusterLengths| CBlockTransfer|
//###################| Layout| Layout| Layout| Layout| Type| Type| Type| DataType| Type| Type| Elementwise| Elementwise| Elementwise| Spacialization| Prefetch| Size| Block| Block| Block| | | XDL| XDL| Per| Per| ThreadCluster| ThreadCluster| SrcAccessOrder| SrcVectorDim| SrcScalar| DstScalar| AddExtraM| ThreadCluster| ThreadCluster| SrcAccessOrder| SrcVectorDim| SrcScalar| DstScalar| AddExtraN| MXdlPerWave| NXdlPerWave| _MBlock_MWaveMPerXdl| ScalarPerVector|
//###################| | | | | | | | | | | Operation| Operation| Operation| | Stage| | | | | | | | | Wave| Wave| Lengths_K0_M_K1| ArrangeOrder| | | PerVector| PerVector_K1| | Lengths_K0_N_K1| ArrangeOrder| | | PerVector| PerVector_K1| | PerShuffle| PerShuffle| _NBlock_NWaveNPerXdl| _NWaveNPerXdl|
//###################| | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | |
DeviceGroupedGemm_Xdl< Col, Row, Empty_Tuple, Row, F16, F16, F32, F16, Empty_Tuple, F16, PassThrough, PassThrough, FastGelu, GemmMNKPadding, 1, 64, 16, 16, 32, 8, 8, 16, 16, 1, 1, S<4, 16, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 1, 8, 1, S<4, 16, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 1, 8, 1, 1, 1, S<1, 16, 1, 4>, 1>,
DeviceGroupedGemm_Xdl< Col, Row, Empty_Tuple, Row, F16, F16, F32, F16, Empty_Tuple, F16, PassThrough, PassThrough, FastGelu, GemmMNKPadding, 1, 256, 16, 64, 32, 8, 8, 16, 16, 1, 1, S<4, 16, 4>, S<0, 2, 1>, S<0, 2, 1>, 1, 1, 2, 1, S<4, 64, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 1, 8, 1, 1, 1, S<1, 16, 1, 4>, 1>,
DeviceGroupedGemm_Xdl< Col, Row, Empty_Tuple, Row, F16, F16, F32, F16, Empty_Tuple, F16, PassThrough, PassThrough, FastGelu, GemmMNKPadding, 1, 128, 128, 64, 32, 2, 2, 32, 32, 2, 2, S<4, 32, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 2, 0, S<4, 16, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 2, 0, 1, 1, S<1, 32, 1, 4>, 8>,
DeviceGroupedGemm_Xdl< Col, Row, Empty_Tuple, Row, F16, F16, F32, F16, Empty_Tuple, F16, PassThrough, PassThrough, FastGelu, GemmMNKPadding, 1, 128, 128, 64, 32, 8, 8, 32, 32, 2, 2, S<4, 32, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 8, 1, S<4, 32, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 2, 8, 1, 1, 1, S<1, 32, 1, 4>, 8>,
DeviceGroupedGemm_Xdl< Col, Row, Empty_Tuple, Row, F16, F16, F32, F16, Empty_Tuple, F16, PassThrough, PassThrough, FastGelu, GemmMNKPadding, 1, 128, 64, 128, 32, 2, 2, 32, 32, 2, 2, S<8, 16, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 2, 0, S<4, 32, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 2, 0, 1, 1, S<1, 16, 1, 8>, 8>,
DeviceGroupedGemm_Xdl< Col, Row, Empty_Tuple, Row, F16, F16, F32, F16, Empty_Tuple, F16, PassThrough, PassThrough, FastGelu, GemmMNKPadding, 1, 128, 64, 128, 32, 8, 8, 32, 32, 2, 2, S<4, 32, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 2, 8, 1, S<4, 32, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 8, 1, 1, 1, S<1, 16, 1, 8>, 8>,
DeviceGroupedGemm_Xdl< Col, Row, Empty_Tuple, Row, F16, F16, F32, F16, Empty_Tuple, F16, PassThrough, PassThrough, FastGelu, GemmMNKPadding, 1, 256, 128, 64, 32, 2, 2, 32, 32, 2, 1, S<8, 32, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 2, 0, S<16,16, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 2, 0, 1, 1, S<1, 32, 1, 8>, 8>,
DeviceGroupedGemm_Xdl< Col, Row, Empty_Tuple, Row, F16, F16, F32, F16, Empty_Tuple, F16, PassThrough, PassThrough, FastGelu, GemmMNKPadding, 1, 256, 128, 64, 32, 8, 8, 32, 32, 2, 1, S<4, 64, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 2, 8, 1, S<4, 64, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 1, 8, 1, 1, 1, S<1, 32, 1, 8>, 8>,
DeviceGroupedGemm_Xdl< Col, Row, Empty_Tuple, Row, F16, F16, F32, F16, Empty_Tuple, F16, PassThrough, PassThrough, FastGelu, GemmMNKPadding, 1, 256, 64, 128, 32, 2, 2, 32, 32, 1, 2, S<16,16, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 2, 0, S<8, 32, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 2, 0, 1, 1, S<1, 32, 1, 8>, 8>,
DeviceGroupedGemm_Xdl< Col, Row, Empty_Tuple, Row, F16, F16, F32, F16, Empty_Tuple, F16, PassThrough, PassThrough, FastGelu, GemmMNKPadding, 1, 256, 64, 128, 32, 8, 8, 32, 32, 1, 2, S<4, 64, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 1, 8, 1, S<4, 64, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 2, 8, 1, 1, 1, S<1, 32, 1, 8>, 8>
// clang-format on
>;
void add_device_grouped_gemm_fastgelu_xdl_f16_f16_f16_km_kn_mn_instances(
std::vector<std::unique_ptr<DeviceGroupedGemm<Col,
Row,
Empty_Tuple,
Row,
F16,
F16,
Empty_Tuple,
F16,
PassThrough,
PassThrough,
FastGelu>>>& instances)
{
add_device_operation_instances(
instances, device_grouped_gemm_fastgelu_xdl_f16_f16_f16_km_kn_mn_instances{});
add_device_operation_instances(
instances,
device_grouped_gemm_fastgelu_xdl_f16_f16_f16_km_kn_mn_irregular_tile_instances{});
}
} // namespace instance
} // namespace device
} // namespace tensor_operation
} // namespace ck
library/src/tensor_operation_instance/gpu/grouped_gemm_fastgelu/device_grouped_gemm_fastgelu_xdl_f16_f16_f16_km_nk_mn_instance.cpp 0 → 100644
View file @ bf5fe7b3
// SPDX-License-Identifier: MIT
// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
#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_grouped_gemm_xdl.hpp"
#include "ck/library/tensor_operation_instance/add_device_operation_instance.hpp"
namespace ck {
namespace tensor_operation {
namespace device {
namespace instance {
using F16 = ck::half_t;
using F32 = float;
using Row = ck::tensor_layout::gemm::RowMajor;
using Col = ck::tensor_layout::gemm::ColumnMajor;
template <ck::index_t... Is>
using S = ck::Sequence<Is...>;
using Empty_Tuple = ck::Tuple<>;
using PassThrough = ck::tensor_operation::element_wise::PassThrough;
using FastGelu = ck::tensor_operation::element_wise::FastGelu;
static constexpr auto GemmDefault = ck::tensor_operation::device::GemmSpecialization::Default;
// a[k, m] * b[n, k] = e[m, n]
using device_grouped_gemm_fastgelu_xdl_f16_f16_f16_km_nk_mn_instances = std::tuple<
// clang-format off
//###################| A| B| Ds| E| AData| BData| AccData| CShuffle| DsData| EData| A| B| C| GEMM| NumGemmK| Block| MPer| NPer| KPer| AK1| BK1| MPer| NPer| MXdl| NXdl| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockLds| BBlockTransfer| BBlockTransfer| BBlockTransfer| BlockTransfer| BBlockTransfer| BBlockTransfer| BBlockLds| CShuffle| CShuffle| CBlockTransferClusterLengths| CBlockTransfer|
//###################| Layout| Layout| Layout| Layout| Type| Type| Type| DataType| Type| Type| Elementwise| Elementwise| Elementwise| Spacialization| Prefetch| Size| Block| Block| Block| | | XDL| XDL| Per| Per| ThreadCluster| ThreadCluster| SrcAccessOrder| SrcVectorDim| SrcScalar| DstScalar| AddExtraM| ThreadCluster| ThreadCluster| SrcAccessOrder| SrcVectorDim| SrcScalar| DstScalar| AddExtraN| MXdlPerWave| NXdlPerWave| _MBlock_MWaveMPerXdl| ScalarPerVector|
//###################| | | | | | | | | | | Operation| Operation| Operation| | Stage| | | | | | | | | Wave| Wave| Lengths_K0_M_K1| ArrangeOrder| | | PerVector| PerVector_K1| | Lengths_K0_N_K1| ArrangeOrder| | | PerVector| PerVector_K1| | PerShuffle| PerShuffle| _NBlock_NWaveNPerXdl| _NWaveNPerXdl|
//###################| | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | |
DeviceGroupedGemm_Xdl< Col, Col, Empty_Tuple, Row, F16, F16, F32, F16, Empty_Tuple, F16, PassThrough, PassThrough, FastGelu, GemmDefault, 1, 256, 256, 128, 32, 2, 8, 32, 32, 4, 2, S<4, 64, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 2, 0, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, 1, 1, S<1, 32, 1, 8>, 8>,
DeviceGroupedGemm_Xdl< Col, Col, Empty_Tuple, Row, F16, F16, F32, F16, Empty_Tuple, F16, PassThrough, PassThrough, FastGelu, GemmDefault, 1, 256, 256, 128, 32, 8, 8, 32, 32, 4, 2, S<4, 64, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 8, 1, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, 1, 1, S<1, 32, 1, 8>, 8>,
DeviceGroupedGemm_Xdl< Col, Col, Empty_Tuple, Row, F16, F16, F32, F16, Empty_Tuple, F16, PassThrough, PassThrough, FastGelu, GemmDefault, 1, 256, 128, 256, 32, 2, 8, 32, 32, 2, 4, S<8, 32, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 2, 0, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, 1, 1, S<1, 32, 1, 8>, 8>,
DeviceGroupedGemm_Xdl< Col, Col, Empty_Tuple, Row, F16, F16, F32, F16, Empty_Tuple, F16, PassThrough, PassThrough, FastGelu, GemmDefault, 1, 256, 128, 256, 32, 8, 8, 32, 32, 2, 4, S<4, 64, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 2, 8, 1, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, 1, 1, S<1, 32, 1, 8>, 8>,
DeviceGroupedGemm_Xdl< Col, Col, Empty_Tuple, Row, F16, F16, F32, F16, Empty_Tuple, F16, PassThrough, PassThrough, FastGelu, GemmDefault, 1, 128, 128, 128, 32, 2, 8, 32, 32, 4, 2, S<4, 32, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 2, 0, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, 1, 1, S<1, 16, 1, 8>, 8>,
DeviceGroupedGemm_Xdl< Col, Col, Empty_Tuple, Row, F16, F16, F32, F16, Empty_Tuple, F16, PassThrough, PassThrough, FastGelu, GemmDefault, 1, 128, 128, 128, 32, 8, 8, 32, 32, 4, 2, S<4, 32, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 8, 1, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, 1, 1, S<1, 16, 1, 8>, 8>,
DeviceGroupedGemm_Xdl< Col, Col, Empty_Tuple, Row, F16, F16, F32, F16, Empty_Tuple, F16, PassThrough, PassThrough, FastGelu, GemmDefault, 1, 256, 128, 128, 32, 2, 8, 32, 32, 2, 2, S<8, 32, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 2, 0, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, 1, 1, S<1, 32, 1, 8>, 8>,
DeviceGroupedGemm_Xdl< Col, Col, Empty_Tuple, Row, F16, F16, F32, F16, Empty_Tuple, F16, PassThrough, PassThrough, FastGelu, GemmDefault, 1, 256, 128, 128, 32, 8, 8, 32, 32, 2, 2, S<4, 64, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 2, 8, 1, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, 1, 1, S<1, 32, 1, 8>, 8>,
DeviceGroupedGemm_Xdl< Col, Col, Empty_Tuple, Row, F16, F16, F32, F16, Empty_Tuple, F16, PassThrough, PassThrough, FastGelu, GemmDefault, 1, 128, 128, 64, 32, 2, 8, 32, 32, 2, 2, S<4, 32, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 2, 0, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, 1, 1, S<1, 32, 1, 4>, 8>,
DeviceGroupedGemm_Xdl< Col, Col, Empty_Tuple, Row, F16, F16, F32, F16, Empty_Tuple, F16, PassThrough, PassThrough, FastGelu, GemmDefault, 1, 128, 128, 64, 32, 8, 8, 32, 32, 2, 2, S<4, 32, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 8, 1, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, 1, 1, S<1, 32, 1, 4>, 8>,
DeviceGroupedGemm_Xdl< Col, Col, Empty_Tuple, Row, F16, F16, F32, F16, Empty_Tuple, F16, PassThrough, PassThrough, FastGelu, GemmDefault, 1, 128, 64, 128, 32, 2, 8, 32, 32, 2, 2, S<8, 16, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 2, 0, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, 1, 1, S<1, 16, 1, 8>, 8>,
DeviceGroupedGemm_Xdl< Col, Col, Empty_Tuple, Row, F16, F16, F32, F16, Empty_Tuple, F16, PassThrough, PassThrough, FastGelu, GemmDefault, 1, 128, 64, 128, 32, 8, 8, 32, 32, 2, 2, S<4, 32, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 2, 8, 1, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, 1, 1, S<1, 16, 1, 8>, 8>,
DeviceGroupedGemm_Xdl< Col, Col, Empty_Tuple, Row, F16, F16, F32, F16, Empty_Tuple, F16, PassThrough, PassThrough, FastGelu, GemmDefault, 1, 256, 128, 64, 32, 2, 8, 32, 32, 2, 1, S<8, 32, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 2, 0, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, 1, 1, S<1, 32, 1, 8>, 8>,
DeviceGroupedGemm_Xdl< Col, Col, Empty_Tuple, Row, F16, F16, F32, F16, Empty_Tuple, F16, PassThrough, PassThrough, FastGelu, GemmDefault, 1, 256, 128, 64, 32, 8, 8, 32, 32, 2, 1, S<4, 64, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 2, 8, 1, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, 1, 1, S<1, 32, 1, 8>, 8>,
DeviceGroupedGemm_Xdl< Col, Col, Empty_Tuple, Row, F16, F16, F32, F16, Empty_Tuple, F16, PassThrough, PassThrough, FastGelu, GemmDefault, 1, 256, 64, 128, 32, 2, 8, 32, 32, 1, 2, S<16,16, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 2, 0, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, 1, 1, S<1, 32, 1, 8>, 8>,
DeviceGroupedGemm_Xdl< Col, Col, Empty_Tuple, Row, F16, F16, F32, F16, Empty_Tuple, F16, PassThrough, PassThrough, FastGelu, GemmDefault, 1, 256, 64, 128, 32, 8, 8, 32, 32, 1, 2, S<4, 64, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 1, 8, 1, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, 1, 1, S<1, 32, 1, 8>, 8>
// clang-format on
>;
static constexpr auto GemmMNKPadding = ck::tensor_operation::device::GemmSpecialization::MNKPadding;
using device_grouped_gemm_fastgelu_xdl_f16_f16_f16_km_nk_mn_irregular_tile_instances = std::tuple<
// clang-format off
//###################| A| B| Ds| E| AData| BData| AccData| CShuffle| DsData| EData| A| B| C| GEMM| NumGemmK| Block| MPer| NPer| KPer| AK1| BK1| MPer| NPer| MXdl| NXdl| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockLds| BBlockTransfer| BBlockTransfer| BBlockTransfer| BlockTransfer| BBlockTransfer| BBlockTransfer| BBlockLds| CShuffle| CShuffle| CBlockTransferClusterLengths| CBlockTransfer|
//###################| Layout| Layout| Layout| Layout| Type| Type| Type| DataType| Type| Type| Elementwise| Elementwise| Elementwise| Spacialization| Prefetch| Size| Block| Block| Block| | | XDL| XDL| Per| Per| ThreadCluster| ThreadCluster| SrcAccessOrder| SrcVectorDim| SrcScalar| DstScalar| AddExtraM| ThreadCluster| ThreadCluster| SrcAccessOrder| SrcVectorDim| SrcScalar| DstScalar| AddExtraN| MXdlPerWave| NXdlPerWave| _MBlock_MWaveMPerXdl| ScalarPerVector|
//###################| | | | | | | | | | | Operation| Operation| Operation| | Stage| | | | | | | | | Wave| Wave| Lengths_K0_M_K1| ArrangeOrder| | | PerVector| PerVector_K1| | Lengths_K0_N_K1| ArrangeOrder| | | PerVector| PerVector_K1| | PerShuffle| PerShuffle| _NBlock_NWaveNPerXdl| _NWaveNPerXdl|
//###################| | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | |
DeviceGroupedGemm_Xdl< Col, Col, Empty_Tuple, Row, F16, F16, F32, F16, Empty_Tuple, F16, PassThrough, PassThrough, FastGelu, GemmMNKPadding, 1, 64, 16, 16, 32, 8, 8, 16, 16, 1, 1, S<4, 16, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 1, 8, 1, S<4, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 1, 8, 1, 1, 1, S<1, 16, 1, 4>, 1>,
DeviceGroupedGemm_Xdl< Col, Col, Empty_Tuple, Row, F16, F16, F32, F16, Empty_Tuple, F16, PassThrough, PassThrough, FastGelu, GemmMNKPadding, 1, 256, 16, 64, 32, 8, 8, 16, 16, 1, 1, S<4, 16, 4>, S<0, 2, 1>, S<0, 2, 1>, 1, 1, 2, 1, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 1, 8, 1, 1, 1, S<1, 16, 1, 4>, 1>,
DeviceGroupedGemm_Xdl< Col, Col, Empty_Tuple, Row, F16, F16, F32, F16, Empty_Tuple, F16, PassThrough, PassThrough, FastGelu, GemmMNKPadding, 1, 128, 128, 64, 32, 2, 8, 32, 32, 2, 2, S<4, 32, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 2, 0, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, 1, 1, S<1, 32, 1, 4>, 8>,
DeviceGroupedGemm_Xdl< Col, Col, Empty_Tuple, Row, F16, F16, F32, F16, Empty_Tuple, F16, PassThrough, PassThrough, FastGelu, GemmMNKPadding, 1, 128, 128, 64, 32, 8, 8, 32, 32, 2, 2, S<4, 32, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 8, 1, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, 1, 1, S<1, 32, 1, 4>, 8>,
DeviceGroupedGemm_Xdl< Col, Col, Empty_Tuple, Row, F16, F16, F32, F16, Empty_Tuple, F16, PassThrough, PassThrough, FastGelu, GemmMNKPadding, 1, 128, 64, 128, 32, 2, 8, 32, 32, 2, 2, S<8, 16, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 2, 0, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, 1, 1, S<1, 16, 1, 8>, 8>,
DeviceGroupedGemm_Xdl< Col, Col, Empty_Tuple, Row, F16, F16, F32, F16, Empty_Tuple, F16, PassThrough, PassThrough, FastGelu, GemmMNKPadding, 1, 128, 64, 128, 32, 8, 8, 32, 32, 2, 2, S<4, 32, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 2, 8, 1, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, 1, 1, S<1, 16, 1, 8>, 8>,
DeviceGroupedGemm_Xdl< Col, Col, Empty_Tuple, Row, F16, F16, F32, F16, Empty_Tuple, F16, PassThrough, PassThrough, FastGelu, GemmMNKPadding, 1, 256, 128, 64, 32, 2, 8, 32, 32, 2, 1, S<8, 32, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 2, 0, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, 1, 1, S<1, 32, 1, 8>, 8>,
DeviceGroupedGemm_Xdl< Col, Col, Empty_Tuple, Row, F16, F16, F32, F16, Empty_Tuple, F16, PassThrough, PassThrough, FastGelu, GemmMNKPadding, 1, 256, 128, 64, 32, 8, 8, 32, 32, 2, 1, S<4, 64, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 2, 8, 1, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, 1, 1, S<1, 32, 1, 8>, 8>,
DeviceGroupedGemm_Xdl< Col, Col, Empty_Tuple, Row, F16, F16, F32, F16, Empty_Tuple, F16, PassThrough, PassThrough, FastGelu, GemmMNKPadding, 1, 256, 64, 128, 32, 2, 8, 32, 32, 1, 2, S<16,16, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 2, 0, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, 1, 1, S<1, 32, 1, 8>, 8>,
DeviceGroupedGemm_Xdl< Col, Col, Empty_Tuple, Row, F16, F16, F32, F16, Empty_Tuple, F16, PassThrough, PassThrough, FastGelu, GemmMNKPadding, 1, 256, 64, 128, 32, 8, 8, 32, 32, 1, 2, S<4, 64, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 1, 8, 1, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, 1, 1, S<1, 32, 1, 8>, 8>
// clang-format on
>;
void add_device_grouped_gemm_fastgelu_xdl_f16_f16_f16_km_nk_mn_instances(
std::vector<std::unique_ptr<DeviceGroupedGemm<Col,
Col,
Empty_Tuple,
Row,
F16,
F16,
Empty_Tuple,
F16,
PassThrough,
PassThrough,
FastGelu>>>& instances)
{
add_device_operation_instances(
instances, device_grouped_gemm_fastgelu_xdl_f16_f16_f16_km_nk_mn_instances{});
add_device_operation_instances(
instances,
device_grouped_gemm_fastgelu_xdl_f16_f16_f16_km_nk_mn_irregular_tile_instances{});
}
} // namespace instance
} // namespace device
} // namespace tensor_operation
} // namespace ck
Markdown is supported
0% or .
You are about to add 0 people to the discussion. Proceed with caution.
Finish editing this message first!
Please register or to comment