Skip to content

GitLab

Unverified Commit bf80ceee authored by guangzlu's avatar guangzlu Committed by GitHub
Browse files

Merge branch 'attn-fwd-dropout-verify' into attn-bwd-dropout

parents 9dc05eaa a2887630
Hide whitespace changes
Inline Side-by-side
Showing with 1256 additions and 1 deletion
example/32_batched_gemm_scale_softmax_gemm/grouped_multihead_attention_forward_fp16.cpp 100755 → 100644
View file @ bf80ceee
...@@ -33,6 +33,7 @@ using S = ck::Sequence<Is...>; ...@@ -33,6 +33,7 @@ using S = ck::Sequence<Is...>;
using F16 = ck::half_t; using F16 = ck::half_t;
using F32 = float; using F32 = float;
using U16 = unsigned short;
using PassThrough = ck::tensor_operation::element_wise::PassThrough; using PassThrough = ck::tensor_operation::element_wise::PassThrough;
...@@ -42,6 +43,7 @@ using B1DataType = F16; ...@@ -42,6 +43,7 @@ using B1DataType = F16;
using AccDataType = F32; using AccDataType = F32;
using CShuffleDataType = F32; using CShuffleDataType = F32;
using CDataType = F16; using CDataType = F16;
using ZDataType = U16;
using LSEDataType = F32; using LSEDataType = F32;
using Acc0BiasDataType = ck::Tuple<>; using Acc0BiasDataType = ck::Tuple<>;
using Acc1BiasDataType = ck::Tuple<>; using Acc1BiasDataType = ck::Tuple<>;
...@@ -69,6 +71,7 @@ static constexpr auto TensorSpecC = ck::tensor_operation::device::TensorSpecial ...@@ -69,6 +71,7 @@ static constexpr auto TensorSpecC = ck::tensor_operation::device::TensorSpecial
using DeviceGemmInstance = using DeviceGemmInstance =
ck::tensor_operation::device::DeviceGroupedMultiheadAttentionForward_Xdl_CShuffle< ck::tensor_operation::device::DeviceGroupedMultiheadAttentionForward_Xdl_CShuffle<
NumDimG, NumDimG,
NumDimM, NumDimM,
NumDimN, NumDimN,
...@@ -78,6 +81,7 @@ using DeviceGemmInstance = ...@@ -78,6 +81,7 @@ using DeviceGemmInstance =
B0DataType, B0DataType,
B1DataType, B1DataType,
CDataType, CDataType,
ZDataType,
LSEDataType, LSEDataType,
Acc0BiasDataType, Acc0BiasDataType,
Acc1BiasDataType, Acc1BiasDataType,
...@@ -159,4 +163,5 @@ using ReferenceGemm1Instance = ck::tensor_operation::host::ReferenceBatchedGemm< ...@@ -159,4 +163,5 @@ using ReferenceGemm1Instance = ck::tensor_operation::host::ReferenceBatchedGemm<
#include "run_grouped_multihead_attention_forward.inc" #include "run_grouped_multihead_attention_forward.inc"
int main(int argc, char* argv[]) { return run(argc, argv); } int main(int argc, char* argv[]) { return run(argc, argv); }
example/32_batched_gemm_scale_softmax_gemm/run_grouped_multihead_attention_forward.inc
View file @ bf80ceee
...@@ -48,6 +48,7 @@ int run(int argc, char* argv[]) ...@@ -48,6 +48,7 @@ int run(int argc, char* argv[])
std::vector<const void*> p_b0; std::vector<const void*> p_b0;
std::vector<const void*> p_b1; std::vector<const void*> p_b1;
std::vector<void*> p_c; std::vector<void*> p_c;
std::vector<void*> p_z;
std::vector<void*> p_lse; std::vector<void*> p_lse;
std::vector<std::vector<int>> g0_g1_m_n_k_o; std::vector<std::vector<int>> g0_g1_m_n_k_o;
...@@ -55,6 +56,7 @@ int run(int argc, char* argv[]) ...@@ -55,6 +56,7 @@ int run(int argc, char* argv[])
std::vector<Tensor<B0DataType>> b0_tensors; std::vector<Tensor<B0DataType>> b0_tensors;
std::vector<Tensor<B1DataType>> b1_tensors; std::vector<Tensor<B1DataType>> b1_tensors;
std::vector<Tensor<CDataType>> c_tensors; std::vector<Tensor<CDataType>> c_tensors;
std::vector<Tensor<ZDataType>> z_tensors;
std::vector<Tensor<LSEDataType>> lse_tensors; std::vector<Tensor<LSEDataType>> lse_tensors;
using DeviceMemPtr = std::unique_ptr<DeviceMem>; using DeviceMemPtr = std::unique_ptr<DeviceMem>;
...@@ -62,6 +64,7 @@ int run(int argc, char* argv[]) ...@@ -62,6 +64,7 @@ int run(int argc, char* argv[])
std::vector<DeviceMemPtr> b0_tensors_device; std::vector<DeviceMemPtr> b0_tensors_device;
std::vector<DeviceMemPtr> b1_tensors_device; std::vector<DeviceMemPtr> b1_tensors_device;
std::vector<DeviceMemPtr> c_tensors_device; std::vector<DeviceMemPtr> c_tensors_device;
std::vector<DeviceMemPtr> z_tensors_device;
std::vector<DeviceMemPtr> lse_tensors_device; std::vector<DeviceMemPtr> lse_tensors_device;
std::size_t flop = 0, num_byte = 0; std::size_t flop = 0, num_byte = 0;
...@@ -101,6 +104,12 @@ int run(int argc, char* argv[]) ...@@ -101,6 +104,12 @@ int run(int argc, char* argv[])
output_permute output_permute
? std::vector<ck::index_t>{M * G1 * O, O, G1 * O, 1} // C layout [G0, M, G1, O] ? std::vector<ck::index_t>{M * G1 * O, O, G1 * O, 1} // C layout [G0, M, G1, O]
: std::vector<ck::index_t>{G1 * M * O, M * O, O, 1}; // C layout [G0, G1, M, O] : std::vector<ck::index_t>{G1 * M * O, M * O, O, 1}; // C layout [G0, G1, M, O]
std::vector<ck::index_t> z_gs_ms_ns_lengths{G0, G1, M, N};
std::vector<ck::index_t> z_gs_ms_ns_strides =
output_permute
? std::vector<ck::index_t>{M * G1 * N, N, G1 * N, 1} // Z layout [G0, M, G1, N]
: std::vector<ck::index_t>{G1 * M * N, M * N, N, 1}; // Z layout [G0, G1, M, N]
std::vector<ck::index_t> lse_gs_ms_lengths{G0, G1, M}; std::vector<ck::index_t> lse_gs_ms_lengths{G0, G1, M};
std::vector<ck::index_t> lse_gs_ms_strides = std::vector<ck::index_t> lse_gs_ms_strides =
...@@ -114,6 +123,8 @@ int run(int argc, char* argv[]) ...@@ -114,6 +123,8 @@ int run(int argc, char* argv[])
b1_gs_os_ns_strides, b1_gs_os_ns_strides,
c_gs_ms_os_lengths, c_gs_ms_os_lengths,
c_gs_ms_os_strides, c_gs_ms_os_strides,
z_gs_ms_ns_lengths,
z_gs_ms_ns_strides,
lse_gs_ms_lengths, lse_gs_ms_lengths,
lse_gs_ms_strides, lse_gs_ms_strides,
{}, // acc0_biases_gs_ms_ns_lengths {}, // acc0_biases_gs_ms_ns_lengths
...@@ -126,6 +137,7 @@ int run(int argc, char* argv[]) ...@@ -126,6 +137,7 @@ int run(int argc, char* argv[])
Tensor<B0DataType> b0_gs_ns_ks(b0_gs_ns_ks_lengths, b0_gs_ns_ks_strides); Tensor<B0DataType> b0_gs_ns_ks(b0_gs_ns_ks_lengths, b0_gs_ns_ks_strides);
Tensor<B1DataType> b1_gs_os_ns(b1_gs_os_ns_lengths, b1_gs_os_ns_strides); Tensor<B1DataType> b1_gs_os_ns(b1_gs_os_ns_lengths, b1_gs_os_ns_strides);
Tensor<CDataType> c_gs_ms_os_device_result(c_gs_ms_os_lengths, c_gs_ms_os_strides); Tensor<CDataType> c_gs_ms_os_device_result(c_gs_ms_os_lengths, c_gs_ms_os_strides);
Tensor<ZDataType> z_gs_ms_ns(z_gs_ms_ns_lengths, z_gs_ms_ns_strides);
Tensor<LSEDataType> lse_gs_ms_device_result(lse_gs_ms_lengths, lse_gs_ms_strides); Tensor<LSEDataType> lse_gs_ms_device_result(lse_gs_ms_lengths, lse_gs_ms_strides);
int Batch = G0 * G1; int Batch = G0 * G1;
...@@ -140,10 +152,13 @@ int run(int argc, char* argv[]) ...@@ -140,10 +152,13 @@ int run(int argc, char* argv[])
<< "b0_gs_ns_ks[" << i << "]: " << b0_gs_ns_ks.mDesc << ", " << "b0_gs_ns_ks[" << i << "]: " << b0_gs_ns_ks.mDesc << ", "
<< "b1_gs_os_ns[" << i << "]: " << b1_gs_os_ns.mDesc << ", " << "b1_gs_os_ns[" << i << "]: " << b1_gs_os_ns.mDesc << ", "
<< "c_gs_ms_os[" << i << "]: " << c_gs_ms_os_device_result.mDesc << ", " << "c_gs_ms_os[" << i << "]: " << c_gs_ms_os_device_result.mDesc << ", "
<< "c_gs_ms_os[" << i << "]: " << c_gs_ms_os_device_result.mDesc << ", "
<< "lse_gs_ms_os[" << i << "]: " << lse_gs_ms_device_result.mDesc << "lse_gs_ms_os[" << i << "]: " << lse_gs_ms_device_result.mDesc
<< std::endl; << std::endl;
} }
z_gs_ms_ns.GenerateTensorValue(GeneratorTensor_1<DataType>{0});
switch(init_method) switch(init_method)
{ {
case 0: break; case 0: break;
...@@ -172,6 +187,7 @@ int run(int argc, char* argv[]) ...@@ -172,6 +187,7 @@ int run(int argc, char* argv[])
b0_tensors.push_back(b0_gs_ns_ks); b0_tensors.push_back(b0_gs_ns_ks);
b1_tensors.push_back(b1_gs_os_ns); b1_tensors.push_back(b1_gs_os_ns);
c_tensors.push_back(c_gs_ms_os_device_result); c_tensors.push_back(c_gs_ms_os_device_result);
z_tensors.push_back(z_gs_ms_ns);
lse_tensors.push_back(lse_gs_ms_device_result); lse_tensors.push_back(lse_gs_ms_device_result);
a_tensors_device.emplace_back(std::make_unique<DeviceMem>( a_tensors_device.emplace_back(std::make_unique<DeviceMem>(
...@@ -182,6 +198,8 @@ int run(int argc, char* argv[]) ...@@ -182,6 +198,8 @@ int run(int argc, char* argv[])
sizeof(B1DataType) * b1_gs_os_ns.mDesc.GetElementSpaceSize())); sizeof(B1DataType) * b1_gs_os_ns.mDesc.GetElementSpaceSize()));
c_tensors_device.emplace_back(std::make_unique<DeviceMem>( c_tensors_device.emplace_back(std::make_unique<DeviceMem>(
sizeof(CDataType) * c_gs_ms_os_device_result.mDesc.GetElementSpaceSize())); sizeof(CDataType) * c_gs_ms_os_device_result.mDesc.GetElementSpaceSize()));
z_tensors_device.emplace_back(std::make_unique<DeviceMem>(
sizeof(ZDataType) * z_gs_ms_ns.mDesc.GetElementSpaceSize()));
lse_tensors_device.emplace_back(std::make_unique<DeviceMem>( lse_tensors_device.emplace_back(std::make_unique<DeviceMem>(
sizeof(LSEDataType) * lse_gs_ms_device_result.mDesc.GetElementSpaceSize())); sizeof(LSEDataType) * lse_gs_ms_device_result.mDesc.GetElementSpaceSize()));
...@@ -193,6 +211,7 @@ int run(int argc, char* argv[]) ...@@ -193,6 +211,7 @@ int run(int argc, char* argv[])
p_b0.push_back(b0_tensors_device[i]->GetDeviceBuffer()); p_b0.push_back(b0_tensors_device[i]->GetDeviceBuffer());
p_b1.push_back(b1_tensors_device[i]->GetDeviceBuffer()); p_b1.push_back(b1_tensors_device[i]->GetDeviceBuffer());
p_c.push_back(c_tensors_device[i]->GetDeviceBuffer()); p_c.push_back(c_tensors_device[i]->GetDeviceBuffer());
p_z.push_back(z_tensors_device[i]->GetDeviceBuffer());
p_lse.push_back(lse_tensors_device[i]->GetDeviceBuffer()); p_lse.push_back(lse_tensors_device[i]->GetDeviceBuffer());
} }
...@@ -209,6 +228,7 @@ int run(int argc, char* argv[]) ...@@ -209,6 +228,7 @@ int run(int argc, char* argv[])
p_b0, p_b0,
p_b1, p_b1,
p_c, p_c,
p_z,
p_lse, p_lse,
{}, // p_acc0_biases {}, // p_acc0_biases
{}, // p_acc1_biases {}, // p_acc1_biases
......
include/ck/tensor_operation/gpu/device/device_grouped_gemm_softmax_gemm_permute.hpp
View file @ bf80ceee
...@@ -79,6 +79,7 @@ template <index_t NumDimG, ...@@ -79,6 +79,7 @@ template <index_t NumDimG,
typename B0DataType, typename B0DataType,
typename B1DataType, typename B1DataType,
typename CDataType, typename CDataType,
typename ZDataType,
typename LSEDataType, typename LSEDataType,
typename Acc0BiasDataType, typename Acc0BiasDataType,
typename Acc1BiasDataType, typename Acc1BiasDataType,
...@@ -104,6 +105,9 @@ struct DeviceGroupedMultiheadAttentionForward : public BaseOperator ...@@ -104,6 +105,9 @@ struct DeviceGroupedMultiheadAttentionForward : public BaseOperator
std::vector<index_t> c_gs_ms_os_lengths; std::vector<index_t> c_gs_ms_os_lengths;
std::vector<index_t> c_gs_ms_os_strides; std::vector<index_t> c_gs_ms_os_strides;
std::vector<index_t> z_gs_ms_ns_lengths;
std::vector<index_t> z_gs_ms_ns_strides;
std::vector<index_t> lse_gs_ms_lengths; std::vector<index_t> lse_gs_ms_lengths;
std::vector<index_t> lse_gs_ms_strides; std::vector<index_t> lse_gs_ms_strides;
...@@ -119,6 +123,7 @@ struct DeviceGroupedMultiheadAttentionForward : public BaseOperator ...@@ -119,6 +123,7 @@ struct DeviceGroupedMultiheadAttentionForward : public BaseOperator
std::vector<const void*> p_b0_vec, std::vector<const void*> p_b0_vec,
std::vector<const void*> p_b1_vec, std::vector<const void*> p_b1_vec,
std::vector<void*> p_c_vec, std::vector<void*> p_c_vec,
std::vector<void*> p_z_vec,
std::vector<void*> p_lse_vec, std::vector<void*> p_lse_vec,
std::vector<std::vector<const void*>> p_acc0_biases_vec, std::vector<std::vector<const void*>> p_acc0_biases_vec,
std::vector<std::vector<const void*>> p_acc1_biases_vec, std::vector<std::vector<const void*>> p_acc1_biases_vec,
......
include/ck/tensor_operation/gpu/device/impl/device_grouped_multihead_attention_forward_xdl_cshuffle 0 → 100644
View file @ bf80ceee
// SPDX-License-Identifier: MIT
// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
#pragma once
#include <iostream>
#include <sstream>
#include "ck/utility/common_header.hpp"
#include "ck/utility/philox_rand.hpp"
#include "ck/tensor_description/tensor_descriptor.hpp"
#include "ck/tensor_description/tensor_descriptor_helper.hpp"
#include "ck/tensor_operation/gpu/device/tensor_layout.hpp"
#include "ck/tensor_operation/gpu/device/device_grouped_gemm_softmax_gemm_permute.hpp"
#include "ck/tensor_operation/gpu/device/gemm_specialization.hpp"
#include "ck/tensor_operation/gpu/device/matrix_padder.hpp"
#include "ck/tensor_operation/gpu/grid/gridwise_batched_gemm_softmax_gemm_xdl_cshuffle_v2.hpp"
#include "ck/tensor_operation/operator_transform/transform_contraction_to_gemm.hpp"
#include "ck/host_utility/device_prop.hpp"
#include "ck/host_utility/kernel_launch.hpp"
namespace ck {
namespace tensor_operation {
namespace device {
template <typename GridwiseGemm,
typename GemmAccDataType,
typename GroupKernelArg,
typename AElementwiseOperation,
typename BElementwiseOperation,
typename AccElementwiseOperation,
typename B1ElementwiseOperation,
typename CElementwiseOperation,
bool HasMainKBlockLoop,
bool IsDropout>
__global__ void
#if CK_USE_LAUNCH_BOUNDS
__launch_bounds__(CK_MAX_THREAD_PER_BLOCK, CK_MIN_BLOCK_PER_CU)
#endif
kernel_grouped_gemm_softmax_gemm_xdl_cshuffle_v2(
const void CK_CONSTANT_ADDRESS_SPACE* group_kernel_args,
const index_t group_count,
const AElementwiseOperation a_element_op,
const BElementwiseOperation b_element_op,
const AccElementwiseOperation acc_element_op,
const B1ElementwiseOperation b1_element_op,
const CElementwiseOperation c_element_op,
const ushort p_dropout_in_16bits,
const GemmAccDataType p_dropout_rescale,
const unsigned long long seed,
const unsigned long long offset)
{
#if(!defined(__HIP_DEVICE_COMPILE__) || defined(__gfx908__) || defined(__gfx90a__))
__shared__ char p_shared[GridwiseGemm::GetSharedMemoryNumberOfByte()];
const index_t block_id = get_block_1d_id();
const index_t global_thread_id = get_thread_global_1d_id();
ck::philox ph(seed, global_thread_id, offset);
const auto arg_ptr = reinterpret_cast<const GroupKernelArg*>(
cast_pointer_to_generic_address_space(group_kernel_args));
index_t left = 0;
index_t right = group_count;
index_t group_id = index_t((left + right) / 2);
while(
(!(block_id >= arg_ptr[group_id].block_start_ && block_id < arg_ptr[group_id].block_end_)))
{
if(block_id < arg_ptr[group_id].block_start_)
{
right = group_id;
}
else
{
left = group_id;
}
group_id = index_t((left + right) / 2);
}
// per-group batch offset
const index_t num_blocks_per_batch = arg_ptr[group_id].num_blocks_per_batch_;
const index_t g_idx = __builtin_amdgcn_readfirstlane(
(block_id - arg_ptr[group_id].block_start_) / num_blocks_per_batch);
const long_index_t a_batch_offset = __builtin_amdgcn_readfirstlane(
static_cast<long_index_t>(arg_ptr[group_id].compute_base_ptr_of_batch_.GetABasePtr(g_idx)));
const long_index_t b_batch_offset = __builtin_amdgcn_readfirstlane(
static_cast<long_index_t>(arg_ptr[group_id].compute_base_ptr_of_batch_.GetBBasePtr(g_idx)));
const long_index_t b1_batch_offset = __builtin_amdgcn_readfirstlane(static_cast<long_index_t>(
arg_ptr[group_id].compute_base_ptr_of_batch_.GetB1BasePtr(g_idx)));
const long_index_t c_batch_offset = __builtin_amdgcn_readfirstlane(
static_cast<long_index_t>(arg_ptr[group_id].compute_base_ptr_of_batch_.GetCBasePtr(g_idx)));
const long_index_t z_batch_offset = __builtin_amdgcn_readfirstlane(
static_cast<long_index_t>(arg_ptr[group_id].compute_base_ptr_of_batch_.GetZBasePtr(g_idx)));
const long_index_t lse_batch_offset = __builtin_amdgcn_readfirstlane(static_cast<long_index_t>(
arg_ptr[group_id].compute_base_ptr_of_batch_.GetLSEBasePtr(g_idx)));
//unsigned short* p_z_grid_in = //
// (arg_ptr[group_id].p_z_grid_ == nullptr ? nullptr
// : arg_ptr[group_id].p_z_grid_ + z_batch_offset);
GridwiseGemm::template Run<HasMainKBlockLoop, IsDropout>(
arg_ptr[group_id].p_a_grid_ + a_batch_offset,
arg_ptr[group_id].p_b_grid_ + b_batch_offset,
arg_ptr[group_id].p_b1_grid_ + b1_batch_offset,
arg_ptr[group_id].p_c_grid_ + c_batch_offset,
arg_ptr[group_id].p_z_grid_ == nullptr ? nullptr
: arg_ptr[group_id].p_z_grid_ + z_batch_offset,
arg_ptr[group_id].p_lse_grid_ + lse_batch_offset,
p_shared,
a_element_op,
b_element_op,
acc_element_op,
b1_element_op,
c_element_op,
arg_ptr[group_id].a_grid_desc_ak0_m_ak1_,
arg_ptr[group_id].b_grid_desc_bk0_n_bk1_,
arg_ptr[group_id].b1_grid_desc_bk0_n_bk1_,
arg_ptr[group_id].c_grid_desc_mblock_mperblock_nblock_nperblock_,
arg_ptr[group_id].z_grid_desc_m0_n0_m1_n1_m2_n2_m3_n3_n4_n5_, ////////
arg_ptr[group_id].lse_grid_desc_m_,
arg_ptr[group_id].block_2_ctile_map_,
arg_ptr[group_id].c0_matrix_mask_,
p_dropout_in_16bits,
p_dropout_rescale,
ph);
#else
ignore = group_kernel_args;
ignore = group_count;
ignore = a_element_op;
ignore = b_element_op;
ignore = acc_element_op;
ignore = b1_element_op;
ignore = c_element_op;
#endif // end of if (defined(__gfx908__) || defined(__gfx90a__))
}
// Computes C = A * B0 * B1
// ^^^^^^ (Acc0)
// ^^^^^^^^^^^ (Acc1)
template <index_t NumDimG,
index_t NumDimM,
index_t NumDimN,
index_t NumDimK,
index_t NumDimO, // NumDimGemm1N
typename ADataType,
typename BDataType,
typename B1DataType,
typename CDataType,
typename ZDataType,
typename LSEDataType,
typename Acc0BiasDataType,
typename Acc1BiasDataType,
typename GemmAccDataType,
typename CShuffleDataType,
typename AElementwiseOperation,
typename BElementwiseOperation,
typename AccElementwiseOperation,
typename B1ElementwiseOperation,
typename CElementwiseOperation,
GemmSpecialization GemmSpec,
TensorSpecialization ASpec,
TensorSpecialization BSpec,
TensorSpecialization B1Spec,
TensorSpecialization CSpec,
index_t NumGemmKPrefetchStage,
index_t BlockSize,
index_t MPerBlock,
index_t NPerBlock, // Gemm0NPerBlock
index_t KPerBlock, // Gemm0KPerBlock
index_t Gemm1NPerBlock,
index_t Gemm1KPerBlock,
index_t AK1,
index_t BK1,
index_t B1K1,
index_t MPerXDL,
index_t NPerXDL,
index_t MXdlPerWave,
index_t NXdlPerWave,
index_t Gemm1NXdlPerWave,
typename ABlockTransferThreadClusterLengths_AK0_M_AK1,
typename ABlockTransferThreadClusterArrangeOrder,
typename ABlockTransferSrcAccessOrder,
index_t ABlockTransferSrcVectorDim,
index_t ABlockTransferSrcScalarPerVector,
index_t ABlockTransferDstScalarPerVector_AK1,
bool ABlockLdsExtraM,
typename BBlockTransferThreadClusterLengths_BK0_N_BK1,
typename BBlockTransferThreadClusterArrangeOrder,
typename BBlockTransferSrcAccessOrder,
index_t BBlockTransferSrcVectorDim,
index_t BBlockTransferSrcScalarPerVector,
index_t BBlockTransferDstScalarPerVector_BK1,
bool BBlockLdsExtraN,
typename B1BlockTransferThreadClusterLengths_BK0_N_BK1,
typename B1BlockTransferThreadClusterArrangeOrder,
typename B1BlockTransferSrcAccessOrder,
index_t B1BlockTransferSrcVectorDim,
index_t B1BlockTransferSrcScalarPerVector,
index_t B1BlockTransferDstScalarPerVector_BK1,
bool B1BlockLdsExtraN,
index_t CShuffleMXdlPerWavePerShuffle,
index_t CShuffleNXdlPerWavePerShuffle,
typename CShuffleBlockTransferClusterLengths_MBlock_MPerBlock_NBlock_NPerBlock,
index_t CShuffleBlockTransferScalarPerVector_NPerBlock,
MaskingSpecialization MaskingSpec,
LoopScheduler LoopSched = LoopScheduler::Default>
struct DeviceGroupedGemmSoftmaxGemmPermute_Train_Xdl_CShuffle
: public DeviceGroupedGemmSoftmaxGemmPermuteTrain<NumDimG,
NumDimM,
NumDimN,
NumDimK,
NumDimO,
ADataType,
BDataType,
B1DataType,
CDataType,
ZDataType,
LSEDataType,
Acc0BiasDataType,
Acc1BiasDataType,
AElementwiseOperation,
BElementwiseOperation,
AccElementwiseOperation,
B1ElementwiseOperation,
CElementwiseOperation,
MaskingSpec>
{
static_assert(NumDimG > 0 && NumDimM > 0 && NumDimN > 0 && NumDimK > 0 && NumDimO > 0,
"Number of dimension must be greater than 0");
static constexpr index_t NumAcc0Bias = Acc0BiasDataType::Size();
static constexpr index_t NumAcc1Bias = Acc1BiasDataType::Size();
// TODO ANT: implement bias combination
static_assert(NumAcc0Bias == 0 && NumAcc0Bias == 0, "Bias addition is unimplemented");
#if 0
// TODO ANT: use alias
static constexpr index_t NumDimGemm0M = NumDimM;
static constexpr index_t NumDimGemm0N = NumDimN;
static constexpr index_t NumDimGemm0K = NumDimK;
static constexpr index_t NumDimGemm1M = NumDimM;
static constexpr index_t NumDimGemm1N = NumDimO;
static constexpr index_t NumDimGemm1K = NumDimN;
#endif
using DeviceOp = DeviceGroupedGemmSoftmaxGemmPermute_Train_Xdl_CShuffle;
using ProblemDesc = typename DeviceGroupedGemmSoftmaxGemmPermuteTrain<NumDimG,
NumDimM,
NumDimN,
NumDimK,
NumDimO,
ADataType,
BDataType,
B1DataType,
CDataType,
ZDataType,
LSEDataType,
Acc0BiasDataType,
Acc1BiasDataType,
AElementwiseOperation,
BElementwiseOperation,
AccElementwiseOperation,
B1ElementwiseOperation,
CElementwiseOperation,
MaskingSpec>::ProblemDesc;
static constexpr auto I0 = Number<0>{};
static constexpr auto I1 = Number<1>{};
static constexpr auto I2 = Number<2>{};
using Transform = TransformBatchedContractionContractionToBatchedGemmGemm<
Sequence<NumDimG, NumDimM, NumDimN, NumDimK, NumDimO>,
Sequence<MPerBlock, NPerBlock, KPerBlock, Gemm1NPerBlock>,
GemmSpec,
ASpec,
BSpec,
B1Spec,
CSpec>;
static auto MakeAGridDescriptor_AK0_M_AK1(const std::vector<index_t>& a_gs_ms_ks_lengths_vec,
const std::vector<index_t>& a_gs_ms_ks_strides_vec)
{
return Transform::MakeAGridDescriptor_AK0_M_AK1(
Transform::MakeAGridDescriptor_M_K(a_gs_ms_ks_lengths_vec, a_gs_ms_ks_strides_vec),
Number<AK1>{});
}
static auto MakeBGridDescriptor_BK0_N_BK1(const std::vector<index_t>& b_gs_ns_ks_lengths_vec,
const std::vector<index_t>& b_gs_ns_ks_strides_vec)
{
return Transform::MakeB0GridDescriptor_BK0_N_BK1(
Transform::MakeB0GridDescriptor_N_K(b_gs_ns_ks_lengths_vec, b_gs_ns_ks_strides_vec),
Number<BK1>{});
}
static auto
MakeB1GridDescriptor_BK0_N_BK1(const std::vector<index_t>& b1_gs_gemm1ns_gemm1ks_lengths_vec,
const std::vector<index_t>& b1_gs_gemm1ns_gemm1ks_strides_vec)
{
return Transform::MakeB1GridDescriptor_BK0_N_BK1(
Transform::MakeB1GridDescriptor_N_K(b1_gs_gemm1ns_gemm1ks_lengths_vec,
b1_gs_gemm1ns_gemm1ks_strides_vec),
Number<B1K1>{});
}
static auto MakeZGridDescriptor_M_N(const std::vector<index_t>& z_gs_ms_ns_lengths_vec,
const std::vector<index_t>& z_gs_ms_ns_strides_vec)
{
return Transform::MakeCGridDescriptor_M_N(z_gs_ms_ns_lengths_vec, z_gs_ms_ns_strides_vec);
}
static auto MakeLSEGridDescriptor_M(index_t MRaw)
{
const auto lse_grid_desc_mraw = make_naive_tensor_descriptor_packed(make_tuple(MRaw));
const auto M = math::integer_divide_ceil(MRaw, MPerBlock) * MPerBlock;
const auto MPad = M - MRaw;
if constexpr(GemmSpec == GemmSpecialization::MPadding ||
GemmSpec == GemmSpecialization::MNPadding ||
GemmSpec == GemmSpecialization::MKPadding ||
GemmSpec == GemmSpecialization::MNKPadding)
{
// pad M
return transform_tensor_descriptor(lse_grid_desc_mraw,
make_tuple(make_right_pad_transform(MRaw, MPad)),
make_tuple(Sequence<0>{}),
make_tuple(Sequence<0>{}));
}
else
{
// not pad M
return lse_grid_desc_mraw;
}
}
using AGridDesc_AK0_M_AK1 = decltype(MakeAGridDescriptor_AK0_M_AK1({}, {}));
using BGridDesc_BK0_N_BK1 = decltype(MakeBGridDescriptor_BK0_N_BK1({}, {}));
using B1GridDesc_BK0_N_BK1 = decltype(MakeB1GridDescriptor_BK0_N_BK1({}, {}));
using CGridDesc_M_N = decltype(Transform::MakeCGridDescriptor_M_N({}, {}));
using LSEGridDesc_M = decltype(MakeLSEGridDescriptor_M(1));
using ZGridDesc_M_N = decltype(MakeZGridDescriptor_M_N({}, {}));
using AGridDesc_G_M_K = decltype(Transform::MakeAGridDescriptor_G_M_K({}, {}));
using BGridDesc_G_N_K = decltype(Transform::MakeB0GridDescriptor_G_N_K({}, {}));
using B1GridDesc_G_N_K = decltype(Transform::MakeB1GridDescriptor_G_N_K({}, {}));
using CGridDesc_G_M_N = decltype(Transform::MakeCGridDescriptor_G_M_N({}, {}));
using ZGridDesc_G_M_N = decltype(Transform::MakeCGridDescriptor_G_M_N({}, {}));
constexpr static auto make_MaskOutPredicate()
{
if constexpr(MaskingSpec == MaskingSpecialization::MaskDisabled)
{
return MaskDisabledPredicate{};
}
else if constexpr(MaskingSpec == MaskingSpecialization::MaskOutUpperTriangle)
{
return MaskOutUpperTrianglePredicate{};
}
}
using C0MatrixMask = C0MatrixMask_impl<decltype(make_MaskOutPredicate())>;
struct ComputeBasePtrOfStridedBatch
{
ComputeBasePtrOfStridedBatch(const AGridDesc_G_M_K& a_grid_desc_g_m_k,
const BGridDesc_G_N_K& b_grid_desc_g_n_k,
const B1GridDesc_G_N_K& b1_grid_desc_g_n_k,
const CGridDesc_G_M_N& c_grid_desc_g_m_n,
const ZGridDesc_G_M_N& z_grid_desc_g_m_n,
index_t BatchStrideLSE)
: a_grid_desc_g_m_k_(a_grid_desc_g_m_k),
b_grid_desc_g_n_k_(b_grid_desc_g_n_k),
b1_grid_desc_g_n_k_(b1_grid_desc_g_n_k),
c_grid_desc_g_m_n_(c_grid_desc_g_m_n),
z_grid_desc_g_m_n_(z_grid_desc_g_m_n),
BatchStrideLSE_(BatchStrideLSE)
{
}
__host__ __device__ constexpr long_index_t GetABasePtr(index_t g_idx) const
{
return a_grid_desc_g_m_k_.CalculateOffset(make_multi_index(g_idx, 0, 0));
}
__host__ __device__ constexpr long_index_t GetBBasePtr(index_t g_idx) const
{
return b_grid_desc_g_n_k_.CalculateOffset(make_multi_index(g_idx, 0, 0));
}
__host__ __device__ constexpr long_index_t GetB1BasePtr(index_t g_idx) const
{
return b1_grid_desc_g_n_k_.CalculateOffset(make_multi_index(g_idx, 0, 0));
}
__host__ __device__ constexpr long_index_t GetCBasePtr(index_t g_idx) const
{
return c_grid_desc_g_m_n_.CalculateOffset(make_multi_index(g_idx, 0, 0));
}
__host__ __device__ constexpr long_index_t GetZBasePtr(index_t g_idx) const
{
return z_grid_desc_g_m_n_.CalculateOffset(make_multi_index(g_idx, 0, 0));
}
__host__ __device__ constexpr long_index_t GetLSEBasePtr(index_t g_idx) const
{
return g_idx * static_cast<long_index_t>(BatchStrideLSE_);
}
private:
AGridDesc_G_M_K a_grid_desc_g_m_k_;
BGridDesc_G_N_K b_grid_desc_g_n_k_;
B1GridDesc_G_N_K b1_grid_desc_g_n_k_;
CGridDesc_G_M_N c_grid_desc_g_m_n_;
ZGridDesc_G_M_N z_grid_desc_g_m_n_;
index_t BatchStrideLSE_;
};
// GridwiseGemm
using GridwiseGemm = GridwiseBatchedGemmSoftmaxGemmTrain_Xdl_CShuffle<
ADataType, // TODO: distinguish A/B datatype
GemmAccDataType,
CShuffleDataType,
CDataType,
LSEDataType,
AElementwiseOperation,
BElementwiseOperation,
AccElementwiseOperation,
B1ElementwiseOperation,
CElementwiseOperation,
InMemoryDataOperationEnum::Set,
AGridDesc_AK0_M_AK1,
BGridDesc_BK0_N_BK1,
B1GridDesc_BK0_N_BK1,
CGridDesc_M_N,
ZGridDesc_M_N,
LSEGridDesc_M,
NumGemmKPrefetchStage,
BlockSize,
MPerBlock,
NPerBlock,
KPerBlock,
Gemm1NPerBlock,
Gemm1KPerBlock,
AK1,
BK1,
B1K1,
MPerXDL,
NPerXDL,
MXdlPerWave,
NXdlPerWave,
Gemm1NXdlPerWave,
ABlockTransferThreadClusterLengths_AK0_M_AK1,
ABlockTransferThreadClusterArrangeOrder,
ABlockTransferSrcAccessOrder,
ABlockTransferSrcVectorDim,
ABlockTransferSrcScalarPerVector,
ABlockTransferDstScalarPerVector_AK1,
true,
ABlockLdsExtraM,
BBlockTransferThreadClusterLengths_BK0_N_BK1,
BBlockTransferThreadClusterArrangeOrder,
BBlockTransferSrcAccessOrder,
BBlockTransferSrcVectorDim,
BBlockTransferSrcScalarPerVector,
BBlockTransferDstScalarPerVector_BK1,
true,
BBlockLdsExtraN,
B1BlockTransferThreadClusterLengths_BK0_N_BK1,
B1BlockTransferThreadClusterArrangeOrder,
B1BlockTransferSrcAccessOrder,
B1BlockTransferSrcVectorDim,
B1BlockTransferSrcScalarPerVector,
B1BlockTransferDstScalarPerVector_BK1,
false,
B1BlockLdsExtraN,
CShuffleMXdlPerWavePerShuffle,
CShuffleNXdlPerWavePerShuffle,
CShuffleBlockTransferClusterLengths_MBlock_MPerBlock_NBlock_NPerBlock,
CShuffleBlockTransferScalarPerVector_NPerBlock,
LoopSched,
Transform::matrix_padder.PadN,
MaskingSpec == MaskingSpecialization::MaskOutUpperTriangle>;
using Block2CTileMap = OffsettedBlockToCTileMap<typename GridwiseGemm::DefaultBlock2CTileMap>;
struct GroupKernelArg
{
// pointers
const ADataType* p_a_grid_;
const BDataType* p_b_grid_;
const B1DataType* p_b1_grid_;
CDataType* p_c_grid_;
ZDataType* p_z_grid_;
LSEDataType* p_lse_grid_;
// tensor descriptors for block/thread-wise copy
AGridDesc_AK0_M_AK1 a_grid_desc_ak0_m_ak1_;
BGridDesc_BK0_N_BK1 b_grid_desc_bk0_n_bk1_;
B1GridDesc_BK0_N_BK1 b1_grid_desc_bk0_n_bk1_;
typename GridwiseGemm::CGridDescriptor_MBlock_MPerBlock_NBlock_NPerBlock
c_grid_desc_mblock_mperblock_nblock_nperblock_;
typename GridwiseGemm::ZGridDescriptor_M0_N0_M1_N1_M2_N2_M3_N3_N4_N5
z_grid_desc_m0_n0_m1_n1_m2_n2_m3_n3_n4_n5_;
ZGridDesc_M_N z_grid_desc_m_n_;
LSEGridDesc_M lse_grid_desc_m_;
// batch & stride
index_t num_blocks_per_batch_;
ComputeBasePtrOfStridedBatch compute_base_ptr_of_batch_;
// check C0 masking and padding
C0MatrixMask c0_matrix_mask_;
// block-to-c-tile map
Block2CTileMap block_2_ctile_map_;
index_t block_start_, block_end_;
};
struct GroupDeviceArg
{
// lengths for the last dimensions of overall problem for sanity check of vector load/store
std::vector<index_t> raw_lengths_mz_nz_kz_gemm1nz_;
// strides for the last dimensions of each tensor for sanity check of vector load/store
std::vector<index_t> a_mz_kz_strides_;
std::vector<index_t> b_nz_kz_strides_;
std::vector<index_t> b1_nz_kz_strides_;
std::vector<index_t> c_mz_gemm1nz_strides_;
// for gridwise gemm check
CGridDesc_M_N c_grid_desc_m_n_;
};
// Argument
// FIXME: constness
struct Argument : public BaseArgument
{
Argument(std::vector<const void*> p_a_vec,
std::vector<const void*> p_b_vec,
std::vector<const void*> p_b1_vec,
std::vector<void*> p_c_vec,
std::vector<void*> p_z_vec,
std::vector<void*> p_lse_vec,
std::vector<std::vector<const void*>> p_acc0_biases_vec,
std::vector<std::vector<const void*>> p_acc1_biases_vec,
std::vector<ProblemDesc> problem_desc_vec,
AElementwiseOperation a_element_op,
BElementwiseOperation b_element_op,
AccElementwiseOperation acc_element_op,
B1ElementwiseOperation b1_element_op,
CElementwiseOperation c_element_op,
float p_dropout,
std::tuple<unsigned long long, unsigned long long> seeds)
: a_element_op_{a_element_op},
b_element_op_{b_element_op},
acc_element_op_{acc_element_op},
b1_element_op_{b1_element_op},
c_element_op_{c_element_op}
{
// TODO ANT: implement bias addition
group_count_ = problem_desc_vec.size();
if(!(group_count_ == p_a_vec.size() && group_count_ == p_b_vec.size() &&
group_count_ == p_b1_vec.size() && group_count_ == p_c_vec.size()))
{
throw std::runtime_error("wrong! group_count_ != a/b/b1/c_vec.size");
}
if(!(p_acc0_biases_vec.size() == p_acc1_biases_vec.size()))
{
throw std::runtime_error("wrong! acc0_bias_vec.size != acc1_bias_vec.size");
}
grid_size_ = 0;
for(std::size_t i = 0; i < group_count_; i++)
{
const auto p_a_grid = static_cast<const ADataType*>(p_a_vec[i]);
const auto p_b_grid = static_cast<const BDataType*>(p_b_vec[i]);
const auto p_b1_grid = static_cast<const B1DataType*>(p_b1_vec[i]);
const auto p_c_grid = static_cast<CDataType*>(p_c_vec[i]);
const auto p_z_grid = static_cast<ZDataType*>(p_z_vec[i]);
const auto p_lse_grid = static_cast<LSEDataType*>(p_lse_vec[i]);
const auto& problem_desc = problem_desc_vec[i];
const auto a_grid_desc_ak0_m_ak1 = MakeAGridDescriptor_AK0_M_AK1(
problem_desc.a_gs_ms_ks_lengths, problem_desc.a_gs_ms_ks_strides);
const auto b_grid_desc_bk0_n_bk1 = MakeBGridDescriptor_BK0_N_BK1(
problem_desc.b0_gs_ns_ks_lengths, problem_desc.b0_gs_ns_ks_strides);
const auto b1_grid_desc_bk0_n_bk1 = MakeB1GridDescriptor_BK0_N_BK1(
problem_desc.b1_gs_os_ns_lengths, problem_desc.b1_gs_os_ns_strides);
const auto c_grid_desc_m_n = Transform::MakeCGridDescriptor_M_N(
problem_desc.c_gs_ms_os_lengths, problem_desc.c_gs_ms_os_strides);
const auto z_grid_desc_m_n = MakeZGridDescriptor_M_N(
problem_desc.z_gs_ms_ns_lengths, problem_desc.z_gs_ms_ns_strides);
const auto lse_grid_desc_m =
DeviceOp::MakeLSEGridDescriptor_M(problem_desc.lse_gs_ms_lengths[NumDimG]);
const auto a_grid_desc_g_m_k = Transform::MakeAGridDescriptor_G_M_K(
problem_desc.a_gs_ms_ks_lengths, problem_desc.a_gs_ms_ks_strides);
const auto b_grid_desc_g_n_k = Transform::MakeB0GridDescriptor_G_N_K(
problem_desc.b0_gs_ns_ks_lengths, problem_desc.b0_gs_ns_ks_strides);
const auto b1_grid_desc_g_n_k = Transform::MakeB1GridDescriptor_G_N_K(
problem_desc.b1_gs_os_ns_lengths, problem_desc.b1_gs_os_ns_strides);
const auto c_grid_desc_g_m_n = Transform::MakeCGridDescriptor_G_M_N(
problem_desc.c_gs_ms_os_lengths, problem_desc.c_gs_ms_os_strides);
const auto z_grid_desc_g_m_n = Transform::MakeCGridDescriptor_G_M_N(
problem_desc.z_gs_ms_ns_lengths, problem_desc.z_gs_ms_ns_strides);
const auto c_grid_desc_mblock_mperblock_nblock_nperblock =
GridwiseGemm::MakeCGridDescriptor_MBlock_MPerBlock_NBlock_NPerBlock(
c_grid_desc_m_n);
//typename GridwiseGemm::ZGridDescriptor_M0_N0_M1_N1_M2_N2_M3_N3_N4_N5
// z_grid_desc_m0_n0_m1_n1_m2_n2_m3_n3_n4_n5;
auto z_grid_desc_m0_n0_m1_n1_m2_n2_m3_n3_n4_n5 =
GridwiseGemm::MakeCGridDescriptor_M0_N0_M1_N1_M2_N2_M3_N3_N4_N5(
z_grid_desc_m_n);
const index_t BlockStart = grid_size_;
const auto block_2_ctile_map = Block2CTileMap(c_grid_desc_m_n, BlockStart);
const index_t batch_count = c_grid_desc_g_m_n.GetLength(I0);
const index_t grid_size_grp =
block_2_ctile_map.CalculateGridSize(c_grid_desc_m_n) * batch_count;
const index_t BlockEnd = grid_size_ + grid_size_grp;
// batch stride
const auto compute_base_ptr_of_batch = ComputeBasePtrOfStridedBatch(
a_grid_desc_g_m_k,
b_grid_desc_g_n_k,
b1_grid_desc_g_n_k,
c_grid_desc_g_m_n,
z_grid_desc_g_m_n,
type_convert<index_t>(lse_grid_desc_m.GetElementSpaceSize()));
// C0 mask
const auto c0_matrix_mask = C0MatrixMask(b_grid_desc_g_n_k.GetLength(I1));
grid_size_ += grid_size_grp;
// for each group, make sure acc0_biases_gs_ms_ns_lengths.size() == NumAcc0Bias and
// so on
if(!(problem_desc.acc0_biases_gs_ms_ns_lengths.size() == NumAcc0Bias &&
problem_desc.acc0_biases_gs_ms_ns_strides.size() == NumAcc0Bias &&
problem_desc.acc1_biases_gs_ms_os_lengths.size() == NumAcc1Bias &&
problem_desc.acc1_biases_gs_ms_os_strides.size() == NumAcc1Bias))
{
throw std::runtime_error(
"wrong! number of biases in function argument does not "
"match that in template argument");
}
group_kernel_args_.push_back({p_a_grid,
p_b_grid,
p_b1_grid,
p_c_grid,
p_z_grid,
p_lse_grid,
a_grid_desc_ak0_m_ak1,
b_grid_desc_bk0_n_bk1,
b1_grid_desc_bk0_n_bk1,
c_grid_desc_mblock_mperblock_nblock_nperblock,
z_grid_desc_m0_n0_m1_n1_m2_n2_m3_n3_n4_n5,
z_grid_desc_m_n,
lse_grid_desc_m,
block_2_ctile_map.CalculateGridSize(c_grid_desc_m_n),
compute_base_ptr_of_batch,
c0_matrix_mask,
block_2_ctile_map,
BlockStart,
BlockEnd});
group_device_args_.push_back(
{{problem_desc.a_gs_ms_ks_lengths[NumDimG + NumDimM - 1],
problem_desc.b0_gs_ns_ks_lengths[NumDimG + NumDimN - 1],
problem_desc.b0_gs_ns_ks_lengths[NumDimG + NumDimN + NumDimK - 1],
problem_desc.b1_gs_os_ns_lengths[NumDimG + NumDimO - 1]},
{problem_desc.a_gs_ms_ks_strides[NumDimG + NumDimM - 1],
problem_desc.a_gs_ms_ks_strides[NumDimG + NumDimM + NumDimK - 1]},
{problem_desc.b0_gs_ns_ks_strides[NumDimG + NumDimN - 1],
problem_desc.b0_gs_ns_ks_strides[NumDimG + NumDimN + NumDimK - 1]},
{problem_desc.b1_gs_os_ns_strides[NumDimG + NumDimO - 1],
problem_desc.b1_gs_os_ns_strides[NumDimG + NumDimO + NumDimN - 1]},
{problem_desc.c_gs_ms_os_strides[NumDimG + NumDimM - 1],
problem_desc.c_gs_ms_os_strides[NumDimG + NumDimM + NumDimO - 1]},
c_grid_desc_m_n});
}
is_dropout_ = p_dropout > 0.0; //
p_dropout_ = 1.f - p_dropout;
p_dropout_in_16bits_ = uint16_t(std::floor(p_dropout_ * 65535.0));
p_dropout_ = 1.f / p_dropout_;
p_dropout_rescale_ = type_convert<GemmAccDataType>(p_dropout_);
seed_ = std::get<0>(seeds);
offset_ = std::get<1>(seeds);
}
std::vector<GroupKernelArg> group_kernel_args_;
std::vector<GroupDeviceArg> group_device_args_;
std::size_t group_count_;
index_t grid_size_;
AElementwiseOperation a_element_op_;
BElementwiseOperation b_element_op_;
AccElementwiseOperation acc_element_op_;
B1ElementwiseOperation b1_element_op_;
CElementwiseOperation c_element_op_;
float p_dropout_;
ushort p_dropout_in_16bits_;
unsigned long long seed_;
unsigned long long offset_;
GemmAccDataType p_dropout_rescale_;
bool is_dropout_;
};
// Invoker
struct Invoker : public BaseInvoker
{
using Argument = DeviceOp::Argument;
float Run(const Argument& arg, const StreamConfig& stream_config = StreamConfig{})
{
if(!DeviceOp::IsSupportedArgument(arg))
{
throw std::runtime_error("wrong! unsupported argument");
}
bool all_has_main_k_block_loop = true;
bool some_has_main_k_block_loop = false;
for(std::size_t i = 0; i < arg.group_count_; i++)
{
const auto K = arg.group_kernel_args_[i].a_grid_desc_ak0_m_ak1_.GetLength(I0) *
arg.group_kernel_args_[i].a_grid_desc_ak0_m_ak1_.GetLength(I2);
const bool y = GridwiseGemm::CalculateHasMainKBlockLoop(K);
all_has_main_k_block_loop &= y;
some_has_main_k_block_loop |= y;
}
hipGetErrorString(hipMemcpy(arg.p_workspace_,
arg.group_kernel_args_.data(),
arg.group_kernel_args_.size() * sizeof(GroupKernelArg),
hipMemcpyHostToDevice));
float ave_time = 0;
auto launch_kernel = [&](auto has_main_k_block_loop_, auto is_dropout_) {
const auto kernel =
kernel_grouped_gemm_softmax_gemm_xdl_cshuffle_v2<GridwiseGemm,
GemmAccDataType,
GroupKernelArg,
AElementwiseOperation,
BElementwiseOperation,
AccElementwiseOperation,
B1ElementwiseOperation,
CElementwiseOperation,
has_main_k_block_loop_,
is_dropout_>;
return launch_and_time_kernel(
stream_config,
kernel,
dim3(arg.grid_size_),
dim3(BlockSize),
0,
cast_pointer_to_constant_address_space(arg.p_workspace_),
arg.group_count_,
arg.a_element_op_,
arg.b_element_op_,
arg.acc_element_op_,
arg.b1_element_op_,
arg.c_element_op_,
arg.p_dropout_in_16bits_,
arg.p_dropout_rescale_,
arg.seed_,
arg.offset_);
};
// Gemm1_K is split into Gemm1_K0/K1 where K1 is known at compile time, so we only need
// to concern Gemm0's loop
if(all_has_main_k_block_loop)
{
if(arg.is_dropout_)
{
ave_time = launch_kernel(integral_constant<bool, true>{},
integral_constant<bool, true>{});
}
else
{
ave_time = launch_kernel(integral_constant<bool, true>{},
integral_constant<bool, false>{});
}
}
else if(!some_has_main_k_block_loop)
{
if(arg.is_dropout_)
{
ave_time = launch_kernel(integral_constant<bool, false>{},
integral_constant<bool, true>{});
}
else
{
ave_time = launch_kernel(integral_constant<bool, false>{},
integral_constant<bool, false>{});
}
}
else
{
throw std::runtime_error("wrong! all gemm problems have to simultaneously meet "
"has_main_k_block_loop or no_main_k_block_loop");
}
return ave_time;
}
// polymorphic
float Run(const BaseArgument* p_arg,
const StreamConfig& stream_config = StreamConfig{}) override
{
return Run(*dynamic_cast<const Argument*>(p_arg), stream_config);
}
};
static constexpr bool IsValidCompilationParameter()
{
// TODO: properly implement this check
return true;
}
static bool IsSupportedArgument(const Argument& arg)
{
if(!(ck::get_device_name() == "gfx908" || ck::get_device_name() == "gfx90a"))
{
return false;
}
// TODO ANT: Check if tensor specialization & strides mismatch
bool all_has_main_k_block_loop = true;
bool some_has_main_k_block_loop = false;
for(std::size_t i = 0; i < arg.group_count_; i++)
{
const auto& kernel_arg = arg.group_kernel_args_[i];
const auto& device_arg = arg.group_device_args_[i];
// Check if C permute dimension matches GEMM + GEMM shape
const index_t c_m = device_arg.c_grid_desc_m_n_.GetLength(I0);
const index_t c_gemm1n = device_arg.c_grid_desc_m_n_.GetLength(I1);
const index_t a_m = kernel_arg.a_grid_desc_ak0_m_ak1_.GetLength(I1);
const index_t b1_gemm1n = kernel_arg.b1_grid_desc_bk0_n_bk1_.GetLength(I1);
if(!(c_m == a_m && c_gemm1n == b1_gemm1n))
{
return false;
}
// Check if having main loop
const auto K = kernel_arg.a_grid_desc_ak0_m_ak1_.GetLength(I0) *
kernel_arg.a_grid_desc_ak0_m_ak1_.GetLength(I2);
const bool y = GridwiseGemm::CalculateHasMainKBlockLoop(K);
all_has_main_k_block_loop &= y;
some_has_main_k_block_loop |= y;
// Note: we need raw lengths since threadwise copy can not handle vector load when
// part of vector is out of bounds
const auto MzRaw = device_arg.raw_lengths_mz_nz_kz_gemm1nz_[0];
const auto NzRaw = device_arg.raw_lengths_mz_nz_kz_gemm1nz_[1];
const auto KzRaw = device_arg.raw_lengths_mz_nz_kz_gemm1nz_[2];
const auto Gemm1NzRaw = device_arg.raw_lengths_mz_nz_kz_gemm1nz_[3];
// Check scalar per vector requirement
const auto a_extent_lowest = ABlockTransferSrcVectorDim == 2 ? KzRaw : MzRaw;
const auto b_extent_lowest = BBlockTransferSrcVectorDim == 2 ? KzRaw : NzRaw;
const auto b1_extent_lowest = B1BlockTransferSrcVectorDim == 2 ? NzRaw : Gemm1NzRaw;
const auto c_extent_lowest = Gemm1NzRaw;
if(!(a_extent_lowest % ABlockTransferSrcScalarPerVector == 0 &&
b_extent_lowest % BBlockTransferSrcScalarPerVector == 0 &&
b1_extent_lowest % B1BlockTransferSrcScalarPerVector == 0 &&
c_extent_lowest % CShuffleBlockTransferScalarPerVector_NPerBlock == 0))
{
return false;
}
// Check vector load/store requirement
const auto a_stride_lowest = ABlockTransferSrcVectorDim == 2
? device_arg.a_mz_kz_strides_[1]
: device_arg.a_mz_kz_strides_[0];
const auto b_stride_lowest = BBlockTransferSrcVectorDim == 2
? device_arg.b_nz_kz_strides_[1]
: device_arg.b_nz_kz_strides_[0];
const auto b1_stride_lowest = B1BlockTransferSrcVectorDim == 2
? device_arg.b1_nz_kz_strides_[1]
: device_arg.b1_nz_kz_strides_[0];
const auto c_stride_lowest =
device_arg.c_mz_gemm1nz_strides_[1]; // cshuffle assumes lowest dim in Gemm1Ns to be
// contiguous
if(!(a_stride_lowest == 1 || b_stride_lowest == 1 || b1_stride_lowest == 1 ||
c_stride_lowest == 1))
{
return false;
}
if(!GridwiseGemm::CheckValidity(kernel_arg.a_grid_desc_ak0_m_ak1_,
kernel_arg.b_grid_desc_bk0_n_bk1_,
kernel_arg.b1_grid_desc_bk0_n_bk1_,
device_arg.c_grid_desc_m_n_,
kernel_arg.block_2_ctile_map_))
{
return false;
}
}
// all gemm problems have to simultaneously meet has_main_k_block_loop or
// no_main_k_block_loop
if(!(all_has_main_k_block_loop || !some_has_main_k_block_loop))
{
return false;
}
return true;
}
// polymorphic
bool IsSupportedArgument(const BaseArgument* p_arg) override
{
return IsSupportedArgument(*dynamic_cast<const Argument*>(p_arg));
}
static auto MakeArgument(std::vector<const void*> p_a_vec,
std::vector<const void*> p_b_vec,
std::vector<const void*> p_b1_vec,
std::vector<void*> p_c_vec,
std::vector<void*> p_z_vec,
std::vector<void*> p_lse_vec,
std::vector<std::vector<const void*>> p_acc0_biases_vec,
std::vector<std::vector<const void*>> p_acc1_biases_vec,
std::vector<ProblemDesc> problem_desc_vec,
AElementwiseOperation a_element_op,
BElementwiseOperation b_element_op,
AccElementwiseOperation acc_element_op,
B1ElementwiseOperation b1_element_op,
CElementwiseOperation c_element_op,
float p_dropout,
std::tuple<unsigned long long, unsigned long long> seeds)
{
return Argument{p_a_vec,
p_b_vec,
p_b1_vec,
p_c_vec,
p_z_vec,
p_lse_vec,
p_acc0_biases_vec,
p_acc1_biases_vec,
problem_desc_vec,
a_element_op,
b_element_op,
acc_element_op,
b1_element_op,
c_element_op,
p_dropout,
seeds};
}
static auto MakeInvoker() { return Invoker{}; }
// polymorphic
std::unique_ptr<BaseArgument>
MakeArgumentPointer(std::vector<const void*> p_a_vec,
std::vector<const void*> p_b_vec,
std::vector<const void*> p_b1_vec,
std::vector<void*> p_c_vec,
std::vector<void*> p_z_vec,
std::vector<void*> p_lse_vec,
std::vector<std::vector<const void*>> p_acc0_biases_vec,
std::vector<std::vector<const void*>> p_acc1_biases_vec,
std::vector<ProblemDesc> problem_desc_vec,
AElementwiseOperation a_element_op,
BElementwiseOperation b_element_op,
AccElementwiseOperation acc_element_op,
B1ElementwiseOperation b1_element_op,
CElementwiseOperation c_element_op,
float p_dropout,
std::tuple<unsigned long long, unsigned long long> seeds) override
{
return std::make_unique<Argument>(p_a_vec,
p_b_vec,
p_b1_vec,
p_c_vec,
p_z_vec,
p_lse_vec,
p_acc0_biases_vec,
p_acc1_biases_vec,
problem_desc_vec,
a_element_op,
b_element_op,
acc_element_op,
b1_element_op,
c_element_op,
p_dropout,
seeds);
}
// polymorphic
std::unique_ptr<BaseInvoker> MakeInvokerPointer() override
{
return std::make_unique<Invoker>(Invoker{});
}
// polymorphic
std::string GetTypeString() const override
{
auto str = std::stringstream();
// clang-format off
str << "DeviceGroupedGemmSoftmaxGemmPermute_Train_Xdl_CShuffle"
<< "<"
<< BlockSize << ", "
<< MPerBlock << ", "
<< NPerBlock << ", "
<< KPerBlock << ", "
<< AK1 << ", "
<< BK1 << ", "
<< MPerBlock << ", "
<< Gemm1NPerBlock << ", "
<< Gemm1KPerBlock << ", "
<< B1K1 << ", "
<< getGemmSpecializationString(GemmSpec) << ", "
<< "ASpec" << getTensorSpecializationString(ASpec) << ", "
<< "B0Spec" << getTensorSpecializationString(BSpec) << ", "
<< "B1Spec" << getTensorSpecializationString(B1Spec) << ", "
<< "CSpec" << getTensorSpecializationString(CSpec) << ", "
<< getMaskingSpecializationString(MaskingSpec) << ">";
// clang-format on
return str.str();
}
size_t GetWorkSpaceSize(const BaseArgument* p_arg) const override
{
return dynamic_cast<const Argument*>(p_arg)->group_count_ * sizeof(GroupKernelArg);
}
};
} // namespace device
} // namespace tensor_operation
} // namespace ck
include/ck/tensor_operation/gpu/grid/gridwise_batched_multihead_attention_forward_xdl_cshuffle.hpp
View file @ bf80ceee
...@@ -35,6 +35,7 @@ template <typename FloatAB, ...@@ -35,6 +35,7 @@ template <typename FloatAB,
typename BGridDesc_BK0_N_BK1, typename BGridDesc_BK0_N_BK1,
typename B1GridDesc_BK0_N_BK1, typename B1GridDesc_BK0_N_BK1,
typename CGridDesc_M_N, typename CGridDesc_M_N,
typename ZGridDesc_M_N,
typename LSEGridDesc_M, typename LSEGridDesc_M,
index_t NumGemmKPrefetchStage, index_t NumGemmKPrefetchStage,
index_t BlockSize, index_t BlockSize,
...@@ -97,6 +98,8 @@ struct GridwiseBatchedMultiheadAttentionForward_Xdl_CShuffle ...@@ -97,6 +98,8 @@ struct GridwiseBatchedMultiheadAttentionForward_Xdl_CShuffle
static constexpr auto I6 = Number<6>{}; static constexpr auto I6 = Number<6>{};
static constexpr auto I7 = Number<7>{}; static constexpr auto I7 = Number<7>{};
static constexpr auto WaveSize = 64;
// K1 should be Number<...> // K1 should be Number<...>
// Gemm0 // Gemm0
static constexpr auto AK0 = Number<KPerBlock / AK1Value>{}; static constexpr auto AK0 = Number<KPerBlock / AK1Value>{};
...@@ -116,6 +119,65 @@ struct GridwiseBatchedMultiheadAttentionForward_Xdl_CShuffle ...@@ -116,6 +119,65 @@ struct GridwiseBatchedMultiheadAttentionForward_Xdl_CShuffle
using GridwiseGemmPipe = remove_cvref_t<decltype( using GridwiseGemmPipe = remove_cvref_t<decltype(
GridwiseGemmPipeline_Selector<PipelineVer, NumGemmKPrefetchStage>())>; GridwiseGemmPipeline_Selector<PipelineVer, NumGemmKPrefetchStage>())>;
// C desc for source in blockwise copy
__host__ __device__ static constexpr auto
MakeCGridDescriptor_M0_N0_M1_N1_M2_N2_M3_N3_N4_N5(const ZGridDesc_M_N& z_grid_desc_m_n) ////=> for z use
{
const auto M = z_grid_desc_m_n.GetLength(I0);
const auto N = z_grid_desc_m_n.GetLength(I1);
constexpr auto mfma = MfmaSelector<FloatAB, MPerXdl, NPerXdl>::selected_mfma;
constexpr auto N3 = mfma.num_groups_per_blk;
constexpr auto N4 = mfma.num_input_blks;
constexpr auto N5 = mfma.group_size;
return transform_tensor_descriptor(
z_grid_desc_m_n,
make_tuple(make_unmerge_transform(
make_tuple(M / MPerBlock, MXdlPerWave, Gemm0MWaves, MPerXdl)),
make_unmerge_transform(
make_tuple(N / NPerBlock, NXdlPerWave, Gemm0NWaves, N3, N4, N5))),
make_tuple(Sequence<0>{}, Sequence<1>{}),
make_tuple(Sequence<0, 2, 4, 6>{}, Sequence<1, 3, 5, 7, 8, 9>{}));
}
__host__ __device__ static constexpr auto
MakeZGridDescriptor_M0_N0_M1_N1_M2_N2_M3_N3_N4_N5(const index_t M, const index_t N) ////=> for z use
{
constexpr auto mfma = MfmaSelector<FloatAB, MPerXdl, NPerXdl>::selected_mfma;
constexpr auto N3 = mfma.num_groups_per_blk;
constexpr auto N4 = mfma.num_input_blks;
constexpr auto N5 = mfma.group_size;
return transform_tensor_descriptor(
make_naive_tensor_descriptor_packed(make_tuple(M, N)),
make_tuple(make_unmerge_transform(
make_tuple(M / MPerBlock, MXdlPerWave, Gemm0MWaves, MPerXdl)),
make_unmerge_transform(
make_tuple(N / NPerBlock, NXdlPerWave, Gemm0NWaves, N3, N4, N5))),
make_tuple(Sequence<0>{}, Sequence<1>{}),
make_tuple(Sequence<0, 2, 4, 6>{}, Sequence<1, 3, 5, 7, 8, 9>{}));
}
__device__ static auto GetGemm0WaveIdx()
{
const index_t thread_id = get_thread_local_1d_id();
constexpr auto threadid_to_wave_idx_adaptor = make_single_stage_tensor_adaptor(
make_tuple(make_merge_transform(make_tuple(Gemm0MWaves, Gemm0NWaves, WaveSize))),
make_tuple(Sequence<0, 1, 2>{}),
make_tuple(Sequence<0>{}));
return threadid_to_wave_idx_adaptor.CalculateBottomIndex(make_multi_index(thread_id));
}
__device__ static auto GetGemm0WaveMNIdx(const index_t thread_id)
{
constexpr auto wave_threadid_to_mn_idx_adaptor = make_single_stage_tensor_adaptor(
make_tuple(make_merge_transform(make_tuple(WaveSize / MPerXdl, MPerXdl))),
make_tuple(Sequence<0, 1>{}),
make_tuple(Sequence<0>{}));
return wave_threadid_to_mn_idx_adaptor.CalculateBottomIndex(make_multi_index(thread_id));
}
template <typename ABlockDesc_AK0_M_AK1> template <typename ABlockDesc_AK0_M_AK1>
__host__ __device__ static constexpr auto __host__ __device__ static constexpr auto
MakeGemm0AMmaTileDescriptor_M0_M1_M2_K(const ABlockDesc_AK0_M_AK1&) MakeGemm0AMmaTileDescriptor_M0_M1_M2_K(const ABlockDesc_AK0_M_AK1&)
...@@ -323,6 +385,9 @@ struct GridwiseBatchedMultiheadAttentionForward_Xdl_CShuffle ...@@ -323,6 +385,9 @@ struct GridwiseBatchedMultiheadAttentionForward_Xdl_CShuffle
using DefaultBlock2CTileMap = using DefaultBlock2CTileMap =
remove_cvref_t<decltype(MakeDefaultBlock2CTileMap(CGridDesc_M_N{}))>; remove_cvref_t<decltype(MakeDefaultBlock2CTileMap(CGridDesc_M_N{}))>;
using ZGridDescriptor_M0_N0_M1_N1_M2_N2_M3_N3_N4_N5 = remove_cvref_t<decltype(
MakeCGridDescriptor_M0_N0_M1_N1_M2_N2_M3_N3_N4_N5(ZGridDesc_M_N{}))>;
struct SharedMemTrait struct SharedMemTrait
{ {
// LDS allocation for A and B: be careful of alignment // LDS allocation for A and B: be careful of alignment
...@@ -367,6 +432,7 @@ struct GridwiseBatchedMultiheadAttentionForward_Xdl_CShuffle ...@@ -367,6 +432,7 @@ struct GridwiseBatchedMultiheadAttentionForward_Xdl_CShuffle
const FloatAB* __restrict__ p_b_grid, const FloatAB* __restrict__ p_b_grid,
const FloatAB* __restrict__ p_b1_grid, const FloatAB* __restrict__ p_b1_grid,
FloatC* __restrict__ p_c_grid, FloatC* __restrict__ p_c_grid,
unsigned short* __restrict__ p_z_grid,
FloatLSE* __restrict__ p_lse_grid, FloatLSE* __restrict__ p_lse_grid,
void* __restrict__ p_shared, void* __restrict__ p_shared,
const AElementwiseOperation& a_element_op, const AElementwiseOperation& a_element_op,
...@@ -379,6 +445,8 @@ struct GridwiseBatchedMultiheadAttentionForward_Xdl_CShuffle ...@@ -379,6 +445,8 @@ struct GridwiseBatchedMultiheadAttentionForward_Xdl_CShuffle
const B1GridDesc_BK0_N_BK1& b1_grid_desc_bk0_n_bk1, const B1GridDesc_BK0_N_BK1& b1_grid_desc_bk0_n_bk1,
const CGridDescriptor_MBlock_MPerBlock_NBlock_NPerBlock& const CGridDescriptor_MBlock_MPerBlock_NBlock_NPerBlock&
c_grid_desc_mblock_mperblock_nblock_nperblock, c_grid_desc_mblock_mperblock_nblock_nperblock,
const ZGridDescriptor_M0_N0_M1_N1_M2_N2_M3_N3_N4_N5&
z_grid_desc_m0_n0_m1_n1_m2_n2_m3_n3_n4_n5,
const LSEGridDesc_M& lse_grid_desc_m, const LSEGridDesc_M& lse_grid_desc_m,
const Block2CTileMap& block_2_ctile_map, const Block2CTileMap& block_2_ctile_map,
const C0MatrixMask& c0_matrix_mask, const C0MatrixMask& c0_matrix_mask,
...@@ -782,6 +850,79 @@ struct GridwiseBatchedMultiheadAttentionForward_Xdl_CShuffle ...@@ -782,6 +850,79 @@ struct GridwiseBatchedMultiheadAttentionForward_Xdl_CShuffle
// gemm1 K loop // gemm1 K loop
index_t gemm1_k_block_outer_index = 0; index_t gemm1_k_block_outer_index = 0;
///////////////////=>z for dropout
//
// z vgpr copy to global
//
// z matrix threadwise desc
constexpr auto z_thread_desc_m0_n0_m1_n1_m2_n2_m3_n3_n4_n5 =
make_naive_tensor_descriptor_packed(make_tuple(I1, // MBlockId
I1, // NBlockID
m0, // MRepeat
n0, // NRepeat
m1, // MWaveId
n1, // NWaveId
m2, // MPerXdl
n2, // NGroupNum
n3, // NInputNum
n4)); // registerNum
StaticBuffer<AddressSpaceEnum::Vgpr,
unsigned short,
z_thread_desc_m0_n0_m1_n1_m2_n2_m3_n3_n4_n5.GetElementSpaceSize(),
true>
z_tenor_buffer;
z_tenor_buffer.Clear();
// z matrix global desc
/*const auto M = q_grid_desc_k0_m_k1.GetLength(I1);
const auto N = k_grid_desc_k0_n_k1.GetLength(I1);
auto z_grid_desc_m0_n0_m1_n1_m2_n2_m3_n3_n4_n5 =
MakeZGridDescriptor_M0_N0_M1_N1_M2_N2_M3_N3_N4_N5(M, N);*/
auto z_grid_buf = make_dynamic_buffer<AddressSpaceEnum::Global>(
p_z_grid, z_grid_desc_m0_n0_m1_n1_m2_n2_m3_n3_n4_n5.GetElementSpaceSize());
const auto wave_id = GetGemm0WaveIdx();
const auto wave_m_n_id = GetGemm0WaveMNIdx(wave_id[I2]); // I2: 0~63
auto z_thread_copy_vgpr_to_global = ThreadwiseTensorSliceTransfer_v1r3<
ushort,
ushort,
decltype(z_thread_desc_m0_n0_m1_n1_m2_n2_m3_n3_n4_n5),
decltype(z_grid_desc_m0_n0_m1_n1_m2_n2_m3_n3_n4_n5),
tensor_operation::element_wise::PassThrough,
Sequence<I1, // MBlockId
I1, // NBlockID
m0, // MRepeat
n0, // NRepeat
m1, // MWaveId
n1, // NWaveId
m2, // MPerXdl
n2, // NGroupNum
n3, // NInputNum
n4>,
Sequence<0, 1, 2, 3, 4, 5, 6, 7, 8, 9>,
9, // DstVectorDim
n4, // DstScalarPerVector
InMemoryDataOperationEnum::Set,
1, // DstScalarStrideInVector
true>{z_grid_desc_m0_n0_m1_n1_m2_n2_m3_n3_n4_n5,
make_multi_index(block_work_idx[I0], // MBlockId
0, // NBlockId
0, // mrepeat
0, // nrepeat
wave_id[I0], // MWaveId
wave_id[I1], // NWaveId
wave_m_n_id[I1], // MPerXdl
0, // group
wave_m_n_id[I0], // NInputIndex
0),
tensor_operation::element_wise::PassThrough{}};
///////////////////=>z for dropout
do do
{ {
auto n_block_data_idx_on_grid = auto n_block_data_idx_on_grid =
...@@ -876,9 +1017,35 @@ struct GridwiseBatchedMultiheadAttentionForward_Xdl_CShuffle ...@@ -876,9 +1017,35 @@ struct GridwiseBatchedMultiheadAttentionForward_Xdl_CShuffle
if constexpr(IsDropout) // dropout if constexpr(IsDropout) // dropout
{ {
blockwise_dropout.ApplyDropout(acc_thread_buf, ph);
// save z to global
if(p_z_grid)
{
// P_dropped
blockwise_dropout.template ApplyDropout<decltype(acc_thread_buf),
decltype(z_tenor_buffer),
true>(
acc_thread_buf, ph, z_tenor_buffer);
z_thread_copy_vgpr_to_global.Run(z_thread_desc_m0_n0_m1_n1_m2_n2_m3_n3_n4_n5,
make_tuple(I0, I0, I0, I0, I0, I0, I0, I0, I0, I0),
z_tenor_buffer,
z_grid_desc_m0_n0_m1_n1_m2_n2_m3_n3_n4_n5,
z_grid_buf);
}
else
{
// P_dropped
blockwise_dropout.template ApplyDropout<decltype(acc_thread_buf), true>(
acc_thread_buf, ph);
}
} }
//if constexpr(IsDropout) // dropout
//{
// blockwise_dropout.ApplyDropout(acc_thread_buf, ph);
//}
// TODO: may convert to log domain // TODO: may convert to log domain
running_max_new = mathext::max(max, running_max); running_max_new = mathext::max(max, running_max);
running_sum_new = mathext::exp(running_max - running_max_new) * running_sum + running_sum_new = mathext::exp(running_max - running_max_new) * running_sum +
......
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