Skip to content

GitLab

Commit 0eb75e21 authored by carlushuang's avatar carlushuang
Browse files

Merge remote-tracking branch 'origin/develop' into ck_tile/moe

parents 1b4b640b c8b6b642
Expand all Hide whitespace changes
Inline Side-by-side
Showing with 707 additions and 261 deletions
codegen/test/rtc/src/tmp_dir.cpp
View file @ 0eb75e21
......@@ -45,4 +45,4 @@ void tmp_dir::execute(const std::string& cmd) const
tmp_dir::~tmp_dir() { std::filesystem::remove_all(this->path); }
} // namespace rtc
\ No newline at end of file
} // namespace rtc
docs/sphinx/requirements.in
View file @ 0eb75e21
rocm-docs-core==1.6.1
rocm-docs-core==1.7.0
sphinxcontrib-bibtex==2.6.2
docs/sphinx/requirements.txt
View file @ 0eb75e21
......@@ -103,7 +103,7 @@ requests==2.32.3
# via
# pygithub
# sphinx
rocm-docs-core==1.6.1
rocm-docs-core==1.7.0
# via -r requirements.in
six==1.16.0
# via pybtex
......
example/10_convnd_fwd_multiple_d_multiple_reduce/common.hpp
View file @ 0eb75e21
// SPDX-License-Identifier: MIT
// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
#include <algorithm>
#include <cassert>
......@@ -139,7 +139,7 @@ inline bool parse_cmd_args(int argc,
inline HostTensorDescriptor
make_r0_host_tensor_descriptor(const ck::utils::conv::ConvParam& problem_size)
{
std::vector<ck::index_t> dimensions{problem_size.G_, problem_size.N_};
std::vector<ck::long_index_t> dimensions{problem_size.G_, problem_size.N_};
ck::ranges::copy(problem_size.output_spatial_lengths_, std::back_inserter(dimensions));
......
example/12_reduce/reduce_blockwise.cpp
View file @ 0eb75e21
// SPDX-License-Identifier: MIT
// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
#include <iostream>
#include <initializer_list>
......@@ -255,34 +255,61 @@ int main(int argc, char* argv[])
else
{
// for testing half_t
pass =
pass && reduce_blockwise_test<ck::half_t, float, ReduceOpId, PropagateNan, OutputIndex>(
true, 2, true, {3, 3, 3, 3, 3, 3, 3, 3, 3, 3, 3, 3}, {0, 1, 2}, 1.0f, 0.0f);
pass =
pass && reduce_blockwise_test<ck::half_t, float, ReduceOpId, PropagateNan, OutputIndex>(
true, 2, true, {16, 64, 32, 960}, {0, 1, 2}, 1.0f, 0.0f);
// for testing float
pass =
pass && reduce_blockwise_test<float, float, ReduceOpId, PropagateNan, OutputIndex>(
true, 2, true, {3, 3, 3, 3, 3, 3, 3, 3, 3, 3, 3, 3}, {0, 1, 2}, 1.0f, 0.0f);
pass = pass && reduce_blockwise_test<float, float, ReduceOpId, PropagateNan, OutputIndex>(
true, 2, true, {16, 64, 32, 960}, {0, 1, 2}, 1.0f, 0.0f);
// for testing double
pass =
pass && reduce_blockwise_test<float, float, ReduceOpId, PropagateNan, OutputIndex>(
true, 2, true, {3, 3, 3, 3, 3, 3, 3, 3, 3, 3, 3, 3}, {0, 1, 2}, 1.0f, 0.0f);
pass = pass && reduce_blockwise_test<float, float, ReduceOpId, PropagateNan, OutputIndex>(
true, 2, true, {16, 64, 32, 960}, {0, 1, 2}, 1.0f, 0.0f);
// for testing bhalf_t
pass = pass &&
reduce_blockwise_test<ck::bhalf_t, float, ReduceOpId, PropagateNan, OutputIndex>(
true, 2, true, {3, 3, 3, 3, 3, 3, 3, 3, 3, 3, 3, 3}, {0, 1, 2}, 1.0f, 0.0f);
pass = pass &&
reduce_blockwise_test<ck::bhalf_t, float, ReduceOpId, PropagateNan, OutputIndex>(
true, 2, true, {16, 64, 32, 960}, {0, 1, 2}, 1.0f, 0.0f);
// for testing int8_t
pass =
pass && reduce_blockwise_test<int8_t, int32_t, ReduceOpId, PropagateNan, OutputIndex>(
true, 2, true, {3, 3, 3, 3, 3, 3, 3, 3, 3, 3, 3, 3}, {0, 1, 2}, 1.0f, 0.0f);
pass =
pass && reduce_blockwise_test<int8_t, int32_t, ReduceOpId, PropagateNan, OutputIndex>(
true, 2, true, {16, 64, 32, 960}, {0, 1, 2}, 1.0f, 0.0f);
#ifdef CK_EXPERIMENTAL_BIT_INT_EXTENSION_INT4
// for testing int4_t using AVG operation
pass =
pass && reduce_blockwise_test<int4_t, int32_t, ReduceTensorOp::AVG, false, false>(
true, 2, true, {3, 3, 3, 3, 3, 3, 3, 3, 3, 3, 3, 3}, {0, 1, 2}, 1.0f, 0.0f);
pass = pass && reduce_blockwise_test<int4_t, int32_t, ReduceTensorOp::AVG, false, false>(
true, 2, true, {16, 64, 32, 960}, {0, 1, 2}, 1.0f, 0.0f);
// for testing int4_t using MAX operation
pass =
pass && reduce_blockwise_test<int4_t, int8_t, ReduceTensorOp::MAX, false, false>(
true, 2, true, {3, 3, 3, 3, 3, 3, 3, 3, 3, 3, 3, 3}, {0, 1, 2}, 1.0f, 0.0f);
pass = pass && reduce_blockwise_test<int4_t, int8_t, ReduceTensorOp::MAX, false, false>(
true, 2, true, {16, 64, 32, 960}, {0, 1, 2}, 1.0f, 0.0f);
#endif
......
example/12_reduce/reduce_blockwise_impl.hpp
View file @ 0eb75e21
// SPDX-License-Identifier: MIT
// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
#pragma once
......@@ -316,7 +316,17 @@ int reduce_blockwise_impl(bool do_verification,
auto invoker_ptr = reduce.MakeInvokerPointer();
float avg_time = invoker_ptr->Run(argument_ptr.get(), StreamConfig{nullptr, time_kernel});
int log_level = 0, cold_niters = 5, nrepeat = 50;
if(beta != 0.0f)
{
std::cerr << "Warning: With beta != 0.0f there must be only one repeat for correct results "
"since out memory is being overwritten."
<< std::endl;
cold_niters = 0;
nrepeat = 1;
}
float avg_time = invoker_ptr->Run(
argument_ptr.get(), StreamConfig{nullptr, time_kernel, log_level, cold_niters, nrepeat});
std::size_t num_bytes = invariant_total_length * reduce_total_length * sizeof(InOutDataType) +
invariant_total_length * sizeof(InOutDataType);
......
example/12_reduce/reduce_example_common.hpp
View file @ 0eb75e21
// SPDX-License-Identifier: MIT
// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
#pragma once
......@@ -38,7 +38,8 @@ struct ReduceShape
static constexpr ck::index_t NumReduceDim_ = NumReduceDim;
};
using reduce_shape_instances = std::tuple<ReduceShape<3, 1>,
using reduce_shape_instances = std::tuple<ReduceShape<12, 3>,
ReduceShape<3, 1>,
ReduceShape<3, 2>,
ReduceShape<4, 1>,
ReduceShape<4, 2>,
......
example/17_convnd_bwd_data/convnd_bwd_data_common.hpp
View file @ 0eb75e21
// SPDX-License-Identifier: MIT
// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
#include <iostream>
#include <numeric>
......@@ -80,6 +80,29 @@ int run_conv_bwd_data(bool do_verification,
// reset input to zero
in_device_buf.SetZero();
std::vector<ck::index_t> input_spatial_lengths_i32(NDimSpatial);
std::vector<ck::index_t> filter_spatial_lengths_i32(NDimSpatial);
std::vector<ck::index_t> output_spatial_lengths_i32(NDimSpatial);
std::vector<ck::index_t> conv_filter_strides_i32(NDimSpatial);
std::vector<ck::index_t> conv_filter_dilations_i32(NDimSpatial);
std::vector<ck::index_t> input_left_pads_i32(NDimSpatial);
std::vector<ck::index_t> input_right_pads_i32(NDimSpatial);
for(ck::index_t d = 0; d < NDimSpatial; d++)
{
input_spatial_lengths_i32[d] =
static_cast<ck::index_t>(conv_param.input_spatial_lengths_[d]);
filter_spatial_lengths_i32[d] =
static_cast<ck::index_t>(conv_param.filter_spatial_lengths_[d]);
output_spatial_lengths_i32[d] =
static_cast<ck::index_t>(conv_param.GetOutputSpatialLengths()[d]);
conv_filter_strides_i32[d] = static_cast<ck::index_t>(conv_param.conv_filter_strides_[d]);
conv_filter_dilations_i32[d] =
static_cast<ck::index_t>(conv_param.conv_filter_dilations_[d]);
input_left_pads_i32[d] = static_cast<ck::index_t>(conv_param.input_left_pads_[d]);
input_right_pads_i32[d] = static_cast<ck::index_t>(conv_param.input_right_pads_[d]);
}
// do GEMM
auto conv = DeviceConvNdBwdDataInstance{};
auto invoker = conv.MakeInvoker();
......@@ -87,16 +110,16 @@ int run_conv_bwd_data(bool do_verification,
conv.MakeArgumentPointer(static_cast<InDataType*>(in_device_buf.GetDeviceBuffer()),
static_cast<WeiDataType*>(wei_device_buf.GetDeviceBuffer()),
static_cast<OutDataType*>(out_device_buf.GetDeviceBuffer()),
conv_param.N_,
conv_param.K_,
conv_param.C_,
conv_param.input_spatial_lengths_,
conv_param.filter_spatial_lengths_,
conv_param.GetOutputSpatialLengths(),
conv_param.conv_filter_strides_,
conv_param.conv_filter_dilations_,
conv_param.input_left_pads_,
conv_param.input_right_pads_,
static_cast<ck::index_t>(conv_param.N_),
static_cast<ck::index_t>(conv_param.K_),
static_cast<ck::index_t>(conv_param.C_),
input_spatial_lengths_i32,
filter_spatial_lengths_i32,
output_spatial_lengths_i32,
conv_filter_strides_i32,
conv_filter_dilations_i32,
input_left_pads_i32,
input_right_pads_i32,
in_element_op,
wei_element_op,
out_element_op);
......
example/20_grouped_conv_bwd_weight/common.hpp
View file @ 0eb75e21
......@@ -23,12 +23,8 @@
using BF16 = ck::bhalf_t;
using F16 = ck::half_t;
using F32 = float;
#ifdef CK_ENABLE_FP8
using F8 = ck::f8_t;
#endif
#ifdef CK_ENABLE_BF8
using BF8 = ck::bf8_t;
#endif
using F8 = ck::f8_t;
using BF8 = ck::bf8_t;
template <ck::index_t... Is>
using S = ck::Sequence<Is...>;
......
example/65_gemm_multiply_multiply/gemm_add_add_xdl_fp16.cpp
View file @ 0eb75e21
......@@ -208,6 +208,7 @@ int main(int argc, char* argv[])
StrideB,
std::array<ck::index_t, NumDTensor>{StrideD, StrideD},
StrideE,
1,
a_element_op,
b_element_op,
cde_element_op);
......
example/65_gemm_multiply_multiply/gemm_multiply_multiply_xdl_fp8.cpp
View file @ 0eb75e21
......@@ -69,7 +69,7 @@ using AElementOp = PassThrough;
using BElementOp = PassThrough;
using CDEElementOp = MultiplyMultiply;
static constexpr auto GemmSpec = ck::tensor_operation::device::GemmSpecialization::Default;
static constexpr auto GemmSpec = ck::tensor_operation::device::GemmSpecialization::MNPadding;
using DeviceOpInstance = ck::tensor_operation::device::DeviceGemmMultiD_Xdl_CShuffle_V3
// clang-format off
......@@ -99,6 +99,8 @@ int main(int argc, char* argv[])
ck::index_t StrideD = 0;
ck::index_t StrideE = N;
ck::index_t KBatch = 1;
if(argc == 1)
{
// use default case
......@@ -109,7 +111,7 @@ int main(int argc, char* argv[])
init_method = std::stoi(argv[2]);
time_kernel = std::stoi(argv[3]);
}
else if(argc == 11)
else if(argc == 12)
{
do_verification = std::stoi(argv[1]);
init_method = std::stoi(argv[2]);
......@@ -123,13 +125,16 @@ int main(int argc, char* argv[])
StrideB = std::stoi(argv[8]);
StrideD = std::stoi(argv[9]);
StrideE = std::stoi(argv[10]);
KBatch = std::stoi(argv[11]);
}
else
{
printf("arg1: verification (0=no, 1=yes)\n");
printf("arg2: initialization (0=no init, 1=integer value, 2=decimal value)\n");
printf("arg3: time kernel (0=no, 1=yes)\n");
printf("arg4 to 9: M (256x), N(128x), K(32x), StrideA, StrideB, StrideD, StrideE\n");
printf(
"arg4 to 9: M (256x), N(128x), K(32x), StrideA, StrideB, StrideD, StrideE, KBatch\n");
exit(0);
}
......@@ -212,6 +217,7 @@ int main(int argc, char* argv[])
StrideB,
std::array<ck::index_t, NumDTensor>{I0, I0},
StrideE,
KBatch,
a_element_op,
b_element_op,
cde_element_op);
......@@ -236,10 +242,12 @@ int main(int argc, char* argv[])
std::cout << "Perf: " << ave_time << " ms, " << tflops << " TFlops, " << gb_per_sec << " GB/s"
<< std::endl;
e_device_buf.FromDevice(e_m_n_device_result.mData.data());
if(do_verification)
{
invoker.Run(argument, StreamConfig{nullptr, false});
e_device_buf.FromDevice(e_m_n_device_result.mData.data());
Tensor<CShuffleDataType> c_m_n({M, N});
using ReferenceGemmInstance = ck::tensor_operation::host::ReferenceGemm<A0DataType,
......
example/CMakeLists.txt
View file @ 0eb75e21
......@@ -72,10 +72,24 @@ function(add_example_executable EXAMPLE_NAME FILE_NAME)
list(REMOVE_ITEM FILE_NAME "${source}")
endif()
endforeach()
#Do not build any FP8 examples if CK_ENABLE_FP8 not set
foreach(source IN LISTS FILE_NAME)
if(NOT DEFINED CK_ENABLE_FP8 AND source MATCHES "_fp8")
message("removing fp8 example ${source} ")
list(REMOVE_ITEM FILE_NAME "${source}")
endif()
endforeach()
#Do not build any BF8 examples if CK_ENABLE_BF8 not set
foreach(source IN LISTS FILE_NAME)
if(NOT DEFINED CK_ENABLE_BF8 AND source MATCHES "_bf8")
message("removing bf8 example ${source} ")
list(REMOVE_ITEM FILE_NAME "${source}")
endif()
endforeach()
#only continue if there are some source files left on the list
if(FILE_NAME)
if(FILE_NAME MATCHES "_xdl")
list(REMOVE_ITEM EX_TARGETS gfx1030 gfx1100 gfx1101 gfx1102 gfx1103)
list(REMOVE_ITEM EX_TARGETS gfx1030 gfx1100 gfx1101 gfx1102 gfx1103 gfx1200 gfx1201)
elseif(FILE_NAME MATCHES "_wmma")
list(REMOVE_ITEM EX_TARGETS gfx908 gfx90a gfx940 gfx941 gfx942 gfx1030)
endif()
......@@ -162,7 +176,7 @@ function(add_example_executable_no_testing EXAMPLE_NAME FILE_NAME)
#only continue if there are some source files left on the list
if(FILE_NAME)
if(FILE_NAME MATCHES "_xdl")
list(REMOVE_ITEM EX_TARGETS gfx900 gfx906 gfx1030 gfx1100 gfx1101 gfx1102 gfx1103)
list(REMOVE_ITEM EX_TARGETS gfx900 gfx906 gfx1030 gfx1100 gfx1101 gfx1102 gfx1103 gfx1200 gfx1201)
elseif(FILE_NAME MATCHES "_wmma")
list(REMOVE_ITEM EX_TARGETS gfx900 gfx906 gfx908 gfx90a gfx940 gfx941 gfx942 gfx1030)
endif()
......
example/ck_tile/01_fmha/CMakeLists.txt
View file @ 0eb75e21
......@@ -6,7 +6,7 @@ execute_process(
execute_process(
COMMAND ${Python3_EXECUTABLE} ${CMAKE_CURRENT_LIST_DIR}/generate.py
--api bwd --list_blobs ${CMAKE_CURRENT_BINARY_DIR}/bwd_blob_list.txt
--api bwd --list_blobs ${CMAKE_CURRENT_BINARY_DIR}/bwd_blob_list.txt --receipt 3
)
# NOTE: for cmake, the FMHA_FWD_GEN_BLOBS/FMHA_BWD_GEN_BLOBS files must be in the same directory
......@@ -23,7 +23,7 @@ add_custom_command(
add_custom_command(
OUTPUT ${FMHA_BWD_GEN_BLOBS}
COMMAND ${Python3_EXECUTABLE} ${CMAKE_CURRENT_LIST_DIR}/generate.py
--api bwd --output_dir ${CMAKE_CURRENT_BINARY_DIR}
--api bwd --output_dir ${CMAKE_CURRENT_BINARY_DIR} --receipt 3
)
set(EXAMPLE_FMHA_FWD "tile_example_fmha_fwd")
......@@ -55,11 +55,10 @@ set(EXAMPLE_FMHA_BWD_COMPILE_OPTIONS)
# ... because they are auto-generated
if(FMHA_FWD_FAST_EXP2)
list(APPEND EXAMPLE_FMHA_FWD_COMPILE_OPTIONS -Wno-undefined-func-template -DCK_TILE_FMHA_FWD_FAST_EXP2=1 -fgpu-flush-denormals-to-zero)
list(APPEND EXAMPLE_FMHA_BWD_COMPILE_OPTIONS -Wno-undefined-func-template -DCK_TILE_FMHA_FWD_FAST_EXP2=1 -fgpu-flush-denormals-to-zero)
else()
list(APPEND EXAMPLE_FMHA_FWD_COMPILE_OPTIONS -Wno-undefined-func-template -DCK_TILE_FMHA_FWD_FAST_EXP2=0)
list(APPEND EXAMPLE_FMHA_BWD_COMPILE_OPTIONS -Wno-undefined-func-template -DCK_TILE_FMHA_FWD_FAST_EXP2=0)
endif()
list(APPEND EXAMPLE_FMHA_BWD_COMPILE_OPTIONS -Wno-undefined-func-template -fgpu-flush-denormals-to-zero)
# Allow comparing floating points directly in order to check sentinel values
list(APPEND EXAMPLE_FMHA_FWD_COMPILE_OPTIONS -Wno-float-equal)
......
example/ck_tile/01_fmha/codegen/cpp_symbol_map.py
View file @ 0eb75e21
......@@ -66,6 +66,22 @@ BIAS_CHECK_MAP = {
"alibi" : "bias_enum::alibi"
}
DROPOUT_MAP = {
"no" : "ck_tile::BlockDropoutBwd<false, true, false>",
"dropout_wg32" : "ck_tile::BlockDropoutBwd<true, true, false>",
"dropout_wg32_storerandval" : "ck_tile::BlockDropoutBwd<true, true, true >",
"dropout_wg16" : "ck_tile::BlockDropoutBwd<true, false, false>",
"dropout_wg16_storerandval" : "ck_tile::BlockDropoutBwd<true, false, true >"
}
DROPOUT_CHECK_MAP = {
"no" : "t.has_dropout == false",
"dropout_wg32" : "t.has_dropout == true && t.is_store_randval == false",
"dropout_wg32_storerandval" : "t.has_dropout == true && t.is_store_randval == true",
"dropout_wg16" : "t.has_dropout == true && t.is_store_randval == false",
"dropout_wg16_storerandval" : "t.has_dropout == true && t.is_store_randval == true",
}
MODE_MAP = {
"batch" : "false",
"group" : "true"
......
example/ck_tile/01_fmha/fmha_bwd.cpp
View file @ 0eb75e21
......@@ -87,7 +87,11 @@ auto create_args(int argc, char* argv[])
.insert("drop_offset", "0", "offset for random number generator")
.insert("timer", "gpu", "gpu:gpu timer, cpu:cpu timer")
.insert("warmup", "5", "number of iterations before benchmark the kernel")
.insert("repeat", "20", "number of iterations to benchmark the kernel");
.insert("repeat", "20", "number of iterations to benchmark the kernel")
.insert("deterministic",
"0",
"if set to 1 will use multi-buffer reduction strategy for dq, atomic opeartion "
"will not be used");
bool result = arg_parser.parse(argc, argv);
return std::make_tuple(result, arg_parser);
......@@ -128,11 +132,6 @@ bool run(const ck_tile::ArgParser& arg_parser)
ck_tile::index_t hdim_v = arg_parser.get_int("d_v");
if(hdim_v < 0)
hdim_v = hdim_q;
if(hdim_q % 2 != 0 || hdim_v % 2 != 0)
{
std::cerr << "FMHA Bwd kernel currently only supports even headdim" << std::endl;
return false;
}
bool i_perm = arg_parser.get_bool("iperm"); // if true, will be batch * nhead * seqlen * hdim
bool o_perm = arg_parser.get_bool("operm"); // if false, will be batch * seqlen * nhead * hdim
......@@ -177,9 +176,10 @@ bool run(const ck_tile::ArgParser& arg_parser)
seed.reset();
}
int stream_warmup = arg_parser.get_int("warmup");
int stream_repeat = arg_parser.get_int("repeat");
bool kname = arg_parser.get_bool("kname");
int stream_warmup = arg_parser.get_int("warmup");
int stream_repeat = arg_parser.get_int("repeat");
bool kname = arg_parser.get_bool("kname");
bool deterministic = arg_parser.get_bool("deterministic");
ck_tile::stream_config stream_config{nullptr,
true,
......@@ -265,6 +265,9 @@ bool run(const ck_tile::ArgParser& arg_parser)
(mode == mode_enum::batch ? seqlen_q : seqstart_q_host.back());
const ck_tile::index_t shape_seqlen_k =
(mode == mode_enum::batch ? seqlen_k : seqstart_k_host.back());
const ck_tile::index_t kN0 = (hdim_q <= 128) ? 128 : 64;
const ck_tile::index_t nsplits =
deterministic ? ck_tile::integer_divide_ceil(max_seqlen_k, kN0) : 1;
ck_tile::HostTensor<QDataType> q_host(
get_lengths(i_perm, shape_batch, nhead, shape_seqlen_q, hdim_q));
......@@ -284,9 +287,9 @@ bool run(const ck_tile::ArgParser& arg_parser)
ck_tile::HostTensor<ODataType> o_host(
get_lengths(o_perm, shape_batch, nhead, shape_seqlen_q, hdim_v));
ck_tile::HostTensor<LSEDataType> lse_host(
std::array<ck_tile::index_t, 3>{batch, nhead, max_seqlen_q});
std::array<ck_tile::index_t, 3>{shape_batch, nhead, shape_seqlen_q});
ck_tile::HostTensor<DDataType> d_host(
std::array<ck_tile::index_t, 3>{batch, nhead, max_seqlen_q});
std::array<ck_tile::index_t, 3>{shape_batch, nhead, shape_seqlen_q});
ck_tile::HostTensor<RandValOutputDataType> randval_host(
p_drop > 0 ? get_lengths(true, shape_batch, nhead, shape_seqlen_q, max_seqlen_k)
: std::array<ck_tile::index_t, 4>{1, 1, 1, 1});
......@@ -302,6 +305,10 @@ bool run(const ck_tile::ArgParser& arg_parser)
use_dbias
? get_lengths(i_perm, shape_batch, nhead, shape_seqlen_q, max_seqlen_k)
: std::array<ck_tile::index_t, 4>{1, 1, 1, 1} /* dummy shape for simplifying code */);
ck_tile::HostTensor<AccDataType> dq_acc_host(
i_perm
? std::array<ck_tile::index_t, 5>{nsplits, shape_batch, nhead, shape_seqlen_q, hdim_q}
: std::array<ck_tile::index_t, 5>{nsplits, shape_batch, shape_seqlen_q, nhead, hdim_q});
if(init_method == 0)
{
......@@ -362,6 +369,7 @@ bool run(const ck_tile::ArgParser& arg_parser)
ck_tile::DeviceMem seqstart_q(seqstart_q_host.size() * sizeof(int32_t));
ck_tile::DeviceMem seqstart_k(seqstart_k_host.size() * sizeof(int32_t));
ck_tile::DeviceMem alibi_slope_buf(alibi_slope_host.get_element_space_size_in_bytes());
ck_tile::DeviceMem dq_acc_buf(dq_acc_host.get_element_space_size_in_bytes());
q_buf.ToDevice(q_host.data());
k_buf.ToDevice(k_host.data());
......@@ -387,8 +395,17 @@ bool run(const ck_tile::ArgParser& arg_parser)
std::cout << "[" << prec << "|" << mode << "|" << io_layout(i_perm, o_perm) << "] b:" << batch
<< ", h:" << nhead << "/" << nhead_k << ", s:" << seqlen_q << "/" << seqlen_k
<< ", d:" << hdim_q << "/" << hdim_v << ", scale:" << scale << ", bias:" << bias
<< ", dbias:" << use_dbias << ", p_drop:" << p_drop << ", mask:" << mask
<< std::flush;
<< ", dbias:" << use_dbias << ", p_drop:" << p_drop << ", s_randval:" << s_randval
<< ", deterministic:" << deterministic << ", mask:" << mask << std::flush;
std::size_t workspace_size =
dq_acc_host.get_element_space_size_in_bytes() * sizeof(AccDataType) / (1024 * 1024);
if(deterministic == 1)
{
std::cout << "\nDeterministic mode ON: " << workspace_size
<< " MByte memory workspace allocated" << std::endl;
}
auto fmha_traits = fmha_bwd_traits{hdim_q,
hdim_v,
......@@ -397,7 +414,9 @@ bool run(const ck_tile::ArgParser& arg_parser)
mask.type,
bias.type,
use_dbias,
p_drop > 0.0f};
p_drop > 0.0f,
s_randval,
deterministic};
auto fmha_args = [&]() {
assert(nhead % nhead_k == 0);
/// NOTE: we broadcast bias from [1, 1, seqlen_q, seqlen_k] to [batch, nhead, seqlen_q,
......@@ -422,7 +441,7 @@ bool run(const ck_tile::ArgParser& arg_parser)
const ck_tile::index_t nhead_stride_o = (o_perm ? shape_seqlen_q * hdim_v : hdim_v);
const ck_tile::index_t nhead_stride_randval = (shape_seqlen_q * max_seqlen_k);
const ck_tile::index_t nhead_stride_do = (o_perm ? shape_seqlen_q * hdim_v : hdim_v);
const ck_tile::index_t nhead_stride_lsed = max_seqlen_q;
const ck_tile::index_t nhead_stride_lsed = shape_seqlen_q;
const ck_tile::index_t nhead_stride_dbias =
(i_perm ? shape_seqlen_q * max_seqlen_k : max_seqlen_k);
// setup batch_stride_* arguments
......@@ -433,10 +452,12 @@ bool run(const ck_tile::ArgParser& arg_parser)
const ck_tile::index_t batch_stride_o = (nhead * shape_seqlen_q * hdim_v);
const ck_tile::index_t batch_stride_randval = (nhead * shape_seqlen_q * max_seqlen_k);
const ck_tile::index_t batch_stride_do = (nhead * shape_seqlen_q * hdim_v);
const ck_tile::index_t batch_stride_lsed = (nhead * max_seqlen_q);
const ck_tile::index_t batch_stride_lsed = (nhead * shape_seqlen_q);
const ck_tile::index_t batch_stride_dk = (nhead * shape_seqlen_k * hdim_q);
const ck_tile::index_t batch_stride_dv = (nhead * shape_seqlen_k * hdim_v);
const ck_tile::index_t batch_stride_dbias = (nhead * shape_seqlen_q * max_seqlen_k);
const ck_tile::index_t split_stride_dq_acc =
(shape_batch * nhead * shape_seqlen_q * hdim_q);
return fmha_bwd_args{q_buf.GetDeviceBuffer(),
k_buf.GetDeviceBuffer(),
......@@ -452,6 +473,7 @@ bool run(const ck_tile::ArgParser& arg_parser)
dk_buf.GetDeviceBuffer(),
dv_buf.GetDeviceBuffer(),
dbias_buf.GetDeviceBuffer(),
dq_acc_buf.GetDeviceBuffer(),
seqstart_q.GetDeviceBuffer(),
seqstart_k.GetDeviceBuffer(),
nullptr,
......@@ -473,6 +495,8 @@ bool run(const ck_tile::ArgParser& arg_parser)
stride_o,
stride_randval,
stride_do,
stride_q, // stride_dq_acc
stride_q, // stride_dq
stride_dk,
stride_dv,
stride_dbias,
......@@ -484,6 +508,10 @@ bool run(const ck_tile::ArgParser& arg_parser)
nhead_stride_randval,
nhead_stride_do,
nhead_stride_lsed,
nhead_stride_q, // nhead_stride_dq_acc
nhead_stride_q, // nhead_stride_dq
nhead_stride_k, // nhead_stride_dk
nhead_stride_v, // nhead_stride_dv
nhead_stride_dbias,
batch_stride_q,
batch_stride_k,
......@@ -493,15 +521,17 @@ bool run(const ck_tile::ArgParser& arg_parser)
batch_stride_randval,
batch_stride_do,
batch_stride_lsed,
batch_stride_q, // batch_stride_dq_acc
batch_stride_q, // batch_stride_dq
batch_stride_dk,
batch_stride_dv,
batch_stride_dbias,
split_stride_dq_acc,
mask.left,
mask.right,
static_cast<ck_tile::index_t>(mask.type),
p_drop,
p_undrop,
s_randval,
{drop_seed, drop_offset}};
}();
......@@ -719,7 +749,7 @@ bool run(const ck_tile::ArgParser& arg_parser)
if(o_perm) o_host_ref.ForEach([&](auto& self, auto idx) { o_host(b, idx[0], idx[1] + query_offset, idx[2]) = self(idx); });
else o_host_ref.ForEach([&](auto& self, auto idx) { o_host(b, idx[1] + query_offset, idx[0], idx[2]) = self(idx); });
lse_host_ref.ForEach([&](auto& self, auto idx) { lse_host(wb, idx[0], idx[1]) = self(idx); });
lse_host_ref.ForEach([&](auto& self, auto idx) { lse_host(b, idx[0], idx[1] + query_offset) = self(idx); });
// clang-format on
q_host_refs.push_back(q_host_ref);
......@@ -738,6 +768,7 @@ bool run(const ck_tile::ArgParser& arg_parser)
lse_buf.ToDevice(lse_host.data());
dq_buf.SetZero();
dbias_buf.SetZero();
dq_acc_buf.SetZero();
ck_tile::stream_config stream_config_v{
nullptr, true, 0, 0, 1, arg_parser.get_str("timer") == std::string("gpu")};
......
example/ck_tile/01_fmha/fmha_bwd.hpp
View file @ 0eb75e21
......@@ -77,6 +77,7 @@ struct fmha_bwd_args
void* dk_ptr;
void* dv_ptr;
void* dbias_ptr;
void* dq_acc_ptr;
const void* seqstart_q_ptr;
const void* seqstart_k_ptr;
const void* seqlen_k_ptr;
......@@ -97,6 +98,8 @@ struct fmha_bwd_args
ck_tile::index_t stride_o;
ck_tile::index_t stride_randval;
ck_tile::index_t stride_do;
ck_tile::index_t stride_dq_acc;
ck_tile::index_t stride_dq;
ck_tile::index_t stride_dk;
ck_tile::index_t stride_dv;
ck_tile::index_t stride_dbias;
......@@ -108,6 +111,10 @@ struct fmha_bwd_args
ck_tile::index_t nhead_stride_randval;
ck_tile::index_t nhead_stride_do;
ck_tile::index_t nhead_stride_lsed;
ck_tile::index_t nhead_stride_dq_acc;
ck_tile::index_t nhead_stride_dq;
ck_tile::index_t nhead_stride_dk;
ck_tile::index_t nhead_stride_dv;
ck_tile::index_t nhead_stride_dbias;
ck_tile::index_t batch_stride_q;
ck_tile::index_t batch_stride_k;
......@@ -117,15 +124,17 @@ struct fmha_bwd_args
ck_tile::index_t batch_stride_randval;
ck_tile::index_t batch_stride_do;
ck_tile::index_t batch_stride_lsed;
ck_tile::index_t batch_stride_dq_acc;
ck_tile::index_t batch_stride_dq;
ck_tile::index_t batch_stride_dk;
ck_tile::index_t batch_stride_dv;
ck_tile::index_t batch_stride_dbias;
ck_tile::index_t split_stride_dq_acc;
ck_tile::index_t window_size_left;
ck_tile::index_t window_size_right;
ck_tile::index_t mask_type;
float p_drop;
float p_undrop;
bool s_randval;
std::tuple<uint64_t, uint64_t> drop_seed_offset;
};
......@@ -145,10 +154,10 @@ auto fmha_bwd_dq_dk_dv_create_kargs_and_grids(fmha_bwd_args args)
args.do_ptr,
args.d_ptr,
args.rand_val_ptr,
args.dq_ptr,
args.dk_ptr,
args.dv_ptr,
args.dbias_ptr,
args.dq_acc_ptr,
args.seqstart_q_ptr,
args.seqstart_k_ptr,
args.seqlen_k_ptr,
......@@ -163,6 +172,7 @@ auto fmha_bwd_dq_dk_dv_create_kargs_and_grids(fmha_bwd_args args)
args.stride_bias,
args.stride_randval,
args.stride_do,
args.stride_dq_acc,
args.stride_dk,
args.stride_dv,
args.stride_dbias,
......@@ -173,13 +183,15 @@ auto fmha_bwd_dq_dk_dv_create_kargs_and_grids(fmha_bwd_args args)
args.nhead_stride_randval,
args.nhead_stride_do,
args.nhead_stride_lsed,
args.nhead_stride_dq_acc,
args.nhead_stride_dk,
args.nhead_stride_dv,
args.nhead_stride_dbias,
args.batch_stride_lsed,
args.split_stride_dq_acc,
args.window_size_left,
args.window_size_right,
args.mask_type,
args.p_drop,
args.s_randval,
args.drop_seed_offset);
}
else
......@@ -192,10 +204,10 @@ auto fmha_bwd_dq_dk_dv_create_kargs_and_grids(fmha_bwd_args args)
args.do_ptr,
args.d_ptr,
args.rand_val_ptr,
args.dq_ptr,
args.dk_ptr,
args.dv_ptr,
args.dbias_ptr,
args.dq_acc_ptr,
args.seqlen_q,
args.seqlen_k,
args.hdim_q,
......@@ -209,6 +221,7 @@ auto fmha_bwd_dq_dk_dv_create_kargs_and_grids(fmha_bwd_args args)
args.stride_bias,
args.stride_randval,
args.stride_do,
args.stride_dq_acc,
args.stride_dk,
args.stride_dv,
args.stride_dbias,
......@@ -219,6 +232,9 @@ auto fmha_bwd_dq_dk_dv_create_kargs_and_grids(fmha_bwd_args args)
args.nhead_stride_randval,
args.nhead_stride_do,
args.nhead_stride_lsed,
args.nhead_stride_dq_acc,
args.nhead_stride_dk,
args.nhead_stride_dv,
args.nhead_stride_dbias,
args.batch_stride_q,
args.batch_stride_k,
......@@ -227,14 +243,15 @@ auto fmha_bwd_dq_dk_dv_create_kargs_and_grids(fmha_bwd_args args)
args.batch_stride_randval,
args.batch_stride_do,
args.batch_stride_lsed,
args.batch_stride_dq_acc,
args.batch_stride_dk,
args.batch_stride_dv,
args.batch_stride_dbias,
args.split_stride_dq_acc,
args.window_size_left,
args.window_size_right,
args.mask_type,
args.p_drop,
args.s_randval,
args.drop_seed_offset);
}
}();
......@@ -260,8 +277,7 @@ auto fmha_bwd_dot_do_o_create_kargs_and_grids(fmha_bwd_args args)
args.stride_o,
args.nhead_stride_do,
args.nhead_stride_o,
args.nhead_stride_lsed,
args.batch_stride_lsed);
args.nhead_stride_lsed);
}
else
{ // create batch mode kernel arguments
......@@ -286,19 +302,59 @@ auto fmha_bwd_dot_do_o_create_kargs_and_grids(fmha_bwd_args args)
return ck_tile::make_tuple(kargs, grids);
}
template <typename FmhaBwdConvertQGradKernel>
auto fmha_bwd_convert_dq_create_kargs_and_grids(fmha_bwd_args args)
{
auto kargs = [&] {
// create group mode kernel arguments
if constexpr(FmhaBwdConvertQGradKernel::kIsGroupMode)
{
return FmhaBwdConvertQGradKernel::MakeKargs(args.dq_acc_ptr,
args.dq_ptr,
args.seqstart_q_ptr,
args.seqstart_k_ptr,
args.hdim_q,
args.stride_dq,
args.stride_dq_acc,
args.nhead_stride_dq,
args.nhead_stride_dq_acc,
args.split_stride_dq_acc);
}
else
{ // create batch mode kernel arguments
return FmhaBwdConvertQGradKernel::MakeKargs(args.dq_acc_ptr,
args.dq_ptr,
args.seqlen_q,
args.seqlen_k,
args.hdim_q,
args.stride_dq,
args.stride_dq_acc,
args.nhead_stride_dq,
args.nhead_stride_dq_acc,
args.batch_stride_dq,
args.batch_stride_dq_acc,
args.split_stride_dq_acc);
}
}();
dim3 grids = FmhaBwdConvertQGradKernel::GridSize(args.batch, args.nhead_q, args.max_seqlen_q);
return ck_tile::make_tuple(kargs, grids);
}
// this is used to pattern-match internl kernel implementation, not to instantiate kernel
template <ck_tile::index_t HDim_,
typename DataType_,
bool kIsGroupMode_,
ck_tile::BlockFmhaBwdPipelineEnum FmhaBwdPipelineEnum_,
typename FmhaMask_,
typename FmhaDropout_,
ck_tile::BlockAttentionBiasEnum BiasEnum_,
bool kHasBiasGrad_,
bool kHasDropout_,
bool kPadS_,
bool kPadSK_,
bool kPadD_,
bool kPadDv_>
bool kPadDv_,
bool kIsDeterministic_>
struct fmha_bwd_dq_dk_dv_traits_
{
static constexpr ck_tile::index_t HDim = HDim_;
......@@ -306,13 +362,14 @@ struct fmha_bwd_dq_dk_dv_traits_
static constexpr bool kIsGroupMode = kIsGroupMode_;
static constexpr auto FmhaBwdPipelineEnum = FmhaBwdPipelineEnum_;
using FmhaMask = ck_tile::remove_cvref_t<FmhaMask_>;
using FmhaDropout = ck_tile::remove_cvref_t<FmhaDropout_>;
static constexpr auto BiasEnum = BiasEnum_;
static constexpr bool kHasBiasGrad = kHasBiasGrad_;
static constexpr bool kHasDropout = kHasDropout_;
static constexpr bool kPadS = kPadS_;
static constexpr bool kPadSK = kPadSK_;
static constexpr bool kPadD = kPadD_;
static constexpr bool kPadDv = kPadDv_;
static constexpr bool kIsDeterministic = kIsDeterministic_;
};
template <typename Traits_>
......@@ -343,6 +400,31 @@ void fmha_bwd_dot_do_o_oneshot_(const ck_tile::stream_config&, fmha_bwd_args);
template <typename Traits_>
std::string fmha_bwd_dot_do_o_get_name_();
template <ck_tile::index_t HDim_,
typename DataType_,
bool kIsGroupMode_,
bool kPadS_,
bool kPadD_,
bool kIsDeterministic_>
struct fmha_bwd_convert_dq_traits_
{
static constexpr ck_tile::index_t HDim = HDim_;
using DataType = ck_tile::remove_cvref_t<DataType_>;
static constexpr bool kIsGroupMode = kIsGroupMode_;
static constexpr bool kPadS = kPadS_;
static constexpr bool kPadD = kPadD_;
static constexpr bool kIsDeterministic = kIsDeterministic_;
};
template <typename Traits_>
float fmha_bwd_convert_dq_(const ck_tile::stream_config&, fmha_bwd_args);
template <typename Traits_>
void fmha_bwd_convert_dq_oneshot_(const ck_tile::stream_config&, fmha_bwd_args);
template <typename Traits_>
std::string fmha_bwd_convert_dq_get_name_();
// This is the public API, will be generated by script
struct fmha_bwd_traits
{
......@@ -354,6 +436,8 @@ struct fmha_bwd_traits
bias_enum bias_type; // 0:no bias, 1:elementwise bias, 2:alibi. sync with BlockAttentionBiasEnum
bool has_dbias;
bool has_dropout;
bool is_store_randval;
bool is_deterministic;
// TODO: padding check is inside this api
};
float fmha_bwd(fmha_bwd_traits, fmha_bwd_args, const ck_tile::stream_config&);
example/ck_tile/01_fmha/fmha_fwd.cpp
View file @ 0eb75e21
......@@ -479,16 +479,18 @@ bool run(const ck_tile::ArgParser& arg_parser)
: std::array<ck_tile::index_t, 2>{1, 1});
ck_tile::HostTensor<LSEDataType> lse_acc_host(
1 < num_splits ? std::array<ck_tile::index_t, 4>{num_splits, batch, nhead, max_seqlen_q}
: std::array<ck_tile::index_t, 4>{1, 1, 1, 1});
1 < num_splits
? std::array<ck_tile::index_t, 4>{num_splits, shape_batch, nhead, shape_seqlen_q}
: std::array<ck_tile::index_t, 4>{1, 1, 1, 1});
ck_tile::HostTensor<OaccDataType> o_acc_host(
1 < num_splits
? std::array<ck_tile::index_t, 5>{num_splits, batch, nhead, max_seqlen_q, hdim_v}
: std::array<ck_tile::index_t, 5>{1, 1, 1, 1, 1});
// self define lse data layout as [batch, nhead, max_seqlen_q]
// batch mode of lse data layout is [batch, nhead, seqlen_q]
// group mode of lse data layout is [nhead, total_seqlen_q]
ck_tile::HostTensor<LSEDataType> lse_host(
lse ? std::array<ck_tile::index_t, 3>{batch, nhead, max_seqlen_q}
lse ? std::array<ck_tile::index_t, 3>{shape_batch, nhead, shape_seqlen_q}
: std::array<ck_tile::index_t, 3>{1, 1, 1} /* dummy shape for simplifying code */);
ck_tile::HostTensor<ODataType> o_host(
......@@ -669,8 +671,8 @@ bool run(const ck_tile::ArgParser& arg_parser)
const ck_tile::index_t nhead_stride_bias =
(i_perm ? 0 * shape_seqlen_q * shape_seqlen_k : 0 * shape_seqlen_k);
const ck_tile::index_t nhead_stride_randval = (shape_seqlen_q * max_seqlen_k);
const ck_tile::index_t nhead_stride_lse = max_seqlen_q;
const ck_tile::index_t nhead_stride_lse_acc = max_seqlen_q;
const ck_tile::index_t nhead_stride_lse = shape_seqlen_q;
const ck_tile::index_t nhead_stride_lse_acc = shape_seqlen_q;
const ck_tile::index_t nhead_stride_o_acc = (max_seqlen_q * hdim_v);
const ck_tile::index_t nhead_stride_o = (o_perm ? shape_seqlen_q * hdim_v : hdim_v);
// setup batch_stride_* arguments
......@@ -679,12 +681,12 @@ bool run(const ck_tile::ArgParser& arg_parser)
const ck_tile::index_t batch_stride_v = (nhead_k * hdim_v * shape_seqlen_k);
const ck_tile::index_t batch_stride_bias = (0 * nhead * shape_seqlen_q * shape_seqlen_k);
const ck_tile::index_t batch_stride_randval = (nhead * shape_seqlen_q * max_seqlen_k);
const ck_tile::index_t batch_stride_lse = (nhead * max_seqlen_q);
const ck_tile::index_t batch_stride_lse_acc = (nhead * max_seqlen_q);
const ck_tile::index_t batch_stride_lse = (nhead * shape_seqlen_q);
const ck_tile::index_t batch_stride_lse_acc = (nhead * shape_seqlen_q);
const ck_tile::index_t batch_stride_o_acc = (nhead * max_seqlen_q * hdim_v);
const ck_tile::index_t batch_stride_o = (nhead * shape_seqlen_q * hdim_v);
// setup split_stride_* arguments (only used in split-kv kernel)
const ck_tile::index_t split_stride_lse_acc = (batch * nhead * max_seqlen_q);
const ck_tile::index_t split_stride_lse_acc = (shape_batch * nhead * shape_seqlen_q);
const ck_tile::index_t split_stride_o_acc = (batch * nhead * max_seqlen_q * hdim_v);
return fmha_fwd_args{q_buf.GetDeviceBuffer(),
......@@ -996,8 +998,9 @@ bool run(const ck_tile::ArgParser& arg_parser)
if(lse)
{
ck_tile::HostTensor<SMPLComputeDataType> lse_host_result({nhead, real_seqlen_q});
lse_host_result.ForEach(
[&](auto& self, auto idx) { self(idx) = lse_host(wb, idx[0], idx[1]); });
lse_host_result.ForEach([&](auto& self, auto idx) {
self(idx) = lse_host(b, idx[0], idx[1] + query_offset);
});
cur_pass = ck_tile::check_err(lse_host_result,
lse_host_ref,
......
example/ck_tile/01_fmha/fmha_fwd.hpp
View file @ 0eb75e21
......@@ -185,7 +185,6 @@ auto fmha_fwd_create_kargs_and_grids(fmha_fwd_args args)
args.nhead_stride_randval,
args.nhead_stride_lse,
args.nhead_stride_o,
args.batch_stride_lse,
args.window_size_left,
args.window_size_right,
args.mask_type,
......@@ -284,7 +283,6 @@ auto fmha_fwd_splitkv_create_kargs_and_grids(fmha_fwd_args args)
args.nhead_stride_randval,
args.nhead_stride_lse_acc,
args.nhead_stride_o_acc,
args.batch_stride_lse_acc,
args.batch_stride_o_acc,
args.split_stride_lse_acc,
args.split_stride_o_acc,
......@@ -376,9 +374,7 @@ auto fmha_fwd_splitkv_combine_create_kargs_and_grids(fmha_fwd_args args)
args.nhead_stride_o_acc,
args.nhead_stride_lse,
args.nhead_stride_o,
args.batch_stride_lse_acc,
args.batch_stride_o_acc,
args.batch_stride_lse,
args.split_stride_lse_acc,
args.split_stride_o_acc);
}
......
example/ck_tile/01_fmha/script/run_full_test.sh 0 → 100755
View file @ 0eb75e21
#!/bin/bash
#
# in order to run this script you'd first need to build the tile_example_fmha_fwd and tile_eaxmple_fmha_bwd executables in ../build/bin/
#
# run the script as "./run_full_test.sh <tag for your test environment> <branch name> <host name> <gpu_arch>
# input arguments:
# environment tag : a string describing the specifics of your test environment
# branch name : name of the branch in git repo (git status | grep -e 'On branch')
# host name : $hostname
# gpu architecture: e.g., gfx90a, or gfx942, etc.
#get the command line arguments:
export env_type=$1
echo 'Environment type: ' $env_type
export branch=$2
echo 'Branch name: ' $branch
export host_name=$3
echo 'Host name: ' $host_name
export GPU_arch=$4
echo 'GPU_arch: ' $GPU_arch
function print_log_header(){
rm -f $1;
echo 'On branch ' $3 &> $1;
echo 'Node name: ' $4 >> $1;
#get GPU_arch and number of compute units from rocminfo
echo -n "GPU_arch: " >> $1; rocminfo | grep "Name:" | grep "gfx" >> $1;
rocminfo | grep "Compute Unit:" >> $1;
hipcc --version | grep -e 'HIP version' >> $1;
echo 'Environment type: ' $2 >> $1;
/opt/rocm/bin/amdclang++ --version | grep -e 'InstalledDir' >> $1;
}
#run verification tests
example/ck_tile/01_fmha/script/smoke_test_fwd.sh
example/ck_tile/01_fmha/script/smoke_test_bwd.sh
#run performance benchmarks
export fmha_fwd_log="perf_fmha_fwd_$GPU_arch.log"
print_log_header $fmha_fwd_log $env_type $branch $host_name
example/ck_tile/01_fmha/script/benchmark_fwd.sh 2>&1 | tee -a $fmha_fwd_log
export fmha_bwd_log="perf_fmha_bwd_$GPU_arch.log"
print_log_header $fmha_bwd_log $env_type $branch $host_name
example/ck_tile/01_fmha/script/benchmark_bwd.sh 2>&1 | tee -a $fmha_bwd_log
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