mirror of
https://github.com/ROCm/composable_kernel.git
synced 2026-05-13 09:45:56 +00:00
batched_gemm + multiple_d + gemm + multiple_d (#394)
* refactor * start * add device gemm file * add BatchStrideD0 * add stridd0 * add gridwise file * add d0 parameters to gridwise gemm * add c layout transformer * add d0 threadwise copy * init kernel * init kernel * regular code * nm desc put to out * kernel parameter can not use reference * host add bias+gelu * run right for bias+gelu * change AddFastGelu into another file * interface add d1 bias parameters * add d1 parameter to argument * add d1 parameter to gridwise * first all code,not verify * gelu change to relu and GetElementSpaceSize bug * add instance * start add to ckprofiler * ckprofiler finish code * change input parameter for ckProfiler * fix host bias+gelu bug * show help for ckProfiler * fix bug for lunch kernel ignore parametes * add pad and fix about bug * mutiple d0 * add dynamic d0_element_op * change profiler and instance to mutiple d0 * example have 2 d0 * remove some comments not using * change 2 d0 have self parameters * change d element_op name * change class name(multiple_d) * fix bug * fix bug that don't find file * update profiler * refactor * update profiler * clean * revert example change * add gon layout * optimize parameter for gno * add gon to gemm+gemm * change helping input parameters * change to GemmPadder_v2 * using ForEach * fix gb_per_sec Co-authored-by: Chao Liu <lc.roy86@gmail.com> Co-authored-by: ltqin <letaoqin@amd.com>
This commit is contained in:
ltqin
@@ -0,0 +1,139 @@
|
||||
// SPDX-License-Identifier: MIT
|
||||
// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
|
||||
|
||||
#pragma once
|
||||
|
||||
#include <cstdlib>
|
||||
|
||||
#include "ck/ck.hpp"
|
||||
#include "ck/tensor_operation/gpu/device/tensor_layout.hpp"
|
||||
#include "ck/tensor_operation/gpu/device/device_batched_gemm_multiple_d_gemm_multiple_d.hpp"
|
||||
#include "ck/tensor_operation/gpu/element/element_wise_operation.hpp"
|
||||
|
||||
#include "ck/library/tensor_operation_instance/device_operation_instance_factory.hpp"
|
||||
|
||||
using CDE0ElementOp = ck::tensor_operation::element_wise::AddRelu;
|
||||
using CDE1ElementOp = ck::tensor_operation::element_wise::Add;
|
||||
|
||||
namespace ck {
|
||||
namespace tensor_operation {
|
||||
namespace device {
|
||||
namespace instance {
|
||||
|
||||
void add_device_batched_gemm_add_relu_gemm_add_xdl_cshuffle_f16_f16_f16_f16_gmk_gnk_gno_gmo_instance(
|
||||
std::vector<std::unique_ptr<DeviceBatchedGemmMultipleDGemmMultipleD<Row,
|
||||
Col,
|
||||
ck::Tuple<Row>,
|
||||
Row,
|
||||
ck::Tuple<Row>,
|
||||
Row,
|
||||
F16,
|
||||
F16,
|
||||
ck::Tuple<F16>,
|
||||
F16,
|
||||
ck::Tuple<F16>,
|
||||
F16,
|
||||
PassThrough,
|
||||
PassThrough,
|
||||
CDE0ElementOp,
|
||||
PassThrough,
|
||||
CDE1ElementOp>>>&
|
||||
instances);
|
||||
|
||||
void add_device_batched_gemm_add_relu_gemm_add_xdl_cshuffle_f16_f16_f16_f16_gmk_gnk_gon_gmo_instance(
|
||||
std::vector<std::unique_ptr<DeviceBatchedGemmMultipleDGemmMultipleD<Row,
|
||||
Col,
|
||||
ck::Tuple<Row>,
|
||||
Col,
|
||||
ck::Tuple<Row>,
|
||||
Row,
|
||||
F16,
|
||||
F16,
|
||||
ck::Tuple<F16>,
|
||||
F16,
|
||||
ck::Tuple<F16>,
|
||||
F16,
|
||||
PassThrough,
|
||||
PassThrough,
|
||||
CDE0ElementOp,
|
||||
PassThrough,
|
||||
CDE1ElementOp>>>&
|
||||
instances);
|
||||
|
||||
template <typename A0Layout,
|
||||
typename B0Layout,
|
||||
typename D0sLayout,
|
||||
typename B1Layout,
|
||||
typename D1sLayout,
|
||||
typename E1Layout,
|
||||
typename A0DataType,
|
||||
typename B0DataType,
|
||||
typename D0sDataType,
|
||||
typename B1DataType,
|
||||
typename D1sDataType,
|
||||
typename E1DataType>
|
||||
struct DeviceOperationInstanceFactory<
|
||||
ck::tensor_operation::device::DeviceBatchedGemmMultipleDGemmMultipleD<A0Layout,
|
||||
B0Layout,
|
||||
D0sLayout,
|
||||
B1Layout,
|
||||
D1sLayout,
|
||||
E1Layout,
|
||||
A0DataType,
|
||||
B0DataType,
|
||||
D0sDataType,
|
||||
B1DataType,
|
||||
D1sDataType,
|
||||
E1DataType,
|
||||
PassThrough,
|
||||
PassThrough,
|
||||
CDE0ElementOp,
|
||||
PassThrough,
|
||||
CDE1ElementOp>>
|
||||
{
|
||||
using DeviceOp = DeviceBatchedGemmMultipleDGemmMultipleD<A0Layout,
|
||||
B0Layout,
|
||||
D0sLayout,
|
||||
B1Layout,
|
||||
D1sLayout,
|
||||
E1Layout,
|
||||
A0DataType,
|
||||
B0DataType,
|
||||
D0sDataType,
|
||||
B1DataType,
|
||||
D1sDataType,
|
||||
E1DataType,
|
||||
PassThrough,
|
||||
PassThrough,
|
||||
CDE0ElementOp,
|
||||
PassThrough,
|
||||
CDE1ElementOp>;
|
||||
|
||||
static auto GetInstances()
|
||||
{
|
||||
std::vector<std::unique_ptr<DeviceOp>> op_ptrs;
|
||||
|
||||
if constexpr(is_same_v<A0DataType, half_t> && is_same_v<B0DataType, half_t> &&
|
||||
is_same_v<B1DataType, half_t> && is_same_v<E1DataType, half_t>)
|
||||
{
|
||||
if constexpr(is_same_v<A0Layout, Row> && is_same_v<B0Layout, Col> &&
|
||||
is_same_v<B1Layout, Row> && is_same_v<E1Layout, Row>)
|
||||
{
|
||||
add_device_batched_gemm_add_relu_gemm_add_xdl_cshuffle_f16_f16_f16_f16_gmk_gnk_gno_gmo_instance(
|
||||
op_ptrs);
|
||||
}
|
||||
else if constexpr(is_same_v<A0Layout, Row> && is_same_v<B0Layout, Col> &&
|
||||
is_same_v<B1Layout, Col> && is_same_v<E1Layout, Row>)
|
||||
{
|
||||
add_device_batched_gemm_add_relu_gemm_add_xdl_cshuffle_f16_f16_f16_f16_gmk_gnk_gon_gmo_instance(
|
||||
op_ptrs);
|
||||
}
|
||||
}
|
||||
return op_ptrs;
|
||||
}
|
||||
};
|
||||
|
||||
} // namespace instance
|
||||
} // namespace device
|
||||
} // namespace tensor_operation
|
||||
} // namespace ck
|
||||
Reference in New Issue
Block a user