mirror of
https://github.com/ROCm/composable_kernel.git
synced 2026-05-11 17:00:18 +00:00
[GEMM] F8 GEMM, performance optimized. (#1384)
* add ab_scale init support
* enabled interwave
* add scale type; update isSupport
* adjust example
* clean
* enable f8 pure gemm rcr ckprofiler
* Add gemm_multiply_multiply instances
* clang format
* Optimize for ScaleBlockMNK=128
* enable abscale f8 gemm ck profiler
* Add pure f8 gemm test suite
* Reverting to the state of project at f60fd77
* update copyright
* clang format
* update copyright
---------
Co-authored-by: root <jizhan@amd.com>
This commit is contained in:
Haocong WANG
@@ -108,6 +108,7 @@ using FastGelu = ck::tensor_operation::element_wise::FastGelu;
|
||||
using MultiplyFastGelu = ck::tensor_operation::element_wise::MultiplyFastGelu;
|
||||
using AddMultiply = ck::tensor_operation::element_wise::AddMultiply;
|
||||
using MultiplyAdd = ck::tensor_operation::element_wise::MultiplyAdd;
|
||||
using MultiplyMultiply = ck::tensor_operation::element_wise::MultiplyMultiply;
|
||||
using ScaleAdd = ck::tensor_operation::element_wise::ScaleAdd;
|
||||
using Gelu = ck::tensor_operation::element_wise::Gelu;
|
||||
using Swish = ck::tensor_operation::element_wise::Swish;
|
||||
|
||||
@@ -0,0 +1,226 @@
|
||||
// SPDX-License-Identifier: MIT
|
||||
// Copyright (c) 2024, Advanced Micro Devices, Inc. All rights reserved.
|
||||
|
||||
#pragma once
|
||||
|
||||
#include <vector>
|
||||
#include <memory>
|
||||
#include "ck/ck.hpp"
|
||||
#include "ck/tensor_operation/gpu/device/tensor_layout.hpp"
|
||||
#include "ck/tensor_operation/gpu/device/impl/device_gemm_multiple_d_xdl_cshuffle_v3_ab_scale.hpp"
|
||||
#include "ck/tensor_operation/gpu/element/element_wise_operation.hpp"
|
||||
|
||||
#include "ck/library/tensor_operation_instance/device_operation_instance_factory.hpp"
|
||||
|
||||
namespace ck {
|
||||
namespace tensor_operation {
|
||||
namespace device {
|
||||
namespace instance {
|
||||
#if(defined(CK_ENABLE_BF16) || defined(CK_ENABLE_FP8))
|
||||
void add_device_gemm_ab_scale_xdl_f8_f8_bf16_mk_nk_mn_128_128_128_comp_default_instances(
|
||||
std::vector<std::unique_ptr<DeviceGemmMultipleD_ABScale<Row,
|
||||
Col,
|
||||
Tuple<>,
|
||||
Row,
|
||||
F8,
|
||||
F32,
|
||||
F8,
|
||||
F32,
|
||||
Tuple<>,
|
||||
BF16,
|
||||
128,
|
||||
128,
|
||||
128,
|
||||
PassThrough,
|
||||
PassThrough,
|
||||
PassThrough>>>& instances);
|
||||
|
||||
void add_device_gemm_ab_scale_xdl_f8_f8_bf16_mk_nk_mn_128_128_128_comp_kpadding_instances(
|
||||
std::vector<std::unique_ptr<DeviceGemmMultipleD_ABScale<Row,
|
||||
Col,
|
||||
Tuple<>,
|
||||
Row,
|
||||
F8,
|
||||
F32,
|
||||
F8,
|
||||
F32,
|
||||
Tuple<>,
|
||||
BF16,
|
||||
128,
|
||||
128,
|
||||
128,
|
||||
PassThrough,
|
||||
PassThrough,
|
||||
PassThrough>>>& instances);
|
||||
|
||||
void add_device_gemm_ab_scale_xdl_f8_f8_bf16_mk_nk_mn_128_128_128_comp_mnpadding_instances(
|
||||
std::vector<std::unique_ptr<DeviceGemmMultipleD_ABScale<Row,
|
||||
Col,
|
||||
Tuple<>,
|
||||
Row,
|
||||
F8,
|
||||
F32,
|
||||
F8,
|
||||
F32,
|
||||
Tuple<>,
|
||||
BF16,
|
||||
128,
|
||||
128,
|
||||
128,
|
||||
PassThrough,
|
||||
PassThrough,
|
||||
PassThrough>>>& instances);
|
||||
|
||||
void add_device_gemm_ab_scale_xdl_f8_f8_bf16_mk_nk_mn_128_128_128_comp_mnkpadding_instances(
|
||||
std::vector<std::unique_ptr<DeviceGemmMultipleD_ABScale<Row,
|
||||
Col,
|
||||
Tuple<>,
|
||||
Row,
|
||||
F8,
|
||||
F32,
|
||||
F8,
|
||||
F32,
|
||||
Tuple<>,
|
||||
BF16,
|
||||
128,
|
||||
128,
|
||||
128,
|
||||
PassThrough,
|
||||
PassThrough,
|
||||
PassThrough>>>& instances);
|
||||
|
||||
void add_device_gemm_ab_scale_xdl_f8_f8_bf16_mk_nk_mn_128_128_128_mem_v1_default_instances(
|
||||
std::vector<std::unique_ptr<DeviceGemmMultipleD_ABScale<Row,
|
||||
Col,
|
||||
Tuple<>,
|
||||
Row,
|
||||
F8,
|
||||
F32,
|
||||
F8,
|
||||
F32,
|
||||
Tuple<>,
|
||||
BF16,
|
||||
128,
|
||||
128,
|
||||
128,
|
||||
PassThrough,
|
||||
PassThrough,
|
||||
PassThrough>>>& instances);
|
||||
|
||||
void add_device_gemm_ab_scale_xdl_f8_f8_bf16_mk_nk_mn_128_128_128_mem_v1_kpadding_instances(
|
||||
std::vector<std::unique_ptr<DeviceGemmMultipleD_ABScale<Row,
|
||||
Col,
|
||||
Tuple<>,
|
||||
Row,
|
||||
F8,
|
||||
F32,
|
||||
F8,
|
||||
F32,
|
||||
Tuple<>,
|
||||
BF16,
|
||||
128,
|
||||
128,
|
||||
128,
|
||||
PassThrough,
|
||||
PassThrough,
|
||||
PassThrough>>>& instances);
|
||||
|
||||
void add_device_gemm_ab_scale_xdl_f8_f8_bf16_mk_nk_mn_128_128_128_mem_v1_mnkpadding_instances(
|
||||
std::vector<std::unique_ptr<DeviceGemmMultipleD_ABScale<Row,
|
||||
Col,
|
||||
Tuple<>,
|
||||
Row,
|
||||
F8,
|
||||
F32,
|
||||
F8,
|
||||
F32,
|
||||
Tuple<>,
|
||||
BF16,
|
||||
128,
|
||||
128,
|
||||
128,
|
||||
PassThrough,
|
||||
PassThrough,
|
||||
PassThrough>>>& instances);
|
||||
#endif
|
||||
|
||||
template <typename A0DataType,
|
||||
typename A1DataType,
|
||||
typename B0DataType,
|
||||
typename B1DataType,
|
||||
typename CDataType,
|
||||
typename ALayout,
|
||||
typename BLayout,
|
||||
typename CLayout>
|
||||
struct DeviceOperationInstanceFactory<ck::tensor_operation::device::DeviceGemmMultipleD_ABScale<
|
||||
ALayout,
|
||||
BLayout,
|
||||
Tuple<>,
|
||||
CLayout,
|
||||
A0DataType,
|
||||
A1DataType,
|
||||
B0DataType,
|
||||
B1DataType,
|
||||
Tuple<>,
|
||||
CDataType,
|
||||
128,
|
||||
128,
|
||||
128,
|
||||
ck::tensor_operation::element_wise::PassThrough,
|
||||
ck::tensor_operation::element_wise::PassThrough,
|
||||
ck::tensor_operation::element_wise::PassThrough>>
|
||||
{
|
||||
using DeviceOp = DeviceGemmMultipleD_ABScale<ALayout,
|
||||
BLayout,
|
||||
Tuple<>,
|
||||
CLayout,
|
||||
A0DataType,
|
||||
A1DataType,
|
||||
B0DataType,
|
||||
B1DataType,
|
||||
Tuple<>,
|
||||
CDataType,
|
||||
128,
|
||||
128,
|
||||
128,
|
||||
ck::tensor_operation::element_wise::PassThrough,
|
||||
ck::tensor_operation::element_wise::PassThrough,
|
||||
ck::tensor_operation::element_wise::PassThrough>;
|
||||
|
||||
static auto GetInstances()
|
||||
{
|
||||
std::vector<std::unique_ptr<DeviceOp>> op_ptrs;
|
||||
|
||||
#if(defined(CK_ENABLE_BF16) || defined(CK_ENABLE_FP8))
|
||||
if constexpr(is_same_v<A0DataType, f8_t> && is_same_v<B0DataType, f8_t> &&
|
||||
is_same_v<CDataType, bhalf_t>)
|
||||
{
|
||||
if constexpr(is_same_v<ALayout, Row> && is_same_v<BLayout, Col> &&
|
||||
is_same_v<CLayout, Row>)
|
||||
{
|
||||
add_device_gemm_ab_scale_xdl_f8_f8_bf16_mk_nk_mn_128_128_128_comp_default_instances(
|
||||
op_ptrs);
|
||||
add_device_gemm_ab_scale_xdl_f8_f8_bf16_mk_nk_mn_128_128_128_comp_kpadding_instances(
|
||||
op_ptrs);
|
||||
add_device_gemm_ab_scale_xdl_f8_f8_bf16_mk_nk_mn_128_128_128_comp_mnpadding_instances(
|
||||
op_ptrs);
|
||||
add_device_gemm_ab_scale_xdl_f8_f8_bf16_mk_nk_mn_128_128_128_comp_mnkpadding_instances(
|
||||
op_ptrs);
|
||||
|
||||
add_device_gemm_ab_scale_xdl_f8_f8_bf16_mk_nk_mn_128_128_128_mem_v1_default_instances(
|
||||
op_ptrs);
|
||||
add_device_gemm_ab_scale_xdl_f8_f8_bf16_mk_nk_mn_128_128_128_mem_v1_kpadding_instances(
|
||||
op_ptrs);
|
||||
add_device_gemm_ab_scale_xdl_f8_f8_bf16_mk_nk_mn_128_128_128_mem_v1_mnkpadding_instances(
|
||||
op_ptrs);
|
||||
}
|
||||
}
|
||||
#endif
|
||||
return op_ptrs;
|
||||
}
|
||||
};
|
||||
|
||||
} // namespace instance
|
||||
} // namespace device
|
||||
} // namespace tensor_operation
|
||||
} // namespace ck
|
||||
@@ -0,0 +1,225 @@
|
||||
// SPDX-License-Identifier: MIT
|
||||
// Copyright (c) 2024, Advanced Micro Devices, Inc. All rights reserved.
|
||||
|
||||
#pragma once
|
||||
|
||||
#include <vector>
|
||||
#include <memory>
|
||||
#include "ck/ck.hpp"
|
||||
#include "ck/tensor_operation/gpu/device/tensor_layout.hpp"
|
||||
#include "ck/tensor_operation/gpu/device/impl/device_gemm_multiple_d_xdl_cshuffle_v3.hpp"
|
||||
#include "ck/tensor_operation/gpu/element/element_wise_operation.hpp"
|
||||
|
||||
#include "ck/library/tensor_operation_instance/device_operation_instance_factory.hpp"
|
||||
|
||||
namespace ck {
|
||||
namespace tensor_operation {
|
||||
namespace device {
|
||||
namespace instance {
|
||||
#if(defined(CK_ENABLE_BF16) || defined(CK_ENABLE_FP8))
|
||||
void add_device_gemm_multiply_multiply_xdl_f8_f8_bf16_mk_nk_mn_comp_default_instances(
|
||||
std::vector<std::unique_ptr<DeviceGemmMultipleD<Row,
|
||||
Col,
|
||||
Tuple<Row, Col>,
|
||||
Row,
|
||||
F8,
|
||||
F8,
|
||||
Tuple<F32, F32>,
|
||||
BF16,
|
||||
PassThrough,
|
||||
PassThrough,
|
||||
MultiplyMultiply>>>& instances);
|
||||
|
||||
void add_device_gemm_multiply_multiply_xdl_f8_f8_bf16_mk_nk_mn_comp_kpadding_instances(
|
||||
std::vector<std::unique_ptr<DeviceGemmMultipleD<Row,
|
||||
Col,
|
||||
Tuple<Row, Col>,
|
||||
Row,
|
||||
F8,
|
||||
F8,
|
||||
Tuple<F32, F32>,
|
||||
BF16,
|
||||
PassThrough,
|
||||
PassThrough,
|
||||
MultiplyMultiply>>>& instances);
|
||||
|
||||
void add_device_gemm_multiply_multiply_xdl_f8_f8_bf16_mk_nk_mn_comp_mnpadding_instances(
|
||||
std::vector<std::unique_ptr<DeviceGemmMultipleD<Row,
|
||||
Col,
|
||||
Tuple<Row, Col>,
|
||||
Row,
|
||||
F8,
|
||||
F8,
|
||||
Tuple<F32, F32>,
|
||||
BF16,
|
||||
PassThrough,
|
||||
PassThrough,
|
||||
MultiplyMultiply>>>& instances);
|
||||
|
||||
void add_device_gemm_multiply_multiply_xdl_f8_f8_bf16_mk_nk_mn_comp_mnkpadding_instances(
|
||||
std::vector<std::unique_ptr<DeviceGemmMultipleD<Row,
|
||||
Col,
|
||||
Tuple<Row, Col>,
|
||||
Row,
|
||||
F8,
|
||||
F8,
|
||||
Tuple<F32, F32>,
|
||||
BF16,
|
||||
PassThrough,
|
||||
PassThrough,
|
||||
MultiplyMultiply>>>& instances);
|
||||
|
||||
void add_device_gemm_multiply_multiply_xdl_f8_f8_bf16_mk_nk_mn_mem_v1_default_instances(
|
||||
std::vector<std::unique_ptr<DeviceGemmMultipleD<Row,
|
||||
Col,
|
||||
Tuple<Row, Col>,
|
||||
Row,
|
||||
F8,
|
||||
F8,
|
||||
Tuple<F32, F32>,
|
||||
BF16,
|
||||
PassThrough,
|
||||
PassThrough,
|
||||
MultiplyMultiply>>>& instances);
|
||||
|
||||
void add_device_gemm_multiply_multiply_xdl_f8_f8_bf16_mk_nk_mn_mem_v1_kpadding_instances(
|
||||
std::vector<std::unique_ptr<DeviceGemmMultipleD<Row,
|
||||
Col,
|
||||
Tuple<Row, Col>,
|
||||
Row,
|
||||
F8,
|
||||
F8,
|
||||
Tuple<F32, F32>,
|
||||
BF16,
|
||||
PassThrough,
|
||||
PassThrough,
|
||||
MultiplyMultiply>>>& instances);
|
||||
|
||||
void add_device_gemm_multiply_multiply_xdl_f8_f8_bf16_mk_nk_mn_mem_v1_mnkpadding_instances(
|
||||
std::vector<std::unique_ptr<DeviceGemmMultipleD<Row,
|
||||
Col,
|
||||
Tuple<Row, Col>,
|
||||
Row,
|
||||
F8,
|
||||
F8,
|
||||
Tuple<F32, F32>,
|
||||
BF16,
|
||||
PassThrough,
|
||||
PassThrough,
|
||||
MultiplyMultiply>>>& instances);
|
||||
|
||||
void add_device_gemm_multiply_multiply_xdl_f8_f8_bf16_mk_nk_mn_mem_v2_default_instances(
|
||||
std::vector<std::unique_ptr<DeviceGemmMultipleD<Row,
|
||||
Col,
|
||||
Tuple<Row, Col>,
|
||||
Row,
|
||||
F8,
|
||||
F8,
|
||||
Tuple<F32, F32>,
|
||||
BF16,
|
||||
PassThrough,
|
||||
PassThrough,
|
||||
MultiplyMultiply>>>& instances);
|
||||
|
||||
void add_device_gemm_multiply_multiply_xdl_f8_f8_bf16_mk_nk_mn_mem_v2_kpadding_instances(
|
||||
std::vector<std::unique_ptr<DeviceGemmMultipleD<Row,
|
||||
Col,
|
||||
Tuple<Row, Col>,
|
||||
Row,
|
||||
F8,
|
||||
F8,
|
||||
Tuple<F32, F32>,
|
||||
BF16,
|
||||
PassThrough,
|
||||
PassThrough,
|
||||
MultiplyMultiply>>>& instances);
|
||||
|
||||
void add_device_gemm_multiply_multiply_xdl_f8_f8_bf16_mk_nk_mn_mem_v2_mnkpadding_instances(
|
||||
std::vector<std::unique_ptr<DeviceGemmMultipleD<Row,
|
||||
Col,
|
||||
Tuple<Row, Col>,
|
||||
Row,
|
||||
F8,
|
||||
F8,
|
||||
Tuple<F32, F32>,
|
||||
BF16,
|
||||
PassThrough,
|
||||
PassThrough,
|
||||
MultiplyMultiply>>>& instances);
|
||||
#endif
|
||||
|
||||
template <typename ADataType,
|
||||
typename BDataType,
|
||||
typename CDataType,
|
||||
typename ALayout,
|
||||
typename BLayout,
|
||||
typename CLayout>
|
||||
struct DeviceOperationInstanceFactory<ck::tensor_operation::device::DeviceGemmMultipleD<
|
||||
ALayout,
|
||||
BLayout,
|
||||
Tuple<Row, Col>,
|
||||
CLayout,
|
||||
ADataType,
|
||||
BDataType,
|
||||
Tuple<F32, F32>,
|
||||
CDataType,
|
||||
ck::tensor_operation::element_wise::PassThrough,
|
||||
ck::tensor_operation::element_wise::PassThrough,
|
||||
ck::tensor_operation::element_wise::MultiplyMultiply>>
|
||||
{
|
||||
using DeviceOp = DeviceGemmMultipleD<ALayout,
|
||||
BLayout,
|
||||
Tuple<Row, Col>,
|
||||
CLayout,
|
||||
ADataType,
|
||||
BDataType,
|
||||
Tuple<F32, F32>,
|
||||
CDataType,
|
||||
ck::tensor_operation::element_wise::PassThrough,
|
||||
ck::tensor_operation::element_wise::PassThrough,
|
||||
ck::tensor_operation::element_wise::MultiplyMultiply>;
|
||||
|
||||
static auto GetInstances()
|
||||
{
|
||||
std::vector<std::unique_ptr<DeviceOp>> op_ptrs;
|
||||
|
||||
#if(defined(CK_ENABLE_BF16) || defined(CK_ENABLE_FP8))
|
||||
if constexpr(is_same_v<ADataType, f8_t> && is_same_v<BDataType, f8_t> &&
|
||||
is_same_v<CDataType, bhalf_t>)
|
||||
{
|
||||
if constexpr(is_same_v<ALayout, Row> && is_same_v<BLayout, Col> &&
|
||||
is_same_v<CLayout, Row>)
|
||||
{
|
||||
add_device_gemm_multiply_multiply_xdl_f8_f8_bf16_mk_nk_mn_comp_default_instances(
|
||||
op_ptrs);
|
||||
add_device_gemm_multiply_multiply_xdl_f8_f8_bf16_mk_nk_mn_comp_kpadding_instances(
|
||||
op_ptrs);
|
||||
add_device_gemm_multiply_multiply_xdl_f8_f8_bf16_mk_nk_mn_comp_mnpadding_instances(
|
||||
op_ptrs);
|
||||
add_device_gemm_multiply_multiply_xdl_f8_f8_bf16_mk_nk_mn_comp_mnkpadding_instances(
|
||||
op_ptrs);
|
||||
|
||||
add_device_gemm_multiply_multiply_xdl_f8_f8_bf16_mk_nk_mn_mem_v1_default_instances(
|
||||
op_ptrs);
|
||||
add_device_gemm_multiply_multiply_xdl_f8_f8_bf16_mk_nk_mn_mem_v1_kpadding_instances(
|
||||
op_ptrs);
|
||||
add_device_gemm_multiply_multiply_xdl_f8_f8_bf16_mk_nk_mn_mem_v1_mnkpadding_instances(
|
||||
op_ptrs);
|
||||
|
||||
add_device_gemm_multiply_multiply_xdl_f8_f8_bf16_mk_nk_mn_mem_v2_default_instances(
|
||||
op_ptrs);
|
||||
add_device_gemm_multiply_multiply_xdl_f8_f8_bf16_mk_nk_mn_mem_v2_kpadding_instances(
|
||||
op_ptrs);
|
||||
add_device_gemm_multiply_multiply_xdl_f8_f8_bf16_mk_nk_mn_mem_v2_mnkpadding_instances(
|
||||
op_ptrs);
|
||||
}
|
||||
}
|
||||
#endif
|
||||
return op_ptrs;
|
||||
}
|
||||
};
|
||||
|
||||
} // namespace instance
|
||||
} // namespace device
|
||||
} // namespace tensor_operation
|
||||
} // namespace ck
|
||||
@@ -1,5 +1,5 @@
|
||||
// SPDX-License-Identifier: MIT
|
||||
// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
|
||||
// Copyright (c) 2024, Advanced Micro Devices, Inc. All rights reserved.
|
||||
|
||||
#pragma once
|
||||
|
||||
@@ -315,7 +315,7 @@ void add_device_gemm_xdl_universal_f8_f16_f16_mk_nk_mn_mem_v2_mnkpadding_instanc
|
||||
DeviceGemmV2<Row, Col, Row, F8, F16, F16, PassThrough, PassThrough, PassThrough>>>&
|
||||
instances);
|
||||
#endif
|
||||
#ifdef CK_ENABLE_FP16
|
||||
#ifdef CK_ENABLE_BF16
|
||||
void add_device_gemm_xdl_universal_bf16_bf16_bf16_mk_kn_mn_comp_default_instances(
|
||||
std::vector<std::unique_ptr<
|
||||
DeviceGemmV2<Row, Row, Row, BF16, BF16, BF16, PassThrough, PassThrough, PassThrough>>>&
|
||||
@@ -416,6 +416,57 @@ void add_device_gemm_xdl_universal_bf16_bf16_bf16_mk_nk_mn_mem_v2_mnkpadding_ins
|
||||
DeviceGemmV2<Row, Col, Row, BF16, BF16, BF16, PassThrough, PassThrough, PassThrough>>>&
|
||||
instances);
|
||||
#endif
|
||||
#if(defined(CK_ENABLE_BF16) || defined(CK_ENABLE_FP8))
|
||||
void add_device_gemm_xdl_universal_f8_f8_bf16_mk_nk_mn_comp_default_instances(
|
||||
std::vector<std::unique_ptr<
|
||||
DeviceGemmV2<Row, Col, Row, F8, F8, BF16, PassThrough, PassThrough, PassThrough>>>&
|
||||
instances);
|
||||
|
||||
void add_device_gemm_xdl_universal_f8_f8_bf16_mk_nk_mn_comp_kpadding_instances(
|
||||
std::vector<std::unique_ptr<
|
||||
DeviceGemmV2<Row, Col, Row, F8, F8, BF16, PassThrough, PassThrough, PassThrough>>>&
|
||||
instances);
|
||||
|
||||
void add_device_gemm_xdl_universal_f8_f8_bf16_mk_nk_mn_comp_mnpadding_instances(
|
||||
std::vector<std::unique_ptr<
|
||||
DeviceGemmV2<Row, Col, Row, F8, F8, BF16, PassThrough, PassThrough, PassThrough>>>&
|
||||
instances);
|
||||
|
||||
void add_device_gemm_xdl_universal_f8_f8_bf16_mk_nk_mn_comp_mnkpadding_instances(
|
||||
std::vector<std::unique_ptr<
|
||||
DeviceGemmV2<Row, Col, Row, F8, F8, BF16, PassThrough, PassThrough, PassThrough>>>&
|
||||
instances);
|
||||
|
||||
void add_device_gemm_xdl_universal_f8_f8_bf16_mk_nk_mn_mem_v1_default_instances(
|
||||
std::vector<std::unique_ptr<
|
||||
DeviceGemmV2<Row, Col, Row, F8, F8, BF16, PassThrough, PassThrough, PassThrough>>>&
|
||||
instances);
|
||||
|
||||
void add_device_gemm_xdl_universal_f8_f8_bf16_mk_nk_mn_mem_v1_kpadding_instances(
|
||||
std::vector<std::unique_ptr<
|
||||
DeviceGemmV2<Row, Col, Row, F8, F8, BF16, PassThrough, PassThrough, PassThrough>>>&
|
||||
instances);
|
||||
|
||||
void add_device_gemm_xdl_universal_f8_f8_bf16_mk_nk_mn_mem_v1_mnkpadding_instances(
|
||||
std::vector<std::unique_ptr<
|
||||
DeviceGemmV2<Row, Col, Row, F8, F8, BF16, PassThrough, PassThrough, PassThrough>>>&
|
||||
instances);
|
||||
|
||||
void add_device_gemm_xdl_universal_f8_f8_bf16_mk_nk_mn_mem_v2_default_instances(
|
||||
std::vector<std::unique_ptr<
|
||||
DeviceGemmV2<Row, Col, Row, F8, F8, BF16, PassThrough, PassThrough, PassThrough>>>&
|
||||
instances);
|
||||
|
||||
void add_device_gemm_xdl_universal_f8_f8_bf16_mk_nk_mn_mem_v2_kpadding_instances(
|
||||
std::vector<std::unique_ptr<
|
||||
DeviceGemmV2<Row, Col, Row, F8, F8, BF16, PassThrough, PassThrough, PassThrough>>>&
|
||||
instances);
|
||||
|
||||
void add_device_gemm_xdl_universal_f8_f8_bf16_mk_nk_mn_mem_v2_mnkpadding_instances(
|
||||
std::vector<std::unique_ptr<
|
||||
DeviceGemmV2<Row, Col, Row, F8, F8, BF16, PassThrough, PassThrough, PassThrough>>>&
|
||||
instances);
|
||||
#endif
|
||||
|
||||
template <typename ADataType,
|
||||
typename BDataType,
|
||||
@@ -596,7 +647,7 @@ struct DeviceOperationInstanceFactory<
|
||||
}
|
||||
}
|
||||
#endif
|
||||
#ifdef CK_ENABLE_FP16
|
||||
#ifdef CK_ENABLE_BF16
|
||||
if constexpr(is_same_v<ADataType, bhalf_t> && is_same_v<BDataType, bhalf_t> &&
|
||||
is_same_v<CDataType, bhalf_t>)
|
||||
{
|
||||
@@ -653,6 +704,33 @@ struct DeviceOperationInstanceFactory<
|
||||
op_ptrs);
|
||||
}
|
||||
}
|
||||
#endif
|
||||
#if(defined(CK_ENABLE_BF16) || defined(CK_ENABLE_FP8))
|
||||
if constexpr(is_same_v<ADataType, f8_t> && is_same_v<BDataType, f8_t> &&
|
||||
is_same_v<CDataType, bhalf_t>)
|
||||
{
|
||||
if constexpr(is_same_v<ALayout, Row> && is_same_v<BLayout, Col> &&
|
||||
is_same_v<CLayout, Row>)
|
||||
{
|
||||
add_device_gemm_xdl_universal_f8_f8_bf16_mk_nk_mn_comp_default_instances(op_ptrs);
|
||||
add_device_gemm_xdl_universal_f8_f8_bf16_mk_nk_mn_comp_kpadding_instances(op_ptrs);
|
||||
add_device_gemm_xdl_universal_f8_f8_bf16_mk_nk_mn_comp_mnpadding_instances(op_ptrs);
|
||||
add_device_gemm_xdl_universal_f8_f8_bf16_mk_nk_mn_comp_mnkpadding_instances(
|
||||
op_ptrs);
|
||||
|
||||
add_device_gemm_xdl_universal_f8_f8_bf16_mk_nk_mn_mem_v1_default_instances(op_ptrs);
|
||||
add_device_gemm_xdl_universal_f8_f8_bf16_mk_nk_mn_mem_v1_kpadding_instances(
|
||||
op_ptrs);
|
||||
add_device_gemm_xdl_universal_f8_f8_bf16_mk_nk_mn_mem_v1_mnkpadding_instances(
|
||||
op_ptrs);
|
||||
|
||||
add_device_gemm_xdl_universal_f8_f8_bf16_mk_nk_mn_mem_v2_default_instances(op_ptrs);
|
||||
add_device_gemm_xdl_universal_f8_f8_bf16_mk_nk_mn_mem_v2_kpadding_instances(
|
||||
op_ptrs);
|
||||
add_device_gemm_xdl_universal_f8_f8_bf16_mk_nk_mn_mem_v2_mnkpadding_instances(
|
||||
op_ptrs);
|
||||
}
|
||||
}
|
||||
#endif
|
||||
return op_ptrs;
|
||||
}
|
||||
|
||||
Reference in New Issue
Block a user