Skip to content

GitLab

Unverified Commit 3d61f89a authored by Illia Silin's avatar Illia Silin Committed by GitHub
Browse files

Merge pull request #134 from ROCm/merge_from_public 

Merge from public
parents c160c6cf 4558a3f8
Hide whitespace changes
Inline Side-by-side
Showing with 1829 additions and 157 deletions
library/include/ck/library/tensor_operation_instance/gpu/gemm_ab_scale.hpp 0 → 100644
View file @ 3d61f89a
// 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
library/include/ck/library/tensor_operation_instance/gpu/gemm_multiply_multiply.hpp 0 → 100644
View file @ 3d61f89a
// 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<DeviceGemmMultipleDSplitK<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<DeviceGemmMultipleDSplitK<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<DeviceGemmMultipleDSplitK<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<DeviceGemmMultipleDSplitK<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<DeviceGemmMultipleDSplitK<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<DeviceGemmMultipleDSplitK<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::DeviceGemmMultipleDSplitK<
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 =
DeviceGemmMultipleDSplitK<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_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_v2_default_instances(
op_ptrs);
add_device_gemm_multiply_multiply_xdl_f8_f8_bf16_mk_nk_mn_mem_v2_kpadding_instances(
op_ptrs);
}
}
#endif
return op_ptrs;
}
};
} // namespace instance
} // namespace device
} // namespace tensor_operation
} // namespace ck
library/include/ck/library/tensor_operation_instance/gpu/gemm_universal.hpp
View file @ 3d61f89a
// 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
......@@ -77,16 +77,6 @@ void add_device_gemm_xdl_universal_f16_f16_f16_mk_nk_mn_comp_kpadding_instances(
DeviceGemmV2<Row, Col, Row, F16, F16, F16, PassThrough, PassThrough, PassThrough>>>&
instances);
void add_device_gemm_xdl_universal_f16_f16_f16_mk_nk_mn_comp_mnpadding_instances(
std::vector<std::unique_ptr<
DeviceGemmV2<Row, Col, Row, F16, F16, F16, PassThrough, PassThrough, PassThrough>>>&
instances);
void add_device_gemm_xdl_universal_f16_f16_f16_mk_nk_mn_comp_mnkpadding_instances(
std::vector<std::unique_ptr<
DeviceGemmV2<Row, Col, Row, F16, F16, F16, PassThrough, PassThrough, PassThrough>>>&
instances);
void add_device_gemm_xdl_universal_f16_f16_f16_mk_nk_mn_mem_v1_default_instances(
std::vector<std::unique_ptr<
DeviceGemmV2<Row, Col, Row, F16, F16, F16, PassThrough, PassThrough, PassThrough>>>&
......@@ -97,11 +87,6 @@ void add_device_gemm_xdl_universal_f16_f16_f16_mk_nk_mn_mem_v1_kpadding_instance
DeviceGemmV2<Row, Col, Row, F16, F16, F16, PassThrough, PassThrough, PassThrough>>>&
instances);
void add_device_gemm_xdl_universal_f16_f16_f16_mk_nk_mn_mem_v1_mnkpadding_instances(
std::vector<std::unique_ptr<
DeviceGemmV2<Row, Col, Row, F16, F16, F16, PassThrough, PassThrough, PassThrough>>>&
instances);
void add_device_gemm_xdl_universal_f16_f16_f16_mk_nk_mn_mem_v2_default_instances(
std::vector<std::unique_ptr<
DeviceGemmV2<Row, Col, Row, F16, F16, F16, PassThrough, PassThrough, PassThrough>>>&
......@@ -111,13 +96,8 @@ void add_device_gemm_xdl_universal_f16_f16_f16_mk_nk_mn_mem_v2_kpadding_instance
std::vector<std::unique_ptr<
DeviceGemmV2<Row, Col, Row, F16, F16, F16, PassThrough, PassThrough, PassThrough>>>&
instances);
void add_device_gemm_xdl_universal_f16_f16_f16_mk_nk_mn_mem_v2_mnkpadding_instances(
std::vector<std::unique_ptr<
DeviceGemmV2<Row, Col, Row, F16, F16, F16, PassThrough, PassThrough, PassThrough>>>&
instances);
#endif
#if(defined(CK_ENABLE_FP16) || defined(CK_ENABLE_FP8))
#if(defined(CK_ENABLE_FP16) && defined(CK_ENABLE_FP8))
void add_device_gemm_xdl_universal_f16_f8_f16_mk_kn_mn_comp_default_instances(
std::vector<std::unique_ptr<
DeviceGemmV2<Row, Row, Row, F16, F8, F16, PassThrough, PassThrough, PassThrough>>>&
......@@ -177,16 +157,6 @@ void add_device_gemm_xdl_universal_f16_f8_f16_mk_nk_mn_comp_kpadding_instances(
DeviceGemmV2<Row, Col, Row, F16, F8, F16, PassThrough, PassThrough, PassThrough>>>&
instances);
void add_device_gemm_xdl_universal_f16_f8_f16_mk_nk_mn_comp_mnpadding_instances(
std::vector<std::unique_ptr<
DeviceGemmV2<Row, Col, Row, F16, F8, F16, PassThrough, PassThrough, PassThrough>>>&
instances);
void add_device_gemm_xdl_universal_f16_f8_f16_mk_nk_mn_comp_mnkpadding_instances(
std::vector<std::unique_ptr<
DeviceGemmV2<Row, Col, Row, F16, F8, F16, PassThrough, PassThrough, PassThrough>>>&
instances);
void add_device_gemm_xdl_universal_f16_f8_f16_mk_nk_mn_mem_v1_default_instances(
std::vector<std::unique_ptr<
DeviceGemmV2<Row, Col, Row, F16, F8, F16, PassThrough, PassThrough, PassThrough>>>&
......@@ -196,12 +166,6 @@ void add_device_gemm_xdl_universal_f16_f8_f16_mk_nk_mn_mem_v1_kpadding_instances
std::vector<std::unique_ptr<
DeviceGemmV2<Row, Col, Row, F16, F8, F16, PassThrough, PassThrough, PassThrough>>>&
instances);
void add_device_gemm_xdl_universal_f16_f8_f16_mk_nk_mn_mem_v1_mnkpadding_instances(
std::vector<std::unique_ptr<
DeviceGemmV2<Row, Col, Row, F16, F8, F16, PassThrough, PassThrough, PassThrough>>>&
instances);
void add_device_gemm_xdl_universal_f16_f8_f16_mk_nk_mn_mem_v2_default_instances(
std::vector<std::unique_ptr<
DeviceGemmV2<Row, Col, Row, F16, F8, F16, PassThrough, PassThrough, PassThrough>>>&
......@@ -212,10 +176,6 @@ void add_device_gemm_xdl_universal_f16_f8_f16_mk_nk_mn_mem_v2_kpadding_instances
DeviceGemmV2<Row, Col, Row, F16, F8, F16, PassThrough, PassThrough, PassThrough>>>&
instances);
void add_device_gemm_xdl_universal_f16_f8_f16_mk_nk_mn_mem_v2_mnkpadding_instances(
std::vector<std::unique_ptr<
DeviceGemmV2<Row, Col, Row, F16, F8, F16, PassThrough, PassThrough, PassThrough>>>&
instances);
void add_device_gemm_xdl_universal_f8_f16_f16_mk_kn_mn_comp_default_instances(
std::vector<std::unique_ptr<
DeviceGemmV2<Row, Row, Row, F8, F16, F16, PassThrough, PassThrough, PassThrough>>>&
......@@ -275,16 +235,6 @@ void add_device_gemm_xdl_universal_f8_f16_f16_mk_nk_mn_comp_kpadding_instances(
DeviceGemmV2<Row, Col, Row, F8, F16, F16, PassThrough, PassThrough, PassThrough>>>&
instances);
void add_device_gemm_xdl_universal_f8_f16_f16_mk_nk_mn_comp_mnpadding_instances(
std::vector<std::unique_ptr<
DeviceGemmV2<Row, Col, Row, F8, F16, F16, PassThrough, PassThrough, PassThrough>>>&
instances);
void add_device_gemm_xdl_universal_f8_f16_f16_mk_nk_mn_comp_mnkpadding_instances(
std::vector<std::unique_ptr<
DeviceGemmV2<Row, Col, Row, F8, F16, F16, PassThrough, PassThrough, PassThrough>>>&
instances);
void add_device_gemm_xdl_universal_f8_f16_f16_mk_nk_mn_mem_v1_default_instances(
std::vector<std::unique_ptr<
DeviceGemmV2<Row, Col, Row, F8, F16, F16, PassThrough, PassThrough, PassThrough>>>&
......@@ -295,11 +245,6 @@ void add_device_gemm_xdl_universal_f8_f16_f16_mk_nk_mn_mem_v1_kpadding_instances
DeviceGemmV2<Row, Col, Row, F8, F16, F16, PassThrough, PassThrough, PassThrough>>>&
instances);
void add_device_gemm_xdl_universal_f8_f16_f16_mk_nk_mn_mem_v1_mnkpadding_instances(
std::vector<std::unique_ptr<
DeviceGemmV2<Row, Col, Row, F8, F16, F16, PassThrough, PassThrough, PassThrough>>>&
instances);
void add_device_gemm_xdl_universal_f8_f16_f16_mk_nk_mn_mem_v2_default_instances(
std::vector<std::unique_ptr<
DeviceGemmV2<Row, Col, Row, F8, F16, F16, PassThrough, PassThrough, PassThrough>>>&
......@@ -309,13 +254,8 @@ void add_device_gemm_xdl_universal_f8_f16_f16_mk_nk_mn_mem_v2_kpadding_instances
std::vector<std::unique_ptr<
DeviceGemmV2<Row, Col, Row, F8, F16, F16, PassThrough, PassThrough, PassThrough>>>&
instances);
void add_device_gemm_xdl_universal_f8_f16_f16_mk_nk_mn_mem_v2_mnkpadding_instances(
std::vector<std::unique_ptr<
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>>>&
......@@ -376,44 +316,100 @@ void add_device_gemm_xdl_universal_bf16_bf16_bf16_mk_nk_mn_comp_kpadding_instanc
DeviceGemmV2<Row, Col, Row, BF16, BF16, BF16, PassThrough, PassThrough, PassThrough>>>&
instances);
void add_device_gemm_xdl_universal_bf16_bf16_bf16_mk_nk_mn_comp_mnpadding_instances(
void add_device_gemm_xdl_universal_bf16_bf16_bf16_mk_nk_mn_mem_v1_default_instances(
std::vector<std::unique_ptr<
DeviceGemmV2<Row, Col, Row, BF16, BF16, BF16, PassThrough, PassThrough, PassThrough>>>&
instances);
void add_device_gemm_xdl_universal_bf16_bf16_bf16_mk_nk_mn_comp_mnkpadding_instances(
void add_device_gemm_xdl_universal_bf16_bf16_bf16_mk_nk_mn_mem_v1_kpadding_instances(
std::vector<std::unique_ptr<
DeviceGemmV2<Row, Col, Row, BF16, BF16, BF16, PassThrough, PassThrough, PassThrough>>>&
instances);
void add_device_gemm_xdl_universal_bf16_bf16_bf16_mk_nk_mn_mem_v1_default_instances(
void add_device_gemm_xdl_universal_bf16_bf16_bf16_mk_nk_mn_mem_v2_default_instances(
std::vector<std::unique_ptr<
DeviceGemmV2<Row, Col, Row, BF16, BF16, BF16, PassThrough, PassThrough, PassThrough>>>&
instances);
void add_device_gemm_xdl_universal_bf16_bf16_bf16_mk_nk_mn_mem_v1_kpadding_instances(
void add_device_gemm_xdl_universal_bf16_bf16_bf16_mk_nk_mn_mem_v2_kpadding_instances(
std::vector<std::unique_ptr<
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_kn_mn_comp_default_instances(
std::vector<std::unique_ptr<
DeviceGemmV2<Row, Row, Row, F8, F8, BF16, PassThrough, PassThrough, PassThrough>>>&
instances);
void add_device_gemm_xdl_universal_bf16_bf16_bf16_mk_nk_mn_mem_v1_mnkpadding_instances(
void add_device_gemm_xdl_universal_f8_f8_bf16_mk_kn_mn_comp_kpadding_instances(
std::vector<std::unique_ptr<
DeviceGemmV2<Row, Col, Row, BF16, BF16, BF16, PassThrough, PassThrough, PassThrough>>>&
DeviceGemmV2<Row, Row, Row, F8, F8, BF16, PassThrough, PassThrough, PassThrough>>>&
instances);
void add_device_gemm_xdl_universal_bf16_bf16_bf16_mk_nk_mn_mem_v2_default_instances(
void add_device_gemm_xdl_universal_f8_f8_bf16_mk_kn_mn_comp_nkpadding_instances(
std::vector<std::unique_ptr<
DeviceGemmV2<Row, Col, Row, BF16, BF16, BF16, PassThrough, PassThrough, PassThrough>>>&
DeviceGemmV2<Row, Row, Row, F8, F8, BF16, PassThrough, PassThrough, PassThrough>>>&
instances);
void add_device_gemm_xdl_universal_bf16_bf16_bf16_mk_nk_mn_mem_v2_kpadding_instances(
void add_device_gemm_xdl_universal_f8_f8_bf16_mk_kn_mn_mem_v1_default_instances(
std::vector<std::unique_ptr<
DeviceGemmV2<Row, Col, Row, BF16, BF16, BF16, PassThrough, PassThrough, PassThrough>>>&
DeviceGemmV2<Row, Row, Row, F8, F8, BF16, PassThrough, PassThrough, PassThrough>>>&
instances);
void add_device_gemm_xdl_universal_bf16_bf16_bf16_mk_nk_mn_mem_v2_mnkpadding_instances(
void add_device_gemm_xdl_universal_f8_f8_bf16_mk_kn_mn_mem_v1_kpadding_instances(
std::vector<std::unique_ptr<
DeviceGemmV2<Row, Col, Row, BF16, BF16, BF16, PassThrough, PassThrough, PassThrough>>>&
DeviceGemmV2<Row, Row, Row, F8, F8, BF16, PassThrough, PassThrough, PassThrough>>>&
instances);
void add_device_gemm_xdl_universal_f8_f8_bf16_mk_kn_mn_mem_v1_nkpadding_instances(
std::vector<std::unique_ptr<
DeviceGemmV2<Row, Row, Row, F8, F8, BF16, PassThrough, PassThrough, PassThrough>>>&
instances);
void add_device_gemm_xdl_universal_f8_f8_bf16_mk_kn_mn_mem_v2_default_instances(
std::vector<std::unique_ptr<
DeviceGemmV2<Row, Row, Row, F8, F8, BF16, PassThrough, PassThrough, PassThrough>>>&
instances);
void add_device_gemm_xdl_universal_f8_f8_bf16_mk_kn_mn_mem_v2_kpadding_instances(
std::vector<std::unique_ptr<
DeviceGemmV2<Row, Row, Row, F8, F8, BF16, PassThrough, PassThrough, PassThrough>>>&
instances);
void add_device_gemm_xdl_universal_f8_f8_bf16_mk_kn_mn_mem_v2_nkpadding_instances(
std::vector<std::unique_ptr<
DeviceGemmV2<Row, Row, Row, F8, F8, BF16, PassThrough, PassThrough, PassThrough>>>&
instances);
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_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_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);
#endif
......@@ -481,28 +477,20 @@ struct DeviceOperationInstanceFactory<
{
add_device_gemm_xdl_universal_f16_f16_f16_mk_nk_mn_comp_default_instances(op_ptrs);
add_device_gemm_xdl_universal_f16_f16_f16_mk_nk_mn_comp_kpadding_instances(op_ptrs);
add_device_gemm_xdl_universal_f16_f16_f16_mk_nk_mn_comp_mnpadding_instances(
op_ptrs);
add_device_gemm_xdl_universal_f16_f16_f16_mk_nk_mn_comp_mnkpadding_instances(
op_ptrs);
add_device_gemm_xdl_universal_f16_f16_f16_mk_nk_mn_mem_v1_default_instances(
op_ptrs);
add_device_gemm_xdl_universal_f16_f16_f16_mk_nk_mn_mem_v1_kpadding_instances(
op_ptrs);
add_device_gemm_xdl_universal_f16_f16_f16_mk_nk_mn_mem_v1_mnkpadding_instances(
op_ptrs);
add_device_gemm_xdl_universal_f16_f16_f16_mk_nk_mn_mem_v2_default_instances(
op_ptrs);
add_device_gemm_xdl_universal_f16_f16_f16_mk_nk_mn_mem_v2_kpadding_instances(
op_ptrs);
add_device_gemm_xdl_universal_f16_f16_f16_mk_nk_mn_mem_v2_mnkpadding_instances(
op_ptrs);
}
}
#endif
#if(defined(CK_ENABLE_FP16) || defined(CK_ENABLE_FP8))
#if(defined(CK_ENABLE_FP16) && defined(CK_ENABLE_FP8))
if constexpr(is_same_v<ADataType, half_t> && is_same_v<BDataType, f8_t> &&
is_same_v<CDataType, half_t>)
{
......@@ -511,21 +499,14 @@ struct DeviceOperationInstanceFactory<
{
add_device_gemm_xdl_universal_f16_f8_f16_mk_nk_mn_comp_default_instances(op_ptrs);
add_device_gemm_xdl_universal_f16_f8_f16_mk_nk_mn_comp_kpadding_instances(op_ptrs);
add_device_gemm_xdl_universal_f16_f8_f16_mk_nk_mn_comp_mnpadding_instances(op_ptrs);
add_device_gemm_xdl_universal_f16_f8_f16_mk_nk_mn_comp_mnkpadding_instances(
op_ptrs);
add_device_gemm_xdl_universal_f16_f8_f16_mk_nk_mn_mem_v1_default_instances(op_ptrs);
add_device_gemm_xdl_universal_f16_f8_f16_mk_nk_mn_mem_v1_kpadding_instances(
op_ptrs);
add_device_gemm_xdl_universal_f16_f8_f16_mk_nk_mn_mem_v1_mnkpadding_instances(
op_ptrs);
add_device_gemm_xdl_universal_f16_f8_f16_mk_nk_mn_mem_v2_default_instances(op_ptrs);
add_device_gemm_xdl_universal_f16_f8_f16_mk_nk_mn_mem_v2_kpadding_instances(
op_ptrs);
add_device_gemm_xdl_universal_f16_f8_f16_mk_nk_mn_mem_v2_mnkpadding_instances(
op_ptrs);
}
else if constexpr(is_same_v<ALayout, Row> && is_same_v<BLayout, Row> &&
is_same_v<CLayout, Row>)
......@@ -557,21 +538,14 @@ struct DeviceOperationInstanceFactory<
{
add_device_gemm_xdl_universal_f8_f16_f16_mk_nk_mn_comp_default_instances(op_ptrs);
add_device_gemm_xdl_universal_f8_f16_f16_mk_nk_mn_comp_kpadding_instances(op_ptrs);
add_device_gemm_xdl_universal_f8_f16_f16_mk_nk_mn_comp_mnpadding_instances(op_ptrs);
add_device_gemm_xdl_universal_f8_f16_f16_mk_nk_mn_comp_mnkpadding_instances(
op_ptrs);
add_device_gemm_xdl_universal_f8_f16_f16_mk_nk_mn_mem_v1_default_instances(op_ptrs);
add_device_gemm_xdl_universal_f8_f16_f16_mk_nk_mn_mem_v1_kpadding_instances(
op_ptrs);
add_device_gemm_xdl_universal_f8_f16_f16_mk_nk_mn_mem_v1_mnkpadding_instances(
op_ptrs);
add_device_gemm_xdl_universal_f8_f16_f16_mk_nk_mn_mem_v2_default_instances(op_ptrs);
add_device_gemm_xdl_universal_f8_f16_f16_mk_nk_mn_mem_v2_kpadding_instances(
op_ptrs);
add_device_gemm_xdl_universal_f8_f16_f16_mk_nk_mn_mem_v2_mnkpadding_instances(
op_ptrs);
}
else if constexpr(is_same_v<ALayout, Row> && is_same_v<BLayout, Row> &&
is_same_v<CLayout, Row>)
......@@ -596,7 +570,7 @@ 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>)
{
......@@ -633,23 +607,54 @@ struct DeviceOperationInstanceFactory<
op_ptrs);
add_device_gemm_xdl_universal_bf16_bf16_bf16_mk_nk_mn_comp_kpadding_instances(
op_ptrs);
add_device_gemm_xdl_universal_bf16_bf16_bf16_mk_nk_mn_comp_mnpadding_instances(
op_ptrs);
add_device_gemm_xdl_universal_bf16_bf16_bf16_mk_nk_mn_comp_mnkpadding_instances(
op_ptrs);
add_device_gemm_xdl_universal_bf16_bf16_bf16_mk_nk_mn_mem_v1_default_instances(
op_ptrs);
add_device_gemm_xdl_universal_bf16_bf16_bf16_mk_nk_mn_mem_v1_kpadding_instances(
op_ptrs);
add_device_gemm_xdl_universal_bf16_bf16_bf16_mk_nk_mn_mem_v1_mnkpadding_instances(
op_ptrs);
add_device_gemm_xdl_universal_bf16_bf16_bf16_mk_nk_mn_mem_v2_default_instances(
op_ptrs);
add_device_gemm_xdl_universal_bf16_bf16_bf16_mk_nk_mn_mem_v2_kpadding_instances(
op_ptrs);
add_device_gemm_xdl_universal_bf16_bf16_bf16_mk_nk_mn_mem_v2_mnkpadding_instances(
}
}
#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, Row> &&
is_same_v<CLayout, Row>)
{
add_device_gemm_xdl_universal_f8_f8_bf16_mk_kn_mn_comp_default_instances(op_ptrs);
add_device_gemm_xdl_universal_f8_f8_bf16_mk_kn_mn_comp_kpadding_instances(op_ptrs);
add_device_gemm_xdl_universal_f8_f8_bf16_mk_kn_mn_comp_nkpadding_instances(op_ptrs);
add_device_gemm_xdl_universal_f8_f8_bf16_mk_kn_mn_mem_v1_default_instances(op_ptrs);
add_device_gemm_xdl_universal_f8_f8_bf16_mk_kn_mn_mem_v1_kpadding_instances(
op_ptrs);
add_device_gemm_xdl_universal_f8_f8_bf16_mk_kn_mn_mem_v1_nkpadding_instances(
op_ptrs);
add_device_gemm_xdl_universal_f8_f8_bf16_mk_kn_mn_mem_v2_default_instances(op_ptrs);
add_device_gemm_xdl_universal_f8_f8_bf16_mk_kn_mn_mem_v2_kpadding_instances(
op_ptrs);
add_device_gemm_xdl_universal_f8_f8_bf16_mk_kn_mn_mem_v2_nkpadding_instances(
op_ptrs);
}
else 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_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_v2_default_instances(op_ptrs);
add_device_gemm_xdl_universal_f8_f8_bf16_mk_nk_mn_mem_v2_kpadding_instances(
op_ptrs);
}
}
......
library/include/ck/library/tensor_operation_instance/gpu/gemm_universal_reduce.hpp 0 → 100644
View file @ 3d61f89a
// SPDX-License-Identifier: MIT
// Copyright (c) 2018-2024, 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_xdl_cshuffle_v3r1.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 {
using DsLayout = ck::Tuple<>;
using DsDataType = ck::Tuple<>;
#ifdef CK_ENABLE_FP16
void add_device_gemm_xdl_universal_reduce_f16_f16_f16_mk_kn_mn_comp_default_instances(
std::vector<std::unique_ptr<DeviceGemmV2R1<Row,
Row,
DsLayout,
Row,
F16,
F16,
DsDataType,
F16,
PassThrough,
PassThrough,
PassThrough>>>& instances);
void add_device_gemm_xdl_universal_reduce_f16_f16_f16_mk_kn_mn_comp_kpadding_instances(
std::vector<std::unique_ptr<DeviceGemmV2R1<Row,
Row,
DsLayout,
Row,
F16,
F16,
DsDataType,
F16,
PassThrough,
PassThrough,
PassThrough>>>& instances);
void add_device_gemm_xdl_universal_reduce_f16_f16_f16_mk_kn_mn_comp_mnpadding_instances(
std::vector<std::unique_ptr<DeviceGemmV2R1<Row,
Row,
DsLayout,
Row,
F16,
F16,
DsDataType,
F16,
PassThrough,
PassThrough,
PassThrough>>>& instances);
void add_device_gemm_xdl_universal_reduce_f16_f16_f16_mk_kn_mn_comp_mnkpadding_instances(
std::vector<std::unique_ptr<DeviceGemmV2R1<Row,
Row,
DsLayout,
Row,
F16,
F16,
DsDataType,
F16,
PassThrough,
PassThrough,
PassThrough>>>& instances);
void add_device_gemm_xdl_universal_reduce_f16_f16_f16_mk_kn_mn_mem_v1_default_instances(
std::vector<std::unique_ptr<DeviceGemmV2R1<Row,
Row,
DsLayout,
Row,
F16,
F16,
DsDataType,
F16,
PassThrough,
PassThrough,
PassThrough>>>& instances);
void add_device_gemm_xdl_universal_reduce_f16_f16_f16_mk_kn_mn_mem_v1_kpadding_instances(
std::vector<std::unique_ptr<DeviceGemmV2R1<Row,
Row,
DsLayout,
Row,
F16,
F16,
DsDataType,
F16,
PassThrough,
PassThrough,
PassThrough>>>& instances);
void add_device_gemm_xdl_universal_reduce_f16_f16_f16_mk_kn_mn_mem_v1_mnkpadding_instances(
std::vector<std::unique_ptr<DeviceGemmV2R1<Row,
Row,
DsLayout,
Row,
F16,
F16,
DsDataType,
F16,
PassThrough,
PassThrough,
PassThrough>>>& instances);
void add_device_gemm_xdl_universal_reduce_f16_f16_f16_mk_kn_mn_mem_v2_default_instances(
std::vector<std::unique_ptr<DeviceGemmV2R1<Row,
Row,
DsLayout,
Row,
F16,
F16,
DsDataType,
F16,
PassThrough,
PassThrough,
PassThrough>>>& instances);
void add_device_gemm_xdl_universal_reduce_f16_f16_f16_mk_kn_mn_mem_v2_kpadding_instances(
std::vector<std::unique_ptr<DeviceGemmV2R1<Row,
Row,
DsLayout,
Row,
F16,
F16,
DsDataType,
F16,
PassThrough,
PassThrough,
PassThrough>>>& instances);
void add_device_gemm_xdl_universal_reduce_f16_f16_f16_mk_kn_mn_mem_v2_mnkpadding_instances(
std::vector<std::unique_ptr<DeviceGemmV2R1<Row,
Row,
DsLayout,
Row,
F16,
F16,
DsDataType,
F16,
PassThrough,
PassThrough,
PassThrough>>>& instances);
#endif
#if(defined(CK_ENABLE_BF16) || defined(CK_ENABLE_INT8))
void add_device_gemm_xdl_universal_reduce_bf16_i8_bf16_mk_kn_mn_comp_default_instances(
std::vector<std::unique_ptr<DeviceGemmV2R1<Row,
Row,
DsLayout,
Row,
BF16,
I8,
DsDataType,
BF16,
PassThrough,
PassThrough,
PassThrough>>>& instances);
void add_device_gemm_xdl_universal_reduce_bf16_i8_bf16_mk_kn_mn_comp_kpadding_instances(
std::vector<std::unique_ptr<DeviceGemmV2R1<Row,
Row,
DsLayout,
Row,
BF16,
I8,
DsDataType,
BF16,
PassThrough,
PassThrough,
PassThrough>>>& instances);
void add_device_gemm_xdl_universal_reduce_bf16_i8_bf16_mk_kn_mn_comp_mnkpadding_instances(
std::vector<std::unique_ptr<DeviceGemmV2R1<Row,
Row,
DsLayout,
Row,
BF16,
I8,
DsDataType,
BF16,
PassThrough,
PassThrough,
PassThrough>>>& instances);
void add_device_gemm_xdl_universal_reduce_bf16_i8_bf16_mk_kn_mn_mem_v2_default_instances(
std::vector<std::unique_ptr<DeviceGemmV2R1<Row,
Row,
DsLayout,
Row,
BF16,
I8,
DsDataType,
BF16,
PassThrough,
PassThrough,
PassThrough>>>& instances);
void add_device_gemm_xdl_universal_reduce_bf16_i8_bf16_mk_kn_mn_mem_v2_kpadding_instances(
std::vector<std::unique_ptr<DeviceGemmV2R1<Row,
Row,
DsLayout,
Row,
BF16,
I8,
DsDataType,
BF16,
PassThrough,
PassThrough,
PassThrough>>>& instances);
void add_device_gemm_xdl_universal_reduce_bf16_i8_bf16_mk_kn_mn_comp_mnpadding_instances(
std::vector<std::unique_ptr<DeviceGemmV2R1<Row,
Row,
DsLayout,
Row,
BF16,
I8,
DsDataType,
BF16,
PassThrough,
PassThrough,
PassThrough>>>& instances);
void add_device_gemm_xdl_universal_reduce_bf16_i8_bf16_mk_kn_mn_mem_v2_mnkpadding_instances(
std::vector<std::unique_ptr<DeviceGemmV2R1<Row,
Row,
DsLayout,
Row,
BF16,
I8,
DsDataType,
BF16,
PassThrough,
PassThrough,
PassThrough>>>& instances);
#endif
#ifdef CK_ENABLE_BF16
void add_device_gemm_xdl_universal_reduce_bf16_bf16_bf16_mk_kn_mn_comp_default_instances(
std::vector<std::unique_ptr<DeviceGemmV2R1<Row,
Row,
DsLayout,
Row,
BF16,
BF16,
DsDataType,
BF16,
PassThrough,
PassThrough,
PassThrough>>>& instances);
void add_device_gemm_xdl_universal_reduce_bf16_bf16_bf16_mk_kn_mn_comp_kpadding_instances(
std::vector<std::unique_ptr<DeviceGemmV2R1<Row,
Row,
DsLayout,
Row,
BF16,
BF16,
DsDataType,
BF16,
PassThrough,
PassThrough,
PassThrough>>>& instances);
void add_device_gemm_xdl_universal_reduce_bf16_bf16_bf16_mk_kn_mn_comp_mnkpadding_instances(
std::vector<std::unique_ptr<DeviceGemmV2R1<Row,
Row,
DsLayout,
Row,
BF16,
BF16,
DsDataType,
BF16,
PassThrough,
PassThrough,
PassThrough>>>& instances);
void add_device_gemm_xdl_universal_reduce_bf16_bf16_bf16_mk_kn_mn_mem_v2_default_instances(
std::vector<std::unique_ptr<DeviceGemmV2R1<Row,
Row,
DsLayout,
Row,
BF16,
BF16,
DsDataType,
BF16,
PassThrough,
PassThrough,
PassThrough>>>& instances);
void add_device_gemm_xdl_universal_reduce_bf16_bf16_bf16_mk_kn_mn_mem_v2_kpadding_instances(
std::vector<std::unique_ptr<DeviceGemmV2R1<Row,
Row,
DsLayout,
Row,
BF16,
BF16,
DsDataType,
BF16,
PassThrough,
PassThrough,
PassThrough>>>& instances);
void add_device_gemm_xdl_universal_reduce_bf16_bf16_bf16_mk_kn_mn_comp_mnpadding_instances(
std::vector<std::unique_ptr<DeviceGemmV2R1<Row,
Row,
DsLayout,
Row,
BF16,
BF16,
DsDataType,
BF16,
PassThrough,
PassThrough,
PassThrough>>>& instances);
void add_device_gemm_xdl_universal_reduce_bf16_bf16_bf16_mk_kn_mn_mem_v2_mnkpadding_instances(
std::vector<std::unique_ptr<DeviceGemmV2R1<Row,
Row,
DsLayout,
Row,
BF16,
BF16,
DsDataType,
BF16,
PassThrough,
PassThrough,
PassThrough>>>& instances);
#endif
template <typename ADataType,
typename BDataType,
typename DsDataType,
typename CDataType,
typename ALayout,
typename BLayout,
typename DsLayout,
typename CLayout>
struct DeviceOperationInstanceFactory<
ck::tensor_operation::device::DeviceGemmV2R1<ALayout,
BLayout,
DsLayout,
CLayout,
ADataType,
BDataType,
DsDataType,
CDataType,
ck::tensor_operation::element_wise::PassThrough,
ck::tensor_operation::element_wise::PassThrough,
ck::tensor_operation::element_wise::PassThrough>>
{
using DeviceOp = DeviceGemmV2R1<ALayout,
BLayout,
DsLayout,
CLayout,
ADataType,
BDataType,
DsDataType,
CDataType,
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;
#ifdef CK_ENABLE_FP16
if constexpr(is_same_v<ADataType, half_t> && is_same_v<BDataType, half_t> &&
is_same_v<CDataType, half_t>)
{
if constexpr(is_same_v<ALayout, Row> && is_same_v<BLayout, Row> &&
is_same_v<CLayout, Row>)
{
add_device_gemm_xdl_universal_reduce_f16_f16_f16_mk_kn_mn_comp_default_instances(
op_ptrs);
add_device_gemm_xdl_universal_reduce_f16_f16_f16_mk_kn_mn_comp_kpadding_instances(
op_ptrs);
add_device_gemm_xdl_universal_reduce_f16_f16_f16_mk_kn_mn_comp_mnpadding_instances(
op_ptrs);
add_device_gemm_xdl_universal_reduce_f16_f16_f16_mk_kn_mn_comp_mnkpadding_instances(
op_ptrs);
add_device_gemm_xdl_universal_reduce_f16_f16_f16_mk_kn_mn_mem_v1_default_instances(
op_ptrs);
add_device_gemm_xdl_universal_reduce_f16_f16_f16_mk_kn_mn_mem_v1_kpadding_instances(
op_ptrs);
add_device_gemm_xdl_universal_reduce_f16_f16_f16_mk_kn_mn_mem_v1_mnkpadding_instances(
op_ptrs);
add_device_gemm_xdl_universal_reduce_f16_f16_f16_mk_kn_mn_mem_v2_default_instances(
op_ptrs);
add_device_gemm_xdl_universal_reduce_f16_f16_f16_mk_kn_mn_mem_v2_kpadding_instances(
op_ptrs);
add_device_gemm_xdl_universal_reduce_f16_f16_f16_mk_kn_mn_mem_v2_mnkpadding_instances(
op_ptrs);
}
}
#endif
#if(defined(CK_ENABLE_BF16) || defined(CK_ENABLE_INT8))
if constexpr(is_same_v<ADataType, bhalf_t> && is_same_v<BDataType, int8_t> &&
is_same_v<CDataType, bhalf_t>)
{
if constexpr(is_same_v<ALayout, Row> && is_same_v<BLayout, Row> &&
is_same_v<CLayout, Row>)
{
add_device_gemm_xdl_universal_reduce_bf16_i8_bf16_mk_kn_mn_comp_default_instances(
op_ptrs);
add_device_gemm_xdl_universal_reduce_bf16_i8_bf16_mk_kn_mn_comp_kpadding_instances(
op_ptrs);
add_device_gemm_xdl_universal_reduce_bf16_i8_bf16_mk_kn_mn_comp_mnkpadding_instances(
op_ptrs);
add_device_gemm_xdl_universal_reduce_bf16_i8_bf16_mk_kn_mn_mem_v2_default_instances(
op_ptrs);
add_device_gemm_xdl_universal_reduce_bf16_i8_bf16_mk_kn_mn_mem_v2_kpadding_instances(
op_ptrs);
add_device_gemm_xdl_universal_reduce_bf16_i8_bf16_mk_kn_mn_comp_mnpadding_instances(
op_ptrs);
add_device_gemm_xdl_universal_reduce_bf16_i8_bf16_mk_kn_mn_mem_v2_mnkpadding_instances(
op_ptrs);
}
}
#endif
#ifdef CK_ENABLE_BF16
if constexpr(is_same_v<ADataType, bhalf_t> && is_same_v<BDataType, bhalf_t> &&
is_same_v<CDataType, bhalf_t>)
{
if constexpr(is_same_v<ALayout, Row> && is_same_v<BLayout, Row> &&
is_same_v<CLayout, Row>)
{
add_device_gemm_xdl_universal_reduce_bf16_bf16_bf16_mk_kn_mn_comp_default_instances(
op_ptrs);
add_device_gemm_xdl_universal_reduce_bf16_bf16_bf16_mk_kn_mn_comp_kpadding_instances(
op_ptrs);
add_device_gemm_xdl_universal_reduce_bf16_bf16_bf16_mk_kn_mn_comp_mnkpadding_instances(
op_ptrs);
add_device_gemm_xdl_universal_reduce_bf16_bf16_bf16_mk_kn_mn_mem_v2_default_instances(
op_ptrs);
add_device_gemm_xdl_universal_reduce_bf16_bf16_bf16_mk_kn_mn_mem_v2_kpadding_instances(
op_ptrs);
add_device_gemm_xdl_universal_reduce_bf16_bf16_bf16_mk_kn_mn_comp_mnpadding_instances(
op_ptrs);
add_device_gemm_xdl_universal_reduce_bf16_bf16_bf16_mk_kn_mn_mem_v2_mnkpadding_instances(
op_ptrs);
}
}
#endif
return op_ptrs;
}
};
} // namespace instance
} // namespace device
} // namespace tensor_operation
} // namespace ck
library/include/ck/library/tensor_operation_instance/gpu/grouped_conv_bwd_weight/device_grouped_conv_bwd_weight_two_stage_xdl_instance.hpp
View file @ 3d61f89a
......@@ -40,10 +40,10 @@ template <ck::index_t NDimSpatial,
BlockGemmPipelineVersion PipelineVersion>
using device_grouped_conv_bwd_weight_two_stage_xdl_c_shuffle_f16_instances = std::tuple<
// clang-format off
//#########################################| Num| InLayout| WeiLayout| OutLayout| InData| WeiData| OutData| AccData| In| Wei| Out| ConvBackward| Block| MPer| NPer| K0Per| K1| MPer| NPer| MXdl| NXdl| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockLds| BBlockTransfer| BBlockTransfer| BBlockTransfer| BlockTransfer| BBlockTransfer| BBlockTransfer| BBlockLds| CShuffle| CShuffle| CBlockTransfer| CBlockTransfer| BlockGemm| BlockGemm| NumBatch|
//#########################################| Dim| | | | Type| Type| Type| Type| Elementwise| Elementwise| Elementwise| Weight| Size| Block| Block| Block| | XDL| XDL| Per| Per| ThreadCluster| ThreadCluster| SrcAccessOrder| SrcVectorDim| SrcScalar| DstScalar| AddExtraM| ThreadCluster| ThreadCluster| SrcAccessOrder| SrcVectorDim| SrcScalar| DstScalar| AddExtraN| MXdlPerWave| NXdlPerWave| ClusterLengths| ScalarPerVector| Pipeline| Pipeline| ToMerge|
//#########################################| Spatial| | | | | | | | Operation| Operation| Operation| Specialization| | | | | | | | Wave| Wave| Lengths_K0_M_K1| ArrangeOrder| | | PerVector| PerVector_K1| | Lengths_K0_N_K1| ArrangeOrder| | | PerVector| PerVector_K1| | PerShuffle| PerShuffle| MBlock_MPerBlock| NWaveNPerXdl| Scheduler| Version| |
//#########################################| | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | NBlock_NPerBlock| | | | |
//#########################################| Num| InLayout| WeiLayout| OutLayout| InData| WeiData| OutData| AccData| In| Wei| Out| ConvBackward| Block| MPer| NPer| K0Per| K1| MPer| NPer| MXdl| NXdl| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockLds| BBlockTransfer| BBlockTransfer| BBlockTransfer| BlockTransfer| BBlockTransfer| BBlockTransfer| BBlockLds| CShuffle| CShuffle| CBlockTransfer| CBlockTransfer| BlockGemm| BlockGemm| NumGroups|
//#########################################| Dim| | | | Type| Type| Type| Type| Elementwise| Elementwise| Elementwise| Weight| Size| Block| Block| Block| | XDL| XDL| Per| Per| ThreadCluster| ThreadCluster| SrcAccessOrder| SrcVectorDim| SrcScalar| DstScalar| AddExtraM| ThreadCluster| ThreadCluster| SrcAccessOrder| SrcVectorDim| SrcScalar| DstScalar| AddExtraN| MXdlPerWave| NXdlPerWave| ClusterLengths| ScalarPerVector| Pipeline| Pipeline| ToMerge|
//#########################################| Spatial| | | | | | | | Operation| Operation| Operation| Specialization| | | | | | | | Wave| Wave| Lengths_K0_M_K1| ArrangeOrder| | | PerVector| PerVector_K1| | Lengths_K0_N_K1| ArrangeOrder| | | PerVector| PerVector_K1| | PerShuffle| PerShuffle| MBlock_MPerBlock| NWaveNPerXdl| Scheduler| Version| |
//#########################################| | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | NBlock_NPerBlock| | | | |
DeviceGroupedConvBwdWeightTwoStage_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, F16, F16, F16, F32, PassThrough, PassThrough, PassThrough, ConvSpec, 64, 16, 16, 32, 8, 16, 16, 1, 1, S<4, 8, 1>, S<2, 0, 1>, S<1, 0, 2>, 1, 1, 4, false, S<4, 8, 1>, S<2, 0, 1>, S<1, 0, 2>, 1, 1, 4, false, 1, 1, S<1, 8, 1, 8>, 1, Scheduler, PipelineVersion, 1>,
DeviceGroupedConvBwdWeightTwoStage_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, F16, F16, F16, F32, PassThrough, PassThrough, PassThrough, ConvSpec, 64, 32, 32, 32, 8, 32, 32, 1, 1, S<4, 8, 1>, S<2, 0, 1>, S<1, 0, 2>, 1, 2, 2, false, S<4, 16, 1>, S<2, 0, 1>, S<1, 0, 2>, 1, 2, 2, false, 1, 1, S<1, 8, 1, 8>, 1, Scheduler, PipelineVersion, 2>,
DeviceGroupedConvBwdWeightTwoStage_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, F16, F16, F16, F32, PassThrough, PassThrough, PassThrough, ConvSpec, 64, 32, 64, 32, 8, 32, 32, 1, 2, S<4, 8, 1>, S<2, 0, 1>, S<1, 0, 2>, 1, 4, 4, false, S<4, 16, 1>, S<2, 0, 1>, S<1, 0, 2>, 1, 4, 4, false, 1, 1, S<1, 8, 1, 8>, 1, Scheduler, PipelineVersion, 4>,
......
library/include/ck/library/tensor_operation_instance/gpu/grouped_conv_fwd/device_grouped_conv_fwd_xdl_binary_outelementop_instance.hpp 0 → 100644
View file @ 3d61f89a
// SPDX-License-Identifier: MIT
// Copyright (c) 2024, Advanced Micro Devices, Inc. All rights reserved.
#include "ck/ck.hpp"
#include "ck/tensor_operation/gpu/device/tensor_layout.hpp"
#include "ck/tensor_operation/gpu/device/impl/device_grouped_conv_fwd_multiple_abd_xdl_cshuffle.hpp"
#include "ck/tensor_operation/gpu/device/convolution_forward_specialization.hpp"
#include "ck/tensor_operation/gpu/device/gemm_specialization.hpp"
#include "ck/tensor_operation/gpu/element/element_wise_operation.hpp"
namespace ck {
namespace tensor_operation {
namespace device {
namespace instance {
using F32 = float;
using F8 = ck::f8_t;
template <ck::index_t... Is>
using S = ck::Sequence<Is...>;
using namespace ck::tensor_layout::convolution;
using PassThrough = ck::tensor_operation::element_wise::PassThrough;
static constexpr auto ConvFwdDefault =
ck::tensor_operation::device::ConvolutionForwardSpecialization::Default;
static constexpr auto ConvFwd1x1P0 = ConvolutionForwardSpecialization::Filter1x1Pad0;
static constexpr auto ConvFwd1x1S1P0 = ConvolutionForwardSpecialization::Filter1x1Stride1Pad0;
static constexpr auto ConvFwdOddC =
ck::tensor_operation::device::ConvolutionForwardSpecialization::OddC;
static constexpr auto GemmMNKPadding = GemmSpecialization::MNKPadding;
template <index_t NDimSpatial,
typename ALayout,
typename BLayout,
typename DsLayout,
typename ELayout,
ConvolutionForwardSpecialization ConvSpec,
typename OutElementOp>
using device_grouped_conv_fwd_xdl_binary_outelementop_f8_instances = std::tuple<
// clang-format off
//########################################| NumDim| A| B| Ds| E| AData| BData| AccData| CShuffle| Ds| EData| A| B| CDE| ConvForward| GEMM| NumGemmK| Block| MPer| NPer| KPer| AK1| BK1| MPer| NPer| MXdl| NXdl| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockLds| BBlockTransfer| BBlockTransfer| BBlockTransfer| BlockTransfer| BBlockTransfer| BBlockTransfer| BBlockLds| CShuffle| CShuffle| CBlockTransferClusterLengths| CBlockTransfer| Compute| Compute|
//########################################| Spatial| Layout| Layout| Layout| Layout| Type| Type| Type| DataType| DataType| Type| Elementwise| Elementwise| Elementwise| Specialization| Specialization| Prefetch| Size| Block| Block| Block| | | XDL| XDL| Per| Per| ThreadCluster| ThreadCluster| SrcAccessOrder| SrcVectorDim| SrcScalar| DstScalar| AddExtraM| ThreadCluster| ThreadCluster| SrcAccessOrder| SrcVectorDim| SrcScalar| DstScalar| AddExtraN| MXdlPerWave| NXdlPerWave| _MBlock_MWaveMPerXdl| ScalarPerVector| TypeA| TypeB|
//########################################| | | | | | | | | | | | Operation| Operation| Operation| | | Stage| | | | | | | | | Wave| Wave| Lengths_K0_M_K1| ArrangeOrder| | | PerVector| PerVector_K1| | Lengths_K0_N_K1| ArrangeOrder| | | PerVector| PerVector_K1| | PerShuffle| PerShuffle| _NBlock_NWaveNPerXdl| _NWaveNPerXdl| | |
//########################################| | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | |
#ifdef CK_ENABLE_FP8
// generic instance
DeviceGroupedConvFwdMultipleABD_Xdl_CShuffle<NDimSpatial,ALayout,BLayout, DsLayout,ELayout, F8, F8, F32, F32, Tuple<F32>, F8, PassThrough, PassThrough, OutElementOp, ConvSpec, GemmMNKPadding, 1, 64, 64, 64, 32, 8, 8, 32, 32, 2, 2, S<4, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 1, 8, 1, S<4, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 1, 8, 1, 1, 1, S<1, 16, 1, 4>, 1, F8, F8>,
// instances for small conv.K and conv.C
DeviceGroupedConvFwdMultipleABD_Xdl_CShuffle<NDimSpatial,ALayout,BLayout, DsLayout,ELayout, F8, F8, F32, F32, Tuple<F32>, F8, PassThrough, PassThrough, OutElementOp, ConvSpec, GemmMNKPadding, 1, 64, 64, 32, 32, 8, 8, 32, 32, 2, 1, S<4, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, S<4, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, 1, 1, S<1, 16, 1, 4>, 1, F8, F8>,
DeviceGroupedConvFwdMultipleABD_Xdl_CShuffle<NDimSpatial,ALayout,BLayout, DsLayout,ELayout, F8, F8, F32, F32, Tuple<F32>, F8, PassThrough, PassThrough, OutElementOp, ConvSpec, GemmMNKPadding, 1, 256, 128, 128, 32, 8, 8, 32, 32, 2, 2, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 1, 8, 1, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 1, 8, 1, 1, 1, S<1, 32, 1, 8>, 8, F8, F8>,
DeviceGroupedConvFwdMultipleABD_Xdl_CShuffle<NDimSpatial,ALayout,BLayout, DsLayout,ELayout, F8, F8, F32, F32, Tuple<F32>, F8, PassThrough, PassThrough, OutElementOp, ConvSpec, GemmMNKPadding, 1, 256, 256, 128, 32, 8, 8, 32, 32, 4, 2, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, 1, 1, S<1, 32, 1, 8>, 8, F8, F8>,
DeviceGroupedConvFwdMultipleABD_Xdl_CShuffle<NDimSpatial,ALayout,BLayout, DsLayout,ELayout, F8, F8, F32, F32, Tuple<F32>, F8, PassThrough, PassThrough, OutElementOp, ConvSpec, GemmMNKPadding, 1, 256, 128, 256, 32, 8, 8, 32, 32, 2, 4, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, 1, 1, S<1, 32, 1, 8>, 8, F8, F8>,
DeviceGroupedConvFwdMultipleABD_Xdl_CShuffle<NDimSpatial,ALayout,BLayout, DsLayout,ELayout, F8, F8, F32, F32, Tuple<F32>, F8, PassThrough, PassThrough, OutElementOp, ConvSpec, GemmMNKPadding, 1, 128, 128, 128, 32, 8, 8, 32, 32, 4, 2, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, 1, 1, S<1, 16, 1, 8>, 8, F8, F8>,
DeviceGroupedConvFwdMultipleABD_Xdl_CShuffle<NDimSpatial,ALayout,BLayout, DsLayout,ELayout, F8, F8, F32, F32, Tuple<F32>, F8, PassThrough, PassThrough, OutElementOp, ConvSpec, GemmMNKPadding, 1, 256, 128, 128, 32, 8, 8, 32, 32, 2, 2, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, 1, 1, S<1, 32, 1, 8>, 8, F8, F8>,
DeviceGroupedConvFwdMultipleABD_Xdl_CShuffle<NDimSpatial,ALayout,BLayout, DsLayout,ELayout, F8, F8, F32, F32, Tuple<F32>, F8, PassThrough, PassThrough, OutElementOp, ConvSpec, GemmMNKPadding, 1, 128, 128, 64, 32, 8, 8, 32, 32, 2, 2, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, 1, 1, S<1, 32, 1, 4>, 8, F8, F8>,
DeviceGroupedConvFwdMultipleABD_Xdl_CShuffle<NDimSpatial,ALayout,BLayout, DsLayout,ELayout, F8, F8, F32, F32, Tuple<F32>, F8, PassThrough, PassThrough, OutElementOp, ConvSpec, GemmMNKPadding, 1, 128, 64, 128, 32, 8, 8, 32, 32, 2, 2, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, 1, 1, S<1, 16, 1, 8>, 8, F8, F8>,
DeviceGroupedConvFwdMultipleABD_Xdl_CShuffle<NDimSpatial,ALayout,BLayout, DsLayout,ELayout, F8, F8, F32, F32, Tuple<F32>, F8, PassThrough, PassThrough, OutElementOp, ConvSpec, GemmMNKPadding, 1, 64, 64, 64, 32, 8, 8, 32, 32, 2, 2, S<4, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, S<4, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, 1, 1, S<1, 16, 1, 4>, 8, F8, F8>,
DeviceGroupedConvFwdMultipleABD_Xdl_CShuffle<NDimSpatial,ALayout,BLayout, DsLayout,ELayout, F8, F8, F32, F32, Tuple<F32>, F8, PassThrough, PassThrough, OutElementOp, ConvSpec, GemmMNKPadding, 1, 256, 128, 64, 32, 8, 8, 32, 32, 2, 1, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, 1, 1, S<1, 32, 1, 8>, 8, F8, F8>,
DeviceGroupedConvFwdMultipleABD_Xdl_CShuffle<NDimSpatial,ALayout,BLayout, DsLayout,ELayout, F8, F8, F32, F32, Tuple<F32>, F8, PassThrough, PassThrough, OutElementOp, ConvSpec, GemmMNKPadding, 1, 256, 64, 128, 32, 8, 8, 32, 32, 1, 2, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, 1, 1, S<1, 32, 1, 8>, 8, F8, F8>,
DeviceGroupedConvFwdMultipleABD_Xdl_CShuffle<NDimSpatial,ALayout,BLayout, DsLayout,ELayout, F8, F8, F32, F32, Tuple<F32>, F8, PassThrough, PassThrough, OutElementOp, ConvSpec, GemmMNKPadding, 1, 128, 128, 32, 32, 8, 8, 32, 32, 2, 1, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, 1, 1, S<1, 32, 1, 4>, 8, F8, F8>,
DeviceGroupedConvFwdMultipleABD_Xdl_CShuffle<NDimSpatial,ALayout,BLayout, DsLayout,ELayout, F8, F8, F32, F32, Tuple<F32>, F8, PassThrough, PassThrough, OutElementOp, ConvSpec, GemmMNKPadding, 1, 128, 32, 128, 32, 8, 8, 32, 32, 1, 2, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, 1, 1, S<1, 16, 1, 8>, 8, F8, F8>,
DeviceGroupedConvFwdMultipleABD_Xdl_CShuffle<NDimSpatial,ALayout,BLayout, DsLayout,ELayout, F8, F8, F32, F32, Tuple<F32>, F8, PassThrough, PassThrough, OutElementOp, ConvSpec, GemmMNKPadding, 1, 64, 64, 32, 32, 8, 8, 32, 32, 2, 1, S<4, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, S<4, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, 1, 1, S<1, 16, 1, 4>, 8, F8, F8>,
DeviceGroupedConvFwdMultipleABD_Xdl_CShuffle<NDimSpatial,ALayout,BLayout, DsLayout,ELayout, F8, F8, F32, F32, Tuple<F32>, F8, PassThrough, PassThrough, OutElementOp, ConvSpec, GemmMNKPadding, 1, 64, 32, 64, 32, 8, 8, 32, 32, 1, 2, S<4, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, S<4, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, 1, 1, S<1, 16, 1, 4>, 8, F8, F8>
#endif
// clang-format on
>;
} // namespace instance
} // namespace device
} // namespace tensor_operation
} // namespace ck
library/include/ck/library/tensor_operation_instance/gpu/grouped_conv_fwd/device_grouped_conv_fwd_xdl_large_tensor_instance.hpp 0 → 100644
View file @ 3d61f89a
// SPDX-License-Identifier: MIT
// Copyright (c) 2024, Advanced Micro Devices, Inc. All rights reserved.
#include "ck/ck.hpp"
#include "ck/tensor_operation/gpu/device/tensor_layout.hpp"
#include "ck/tensor_operation/gpu/device/impl/device_grouped_conv_fwd_multiple_d_xdl_large_tensor_cshuffle.hpp"
#include "ck/tensor_operation/gpu/device/convolution_forward_specialization.hpp"
#include "ck/tensor_operation/gpu/device/gemm_specialization.hpp"
#include "ck/tensor_operation/gpu/element/element_wise_operation.hpp"
namespace ck {
namespace tensor_operation {
namespace device {
namespace instance {
using BF16 = ck::bhalf_t;
using F16 = ck::half_t;
using F32 = float;
template <ck::index_t... Is>
using S = ck::Sequence<Is...>;
using Empty_Tuple = ck::Tuple<>;
using namespace ck::tensor_layout::convolution;
using PassThrough = ck::tensor_operation::element_wise::PassThrough;
static constexpr auto ConvFwdDefault =
ck::tensor_operation::device::ConvolutionForwardSpecialization::Default;
static constexpr auto GemmMNKPadding = GemmSpecialization::MNKPadding;
template <index_t NDimSpatial,
typename ALayout,
typename BLayout,
typename DsLayout,
typename ELayout,
ConvolutionForwardSpecialization ConvSpec>
using device_grouped_conv_fwd_xdl_large_tensor_bf16_instances = std::tuple<
// clang-format off
//########################################| NumDim| A| B| Ds| E| AData| BData| AccData| CShuffle| Ds| EData| A| B| CDE| ConvForward| GEMM| NumGemmK| Block| MPer| NPer| KPer| AK1| BK1| MPer| NPer| MXdl| NXdl| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockLds| BBlockTransfer| BBlockTransfer| BBlockTransfer| BlockTransfer| BBlockTransfer| BBlockTransfer| BBlockLds| CShuffle| CShuffle| CBlockTransferClusterLengths| CBlockTransfer|
//########################################| Spatial| Layout| Layout| Layout| Layout| Type| Type| Type| DataType| DataType| Type| Elementwise| Elementwise| Elementwise| Specialization| Specialization| Prefetch| Size| Block| Block| Block| | | XDL| XDL| Per| Per| ThreadCluster| ThreadCluster| SrcAccessOrder| SrcVectorDim| SrcScalar| DstScalar| AddExtraM| ThreadCluster| ThreadCluster| SrcAccessOrder| SrcVectorDim| SrcScalar| DstScalar| AddExtraN| MXdlPerWave| NXdlPerWave| _MBlock_MWaveMPerXdl| ScalarPerVector|
//########################################| | | | | | | | | | | | Operation| Operation| Operation| | | Stage| | | | | | | | | Wave| Wave| Lengths_K0_M_K1| ArrangeOrder| | | PerVector| PerVector_K1| | Lengths_K0_N_K1| ArrangeOrder| | | PerVector| PerVector_K1| | PerShuffle| PerShuffle| _NBlock_NWaveNPerXdl| _NWaveNPerXdl|
//########################################| | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | |
// generic instance
DeviceGroupedConvFwdMultipleD_Xdl_CShuffle_Large_Tensor<NDimSpatial,ALayout,BLayout, DsLayout,ELayout, BF16, BF16, F32, BF16, DsLayout, BF16, PassThrough, PassThrough, PassThrough, ConvSpec, GemmMNKPadding, 1, 64, 64, 64, 32, 8, 8, 32, 32, 2, 2, S<4, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 1, 8, 1, S<4, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 1, 8, 1, 1, 1, S<1, 16, 1, 4>, 1>,
DeviceGroupedConvFwdMultipleD_Xdl_CShuffle_Large_Tensor<NDimSpatial,ALayout,BLayout, DsLayout,ELayout, BF16, BF16, F32, BF16, DsLayout, BF16, PassThrough, PassThrough, PassThrough, ConvSpec, GemmMNKPadding, 1, 256, 256, 128, 32, 8, 8, 32, 32, 4, 2, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, 1, 1, S<1, 32, 1, 8>, 8>
// clang-format on
>;
template <index_t NDimSpatial,
typename ALayout,
typename BLayout,
typename DsLayout,
typename ELayout,
ConvolutionForwardSpecialization ConvSpec>
using device_grouped_conv_fwd_xdl_large_tensor_f16_instances = std::tuple<
// clang-format off
//########################################| NumDim| A| B| Ds| E| AData| BData| AccData| CShuffle| Ds| EData| A| B| CDE| ConvForward| GEMM| NumGemmK| Block| MPer| NPer| KPer| AK1| BK1| MPer| NPer| MXdl| NXdl| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockLds| BBlockTransfer| BBlockTransfer| BBlockTransfer| BlockTransfer| BBlockTransfer| BBlockTransfer| BBlockLds| CShuffle| CShuffle| CBlockTransferClusterLengths| CBlockTransfer|
//########################################| Spatial| Layout| Layout| Layout| Layout| Type| Type| Type| DataType| DataType| Type| Elementwise| Elementwise| Elementwise| Specialization| Specialization| Prefetch| Size| Block| Block| Block| | | XDL| XDL| Per| Per| ThreadCluster| ThreadCluster| SrcAccessOrder| SrcVectorDim| SrcScalar| DstScalar| AddExtraM| ThreadCluster| ThreadCluster| SrcAccessOrder| SrcVectorDim| SrcScalar| DstScalar| AddExtraN| MXdlPerWave| NXdlPerWave| _MBlock_MWaveMPerXdl| ScalarPerVector|
//########################################| | | | | | | | | | | | Operation| Operation| Operation| | | Stage| | | | | | | | | Wave| Wave| Lengths_K0_M_K1| ArrangeOrder| | | PerVector| PerVector_K1| | Lengths_K0_N_K1| ArrangeOrder| | | PerVector| PerVector_K1| | PerShuffle| PerShuffle| _NBlock_NWaveNPerXdl| _NWaveNPerXdl|
//########################################| | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | |
// generic instance
DeviceGroupedConvFwdMultipleD_Xdl_CShuffle_Large_Tensor<NDimSpatial,ALayout,BLayout, DsLayout,ELayout, F16, F16, F32, F16, DsLayout, F16, PassThrough, PassThrough, PassThrough, ConvSpec, GemmMNKPadding, 1, 64, 64, 64, 32, 8, 8, 32, 32, 2, 2, S<4, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 1, 8, 1, S<4, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 1, 8, 1, 1, 1, S<1, 16, 1, 4>, 1>,
DeviceGroupedConvFwdMultipleD_Xdl_CShuffle_Large_Tensor<NDimSpatial,ALayout,BLayout, DsLayout,ELayout, F16, F16, F32, F16, DsLayout, F16, PassThrough, PassThrough, PassThrough, ConvSpec, GemmMNKPadding, 1, 256, 256, 128, 32, 8, 8, 32, 32, 4, 2, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, 1, 1, S<1, 32, 1, 8>, 8>
// clang-format on
>;
template <index_t NDimSpatial,
typename ALayout,
typename BLayout,
typename DsLayout,
typename ELayout,
ConvolutionForwardSpecialization ConvSpec>
using device_grouped_conv_fwd_xdl_large_tensor_f32_instances = std::tuple<
// clang-format off
//########################################| NumDim| A| B| Ds| E| AData| BData| AccData| CShuffle| Ds| EData| A| B| CDE| ConvForward| GEMM| NumGemmK| Block| MPer| NPer| KPer| AK1| BK1| MPer| NPer| MXdl| NXdl| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockLds| BBlockTransfer| BBlockTransfer| BBlockTransfer| BlockTransfer| BBlockTransfer| BBlockTransfer| BBlockLds| CShuffle| CShuffle| CBlockTransferClusterLengths| CBlockTransfer|
//########################################| Spatial| Layout| Layout| Layout| Layout| Type| Type| Type| DataType| DataType| Type| Elementwise| Elementwise| Elementwise| Specialization| Specialization| Prefetch| Size| Block| Block| Block| | | XDL| XDL| Per| Per| ThreadCluster| ThreadCluster| SrcAccessOrder| SrcVectorDim| SrcScalar| DstScalar| AddExtraM| ThreadCluster| ThreadCluster| SrcAccessOrder| SrcVectorDim| SrcScalar| DstScalar| AddExtraN| MXdlPerWave| NXdlPerWave| _MBlock_MWaveMPerXdl| ScalarPerVector|
//########################################| | | | | | | | | | | | Operation| Operation| Operation| | | Stage| | | | | | | | | Wave| Wave| Lengths_K0_M_K1| ArrangeOrder| | | PerVector| PerVector_K1| | Lengths_K0_N_K1| ArrangeOrder| | | PerVector| PerVector_K1| | PerShuffle| PerShuffle| _NBlock_NWaveNPerXdl| _NWaveNPerXdl|
//########################################| | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | |
// generic instance
DeviceGroupedConvFwdMultipleD_Xdl_CShuffle_Large_Tensor<NDimSpatial,ALayout,BLayout, DsLayout,ELayout, F32, F32, F32, F32, DsLayout, F32, PassThrough, PassThrough, PassThrough, ConvSpec, GemmMNKPadding, 1, 64, 64, 64, 16, 4, 4, 32, 32, 2, 2, S<4, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 1, 4, 1, S<4, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 1, 4, 1, 1, 1, S<1, 8, 1, 8>, 1>,
DeviceGroupedConvFwdMultipleD_Xdl_CShuffle_Large_Tensor<NDimSpatial,ALayout,BLayout, DsLayout,ELayout, F32, F32, F32, F32, DsLayout, F32, PassThrough, PassThrough, PassThrough, ConvSpec, GemmMNKPadding, 1, 256, 256, 128, 16, 4, 4, 32, 32, 4, 2, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 4, 4, 1, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 4, 4, 1, 1, 1, S<1, 16, 1, 16>, 4>
// clang-format on
>;
} // namespace instance
} // namespace device
} // namespace tensor_operation
} // namespace ck
library/include/ck/library/tensor_operation_instance/gpu/grouped_conv_fwd/device_grouped_conv_fwd_xdl_outelementop_instance.hpp
View file @ 3d61f89a
......@@ -147,6 +147,80 @@ using device_grouped_conv_fwd_xdl_outelementop_f8_bf8_instances = std::tuple<
// clang-format on
>;
template <index_t NDimSpatial,
typename ALayout,
typename BLayout,
typename DsLayout,
typename ELayout,
ConvolutionForwardSpecialization ConvSpec,
typename OutElementOp>
using device_grouped_conv_fwd_xdl_outelementop_bf8_f8_instances = std::tuple<
// clang-format off
//########################################| NumDim| A| B| Ds| E| AData| BData| AccData| CShuffle| Ds| EData| A| B| CDE| ConvForward| GEMM| NumGemmK| Block| MPer| NPer| KPer| AK1| BK1| MPer| NPer| MXdl| NXdl| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockLds| BBlockTransfer| BBlockTransfer| BBlockTransfer| BlockTransfer| BBlockTransfer| BBlockTransfer| BBlockLds| CShuffle| CShuffle| CBlockTransferClusterLengths| CBlockTransfer| Compute| Compute|
//########################################| Spatial| Layout| Layout| Layout| Layout| Type| Type| Type| DataType| DataType| Type| Elementwise| Elementwise| Elementwise| Specialization| Specialization| Prefetch| Size| Block| Block| Block| | | XDL| XDL| Per| Per| ThreadCluster| ThreadCluster| SrcAccessOrder| SrcVectorDim| SrcScalar| DstScalar| AddExtraM| ThreadCluster| ThreadCluster| SrcAccessOrder| SrcVectorDim| SrcScalar| DstScalar| AddExtraN| MXdlPerWave| NXdlPerWave| _MBlock_MWaveMPerXdl| ScalarPerVector| TypeA| TypeB|
//########################################| | | | | | | | | | | | Operation| Operation| Operation| | | Stage| | | | | | | | | Wave| Wave| Lengths_K0_M_K1| ArrangeOrder| | | PerVector| PerVector_K1| | Lengths_K0_N_K1| ArrangeOrder| | | PerVector| PerVector_K1| | PerShuffle| PerShuffle| _NBlock_NWaveNPerXdl| _NWaveNPerXdl| | |
//########################################| | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | |
#if defined(CK_ENABLE_FP8) && defined(CK_ENABLE_BF8)
// generic instance
DeviceGroupedConvFwdMultipleABD_Xdl_CShuffle<NDimSpatial,ALayout,BLayout, DsLayout,ELayout, BF8, F8, F32, F32, Tuple<>, F8, PassThrough, PassThrough, OutElementOp, ConvSpec, GemmMNKPadding, 1, 64, 64, 64, 32, 8, 8, 32, 32, 2, 2, S<4, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 1, 8, 1, S<4, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 1, 8, 1, 1, 1, S<1, 16, 1, 4>, 1, BF8, F8>,
// instances for small conv.K and conv.C
DeviceGroupedConvFwdMultipleABD_Xdl_CShuffle<NDimSpatial,ALayout,BLayout, DsLayout,ELayout, BF8, F8, F32, F32, Tuple<>, F8, PassThrough, PassThrough, OutElementOp, ConvSpec, GemmMNKPadding, 1, 64, 64, 32, 32, 8, 8, 32, 32, 2, 1, S<4, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, S<4, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, 1, 1, S<1, 16, 1, 4>, 1, BF8, F8>,
DeviceGroupedConvFwdMultipleABD_Xdl_CShuffle<NDimSpatial,ALayout,BLayout, DsLayout,ELayout, BF8, F8, F32, F32, Tuple<>, F8, PassThrough, PassThrough, OutElementOp, ConvSpec, GemmMNKPadding, 1, 256, 128, 128, 32, 8, 8, 32, 32, 2, 2, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 1, 8, 1, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 1, 8, 1, 1, 1, S<1, 32, 1, 8>, 8, BF8, F8>,
DeviceGroupedConvFwdMultipleABD_Xdl_CShuffle<NDimSpatial,ALayout,BLayout, DsLayout,ELayout, BF8, F8, F32, F32, Tuple<>, F8, PassThrough, PassThrough, OutElementOp, ConvSpec, GemmMNKPadding, 1, 256, 256, 128, 32, 8, 8, 32, 32, 4, 2, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, 1, 1, S<1, 32, 1, 8>, 8, BF8, F8>,
DeviceGroupedConvFwdMultipleABD_Xdl_CShuffle<NDimSpatial,ALayout,BLayout, DsLayout,ELayout, BF8, F8, F32, F32, Tuple<>, F8, PassThrough, PassThrough, OutElementOp, ConvSpec, GemmMNKPadding, 1, 256, 128, 256, 32, 8, 8, 32, 32, 2, 4, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, 1, 1, S<1, 32, 1, 8>, 8, BF8, F8>,
DeviceGroupedConvFwdMultipleABD_Xdl_CShuffle<NDimSpatial,ALayout,BLayout, DsLayout,ELayout, BF8, F8, F32, F32, Tuple<>, F8, PassThrough, PassThrough, OutElementOp, ConvSpec, GemmMNKPadding, 1, 128, 128, 128, 32, 8, 8, 32, 32, 4, 2, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, 1, 1, S<1, 16, 1, 8>, 8, BF8, F8>,
DeviceGroupedConvFwdMultipleABD_Xdl_CShuffle<NDimSpatial,ALayout,BLayout, DsLayout,ELayout, BF8, F8, F32, F32, Tuple<>, F8, PassThrough, PassThrough, OutElementOp, ConvSpec, GemmMNKPadding, 1, 256, 128, 128, 32, 8, 8, 32, 32, 2, 2, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, 1, 1, S<1, 32, 1, 8>, 8, BF8, F8>,
DeviceGroupedConvFwdMultipleABD_Xdl_CShuffle<NDimSpatial,ALayout,BLayout, DsLayout,ELayout, BF8, F8, F32, F32, Tuple<>, F8, PassThrough, PassThrough, OutElementOp, ConvSpec, GemmMNKPadding, 1, 128, 128, 64, 32, 8, 8, 32, 32, 2, 2, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, 1, 1, S<1, 32, 1, 4>, 8, BF8, F8>,
DeviceGroupedConvFwdMultipleABD_Xdl_CShuffle<NDimSpatial,ALayout,BLayout, DsLayout,ELayout, BF8, F8, F32, F32, Tuple<>, F8, PassThrough, PassThrough, OutElementOp, ConvSpec, GemmMNKPadding, 1, 128, 64, 128, 32, 8, 8, 32, 32, 2, 2, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, 1, 1, S<1, 16, 1, 8>, 8, BF8, F8>,
DeviceGroupedConvFwdMultipleABD_Xdl_CShuffle<NDimSpatial,ALayout,BLayout, DsLayout,ELayout, BF8, F8, F32, F32, Tuple<>, F8, PassThrough, PassThrough, OutElementOp, ConvSpec, GemmMNKPadding, 1, 64, 64, 64, 32, 8, 8, 32, 32, 2, 2, S<4, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, S<4, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, 1, 1, S<1, 16, 1, 4>, 8, BF8, F8>,
DeviceGroupedConvFwdMultipleABD_Xdl_CShuffle<NDimSpatial,ALayout,BLayout, DsLayout,ELayout, BF8, F8, F32, F32, Tuple<>, F8, PassThrough, PassThrough, OutElementOp, ConvSpec, GemmMNKPadding, 1, 256, 128, 64, 32, 8, 8, 32, 32, 2, 1, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, 1, 1, S<1, 32, 1, 8>, 8, BF8, F8>,
DeviceGroupedConvFwdMultipleABD_Xdl_CShuffle<NDimSpatial,ALayout,BLayout, DsLayout,ELayout, BF8, F8, F32, F32, Tuple<>, F8, PassThrough, PassThrough, OutElementOp, ConvSpec, GemmMNKPadding, 1, 256, 64, 128, 32, 8, 8, 32, 32, 1, 2, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, 1, 1, S<1, 32, 1, 8>, 8, BF8, F8>,
DeviceGroupedConvFwdMultipleABD_Xdl_CShuffle<NDimSpatial,ALayout,BLayout, DsLayout,ELayout, BF8, F8, F32, F32, Tuple<>, F8, PassThrough, PassThrough, OutElementOp, ConvSpec, GemmMNKPadding, 1, 128, 128, 32, 32, 8, 8, 32, 32, 2, 1, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, 1, 1, S<1, 32, 1, 4>, 8, BF8, F8>,
DeviceGroupedConvFwdMultipleABD_Xdl_CShuffle<NDimSpatial,ALayout,BLayout, DsLayout,ELayout, BF8, F8, F32, F32, Tuple<>, F8, PassThrough, PassThrough, OutElementOp, ConvSpec, GemmMNKPadding, 1, 128, 32, 128, 32, 8, 8, 32, 32, 1, 2, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, 1, 1, S<1, 16, 1, 8>, 8, BF8, F8>,
DeviceGroupedConvFwdMultipleABD_Xdl_CShuffle<NDimSpatial,ALayout,BLayout, DsLayout,ELayout, BF8, F8, F32, F32, Tuple<>, F8, PassThrough, PassThrough, OutElementOp, ConvSpec, GemmMNKPadding, 1, 64, 64, 32, 32, 8, 8, 32, 32, 2, 1, S<4, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, S<4, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, 1, 1, S<1, 16, 1, 4>, 8, BF8, F8>,
DeviceGroupedConvFwdMultipleABD_Xdl_CShuffle<NDimSpatial,ALayout,BLayout, DsLayout,ELayout, BF8, F8, F32, F32, Tuple<>, F8, PassThrough, PassThrough, OutElementOp, ConvSpec, GemmMNKPadding, 1, 64, 32, 64, 32, 8, 8, 32, 32, 1, 2, S<4, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, S<4, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, 1, 1, S<1, 16, 1, 4>, 8, BF8, F8>
#endif
// clang-format on
>;
template <index_t NDimSpatial,
typename ALayout,
typename BLayout,
typename DsLayout,
typename ELayout,
ConvolutionForwardSpecialization ConvSpec,
typename OutElementOp>
using device_grouped_conv_fwd_xdl_outelementop_f8_f8_f32_instances = std::tuple<
// clang-format off
//########################################| NumDim| A| B| Ds| E| AData| BData| AccData| CShuffle| Ds| EData| A| B| CDE| ConvForward| GEMM| NumGemmK| Block| MPer| NPer| KPer| AK1| BK1| MPer| NPer| MXdl| NXdl| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockLds| BBlockTransfer| BBlockTransfer| BBlockTransfer| BlockTransfer| BBlockTransfer| BBlockTransfer| BBlockLds| CShuffle| CShuffle| CBlockTransferClusterLengths| CBlockTransfer| Compute| Compute|
//########################################| Spatial| Layout| Layout| Layout| Layout| Type| Type| Type| DataType| DataType| Type| Elementwise| Elementwise| Elementwise| Specialization| Specialization| Prefetch| Size| Block| Block| Block| | | XDL| XDL| Per| Per| ThreadCluster| ThreadCluster| SrcAccessOrder| SrcVectorDim| SrcScalar| DstScalar| AddExtraM| ThreadCluster| ThreadCluster| SrcAccessOrder| SrcVectorDim| SrcScalar| DstScalar| AddExtraN| MXdlPerWave| NXdlPerWave| _MBlock_MWaveMPerXdl| ScalarPerVector| TypeA| TypeB|
//########################################| | | | | | | | | | | | Operation| Operation| Operation| | | Stage| | | | | | | | | Wave| Wave| Lengths_K0_M_K1| ArrangeOrder| | | PerVector| PerVector_K1| | Lengths_K0_N_K1| ArrangeOrder| | | PerVector| PerVector_K1| | PerShuffle| PerShuffle| _NBlock_NWaveNPerXdl| _NWaveNPerXdl| | |
//########################################| | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | |
#ifdef CK_ENABLE_FP8
// generic instance
DeviceGroupedConvFwdMultipleABD_Xdl_CShuffle<NDimSpatial,ALayout,BLayout, DsLayout,ELayout, F8, F8, F32, F32, Tuple<>, F32, PassThrough, PassThrough, OutElementOp, ConvSpec, GemmMNKPadding, 1, 64, 64, 64, 32, 8, 8, 32, 32, 2, 2, S<4, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 1, 8, 1, S<4, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 1, 8, 1, 1, 1, S<1, 16, 1, 4>, 1, F8, F8>,
// instances for small conv.K and conv.C
DeviceGroupedConvFwdMultipleABD_Xdl_CShuffle<NDimSpatial,ALayout,BLayout, DsLayout,ELayout, F8, F8, F32, F32, Tuple<>, F32, PassThrough, PassThrough, OutElementOp, ConvSpec, GemmMNKPadding, 1, 64, 64, 32, 32, 8, 8, 32, 32, 2, 1, S<4, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, S<4, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, 1, 1, S<1, 16, 1, 4>, 1, F8, F8>,
DeviceGroupedConvFwdMultipleABD_Xdl_CShuffle<NDimSpatial,ALayout,BLayout, DsLayout,ELayout, F8, F8, F32, F32, Tuple<>, F32, PassThrough, PassThrough, OutElementOp, ConvSpec, GemmMNKPadding, 1, 256, 128, 128, 32, 8, 8, 32, 32, 2, 2, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 1, 8, 1, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 1, 8, 1, 1, 1, S<1, 32, 1, 8>, 8, F8, F8>,
DeviceGroupedConvFwdMultipleABD_Xdl_CShuffle<NDimSpatial,ALayout,BLayout, DsLayout,ELayout, F8, F8, F32, F32, Tuple<>, F32, PassThrough, PassThrough, OutElementOp, ConvSpec, GemmMNKPadding, 1, 256, 256, 128, 32, 8, 8, 32, 32, 4, 2, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, 1, 1, S<1, 32, 1, 8>, 8, F8, F8>,
DeviceGroupedConvFwdMultipleABD_Xdl_CShuffle<NDimSpatial,ALayout,BLayout, DsLayout,ELayout, F8, F8, F32, F32, Tuple<>, F32, PassThrough, PassThrough, OutElementOp, ConvSpec, GemmMNKPadding, 1, 256, 128, 256, 32, 8, 8, 32, 32, 2, 4, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, 1, 1, S<1, 32, 1, 8>, 8, F8, F8>,
DeviceGroupedConvFwdMultipleABD_Xdl_CShuffle<NDimSpatial,ALayout,BLayout, DsLayout,ELayout, F8, F8, F32, F32, Tuple<>, F32, PassThrough, PassThrough, OutElementOp, ConvSpec, GemmMNKPadding, 1, 128, 128, 128, 32, 8, 8, 32, 32, 4, 2, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, 1, 1, S<1, 16, 1, 8>, 8, F8, F8>,
DeviceGroupedConvFwdMultipleABD_Xdl_CShuffle<NDimSpatial,ALayout,BLayout, DsLayout,ELayout, F8, F8, F32, F32, Tuple<>, F32, PassThrough, PassThrough, OutElementOp, ConvSpec, GemmMNKPadding, 1, 256, 128, 128, 32, 8, 8, 32, 32, 2, 2, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, 1, 1, S<1, 32, 1, 8>, 8, F8, F8>,
DeviceGroupedConvFwdMultipleABD_Xdl_CShuffle<NDimSpatial,ALayout,BLayout, DsLayout,ELayout, F8, F8, F32, F32, Tuple<>, F32, PassThrough, PassThrough, OutElementOp, ConvSpec, GemmMNKPadding, 1, 128, 128, 64, 32, 8, 8, 32, 32, 2, 2, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, 1, 1, S<1, 32, 1, 4>, 8, F8, F8>,
DeviceGroupedConvFwdMultipleABD_Xdl_CShuffle<NDimSpatial,ALayout,BLayout, DsLayout,ELayout, F8, F8, F32, F32, Tuple<>, F32, PassThrough, PassThrough, OutElementOp, ConvSpec, GemmMNKPadding, 1, 128, 64, 128, 32, 8, 8, 32, 32, 2, 2, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, 1, 1, S<1, 16, 1, 8>, 8, F8, F8>,
DeviceGroupedConvFwdMultipleABD_Xdl_CShuffle<NDimSpatial,ALayout,BLayout, DsLayout,ELayout, F8, F8, F32, F32, Tuple<>, F32, PassThrough, PassThrough, OutElementOp, ConvSpec, GemmMNKPadding, 1, 64, 64, 64, 32, 8, 8, 32, 32, 2, 2, S<4, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, S<4, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, 1, 1, S<1, 16, 1, 4>, 8, F8, F8>,
DeviceGroupedConvFwdMultipleABD_Xdl_CShuffle<NDimSpatial,ALayout,BLayout, DsLayout,ELayout, F8, F8, F32, F32, Tuple<>, F32, PassThrough, PassThrough, OutElementOp, ConvSpec, GemmMNKPadding, 1, 256, 128, 64, 32, 8, 8, 32, 32, 2, 1, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, 1, 1, S<1, 32, 1, 8>, 8, F8, F8>,
DeviceGroupedConvFwdMultipleABD_Xdl_CShuffle<NDimSpatial,ALayout,BLayout, DsLayout,ELayout, F8, F8, F32, F32, Tuple<>, F32, PassThrough, PassThrough, OutElementOp, ConvSpec, GemmMNKPadding, 1, 256, 64, 128, 32, 8, 8, 32, 32, 1, 2, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, 1, 1, S<1, 32, 1, 8>, 8, F8, F8>,
DeviceGroupedConvFwdMultipleABD_Xdl_CShuffle<NDimSpatial,ALayout,BLayout, DsLayout,ELayout, F8, F8, F32, F32, Tuple<>, F32, PassThrough, PassThrough, OutElementOp, ConvSpec, GemmMNKPadding, 1, 128, 128, 32, 32, 8, 8, 32, 32, 2, 1, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, 1, 1, S<1, 32, 1, 4>, 8, F8, F8>,
DeviceGroupedConvFwdMultipleABD_Xdl_CShuffle<NDimSpatial,ALayout,BLayout, DsLayout,ELayout, F8, F8, F32, F32, Tuple<>, F32, PassThrough, PassThrough, OutElementOp, ConvSpec, GemmMNKPadding, 1, 128, 32, 128, 32, 8, 8, 32, 32, 1, 2, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, 1, 1, S<1, 16, 1, 8>, 8, F8, F8>,
DeviceGroupedConvFwdMultipleABD_Xdl_CShuffle<NDimSpatial,ALayout,BLayout, DsLayout,ELayout, F8, F8, F32, F32, Tuple<>, F32, PassThrough, PassThrough, OutElementOp, ConvSpec, GemmMNKPadding, 1, 64, 64, 32, 32, 8, 8, 32, 32, 2, 1, S<4, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, S<4, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, 1, 1, S<1, 16, 1, 4>, 8, F8, F8>,
DeviceGroupedConvFwdMultipleABD_Xdl_CShuffle<NDimSpatial,ALayout,BLayout, DsLayout,ELayout, F8, F8, F32, F32, Tuple<>, F32, PassThrough, PassThrough, OutElementOp, ConvSpec, GemmMNKPadding, 1, 64, 32, 64, 32, 8, 8, 32, 32, 1, 2, S<4, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, S<4, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, 1, 1, S<1, 16, 1, 4>, 8, F8, F8>
#endif
// clang-format on
>;
} // namespace instance
} // namespace device
} // namespace tensor_operation
......
library/include/ck/library/tensor_operation_instance/gpu/grouped_convolution_forward.hpp
View file @ 3d61f89a
......@@ -17,6 +17,7 @@
#endif
#ifdef CK_USE_XDL
#include "grouped_convolution_forward_xdl.inc"
#include "grouped_convolution_forward_xdl_large_tensor.inc"
#include "grouped_convolution_forward_comp_xdl.inc"
#include "grouped_convolution_forward_mem_inter_xdl.inc"
#include "grouped_convolution_forward_mem_intra_xdl.inc"
......@@ -199,6 +200,8 @@ struct DeviceOperationInstanceFactory<ck::tensor_operation::device::DeviceGroupe
is_same_v<BComputeType, float>)
{
add_device_grouped_conv2d_fwd_xdl_nhwgc_gkyxc_nhwgk_f32_instances(op_ptrs);
add_device_grouped_conv2d_fwd_xdl_large_tensor_nhwgc_gkyxc_nhwgk_f32_instances(
op_ptrs);
add_device_grouped_conv2d_fwd_xdl_nhwgc_gkyxc_nhwgk_f32_comp_instances(op_ptrs);
add_device_grouped_conv2d_fwd_xdl_nhwgc_gkyxc_nhwgk_f32_mem_intra_instances(
op_ptrs);
......@@ -212,6 +215,8 @@ struct DeviceOperationInstanceFactory<ck::tensor_operation::device::DeviceGroupe
is_same_v<BComputeType, half_t>)
{
add_device_grouped_conv2d_fwd_xdl_nhwgc_gkyxc_nhwgk_f16_instances(op_ptrs);
add_device_grouped_conv2d_fwd_xdl_large_tensor_nhwgc_gkyxc_nhwgk_f16_instances(
op_ptrs);
add_device_grouped_conv2d_fwd_xdl_nhwgc_gkyxc_nhwgk_f16_comp_instances(op_ptrs);
add_device_grouped_conv2d_fwd_xdl_nhwgc_gkyxc_nhwgk_f16_mem_intra_instances(
op_ptrs);
......@@ -227,6 +232,8 @@ struct DeviceOperationInstanceFactory<ck::tensor_operation::device::DeviceGroupe
is_same_v<BComputeType, ck::bhalf_t>)
{
add_device_grouped_conv2d_fwd_xdl_nhwgc_gkyxc_nhwgk_bf16_instances(op_ptrs);
add_device_grouped_conv2d_fwd_xdl_large_tensor_nhwgc_gkyxc_nhwgk_bf16_instances(
op_ptrs);
add_device_grouped_conv2d_fwd_xdl_nhwgc_gkyxc_nhwgk_bf16_comp_instances(op_ptrs);
add_device_grouped_conv2d_fwd_xdl_nhwgc_gkyxc_nhwgk_bf16_mem_intra_instances(
op_ptrs);
......@@ -284,6 +291,8 @@ struct DeviceOperationInstanceFactory<ck::tensor_operation::device::DeviceGroupe
is_same_v<BComputeType, float>)
{
add_device_grouped_conv3d_fwd_xdl_ndhwgc_gkzyxc_ndhwgk_f32_instances(op_ptrs);
add_device_grouped_conv3d_fwd_xdl_large_tensor_ndhwgc_gkzyxc_ndhwgk_f32_instances(
op_ptrs);
add_device_grouped_conv3d_fwd_xdl_ndhwgc_gkzyxc_ndhwgk_f32_comp_instances(op_ptrs);
add_device_grouped_conv3d_fwd_xdl_ndhwgc_gkzyxc_ndhwgk_f32_mem_intra_instances(
op_ptrs);
......@@ -338,6 +347,8 @@ struct DeviceOperationInstanceFactory<ck::tensor_operation::device::DeviceGroupe
is_same_v<BComputeType, half_t>)
{
add_device_grouped_conv3d_fwd_xdl_ndhwgc_gkzyxc_ndhwgk_f16_instances(op_ptrs);
add_device_grouped_conv3d_fwd_xdl_large_tensor_ndhwgc_gkzyxc_ndhwgk_f16_instances(
op_ptrs);
add_device_grouped_conv3d_fwd_xdl_ndhwgc_gkzyxc_ndhwgk_f16_comp_instances(op_ptrs);
add_device_grouped_conv3d_fwd_xdl_ndhwgc_gkzyxc_ndhwgk_f16_mem_intra_instances(
op_ptrs);
......@@ -353,6 +364,8 @@ struct DeviceOperationInstanceFactory<ck::tensor_operation::device::DeviceGroupe
is_same_v<BComputeType, ck::bhalf_t>)
{
add_device_grouped_conv3d_fwd_xdl_ndhwgc_gkzyxc_ndhwgk_bf16_instances(op_ptrs);
add_device_grouped_conv3d_fwd_xdl_large_tensor_ndhwgc_gkzyxc_ndhwgk_bf16_instances(
op_ptrs);
add_device_grouped_conv3d_fwd_xdl_ndhwgc_gkzyxc_ndhwgk_bf16_comp_instances(op_ptrs);
add_device_grouped_conv3d_fwd_xdl_ndhwgc_gkzyxc_ndhwgk_bf16_mem_intra_instances(
op_ptrs);
......
library/include/ck/library/tensor_operation_instance/gpu/grouped_convolution_forward_convscale.hpp
View file @ 3d61f89a
......@@ -70,6 +70,22 @@ void add_device_grouped_conv3d_fwd_xdl_convscale_ndhwgc_gkzyxc_ndhwgk_f8_bf8_ins
ConvScale,
F8,
BF8>>>& instances);
void add_device_grouped_conv3d_fwd_xdl_convscale_ndhwgc_gkzyxc_ndhwgk_bf8_f8_instances(
std::vector<std::unique_ptr<DeviceGroupedConvFwdMultipleABD<3,
NDHWGC,
GKZYXC,
ck::Tuple<>,
NDHWGK,
BF8,
F8,
ck::Tuple<>,
F8,
PassThrough,
PassThrough,
ConvScale,
BF8,
F8>>>& instances);
#endif
template <ck::index_t NumDimSpatial,
......@@ -147,6 +163,14 @@ struct DeviceOperationInstanceFactory<ck::tensor_operation::device::DeviceGroupe
add_device_grouped_conv3d_fwd_xdl_convscale_ndhwgc_gkzyxc_ndhwgk_f8_bf8_instances(
op_ptrs);
}
if constexpr(is_same_v<InDataType, bf8_t> && is_same_v<WeiDataType, f8_t> &&
is_same_v<OutDataType, f8_t> && is_same_v<AComputeType, bf8_t> &&
is_same_v<BComputeType, f8_t>)
{
add_device_grouped_conv3d_fwd_xdl_convscale_ndhwgc_gkzyxc_ndhwgk_bf8_f8_instances(
op_ptrs);
}
#endif
}
return op_ptrs;
......
library/include/ck/library/tensor_operation_instance/gpu/grouped_convolution_forward_convscale_add.hpp 0 → 100644
View file @ 3d61f89a
// 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/device_grouped_conv_fwd_multiple_abd.hpp"
#include "ck/tensor_operation/gpu/element/binary_element_wise_operation.hpp"
#include "ck/library/tensor_operation_instance/device_operation_instance_factory.hpp"
namespace ck {
namespace tensor_operation {
namespace device {
namespace instance {
using PassThrough = ck::tensor_operation::element_wise::PassThrough;
using ConvScaleAdd = ck::tensor_operation::element_wise::ConvScaleAdd;
#ifdef CK_ENABLE_FP8
void add_device_grouped_conv3d_fwd_xdl_convscale_add_ndhwgc_gkzyxc_ndhwgk_f8_instances(
std::vector<std::unique_ptr<DeviceGroupedConvFwdMultipleABD<3,
NDHWGC,
GKZYXC,
ck::Tuple<NDHWGK>,
NDHWGK,
F8,
F8,
ck::Tuple<F32>,
F8,
PassThrough,
PassThrough,
ConvScaleAdd,
F8,
F8>>>& instances);
#endif
template <ck::index_t NumDimSpatial,
typename InLayout,
typename WeiLayout,
typename DLayouts,
typename OutLayout,
typename InDataType,
typename WeiDataType,
typename DDataTypes,
typename OutDataType,
typename AComputeType,
typename BComputeType>
struct DeviceOperationInstanceFactory<
ck::tensor_operation::device::DeviceGroupedConvFwdMultipleABD<NumDimSpatial,
InLayout,
WeiLayout,
DLayouts,
OutLayout,
InDataType,
WeiDataType,
DDataTypes,
OutDataType,
PassThrough,
PassThrough,
ConvScaleAdd,
AComputeType,
BComputeType>>
{
using DeviceOp = DeviceGroupedConvFwdMultipleABD<NumDimSpatial,
InLayout,
WeiLayout,
DLayouts,
OutLayout,
InDataType,
WeiDataType,
DDataTypes,
OutDataType,
PassThrough,
PassThrough,
ConvScaleAdd,
AComputeType,
BComputeType>;
static auto GetInstances()
{
std::vector<std::unique_ptr<DeviceOp>> op_ptrs;
if constexpr(NumDimSpatial == 3 && is_same_v<InLayout, NDHWGC> &&
is_same_v<WeiLayout, GKZYXC> && is_same_v<OutLayout, NDHWGK>)
{
#ifdef CK_ENABLE_FP8
if constexpr(is_same_v<InDataType, f8_t> && is_same_v<WeiDataType, f8_t> &&
is_same_v<OutDataType, f8_t> && is_same_v<AComputeType, f8_t> &&
is_same_v<BComputeType, f8_t>)
{
add_device_grouped_conv3d_fwd_xdl_convscale_add_ndhwgc_gkzyxc_ndhwgk_f8_instances(
op_ptrs);
}
#endif
}
return op_ptrs;
}
};
} // namespace instance
} // namespace device
} // namespace tensor_operation
} // namespace ck
library/include/ck/library/tensor_operation_instance/gpu/grouped_convolution_forward_convscale_relu.hpp 0 → 100644
View file @ 3d61f89a
// 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/device_grouped_conv_fwd_multiple_abd.hpp"
#include "ck/tensor_operation/gpu/element/combined_element_wise_operation.hpp"
#include "ck/tensor_operation/gpu/element/unary_element_wise_operation.hpp"
#include "ck/library/tensor_operation_instance/device_operation_instance_factory.hpp"
namespace ck {
namespace tensor_operation {
namespace device {
namespace instance {
using PassThrough = ck::tensor_operation::element_wise::PassThrough;
using ConvScaleRelu = ck::tensor_operation::element_wise::ConvScaleRelu;
#ifdef CK_ENABLE_FP8
void add_device_grouped_conv3d_fwd_xdl_convscale_relu_ndhwgc_gkzyxc_ndhwgk_f8_instances(
std::vector<std::unique_ptr<DeviceGroupedConvFwdMultipleABD<3,
NDHWGC,
GKZYXC,
ck::Tuple<>,
NDHWGK,
F8,
F8,
ck::Tuple<>,
F8,
PassThrough,
PassThrough,
ConvScaleRelu,
F8,
F8>>>& instances);
#endif
template <ck::index_t NumDimSpatial,
typename InLayout,
typename WeiLayout,
typename DLayouts,
typename OutLayout,
typename InDataType,
typename WeiDataType,
typename DDataTypes,
typename OutDataType,
typename AComputeType,
typename BComputeType>
struct DeviceOperationInstanceFactory<
ck::tensor_operation::device::DeviceGroupedConvFwdMultipleABD<NumDimSpatial,
InLayout,
WeiLayout,
DLayouts,
OutLayout,
InDataType,
WeiDataType,
DDataTypes,
OutDataType,
PassThrough,
PassThrough,
ConvScaleRelu,
AComputeType,
BComputeType>>
{
using DeviceOp = DeviceGroupedConvFwdMultipleABD<NumDimSpatial,
InLayout,
WeiLayout,
DLayouts,
OutLayout,
InDataType,
WeiDataType,
DDataTypes,
OutDataType,
PassThrough,
PassThrough,
ConvScaleRelu,
AComputeType,
BComputeType>;
static auto GetInstances()
{
std::vector<std::unique_ptr<DeviceOp>> op_ptrs;
if constexpr(NumDimSpatial == 3 && is_same_v<InLayout, NDHWGC> &&
is_same_v<WeiLayout, GKZYXC> && is_same_v<OutLayout, NDHWGK>)
{
#ifdef CK_ENABLE_FP8
if constexpr(is_same_v<InDataType, f8_t> && is_same_v<WeiDataType, f8_t> &&
is_same_v<OutDataType, f8_t> && is_same_v<AComputeType, f8_t> &&
is_same_v<BComputeType, f8_t>)
{
add_device_grouped_conv3d_fwd_xdl_convscale_relu_ndhwgc_gkzyxc_ndhwgk_f8_instances(
op_ptrs);
}
#endif
}
return op_ptrs;
}
};
namespace ew = ck::tensor_operation::element_wise;
using CombConvScaleRelu = ew::UnaryCombinedOp<ew::Scale, ew::Scale, ew::Relu>;
#ifdef CK_ENABLE_FP8
void add_device_grouped_conv3d_fwd_xdl_combconvscale_relu_ndhwgc_gkzyxc_ndhwgk_f8_f8_f32_instances(
std::vector<std::unique_ptr<DeviceGroupedConvFwdMultipleABD<3,
NDHWGC,
GKZYXC,
ck::Tuple<>,
NDHWGK,
F8,
F8,
ck::Tuple<>,
F32,
PassThrough,
PassThrough,
CombConvScaleRelu,
F8,
F8>>>& instances);
#endif
template <ck::index_t NumDimSpatial,
typename InLayout,
typename WeiLayout,
typename DLayouts,
typename OutLayout,
typename InDataType,
typename WeiDataType,
typename DDataTypes,
typename OutDataType,
typename AComputeType,
typename BComputeType>
struct DeviceOperationInstanceFactory<
ck::tensor_operation::device::DeviceGroupedConvFwdMultipleABD<NumDimSpatial,
InLayout,
WeiLayout,
DLayouts,
OutLayout,
InDataType,
WeiDataType,
DDataTypes,
OutDataType,
PassThrough,
PassThrough,
CombConvScaleRelu,
AComputeType,
BComputeType>>
{
using DeviceOp = DeviceGroupedConvFwdMultipleABD<NumDimSpatial,
InLayout,
WeiLayout,
DLayouts,
OutLayout,
InDataType,
WeiDataType,
DDataTypes,
OutDataType,
PassThrough,
PassThrough,
CombConvScaleRelu,
AComputeType,
BComputeType>;
static auto GetInstances()
{
std::vector<std::unique_ptr<DeviceOp>> op_ptrs;
if constexpr(NumDimSpatial == 3 && is_same_v<InLayout, NDHWGC> &&
is_same_v<WeiLayout, GKZYXC> && is_same_v<OutLayout, NDHWGK>)
{
#ifdef CK_ENABLE_FP8
if constexpr(is_same_v<InDataType, f8_t> && is_same_v<WeiDataType, f8_t> &&
is_same_v<OutDataType, F32> && is_same_v<AComputeType, f8_t> &&
is_same_v<BComputeType, f8_t>)
{
add_device_grouped_conv3d_fwd_xdl_combconvscale_relu_ndhwgc_gkzyxc_ndhwgk_f8_f8_f32_instances(
op_ptrs);
}
#endif
}
return op_ptrs;
}
};
} // namespace instance
} // namespace device
} // namespace tensor_operation
} // namespace ck
library/include/ck/library/tensor_operation_instance/gpu/grouped_convolution_forward_xdl_large_tensor.inc 0 → 100644
View file @ 3d61f89a
// SPDX-License-Identifier: MIT
// Copyright (c) 2024, Advanced Micro Devices, Inc. All rights reserved.
#pragma once
namespace ck {
namespace tensor_operation {
namespace device {
namespace instance {
// grouped conv2d forward, NHWGC/GKYXC/NHWGK
#ifdef CK_ENABLE_BF16
void add_device_grouped_conv2d_fwd_xdl_large_tensor_nhwgc_gkyxc_nhwgk_bf16_instances(
std::vector<std::unique_ptr<DeviceGroupedConvFwdMultipleABD<2,
NHWGC,
GKYXC,
Empty_Tuple,
NHWGK,
BF16,
BF16,
Empty_Tuple,
BF16,
PassThrough,
PassThrough,
PassThrough>>>& instances);
#endif
#ifdef CK_ENABLE_FP16
void add_device_grouped_conv2d_fwd_xdl_large_tensor_nhwgc_gkyxc_nhwgk_f16_instances(
std::vector<std::unique_ptr<DeviceGroupedConvFwdMultipleABD<2,
NHWGC,
GKYXC,
Empty_Tuple,
NHWGK,
F16,
F16,
Empty_Tuple,
F16,
PassThrough,
PassThrough,
PassThrough>>>& instances);
#endif
#ifdef CK_ENABLE_FP32
void add_device_grouped_conv2d_fwd_xdl_large_tensor_nhwgc_gkyxc_nhwgk_f32_instances(
std::vector<std::unique_ptr<DeviceGroupedConvFwdMultipleABD<2,
NHWGC,
GKYXC,
Empty_Tuple,
NHWGK,
F32,
F32,
Empty_Tuple,
F32,
PassThrough,
PassThrough,
PassThrough>>>& instances);
#endif
#ifdef CK_ENABLE_BF16
// grouped conv3d forward, NDHWGC/GKZYXC/NDHWGK
void add_device_grouped_conv3d_fwd_xdl_large_tensor_ndhwgc_gkzyxc_ndhwgk_bf16_instances(
std::vector<std::unique_ptr<DeviceGroupedConvFwdMultipleABD<3,
NDHWGC,
GKZYXC,
Empty_Tuple,
NDHWGK,
BF16,
BF16,
Empty_Tuple,
BF16,
PassThrough,
PassThrough,
PassThrough>>>& instances);
#endif
#ifdef CK_ENABLE_FP16
void add_device_grouped_conv3d_fwd_xdl_large_tensor_ndhwgc_gkzyxc_ndhwgk_f16_instances(
std::vector<std::unique_ptr<DeviceGroupedConvFwdMultipleABD<3,
NDHWGC,
GKZYXC,
Empty_Tuple,
NDHWGK,
F16,
F16,
Empty_Tuple,
F16,
PassThrough,
PassThrough,
PassThrough>>>& instances);
#endif
#ifdef CK_ENABLE_FP32
void add_device_grouped_conv3d_fwd_xdl_large_tensor_ndhwgc_gkzyxc_ndhwgk_f32_instances(
std::vector<std::unique_ptr<DeviceGroupedConvFwdMultipleABD<3,
NDHWGC,
GKZYXC,
Empty_Tuple,
NDHWGK,
F32,
F32,
Empty_Tuple,
F32,
PassThrough,
PassThrough,
PassThrough>>>& instances);
#endif
} // namespace instance
} // namespace device
} // namespace tensor_operation
} // namespace ck
library/include/ck/library/tensor_operation_instance/gpu/permute_scale.hpp
View file @ 3d61f89a
......@@ -70,6 +70,12 @@ void add_device_permute_scale_6d_f32_instances(
DeviceElementwise<ck::Tuple<F32>, ck::Tuple<F32>, element_wise::Scale, 6>>>&);
#endif
#ifdef CK_ENABLE_FP8
void add_device_permute_scale_6d_f32_f8_instances(
std::vector<std::unique_ptr<
DeviceElementwise<ck::Tuple<F32>, ck::Tuple<F8>, element_wise::Scale, 6>>>&);
#endif
template <typename InDataTypeTuple,
typename OutDataTypeTuple,
typename ElementwiseOperation,
......@@ -184,6 +190,13 @@ struct DeviceOperationInstanceFactory<
{
add_device_permute_scale_6d_f16_instances(op_ptrs);
}
#endif
#ifdef CK_ENABLE_FP8
if constexpr(is_same_v<InDataTypeTuple, ck::Tuple<F32>> &&
is_same_v<OutDataTypeTuple, ck::Tuple<F8>>)
{
add_device_permute_scale_6d_f32_f8_instances(op_ptrs);
}
#endif
}
return op_ptrs;
......
library/include/ck/library/tensor_operation_instance/gpu/permute_scale/device_permute_scale_instances.hpp
View file @ 3d61f89a
......@@ -10,6 +10,7 @@ namespace tensor_operation {
namespace device {
namespace instance {
using F8 = ck::f8_t;
using F16 = ck::half_t;
using F32 = float;
......@@ -183,6 +184,51 @@ using device_permute_scale_f32_instances = std::tuple<
DeviceElementwiseImpl<ck::Tuple<F32>, ck::Tuple<F32>, ElementwiseOp, NDims, 32, 32, 16, 4, 4, ck::Sequence<1, 0>, ck::Sequence<1>, ck::Sequence<1>>,
DeviceElementwiseImpl<ck::Tuple<F32>, ck::Tuple<F32>, ElementwiseOp, NDims, 32, 16, 32, 4, 4, ck::Sequence<1, 0>, ck::Sequence<1>, ck::Sequence<1>>
>;
#ifdef CK_ENABLE_FP8
template <index_t NDims,
typename ElementwiseOp>
using device_permute_scale_f32_f8_instances = std::tuple<
DeviceElementwiseImpl<ck::Tuple<F32>, ck::Tuple<F8>, ElementwiseOp, NDims, 256, 64, 64, 4, 4, ck::Sequence<1, 0>, ck::Sequence<4>, ck::Sequence<4>>,
DeviceElementwiseImpl<ck::Tuple<F32>, ck::Tuple<F8>, ElementwiseOp, NDims, 256, 128, 32, 4, 4, ck::Sequence<1, 0>, ck::Sequence<4>, ck::Sequence<4>>,
DeviceElementwiseImpl<ck::Tuple<F32>, ck::Tuple<F8>, ElementwiseOp, NDims, 256, 32, 128, 4, 4, ck::Sequence<1, 0>, ck::Sequence<4>, ck::Sequence<4>>,
DeviceElementwiseImpl<ck::Tuple<F32>, ck::Tuple<F8>, ElementwiseOp, NDims, 128, 64, 32, 4, 4, ck::Sequence<1, 0>, ck::Sequence<4>, ck::Sequence<4>>,
DeviceElementwiseImpl<ck::Tuple<F32>, ck::Tuple<F8>, ElementwiseOp, NDims, 128, 32, 64, 4, 4, ck::Sequence<1, 0>, ck::Sequence<4>, ck::Sequence<4>>,
DeviceElementwiseImpl<ck::Tuple<F32>, ck::Tuple<F8>, ElementwiseOp, NDims, 128, 16, 128, 4, 4, ck::Sequence<1, 0>, ck::Sequence<4>, ck::Sequence<4>>,
DeviceElementwiseImpl<ck::Tuple<F32>, ck::Tuple<F8>, ElementwiseOp, NDims, 128, 128, 16, 4, 4, ck::Sequence<1, 0>, ck::Sequence<4>, ck::Sequence<4>>,
DeviceElementwiseImpl<ck::Tuple<F32>, ck::Tuple<F8>, ElementwiseOp, NDims, 64, 32, 32, 4, 4, ck::Sequence<1, 0>, ck::Sequence<4>, ck::Sequence<4>>,
DeviceElementwiseImpl<ck::Tuple<F32>, ck::Tuple<F8>, ElementwiseOp, NDims, 64, 16, 64, 4, 4, ck::Sequence<1, 0>, ck::Sequence<4>, ck::Sequence<4>>,
DeviceElementwiseImpl<ck::Tuple<F32>, ck::Tuple<F8>, ElementwiseOp, NDims, 64, 64, 16, 4, 4, ck::Sequence<1, 0>, ck::Sequence<4>, ck::Sequence<4>>,
DeviceElementwiseImpl<ck::Tuple<F32>, ck::Tuple<F8>, ElementwiseOp, NDims, 32, 32, 16, 4, 4, ck::Sequence<1, 0>, ck::Sequence<4>, ck::Sequence<4>>,
DeviceElementwiseImpl<ck::Tuple<F32>, ck::Tuple<F8>, ElementwiseOp, NDims, 32, 16, 32, 4, 4, ck::Sequence<1, 0>, ck::Sequence<4>, ck::Sequence<4>>,
DeviceElementwiseImpl<ck::Tuple<F32>, ck::Tuple<F8>, ElementwiseOp, NDims, 256, 128, 128, 8, 8, ck::Sequence<1, 0>, ck::Sequence<8>, ck::Sequence<8>>,
DeviceElementwiseImpl<ck::Tuple<F32>, ck::Tuple<F8>, ElementwiseOp, NDims, 256, 256, 64, 8, 8, ck::Sequence<1, 0>, ck::Sequence<8>, ck::Sequence<8>>,
DeviceElementwiseImpl<ck::Tuple<F32>, ck::Tuple<F8>, ElementwiseOp, NDims, 256, 64, 256, 8, 8, ck::Sequence<1, 0>, ck::Sequence<8>, ck::Sequence<8>>,
DeviceElementwiseImpl<ck::Tuple<F32>, ck::Tuple<F8>, ElementwiseOp, NDims, 128, 128, 64, 8, 8, ck::Sequence<1, 0>, ck::Sequence<8>, ck::Sequence<8>>,
DeviceElementwiseImpl<ck::Tuple<F32>, ck::Tuple<F8>, ElementwiseOp, NDims, 128, 64, 128, 8, 8, ck::Sequence<1, 0>, ck::Sequence<8>, ck::Sequence<8>>,
DeviceElementwiseImpl<ck::Tuple<F32>, ck::Tuple<F8>, ElementwiseOp, NDims, 128, 32, 256, 8, 8, ck::Sequence<1, 0>, ck::Sequence<8>, ck::Sequence<8>>,
DeviceElementwiseImpl<ck::Tuple<F32>, ck::Tuple<F8>, ElementwiseOp, NDims, 128, 256, 32, 8, 8, ck::Sequence<1, 0>, ck::Sequence<8>, ck::Sequence<8>>,
DeviceElementwiseImpl<ck::Tuple<F32>, ck::Tuple<F8>, ElementwiseOp, NDims, 64, 64, 64, 8, 8, ck::Sequence<1, 0>, ck::Sequence<8>, ck::Sequence<8>>,
DeviceElementwiseImpl<ck::Tuple<F32>, ck::Tuple<F8>, ElementwiseOp, NDims, 64, 32, 128, 8, 8, ck::Sequence<1, 0>, ck::Sequence<8>, ck::Sequence<8>>,
DeviceElementwiseImpl<ck::Tuple<F32>, ck::Tuple<F8>, ElementwiseOp, NDims, 64, 128, 32, 8, 8, ck::Sequence<1, 0>, ck::Sequence<8>, ck::Sequence<8>>,
DeviceElementwiseImpl<ck::Tuple<F32>, ck::Tuple<F8>, ElementwiseOp, NDims, 32, 64, 32, 8, 8, ck::Sequence<1, 0>, ck::Sequence<8>, ck::Sequence<8>>,
DeviceElementwiseImpl<ck::Tuple<F32>, ck::Tuple<F8>, ElementwiseOp, NDims, 32, 32, 64, 8, 8, ck::Sequence<1, 0>, ck::Sequence<8>, ck::Sequence<8>>,
DeviceElementwiseImpl<ck::Tuple<F32>, ck::Tuple<F8>, ElementwiseOp, NDims, 256, 64, 64, 4, 4, ck::Sequence<1, 0>, ck::Sequence<1>, ck::Sequence<1>>,
DeviceElementwiseImpl<ck::Tuple<F32>, ck::Tuple<F8>, ElementwiseOp, NDims, 256, 128, 32, 4, 4, ck::Sequence<1, 0>, ck::Sequence<1>, ck::Sequence<1>>,
DeviceElementwiseImpl<ck::Tuple<F32>, ck::Tuple<F8>, ElementwiseOp, NDims, 256, 32, 128, 4, 4, ck::Sequence<1, 0>, ck::Sequence<1>, ck::Sequence<1>>,
DeviceElementwiseImpl<ck::Tuple<F32>, ck::Tuple<F8>, ElementwiseOp, NDims, 128, 64, 32, 4, 4, ck::Sequence<1, 0>, ck::Sequence<1>, ck::Sequence<1>>,
DeviceElementwiseImpl<ck::Tuple<F32>, ck::Tuple<F8>, ElementwiseOp, NDims, 128, 32, 64, 4, 4, ck::Sequence<1, 0>, ck::Sequence<1>, ck::Sequence<1>>,
DeviceElementwiseImpl<ck::Tuple<F32>, ck::Tuple<F8>, ElementwiseOp, NDims, 128, 16, 128, 4, 4, ck::Sequence<1, 0>, ck::Sequence<1>, ck::Sequence<1>>,
DeviceElementwiseImpl<ck::Tuple<F32>, ck::Tuple<F8>, ElementwiseOp, NDims, 128, 128, 16, 4, 4, ck::Sequence<1, 0>, ck::Sequence<1>, ck::Sequence<1>>,
DeviceElementwiseImpl<ck::Tuple<F32>, ck::Tuple<F8>, ElementwiseOp, NDims, 64, 32, 32, 4, 4, ck::Sequence<1, 0>, ck::Sequence<1>, ck::Sequence<1>>,
DeviceElementwiseImpl<ck::Tuple<F32>, ck::Tuple<F8>, ElementwiseOp, NDims, 64, 16, 64, 4, 4, ck::Sequence<1, 0>, ck::Sequence<1>, ck::Sequence<1>>,
DeviceElementwiseImpl<ck::Tuple<F32>, ck::Tuple<F8>, ElementwiseOp, NDims, 64, 64, 16, 4, 4, ck::Sequence<1, 0>, ck::Sequence<1>, ck::Sequence<1>>,
DeviceElementwiseImpl<ck::Tuple<F32>, ck::Tuple<F8>, ElementwiseOp, NDims, 32, 32, 16, 4, 4, ck::Sequence<1, 0>, ck::Sequence<1>, ck::Sequence<1>>,
DeviceElementwiseImpl<ck::Tuple<F32>, ck::Tuple<F8>, ElementwiseOp, NDims, 32, 16, 32, 4, 4, ck::Sequence<1, 0>, ck::Sequence<1>, ck::Sequence<1>>
>;
#endif
// clang-format on
} // namespace instance
......
library/include/ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_blockwise_f32_f32_f32_amax.hpp
View file @ 3d61f89a
......@@ -14,15 +14,24 @@ namespace device {
namespace instance {
// clang-format off
// InDataType | AccDataType | OutDataType | Rank | NumReduceDim | ReduceOperation | InElementwiseOp | AccElementwiseOp | PropagateNan | UseIndex
extern template void add_device_reduce_instance_blockwise<F32, F32, F32, 4, 3, ReduceAMax, UnaryAbs, PassThrough, false, false>(std::vector<DeviceReducePtr<F32, F32, F32, 4, 3, ReduceAMax, UnaryAbs, PassThrough, false, false>>&);
extern template void add_device_reduce_instance_blockwise<F32, F32, F32, 4, 4, ReduceAMax, UnaryAbs, PassThrough, false, false>(std::vector<DeviceReducePtr<F32, F32, F32, 4, 4, ReduceAMax, UnaryAbs, PassThrough, false, false>>&);
extern template void add_device_reduce_instance_blockwise<F32, F32, F32, 4, 1, ReduceAMax, UnaryAbs, PassThrough, false, false>(std::vector<DeviceReducePtr<F32, F32, F32, 4, 1, ReduceAMax, UnaryAbs, PassThrough, false, false>>&);
extern template void add_device_reduce_instance_blockwise<F32, F32, F32, 2, 1, ReduceAMax, UnaryAbs, PassThrough, false, false>(std::vector<DeviceReducePtr<F32, F32, F32, 2, 1, ReduceAMax, UnaryAbs, PassThrough, false, false>>&);
extern template void add_device_reduce_instance_blockwise<F32, F32, F32, 4, 3, ReduceAMax, UnaryAbs, PassThrough, false, true>(std::vector<DeviceReducePtr<F32, F32, F32, 4, 3, ReduceAMax, UnaryAbs, PassThrough, false, true>>&);
extern template void add_device_reduce_instance_blockwise<F32, F32, F32, 4, 4, ReduceAMax, UnaryAbs, PassThrough, false, true>(std::vector<DeviceReducePtr<F32, F32, F32, 4, 4, ReduceAMax, UnaryAbs, PassThrough, false, true>>&);
extern template void add_device_reduce_instance_blockwise<F32, F32, F32, 4, 1, ReduceAMax, UnaryAbs, PassThrough, false, true>(std::vector<DeviceReducePtr<F32, F32, F32, 4, 1, ReduceAMax, UnaryAbs, PassThrough, false, true>>&);
extern template void add_device_reduce_instance_blockwise<F32, F32, F32, 2, 1, ReduceAMax, UnaryAbs, PassThrough, false, true>(std::vector<DeviceReducePtr<F32, F32, F32, 2, 1, ReduceAMax, UnaryAbs, PassThrough, false, true>>&);
// InDataType | AccDataType | OutDataType | Rank | NumReduceDim | ReduceOperation | InElementwiseOp | AccElementwiseOp | PropagateNan | UseIndex
extern template void add_device_reduce_instance_blockwise< F32, F32, F32, 4, 3, ReduceAMax, UnaryAbs, PassThrough, false, false>(std::vector<DeviceReducePtr<F32, F32, F32, 4, 3, ReduceAMax, UnaryAbs, PassThrough, false, false>>&);
extern template void add_device_reduce_instance_blockwise< F32, F32, F32, 4, 4, ReduceAMax, UnaryAbs, PassThrough, false, false>(std::vector<DeviceReducePtr<F32, F32, F32, 4, 4, ReduceAMax, UnaryAbs, PassThrough, false, false>>&);
extern template void add_device_reduce_instance_blockwise< F32, F32, F32, 4, 1, ReduceAMax, UnaryAbs, PassThrough, false, false>(std::vector<DeviceReducePtr<F32, F32, F32, 4, 1, ReduceAMax, UnaryAbs, PassThrough, false, false>>&);
extern template void add_device_reduce_instance_blockwise< F32, F32, F32, 2, 1, ReduceAMax, UnaryAbs, PassThrough, false, false>(std::vector<DeviceReducePtr<F32, F32, F32, 2, 1, ReduceAMax, UnaryAbs, PassThrough, false, false>>&);
extern template void add_device_reduce_instance_blockwise< F32, F32, F32, 4, 3, ReduceAMax, UnaryAbs, PassThrough, false, true>(std::vector<DeviceReducePtr<F32, F32, F32, 4, 3, ReduceAMax, UnaryAbs, PassThrough, false, true>>&);
extern template void add_device_reduce_instance_blockwise< F32, F32, F32, 4, 4, ReduceAMax, UnaryAbs, PassThrough, false, true>(std::vector<DeviceReducePtr<F32, F32, F32, 4, 4, ReduceAMax, UnaryAbs, PassThrough, false, true>>&);
extern template void add_device_reduce_instance_blockwise< F32, F32, F32, 4, 1, ReduceAMax, UnaryAbs, PassThrough, false, true>(std::vector<DeviceReducePtr<F32, F32, F32, 4, 1, ReduceAMax, UnaryAbs, PassThrough, false, true>>&);
extern template void add_device_reduce_instance_blockwise< F32, F32, F32, 2, 1, ReduceAMax, UnaryAbs, PassThrough, false, true>(std::vector<DeviceReducePtr<F32, F32, F32, 2, 1, ReduceAMax, UnaryAbs, PassThrough, false, true>>&);
extern template void add_device_reduce_instance_blockwise< F32, F32, F32, 6, 6, ReduceAMax, UnaryAbs, PassThrough, true, false>(std::vector<DeviceReducePtr<F32, F32, F32, 6, 6, ReduceAMax, UnaryAbs, PassThrough, true, false>>&);
extern template void add_device_reduce_instance_blockwise< F32, F32, F32, 5, 5, ReduceAMax, UnaryAbs, PassThrough, true, false>(std::vector<DeviceReducePtr<F32, F32, F32, 5, 5, ReduceAMax, UnaryAbs, PassThrough, true, false>>&);
extern template void add_device_reduce_instance_blockwise< F32, F32, F32, 4, 4, ReduceAMax, UnaryAbs, PassThrough, true, false>(std::vector<DeviceReducePtr<F32, F32, F32, 4, 4, ReduceAMax, UnaryAbs, PassThrough, true, false>>&);
extern template void add_device_reduce_instance_blockwise< F32, F32, F32, 6, 3, ReduceAMax, UnaryAbs, PassThrough, true, false>(std::vector<DeviceReducePtr<F32, F32, F32, 6, 3, ReduceAMax, UnaryAbs, PassThrough, true, false>>&);
extern template void add_device_reduce_instance_blockwise< F32, F32, F32, 5, 3, ReduceAMax, UnaryAbs, PassThrough, true, false>(std::vector<DeviceReducePtr<F32, F32, F32, 5, 3, ReduceAMax, UnaryAbs, PassThrough, true, false>>&);
extern template void add_device_reduce_instance_blockwise< F32, F32, F32, 4, 3, ReduceAMax, UnaryAbs, PassThrough, true, false>(std::vector<DeviceReducePtr<F32, F32, F32, 4, 3, ReduceAMax, UnaryAbs, PassThrough, true, false>>&);
extern template void add_device_reduce_instance_blockwise< F32, F32, F32, 3, 3, ReduceAMax, PassThrough, PassThrough, true, false>(std::vector<DeviceReducePtr<F32, F32, F32, 3, 3, ReduceAMax, PassThrough, PassThrough, true, false>>&);
extern template void add_device_reduce_instance_blockwise< F32, F32, F32, 2, 2, ReduceAMax, PassThrough, PassThrough, true, false>(std::vector<DeviceReducePtr<F32, F32, F32, 2, 2, ReduceAMax, PassThrough, PassThrough, true, false>>&);
extern template void add_device_reduce_instance_blockwise< F32, F32, F32, 1, 1, ReduceAMax, PassThrough, PassThrough, true, false>(std::vector<DeviceReducePtr<F32, F32, F32, 1, 1, ReduceAMax, PassThrough, PassThrough, true, false>>&);
// clang-format on
} // namespace instance
......
library/include/ck/library/utility/check_err.hpp
View file @ 3d61f89a
......@@ -146,7 +146,7 @@ check_err(const Range& out,
bool res{true};
int err_count = 0;
double err = 0;
double max_err = std::numeric_limits<ranges::range_value_t<Range>>::min();
double max_err = NumericLimits<ranges::range_value_t<Range>>::Min();
for(std::size_t i = 0; i < ref.size(); ++i)
{
const double o = type_convert<float>(*std::next(std::begin(out), i));
......@@ -178,7 +178,9 @@ check_err(const Range& out,
template <typename Range, typename RefRange>
std::enable_if_t<(std::is_same_v<ranges::range_value_t<Range>, ranges::range_value_t<RefRange>> &&
std::is_integral_v<ranges::range_value_t<Range>> &&
!std::is_same_v<ranges::range_value_t<Range>, bhalf_t>)
!std::is_same_v<ranges::range_value_t<Range>, bhalf_t> &&
!std::is_same_v<ranges::range_value_t<Range>, f8_t> &&
!std::is_same_v<ranges::range_value_t<Range>, bf8_t>)
#ifdef CK_EXPERIMENTAL_BIT_INT_EXTENSION_INT4
|| std::is_same_v<ranges::range_value_t<Range>, int4_t>
#endif
......
library/include/ck/library/utility/convolution_parameter.hpp
View file @ 3d61f89a
// 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
......@@ -31,23 +31,35 @@ struct ConvParam
const std::vector<ck::index_t>& left_pads,
const std::vector<ck::index_t>& right_pads);
ck::index_t num_dim_spatial_;
ck::index_t G_;
ck::index_t N_;
ck::index_t K_;
ck::index_t C_;
std::vector<ck::index_t> filter_spatial_lengths_;
std::vector<ck::index_t> input_spatial_lengths_;
std::vector<ck::index_t> output_spatial_lengths_;
std::vector<ck::index_t> conv_filter_strides_;
std::vector<ck::index_t> conv_filter_dilations_;
std::vector<ck::index_t> input_left_pads_;
std::vector<ck::index_t> input_right_pads_;
std::vector<ck::index_t> GetOutputSpatialLengths() const;
ConvParam(ck::long_index_t n_dim,
ck::long_index_t group_count,
ck::long_index_t n_batch,
ck::long_index_t n_out_channels,
ck::long_index_t n_in_channels,
const std::vector<ck::long_index_t>& filters_len,
const std::vector<ck::long_index_t>& input_len,
const std::vector<ck::long_index_t>& strides,
const std::vector<ck::long_index_t>& dilations,
const std::vector<ck::long_index_t>& left_pads,
const std::vector<ck::long_index_t>& right_pads);
ck::long_index_t num_dim_spatial_;
ck::long_index_t G_;
ck::long_index_t N_;
ck::long_index_t K_;
ck::long_index_t C_;
std::vector<ck::long_index_t> filter_spatial_lengths_;
std::vector<ck::long_index_t> input_spatial_lengths_;
std::vector<ck::long_index_t> output_spatial_lengths_;
std::vector<ck::long_index_t> conv_filter_strides_;
std::vector<ck::long_index_t> conv_filter_dilations_;
std::vector<ck::long_index_t> input_left_pads_;
std::vector<ck::long_index_t> input_right_pads_;
std::vector<ck::long_index_t> GetOutputSpatialLengths() const;
std::size_t GetFlops() const;
......
library/include/ck/library/utility/host_tensor.hpp
View file @ 3d61f89a
......@@ -96,9 +96,16 @@ struct HostTensorDescriptor
this->CalculateStrides();
}
HostTensorDescriptor(const std::initializer_list<ck::long_index_t>& lens)
: mLens(lens.begin(), lens.end())
{
this->CalculateStrides();
}
template <typename Lengths,
typename = std::enable_if_t<
std::is_convertible_v<ck::ranges::range_value_t<Lengths>, std::size_t>>>
std::is_convertible_v<ck::ranges::range_value_t<Lengths>, std::size_t> ||
std::is_convertible_v<ck::ranges::range_value_t<Lengths>, ck::long_index_t>>>
HostTensorDescriptor(const Lengths& lens) : mLens(lens.begin(), lens.end())
{
this->CalculateStrides();
......@@ -114,11 +121,19 @@ struct HostTensorDescriptor
{
}
HostTensorDescriptor(const std::initializer_list<ck::long_index_t>& lens,
const std::initializer_list<ck::long_index_t>& strides)
: mLens(lens.begin(), lens.end()), mStrides(strides.begin(), strides.end())
{
}
template <typename Lengths,
typename Strides,
typename = std::enable_if_t<
std::is_convertible_v<ck::ranges::range_value_t<Lengths>, std::size_t> &&
std::is_convertible_v<ck::ranges::range_value_t<Strides>, std::size_t>>>
(std::is_convertible_v<ck::ranges::range_value_t<Lengths>, std::size_t> &&
std::is_convertible_v<ck::ranges::range_value_t<Strides>, std::size_t>) ||
(std::is_convertible_v<ck::ranges::range_value_t<Lengths>, ck::long_index_t> &&
std::is_convertible_v<ck::ranges::range_value_t<Strides>, ck::long_index_t>)>>
HostTensorDescriptor(const Lengths& lens, const Strides& strides)
: mLens(lens.begin(), lens.end()), mStrides(strides.begin(), strides.end())
{
......
library/src/tensor_operation_instance/gpu/CMakeLists.txt
View file @ 3d61f89a
......@@ -64,6 +64,13 @@ function(add_instance_library INSTANCE_NAME)
list(REMOVE_ITEM ARGN "${source}")
endif()
endforeach()
# Do not build mha instances if gfx94 targets are not on the target list
foreach(source IN LISTS ARGN)
if(NOT INST_TARGETS MATCHES "gfx94" AND source MATCHES "mha")
message("removing mha instance ${source} ")
list(REMOVE_ITEM ARGN "${source}")
endif()
endforeach()
#only continue if there are some source files left on the list
if(ARGN)
set(INST_OBJ)
......@@ -74,9 +81,11 @@ function(add_instance_library INSTANCE_NAME)
set(INST_TARGETS ${GPU_TARGETS})
endif()
if(source MATCHES "_xdl")
list(REMOVE_ITEM INST_TARGETS gfx900 gfx906 gfx1030 gfx1100 gfx1101 gfx1102 gfx1103)
list(REMOVE_ITEM INST_TARGETS gfx900 gfx906 gfx1030 gfx1100 gfx1101 gfx1102 gfx1103 gfx1200 gfx1201)
elseif(ARGN MATCHES "_wmma")
list(REMOVE_ITEM INST_TARGETS gfx900 gfx906 gfx908 gfx90a gfx940 gfx941 gfx942 gfx1030)
elseif(ARGN MATCHES "mha")
list(REMOVE_ITEM INST_TARGETS gfx900 gfx906 gfx908 gfx90a gfx1030 gfx1100 gfx1101 gfx1102 gfx1103 gfx1200 gfx1201)
endif()
set(offload_targets)
foreach(target IN LISTS INST_TARGETS)
......@@ -86,7 +95,29 @@ function(add_instance_library INSTANCE_NAME)
list(APPEND INST_OBJ ${source})
endforeach()
add_library(${INSTANCE_NAME} OBJECT ${INST_OBJ})
# Allow comparing floating points directly in order to check sentinel values
if(${INSTANCE_NAME} STREQUAL "device_mha_instance")
if(NOT DEFINED FMHA_FWD_FAST_EXP2)
set(FMHA_FWD_FAST_EXP2 true)
endif()
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)
else()
list(APPEND EXAMPLE_FMHA_FWD_COMPILE_OPTIONS -Wno-undefined-func-template -DCK_TILE_FMHA_FWD_FAST_EXP2=0)
endif()
list(APPEND EXAMPLE_FMHA_FWD_COMPILE_OPTIONS -Wno-float-equal)
target_compile_options(device_mha_instance PRIVATE ${EXAMPLE_FMHA_FWD_COMPILE_OPTIONS})
endif()
target_compile_features(${INSTANCE_NAME} PUBLIC)
# flags to compress the library
if(NOT WIN32 AND ${hip_VERSION_FLAT} GREATER 600241132)
message("Adding --offload-compress flag for ${INSTANCE_NAME}")
target_compile_options(${INSTANCE_NAME} PRIVATE --offload-compress)
endif()
set_target_properties(${INSTANCE_NAME} PROPERTIES POSITION_INDEPENDENT_CODE ON)
clang_tidy_check(${INSTANCE_NAME})
set(result 0)
......@@ -286,20 +317,22 @@ if(CK_DEVICE_CONV_INSTANCES)
)
endif()
if(CK_DEVICE_MHA_INSTANCES)
add_library(device_mha_operations STATIC ${CK_DEVICE_MHA_INSTANCES})
add_library(composablekernels::device_mha_operations ALIAS device_mha_operations)
target_compile_features(device_mha_operations PUBLIC)
set_target_properties(device_mha_operations PROPERTIES POSITION_INDEPENDENT_CODE ON)
target_include_directories(device_mha_operations PUBLIC
$<INSTALL_INTERFACE:${CMAKE_INSTALL_INCLUDEDIR}/ck/library/tensor_operation_instance/gpu/mha>
)
rocm_install(TARGETS device_mha_operations
EXPORT device_mha_operationsTargets)
rocm_install(EXPORT device_mha_operationsTargets
FILE composable_kerneldevice_mha_operationsTargets.cmake
NAMESPACE composable_kernel::
DESTINATION ${CMAKE_INSTALL_LIBDIR}/cmake/composable_kernel
)
set(gpu_list ${INST_TARGETS})
list(FILTER gpu_list INCLUDE REGEX "^gfx94")
if(gpu_list)
add_library(device_mha_operations STATIC ${CK_DEVICE_MHA_INSTANCES})
add_library(composablekernels::device_mha_operations ALIAS device_mha_operations)
target_compile_features(device_mha_operations PUBLIC)
set_target_properties(device_mha_operations PROPERTIES POSITION_INDEPENDENT_CODE ON)
rocm_install(TARGETS device_mha_operations
EXPORT device_mha_operationsTargets)
rocm_install(EXPORT device_mha_operationsTargets
FILE composable_kerneldevice_mha_operationsTargets.cmake
NAMESPACE composable_kernel::
DESTINATION ${CMAKE_INSTALL_LIBDIR}/cmake/composable_kernel
)
endif()
endif()
if(CK_DEVICE_CONTRACTION_INSTANCES)
add_library(device_contraction_operations STATIC ${CK_DEVICE_CONTRACTION_INSTANCES})
......
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