Skip to content

GitLab

Commit 9fb64dae authored by aska-0096's avatar aska-0096
Browse files

Merge branch 'develop' of... 

Merge branch 'develop' of https://github.com/ROCmSoftwarePlatform/composable_kernel into e2e_kernellib
parents e330961d fe96e8fb
Hide whitespace changes
Inline Side-by-side
Showing with 1142 additions and 143 deletions
library/src/tensor_operation_instance/gpu/quantization/device_conv2d_xdl_bias_perchannel_quantization_int8_instance.cpp → library/src/tensor_operation_instance/gpu/quantization/conv2d_fwd/device_conv2d_xdl_bias_perchannel_quantization_int8_instance.cpp
View file @ 9fb64dae
......@@ -7,7 +7,7 @@ namespace ck {
namespace tensor_operation {
namespace device {
namespace instance {
void add_device_conv2d_bias_perchannel_quantization_int8_instances(
void add_device_conv2d_xdl_bias_perchannel_quantization_int8_instances(
std::vector<std::unique_ptr<DeviceGroupedConvFwdMultipleD<NDimSpatial,
GNHWC,
GKYXC,
......@@ -22,23 +22,26 @@ void add_device_conv2d_bias_perchannel_quantization_int8_instances(
Add_Mul2_Clamp>>>& instances)
{
add_device_operation_instances(instances,
device_conv2d_int8_32Ds_instances<GK_GK_Tuple,
I32_F32_Tuple,
Add_Mul2_Clamp,
ConvFwdDefault>{});
device_grouped_conv2d_xdl_int8_instances<GK_GK_Tuple,
I32_F32_Tuple,
Add_Mul2_Clamp,
ConvFwdDefault,
8>{});
add_device_operation_instances(instances,
device_conv2d_int8_32Ds_instances<GK_GK_Tuple,
I32_F32_Tuple,
Add_Mul2_Clamp,
ConvFwd1x1P0>{});
device_grouped_conv2d_xdl_int8_instances<GK_GK_Tuple,
I32_F32_Tuple,
Add_Mul2_Clamp,
ConvFwd1x1P0,
8>{});
add_device_operation_instances(instances,
device_conv2d_int8_32Ds_instances<GK_GK_Tuple,
I32_F32_Tuple,
Add_Mul2_Clamp,
ConvFwd1x1S1P0>{});
device_grouped_conv2d_xdl_int8_instances<GK_GK_Tuple,
I32_F32_Tuple,
Add_Mul2_Clamp,
ConvFwd1x1S1P0,
8>{});
}
void add_device_conv2d_bias_relu_perchannel_quantization_int8_instances(
void add_device_conv2d_xdl_bias_relu_perchannel_quantization_int8_instances(
std::vector<std::unique_ptr<DeviceGroupedConvFwdMultipleD<NDimSpatial,
GNHWC,
GKYXC,
......@@ -53,20 +56,23 @@ void add_device_conv2d_bias_relu_perchannel_quantization_int8_instances(
Add_Relu_Mul2_Clamp>>>& instances)
{
add_device_operation_instances(instances,
device_conv2d_int8_32Ds_instances<GK_GK_Tuple,
I32_F32_Tuple,
Add_Relu_Mul2_Clamp,
ConvFwdDefault>{});
device_grouped_conv2d_xdl_int8_instances<GK_GK_Tuple,
I32_F32_Tuple,
Add_Relu_Mul2_Clamp,
ConvFwdDefault,
8>{});
add_device_operation_instances(instances,
device_conv2d_int8_32Ds_instances<GK_GK_Tuple,
I32_F32_Tuple,
Add_Relu_Mul2_Clamp,
ConvFwd1x1P0>{});
device_grouped_conv2d_xdl_int8_instances<GK_GK_Tuple,
I32_F32_Tuple,
Add_Relu_Mul2_Clamp,
ConvFwd1x1P0,
8>{});
add_device_operation_instances(instances,
device_conv2d_int8_32Ds_instances<GK_GK_Tuple,
I32_F32_Tuple,
Add_Relu_Mul2_Clamp,
ConvFwd1x1S1P0>{});
device_grouped_conv2d_xdl_int8_instances<GK_GK_Tuple,
I32_F32_Tuple,
Add_Relu_Mul2_Clamp,
ConvFwd1x1S1P0,
8>{});
}
} // namespace instance
} // namespace device
......
library/src/tensor_operation_instance/gpu/quantization/device_conv2d_xdl_bias_perlayer_quantization_int8_instance.cpp → library/src/tensor_operation_instance/gpu/quantization/conv2d_fwd/device_conv2d_xdl_bias_perlayer_quantization_int8_instance.cpp
View file @ 9fb64dae
......@@ -7,7 +7,7 @@ namespace ck {
namespace tensor_operation {
namespace device {
namespace instance {
void add_device_conv2d_bias_perlayer_quantization_int8_instances(
void add_device_conv2d_xdl_bias_perlayer_quantization_int8_instances(
std::vector<std::unique_ptr<DeviceGroupedConvFwdMultipleD<NDimSpatial,
GNHWC,
GKYXC,
......@@ -21,18 +21,27 @@ void add_device_conv2d_bias_perlayer_quantization_int8_instances(
PassThrough,
Add_Mul_Clamp>>>& instances)
{
add_device_operation_instances(
instances,
device_conv2d_int8_32Ds_instances<GK_Tuple, I32_Tuple, Add_Mul_Clamp, ConvFwdDefault>{});
add_device_operation_instances(
instances,
device_conv2d_int8_32Ds_instances<GK_Tuple, I32_Tuple, Add_Mul_Clamp, ConvFwd1x1P0>{});
add_device_operation_instances(
instances,
device_conv2d_int8_32Ds_instances<GK_Tuple, I32_Tuple, Add_Mul_Clamp, ConvFwd1x1S1P0>{});
add_device_operation_instances(instances,
device_grouped_conv2d_xdl_int8_instances<GK_Tuple,
I32_Tuple,
Add_Mul_Clamp,
ConvFwdDefault,
8>{});
add_device_operation_instances(instances,
device_grouped_conv2d_xdl_int8_instances<GK_Tuple,
I32_Tuple,
Add_Mul_Clamp,
ConvFwd1x1P0,
8>{});
add_device_operation_instances(instances,
device_grouped_conv2d_xdl_int8_instances<GK_Tuple,
I32_Tuple,
Add_Mul_Clamp,
ConvFwd1x1S1P0,
8>{});
}
void add_device_conv2d_bias_relu_perlayer_quantization_int8_instances(
void add_device_conv2d_xdl_bias_relu_perlayer_quantization_int8_instances(
std::vector<std::unique_ptr<DeviceGroupedConvFwdMultipleD<NDimSpatial,
GNHWC,
GKYXC,
......@@ -47,20 +56,25 @@ void add_device_conv2d_bias_relu_perlayer_quantization_int8_instances(
Add_Relu_Mul_Clamp>>>& instances)
{
add_device_operation_instances(instances,
device_conv2d_int8_32Ds_instances<GK_Tuple,
I32_Tuple,
Add_Relu_Mul_Clamp,
ConvFwdDefault>{});
device_grouped_conv2d_xdl_int8_instances<GK_Tuple,
I32_Tuple,
Add_Relu_Mul_Clamp,
ConvFwdDefault,
8>{});
add_device_operation_instances(
instances,
device_conv2d_int8_32Ds_instances<GK_Tuple, I32_Tuple, Add_Relu_Mul_Clamp, ConvFwd1x1P0>{});
add_device_operation_instances(instances,
device_grouped_conv2d_xdl_int8_instances<GK_Tuple,
I32_Tuple,
Add_Relu_Mul_Clamp,
ConvFwd1x1P0,
8>{});
add_device_operation_instances(instances,
device_conv2d_int8_32Ds_instances<GK_Tuple,
I32_Tuple,
Add_Relu_Mul_Clamp,
ConvFwd1x1S1P0>{});
device_grouped_conv2d_xdl_int8_instances<GK_Tuple,
I32_Tuple,
Add_Relu_Mul_Clamp,
ConvFwd1x1S1P0,
8>{});
}
} // namespace instance
} // namespace device
......
library/src/tensor_operation_instance/gpu/quantization/conv2d_fwd/device_conv2d_xdl_int8_instance.hpp 0 → 100644
View file @ 9fb64dae
// SPDX-License-Identifier: MIT
// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
#pragma once
#include "conv2d_quantization_common.hpp"
#include "ck/tensor_operation/gpu/device/impl/device_grouped_conv_fwd_multiple_d_xdl_cshuffle.hpp"
namespace ck {
namespace tensor_operation {
namespace device {
namespace instance {
// clang-format off
template <typename DsLayout,
typename DsDatatype,
typename OutElementOp,
ConvolutionForwardSpecialization ConvSpec,
index_t DstScalarPerVector>
using device_grouped_conv2d_xdl_int8_instances =
std::tuple <
//########################################| 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|
//########################################| | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | |
DeviceGroupedConvFwdMultipleD_Xdl_CShuffle<NDimSpatial, GNHWC, GKYXC, DsLayout, GNHWK, int8_t, int8_t, int32_t, int32_t, DsDatatype, int8_t, PassThrough, PassThrough, OutElementOp, ConvSpec, GemmSpec, 1, 256, 256, 128, 64, 16, 16, 32, 32, 4, 2, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 16, 16, 1, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 16, 16, 1, 1, 1, S<1, 64, 1, 4>, DstScalarPerVector>,
DeviceGroupedConvFwdMultipleD_Xdl_CShuffle<NDimSpatial, GNHWC, GKYXC, DsLayout, GNHWK, int8_t, int8_t, int32_t, int32_t, DsDatatype, int8_t, PassThrough, PassThrough, OutElementOp, ConvSpec, GemmSpec, 1, 256, 128, 256, 64, 16, 16, 32, 32, 2, 4, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 16, 16, 1, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 16, 16, 1, 1, 1, S<1, 64, 1, 4>, DstScalarPerVector>,
DeviceGroupedConvFwdMultipleD_Xdl_CShuffle<NDimSpatial, GNHWC, GKYXC, DsLayout, GNHWK, int8_t, int8_t, int32_t, int32_t, DsDatatype, int8_t, PassThrough, PassThrough, OutElementOp, ConvSpec, GemmSpec, 1, 128, 128, 128, 64, 16, 16, 32, 32, 4, 2, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 16, 16, 1, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 16, 16, 1, 1, 1, S<1, 32, 1, 4>, DstScalarPerVector>,
DeviceGroupedConvFwdMultipleD_Xdl_CShuffle<NDimSpatial, GNHWC, GKYXC, DsLayout, GNHWK, int8_t, int8_t, int32_t, int32_t, DsDatatype, int8_t, PassThrough, PassThrough, OutElementOp, ConvSpec, GemmSpec, 1, 256, 128, 128, 64, 16, 16, 32, 32, 2, 2, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 16, 16, 1, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 16, 16, 1, 1, 1, S<1, 64, 1, 4>, DstScalarPerVector>,
DeviceGroupedConvFwdMultipleD_Xdl_CShuffle<NDimSpatial, GNHWC, GKYXC, DsLayout, GNHWK, int8_t, int8_t, int32_t, int32_t, DsDatatype, int8_t, PassThrough, PassThrough, OutElementOp, ConvSpec, GemmSpec, 1, 128, 128, 64, 64, 16, 16, 32, 32, 2, 2, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 16, 16, 1, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 16, 16, 1, 1, 1, S<1, 64, 1, 2>, DstScalarPerVector>,
DeviceGroupedConvFwdMultipleD_Xdl_CShuffle<NDimSpatial, GNHWC, GKYXC, DsLayout, GNHWK, int8_t, int8_t, int32_t, int32_t, DsDatatype, int8_t, PassThrough, PassThrough, OutElementOp, ConvSpec, GemmSpec, 1, 128, 64, 128, 64, 16, 16, 32, 32, 2, 2, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 16, 16, 1, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 16, 16, 1, 1, 1, S<1, 32, 1, 4>, DstScalarPerVector>,
DeviceGroupedConvFwdMultipleD_Xdl_CShuffle<NDimSpatial, GNHWC, GKYXC, DsLayout, GNHWK, int8_t, int8_t, int32_t, int32_t, DsDatatype, int8_t, PassThrough, PassThrough, OutElementOp, ConvSpec, GemmSpec, 1, 64, 64, 64, 64, 16, 16, 32, 32, 2, 2, S<4, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 16, 16, 1, S<4, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 16, 16, 1, 1, 1, S<1, 32, 1, 2>, DstScalarPerVector>,
DeviceGroupedConvFwdMultipleD_Xdl_CShuffle<NDimSpatial, GNHWC, GKYXC, DsLayout, GNHWK, int8_t, int8_t, int32_t, int32_t, DsDatatype, int8_t, PassThrough, PassThrough, OutElementOp, ConvSpec, GemmSpec, 1, 256, 128, 64, 64, 16, 16, 32, 32, 2, 1, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 16, 16, 1, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 16, 16, 1, 1, 1, S<1, 64, 1, 4>, DstScalarPerVector>,
DeviceGroupedConvFwdMultipleD_Xdl_CShuffle<NDimSpatial, GNHWC, GKYXC, DsLayout, GNHWK, int8_t, int8_t, int32_t, int32_t, DsDatatype, int8_t, PassThrough, PassThrough, OutElementOp, ConvSpec, GemmSpec, 1, 256, 64, 128, 64, 16, 16, 32, 32, 1, 2, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 16, 16, 1, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 16, 16, 1, 1, 1, S<1, 64, 1, 4>, DstScalarPerVector>,
DeviceGroupedConvFwdMultipleD_Xdl_CShuffle<NDimSpatial, GNHWC, GKYXC, DsLayout, GNHWK, int8_t, int8_t, int32_t, int32_t, DsDatatype, int8_t, PassThrough, PassThrough, OutElementOp, ConvSpec, GemmSpec, 1, 128, 128, 32, 64, 16, 16, 32, 32, 2, 1, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 16, 16, 1, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 16, 16, 1, 1, 1, S<1, 64, 1, 2>, DstScalarPerVector>,
DeviceGroupedConvFwdMultipleD_Xdl_CShuffle<NDimSpatial, GNHWC, GKYXC, DsLayout, GNHWK, int8_t, int8_t, int32_t, int32_t, DsDatatype, int8_t, PassThrough, PassThrough, OutElementOp, ConvSpec, GemmSpec, 1, 128, 32, 128, 64, 16, 16, 32, 32, 1, 2, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 16, 16, 1, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 16, 16, 1, 1, 1, S<1, 32, 1, 4>, DstScalarPerVector>,
DeviceGroupedConvFwdMultipleD_Xdl_CShuffle<NDimSpatial, GNHWC, GKYXC, DsLayout, GNHWK, int8_t, int8_t, int32_t, int32_t, DsDatatype, int8_t, PassThrough, PassThrough, OutElementOp, ConvSpec, GemmSpec, 1, 64, 64, 32, 64, 16, 16, 32, 32, 2, 1, S<4, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 16, 16, 1, S<4, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 16, 16, 1, 1, 1, S<1, 32, 1, 2>, DstScalarPerVector>,
DeviceGroupedConvFwdMultipleD_Xdl_CShuffle<NDimSpatial, GNHWC, GKYXC, DsLayout, GNHWK, int8_t, int8_t, int32_t, int32_t, DsDatatype, int8_t, PassThrough, PassThrough, OutElementOp, ConvSpec, GemmSpec, 1, 64, 32, 64, 64, 16, 16, 32, 32, 1, 2, S<4, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 16, 16, 1, S<4, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 16, 16, 1, 1, 1, S<1, 32, 1, 2>, DstScalarPerVector>
>;
// clang-format on
} // namespace instance
} // namespace device
} // namespace tensor_operation
} // namespace ck
library/src/tensor_operation_instance/gpu/quantization/device_conv2d_xdl_perchannel_quantization_int8_instance.cpp → library/src/tensor_operation_instance/gpu/quantization/conv2d_fwd/device_conv2d_xdl_perchannel_quantization_int8_instance.cpp
View file @ 9fb64dae
......@@ -7,7 +7,7 @@ namespace ck {
namespace tensor_operation {
namespace device {
namespace instance {
void add_device_conv2d_perchannel_quantization_int8_instances(
void add_device_conv2d_xdl_perchannel_quantization_int8_instances(
std::vector<std::unique_ptr<DeviceGroupedConvFwdMultipleD<NDimSpatial,
GNHWC,
GKYXC,
......@@ -21,18 +21,27 @@ void add_device_conv2d_perchannel_quantization_int8_instances(
PassThrough,
Mul2_Clamp>>>& instances)
{
add_device_operation_instances(
instances,
device_conv2d_int8_32Ds_instances<GK_Tuple, F32_Tuple, Mul2_Clamp, ConvFwdDefault>{});
add_device_operation_instances(
instances,
device_conv2d_int8_32Ds_instances<GK_Tuple, F32_Tuple, Mul2_Clamp, ConvFwd1x1P0>{});
add_device_operation_instances(
instances,
device_conv2d_int8_32Ds_instances<GK_Tuple, F32_Tuple, Mul2_Clamp, ConvFwd1x1S1P0>{});
add_device_operation_instances(instances,
device_grouped_conv2d_xdl_int8_instances<GK_Tuple,
F32_Tuple,
Mul2_Clamp,
ConvFwdDefault,
8>{});
add_device_operation_instances(instances,
device_grouped_conv2d_xdl_int8_instances<GK_Tuple,
F32_Tuple,
Mul2_Clamp,
ConvFwd1x1P0,
8>{});
add_device_operation_instances(instances,
device_grouped_conv2d_xdl_int8_instances<GK_Tuple,
F32_Tuple,
Mul2_Clamp,
ConvFwd1x1S1P0,
8>{});
}
void add_device_conv2d_relu_perchannel_quantization_int8_instances(
void add_device_conv2d_xdl_relu_perchannel_quantization_int8_instances(
std::vector<std::unique_ptr<DeviceGroupedConvFwdMultipleD<NDimSpatial,
GNHWC,
GKYXC,
......@@ -46,15 +55,24 @@ void add_device_conv2d_relu_perchannel_quantization_int8_instances(
PassThrough,
Relu_Mul2_Clamp>>>& instances)
{
add_device_operation_instances(
instances,
device_conv2d_int8_32Ds_instances<GK_Tuple, F32_Tuple, Relu_Mul2_Clamp, ConvFwdDefault>{});
add_device_operation_instances(
instances,
device_conv2d_int8_32Ds_instances<GK_Tuple, F32_Tuple, Relu_Mul2_Clamp, ConvFwd1x1P0>{});
add_device_operation_instances(
instances,
device_conv2d_int8_32Ds_instances<GK_Tuple, F32_Tuple, Relu_Mul2_Clamp, ConvFwd1x1S1P0>{});
add_device_operation_instances(instances,
device_grouped_conv2d_xdl_int8_instances<GK_Tuple,
F32_Tuple,
Relu_Mul2_Clamp,
ConvFwdDefault,
8>{});
add_device_operation_instances(instances,
device_grouped_conv2d_xdl_int8_instances<GK_Tuple,
F32_Tuple,
Relu_Mul2_Clamp,
ConvFwd1x1P0,
8>{});
add_device_operation_instances(instances,
device_grouped_conv2d_xdl_int8_instances<GK_Tuple,
F32_Tuple,
Relu_Mul2_Clamp,
ConvFwd1x1S1P0,
8>{});
}
} // namespace instance
} // namespace device
......
library/src/tensor_operation_instance/gpu/quantization/device_conv2d_xdl_perlayer_quantization_int8_instance.cpp → library/src/tensor_operation_instance/gpu/quantization/conv2d_fwd/device_conv2d_xdl_perlayer_quantization_int8_instance.cpp
View file @ 9fb64dae
......@@ -7,7 +7,7 @@ namespace ck {
namespace tensor_operation {
namespace device {
namespace instance {
void add_device_conv2d_perlayer_quantization_int8_instances(
void add_device_conv2d_xdl_perlayer_quantization_int8_instances(
std::vector<std::unique_ptr<DeviceGroupedConvFwdMultipleD<NDimSpatial,
GNHWC,
GKYXC,
......@@ -21,18 +21,27 @@ void add_device_conv2d_perlayer_quantization_int8_instances(
PassThrough,
Mul_Clamp>>>& instances)
{
add_device_operation_instances(
instances,
device_conv2d_int8_instances<Empty_Tuple, Empty_Tuple, Mul_Clamp, ConvFwdDefault>{});
add_device_operation_instances(
instances,
device_conv2d_int8_instances<Empty_Tuple, Empty_Tuple, Mul_Clamp, ConvFwd1x1P0>{});
add_device_operation_instances(
instances,
device_conv2d_int8_instances<Empty_Tuple, Empty_Tuple, Mul_Clamp, ConvFwd1x1S1P0>{});
add_device_operation_instances(instances,
device_grouped_conv2d_xdl_int8_instances<Empty_Tuple,
Empty_Tuple,
Mul_Clamp,
ConvFwdDefault,
16>{});
add_device_operation_instances(instances,
device_grouped_conv2d_xdl_int8_instances<Empty_Tuple,
Empty_Tuple,
Mul_Clamp,
ConvFwd1x1P0,
16>{});
add_device_operation_instances(instances,
device_grouped_conv2d_xdl_int8_instances<Empty_Tuple,
Empty_Tuple,
Mul_Clamp,
ConvFwd1x1S1P0,
16>{});
}
void add_device_conv2d_relu_perlayer_quantization_int8_instances(
void add_device_conv2d_xdl_relu_perlayer_quantization_int8_instances(
std::vector<std::unique_ptr<DeviceGroupedConvFwdMultipleD<NDimSpatial,
GNHWC,
GKYXC,
......@@ -46,15 +55,24 @@ void add_device_conv2d_relu_perlayer_quantization_int8_instances(
PassThrough,
Relu_Mul_Clamp>>>& instances)
{
add_device_operation_instances(
instances,
device_conv2d_int8_instances<Empty_Tuple, Empty_Tuple, Relu_Mul_Clamp, ConvFwdDefault>{});
add_device_operation_instances(
instances,
device_conv2d_int8_instances<Empty_Tuple, Empty_Tuple, Relu_Mul_Clamp, ConvFwd1x1P0>{});
add_device_operation_instances(
instances,
device_conv2d_int8_instances<Empty_Tuple, Empty_Tuple, Relu_Mul_Clamp, ConvFwd1x1S1P0>{});
add_device_operation_instances(instances,
device_grouped_conv2d_xdl_int8_instances<Empty_Tuple,
Empty_Tuple,
Relu_Mul_Clamp,
ConvFwdDefault,
16>{});
add_device_operation_instances(instances,
device_grouped_conv2d_xdl_int8_instances<Empty_Tuple,
Empty_Tuple,
Relu_Mul_Clamp,
ConvFwd1x1P0,
16>{});
add_device_operation_instances(instances,
device_grouped_conv2d_xdl_int8_instances<Empty_Tuple,
Empty_Tuple,
Relu_Mul_Clamp,
ConvFwd1x1S1P0,
16>{});
}
} // namespace instance
} // namespace device
......
library/src/tensor_operation_instance/gpu/quantization/device_conv2d_xdl_int8_instance.hpp deleted 100644 → 0
View file @ e330961d
// SPDX-License-Identifier: MIT
// Copyright (c) 2018-2022, 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_cshuffle.hpp"
#include "ck/tensor_operation/gpu/element/element_wise_operation.hpp"
#include "ck/library/tensor_operation_instance/add_device_operation_instance.hpp"
namespace ck {
namespace tensor_operation {
namespace device {
namespace instance {
using Empty_Tuple = ck::Tuple<>;
template <ck::index_t... Is>
using S = ck::Sequence<Is...>;
using GNHWC = ck::tensor_layout::convolution::GNHWC;
using GKYXC = ck::tensor_layout::convolution::GKYXC;
using GNHWK = ck::tensor_layout::convolution::GNHWK;
using GK = ck::tensor_layout::convolution::G_K;
using PassThrough = ck::tensor_operation::element_wise::PassThrough;
using Relu = ck::tensor_operation::element_wise::Relu;
using GK_Tuple = ck::Tuple<GK>;
using GK_GK_Tuple = ck::Tuple<GK, GK>;
using I32_Tuple = ck::Tuple<int32_t>;
using F32_Tuple = ck::Tuple<float>;
using I32_F32_Tuple = ck::Tuple<int32_t, float>;
using Mul_Clamp = ck::tensor_operation::element_wise::Activation_Mul_Clamp<PassThrough>;
using Relu_Mul_Clamp = ck::tensor_operation::element_wise::Activation_Mul_Clamp<Relu>;
using Add_Mul_Clamp = ck::tensor_operation::element_wise::Add_Activation_Mul_Clamp<PassThrough>;
using Add_Relu_Mul_Clamp = ck::tensor_operation::element_wise::Add_Activation_Mul_Clamp<Relu>;
using Mul2_Clamp = ck::tensor_operation::element_wise::Activation_Mul2_Clamp<PassThrough>;
using Relu_Mul2_Clamp = ck::tensor_operation::element_wise::Activation_Mul2_Clamp<Relu>;
using Add_Mul2_Clamp = ck::tensor_operation::element_wise::Add_Activation_Mul2_Clamp<PassThrough>;
using Add_Relu_Mul2_Clamp = ck::tensor_operation::element_wise::Add_Activation_Mul2_Clamp<Relu>;
static constexpr ck::index_t NDimSpatial = 2;
static constexpr auto GemmSpec = ck::tensor_operation::device::GemmSpecialization::MNKPadding;
static constexpr auto ConvFwdDefault =
ck::tensor_operation::device::ConvolutionForwardSpecialization::Default;
static constexpr auto ConvFwd1x1P0 =
ck::tensor_operation::device::ConvolutionForwardSpecialization::Filter1x1Pad0;
static constexpr auto ConvFwd1x1S1P0 =
ck::tensor_operation::device::ConvolutionForwardSpecialization::Filter1x1Stride1Pad0;
template <typename DsLayout,
typename DsDatatype,
typename OutElementOp,
ConvolutionForwardSpecialization ConvSpec>
// clang-format off
using device_conv2d_int8_instances =
std::tuple <
//########################################| 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|
//########################################| | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | |
DeviceGroupedConvFwdMultipleD_Xdl_CShuffle< 2, GNHWC, GKYXC, DsLayout, GNHWK, int8_t, int8_t, int32_t, int32_t, DsDatatype, int8_t, PassThrough, PassThrough, OutElementOp, ConvSpec, GemmSpec, 1, 256, 256, 128, 64, 16, 16, 32, 32, 4, 2, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 16, 16, 1, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 16, 16, 1, 1, 1, S<1, 64, 1, 4>, 16>,
DeviceGroupedConvFwdMultipleD_Xdl_CShuffle< 2, GNHWC, GKYXC, DsLayout, GNHWK, int8_t, int8_t, int32_t, int32_t, DsDatatype, int8_t, PassThrough, PassThrough, OutElementOp, ConvSpec, GemmSpec, 1, 256, 128, 256, 64, 16, 16, 32, 32, 2, 4, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 16, 16, 1, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 16, 16, 1, 1, 1, S<1, 64, 1, 4>, 16>,
DeviceGroupedConvFwdMultipleD_Xdl_CShuffle< 2, GNHWC, GKYXC, DsLayout, GNHWK, int8_t, int8_t, int32_t, int32_t, DsDatatype, int8_t, PassThrough, PassThrough, OutElementOp, ConvSpec, GemmSpec, 1, 128, 128, 128, 64, 16, 16, 32, 32, 4, 2, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 16, 16, 1, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 16, 16, 1, 1, 1, S<1, 32, 1, 4>, 16>,
DeviceGroupedConvFwdMultipleD_Xdl_CShuffle< 2, GNHWC, GKYXC, DsLayout, GNHWK, int8_t, int8_t, int32_t, int32_t, DsDatatype, int8_t, PassThrough, PassThrough, OutElementOp, ConvSpec, GemmSpec, 1, 256, 128, 128, 64, 16, 16, 32, 32, 2, 2, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 16, 16, 1, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 16, 16, 1, 1, 1, S<1, 64, 1, 4>, 16>,
DeviceGroupedConvFwdMultipleD_Xdl_CShuffle< 2, GNHWC, GKYXC, DsLayout, GNHWK, int8_t, int8_t, int32_t, int32_t, DsDatatype, int8_t, PassThrough, PassThrough, OutElementOp, ConvSpec, GemmSpec, 1, 128, 128, 64, 64, 16, 16, 32, 32, 2, 2, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 16, 16, 1, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 16, 16, 1, 1, 1, S<1, 64, 1, 2>, 16>,
DeviceGroupedConvFwdMultipleD_Xdl_CShuffle< 2, GNHWC, GKYXC, DsLayout, GNHWK, int8_t, int8_t, int32_t, int32_t, DsDatatype, int8_t, PassThrough, PassThrough, OutElementOp, ConvSpec, GemmSpec, 1, 128, 64, 128, 64, 16, 16, 32, 32, 2, 2, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 16, 16, 1, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 16, 16, 1, 1, 1, S<1, 32, 1, 4>, 16>,
DeviceGroupedConvFwdMultipleD_Xdl_CShuffle< 2, GNHWC, GKYXC, DsLayout, GNHWK, int8_t, int8_t, int32_t, int32_t, DsDatatype, int8_t, PassThrough, PassThrough, OutElementOp, ConvSpec, GemmSpec, 1, 64, 64, 64, 64, 16, 16, 32, 32, 2, 2, S<4, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 16, 16, 1, S<4, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 16, 16, 1, 1, 1, S<1, 32, 1, 2>, 16>,
DeviceGroupedConvFwdMultipleD_Xdl_CShuffle< 2, GNHWC, GKYXC, DsLayout, GNHWK, int8_t, int8_t, int32_t, int32_t, DsDatatype, int8_t, PassThrough, PassThrough, OutElementOp, ConvSpec, GemmSpec, 1, 256, 128, 64, 64, 16, 16, 32, 32, 2, 1, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 16, 16, 1, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 16, 16, 1, 1, 1, S<1, 64, 1, 4>, 16>,
DeviceGroupedConvFwdMultipleD_Xdl_CShuffle< 2, GNHWC, GKYXC, DsLayout, GNHWK, int8_t, int8_t, int32_t, int32_t, DsDatatype, int8_t, PassThrough, PassThrough, OutElementOp, ConvSpec, GemmSpec, 1, 256, 64, 128, 64, 16, 16, 32, 32, 1, 2, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 16, 16, 1, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 16, 16, 1, 1, 1, S<1, 64, 1, 4>, 16>,
DeviceGroupedConvFwdMultipleD_Xdl_CShuffle< 2, GNHWC, GKYXC, DsLayout, GNHWK, int8_t, int8_t, int32_t, int32_t, DsDatatype, int8_t, PassThrough, PassThrough, OutElementOp, ConvSpec, GemmSpec, 1, 128, 128, 32, 64, 16, 16, 32, 32, 2, 1, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 16, 16, 1, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 16, 16, 1, 1, 1, S<1, 64, 1, 2>, 16>,
DeviceGroupedConvFwdMultipleD_Xdl_CShuffle< 2, GNHWC, GKYXC, DsLayout, GNHWK, int8_t, int8_t, int32_t, int32_t, DsDatatype, int8_t, PassThrough, PassThrough, OutElementOp, ConvSpec, GemmSpec, 1, 128, 32, 128, 64, 16, 16, 32, 32, 1, 2, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 16, 16, 1, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 16, 16, 1, 1, 1, S<1, 32, 1, 4>, 16>,
DeviceGroupedConvFwdMultipleD_Xdl_CShuffle< 2, GNHWC, GKYXC, DsLayout, GNHWK, int8_t, int8_t, int32_t, int32_t, DsDatatype, int8_t, PassThrough, PassThrough, OutElementOp, ConvSpec, GemmSpec, 1, 64, 64, 32, 64, 16, 16, 32, 32, 2, 1, S<4, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 16, 16, 1, S<4, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 16, 16, 1, 1, 1, S<1, 32, 1, 2>, 16>,
DeviceGroupedConvFwdMultipleD_Xdl_CShuffle< 2, GNHWC, GKYXC, DsLayout, GNHWK, int8_t, int8_t, int32_t, int32_t, DsDatatype, int8_t, PassThrough, PassThrough, OutElementOp, ConvSpec, GemmSpec, 1, 64, 32, 64, 64, 16, 16, 32, 32, 1, 2, S<4, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 16, 16, 1, S<4, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 16, 16, 1, 1, 1, S<1, 32, 1, 2>, 16>
>;
// clang-format on
// for conv + multiple of 32 bit Ds. bit of Ds will affect the ScalarPerVector of C
template <typename DsLayout,
typename DsDatatype,
typename OutElementOp,
ConvolutionForwardSpecialization ConvSpec>
// clang-format off
using device_conv2d_int8_32Ds_instances =
std::tuple <
//########################################| 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|
//########################################| | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | |
DeviceGroupedConvFwdMultipleD_Xdl_CShuffle< 2, GNHWC, GKYXC, DsLayout, GNHWK, int8_t, int8_t, int32_t, int32_t, DsDatatype, int8_t, PassThrough, PassThrough, OutElementOp, ConvSpec, GemmSpec, 1, 256, 256, 128, 64, 16, 16, 32, 32, 4, 2, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 16, 16, 1, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 16, 16, 1, 1, 1, S<1, 64, 1, 4>, 8>,
DeviceGroupedConvFwdMultipleD_Xdl_CShuffle< 2, GNHWC, GKYXC, DsLayout, GNHWK, int8_t, int8_t, int32_t, int32_t, DsDatatype, int8_t, PassThrough, PassThrough, OutElementOp, ConvSpec, GemmSpec, 1, 256, 128, 256, 64, 16, 16, 32, 32, 2, 4, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 16, 16, 1, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 16, 16, 1, 1, 1, S<1, 64, 1, 4>, 8>,
DeviceGroupedConvFwdMultipleD_Xdl_CShuffle< 2, GNHWC, GKYXC, DsLayout, GNHWK, int8_t, int8_t, int32_t, int32_t, DsDatatype, int8_t, PassThrough, PassThrough, OutElementOp, ConvSpec, GemmSpec, 1, 128, 128, 128, 64, 16, 16, 32, 32, 4, 2, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 16, 16, 1, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 16, 16, 1, 1, 1, S<1, 32, 1, 4>, 8>,
DeviceGroupedConvFwdMultipleD_Xdl_CShuffle< 2, GNHWC, GKYXC, DsLayout, GNHWK, int8_t, int8_t, int32_t, int32_t, DsDatatype, int8_t, PassThrough, PassThrough, OutElementOp, ConvSpec, GemmSpec, 1, 256, 128, 128, 64, 16, 16, 32, 32, 2, 2, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 16, 16, 1, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 16, 16, 1, 1, 1, S<1, 64, 1, 4>, 8>,
DeviceGroupedConvFwdMultipleD_Xdl_CShuffle< 2, GNHWC, GKYXC, DsLayout, GNHWK, int8_t, int8_t, int32_t, int32_t, DsDatatype, int8_t, PassThrough, PassThrough, OutElementOp, ConvSpec, GemmSpec, 1, 128, 128, 64, 64, 16, 16, 32, 32, 2, 2, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 16, 16, 1, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 16, 16, 1, 1, 1, S<1, 64, 1, 2>, 8>,
DeviceGroupedConvFwdMultipleD_Xdl_CShuffle< 2, GNHWC, GKYXC, DsLayout, GNHWK, int8_t, int8_t, int32_t, int32_t, DsDatatype, int8_t, PassThrough, PassThrough, OutElementOp, ConvSpec, GemmSpec, 1, 128, 64, 128, 64, 16, 16, 32, 32, 2, 2, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 16, 16, 1, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 16, 16, 1, 1, 1, S<1, 32, 1, 4>, 8>,
DeviceGroupedConvFwdMultipleD_Xdl_CShuffle< 2, GNHWC, GKYXC, DsLayout, GNHWK, int8_t, int8_t, int32_t, int32_t, DsDatatype, int8_t, PassThrough, PassThrough, OutElementOp, ConvSpec, GemmSpec, 1, 64, 64, 64, 64, 16, 16, 32, 32, 2, 2, S<4, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 16, 16, 1, S<4, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 16, 16, 1, 1, 1, S<1, 32, 1, 2>, 8>,
DeviceGroupedConvFwdMultipleD_Xdl_CShuffle< 2, GNHWC, GKYXC, DsLayout, GNHWK, int8_t, int8_t, int32_t, int32_t, DsDatatype, int8_t, PassThrough, PassThrough, OutElementOp, ConvSpec, GemmSpec, 1, 256, 128, 64, 64, 16, 16, 32, 32, 2, 1, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 16, 16, 1, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 16, 16, 1, 1, 1, S<1, 64, 1, 4>, 8>,
DeviceGroupedConvFwdMultipleD_Xdl_CShuffle< 2, GNHWC, GKYXC, DsLayout, GNHWK, int8_t, int8_t, int32_t, int32_t, DsDatatype, int8_t, PassThrough, PassThrough, OutElementOp, ConvSpec, GemmSpec, 1, 256, 64, 128, 64, 16, 16, 32, 32, 1, 2, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 16, 16, 1, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 16, 16, 1, 1, 1, S<1, 64, 1, 4>, 8>,
DeviceGroupedConvFwdMultipleD_Xdl_CShuffle< 2, GNHWC, GKYXC, DsLayout, GNHWK, int8_t, int8_t, int32_t, int32_t, DsDatatype, int8_t, PassThrough, PassThrough, OutElementOp, ConvSpec, GemmSpec, 1, 128, 128, 32, 64, 16, 16, 32, 32, 2, 1, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 16, 16, 1, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 16, 16, 1, 1, 1, S<1, 64, 1, 2>, 8>,
DeviceGroupedConvFwdMultipleD_Xdl_CShuffle< 2, GNHWC, GKYXC, DsLayout, GNHWK, int8_t, int8_t, int32_t, int32_t, DsDatatype, int8_t, PassThrough, PassThrough, OutElementOp, ConvSpec, GemmSpec, 1, 128, 32, 128, 64, 16, 16, 32, 32, 1, 2, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 16, 16, 1, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 16, 16, 1, 1, 1, S<1, 32, 1, 4>, 8>,
DeviceGroupedConvFwdMultipleD_Xdl_CShuffle< 2, GNHWC, GKYXC, DsLayout, GNHWK, int8_t, int8_t, int32_t, int32_t, DsDatatype, int8_t, PassThrough, PassThrough, OutElementOp, ConvSpec, GemmSpec, 1, 64, 64, 32, 64, 16, 16, 32, 32, 2, 1, S<4, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 16, 16, 1, S<4, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 16, 16, 1, 1, 1, S<1, 32, 1, 2>, 8>,
DeviceGroupedConvFwdMultipleD_Xdl_CShuffle< 2, GNHWC, GKYXC, DsLayout, GNHWK, int8_t, int8_t, int32_t, int32_t, DsDatatype, int8_t, PassThrough, PassThrough, OutElementOp, ConvSpec, GemmSpec, 1, 64, 32, 64, 64, 16, 16, 32, 32, 1, 2, S<4, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 16, 16, 1, S<4, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 16, 16, 1, 1, 1, S<1, 32, 1, 2>, 8>
>;
// clang-format on
} // namespace instance
} // namespace device
} // namespace tensor_operation
} // namespace ck
library/src/tensor_operation_instance/gpu/quantization/gemm/device_gemm_quantization_dl_c_shuffle_i8_i8_i8_instance.hpp 0 → 100644
View file @ 9fb64dae
// SPDX-License-Identifier: MIT
// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
#include "gemm_quantization_common.hpp"
#include "ck/tensor_operation/gpu/device/impl/device_gemm_multiple_d_dl.hpp"
namespace ck {
namespace tensor_operation {
namespace device {
namespace instance {
template <typename OutElementOp>
using device_gemm_quantization_dl_c_shuffle_i8_i8_i8_km_kn_mn_instances = std::tuple<
// clang-format off
//####################| A| B| Ds| E| AData| BData| AccData| DsData| EData| A| B| CDE| GEMM| Block| MPer| NPer| KPer| K1| M1Per| N1Per| KPer| M11N11Thread| M11N11Thread| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| BBlockTransfer| BBlockTransfer| BBlockTransfer| BBlockTransfer| BBlockTransfer| BBlockTransfer| BBlockTransfer| CThreadTransfer| CThreadTransfer| CThreadTransfer|
//####################| Layout| Layout| Layout| Layout| Type| Type| Type| Type| Type| Elementwise| Elementwise| Elementwise| Specialization| Size| Block| Block| Block| | ThreadM| Thread| Thread| ClusterM110Xs| ClusterN110Xs| ThreadSliceLengths| ThreadClusterLengths| ThreadCluster| SrcAccess| SrcVectorTensor| SrcVectorTensor| DstVectorTensor| ThreadSliceLengths| ThreadClusterLengths| ThreadCluster| SrcAccess| SrcVectorTensor| SrcVectorTensor| DstVectorTensor| SrcDstAccess| SrcDstVectorDim| DstScalarPerVector|
//####################| | | | | | | | | | Operation| Operation| Operation| | | | | | | | | | | | K0_M0_M1_K1| K0_M0_M1_K1| ArrangeOrder| Order| Lengths_K0_M0_M1_K1| ContiguousDimOrder| Lengths_K0_M0_M1_K1| K0_N0_N1_K1| K0_N0_N1_K1| ArrangeOrder| Order| Lengths_K0_N0_N1_K1| ContiguousDimOrder| Lengths_K0_N0_N1_K1| Order| | |
//####################| | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | |
DeviceGemmMultipleD_Dl< Col, Row, Empty_Tuple, Row, int8_t, int8_t, int32_t, Empty_Tuple, int8_t, PassThrough, PassThrough, OutElementOp, MNKPadding, 256, 128, 128, 16, 4, 4, 4, 1, S<8, 2>, S<8, 2>, S<2, 1, 4, 4>, S<8, 1, 32, 1>, S<0, 3, 1, 2>, S<0, 3, 1, 2>, S<1, 1, 4, 1>, S<0, 3, 1, 2>, S<1, 1, 4, 4>, S<2, 1, 4, 4>, S<8, 1, 32, 1>, S<0, 3, 1, 2>, S<0, 3, 1, 2>, S<1, 1, 4, 1>, S<0, 3, 1, 2>, S<1, 1, 4, 4>, S<0, 1, 2, 3, 4, 5>, 5, 4>
// clang-format on
>;
template <typename OutElementOp>
using device_gemm_quantization_dl_c_shuffle_i8_i8_i8_km_nk_mn_instances = std::tuple<
// clang-format off
//####################| A| B| Ds| E| AData| BData| AccData| DsData| EData| A| B| CDE| GEMM| Block| MPer| NPer| KPer| K1| M1Per| N1Per| KPer| M11N11Thread| M11N11Thread| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| BBlockTransfer| BBlockTransfer| BBlockTransfer| BBlockTransfer| BBlockTransfer| BBlockTransfer| BBlockTransfer| CThreadTransfer| CThreadTransfer| CThreadTransfer|
//####################| Layout| Layout| Layout| Layout| Type| Type| Type| Type| Type| Elementwise| Elementwise| Elementwise| Specialization| Size| Block| Block| Block| | ThreadM| Thread| Thread| ClusterM110Xs| ClusterN110Xs| ThreadSliceLengths| ThreadClusterLengths| ThreadCluster| SrcAccess| SrcVectorTensor| SrcVectorTensor| DstVectorTensor| ThreadSliceLengths| ThreadClusterLengths| ThreadCluster| SrcAccess| SrcVectorTensor| SrcVectorTensor| DstVectorTensor| SrcDstAccess| SrcDstVectorDim| DstScalarPerVector|
//####################| | | | | | | | | | Operation| Operation| Operation| | | | | | | | | | | | K0_M0_M1_K1| K0_M0_M1_K1| ArrangeOrder| Order| Lengths_K0_M0_M1_K1| ContiguousDimOrder| Lengths_K0_M0_M1_K1| K0_N0_N1_K1| K0_N0_N1_K1| ArrangeOrder| Order| Lengths_K0_N0_N1_K1| ContiguousDimOrder| Lengths_K0_N0_N1_K1| Order| | |
//####################| | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | |
DeviceGemmMultipleD_Dl< Col, Col, Empty_Tuple, Row, int8_t, int8_t, int32_t, Empty_Tuple, int8_t, PassThrough, PassThrough, OutElementOp, MNKPadding, 256, 128, 128, 16, 4, 4, 4, 1, S<8, 2>, S<8, 2>, S<2, 1, 4, 4>, S<8, 1, 32, 1>, S<0, 3, 1, 2>, S<0, 3, 1, 2>, S<1, 1, 4, 1>, S<0, 3, 1, 2>, S<1, 1, 4, 4>, S<8, 1, 1, 4>, S<2, 1, 128, 1>, S<1, 2, 0, 3>, S<1, 2, 0, 3>, S<4, 1, 1, 4>, S<1, 2, 0, 3>, S<1, 1, 1, 4>, S<0, 1, 2, 3, 4, 5>, 5, 4>
// clang-format on
>;
template <typename OutElementOp>
using device_gemm_quantization_dl_c_shuffle_i8_i8_i8_mk_kn_mn_instances = std::tuple<
// clang-format off
//####################| A| B| Ds| E| AData| BData| AccData| DsData| EData| A| B| CDE| GEMM| Block| MPer| NPer| KPer| K1| M1Per| N1Per| KPer| M11N11Thread| M11N11Thread| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| BBlockTransfer| BBlockTransfer| BBlockTransfer| BBlockTransfer| BBlockTransfer| BBlockTransfer| BBlockTransfer| CThreadTransfer| CThreadTransfer| CThreadTransfer|
//####################| Layout| Layout| Layout| Layout| Type| Type| Type| Type| Type| Elementwise| Elementwise| Elementwise| Specialization| Size| Block| Block| Block| | ThreadM| Thread| Thread| ClusterM110Xs| ClusterN110Xs| ThreadSliceLengths| ThreadClusterLengths| ThreadCluster| SrcAccess| SrcVectorTensor| SrcVectorTensor| DstVectorTensor| ThreadSliceLengths| ThreadClusterLengths| ThreadCluster| SrcAccess| SrcVectorTensor| SrcVectorTensor| DstVectorTensor| SrcDstAccess| SrcDstVectorDim| DstScalarPerVector|
//####################| | | | | | | | | | Operation| Operation| Operation| | | | | | | | | | | | K0_M0_M1_K1| K0_M0_M1_K1| ArrangeOrder| Order| Lengths_K0_M0_M1_K1| ContiguousDimOrder| Lengths_K0_M0_M1_K1| K0_N0_N1_K1| K0_N0_N1_K1| ArrangeOrder| Order| Lengths_K0_N0_N1_K1| ContiguousDimOrder| Lengths_K0_N0_N1_K1| Order| | |
//####################| | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | |
DeviceGemmMultipleD_Dl< Row, Row, Empty_Tuple, Row, int8_t, int8_t, int32_t, Empty_Tuple, int8_t, PassThrough, PassThrough, OutElementOp, MNKPadding, 256, 128, 128, 16, 4, 4, 4, 1, S<8, 2>, S<8, 2>, S<8, 1, 1, 4>, S<2, 1, 128, 1>, S<1, 2, 0, 3>, S<1, 2, 0, 3>, S<4, 1, 1, 4>, S<1, 2, 0, 3>, S<1, 1, 1, 4>, S<2, 1, 4, 4>, S<8, 1, 32, 1>, S<0, 3, 1, 2>, S<0, 3, 1, 2>, S<1, 1, 4, 1>, S<0, 3, 1, 2>, S<1, 1, 4, 4>, S<0, 1, 2, 3, 4, 5>, 5, 4>
// clang-format on
>;
template <typename OutElementOp>
using device_gemm_quantization_dl_c_shuffle_i8_i8_i8_mk_nk_mn_instances = std::tuple<
// clang-format off
//####################| A| B| Ds| E| AData| BData| AccData| DsData| EData| A| B| CDE| GEMM| Block| MPer| NPer| KPer| K1| M1Per| N1Per| KPer| M11N11Thread| M11N11Thread| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| BBlockTransfer| BBlockTransfer| BBlockTransfer| BBlockTransfer| BBlockTransfer| BBlockTransfer| BBlockTransfer| CThreadTransfer| CThreadTransfer| CThreadTransfer|
//####################| Layout| Layout| Layout| Layout| Type| Type| Type| Type| Type| Elementwise| Elementwise| Elementwise| Specialization| Size| Block| Block| Block| | ThreadM| Thread| Thread| ClusterM110Xs| ClusterN110Xs| ThreadSliceLengths| ThreadClusterLengths| ThreadCluster| SrcAccess| SrcVectorTensor| SrcVectorTensor| DstVectorTensor| ThreadSliceLengths| ThreadClusterLengths| ThreadCluster| SrcAccess| SrcVectorTensor| SrcVectorTensor| DstVectorTensor| SrcDstAccess| SrcDstVectorDim| DstScalarPerVector|
//####################| | | | | | | | | | Operation| Operation| Operation| | | | | | | | | | | | K0_M0_M1_K1| K0_M0_M1_K1| ArrangeOrder| Order| Lengths_K0_M0_M1_K1| ContiguousDimOrder| Lengths_K0_M0_M1_K1| K0_N0_N1_K1| K0_N0_N1_K1| ArrangeOrder| Order| Lengths_K0_N0_N1_K1| ContiguousDimOrder| Lengths_K0_N0_N1_K1| Order| | |
//####################| | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | |
DeviceGemmMultipleD_Dl< Row, Col, Empty_Tuple, Row, int8_t, int8_t, int32_t, Empty_Tuple, int8_t, PassThrough, PassThrough, OutElementOp, MNKPadding, 256, 128, 128, 16, 4, 4, 4, 1, S<8, 2>, S<8, 2>, S<8, 1, 1, 4>, S<2, 1, 128, 1>, S<1, 2, 0, 3>, S<1, 2, 0, 3>, S<4, 1, 1, 4>, S<1, 2, 0, 3>, S<1, 1, 1, 4>, S<8, 1, 1, 4>, S<2, 1, 128, 1>, S<1, 2, 0, 3>, S<1, 2, 0, 3>, S<4, 1, 1, 4>, S<1, 2, 0, 3>, S<1, 1, 1, 4>, S<0, 1, 2, 3, 4, 5>, 5, 4>
// clang-format on
>;
} // namespace instance
} // namespace device
} // namespace tensor_operation
} // namespace ck
library/src/tensor_operation_instance/gpu/quantization/gemm/device_gemm_quantization_dl_c_shuffle_i8_i8_i8_km_kn_mn_instance.cpp 0 → 100644
View file @ 9fb64dae
// SPDX-License-Identifier: MIT
// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
#include "device_gemm_quantization_dl_c_shuffle_i8_i8_i8_instance.hpp"
namespace ck {
namespace tensor_operation {
namespace device {
namespace instance {
// Layout(A, B, C) = [Col, Row, Row]
void add_device_gemm_quantization_dl_c_shuffle_i8_i8_i8_km_kn_mn_instances(
std::vector<std::unique_ptr<DeviceGemmMultipleD<Col,
Row,
Empty_Tuple,
Row,
int8_t,
int8_t,
Empty_Tuple,
int8_t,
PassThrough,
PassThrough,
Mul_Clamp>>>& instances)
{
add_device_operation_instances(
instances, device_gemm_quantization_dl_c_shuffle_i8_i8_i8_km_kn_mn_instances<Mul_Clamp>{});
}
} // namespace instance
} // namespace device
} // namespace tensor_operation
} // namespace ck
library/src/tensor_operation_instance/gpu/quantization/gemm/device_gemm_quantization_dl_c_shuffle_i8_i8_i8_km_nk_mn_instance.cpp 0 → 100644
View file @ 9fb64dae
// SPDX-License-Identifier: MIT
// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
#include "device_gemm_quantization_dl_c_shuffle_i8_i8_i8_instance.hpp"
namespace ck {
namespace tensor_operation {
namespace device {
namespace instance {
// Layout(A, B, C) = [Col, Col, Row]
void add_device_gemm_quantization_dl_c_shuffle_i8_i8_i8_km_nk_mn_instances(
std::vector<std::unique_ptr<DeviceGemmMultipleD<Col,
Col,
Empty_Tuple,
Row,
int8_t,
int8_t,
Empty_Tuple,
int8_t,
PassThrough,
PassThrough,
Mul_Clamp>>>& instances)
{
add_device_operation_instances(
instances, device_gemm_quantization_dl_c_shuffle_i8_i8_i8_km_nk_mn_instances<Mul_Clamp>{});
}
} // namespace instance
} // namespace device
} // namespace tensor_operation
} // namespace ck
library/src/tensor_operation_instance/gpu/quantization/gemm/device_gemm_quantization_dl_c_shuffle_i8_i8_i8_mk_kn_mn_instance.cpp 0 → 100644
View file @ 9fb64dae
// SPDX-License-Identifier: MIT
// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
#include "device_gemm_quantization_dl_c_shuffle_i8_i8_i8_instance.hpp"
namespace ck {
namespace tensor_operation {
namespace device {
namespace instance {
// Layout(A, B, C) = [Row, Row, Row]
void add_device_gemm_quantization_dl_c_shuffle_i8_i8_i8_mk_kn_mn_instances(
std::vector<std::unique_ptr<DeviceGemmMultipleD<Row,
Row,
Empty_Tuple,
Row,
int8_t,
int8_t,
Empty_Tuple,
int8_t,
PassThrough,
PassThrough,
Mul_Clamp>>>& instances)
{
add_device_operation_instances(
instances, device_gemm_quantization_dl_c_shuffle_i8_i8_i8_mk_kn_mn_instances<Mul_Clamp>{});
}
} // namespace instance
} // namespace device
} // namespace tensor_operation
} // namespace ck
library/src/tensor_operation_instance/gpu/quantization/gemm/device_gemm_quantization_dl_c_shuffle_i8_i8_i8_mk_nk_mn_instance.cpp 0 → 100644
View file @ 9fb64dae
// SPDX-License-Identifier: MIT
// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
#include "device_gemm_quantization_dl_c_shuffle_i8_i8_i8_instance.hpp"
namespace ck {
namespace tensor_operation {
namespace device {
namespace instance {
// Layout(A, B, C) = [Row, Col, Row]
void add_device_gemm_quantization_dl_c_shuffle_i8_i8_i8_mk_nk_mn_instances(
std::vector<std::unique_ptr<DeviceGemmMultipleD<Row,
Col,
Empty_Tuple,
Row,
int8_t,
int8_t,
Empty_Tuple,
int8_t,
PassThrough,
PassThrough,
Mul_Clamp>>>& instances)
{
add_device_operation_instances(
instances, device_gemm_quantization_dl_c_shuffle_i8_i8_i8_mk_nk_mn_instances<Mul_Clamp>{});
}
} // namespace instance
} // namespace device
} // namespace tensor_operation
} // namespace ck
library/src/tensor_operation_instance/gpu/quantization/gemm/device_gemm_quantization_xdl_c_shuffle_i8_i8_i8_instance.hpp 0 → 100644
View file @ 9fb64dae
// SPDX-License-Identifier: MIT
// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
#include "gemm_quantization_common.hpp"
#include "ck/tensor_operation/gpu/device/impl/device_gemm_multiple_d_xdl_cshuffle.hpp"
namespace ck {
namespace tensor_operation {
namespace device {
namespace instance {
template <typename OutElementOp, LoopScheduler GemmLoopScheduler, PipelineVersion GemmPipeline>
using device_gemm_quantization_xdl_c_shuffle_i8_i8_i8_km_kn_mn_instances = std::tuple<
// clang-format off
//##############################| A| B| Ds| E| AData| BData| AccData| CShuffle| DsData| EData| A| B| CDE| 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| LoopScheduler| Pipeline|
//##############################| Layout| Layout| Layout| Layout| Type| Type| Type| DataType| Type| Type| Elementwise| Elementwise| Elementwise| 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| | |
//##############################| | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | |
DeviceGemmMultipleD_Xdl_CShuffle< Col, Row, Empty_Tuple, Row, int8_t, int8_t, int32_t, int32_t, Empty_Tuple, int8_t, PassThrough, PassThrough, OutElementOp, MNKPadding, 1, 256, 256, 128, 64, 4, 4, 32, 32, 4, 2, S<4, 64, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 4, 0, S<8, 32, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 4, 0, 1, 1, S<1, 64, 1, 4>, 16, GemmLoopScheduler, GemmPipeline>,
DeviceGemmMultipleD_Xdl_CShuffle< Col, Row, Empty_Tuple, Row, int8_t, int8_t, int32_t, int32_t, Empty_Tuple, int8_t, PassThrough, PassThrough, OutElementOp, MNKPadding, 1, 256, 256, 128, 64, 16, 16, 32, 32, 4, 2, S<4, 64, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 16, 1, S<4, 64, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 2, 16, 1, 1, 1, S<1, 64, 1, 4>, 16, GemmLoopScheduler, GemmPipeline>,
DeviceGemmMultipleD_Xdl_CShuffle< Col, Row, Empty_Tuple, Row, int8_t, int8_t, int32_t, int32_t, Empty_Tuple, int8_t, PassThrough, PassThrough, OutElementOp, MNKPadding, 1, 256, 128, 256, 64, 4, 4, 32, 32, 2, 4, S<8, 32, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 4, 0, S<4, 64, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 4, 0, 1, 1, S<1, 64, 1, 4>, 16, GemmLoopScheduler, GemmPipeline>,
DeviceGemmMultipleD_Xdl_CShuffle< Col, Row, Empty_Tuple, Row, int8_t, int8_t, int32_t, int32_t, Empty_Tuple, int8_t, PassThrough, PassThrough, OutElementOp, MNKPadding, 1, 256, 128, 256, 64, 16, 16, 32, 32, 2, 4, S<4, 64, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 2, 16, 1, S<4, 64, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 16, 1, 1, 1, S<1, 64, 1, 4>, 16, GemmLoopScheduler, GemmPipeline>,
DeviceGemmMultipleD_Xdl_CShuffle< Col, Row, Empty_Tuple, Row, int8_t, int8_t, int32_t, int32_t, Empty_Tuple, int8_t, PassThrough, PassThrough, OutElementOp, MNKPadding, 1, 128, 128, 128, 64, 4, 4, 32, 32, 4, 2, S<4, 32, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 4, 0, S<4, 32, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 4, 0, 1, 1, S<1, 32, 1, 4>, 16, GemmLoopScheduler, GemmPipeline>,
DeviceGemmMultipleD_Xdl_CShuffle< Col, Row, Empty_Tuple, Row, int8_t, int8_t, int32_t, int32_t, Empty_Tuple, int8_t, PassThrough, PassThrough, OutElementOp, MNKPadding, 1, 128, 128, 128, 64, 16, 16, 32, 32, 4, 2, S<4, 32, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 16, 1, S<4, 32, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 16, 1, 1, 1, S<1, 32, 1, 4>, 16, GemmLoopScheduler, GemmPipeline>,
DeviceGemmMultipleD_Xdl_CShuffle< Col, Row, Empty_Tuple, Row, int8_t, int8_t, int32_t, int32_t, Empty_Tuple, int8_t, PassThrough, PassThrough, OutElementOp, MNKPadding, 1, 256, 128, 128, 64, 4, 4, 32, 32, 2, 2, S<8, 32, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 4, 0, S<8, 32, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 4, 0, 1, 1, S<1, 64, 1, 4>, 16, GemmLoopScheduler, GemmPipeline>,
DeviceGemmMultipleD_Xdl_CShuffle< Col, Row, Empty_Tuple, Row, int8_t, int8_t, int32_t, int32_t, Empty_Tuple, int8_t, PassThrough, PassThrough, OutElementOp, MNKPadding, 1, 256, 128, 128, 64, 16, 16, 32, 32, 2, 2, S<4, 64, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 2, 16, 1, S<4, 64, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 2, 16, 1, 1, 1, S<1, 64, 1, 4>, 16, GemmLoopScheduler, GemmPipeline>,
DeviceGemmMultipleD_Xdl_CShuffle< Col, Row, Empty_Tuple, Row, int8_t, int8_t, int32_t, int32_t, Empty_Tuple, int8_t, PassThrough, PassThrough, OutElementOp, MNKPadding, 1, 128, 128, 64, 64, 4, 4, 32, 32, 2, 2, S<4, 32, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 4, 0, S<4, 16, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 4, 0, 1, 1, S<1, 64, 1, 2>, 16, GemmLoopScheduler, GemmPipeline>,
DeviceGemmMultipleD_Xdl_CShuffle< Col, Row, Empty_Tuple, Row, int8_t, int8_t, int32_t, int32_t, Empty_Tuple, int8_t, PassThrough, PassThrough, OutElementOp, MNKPadding, 1, 128, 128, 64, 64, 16, 16, 32, 32, 2, 2, S<4, 32, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 16, 1, S<4, 32, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 2, 16, 1, 1, 1, S<1, 64, 1, 2>, 16, GemmLoopScheduler, GemmPipeline>,
DeviceGemmMultipleD_Xdl_CShuffle< Col, Row, Empty_Tuple, Row, int8_t, int8_t, int32_t, int32_t, Empty_Tuple, int8_t, PassThrough, PassThrough, OutElementOp, MNKPadding, 1, 128, 64, 128, 64, 4, 4, 32, 32, 2, 2, S<8, 16, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 4, 0, S<4, 32, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 4, 0, 1, 1, S<1, 32, 1, 4>, 16, GemmLoopScheduler, GemmPipeline>,
DeviceGemmMultipleD_Xdl_CShuffle< Col, Row, Empty_Tuple, Row, int8_t, int8_t, int32_t, int32_t, Empty_Tuple, int8_t, PassThrough, PassThrough, OutElementOp, MNKPadding, 1, 128, 64, 128, 64, 16, 16, 32, 32, 2, 2, S<4, 32, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 2, 16, 1, S<4, 32, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 16, 1, 1, 1, S<1, 32, 1, 4>, 16, GemmLoopScheduler, GemmPipeline>,
DeviceGemmMultipleD_Xdl_CShuffle< Col, Row, Empty_Tuple, Row, int8_t, int8_t, int32_t, int32_t, Empty_Tuple, int8_t, PassThrough, PassThrough, OutElementOp, MNKPadding, 1, 256, 128, 64, 64, 4, 4, 32, 32, 2, 1, S<8, 32, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 4, 0, S<16,16, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 4, 0, 1, 1, S<1, 64, 1, 4>, 16, GemmLoopScheduler, GemmPipeline>,
DeviceGemmMultipleD_Xdl_CShuffle< Col, Row, Empty_Tuple, Row, int8_t, int8_t, int32_t, int32_t, Empty_Tuple, int8_t, PassThrough, PassThrough, OutElementOp, MNKPadding, 1, 256, 128, 64, 64, 16, 16, 32, 32, 2, 1, S<4, 64, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 2, 16, 1, S<4, 64, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 1, 16, 1, 1, 1, S<1, 64, 1, 4>, 16, GemmLoopScheduler, GemmPipeline>,
DeviceGemmMultipleD_Xdl_CShuffle< Col, Row, Empty_Tuple, Row, int8_t, int8_t, int32_t, int32_t, Empty_Tuple, int8_t, PassThrough, PassThrough, OutElementOp, MNKPadding, 1, 256, 64, 128, 64, 4, 4, 32, 32, 1, 2, S<16,16, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 4, 0, S<8, 32, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 4, 0, 1, 1, S<1, 64, 1, 4>, 16, GemmLoopScheduler, GemmPipeline>,
DeviceGemmMultipleD_Xdl_CShuffle< Col, Row, Empty_Tuple, Row, int8_t, int8_t, int32_t, int32_t, Empty_Tuple, int8_t, PassThrough, PassThrough, OutElementOp, MNKPadding, 1, 256, 64, 128, 64, 16, 16, 32, 32, 1, 2, S<4, 64, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 1, 16, 1, S<4, 64, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 2, 16, 1, 1, 1, S<1, 64, 1, 4>, 16, GemmLoopScheduler, GemmPipeline>
// clang-format on
>;
template <typename OutElementOp, LoopScheduler GemmLoopScheduler, PipelineVersion GemmPipeline>
using device_gemm_quantization_xdl_c_shuffle_i8_i8_i8_km_nk_mn_instances = std::tuple<
// clang-format off
//##############################| A| B| Ds| E| AData| BData| AccData| CShuffle| DsData| EData| A| B| CDE| 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| LoopScheduler| Pipeline|
//##############################| Layout| Layout| Layout| Layout| Type| Type| Type| DataType| Type| Type| Elementwise| Elementwise| Elementwise| 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| | |
//##############################| | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | |
DeviceGemmMultipleD_Xdl_CShuffle< Col, Col, Empty_Tuple, Row, int8_t, int8_t, int32_t, int32_t, Empty_Tuple, int8_t, PassThrough, PassThrough, OutElementOp, MNKPadding, 1, 256, 256, 128, 64, 4, 16, 32, 32, 4, 2, S<4, 64, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 4, 0, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 16, 16, 1, 1, 1, S<1, 64, 1, 4>, 16, GemmLoopScheduler, GemmPipeline>,
DeviceGemmMultipleD_Xdl_CShuffle< Col, Col, Empty_Tuple, Row, int8_t, int8_t, int32_t, int32_t, Empty_Tuple, int8_t, PassThrough, PassThrough, OutElementOp, MNKPadding, 1, 256, 256, 128, 64, 16, 16, 32, 32, 4, 2, S<4, 64, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 16, 1, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 16, 16, 1, 1, 1, S<1, 64, 1, 4>, 16, GemmLoopScheduler, GemmPipeline>,
DeviceGemmMultipleD_Xdl_CShuffle< Col, Col, Empty_Tuple, Row, int8_t, int8_t, int32_t, int32_t, Empty_Tuple, int8_t, PassThrough, PassThrough, OutElementOp, MNKPadding, 1, 256, 128, 256, 64, 4, 16, 32, 32, 2, 4, S<8, 32, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 4, 0, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 16, 16, 1, 1, 1, S<1, 64, 1, 4>, 16, GemmLoopScheduler, GemmPipeline>,
DeviceGemmMultipleD_Xdl_CShuffle< Col, Col, Empty_Tuple, Row, int8_t, int8_t, int32_t, int32_t, Empty_Tuple, int8_t, PassThrough, PassThrough, OutElementOp, MNKPadding, 1, 256, 128, 256, 64, 16, 16, 32, 32, 2, 4, S<4, 64, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 2, 16, 1, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 16, 16, 1, 1, 1, S<1, 64, 1, 4>, 16, GemmLoopScheduler, GemmPipeline>,
DeviceGemmMultipleD_Xdl_CShuffle< Col, Col, Empty_Tuple, Row, int8_t, int8_t, int32_t, int32_t, Empty_Tuple, int8_t, PassThrough, PassThrough, OutElementOp, MNKPadding, 1, 128, 128, 128, 64, 4, 16, 32, 32, 4, 2, S<4, 32, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 4, 0, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 16, 16, 1, 1, 1, S<1, 32, 1, 4>, 16, GemmLoopScheduler, GemmPipeline>,
DeviceGemmMultipleD_Xdl_CShuffle< Col, Col, Empty_Tuple, Row, int8_t, int8_t, int32_t, int32_t, Empty_Tuple, int8_t, PassThrough, PassThrough, OutElementOp, MNKPadding, 1, 128, 128, 128, 64, 16, 16, 32, 32, 4, 2, S<4, 32, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 16, 1, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 16, 16, 1, 1, 1, S<1, 32, 1, 4>, 16, GemmLoopScheduler, GemmPipeline>,
DeviceGemmMultipleD_Xdl_CShuffle< Col, Col, Empty_Tuple, Row, int8_t, int8_t, int32_t, int32_t, Empty_Tuple, int8_t, PassThrough, PassThrough, OutElementOp, MNKPadding, 1, 256, 128, 128, 64, 4, 16, 32, 32, 2, 2, S<8, 32, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 4, 0, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 16, 16, 1, 1, 1, S<1, 64, 1, 4>, 16, GemmLoopScheduler, GemmPipeline>,
DeviceGemmMultipleD_Xdl_CShuffle< Col, Col, Empty_Tuple, Row, int8_t, int8_t, int32_t, int32_t, Empty_Tuple, int8_t, PassThrough, PassThrough, OutElementOp, MNKPadding, 1, 256, 128, 128, 64, 16, 16, 32, 32, 2, 2, S<4, 64, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 2, 16, 1, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 16, 16, 1, 1, 1, S<1, 64, 1, 4>, 16, GemmLoopScheduler, GemmPipeline>,
DeviceGemmMultipleD_Xdl_CShuffle< Col, Col, Empty_Tuple, Row, int8_t, int8_t, int32_t, int32_t, Empty_Tuple, int8_t, PassThrough, PassThrough, OutElementOp, MNKPadding, 1, 128, 128, 64, 64, 4, 16, 32, 32, 2, 2, S<4, 32, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 4, 0, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 16, 16, 1, 1, 1, S<1, 64, 1, 2>, 16, GemmLoopScheduler, GemmPipeline>,
DeviceGemmMultipleD_Xdl_CShuffle< Col, Col, Empty_Tuple, Row, int8_t, int8_t, int32_t, int32_t, Empty_Tuple, int8_t, PassThrough, PassThrough, OutElementOp, MNKPadding, 1, 128, 128, 64, 64, 16, 16, 32, 32, 2, 2, S<4, 32, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 16, 1, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 16, 16, 1, 1, 1, S<1, 64, 1, 2>, 16, GemmLoopScheduler, GemmPipeline>,
DeviceGemmMultipleD_Xdl_CShuffle< Col, Col, Empty_Tuple, Row, int8_t, int8_t, int32_t, int32_t, Empty_Tuple, int8_t, PassThrough, PassThrough, OutElementOp, MNKPadding, 1, 128, 64, 128, 64, 4, 16, 32, 32, 2, 2, S<8, 16, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 4, 0, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 16, 16, 1, 1, 1, S<1, 32, 1, 4>, 16, GemmLoopScheduler, GemmPipeline>,
DeviceGemmMultipleD_Xdl_CShuffle< Col, Col, Empty_Tuple, Row, int8_t, int8_t, int32_t, int32_t, Empty_Tuple, int8_t, PassThrough, PassThrough, OutElementOp, MNKPadding, 1, 128, 64, 128, 64, 16, 16, 32, 32, 2, 2, S<4, 32, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 2, 16, 1, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 16, 16, 1, 1, 1, S<1, 32, 1, 4>, 16, GemmLoopScheduler, GemmPipeline>,
DeviceGemmMultipleD_Xdl_CShuffle< Col, Col, Empty_Tuple, Row, int8_t, int8_t, int32_t, int32_t, Empty_Tuple, int8_t, PassThrough, PassThrough, OutElementOp, MNKPadding, 1, 256, 128, 64, 64, 4, 16, 32, 32, 2, 1, S<8, 32, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 4, 0, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 16, 16, 1, 1, 1, S<1, 64, 1, 4>, 16, GemmLoopScheduler, GemmPipeline>,
DeviceGemmMultipleD_Xdl_CShuffle< Col, Col, Empty_Tuple, Row, int8_t, int8_t, int32_t, int32_t, Empty_Tuple, int8_t, PassThrough, PassThrough, OutElementOp, MNKPadding, 1, 256, 128, 64, 64, 16, 16, 32, 32, 2, 1, S<4, 64, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 2, 16, 1, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 16, 16, 1, 1, 1, S<1, 64, 1, 4>, 16, GemmLoopScheduler, GemmPipeline>,
DeviceGemmMultipleD_Xdl_CShuffle< Col, Col, Empty_Tuple, Row, int8_t, int8_t, int32_t, int32_t, Empty_Tuple, int8_t, PassThrough, PassThrough, OutElementOp, MNKPadding, 1, 256, 64, 128, 64, 4, 16, 32, 32, 1, 2, S<16,16, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 4, 0, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 16, 16, 1, 1, 1, S<1, 64, 1, 4>, 16, GemmLoopScheduler, GemmPipeline>,
DeviceGemmMultipleD_Xdl_CShuffle< Col, Col, Empty_Tuple, Row, int8_t, int8_t, int32_t, int32_t, Empty_Tuple, int8_t, PassThrough, PassThrough, OutElementOp, MNKPadding, 1, 256, 64, 128, 64, 16, 16, 32, 32, 1, 2, S<4, 64, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 1, 16, 1, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 16, 16, 1, 1, 1, S<1, 64, 1, 4>, 16, GemmLoopScheduler, GemmPipeline>
// clang-format on
>;
template <typename OutElementOp, LoopScheduler GemmLoopScheduler, PipelineVersion GemmPipeline>
using device_gemm_quantization_xdl_c_shuffle_i8_i8_i8_mk_kn_mn_instances = std::tuple<
// clang-format off
//##############################| A| B| Ds| E| AData| BData| AccData| CShuffle| DsData| EData| A| B| CDE| 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| LoopScheduler| Pipeline|
//##############################| Layout| Layout| Layout| Layout| Type| Type| Type| DataType| Type| Type| Elementwise| Elementwise| Elementwise| 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| | |
//##############################| | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | |
DeviceGemmMultipleD_Xdl_CShuffle< Row, Row, Empty_Tuple, Row, int8_t, int8_t, int32_t, int32_t, Empty_Tuple, int8_t, PassThrough, PassThrough, OutElementOp, MNKPadding, 1, 256, 256, 128, 64, 16, 4, 32, 32, 4, 2, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 16, 16, 1, S<8, 32, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 4, 0, 1, 1, S<1, 64, 1, 4>, 16, GemmLoopScheduler, GemmPipeline>,
DeviceGemmMultipleD_Xdl_CShuffle< Row, Row, Empty_Tuple, Row, int8_t, int8_t, int32_t, int32_t, Empty_Tuple, int8_t, PassThrough, PassThrough, OutElementOp, MNKPadding, 1, 256, 256, 128, 64, 16, 16, 32, 32, 4, 2, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 16, 16, 1, S<4, 64, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 2, 16, 1, 1, 1, S<1, 64, 1, 4>, 16, GemmLoopScheduler, GemmPipeline>,
DeviceGemmMultipleD_Xdl_CShuffle< Row, Row, Empty_Tuple, Row, int8_t, int8_t, int32_t, int32_t, Empty_Tuple, int8_t, PassThrough, PassThrough, OutElementOp, MNKPadding, 1, 256, 128, 256, 64, 16, 4, 32, 32, 2, 4, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 16, 16, 1, S<4, 64, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 4, 0, 1, 1, S<1, 64, 1, 4>, 16, GemmLoopScheduler, GemmPipeline>,
DeviceGemmMultipleD_Xdl_CShuffle< Row, Row, Empty_Tuple, Row, int8_t, int8_t, int32_t, int32_t, Empty_Tuple, int8_t, PassThrough, PassThrough, OutElementOp, MNKPadding, 1, 256, 128, 256, 64, 16, 16, 32, 32, 2, 4, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 16, 16, 1, S<4, 64, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 16, 1, 1, 1, S<1, 64, 1, 4>, 16, GemmLoopScheduler, GemmPipeline>,
DeviceGemmMultipleD_Xdl_CShuffle< Row, Row, Empty_Tuple, Row, int8_t, int8_t, int32_t, int32_t, Empty_Tuple, int8_t, PassThrough, PassThrough, OutElementOp, MNKPadding, 1, 128, 128, 128, 64, 16, 4, 32, 32, 4, 2, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 16, 16, 1, S<4, 32, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 4, 0, 1, 1, S<1, 32, 1, 4>, 16, GemmLoopScheduler, GemmPipeline>,
DeviceGemmMultipleD_Xdl_CShuffle< Row, Row, Empty_Tuple, Row, int8_t, int8_t, int32_t, int32_t, Empty_Tuple, int8_t, PassThrough, PassThrough, OutElementOp, MNKPadding, 1, 128, 128, 128, 64, 16, 16, 32, 32, 4, 2, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 16, 16, 1, S<4, 32, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 16, 1, 1, 1, S<1, 32, 1, 4>, 16, GemmLoopScheduler, GemmPipeline>,
DeviceGemmMultipleD_Xdl_CShuffle< Row, Row, Empty_Tuple, Row, int8_t, int8_t, int32_t, int32_t, Empty_Tuple, int8_t, PassThrough, PassThrough, OutElementOp, MNKPadding, 1, 256, 128, 128, 64, 16, 4, 32, 32, 2, 2, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 16, 16, 1, S<8, 32, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 4, 0, 1, 1, S<1, 64, 1, 4>, 16, GemmLoopScheduler, GemmPipeline>,
DeviceGemmMultipleD_Xdl_CShuffle< Row, Row, Empty_Tuple, Row, int8_t, int8_t, int32_t, int32_t, Empty_Tuple, int8_t, PassThrough, PassThrough, OutElementOp, MNKPadding, 1, 256, 128, 128, 64, 16, 16, 32, 32, 2, 2, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 16, 16, 1, S<4, 64, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 2, 16, 1, 1, 1, S<1, 64, 1, 4>, 16, GemmLoopScheduler, GemmPipeline>,
DeviceGemmMultipleD_Xdl_CShuffle< Row, Row, Empty_Tuple, Row, int8_t, int8_t, int32_t, int32_t, Empty_Tuple, int8_t, PassThrough, PassThrough, OutElementOp, MNKPadding, 1, 128, 128, 64, 64, 16, 4, 32, 32, 2, 2, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 16, 16, 1, S<8, 16, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 4, 0, 1, 1, S<1, 64, 1, 2>, 16, GemmLoopScheduler, GemmPipeline>,
DeviceGemmMultipleD_Xdl_CShuffle< Row, Row, Empty_Tuple, Row, int8_t, int8_t, int32_t, int32_t, Empty_Tuple, int8_t, PassThrough, PassThrough, OutElementOp, MNKPadding, 1, 128, 128, 64, 64, 16, 16, 32, 32, 2, 2, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 16, 16, 1, S<4, 32, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 2, 16, 1, 1, 1, S<1, 64, 1, 2>, 16, GemmLoopScheduler, GemmPipeline>,
DeviceGemmMultipleD_Xdl_CShuffle< Row, Row, Empty_Tuple, Row, int8_t, int8_t, int32_t, int32_t, Empty_Tuple, int8_t, PassThrough, PassThrough, OutElementOp, MNKPadding, 1, 128, 64, 128, 64, 16, 4, 32, 32, 2, 2, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 16, 16, 1, S<4, 32, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 4, 0, 1, 1, S<1, 32, 1, 4>, 16, GemmLoopScheduler, GemmPipeline>,
DeviceGemmMultipleD_Xdl_CShuffle< Row, Row, Empty_Tuple, Row, int8_t, int8_t, int32_t, int32_t, Empty_Tuple, int8_t, PassThrough, PassThrough, OutElementOp, MNKPadding, 1, 128, 64, 128, 64, 16, 16, 32, 32, 2, 2, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 16, 16, 1, S<4, 32, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 16, 1, 1, 1, S<1, 32, 1, 4>, 16, GemmLoopScheduler, GemmPipeline>,
DeviceGemmMultipleD_Xdl_CShuffle< Row, Row, Empty_Tuple, Row, int8_t, int8_t, int32_t, int32_t, Empty_Tuple, int8_t, PassThrough, PassThrough, OutElementOp, MNKPadding, 1, 256, 128, 64, 64, 16, 4, 32, 32, 2, 1, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 16, 16, 1, S<16,16, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 4, 0, 1, 1, S<1, 64, 1, 4>, 16, GemmLoopScheduler, GemmPipeline>,
DeviceGemmMultipleD_Xdl_CShuffle< Row, Row, Empty_Tuple, Row, int8_t, int8_t, int32_t, int32_t, Empty_Tuple, int8_t, PassThrough, PassThrough, OutElementOp, MNKPadding, 1, 256, 128, 64, 64, 16, 16, 32, 32, 2, 1, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 16, 16, 1, S<4, 64, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 1, 16, 1, 1, 1, S<1, 64, 1, 4>, 16, GemmLoopScheduler, GemmPipeline>,
DeviceGemmMultipleD_Xdl_CShuffle< Row, Row, Empty_Tuple, Row, int8_t, int8_t, int32_t, int32_t, Empty_Tuple, int8_t, PassThrough, PassThrough, OutElementOp, MNKPadding, 1, 256, 64, 128, 64, 16, 4, 32, 32, 1, 2, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 16, 16, 1, S<8, 32, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 4, 0, 1, 1, S<1, 64, 1, 4>, 16, GemmLoopScheduler, GemmPipeline>,
DeviceGemmMultipleD_Xdl_CShuffle< Row, Row, Empty_Tuple, Row, int8_t, int8_t, int32_t, int32_t, Empty_Tuple, int8_t, PassThrough, PassThrough, OutElementOp, MNKPadding, 1, 256, 64, 128, 64, 16, 16, 32, 32, 1, 2, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 16, 16, 1, S<4, 64, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 2, 16, 1, 1, 1, S<1, 64, 1, 4>, 16, GemmLoopScheduler, GemmPipeline>
// clang-format on
>;
template <typename OutElementOp, LoopScheduler GemmLoopScheduler, PipelineVersion GemmPipeline>
using device_gemm_quantization_xdl_c_shuffle_i8_i8_i8_mk_nk_mn_instances = std::tuple<
// clang-format off
//##############################| A| B| Ds| E| AData| BData| AccData| CShuffle| DsData| EData| A| B| CDE| 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| LoopScheduler| Pipeline|
//##############################| Layout| Layout| Layout| Layout| Type| Type| Type| DataType| Type| Type| Elementwise| Elementwise| Elementwise| 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| | |
//##############################| | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | |
DeviceGemmMultipleD_Xdl_CShuffle< Row, Col, Empty_Tuple, Row, int8_t, int8_t, int32_t, int32_t, Empty_Tuple, int8_t, PassThrough, PassThrough, OutElementOp, MNKPadding, 1, 256, 256, 128, 64, 16, 16, 32, 32, 4, 2, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 16, 16, 1, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, 1, 1, S<1, 32, 1, 4>, 16, GemmLoopScheduler, GemmPipeline>,
DeviceGemmMultipleD_Xdl_CShuffle< Row, Col, Empty_Tuple, Row, int8_t, int8_t, int32_t, int32_t, Empty_Tuple, int8_t, PassThrough, PassThrough, OutElementOp, MNKPadding, 1, 256, 128, 256, 64, 16, 16, 32, 32, 2, 4, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 16, 16, 1, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, 1, 1, S<1, 32, 1, 4>, 16, GemmLoopScheduler, GemmPipeline>,
DeviceGemmMultipleD_Xdl_CShuffle< Row, Col, Empty_Tuple, Row, int8_t, int8_t, int32_t, int32_t, Empty_Tuple, int8_t, PassThrough, PassThrough, OutElementOp, MNKPadding, 1, 128, 128, 128, 64, 16, 16, 32, 32, 4, 2, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 16, 16, 1, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, 1, 1, S<1, 16, 1, 4>, 16, GemmLoopScheduler, GemmPipeline>,
DeviceGemmMultipleD_Xdl_CShuffle< Row, Col, Empty_Tuple, Row, int8_t, int8_t, int32_t, int32_t, Empty_Tuple, int8_t, PassThrough, PassThrough, OutElementOp, MNKPadding, 1, 256, 128, 128, 64, 16, 16, 32, 32, 2, 2, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 16, 16, 1, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, 1, 1, S<1, 32, 1, 4>, 16, GemmLoopScheduler, GemmPipeline>,
DeviceGemmMultipleD_Xdl_CShuffle< Row, Col, Empty_Tuple, Row, int8_t, int8_t, int32_t, int32_t, Empty_Tuple, int8_t, PassThrough, PassThrough, OutElementOp, MNKPadding, 1, 128, 128, 64, 64, 16, 16, 32, 32, 2, 2, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 16, 16, 1, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, 1, 1, S<1, 32, 1, 2>, 16, GemmLoopScheduler, GemmPipeline>,
DeviceGemmMultipleD_Xdl_CShuffle< Row, Col, Empty_Tuple, Row, int8_t, int8_t, int32_t, int32_t, Empty_Tuple, int8_t, PassThrough, PassThrough, OutElementOp, MNKPadding, 1, 128, 64, 128, 64, 16, 16, 32, 32, 2, 2, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 16, 16, 1, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, 1, 1, S<1, 16, 1, 4>, 16, GemmLoopScheduler, GemmPipeline>,
DeviceGemmMultipleD_Xdl_CShuffle< Row, Col, Empty_Tuple, Row, int8_t, int8_t, int32_t, int32_t, Empty_Tuple, int8_t, PassThrough, PassThrough, OutElementOp, MNKPadding, 1, 64, 64, 64, 64, 16, 16, 32, 32, 2, 2, S<4, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 16, 16, 1, S<4, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, 1, 1, S<1, 16, 1, 2>, 16, GemmLoopScheduler, GemmPipeline>,
DeviceGemmMultipleD_Xdl_CShuffle< Row, Col, Empty_Tuple, Row, int8_t, int8_t, int32_t, int32_t, Empty_Tuple, int8_t, PassThrough, PassThrough, OutElementOp, MNKPadding, 1, 256, 128, 64, 64, 16, 16, 32, 32, 2, 1, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 16, 16, 1, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, 1, 1, S<1, 32, 1, 4>, 16, GemmLoopScheduler, GemmPipeline>,
DeviceGemmMultipleD_Xdl_CShuffle< Row, Col, Empty_Tuple, Row, int8_t, int8_t, int32_t, int32_t, Empty_Tuple, int8_t, PassThrough, PassThrough, OutElementOp, MNKPadding, 1, 256, 64, 128, 64, 16, 16, 32, 32, 1, 2, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 16, 16, 1, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, 1, 1, S<1, 32, 1, 4>, 16, GemmLoopScheduler, GemmPipeline>,
DeviceGemmMultipleD_Xdl_CShuffle< Row, Col, Empty_Tuple, Row, int8_t, int8_t, int32_t, int32_t, Empty_Tuple, int8_t, PassThrough, PassThrough, OutElementOp, MNKPadding, 1, 128, 128, 32, 64, 16, 16, 32, 32, 2, 1, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 16, 16, 1, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, 1, 1, S<1, 32, 1, 2>, 16, GemmLoopScheduler, GemmPipeline>,
DeviceGemmMultipleD_Xdl_CShuffle< Row, Col, Empty_Tuple, Row, int8_t, int8_t, int32_t, int32_t, Empty_Tuple, int8_t, PassThrough, PassThrough, OutElementOp, MNKPadding, 1, 128, 32, 128, 64, 16, 16, 32, 32, 1, 2, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 16, 16, 1, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, 1, 1, S<1, 16, 1, 4>, 16, GemmLoopScheduler, GemmPipeline>,
DeviceGemmMultipleD_Xdl_CShuffle< Row, Col, Empty_Tuple, Row, int8_t, int8_t, int32_t, int32_t, Empty_Tuple, int8_t, PassThrough, PassThrough, OutElementOp, MNKPadding, 1, 64, 64, 32, 64, 16, 16, 32, 32, 2, 1, S<4, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 16, 16, 1, S<4, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, 1, 1, S<1, 16, 1, 2>, 16, GemmLoopScheduler, GemmPipeline>,
DeviceGemmMultipleD_Xdl_CShuffle< Row, Col, Empty_Tuple, Row, int8_t, int8_t, int32_t, int32_t, Empty_Tuple, int8_t, PassThrough, PassThrough, OutElementOp, MNKPadding, 1, 64, 32, 64, 64, 16, 16, 32, 32, 1, 2, S<4, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 16, 16, 1, S<4, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, 1, 1, S<1, 16, 1, 2>, 16, GemmLoopScheduler, GemmPipeline>
// clang-format on
>;
} // namespace instance
} // namespace device
} // namespace tensor_operation
} // namespace ck
library/src/tensor_operation_instance/gpu/quantization/gemm/device_gemm_quantization_xdl_c_shuffle_i8_i8_i8_km_kn_mn_instance.cpp 0 → 100644
View file @ 9fb64dae
// SPDX-License-Identifier: MIT
// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
#include "device_gemm_quantization_xdl_c_shuffle_i8_i8_i8_instance.hpp"
namespace ck {
namespace tensor_operation {
namespace device {
namespace instance {
// Layout(A, B, C) = [Col, Row, Row]
void add_device_gemm_quantization_xdl_c_shuffle_i8_i8_i8_km_kn_mn_instances(
std::vector<std::unique_ptr<DeviceGemmMultipleD<Col,
Row,
Empty_Tuple,
Row,
int8_t,
int8_t,
Empty_Tuple,
int8_t,
PassThrough,
PassThrough,
Mul_Clamp>>>& instances)
{
add_device_operation_instances(
instances,
device_gemm_quantization_xdl_c_shuffle_i8_i8_i8_km_kn_mn_instances<Mul_Clamp,
LoopScheduler::Default,
PipelineVersion::v1>{});
#if CK_EXPERIMENTAL_INTER_WAVE_INSTANCES
add_device_operation_instances(
instances,
device_gemm_quantization_xdl_c_shuffle_i8_i8_i8_km_kn_mn_instances<Mul_Clamp,
LoopScheduler::Interwave,
PipelineVersion::v1>{});
#endif
#if CK_EXPERIMENTAL_PIPELINE_V2_INSTANCES
add_device_operation_instances(
instances,
device_gemm_quantization_xdl_c_shuffle_i8_i8_i8_km_kn_mn_instances<Mul_Clamp,
LoopScheduler::Default,
PipelineVersion::v2>{});
#endif
}
} // namespace instance
} // namespace device
} // namespace tensor_operation
} // namespace ck
library/src/tensor_operation_instance/gpu/quantization/gemm/device_gemm_quantization_xdl_c_shuffle_i8_i8_i8_km_nk_mn_instance.cpp 0 → 100644
View file @ 9fb64dae
// SPDX-License-Identifier: MIT
// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
#include "device_gemm_quantization_xdl_c_shuffle_i8_i8_i8_instance.hpp"
namespace ck {
namespace tensor_operation {
namespace device {
namespace instance {
// Layout(A, B, C) = [Col, Col, Row]
void add_device_gemm_quantization_xdl_c_shuffle_i8_i8_i8_km_nk_mn_instances(
std::vector<std::unique_ptr<DeviceGemmMultipleD<Col,
Col,
Empty_Tuple,
Row,
int8_t,
int8_t,
Empty_Tuple,
int8_t,
PassThrough,
PassThrough,
Mul_Clamp>>>& instances)
{
add_device_operation_instances(
instances,
device_gemm_quantization_xdl_c_shuffle_i8_i8_i8_km_nk_mn_instances<Mul_Clamp,
LoopScheduler::Default,
PipelineVersion::v1>{});
#if CK_EXPERIMENTAL_INTER_WAVE_INSTANCES
add_device_operation_instances(
instances,
device_gemm_quantization_xdl_c_shuffle_i8_i8_i8_km_nk_mn_instances<Mul_Clamp,
LoopScheduler::Interwave,
PipelineVersion::v1>{});
#endif
#if CK_EXPERIMENTAL_PIPELINE_V2_INSTANCES
add_device_operation_instances(
instances,
device_gemm_quantization_xdl_c_shuffle_i8_i8_i8_km_nk_mn_instances<Mul_Clamp,
LoopScheduler::Default,
PipelineVersion::v2>{});
#endif
}
} // namespace instance
} // namespace device
} // namespace tensor_operation
} // namespace ck
library/src/tensor_operation_instance/gpu/quantization/gemm/device_gemm_quantization_xdl_c_shuffle_i8_i8_i8_mk_kn_mn_instance.cpp 0 → 100644
View file @ 9fb64dae
// SPDX-License-Identifier: MIT
// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
#include "device_gemm_quantization_xdl_c_shuffle_i8_i8_i8_instance.hpp"
namespace ck {
namespace tensor_operation {
namespace device {
namespace instance {
// Layout(A, B, C) = [Row, Row, Row]
void add_device_gemm_quantization_xdl_c_shuffle_i8_i8_i8_mk_kn_mn_instances(
std::vector<std::unique_ptr<DeviceGemmMultipleD<Row,
Row,
Empty_Tuple,
Row,
int8_t,
int8_t,
Empty_Tuple,
int8_t,
PassThrough,
PassThrough,
Mul_Clamp>>>& instances)
{
add_device_operation_instances(
instances,
device_gemm_quantization_xdl_c_shuffle_i8_i8_i8_mk_kn_mn_instances<Mul_Clamp,
LoopScheduler::Default,
PipelineVersion::v1>{});
#if CK_EXPERIMENTAL_INTER_WAVE_INSTANCES
add_device_operation_instances(
instances,
device_gemm_quantization_xdl_c_shuffle_i8_i8_i8_mk_kn_mn_instances<Mul_Clamp,
LoopScheduler::Interwave,
PipelineVersion::v1>{});
#endif
#if CK_EXPERIMENTAL_PIPELINE_V2_INSTANCES
add_device_operation_instances(
instances,
device_gemm_quantization_xdl_c_shuffle_i8_i8_i8_mk_kn_mn_instances<Mul_Clamp,
LoopScheduler::Default,
PipelineVersion::v2>{});
#endif
}
} // namespace instance
} // namespace device
} // namespace tensor_operation
} // namespace ck
library/src/tensor_operation_instance/gpu/quantization/gemm/device_gemm_quantization_xdl_c_shuffle_i8_i8_i8_mk_nk_mn_instance.cpp 0 → 100644
View file @ 9fb64dae
// SPDX-License-Identifier: MIT
// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
#include "device_gemm_quantization_xdl_c_shuffle_i8_i8_i8_instance.hpp"
namespace ck {
namespace tensor_operation {
namespace device {
namespace instance {
// Layout(A, B, C) = [Row, Col, Row]
void add_device_gemm_quantization_xdl_c_shuffle_i8_i8_i8_mk_nk_mn_instances(
std::vector<std::unique_ptr<DeviceGemmMultipleD<Row,
Col,
Empty_Tuple,
Row,
int8_t,
int8_t,
Empty_Tuple,
int8_t,
PassThrough,
PassThrough,
Mul_Clamp>>>& instances)
{
add_device_operation_instances(
instances,
device_gemm_quantization_xdl_c_shuffle_i8_i8_i8_mk_nk_mn_instances<Mul_Clamp,
LoopScheduler::Default,
PipelineVersion::v1>{});
#if CK_EXPERIMENTAL_INTER_WAVE_INSTANCES
add_device_operation_instances(
instances,
device_gemm_quantization_xdl_c_shuffle_i8_i8_i8_mk_nk_mn_instances<Mul_Clamp,
LoopScheduler::Interwave,
PipelineVersion::v1>{});
#endif
#if CK_EXPERIMENTAL_PIPELINE_V2_INSTANCES
add_device_operation_instances(
instances,
device_gemm_quantization_xdl_c_shuffle_i8_i8_i8_mk_nk_mn_instances<Mul_Clamp,
LoopScheduler::Default,
PipelineVersion::v2>{});
#endif
}
} // namespace instance
} // namespace device
} // namespace tensor_operation
} // namespace ck
library/src/tensor_operation_instance/gpu/quantization/gemm/gemm_quantization_common.hpp 0 → 100644
View file @ 9fb64dae
// SPDX-License-Identifier: MIT
// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
#pragma once
#include "ck/ck.hpp"
#include "ck/tensor_operation/gpu/device/tensor_layout.hpp"
#include "ck/tensor_operation/gpu/element/element_wise_operation.hpp"
#include "ck/library/tensor_operation_instance/add_device_operation_instance.hpp"
#include "ck/tensor_operation/gpu/device/gemm_specialization.hpp"
namespace ck {
namespace tensor_operation {
namespace device {
namespace instance {
template <ck::index_t... Is>
using S = ck::Sequence<Is...>;
using Row = ck::tensor_layout::gemm::RowMajor;
using Col = ck::tensor_layout::gemm::ColumnMajor;
using Empty_Tuple = ck::Tuple<>;
using Row_Row_Tuple = ck::Tuple<Row, Row>;
using Col_Col_Tuple = ck::Tuple<Col, Col>;
using PassThrough = ck::tensor_operation::element_wise::PassThrough;
using Relu = ck::tensor_operation::element_wise::Relu;
using Mul_Clamp = ck::tensor_operation::element_wise::Activation_Mul_Clamp<PassThrough>;
using Relu_Mul_Clamp = ck::tensor_operation::element_wise::Activation_Mul_Clamp<Relu>;
using Add_Mul_Clamp = ck::tensor_operation::element_wise::Add_Activation_Mul_Clamp<PassThrough>;
using Add_Relu_Mul_Clamp = ck::tensor_operation::element_wise::Add_Activation_Mul_Clamp<Relu>;
static constexpr auto MNKPadding = ck::tensor_operation::device::GemmSpecialization::MNKPadding;
} // namespace instance
} // namespace device
} // namespace tensor_operation
} // namespace ck
profiler/include/profiler/profile_grouped_gemm_fastgelu_impl.hpp 0 → 100644
View file @ 9fb64dae
// SPDX-License-Identifier: MIT
// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
#pragma once
#include <iomanip>
#include "ck/ck.hpp"
#include "ck/library/tensor_operation_instance/gpu/grouped_gemm_fastgelu.hpp"
#include "ck/library/utility/check_err.hpp"
#include "ck/library/utility/device_memory.hpp"
#include "ck/library/utility/host_tensor.hpp"
#include "ck/library/utility/fill.hpp"
#include "ck/library/utility/literals.hpp"
#include "ck/library/reference_tensor_operation/cpu/reference_gemm.hpp"
#include "ck/tensor_operation/gpu/device/tensor_layout.hpp"
#include "ck/tensor_operation/gpu/device/device_grouped_gemm.hpp"
#include "ck/tensor_operation/gpu/element/element_wise_operation.hpp"
namespace ck {
namespace profiler {
template <typename ADataType,
typename BDataType,
typename CDataType,
typename AccDataType,
typename ALayout,
typename BLayout,
typename CLayout>
bool profile_grouped_gemm_fastgelu_impl(int do_verification,
int init_method,
bool do_log,
bool time_kernel,
const std::vector<int>& Ms,
const std::vector<int>& Ns,
const std::vector<int>& Ks,
const std::vector<int>& StrideAs,
const std::vector<int>& StrideBs,
const std::vector<int>& StrideCs)
{
bool pass = true;
auto f_host_tensor_descriptor =
[](std::size_t row, std::size_t col, std::size_t stride, auto layout) {
using namespace ck::literals;
if(is_same<decltype(layout), tensor_layout::gemm::RowMajor>::value)
{
return HostTensorDescriptor({row, col}, {stride, 1_uz});
}
else
{
return HostTensorDescriptor({row, col}, {1_uz, stride});
}
};
std::size_t group_count = Ms.size();
if(!(group_count == Ns.size() && group_count == Ks.size() && group_count == StrideAs.size() &&
group_count == StrideBs.size() && group_count == StrideCs.size()))
{
throw std::runtime_error("wrong! inconsistent M/N/Ks, StrideA/B/Cs size\n");
}
std::vector<Tensor<ADataType>> a_m_k;
std::vector<Tensor<BDataType>> b_k_n;
std::vector<Tensor<CDataType>> c_m_n_device_results;
for(std::size_t i = 0; i < group_count; i++)
{
a_m_k.push_back(
Tensor<ADataType>(f_host_tensor_descriptor(Ms[i], Ks[i], StrideAs[i], ALayout{})));
b_k_n.push_back(
Tensor<BDataType>(f_host_tensor_descriptor(Ks[i], Ns[i], StrideBs[i], BLayout{})));
c_m_n_device_results.push_back(
Tensor<CDataType>(f_host_tensor_descriptor(Ms[i], Ns[i], StrideCs[i], CLayout{})));
std::cout << "group: " << i << " a_m_k[" << i << "]:" << a_m_k[i].mDesc << ", b_k_n[" << i
<< "]:" << b_k_n[i].mDesc << ", c_m_n_device_results[" << i
<< "]:" << c_m_n_device_results[i].mDesc << std::endl;
switch(init_method)
{
case 0: break;
case 1:
ck::utils::FillUniformDistributionIntegerValue<ADataType>{}(a_m_k[i]);
ck::utils::FillUniformDistributionIntegerValue<BDataType>{}(b_k_n[i]);
break;
default:
ck::utils::FillUniformDistribution<ADataType>{0.0, 1.0}(a_m_k[i]);
ck::utils::FillUniformDistribution<BDataType>{-0.5, 0.5}(b_k_n[i]);
}
ck::utils::FillConstant<CDataType>{}(c_m_n_device_results[i]);
}
using AElementOp = ck::tensor_operation::element_wise::PassThrough;
using BElementOp = ck::tensor_operation::element_wise::PassThrough;
using CElementOp = ck::tensor_operation::element_wise::FastGelu;
const auto a_element_op = AElementOp{};
const auto b_element_op = BElementOp{};
const auto c_element_op = CElementOp{};
using DeviceMemPtr = std::unique_ptr<DeviceMem>;
std::vector<DeviceMemPtr> a_device_buf, b_device_buf, c_device_buf;
a_device_buf.reserve(group_count);
b_device_buf.reserve(group_count);
c_device_buf.reserve(group_count);
std::vector<const void*> p_a, p_b;
std::vector<void*> p_c;
p_a.reserve(group_count);
p_b.reserve(group_count);
p_c.reserve(group_count);
std::vector<ck::tensor_operation::device::GemmDesc> gemm_descs;
gemm_descs.reserve(group_count);
for(std::size_t i = 0; i < group_count; i++)
{
a_device_buf.emplace_back(
std::make_unique<DeviceMem>(sizeof(ADataType) * a_m_k[i].mDesc.GetElementSpaceSize()));
b_device_buf.emplace_back(
std::make_unique<DeviceMem>(sizeof(BDataType) * b_k_n[i].mDesc.GetElementSpaceSize()));
c_device_buf.emplace_back(std::make_unique<DeviceMem>(
sizeof(CDataType) * c_m_n_device_results[i].mDesc.GetElementSpaceSize()));
a_device_buf[i]->ToDevice(a_m_k[i].mData.data());
b_device_buf[i]->ToDevice(b_k_n[i].mData.data());
c_device_buf[i]->SetZero();
gemm_descs.push_back({Ms[i], Ns[i], Ks[i], StrideAs[i], StrideBs[i], StrideCs[i], {}});
p_a.push_back(a_device_buf[i]->GetDeviceBuffer());
p_b.push_back(b_device_buf[i]->GetDeviceBuffer());
p_c.push_back(c_device_buf[i]->GetDeviceBuffer());
}
using DeviceOp = ck::tensor_operation::device::DeviceGroupedGemm<ALayout,
BLayout,
ck::Tuple<>,
CLayout,
ADataType,
BDataType,
ck::Tuple<>,
CDataType,
AElementOp,
BElementOp,
CElementOp>;
const auto op_ptrs = ck::tensor_operation::device::instance::DeviceOperationInstanceFactory<
DeviceOp>::GetInstances();
if(op_ptrs.size() <= 0)
{
throw std::runtime_error("wrong! no device GEMM instance found");
}
std::string best_gemm_name;
float best_ave_time = 0;
float best_tflops = 0;
float best_gb_per_sec = 0;
auto p_ds = std::vector<std::array<const void*, 0>>{};
// profile device GEMM instances
for(auto& gemm_ptr : op_ptrs)
{
auto argument_ptr = gemm_ptr->MakeArgumentPointer(
p_a, p_b, p_ds, p_c, gemm_descs, a_element_op, b_element_op, c_element_op);
auto invoker_ptr = gemm_ptr->MakeInvokerPointer();
DeviceMem gemm_desc_workspace(gemm_ptr->GetWorkSpaceSize(argument_ptr.get()));
gemm_ptr->SetWorkSpacePointer(argument_ptr.get(), gemm_desc_workspace.GetDeviceBuffer());
if(gemm_ptr->IsSupportedArgument(argument_ptr.get()))
{
std::string gemm_name = gemm_ptr->GetTypeString();
float ave_time =
invoker_ptr->Run(argument_ptr.get(), StreamConfig{nullptr, time_kernel});
std::size_t flop = 0, num_btype = 0;
for(std::size_t i = 0; i < gemm_descs.size(); i++)
{
flop += std::size_t(2) * Ms[i] * Ns[i] * Ks[i];
num_btype += sizeof(ADataType) * Ms[i] * Ks[i] + sizeof(BDataType) * Ks[i] * Ns[i] +
sizeof(CDataType) * Ms[i] * Ns[i];
}
float tflops = static_cast<float>(flop) / 1.E9 / ave_time;
float gb_per_sec = num_btype / 1.E6 / ave_time;
std::cout << "Perf: " << std::setw(10) << ave_time << " ms, " << tflops << " TFlops, "
<< gb_per_sec << " GB/s, " << gemm_name << std::endl;
if(tflops > best_tflops)
{
best_gemm_name = gemm_name;
best_tflops = tflops;
best_ave_time = ave_time;
best_gb_per_sec = gb_per_sec;
}
if(do_verification)
{
for(std::size_t i = 0; i < gemm_descs.size(); i++)
{
c_device_buf[i]->FromDevice(c_m_n_device_results[i].mData.data());
Tensor<CDataType> c_m_n_host_result(
f_host_tensor_descriptor(Ms[i], Ns[i], StrideCs[i], CLayout{}));
using ReferenceGemmInstance =
ck::tensor_operation::host::ReferenceGemm<ADataType,
BDataType,
CDataType,
AccDataType,
AElementOp,
BElementOp,
CElementOp>;
auto ref_gemm = ReferenceGemmInstance{};
auto ref_invoker = ref_gemm.MakeInvoker();
auto ref_argument = ref_gemm.MakeArgument(a_m_k[i],
b_k_n[i],
c_m_n_host_result,
a_element_op,
b_element_op,
c_element_op);
ref_invoker.Run(ref_argument);
bool group_pass =
ck::utils::check_err(c_m_n_device_results[i], c_m_n_host_result);
pass = pass && group_pass;
std::cout << "group: " << i << " verification result: " << std::boolalpha
<< group_pass << std::endl;
if(do_log)
{
LogRangeAsType<float>(std::cout << "a : ", a_m_k[i].mData, ",")
<< std::endl;
LogRangeAsType<float>(std::cout << "b: ", b_k_n[i].mData, ",") << std::endl;
LogRangeAsType<float>(
std::cout << "c_device: ", c_m_n_device_results[i].mData, ",")
<< std::endl;
LogRangeAsType<float>(
std::cout << "c_host : ", c_m_n_host_result.mData, ",")
<< std::endl;
}
}
}
}
else
{
std::cout << "does not support this GEMM problem" << std::endl;
}
}
if(do_verification)
{
std::cout << "Verification: " << (pass ? "SUCCESS" : "FAILURE") << std::endl;
}
std::cout << "Best Perf: " << best_ave_time << " ms, " << best_tflops << " TFlops, "
<< best_gb_per_sec << " GB/s, " << best_gemm_name << std::endl;
return pass;
}
} // namespace profiler
} // namespace ck
profiler/include/profiler/profile_grouped_gemm_impl.hpp
View file @ 9fb64dae
......@@ -108,11 +108,6 @@ bool profile_grouped_gemm_impl(int do_verification,
const auto b_element_op = BElementOp{};
const auto c_element_op = CElementOp{};
// if(do_verification)
// {
// }
using DeviceMemPtr = std::unique_ptr<DeviceMem>;
std::vector<DeviceMemPtr> a_device_buf, b_device_buf, c_device_buf;
......@@ -285,7 +280,7 @@ bool profile_grouped_gemm_impl(int do_verification,
<< best_gb_per_sec << " GB/s, " << best_gemm_name << std::endl;
return pass;
} // namespace profiler
}
} // namespace profiler
} // namespace ck
profiler/src/CMakeLists.txt
View file @ 9fb64dae
......@@ -29,6 +29,7 @@ set(PROFILER_SOURCES
profile_batchnorm_fwd.cpp
profile_batchnorm_bwd.cpp
profile_batchnorm_infer.cpp
profile_grouped_gemm_fastgelu.cpp
)
set(PROFILER_EXECUTABLE ckProfiler)
......@@ -68,4 +69,5 @@ target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_normalization_instan
target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_softmax_instance)
target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_reduce_instance)
target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_batchnorm_instance)
target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_grouped_gemm_fastgelu_instance)
rocm_install(TARGETS ${PROFILER_EXECUTABLE} COMPONENT profiler)
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