ROCm/composable_kernel
1
0
Fork 0
mirror of https://github.com/ROCm/composable_kernel.git synced 2026-05-13 09:45:56 +00:00
Code Issues Packages Projects Releases Wiki Activity

Adding Instances and Examples for FP8-based Scaled Convolution with ReLU Activation and AMAX Reduction. (#1469)

Browse Source 
* Enable CMakePresets build

* Verify Convolution, Scaling and ReLU algorithms.

* Add tensor element-wise scale and type cast operation.

* Reduction implemented but does not work.

* Exploration of Reduction functionality.

* Completed example for Convolution scaled with ReLu activation and AMAX reduction.

* WIP: Add required instances for convolution.

* WIP: Create client example. Implement convolution stage.

* Add elementwise instances.

* Add elementwise scale + convert example.

* Add reduction instances.

* WIP: Client example for AMAX reduction.

* WIP: Add instances for multistage reduction.

* WIP: Implementation of multistage reduction.

* Refactoring.

* Clean up.

* Guard off FP8 instances when the data type is not available.

* Improve output readability.

* Addressing reviewer's comments.
This commit is contained in:
Andriy Roshchenko
2024-08-20 09:30:56 -06:00
committed by GitHub
parent f48529b511
commit a94113a941
17 changed files with 1891 additions and 20 deletions
Download Patch File Download Diff File Expand all files Collapse all files

49
library/src/tensor_operation_instance/gpu/grouped_conv3d_fwd_convscale_relu/xdl/device_grouped_conv3d_fwd_xdl_convscale_relu_ndhwgc_gkzyxc_ndhwgk_f8_instance.cpp
View File

@@ -3,6 +3,7 @@
#include "ck/library/tensor_operation_instance/gpu/grouped_conv_fwd/device_grouped_conv_fwd_xdl_outelementop_instance.hpp"
#include "ck/library/tensor_operation_instance/add_device_operation_instance.hpp"
#include "ck/tensor_operation/gpu/element/combined_element_wise_operation.hpp"
#include "ck/tensor_operation/gpu/element/unary_element_wise_operation.hpp"
namespace ck {
@@ -57,6 +58,54 @@ void add_device_grouped_conv3d_fwd_xdl_convscale_relu_ndhwgc_gkzyxc_ndhwgk_f8_in
ConvScaleRelu>{});
}
namespace ew = ck::tensor_operation::element_wise;
using CombConvScaleRelu = ew::UnaryCombinedOp<ew::Scale, ew::Scale, ew::Relu>;
void add_device_grouped_conv3d_fwd_xdl_combconvscale_relu_ndhwgc_gkzyxc_ndhwgk_f8_f8_f32_instances(
std::vector<std::unique_ptr<DeviceGroupedConvFwdMultipleABD<3,
NDHWGC,
GKZYXC,
ck::Tuple<>,
NDHWGK,
F8,
F8,
ck::Tuple<>,
F32,
PassThrough,
PassThrough,
CombConvScaleRelu,
F8,
F8>>>& instances)
{
add_device_operation_instances(
instances,
device_grouped_conv_fwd_xdl_outelementop_f8_f8_f32_instances<3,
NDHWGC,
GKZYXC,
ck::Tuple<>,
NDHWGK,
ConvFwdDefault,
CombConvScaleRelu>{});
add_device_operation_instances(
instances,
device_grouped_conv_fwd_xdl_outelementop_f8_f8_f32_instances<3,
NDHWGC,
GKZYXC,
ck::Tuple<>,
NDHWGK,
ConvFwd1x1P0,
CombConvScaleRelu>{});
add_device_operation_instances(
instances,
device_grouped_conv_fwd_xdl_outelementop_f8_f8_f32_instances<3,
NDHWGC,
GKZYXC,
ck::Tuple<>,
NDHWGK,
ConvFwd1x1S1P0,
CombConvScaleRelu>{});
}
} // namespace instance
} // namespace device
} // namespace tensor_operation

5
library/src/tensor_operation_instance/gpu/permute_scale/CMakeLists.txt
View File

@@ -1,4 +1,4 @@
add_instance_library(device_permute_scale_instance
add_instance_library(device_permute_scale_instance
device_permute_scale_1d_fp16_instances.cpp
device_permute_scale_2d_fp16_instances.cpp
device_permute_scale_3d_fp16_instances.cpp
@@ -10,4 +10,5 @@ add_instance_library(device_permute_scale_instance
device_permute_scale_3d_fp32_instances.cpp
device_permute_scale_4d_fp32_instances.cpp
device_permute_scale_5d_fp32_instances.cpp
device_permute_scale_6d_fp32_instances.cpp)
device_permute_scale_6d_fp32_instances.cpp
device_permute_scale_6d_fp32_fp8_instances.cpp)

28
library/src/tensor_operation_instance/gpu/permute_scale/device_permute_scale_6d_fp32_fp8_instances.cpp Normal file
View File

@@ -0,0 +1,28 @@
// SPDX-License-Identifier: MIT
// Copyright (c) 2024, Advanced Micro Devices, Inc. All rights reserved.
#include "ck/library/tensor_operation_instance/gpu/permute_scale/device_permute_scale_instances.hpp"
#include "ck/library/tensor_operation_instance/add_device_operation_instance.hpp"
namespace ck {
namespace tensor_operation {
namespace device {
namespace instance {
using Scale = element_wise::Scale;
void add_device_permute_scale_6d_f32_f8_instances(
std::vector<std::unique_ptr<DeviceElementwise<ck::Tuple<F32>, ck::Tuple<F8>, Scale, 6>>>&
instances)
{
#ifdef CK_ENABLE_FP8
add_device_operation_instances(instances, device_permute_scale_f32_f8_instances<6, Scale>{});
#else
ignore = instances;
#endif
}
} // namespace instance
} // namespace device
} // namespace tensor_operation
} // namespace ck

27
library/src/tensor_operation_instance/gpu/reduce/device_reduce_instance_blockwise_f32_f32_f32_amax.cpp
View File

@@ -10,15 +10,24 @@ namespace device {
namespace instance {
// clang-format off
// InDataType | AccDataType | OutDataType | Rank | NumReduceDim | ReduceOperation | InElementwiseOp | AccElementwiseOp | PropagateNan | UseIndex
template void add_device_reduce_instance_blockwise<F32, F32, F32, 4, 3, ReduceAMax, UnaryAbs, PassThrough, false, false>(std::vector<DeviceReducePtr<F32, F32, F32, 4, 3, ReduceAMax, UnaryAbs, PassThrough, false, false>>&);
template void add_device_reduce_instance_blockwise<F32, F32, F32, 4, 4, ReduceAMax, UnaryAbs, PassThrough, false, false>(std::vector<DeviceReducePtr<F32, F32, F32, 4, 4, ReduceAMax, UnaryAbs, PassThrough, false, false>>&);
template void add_device_reduce_instance_blockwise<F32, F32, F32, 4, 1, ReduceAMax, UnaryAbs, PassThrough, false, false>(std::vector<DeviceReducePtr<F32, F32, F32, 4, 1, ReduceAMax, UnaryAbs, PassThrough, false, false>>&);
template void add_device_reduce_instance_blockwise<F32, F32, F32, 2, 1, ReduceAMax, UnaryAbs, PassThrough, false, false>(std::vector<DeviceReducePtr<F32, F32, F32, 2, 1, ReduceAMax, UnaryAbs, PassThrough, false, false>>&);
template void add_device_reduce_instance_blockwise<F32, F32, F32, 4, 3, ReduceAMax, UnaryAbs, PassThrough, false, true>(std::vector<DeviceReducePtr<F32, F32, F32, 4, 3, ReduceAMax, UnaryAbs, PassThrough, false, true>>&);
template void add_device_reduce_instance_blockwise<F32, F32, F32, 4, 4, ReduceAMax, UnaryAbs, PassThrough, false, true>(std::vector<DeviceReducePtr<F32, F32, F32, 4, 4, ReduceAMax, UnaryAbs, PassThrough, false, true>>&);
template void add_device_reduce_instance_blockwise<F32, F32, F32, 4, 1, ReduceAMax, UnaryAbs, PassThrough, false, true>(std::vector<DeviceReducePtr<F32, F32, F32, 4, 1, ReduceAMax, UnaryAbs, PassThrough, false, true>>&);
template void add_device_reduce_instance_blockwise<F32, F32, F32, 2, 1, ReduceAMax, UnaryAbs, PassThrough, false, true>(std::vector<DeviceReducePtr<F32, F32, F32, 2, 1, ReduceAMax, UnaryAbs, PassThrough, false, true>>&);
// InDataType | AccDataType | OutDataType | Rank | NumReduceDim | ReduceOperation | InElementwiseOp | AccElementwiseOp | PropagateNan | UseIndex
template void add_device_reduce_instance_blockwise< F32, F32, F32, 4, 3, ReduceAMax, UnaryAbs, PassThrough, false, false>(std::vector<DeviceReducePtr<F32, F32, F32, 4, 3, ReduceAMax, UnaryAbs, PassThrough, false, false>>&);
template void add_device_reduce_instance_blockwise< F32, F32, F32, 4, 4, ReduceAMax, UnaryAbs, PassThrough, false, false>(std::vector<DeviceReducePtr<F32, F32, F32, 4, 4, ReduceAMax, UnaryAbs, PassThrough, false, false>>&);
template void add_device_reduce_instance_blockwise< F32, F32, F32, 4, 1, ReduceAMax, UnaryAbs, PassThrough, false, false>(std::vector<DeviceReducePtr<F32, F32, F32, 4, 1, ReduceAMax, UnaryAbs, PassThrough, false, false>>&);
template void add_device_reduce_instance_blockwise< F32, F32, F32, 2, 1, ReduceAMax, UnaryAbs, PassThrough, false, false>(std::vector<DeviceReducePtr<F32, F32, F32, 2, 1, ReduceAMax, UnaryAbs, PassThrough, false, false>>&);
template void add_device_reduce_instance_blockwise< F32, F32, F32, 4, 3, ReduceAMax, UnaryAbs, PassThrough, false, true>(std::vector<DeviceReducePtr<F32, F32, F32, 4, 3, ReduceAMax, UnaryAbs, PassThrough, false, true>>&);
template void add_device_reduce_instance_blockwise< F32, F32, F32, 4, 4, ReduceAMax, UnaryAbs, PassThrough, false, true>(std::vector<DeviceReducePtr<F32, F32, F32, 4, 4, ReduceAMax, UnaryAbs, PassThrough, false, true>>&);
template void add_device_reduce_instance_blockwise< F32, F32, F32, 4, 1, ReduceAMax, UnaryAbs, PassThrough, false, true>(std::vector<DeviceReducePtr<F32, F32, F32, 4, 1, ReduceAMax, UnaryAbs, PassThrough, false, true>>&);
template void add_device_reduce_instance_blockwise< F32, F32, F32, 2, 1, ReduceAMax, UnaryAbs, PassThrough, false, true>(std::vector<DeviceReducePtr<F32, F32, F32, 2, 1, ReduceAMax, UnaryAbs, PassThrough, false, true>>&);
template void add_device_reduce_instance_blockwise< F32, F32, F32, 6, 6, ReduceAMax, UnaryAbs, PassThrough, true, false>(std::vector<DeviceReducePtr<F32, F32, F32, 6, 6, ReduceAMax, UnaryAbs, PassThrough, true, false>>&);
template void add_device_reduce_instance_blockwise< F32, F32, F32, 5, 5, ReduceAMax, UnaryAbs, PassThrough, true, false>(std::vector<DeviceReducePtr<F32, F32, F32, 5, 5, ReduceAMax, UnaryAbs, PassThrough, true, false>>&);
template void add_device_reduce_instance_blockwise< F32, F32, F32, 4, 4, ReduceAMax, UnaryAbs, PassThrough, true, false>(std::vector<DeviceReducePtr<F32, F32, F32, 4, 4, ReduceAMax, UnaryAbs, PassThrough, true, false>>&);
template void add_device_reduce_instance_blockwise< F32, F32, F32, 6, 3, ReduceAMax, UnaryAbs, PassThrough, true, false>(std::vector<DeviceReducePtr<F32, F32, F32, 6, 3, ReduceAMax, UnaryAbs, PassThrough, true, false>>&);
template void add_device_reduce_instance_blockwise< F32, F32, F32, 5, 3, ReduceAMax, UnaryAbs, PassThrough, true, false>(std::vector<DeviceReducePtr<F32, F32, F32, 5, 3, ReduceAMax, UnaryAbs, PassThrough, true, false>>&);
template void add_device_reduce_instance_blockwise< F32, F32, F32, 4, 3, ReduceAMax, UnaryAbs, PassThrough, true, false>(std::vector<DeviceReducePtr<F32, F32, F32, 4, 3, ReduceAMax, UnaryAbs, PassThrough, true, false>>&);
template void add_device_reduce_instance_blockwise< F32, F32, F32, 3, 3, ReduceAMax, PassThrough, PassThrough, true, false>(std::vector<DeviceReducePtr<F32, F32, F32, 3, 3, ReduceAMax, PassThrough, PassThrough, true, false>>&);
template void add_device_reduce_instance_blockwise< F32, F32, F32, 2, 2, ReduceAMax, PassThrough, PassThrough, true, false>(std::vector<DeviceReducePtr<F32, F32, F32, 2, 2, ReduceAMax, PassThrough, PassThrough, true, false>>&);
template void add_device_reduce_instance_blockwise< F32, F32, F32, 1, 1, ReduceAMax, PassThrough, PassThrough, true, false>(std::vector<DeviceReducePtr<F32, F32, F32, 1, 1, ReduceAMax, PassThrough, PassThrough, true, false>>&);
// clang-format on
} // namespace instance