diff options
author | Gunes Bayir <gunes.bayir@arm.com> | 2023-01-19 15:56:00 +0000 |
---|---|---|
committer | Gunes Bayir <gunes.bayir@arm.com> | 2023-01-24 09:40:01 +0000 |
commit | cc2877368d5e15d9ea89d31c84ec651fc0fffd13 (patch) | |
tree | c57a3a406125b3a31e2d4aff6126ce99f4ade395 /src/dynamic_fusion/sketch/gpu/operators/GpuCast.cpp | |
parent | a6a153817302793732e28b07c3b4046df3f91a60 (diff) | |
download | ComputeLibrary-cc2877368d5e15d9ea89d31c84ec651fc0fffd13.tar.gz |
Change dynamic fusion API to return destination tensor info
The new dynamic fusion API is introduced in the following patch:
https://review.mlplatform.org/c/ml/ComputeLibrary/+/8906
For each operator (except Conv2D, which is migrated in the above patch), we
- remove destination tensor from is_supported, validate and create calls
- make create_op return ITensorInfo* to the intermediate destination object
Affected operators:
- DepthwiseConv2D
- Cast
- Elementwise Ops
- Clamp
- Reshape
- Resize
Resolves: COMPMID-5777
Change-Id: Ib60ec8a5f081752808455d7a7d790f2ed0627059
Signed-off-by: Gunes Bayir <gunes.bayir@arm.com>
Reviewed-on: https://review.mlplatform.org/c/ml/ComputeLibrary/+/8991
Reviewed-by: Ramy Elgammal <ramy.elgammal@arm.com>
Reviewed-by: Jakub Sujak <jakub.sujak@arm.com>
Dynamic-Fusion: Ramy Elgammal <ramy.elgammal@arm.com>
Comments-Addressed: Arm Jenkins <bsgcomp@arm.com>
Tested-by: Arm Jenkins <bsgcomp@arm.com>
Benchmark: Arm Jenkins <bsgcomp@arm.com>
Diffstat (limited to 'src/dynamic_fusion/sketch/gpu/operators/GpuCast.cpp')
-rw-r--r-- | src/dynamic_fusion/sketch/gpu/operators/GpuCast.cpp | 65 |
1 files changed, 41 insertions, 24 deletions
diff --git a/src/dynamic_fusion/sketch/gpu/operators/GpuCast.cpp b/src/dynamic_fusion/sketch/gpu/operators/GpuCast.cpp index 3a5b64ad9c..33c2d43e07 100644 --- a/src/dynamic_fusion/sketch/gpu/operators/GpuCast.cpp +++ b/src/dynamic_fusion/sketch/gpu/operators/GpuCast.cpp @@ -1,5 +1,5 @@ /* - * Copyright (c) 2022 Arm Limited. + * Copyright (c) 2022-2023 Arm Limited. * * SPDX-License-Identifier: MIT * @@ -38,18 +38,22 @@ namespace dynamic_fusion { namespace { -constexpr GpuOperatorType operator_type = GpuOperatorType::Simple; -} -Status GpuCast::is_supported_op(const GpuWorkloadContext &context, - const ITensorInfo *src, - const ITensorInfo *dst, - const CastAttributes &attributes) +Status is_supported_op_helper(const GpuWorkloadContext &context, + const ITensorInfo *src, + const ITensorInfo *dst, + const CastAttributes &attributes) { ARM_COMPUTE_RETURN_ERROR_ON(src == dst); ARM_COMPUTE_RETURN_ERROR_ON_NULLPTR(src, dst); - // Auto initialize dst tensor info - TensorInfo dst_info_to_validate = *dst; + TensorInfo dst_info_to_validate; + const ITensorInfo *dst_info_to_validate_ptr = &dst_info_to_validate; + + if(dst != nullptr) + { + dst_info_to_validate_ptr = dst; + } + auto_init_if_empty(dst_info_to_validate, src->clone()->set_data_type(attributes.data_type())); // Check support level @@ -59,7 +63,7 @@ Status GpuCast::is_supported_op(const GpuWorkloadContext &context, DataType::U8, DataType::S8, DataType::QASYMM8, DataType::QASYMM8_SIGNED, DataType::QSYMM8_PER_CHANNEL, DataType::S16, DataType::U16, DataType::U32, DataType::S32, DataType::F16, DataType::F32); - ARM_COMPUTE_RETURN_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(&dst_info_to_validate, + ARM_COMPUTE_RETURN_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(dst_info_to_validate_ptr, 1, DataType::U8, DataType::S8, DataType::QASYMM8, DataType::S16, DataType::U16, DataType::U32, DataType::S32, DataType::F16, @@ -76,7 +80,7 @@ Status GpuCast::is_supported_op(const GpuWorkloadContext &context, ArgumentPack<ITensorInfo> arguments; arguments.add_const_tensor(ACL_SRC_0, src); - arguments.add_const_tensor(ACL_DST_0, &dst_info_to_validate); + arguments.add_const_tensor(ACL_DST_0, dst_info_to_validate_ptr); ARM_COMPUTE_RETURN_ON_ERROR(ClComponentCast::validate(properties, arguments, attributes, settings)); } } @@ -87,17 +91,27 @@ Status GpuCast::is_supported_op(const GpuWorkloadContext &context, return Status{}; } +constexpr GpuOperatorType operator_type = GpuOperatorType::Simple; +} // namespace + +Status GpuCast::is_supported_op(const GpuWorkloadContext &context, + const ITensorInfo *src, + const CastAttributes &attributes) +{ + return is_supported_op_helper(context, src, nullptr, attributes); +} Status GpuCast::validate_op(const GpuWorkloadSketch &sketch, const ITensorInfo *src, - const ITensorInfo *dst, const CastAttributes &attributes) { - ARM_COMPUTE_RETURN_ERROR_ON_NULLPTR(src, dst); - ARM_COMPUTE_RETURN_ERROR_ON(!src->has_valid_id() || !dst->has_valid_id()); + ARM_COMPUTE_RETURN_ERROR_ON_NULLPTR(src); + ARM_COMPUTE_RETURN_ERROR_ON(!src->has_valid_id()); + + // Refer to GpuConv2d::validate_op() for id-validness of this TensorInfo object + TensorInfo dst_info_to_validate; // Auto initialize dst tensor info - TensorInfo dst_info_to_validate = *dst; auto_init_if_empty(dst_info_to_validate, src->clone()->set_data_type(attributes.data_type())); // Perform fusion test @@ -110,18 +124,19 @@ Status GpuCast::validate_op(const GpuWorkloadSketch &sketch, "Operator fusion test failed. This operator cannot be fused into the workload"); // Check if configuration is supported - return is_supported_op(*sketch.gpu_context(), src, &dst_info_to_validate, attributes); + return is_supported_op_helper(*sketch.gpu_context(), src, &dst_info_to_validate, attributes); } -void GpuCast::create_op(GpuWorkloadSketch &sketch, - ITensorInfo *src, - ITensorInfo *dst, - const CastAttributes &attributes) +ITensorInfo *GpuCast::create_op(GpuWorkloadSketch &sketch, + ITensorInfo *src, + const CastAttributes &attributes) { - // Assert validation - ARM_COMPUTE_ERROR_THROW_ON(GpuCast::validate_op(sketch, src, dst, attributes)); - ARM_COMPUTE_ERROR_ON_NULLPTR(src, dst); - ARM_COMPUTE_LOG_PARAMS(src, dst, attributes); + ARM_COMPUTE_ERROR_ON_NULLPTR(src); + ARM_COMPUTE_LOG_PARAMS(src, attributes); + ARM_COMPUTE_ERROR_THROW_ON(GpuCast::validate_op(sketch, src, attributes)); + + ITensorInfo *dst = sketch.implementation().create_virtual_tensor(); + ARM_COMPUTE_ERROR_ON_NULLPTR(dst); // Auto initialize dst tensor info if empty auto_init_if_empty(*dst, src->clone()->set_data_type(attributes.data_type())); @@ -160,6 +175,8 @@ void GpuCast::create_op(GpuWorkloadSketch &sketch, const Operator op = sketch.implementation().operator_group().new_operator(operator_type, tensors); sketch.implementation().operator_group().add_operator(op); + + return dst; } } // namespace dynamic_fusion |