From ae1c9fee8bb6073abf2a419af3266f23dbab82a9 Mon Sep 17 00:00:00 2001 From: Suhail Munshi Date: Wed, 14 Apr 2021 12:16:49 +0100 Subject: Added S32 Integer support to DIV operator in CLElementWiseOperations with Tests Partially Resolves : COMPMID-3793 Signed-off-by: Suhail Munshi Change-Id: I14d6884c34f33a6caee11fc1230f9d2d3ae6c4c1 Reviewed-on: https://review.mlplatform.org/c/ml/ComputeLibrary/+/5425 Reviewed-by: Giorgio Arena Reviewed-by: Manuel Bottini Tested-by: Arm Jenkins Comments-Addressed: Arm Jenkins --- src/core/CL/cl_kernels/elementwise_operation.cl | 9 +++++-- src/core/gpu/cl/kernels/ClElementwiseKernel.cpp | 35 +++++++++++++++++++++++-- 2 files changed, 40 insertions(+), 4 deletions(-) (limited to 'src') diff --git a/src/core/CL/cl_kernels/elementwise_operation.cl b/src/core/CL/cl_kernels/elementwise_operation.cl index 11e0612205..a5b486a882 100644 --- a/src/core/CL/cl_kernels/elementwise_operation.cl +++ b/src/core/CL/cl_kernels/elementwise_operation.cl @@ -1,5 +1,5 @@ /* - * Copyright (c) 2018-2020 Arm Limited. + * Copyright (c) 2018-2021 Arm Limited. * * SPDX-License-Identifier: MIT * @@ -38,10 +38,15 @@ #define MAX(x, y) max(x, y) #define MIN(x, y) min(x, y) #define SQUARED_DIFF(x, y) (x - y) * (x - y) -#define DIV(x, y) (x / y) #define POWER(x, y) pow(x, y) #define PRELU(x, y) (select(y * x, x, CONVERT((x > (DATA_TYPE_OUT)0), SELECT_VEC_DATA_TYPE(DATA_TYPE_OUT, VEC_SIZE_OUT)))) +#if defined(S32) +#define DIV(x, y) CONVERT(floor(CONVERT(x , VEC_DATA_TYPE(float, VEC_SIZE_OUT)) / CONVERT(y , VEC_DATA_TYPE(float, VEC_SIZE_OUT))), VEC_DATA_TYPE(DATA_TYPE_OUT, VEC_SIZE_OUT)); +#else /* S32 */ +#define DIV(x, y) (x / y) +#endif /* S32 */ + #define AND(x, y) (CONVERT((x && y), VEC_DATA_TYPE(DATA_TYPE_OUT, VEC_SIZE_OUT)) & ((VEC_DATA_TYPE(DATA_TYPE_OUT, VEC_SIZE_OUT))1)) #define OR(x, y) (CONVERT((x || y), VEC_DATA_TYPE(DATA_TYPE_OUT, VEC_SIZE_OUT)) & ((VEC_DATA_TYPE(DATA_TYPE_OUT, VEC_SIZE_OUT))1)) diff --git a/src/core/gpu/cl/kernels/ClElementwiseKernel.cpp b/src/core/gpu/cl/kernels/ClElementwiseKernel.cpp index 8f12eb2215..335ee9c392 100644 --- a/src/core/gpu/cl/kernels/ClElementwiseKernel.cpp +++ b/src/core/gpu/cl/kernels/ClElementwiseKernel.cpp @@ -98,6 +98,29 @@ Status validate_arguments_with_float_only_supported_rules(const ITensorInfo &src return Status{}; } +Status validate_arguments_divide_operation(const ITensorInfo* src1, const ITensorInfo* src2, const ITensorInfo* dst) +{ + ARM_COMPUTE_RETURN_ERROR_ON_NULLPTR(src1, src2, dst); + ARM_COMPUTE_RETURN_ERROR_ON_F16_UNSUPPORTED(src1); + ARM_COMPUTE_RETURN_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(src1, 1, DataType::F16, DataType::F32, DataType::S32); + ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_DATA_TYPES(src1, src2); + + const TensorShape out_shape = TensorShape::broadcast_shape(src1->tensor_shape(), src2->tensor_shape()); + + ARM_COMPUTE_RETURN_ERROR_ON_MSG(out_shape.total_size() == 0, "Inputs are not broadcast compatible"); + + // Validate in case of configured dst + if(dst->total_size() > 0) + { + ARM_COMPUTE_RETURN_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(dst, 1, DataType::F16, DataType::F32, DataType::S32); + ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_DATA_TYPES(src1, dst); + ARM_COMPUTE_RETURN_ERROR_ON_MSG(detail::have_different_dimensions(out_shape, dst->tensor_shape(), 0), + "Wrong shape for dst"); + } + + return Status{}; +} + Status validate_arguments_with_arithmetic_rules(const ITensorInfo &src1, const ITensorInfo &src2, const ITensorInfo &dst) { ARM_COMPUTE_RETURN_ERROR_ON_F16_UNSUPPORTED(&src1); @@ -180,6 +203,8 @@ CLBuildOptions generate_build_options_with_arithmetic_rules(const ITensorInfo &s build_opts.add_option("-DSCALE_IN2=" + float_to_string_with_full_precision(iq2info.scale)); build_opts.add_option("-DSCALE_OUT=" + float_to_string_with_full_precision(oqinfo.scale)); } + build_opts.add_option_if(src1.data_type() == DataType::S32, "-DS32"); + return build_opts; } @@ -459,9 +484,15 @@ void ClArithmeticKernel::configure(const ClCompileContext &compile_context, Arit Status ClArithmeticKernel::validate(ArithmeticOperation op, const ITensorInfo *src1, const ITensorInfo *src2, const ITensorInfo *dst, const ActivationLayerInfo &act_info) { ARM_COMPUTE_RETURN_ERROR_ON_NULLPTR(src1, src2, dst); - if(op == ArithmeticOperation::DIV || op == ArithmeticOperation::POWER) + if(op == ArithmeticOperation::DIV) { - // Division and Power operators don't support integer arithmetic + // Partial integer support S32/F32/F16 + ARM_COMPUTE_RETURN_ON_ERROR(validate_arguments_divide_operation(src1, src2, dst)); + ARM_COMPUTE_RETURN_ON_ERROR(validate_and_configure_window_for_division(*src1->clone(), *src2->clone(), *dst->clone()).first); + } + else if(op == ArithmeticOperation::POWER) + { + // Power operators doesn't support integer arithmetic ARM_COMPUTE_RETURN_ON_ERROR(validate_arguments_with_float_only_supported_rules(*src1, *src2, *dst)); ARM_COMPUTE_RETURN_ON_ERROR(validate_and_configure_window_for_division(*src1->clone(), *src2->clone(), *dst->clone()).first); } -- cgit v1.2.1