diff options
author | Michele Di Giorgio <michele.digiorgio@arm.com> | 2020-06-10 16:34:50 +0100 |
---|---|---|
committer | Michele Di Giorgio <michele.digiorgio@arm.com> | 2020-06-12 09:26:10 +0000 |
commit | 11c562c94fa6a0399aff798bfd970ed8c1942730 (patch) | |
tree | 6eb2ca628659b1b579a2f99a6dd598d5e6ee5ea2 /src/core | |
parent | d004a7a707feab36e51f51cfc9eb2cb70729d5ad (diff) | |
download | ComputeLibrary-11c562c94fa6a0399aff798bfd970ed8c1942730.tar.gz |
COMPMID-3527: QLSTM should allow nullptr for optional projection bias
Change-Id: I5568ca1e9383f3960886f211d8320fc4525a8804
Signed-off-by: Michele Di Giorgio <michele.digiorgio@arm.com>
Reviewed-on: https://review.mlplatform.org/c/ml/ComputeLibrary/+/3326
Reviewed-by: James Conroy <james.conroy@arm.com>
Reviewed-by: Georgios Pinitas <georgios.pinitas@arm.com>
Comments-Addressed: Arm Jenkins <bsgcomp@arm.com>
Tested-by: Arm Jenkins <bsgcomp@arm.com>
Diffstat (limited to 'src/core')
-rw-r--r-- | src/core/CL/kernels/CLElementwiseOperationKernel.cpp | 12 | ||||
-rw-r--r-- | src/core/NEON/kernels/NEArithmeticAdditionKernel.cpp | 15 |
2 files changed, 22 insertions, 5 deletions
diff --git a/src/core/CL/kernels/CLElementwiseOperationKernel.cpp b/src/core/CL/kernels/CLElementwiseOperationKernel.cpp index 00a97d50e9..4e7d3b3753 100644 --- a/src/core/CL/kernels/CLElementwiseOperationKernel.cpp +++ b/src/core/CL/kernels/CLElementwiseOperationKernel.cpp @@ -93,9 +93,13 @@ Status validate_arguments_with_float_only_supported_rules(const ITensorInfo &inp Status validate_arguments_with_arithmetic_rules(const ITensorInfo &input1, const ITensorInfo &input2, const ITensorInfo &output) { ARM_COMPUTE_RETURN_ERROR_ON_F16_UNSUPPORTED(&input1); - ARM_COMPUTE_RETURN_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(&input1, 1, DataType::U8, DataType::QASYMM8, DataType::QASYMM8_SIGNED, DataType::S16, DataType::QSYMM16, DataType::F16, DataType::F32); + ARM_COMPUTE_RETURN_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(&input1, 1, DataType::U8, DataType::QASYMM8, DataType::QASYMM8_SIGNED, + DataType::S16, DataType::QSYMM16, DataType::F16, + DataType::S32, DataType::F32); ARM_COMPUTE_RETURN_ERROR_ON_F16_UNSUPPORTED(&input2); - ARM_COMPUTE_RETURN_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(&input2, 1, DataType::U8, DataType::QASYMM8, DataType::QASYMM8_SIGNED, DataType::S16, DataType::QSYMM16, DataType::F16, DataType::F32); + ARM_COMPUTE_RETURN_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(&input2, 1, DataType::U8, DataType::QASYMM8, DataType::QASYMM8_SIGNED, + DataType::S16, DataType::QSYMM16, DataType::F16, + DataType::S32, DataType::F32); const bool is_quantized = is_data_type_quantized(input1.data_type()) || is_data_type_quantized(input2.data_type()); if(is_quantized) @@ -119,7 +123,9 @@ Status validate_arguments_with_arithmetic_rules(const ITensorInfo &input1, const if(output.total_size() > 0) { ARM_COMPUTE_RETURN_ERROR_ON_F16_UNSUPPORTED(&output); - ARM_COMPUTE_RETURN_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(&output, 1, DataType::U8, DataType::QASYMM8, DataType::QASYMM8_SIGNED, DataType::S16, DataType::QSYMM16, DataType::F16, DataType::F32); + ARM_COMPUTE_RETURN_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(&output, 1, DataType::U8, DataType::QASYMM8, DataType::QASYMM8_SIGNED, + DataType::S16, DataType::QSYMM16, DataType::F16, + DataType::S32, DataType::F32); ARM_COMPUTE_RETURN_ERROR_ON_MSG((output.data_type() == DataType::U8) && ((input1.data_type() != DataType::U8) || (input2.data_type() != DataType::U8)), "Output can only be U8 if both inputs are U8"); ARM_COMPUTE_RETURN_ERROR_ON_MSG(detail::have_different_dimensions(out_shape, output.tensor_shape(), 0), diff --git a/src/core/NEON/kernels/NEArithmeticAdditionKernel.cpp b/src/core/NEON/kernels/NEArithmeticAdditionKernel.cpp index f8ee578ef8..3878c764a6 100644 --- a/src/core/NEON/kernels/NEArithmeticAdditionKernel.cpp +++ b/src/core/NEON/kernels/NEArithmeticAdditionKernel.cpp @@ -815,8 +815,12 @@ Status validate_arguments(const ITensorInfo &input1, const ITensorInfo &input2, ARM_COMPUTE_UNUSED(policy); ARM_COMPUTE_RETURN_ERROR_ON_CPU_F16_UNSUPPORTED(&input1); - ARM_COMPUTE_RETURN_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(&input1, 1, DataType::U8, DataType::QASYMM8, DataType::QASYMM8_SIGNED, DataType::S16, DataType::QSYMM16, DataType::F16, DataType::F32); - ARM_COMPUTE_RETURN_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(&input2, 1, DataType::U8, DataType::QASYMM8, DataType::QASYMM8_SIGNED, DataType::S16, DataType::QSYMM16, DataType::F16, DataType::F32); + ARM_COMPUTE_RETURN_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(&input1, 1, DataType::U8, DataType::QASYMM8, DataType::QASYMM8_SIGNED, + DataType::S16, DataType::QSYMM16, DataType::F16, + DataType::S32, DataType::F32); + ARM_COMPUTE_RETURN_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(&input2, 1, DataType::U8, DataType::QASYMM8, DataType::QASYMM8_SIGNED, + DataType::S16, DataType::QSYMM16, DataType::F16, + DataType::S32, DataType::F32); const TensorShape out_shape = TensorShape::broadcast_shape(input1.tensor_shape(), input2.tensor_shape()); @@ -834,6 +838,7 @@ Status validate_arguments(const ITensorInfo &input1, const ITensorInfo &input2, && !(input1.data_type() == DataType::U8 && input2.data_type() == DataType::S16 && output.data_type() == DataType::S16) && !(input1.data_type() == DataType::S16 && input2.data_type() == DataType::U8 && output.data_type() == DataType::S16) && !(input1.data_type() == DataType::S16 && input2.data_type() == DataType::S16 && output.data_type() == DataType::S16) + && !(input1.data_type() == DataType::S32 && input2.data_type() == DataType::S32 && output.data_type() == DataType::S32) && !(input1.data_type() == DataType::F32 && input2.data_type() == DataType::F32 && output.data_type() == DataType::F32) && !(input1.data_type() == DataType::F16 && input2.data_type() == DataType::F16 && output.data_type() == DataType::F16) && !(input1.data_type() == DataType::QASYMM8 && input2.data_type() == DataType::QASYMM8 && output.data_type() == DataType::QASYMM8) @@ -862,6 +867,10 @@ std::pair<Status, Window> validate_and_configure_window(ITensorInfo &input1, ITe { set_format_if_unknown(output, Format::S16); } + if(input1.data_type() == DataType::S32 || input2.data_type() == DataType::S32) + { + set_format_if_unknown(output, Format::S32); + } else if(input1.data_type() == DataType::F16 || input2.data_type() == DataType::F16) { set_format_if_unknown(output, Format::F16); @@ -926,6 +935,8 @@ void NEArithmeticAdditionKernel::configure(const ITensor *input1, const ITensor { "add_saturate_U8_U8_S16", &add_U8_U8_S16 }, { "add_wrap_S16_S16_S16", &add_same<int16_t> }, { "add_saturate_S16_S16_S16", &add_same<int16_t> }, + { "add_wrap_S32_S32_S32", &add_same<int32_t> }, + { "add_saturate_S32_S32_S32", &add_same<int32_t> }, { "add_wrap_F32_F32_F32", &add_same<float> }, { "add_saturate_F32_F32_F32", &add_same<float> }, #ifdef __ARM_FEATURE_FP16_VECTOR_ARITHMETIC |