From ea7de7babc319e2fa31c5e1c986e48d6c5370689 Mon Sep 17 00:00:00 2001 From: Giorgio Arena Date: Thu, 10 Dec 2020 16:49:39 +0000 Subject: Enable FFT for FP16 Resolves: COMPMID-4051 Change-Id: I0c0bf97212dd281c19d5081e6247e7dc0c23cd6b Signed-off-by: Giorgio Arena Reviewed-on: https://review.mlplatform.org/c/ml/ComputeLibrary/+/4687 Tested-by: Arm Jenkins Reviewed-by: Gian Marco Iodice Comments-Addressed: Arm Jenkins --- src/runtime/CL/functions/CLConvolutionLayer.cpp | 6 +++--- src/runtime/CL/functions/CLFFT1D.cpp | 2 +- src/runtime/CL/functions/CLFFT2D.cpp | 1 + src/runtime/CL/functions/CLFFTConvolutionLayer.cpp | 17 ++++++++++------- 4 files changed, 15 insertions(+), 11 deletions(-) (limited to 'src/runtime/CL/functions') diff --git a/src/runtime/CL/functions/CLConvolutionLayer.cpp b/src/runtime/CL/functions/CLConvolutionLayer.cpp index edd9298d26..5bfbc7ce57 100644 --- a/src/runtime/CL/functions/CLConvolutionLayer.cpp +++ b/src/runtime/CL/functions/CLConvolutionLayer.cpp @@ -88,7 +88,7 @@ void CLConvolutionLayer::configure(const CLCompileContext &compile_context, ICLT case ConvolutionMethod::FFT: { auto f = std::make_unique(_memory_manager); - f->configure(compile_context, input, weights, biases, output, conv_info, act_info); + f->configure(compile_context, input, weights, biases, output, conv_info, act_info, enable_fast_math); _function = std::move(f); break; } @@ -131,7 +131,7 @@ Status CLConvolutionLayer::validate(const ITensorInfo *input, const ITensorInfo case ConvolutionMethod::FFT: { // Validate FFT-based convolution layer - ARM_COMPUTE_RETURN_ON_ERROR(CLFFTConvolutionLayer::validate(input, weights, nullptr, output, conv_info, act_info)); + ARM_COMPUTE_RETURN_ON_ERROR(CLFFTConvolutionLayer::validate(input, weights, nullptr, output, conv_info, act_info, enable_fast_math)); break; } default: @@ -204,7 +204,7 @@ ConvolutionMethod CLConvolutionLayer::get_convolution_method(const ITensorInfo * { return ConvolutionMethod::DIRECT; } - if((weights->dimension(idx_h) > 7) && (input->dimension(idx_c) > output->dimension(idx_c)) && (CLFFTConvolutionLayer::validate(input, weights, nullptr, output, conv_info, act_info))) + if((weights->dimension(idx_h) > 7) && (input->dimension(idx_c) > output->dimension(idx_c)) && (CLFFTConvolutionLayer::validate(input, weights, nullptr, output, conv_info, act_info, enable_fast_math))) { return ConvolutionMethod::FFT; } diff --git a/src/runtime/CL/functions/CLFFT1D.cpp b/src/runtime/CL/functions/CLFFT1D.cpp index c434b4e570..cf136dc75e 100644 --- a/src/runtime/CL/functions/CLFFT1D.cpp +++ b/src/runtime/CL/functions/CLFFT1D.cpp @@ -118,7 +118,7 @@ void CLFFT1D::configure(const CLCompileContext &compile_context, const ICLTensor Status CLFFT1D::validate(const ITensorInfo *input, const ITensorInfo *output, const FFT1DInfo &config) { ARM_COMPUTE_RETURN_ERROR_ON_NULLPTR(input, output); - ARM_COMPUTE_RETURN_ERROR_ON(input->data_type() != DataType::F32); + ARM_COMPUTE_RETURN_ERROR_ON_DATA_TYPE_NOT_IN(input, DataType::F16, DataType::F32); ARM_COMPUTE_RETURN_ERROR_ON(input->num_channels() != 1 && input->num_channels() != 2); ARM_COMPUTE_RETURN_ERROR_ON(std::set({ 0, 1 }).count(config.axis) == 0); diff --git a/src/runtime/CL/functions/CLFFT2D.cpp b/src/runtime/CL/functions/CLFFT2D.cpp index 1d444bb15d..e0497ca6dc 100644 --- a/src/runtime/CL/functions/CLFFT2D.cpp +++ b/src/runtime/CL/functions/CLFFT2D.cpp @@ -67,6 +67,7 @@ void CLFFT2D::configure(const CLCompileContext &compile_context, const ICLTensor Status CLFFT2D::validate(const ITensorInfo *input, const ITensorInfo *output, const FFT2DInfo &config) { ARM_COMPUTE_RETURN_ERROR_ON_NULLPTR(input, output); + ARM_COMPUTE_RETURN_ERROR_ON_DATA_TYPE_NOT_IN(input, DataType::F16, DataType::F32); // Create intermediate tensor info TensorInfo first_pass_tensor(input->clone()->set_is_resizable(true).reset_padding().set_num_channels(2)); diff --git a/src/runtime/CL/functions/CLFFTConvolutionLayer.cpp b/src/runtime/CL/functions/CLFFTConvolutionLayer.cpp index 97b64b24f3..45e74df703 100644 --- a/src/runtime/CL/functions/CLFFTConvolutionLayer.cpp +++ b/src/runtime/CL/functions/CLFFTConvolutionLayer.cpp @@ -104,14 +104,17 @@ CLFFTConvolutionLayer::CLFFTConvolutionLayer(std::shared_ptr mem } void CLFFTConvolutionLayer::configure(ICLTensor *input, const ICLTensor *weights, const ICLTensor *biases, ICLTensor *output, const PadStrideInfo &conv_info, - const ActivationLayerInfo &act_info) + const ActivationLayerInfo &act_info, bool enable_fast_math) { - configure(CLKernelLibrary::get().get_compile_context(), input, weights, biases, output, conv_info, act_info); + configure(CLKernelLibrary::get().get_compile_context(), input, weights, biases, output, conv_info, act_info, enable_fast_math); } void CLFFTConvolutionLayer::configure(const CLCompileContext &compile_context, ICLTensor *input, const ICLTensor *weights, const ICLTensor *biases, ICLTensor *output, const PadStrideInfo &conv_info, - const ActivationLayerInfo &act_info) + const ActivationLayerInfo &act_info, bool enable_fast_math) { + ARM_COMPUTE_UNUSED(enable_fast_math); + ARM_COMPUTE_ERROR_THROW_ON(CLFFTConvolutionLayer::validate(input->info(), weights->info(), biases != nullptr ? biases->info() : nullptr, output->info(), conv_info, act_info, enable_fast_math)); + _original_weights = weights; _original_bias = biases; @@ -265,9 +268,10 @@ void CLFFTConvolutionLayer::configure(const CLCompileContext &compile_context, I } Status CLFFTConvolutionLayer::validate(const ITensorInfo *input, const ITensorInfo *weights, const ITensorInfo *biases, const ITensorInfo *output, const PadStrideInfo &conv_info, - const ActivationLayerInfo &act_info) + const ActivationLayerInfo &act_info, bool enable_fast_math) { - ARM_COMPUTE_RETURN_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(input, 1, DataType::F32); + ARM_COMPUTE_RETURN_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(input, 1, DataType::F16, DataType::F32); + ARM_COMPUTE_RETURN_ERROR_ON((input->data_type() == DataType::F16) && !enable_fast_math); ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_DATA_TYPES(input, weights); // Get indices for the width and height @@ -287,9 +291,8 @@ Status CLFFTConvolutionLayer::validate(const ITensorInfo *input, const ITensorIn // Validate biases if(biases != nullptr) { - const size_t idx_channels = get_data_layout_dimension_index(input->data_layout(), DataLayoutDimension::CHANNEL); ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_DATA_TYPES(input, biases); - ARM_COMPUTE_RETURN_ERROR_ON(input->tensor_shape()[idx_channels] != biases->tensor_shape().x()); + ARM_COMPUTE_RETURN_ERROR_ON(weights->tensor_shape()[3] != biases->tensor_shape().x()); } // Checks performed when output is configured -- cgit v1.2.1