diff options
Diffstat (limited to 'src/runtime/NEON/functions')
-rw-r--r-- | src/runtime/NEON/functions/NEConvolutionLayer.cpp | 6 | ||||
-rw-r--r-- | src/runtime/NEON/functions/NEDepthwiseConvolutionLayer.cpp | 12 |
2 files changed, 12 insertions, 6 deletions
diff --git a/src/runtime/NEON/functions/NEConvolutionLayer.cpp b/src/runtime/NEON/functions/NEConvolutionLayer.cpp index f790f6a95f..335267522b 100644 --- a/src/runtime/NEON/functions/NEConvolutionLayer.cpp +++ b/src/runtime/NEON/functions/NEConvolutionLayer.cpp @@ -450,11 +450,13 @@ void NEConvolutionLayer::configure(const ITensor *input, const ITensor *weights, // Configure output stage for quantized case if(_is_quantized) { - float multiplier = input->info()->quantization_info().scale * weights->info()->quantization_info().scale / output->info()->quantization_info().scale; + const QuantizationInfo output_quant_info = (output->info()->total_size() == 0) ? input->info()->quantization_info() : output->info()->quantization_info(); + + float multiplier = input->info()->quantization_info().scale * weights->info()->quantization_info().scale / output_quant_info.scale; int output_multiplier, output_shift; quantization::calculate_quantized_multiplier_less_than_one(multiplier, &output_multiplier, &output_shift); _memory_group.manage(&_tmp_output); - _gemmlowp_output_stage.configure(&_gemm_output, biases, &_tmp_output, output_multiplier, output_shift, output->info()->quantization_info().offset); + _gemmlowp_output_stage.configure(&_gemm_output, biases, &_tmp_output, output_multiplier, output_shift, output_quant_info.offset); } // Configure Col2Im diff --git a/src/runtime/NEON/functions/NEDepthwiseConvolutionLayer.cpp b/src/runtime/NEON/functions/NEDepthwiseConvolutionLayer.cpp index d35e3e6026..95fcf8805e 100644 --- a/src/runtime/NEON/functions/NEDepthwiseConvolutionLayer.cpp +++ b/src/runtime/NEON/functions/NEDepthwiseConvolutionLayer.cpp @@ -99,10 +99,12 @@ void NEDepthwiseConvolutionLayer3x3::configure(ITensor *input, const ITensor *we { if(_is_quantized) { - float multiplier = input->info()->quantization_info().scale * weights->info()->quantization_info().scale / output->info()->quantization_info().scale; + const QuantizationInfo output_quant_info = (output->info()->total_size() == 0) ? input->info()->quantization_info() : output->info()->quantization_info(); + + float multiplier = input->info()->quantization_info().scale * weights->info()->quantization_info().scale / output_quant_info.scale; int output_multiplier, output_shift; quantization::calculate_quantized_multiplier_less_than_one(multiplier, &output_multiplier, &output_shift); - _output_stage_kernel.configure(&_accumulator, biases, output, output_multiplier, output_shift, output->info()->quantization_info().offset); + _output_stage_kernel.configure(&_accumulator, biases, output, output_multiplier, output_shift, output_quant_info.offset); _accumulator.allocator()->allocate(); } else @@ -208,10 +210,12 @@ void NEDepthwiseConvolutionLayer::configure(ITensor *input, const ITensor *weigh // Output staged configuration if(_is_quantized) { - float multiplier = input->info()->quantization_info().scale * weights->info()->quantization_info().scale / output->info()->quantization_info().scale; + const QuantizationInfo output_quant_info = (output->info()->total_size() == 0) ? input->info()->quantization_info() : output->info()->quantization_info(); + + float multiplier = input->info()->quantization_info().scale * weights->info()->quantization_info().scale / output_quant_info.scale; int output_multiplier, output_shift; quantization::calculate_quantized_multiplier_less_than_one(multiplier, &output_multiplier, &output_shift); - _output_stage_kernel.configure(&_output_reshaped, biases, output, output_multiplier, output_shift, output->info()->quantization_info().offset); + _output_stage_kernel.configure(&_output_reshaped, biases, output, output_multiplier, output_shift, output_quant_info.offset); _output_reshaped.allocator()->allocate(); } |