From 6b612f5fa1fee9528f2f87491fe7edb3887d9817 Mon Sep 17 00:00:00 2001 From: Michele Di Giorgio Date: Thu, 5 Sep 2019 12:30:22 +0100 Subject: COMPMID-2310: CLGenerateProposalsLayer: support for QASYMM8 Change-Id: I48b77e09857cd43f9498d28e8f4bf346e3d7110d Signed-off-by: Michele Di Giorgio Reviewed-on: https://review.mlplatform.org/c/1969 Reviewed-by: Pablo Marquez Comments-Addressed: Arm Jenkins Tested-by: Arm Jenkins --- .../CPPBoxWithNonMaximaSuppressionLimit.cpp | 69 +++++++++++----------- 1 file changed, 34 insertions(+), 35 deletions(-) (limited to 'src/runtime/CPP/functions/CPPBoxWithNonMaximaSuppressionLimit.cpp') diff --git a/src/runtime/CPP/functions/CPPBoxWithNonMaximaSuppressionLimit.cpp b/src/runtime/CPP/functions/CPPBoxWithNonMaximaSuppressionLimit.cpp index 158f45a320..782771bc50 100644 --- a/src/runtime/CPP/functions/CPPBoxWithNonMaximaSuppressionLimit.cpp +++ b/src/runtime/CPP/functions/CPPBoxWithNonMaximaSuppressionLimit.cpp @@ -30,9 +30,10 @@ namespace arm_compute { namespace { -void dequantize_tensor(const ITensor *input, ITensor *output, DataType data_type) +void dequantize_tensor(const ITensor *input, ITensor *output) { - const UniformQuantizationInfo qinfo = input->info()->quantization_info().uniform(); + const UniformQuantizationInfo qinfo = input->info()->quantization_info().uniform(); + const DataType data_type = input->info()->data_type(); Window window; window.use_tensor_dimensions(input->info()->tensor_shape()); @@ -60,9 +61,10 @@ void dequantize_tensor(const ITensor *input, ITensor *output, DataType data_type } } -void quantize_tensor(const ITensor *input, ITensor *output, DataType data_type) +void quantize_tensor(const ITensor *input, ITensor *output) { - const UniformQuantizationInfo qinfo = input->info()->quantization_info().uniform(); + const UniformQuantizationInfo qinfo = output->info()->quantization_info().uniform(); + const DataType data_type = output->info()->data_type(); Window window; window.use_tensor_dimensions(input->info()->tensor_shape()); @@ -102,7 +104,6 @@ CPPBoxWithNonMaximaSuppressionLimit::CPPBoxWithNonMaximaSuppressionLimit(std::sh _classes(), _batch_splits_out(), _keeps(), - _keeps_size(), _scores_in_f32(), _boxes_in_f32(), _batch_splits_in_f32(), @@ -111,7 +112,6 @@ CPPBoxWithNonMaximaSuppressionLimit::CPPBoxWithNonMaximaSuppressionLimit(std::sh _classes_f32(), _batch_splits_out_f32(), _keeps_f32(), - _keeps_size_f32(), _is_qasymm8(false) { } @@ -119,7 +119,7 @@ CPPBoxWithNonMaximaSuppressionLimit::CPPBoxWithNonMaximaSuppressionLimit(std::sh void CPPBoxWithNonMaximaSuppressionLimit::configure(const ITensor *scores_in, const ITensor *boxes_in, const ITensor *batch_splits_in, ITensor *scores_out, ITensor *boxes_out, ITensor *classes, ITensor *batch_splits_out, ITensor *keeps, ITensor *keeps_size, const BoxNMSLimitInfo info) { - ARM_COMPUTE_ERROR_ON_NULLPTR(scores_in, boxes_in, batch_splits_in, scores_out, boxes_out, classes); + ARM_COMPUTE_ERROR_ON_NULLPTR(scores_in, boxes_in, scores_out, boxes_out, classes); _is_qasymm8 = scores_in->info()->data_type() == DataType::QASYMM8; @@ -131,20 +131,22 @@ void CPPBoxWithNonMaximaSuppressionLimit::configure(const ITensor *scores_in, co _classes = classes; _batch_splits_out = batch_splits_out; _keeps = keeps; - _keeps_size = keeps_size; if(_is_qasymm8) { // Manage intermediate buffers _memory_group.manage(&_scores_in_f32); _memory_group.manage(&_boxes_in_f32); - _memory_group.manage(&_batch_splits_in_f32); _memory_group.manage(&_scores_out_f32); _memory_group.manage(&_boxes_out_f32); _memory_group.manage(&_classes_f32); _scores_in_f32.allocator()->init(scores_in->info()->clone()->set_data_type(DataType::F32)); _boxes_in_f32.allocator()->init(boxes_in->info()->clone()->set_data_type(DataType::F32)); - _batch_splits_in_f32.allocator()->init(batch_splits_in->info()->clone()->set_data_type(DataType::F32)); + if(batch_splits_in != nullptr) + { + _memory_group.manage(&_batch_splits_in_f32); + _batch_splits_in_f32.allocator()->init(batch_splits_in->info()->clone()->set_data_type(DataType::F32)); + } _scores_out_f32.allocator()->init(scores_out->info()->clone()->set_data_type(DataType::F32)); _boxes_out_f32.allocator()->init(boxes_out->info()->clone()->set_data_type(DataType::F32)); _classes_f32.allocator()->init(classes->info()->clone()->set_data_type(DataType::F32)); @@ -158,15 +160,11 @@ void CPPBoxWithNonMaximaSuppressionLimit::configure(const ITensor *scores_in, co _memory_group.manage(&_keeps_f32); _keeps_f32.allocator()->init(keeps->info()->clone()->set_data_type(DataType::F32)); } - if(keeps_size != nullptr) - { - _memory_group.manage(&_keeps_size_f32); - _keeps_size_f32.allocator()->init(keeps_size->info()->clone()->set_data_type(DataType::F32)); - } - _box_with_nms_limit_kernel.configure(&_scores_in_f32, &_boxes_in_f32, &_batch_splits_in_f32, &_scores_out_f32, &_boxes_out_f32, &_classes_f32, + _box_with_nms_limit_kernel.configure(&_scores_in_f32, &_boxes_in_f32, (batch_splits_in != nullptr) ? &_batch_splits_in_f32 : nullptr, + &_scores_out_f32, &_boxes_out_f32, &_classes_f32, (batch_splits_out != nullptr) ? &_batch_splits_out_f32 : nullptr, (keeps != nullptr) ? &_keeps_f32 : nullptr, - (keeps_size != nullptr) ? &_keeps_size_f32 : nullptr, info); + keeps_size, info); } else { @@ -177,7 +175,10 @@ void CPPBoxWithNonMaximaSuppressionLimit::configure(const ITensor *scores_in, co { _scores_in_f32.allocator()->allocate(); _boxes_in_f32.allocator()->allocate(); - _batch_splits_in_f32.allocator()->allocate(); + if(_batch_splits_in != nullptr) + { + _batch_splits_in_f32.allocator()->allocate(); + } _scores_out_f32.allocator()->allocate(); _boxes_out_f32.allocator()->allocate(); _classes_f32.allocator()->allocate(); @@ -189,17 +190,13 @@ void CPPBoxWithNonMaximaSuppressionLimit::configure(const ITensor *scores_in, co { _keeps_f32.allocator()->allocate(); } - if(keeps_size != nullptr) - { - _keeps_size_f32.allocator()->allocate(); - } } } Status validate(const ITensorInfo *scores_in, const ITensorInfo *boxes_in, const ITensorInfo *batch_splits_in, const ITensorInfo *scores_out, const ITensorInfo *boxes_out, const ITensorInfo *classes, const ITensorInfo *batch_splits_out, const ITensorInfo *keeps, const ITensorInfo *keeps_size, const BoxNMSLimitInfo info) { - ARM_COMPUTE_RETURN_ERROR_ON_NULLPTR(scores_in, boxes_in, batch_splits_in, scores_out, boxes_out, classes); + ARM_COMPUTE_RETURN_ERROR_ON_NULLPTR(scores_in, boxes_in, scores_out, boxes_out, classes); ARM_COMPUTE_RETURN_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(scores_in, 1, DataType::QASYMM8, DataType::F16, DataType::F32); const bool is_qasymm8 = scores_in->data_type() == DataType::QASYMM8; @@ -218,31 +215,33 @@ Status validate(const ITensorInfo *scores_in, const ITensorInfo *boxes_in, const void CPPBoxWithNonMaximaSuppressionLimit::run() { + // Acquire all the temporaries + MemoryGroupResourceScope scope_mg(_memory_group); + if(_is_qasymm8) { - dequantize_tensor(_scores_in, &_scores_in_f32, _scores_in->info()->data_type()); - dequantize_tensor(_boxes_in, &_boxes_in_f32, _boxes_in->info()->data_type()); - dequantize_tensor(_batch_splits_in, &_batch_splits_in_f32, _batch_splits_in->info()->data_type()); + dequantize_tensor(_scores_in, &_scores_in_f32); + dequantize_tensor(_boxes_in, &_boxes_in_f32); + if(_batch_splits_in != nullptr) + { + dequantize_tensor(_batch_splits_in, &_batch_splits_in_f32); + } } Scheduler::get().schedule(&_box_with_nms_limit_kernel, Window::DimY); if(_is_qasymm8) { - quantize_tensor(&_scores_out_f32, _scores_out, _scores_out->info()->data_type()); - quantize_tensor(&_boxes_out_f32, _boxes_out, _boxes_out->info()->data_type()); - quantize_tensor(&_classes_f32, _classes, _classes->info()->data_type()); + quantize_tensor(&_scores_out_f32, _scores_out); + quantize_tensor(&_boxes_out_f32, _boxes_out); + quantize_tensor(&_classes_f32, _classes); if(_batch_splits_out != nullptr) { - quantize_tensor(&_batch_splits_out_f32, _batch_splits_out, _batch_splits_out->info()->data_type()); + quantize_tensor(&_batch_splits_out_f32, _batch_splits_out); } if(_keeps != nullptr) { - quantize_tensor(&_keeps_f32, _keeps, _keeps->info()->data_type()); - } - if(_keeps_size != nullptr) - { - quantize_tensor(&_keeps_size_f32, _keeps_size, _keeps_size->info()->data_type()); + quantize_tensor(&_keeps_f32, _keeps); } } } -- cgit v1.2.1