aboutsummaryrefslogtreecommitdiff
path: root/src/gpu/cl/operators/ClGemmConv2d.cpp
diff options
context:
space:
mode:
Diffstat (limited to 'src/gpu/cl/operators/ClGemmConv2d.cpp')
-rw-r--r--src/gpu/cl/operators/ClGemmConv2d.cpp50
1 files changed, 18 insertions, 32 deletions
diff --git a/src/gpu/cl/operators/ClGemmConv2d.cpp b/src/gpu/cl/operators/ClGemmConv2d.cpp
index 682477e4ea..5620471ff9 100644
--- a/src/gpu/cl/operators/ClGemmConv2d.cpp
+++ b/src/gpu/cl/operators/ClGemmConv2d.cpp
@@ -1,5 +1,5 @@
/*
- * Copyright (c) 2017-2021 Arm Limited.
+ * Copyright (c) 2017-2021, 2023 Arm Limited.
*
* SPDX-License-Identifier: MIT
*
@@ -54,14 +54,14 @@ namespace opencl
{
ClGemmConv2d::ClGemmConv2d()
: _weights_reshape_kernel(nullptr), _im2col_kernel(nullptr), _mm_gemm(nullptr), _mm_gemmlowp(nullptr), _col2im_kernel(nullptr), _activation_kernel(nullptr), _im2col_output(), _weights_reshaped(),
- _gemm_output(), _skip_im2col(false), _skip_col2im(false), _is_quantized(false), _fuse_activation(true), _append_bias(false), _is_prepared(false), _use_post_ops(false), _aux_mem(AuxTensorIdx::Count)
+ _gemm_output(), _skip_im2col(false), _skip_col2im(false), _is_quantized(false), _fuse_activation(true), _append_bias(false), _is_prepared(false), _aux_mem(AuxTensorIdx::Count)
{
}
ClGemmConv2d::~ClGemmConv2d() = default;
void ClGemmConv2d::configure_mm(const ClCompileContext &compile_context, const ITensorInfo *src, ITensorInfo *weights, ITensorInfo *biases, ITensorInfo *dst,
const GEMMLowpOutputStageInfo &gemmlowp_output_stage,
- int gemm_3d_depth, const ActivationLayerInfo &act_info, const experimental::PostOpList<ITensorInfo *> &post_ops)
+ int gemm_3d_depth, const ActivationLayerInfo &act_info)
{
ARM_COMPUTE_ERROR_ON_NULLPTR(src, weights);
ARM_COMPUTE_ERROR_THROW_ON(validate_mm(src, weights, biases, dst, gemmlowp_output_stage, gemm_3d_depth, _skip_im2col, act_info));
@@ -76,14 +76,12 @@ void ClGemmConv2d::configure_mm(const ClCompileContext &compile_context, const I
false, // fast_math
false, // fp_mixed_precision
true, // broadcast_bias
- act_info, // activation_info
- post_ops // post ops
+ act_info // activation_info
);
TensorInfo tmp_src{ *src };
if(_is_quantized)
{
- ARM_COMPUTE_ERROR_ON_MSG(post_ops.size() > 0, "ClGemmConv2d quantized types do not support post ops");
// Since we need negative offsets for computing convolution, we need to change QuantizationInfo()
// Extract and negate input and weights offset
const QuantizationInfo input_quantization_info = src->quantization_info();
@@ -118,7 +116,7 @@ void ClGemmConv2d::configure_mm(const ClCompileContext &compile_context, const I
}
Status ClGemmConv2d::validate_mm(const ITensorInfo *src, const ITensorInfo *weights, const ITensorInfo *biases, const ITensorInfo *dst,
- const GEMMLowpOutputStageInfo &gemmlowp_output_stage, int gemm_3d_depth, bool skip_im2col, const ActivationLayerInfo &act_info, const experimental::PostOpList<ITensorInfo *> &post_ops)
+ const GEMMLowpOutputStageInfo &gemmlowp_output_stage, int gemm_3d_depth, bool skip_im2col, const ActivationLayerInfo &act_info)
{
const bool is_quantized = is_data_type_quantized_asymmetric(src->data_type());
@@ -132,13 +130,11 @@ Status ClGemmConv2d::validate_mm(const ITensorInfo *src, const ITensorInfo *weig
false, // fast_math
false, // fp_mixed_precision
true, // broadcast_bias
- act_info, // activation_info
- post_ops // post ops
+ act_info // activation_info
);
if(is_quantized)
{
- ARM_COMPUTE_RETURN_ERROR_ON_MSG(post_ops.size() > 0, "ClGemmConv2d quantized types do not support post ops");
// Since we need negative offsets for computing convolution, we need to change QuantizationInfo()
// Extract and negate input and weights offset
const QuantizationInfo input_quantization_info = src->quantization_info();
@@ -189,19 +185,18 @@ void ClGemmConv2d::configure(const CLCompileContext &compile_context, ITensorInf
// Only for quantize there are few cases where we cannot fuse the activation function in GEMM
_fuse_activation = true;
- _use_post_ops = conv2d_info.post_ops.size() > 0;
const ITensorInfo *gemm_input_to_use = src;
ITensorInfo *gemm_output_to_use = dst;
// Get parameters from conv_info
- unsigned int stride_x = 0;
- unsigned int stride_y = 0;
+ unsigned int stride_x = 0;
+ unsigned int stride_y = 0;
std::tie(stride_x, stride_y) = conv2d_info.conv_info.stride();
// Get convolved dimensions
- unsigned int conv_w = 0;
- unsigned int conv_h = 0;
+ unsigned int conv_w = 0;
+ unsigned int conv_h = 0;
std::tie(conv_w, conv_h) = scaled_dimensions(src->dimension(idx_width),
src->dimension(idx_height),
kernel_width,
@@ -318,11 +313,10 @@ void ClGemmConv2d::configure(const CLCompileContext &compile_context, ITensorInf
// In case of NHWC, we need to run GEMM3D (gemm_3d_depth != 0) in order to avoid reshaping the output matrix
const unsigned int gemm_3d_depth = (data_layout == DataLayout::NHWC) ? conv_h : 0;
- configure_mm(compile_context, gemm_input_to_use, &_weights_reshaped, biases_to_use, gemm_output_to_use, gemmlowp_output_stage, gemm_3d_depth, conv2d_info.act_info, conv2d_info.post_ops);
+ configure_mm(compile_context, gemm_input_to_use, &_weights_reshaped, biases_to_use, gemm_output_to_use, gemmlowp_output_stage, gemm_3d_depth, conv2d_info.act_info);
if(!_skip_col2im)
{
- ARM_COMPUTE_ERROR_ON_MSG(conv2d_info.post_ops.size() > 0, "ClGemmConv2d does not support post ops with col2im operation"); // Post ops must be performed after every other op
// Set the GPU target for col2im
_col2im_kernel = std::make_unique<opencl::kernels::ClCol2ImKernel>();
_col2im_kernel->set_target(CLScheduler::get().target());
@@ -334,8 +328,7 @@ void ClGemmConv2d::configure(const CLCompileContext &compile_context, ITensorInf
ARM_COMPUTE_ERROR_ON_MSG((dst->dimension(idx_width) != conv_w) || (dst->dimension(idx_height) != conv_h),
"Output shape does not match the expected one");
- // Disable running of activation kernel if post ops are used
- if(!_fuse_activation && !_use_post_ops)
+ if(!_fuse_activation)
{
_activation_kernel = std::make_unique<opencl::kernels::ClActivationKernel>();
_activation_kernel->configure(compile_context, dst, nullptr, conv2d_info.act_info);
@@ -383,15 +376,11 @@ Status ClGemmConv2d::validate(const ITensorInfo *src, const ITensorInfo *weights
const bool is_quantized = is_data_type_quantized_asymmetric(data_type);
const bool skip_im2col = (data_layout == DataLayout::NHWC && kernel_width == 1 && kernel_height == 1 && conv2d_info.conv_info.stride().first == 1
&& conv2d_info.conv_info.stride().second == 1);
- const bool skip_col2im = data_layout == DataLayout::NHWC;
- bool fuse_activation = true;
- bool use_post_ops = conv2d_info.post_ops.size() > 0;
+ const bool skip_col2im = data_layout == DataLayout::NHWC;
+ bool fuse_activation = true;
ARM_COMPUTE_RETURN_ERROR_ON((weights->dimension(idx_channel) * conv2d_info.num_groups) != src->dimension(idx_channel));
ARM_COMPUTE_RETURN_ERROR_ON(weights->num_dimensions() > 4);
- ARM_COMPUTE_RETURN_ERROR_ON_MSG(!skip_im2col
- && conv2d_info.post_ops.size() > 0,
- "ClGemmConv2d does not support post ops with col2im or im2col operation"); // Post ops must be performed after every other op
// Validate biases
if(biases != nullptr)
@@ -520,8 +509,7 @@ Status ClGemmConv2d::validate(const ITensorInfo *src, const ITensorInfo *weights
// In case of NHWC, we need to run GEMM3D (gemm_3d_depth != 0) in order to avoid reshaping the output matrix
const unsigned int gemm_3d_depth = (data_layout == DataLayout::NHWC) ? conv_h : 0;
- ARM_COMPUTE_RETURN_ON_ERROR(validate_mm(gemm_input_to_use, weights_to_use, biases_to_use, gemm_output_to_use, gemmlowp_output_stage, gemm_3d_depth, skip_im2col, conv2d_info.act_info,
- conv2d_info.post_ops));
+ ARM_COMPUTE_RETURN_ON_ERROR(validate_mm(gemm_input_to_use, weights_to_use, biases_to_use, gemm_output_to_use, gemmlowp_output_stage, gemm_3d_depth, skip_im2col, conv2d_info.act_info));
// Validate Col2Im
if(!skip_col2im)
@@ -530,8 +518,7 @@ Status ClGemmConv2d::validate(const ITensorInfo *src, const ITensorInfo *weights
}
// Validate Activation Layer
- // Disable running (thus validation) of activation kernel if post ops are used
- if(!fuse_activation && !use_post_ops)
+ if(!fuse_activation)
{
ARM_COMPUTE_RETURN_ON_ERROR(kernels::ClActivationKernel::validate(dst, nullptr, conv2d_info.act_info));
}
@@ -600,8 +587,7 @@ void ClGemmConv2d::run(ITensorPack &tensors)
}
//Run Activation Layer if we cannot fuse in GEMM
- // Disable running of activation kernel if post ops are used
- if(!_fuse_activation && !_use_post_ops)
+ if(!_fuse_activation)
{
ITensorPack pack =
{
@@ -620,7 +606,7 @@ void ClGemmConv2d::prepare(ITensorPack &tensors)
ICLTensor *weights_reshaped_p = utils::cast::polymorphic_downcast<ICLTensor *>(tensors.get_tensor(offset_int_vec(WeightsReshaped)));
CLAuxTensorHandler weights_reshaped(_weights_reshaped, *weights_reshaped_p);
auto weights = tensors.get_const_tensor(TensorType::ACL_SRC_1);
- ITensorPack pack =
+ ITensorPack pack =
{
{ TensorType::ACL_SRC, weights },
{ TensorType::ACL_DST, weights_reshaped.get() }
generated by cgit v1.2.1 (git 2.23.0) at 2024-05-19 22:40:09 +0000