aboutsummaryrefslogtreecommitdiff
path: root/src
diff options
context:
space:
mode:
Diffstat (limited to 'src')
-rw-r--r--src/core/CL/cl_kernels/softmax_layer.cl4
-rw-r--r--src/core/CL/cl_kernels/softmax_layer_quantized.cl4
-rw-r--r--src/core/NEON/NEMath.inl3
-rw-r--r--src/graph/GraphManager.cpp2
-rw-r--r--src/graph/backends/CL/CLFunctionsFactory.cpp1
-rw-r--r--src/runtime/NEON/functions/NEGEMMConvolutionLayer.cpp2
6 files changed, 2 insertions, 14 deletions
diff --git a/src/core/CL/cl_kernels/softmax_layer.cl b/src/core/CL/cl_kernels/softmax_layer.cl
index 01f5de47cf..4d2d89dd73 100644
--- a/src/core/CL/cl_kernels/softmax_layer.cl
+++ b/src/core/CL/cl_kernels/softmax_layer.cl
@@ -1,5 +1,5 @@
/*
- * Copyright (c) 2017-2020 Arm Limited.
+ * Copyright (c) 2017-2021 Arm Limited.
*
* SPDX-License-Identifier: MIT
*
@@ -306,7 +306,6 @@ __kernel void softmax_layer_max_shift_exp_sum_parallel(
}
#ifdef NON_MULTIPLE_OF_GRID_SIZE
// How many work-items needed to complete the computation.
- //TODO: Optimize this calculation (avoid %).
int boundary_workitems = (SRC_WIDTH % (GRID_SIZE * VECTOR_SIZE)) / VECTOR_SIZE;
if(lid < boundary_workitems)
{
@@ -417,7 +416,6 @@ __kernel void softmax_layer_max_shift_exp_sum_parallel(
sum1D += data;
}
#ifdef NON_MULTIPLE_OF_GRID_SIZE
- //TODO: Optimize the calculation (avoid %).
boundary_workitems = (SRC_WIDTH % (GRID_SIZE * VECTOR_SIZE)) / VECTOR_SIZE;
if(lid < boundary_workitems)
{
diff --git a/src/core/CL/cl_kernels/softmax_layer_quantized.cl b/src/core/CL/cl_kernels/softmax_layer_quantized.cl
index b7a6e00dfa..4d5006d804 100644
--- a/src/core/CL/cl_kernels/softmax_layer_quantized.cl
+++ b/src/core/CL/cl_kernels/softmax_layer_quantized.cl
@@ -1,5 +1,5 @@
/*
- * Copyright (c) 2017-2020 Arm Limited.
+ * Copyright (c) 2017-2021 Arm Limited.
*
* SPDX-License-Identifier: MIT
*
@@ -324,7 +324,6 @@ __kernel void softmax_layer_max_shift_exp_sum_quantized_parallel(
}
#ifdef NON_MULTIPLE_OF_GRID_SIZE
// How many work-items needed to complete the computation.
- //TODO: Optimize this calculation (avoid %).
int boundary_workitems = (SRC_WIDTH % (GRID_SIZE * VECTOR_SIZE)) / VECTOR_SIZE;
if(lid < boundary_workitems)
{
@@ -429,7 +428,6 @@ __kernel void softmax_layer_max_shift_exp_sum_quantized_parallel(
sum1D = sum1D + select(0, data_fp, data_diff >= (VEC_INT)(DIFF_MIN));
}
#ifdef NON_MULTIPLE_OF_GRID_SIZE
- //TODO: Optimize the calculation (avoid %).
boundary_workitems = (SRC_WIDTH % (GRID_SIZE * VECTOR_SIZE)) / VECTOR_SIZE;
if(lid < boundary_workitems)
{
diff --git a/src/core/NEON/NEMath.inl b/src/core/NEON/NEMath.inl
index da9d038139..29df5433bb 100644
--- a/src/core/NEON/NEMath.inl
+++ b/src/core/NEON/NEMath.inl
@@ -495,7 +495,6 @@ inline float16x8_t vtaylor_polyq_f16(float16x8_t x, const std::array<float16x8_t
inline float16x8_t vexpq_f16(float16x8_t x)
{
- // TODO (COMPMID-1535) : Revisit FP16 approximations
const float32x4_t x_high = vcvt_f32_f16(vget_high_f16(x));
const float32x4_t x_low = vcvt_f32_f16(vget_low_f16(x));
@@ -505,7 +504,6 @@ inline float16x8_t vexpq_f16(float16x8_t x)
inline float16x8_t vlogq_f16(float16x8_t x)
{
- // TODO (COMPMID-1535) : Revisit FP16 approximations
const float32x4_t x_high = vcvt_f32_f16(vget_high_f16(x));
const float32x4_t x_low = vcvt_f32_f16(vget_low_f16(x));
@@ -515,7 +513,6 @@ inline float16x8_t vlogq_f16(float16x8_t x)
inline float16x8_t vpowq_f16(float16x8_t val, float16x8_t n)
{
- // TODO (giaiod01) - COMPMID-1535
float32x4_t n0_f32 = vcvt_f32_f16(vget_low_f16(n));
float32x4_t n1_f32 = vcvt_f32_f16(vget_high_f16(n));
float32x4_t val0_f32 = vcvt_f32_f16(vget_low_f16(val));
diff --git a/src/graph/GraphManager.cpp b/src/graph/GraphManager.cpp
index ab7aac6230..e357f10401 100644
--- a/src/graph/GraphManager.cpp
+++ b/src/graph/GraphManager.cpp
@@ -55,7 +55,6 @@ void GraphManager::finalize_graph(Graph &graph, GraphContext &ctx, PassManager &
pm.run_type(graph, IGraphMutator::MutationType::IR);
// Force target to all graph construct
- // TODO (COMPMID-2014) : Support heterogeneous execution
Target forced_target = target;
// In case CLVK is selected, use the CL backend and
@@ -77,7 +76,6 @@ void GraphManager::finalize_graph(Graph &graph, GraphContext &ctx, PassManager &
force_target_to_graph(graph, forced_target);
// Setup backend context
- // TODO (COMPMID-2014) : Setup all backends needed by the graph
setup_requested_backend_context(ctx, forced_target);
// Configure all tensors
diff --git a/src/graph/backends/CL/CLFunctionsFactory.cpp b/src/graph/backends/CL/CLFunctionsFactory.cpp
index a3e7261721..1cd3f9f9c7 100644
--- a/src/graph/backends/CL/CLFunctionsFactory.cpp
+++ b/src/graph/backends/CL/CLFunctionsFactory.cpp
@@ -83,7 +83,6 @@ struct CLFusedLayerTypes
using FuseBatchNormalization = CLFuseBatchNormalization;
};
-// TODO (isagot01): Remove once we support heterogeneous scheduling at function level
/** Wrapper for the CPP Function in the OpenCL backend **/
class CPPWrapperFunction : public IFunction
{
diff --git a/src/runtime/NEON/functions/NEGEMMConvolutionLayer.cpp b/src/runtime/NEON/functions/NEGEMMConvolutionLayer.cpp
index 900330e4a6..2876c254fa 100644
--- a/src/runtime/NEON/functions/NEGEMMConvolutionLayer.cpp
+++ b/src/runtime/NEON/functions/NEGEMMConvolutionLayer.cpp
@@ -524,7 +524,6 @@ Status NEGEMMConvolutionLayer::validate(const ITensorInfo *input, const ITensorI
{
// Create tensor info for im2col reshaped inputs
// For CPU, the batch size is on the fourth dimension
- // TODO (giaiod01): Auto-initialize the output shape of im2col COMPMID-1482
TensorShape shape_im2col = input->tensor_shape();
shape_im2col.set(0, mat_weights_rows);
shape_im2col.set(1, conv_w * conv_h);
@@ -532,7 +531,6 @@ Status NEGEMMConvolutionLayer::validate(const ITensorInfo *input, const ITensorI
im2col_reshaped_info = TensorInfo(shape_im2col, 1, data_type);
im2col_reshaped_info.set_quantization_info(input->quantization_info());
-
ARM_COMPUTE_RETURN_ON_ERROR(NEIm2ColKernel::validate(input, &im2col_reshaped_info, Size2D(kernel_width, kernel_height), conv_info, append_bias, dilation));
gemm_input_to_use = &im2col_reshaped_info;
}