COMPMID-3722: Remove OpenCL padding: CLGEMMLowpOffsetContributionKernel

COMPMID-3723: Remove OpenCL padding: CLGEMMLowpOffsetContributionOutputStageKernel

Change-Id: Iac265c2ac4c5749352daa311279a3b8c60ac3b3d
Signed-off-by: Michele Di Giorgio <michele.digiorgio@arm.com>
Reviewed-on: https://review.mlplatform.org/c/ml/ComputeLibrary/+/4228
Tested-by: Arm Jenkins <bsgcomp@arm.com>
Reviewed-by: Giorgio Arena <giorgio.arena@arm.com>
Reviewed-by: Georgios Pinitas <georgios.pinitas@arm.com>
Comments-Addressed: Arm Jenkins <bsgcomp@arm.com>

1 files changed, 13 insertions, 67 deletions

diff --git a/src/core/CL/kernels/CLGEMMLowpOffsetContributionOutputStageKernel.cpp b/src/core/CL/kernels/CLGEMMLowpOffsetContributionOutputStageKernel.cpp
index 85285d6704..afa7bdbfdf 100644
--- a/src/core/CL/kernels/CLGEMMLowpOffsetContributionOutputStageKernel.cpp
+++ b/src/core/CL/kernels/CLGEMMLowpOffsetContributionOutputStageKernel.cpp
@@ -34,9 +34,6 @@
 #include "src/core/helpers/WindowHelpers.h"
 #include "support/StringSupport.h"
 
-#include <cstddef>
-#include <cstdint>
-
 namespace arm_compute
 {
 namespace
@@ -119,52 +116,6 @@ Status validate_arguments(const ITensorInfo *mm_result, const ITensorInfo *vecto
 
     return Status{};
 }
-
-std::pair<Status, Window> validate_and_configure_window(ITensorInfo *mm_result, ITensorInfo *vector_sum_col, ITensorInfo *vector_sum_row, ITensorInfo *bias, ITensorInfo *output,
-                                                        int32_t a_offset, int32_t b_offset, const GEMMLowpOutputStageInfo &output_stage, ITensorInfo *output_multipliers, ITensorInfo *output_shifts)
-{
-    constexpr unsigned int num_elems_processed_per_iteration = 4;
-    bool                   window_changed                    = false;
-
-    // Auto initialize the output
-    auto_init_if_empty(*output, mm_result->clone()->set_data_type(output_stage.output_data_type));
-
-    // Configure kernel window
-    Window win = calculate_max_window(*mm_result, Steps(num_elems_processed_per_iteration));
-
-    AccessWindowHorizontal mm_result_access(mm_result, 0, num_elems_processed_per_iteration);
-    window_changed = window_changed || update_window_and_padding(win, mm_result_access);
-
-    AccessWindowHorizontal output_access(output, 0, num_elems_processed_per_iteration);
-    window_changed = window_changed || update_window_and_padding(win, output_access);
-
-    if(a_offset != 0)
-    {
-        AccessWindowHorizontal vector_sum_col_access(vector_sum_col, 0, num_elems_processed_per_iteration);
-        window_changed = window_changed || update_window_and_padding(win, vector_sum_col_access);
-    }
-    if(b_offset != 0)
-    {
-        AccessWindowStatic vector_sum_row_access(vector_sum_row, 0, 0, vector_sum_row->dimension(0), 0); // NOLINT
-        window_changed = window_changed || update_window_and_padding(win, vector_sum_row_access);
-    }
-
-    if(bias != nullptr)
-    {
-        AccessWindowStatic bias_access(bias, 0, 0, ceil_to_multiple(bias->dimension(0), num_elems_processed_per_iteration), bias->tensor_shape()[1]);
-        window_changed = window_changed || update_window_and_padding(win, bias_access);
-    }
-
-    if(output_multipliers->dimension(0) > 1)
-    {
-        AccessWindowHorizontal output_multipliers_access(output_multipliers, 0, num_elems_processed_per_iteration);
-        AccessWindowHorizontal output_shifts_access(output_shifts, 0, num_elems_processed_per_iteration);
-        window_changed = window_changed || update_window_and_padding(win, output_multipliers_access, output_shifts_access);
-    }
-
-    Status err = (window_changed) ? ARM_COMPUTE_CREATE_ERROR(ErrorCode::RUNTIME_ERROR, "Insufficient Padding!") : Status{};
-    return std::make_pair(err, win);
-}
 } // namespace
 
 CLGEMMLowpOffsetContributionOutputStageKernel::CLGEMMLowpOffsetContributionOutputStageKernel()
@@ -201,6 +152,8 @@ void CLGEMMLowpOffsetContributionOutputStageKernel::configure(const CLCompileCon
                                                   a_offset, b_offset, output_stage,
                                                   output_multipliers->info(), output_shifts->info())); // NOLINT
 
+    auto padding_info = get_padding_info({ mm_result, vector_sum_col, vector_sum_row, bias, output, output_multipliers, output_shifts });
+
     const int min = output_stage.gemmlowp_min_bound;
     const int max = output_stage.gemmlowp_max_bound;
 
@@ -218,8 +171,15 @@ void CLGEMMLowpOffsetContributionOutputStageKernel::configure(const CLCompileCon
                                    && mm_result->info()->num_dimensions() > 1
                                    && mm_result->info()->tensor_shape().y() != vector_sum_row->info()->tensor_shape().x();
 
+    // Auto initialize the output
+    auto_init_if_empty(*output->info(), mm_result->info()->clone()->set_data_type(output_stage.output_data_type));
+
+    const unsigned int num_elems_processed_per_iteration = adjust_vec_size(4, mm_result->info()->dimension(0));
+
     // Set the arguments to pass at compile time
     CLBuildOptions build_opts;
+    build_opts.add_option("-DVEC_SIZE=" + support::cpp11::to_string(num_elems_processed_per_iteration));
+    build_opts.add_option("-DVEC_SIZE_LEFTOVER=" + support::cpp11::to_string(mm_result->info()->dimension(0) % num_elems_processed_per_iteration));
 
     // If a_offset == 0, vector_sum_col can be a nullptr
     if(a_offset != 0)
@@ -252,15 +212,8 @@ void CLGEMMLowpOffsetContributionOutputStageKernel::configure(const CLCompileCon
     _kernel = create_kernel(compile_context, kernel_name, build_opts.options());
 
     // Configure kernel window
-    auto win_config = validate_and_configure_window(mm_result->info(),
-                                                    vector_sum_col != nullptr ? vector_sum_col->info() : nullptr,
-                                                    vector_sum_row != nullptr ? vector_sum_row->info() : nullptr,
-                                                    bias != nullptr ? bias->info() : nullptr,
-                                                    output->info(),
-                                                    a_offset, b_offset, output_stage,
-                                                    output_multipliers->info(), output_shifts->info()); // NOLINT
-    ARM_COMPUTE_ERROR_THROW_ON(win_config.first);
-    ICLKernel::configure_internal(win_config.second);
+    Window win = calculate_max_window(*mm_result->info(), Steps(num_elems_processed_per_iteration));
+    ICLKernel::configure_internal(win);
 
     // Set config_id for enabling LWS tuning
     _config_id = kernel_name + "_";
@@ -269,6 +222,8 @@ void CLGEMMLowpOffsetContributionOutputStageKernel::configure(const CLCompileCon
     _config_id += support::cpp11::to_string(mm_result->info()->dimension(1));
     _config_id += "_";
     _config_id += support::cpp11::to_string(mm_result->info()->dimension(2));
+
+    ARM_COMPUTE_ERROR_ON(has_padding_changed(padding_info));
 }
 
 Status CLGEMMLowpOffsetContributionOutputStageKernel::validate(const ITensorInfo *mm_result, const ITensorInfo *vector_sum_col, const ITensorInfo *vector_sum_row, const ITensorInfo *bias,
@@ -276,15 +231,6 @@ Status CLGEMMLowpOffsetContributionOutputStageKernel::validate(const ITensorInfo
                                                                const ITensorInfo *output_multipliers, const ITensorInfo *output_shifts)
 {
     ARM_COMPUTE_RETURN_ON_ERROR(validate_arguments(mm_result, vector_sum_col, vector_sum_row, bias, output, a_offset, b_offset, output_stage, output_multipliers, output_shifts));
-    ARM_COMPUTE_RETURN_ON_ERROR(validate_and_configure_window(mm_result->clone().get(),
-                                                              vector_sum_col != nullptr ? vector_sum_col->clone().get() : nullptr,
-                                                              vector_sum_row != nullptr ? vector_sum_row->clone().get() : nullptr,
-                                                              bias != nullptr ? bias->clone().get() : nullptr,
-                                                              output->clone().get(),
-                                                              a_offset, b_offset, output_stage,
-                                                              output_multipliers->clone().get(), output_shifts->clone().get())
-                                .first); // NOLINT
-
     return Status{};
 }