Remove padding in FP Cl Gemm kernels

* Remove rhs and bias padding in ClGemmMatrixMultiplyNativeKernel
* Rework ClGemmMatrixMultiplyReshapedOnlyRHSKernel to use the same
  padding boundary condition as the other kernels

Partially resolves COMPMID-4435

Change-Id: I1c17af9cca0b5cb3be087ce160948b7b0e62d297
Signed-off-by: SiCongLi <sicong.li@arm.com>
Reviewed-on: https://review.mlplatform.org/c/ml/ComputeLibrary/+/6549
Reviewed-by: Gian Marco Iodice <gianmarco.iodice@arm.com>
Comments-Addressed: Arm Jenkins <bsgcomp@arm.com>
Tested-by: Arm Jenkins <bsgcomp@arm.com>

1 files changed, 8 insertions, 41 deletions

diff --git a/src/gpu/cl/kernels/ClGemmMatrixMultiplyNativeKernel.cpp b/src/gpu/cl/kernels/ClGemmMatrixMultiplyNativeKernel.cpp
index 6c872fd48c..e389ce5b0c 100644
--- a/src/gpu/cl/kernels/ClGemmMatrixMultiplyNativeKernel.cpp
+++ b/src/gpu/cl/kernels/ClGemmMatrixMultiplyNativeKernel.cpp
@@ -119,15 +119,12 @@ std::pair<Status, Window> validate_and_configure_window(ITensorInfo *src0, ITens
                                                         const GEMMRHSMatrixInfo &rhs_info,
                                                         const GEMMKernelInfo &gemm_info, ElementsProcessed &num_elements_processed)
 {
+    ARM_COMPUTE_UNUSED(src0, src1, src2);
     unsigned int &num_elems_processed_per_iteration_x = num_elements_processed[0];
     unsigned int &num_elems_processed_per_iteration_y = num_elements_processed[1];
     bool          reinterpret_input_as_3d             = gemm_info.reinterpret_input_as_3d;
     bool          reinterpret_output_as_3d            = gemm_info.depth_output_gemm3d != 0;
 
-    Window win{};
-    Window win_out{};
-    bool   window_changed = false;
-
     // In case both input and dst have to be reinterpreted as 3D tensors,
     // force reinterpret_input_as_3d and reinterpret_output_as_3d to be false.
     if(reinterpret_input_as_3d == reinterpret_output_as_3d)
@@ -135,9 +132,6 @@ std::pair<Status, Window> validate_and_configure_window(ITensorInfo *src0, ITens
         reinterpret_output_as_3d = false;
     }
 
-    // dst tensor auto initialization if not yet initialized
-    auto_init_if_empty(*dst, src0->clone()->set_tensor_shape(misc::shape_calculator::compute_mm_shape(*src0, *src1, gemm_info)));
-
     TensorInfo tmp_info(*dst);
 
     if(reinterpret_output_as_3d)
@@ -153,44 +147,14 @@ std::pair<Status, Window> validate_and_configure_window(ITensorInfo *src0, ITens
     num_elems_processed_per_iteration_x = rhs_info.n0;
     num_elems_processed_per_iteration_y = lhs_info.m0;
 
-    win     = calculate_max_window(tmp_info, Steps(num_elems_processed_per_iteration_x, num_elems_processed_per_iteration_y));
-    win_out = calculate_max_window(*dst, Steps(num_elems_processed_per_iteration_x, num_elems_processed_per_iteration_y));
-
-    AccessWindowStatic src0_access(src0, 0, 0,
-                                   src0->dimension(0),
-                                   src0->dimension(1));
-    AccessWindowStatic src1_access(src1, 0, 0,
-                                   ceil_to_multiple(src1->dimension(0), num_elems_processed_per_iteration_x),
-                                   src1->dimension(1));
-    AccessWindowStatic dst_access(dst, 0, 0,
-                                  dst->dimension(0),
-                                  dst->dimension(1));
-
-    if(src2 != nullptr)
-    {
-        const int bias_processed_per_iteration_x = num_elems_processed_per_iteration_x;
-
-        AccessWindowStatic src2_access(src2, 0, 0,
-                                       ceil_to_multiple(src2->dimension(0), bias_processed_per_iteration_x),
-                                       src2->dimension(1));
-
-        window_changed = update_window_and_padding(win, src0_access, src1_access, src2_access) || // window used by the execute_window_loop
-                         update_window_and_padding(win_out, dst_access);                          // window used to update the padding requirements of dst tensor
-    }
-    else
-    {
-        window_changed = update_window_and_padding(win, src0_access, src1_access) || // window used by the execute_window_loop
-                         update_window_and_padding(win_out, dst_access);             // window used to update the padding requirements of dst tensor
-    }
+    Window win = calculate_max_window(tmp_info, Steps(num_elems_processed_per_iteration_x, num_elems_processed_per_iteration_y));
 
     // Collapse along the Z direction
     // This collapse needs to be here in order to tune the Z dimension of LWS
-    Window             collapsed             = win;
     const unsigned int dimension_to_collapse = std::min(static_cast<unsigned int>(dst->num_dimensions()), 2u);
-    collapsed                                = win.collapse(win, dimension_to_collapse);
+    Window             collapsed             = win.collapse(win, dimension_to_collapse);
 
-    Status err = (window_changed) ? ARM_COMPUTE_CREATE_ERROR(ErrorCode::RUNTIME_ERROR, "Insufficient Padding!") : Status{};
-    return std::make_pair(err, collapsed);
+    return std::make_pair(Status{}, collapsed);
 }
 } // namespace
 
@@ -208,7 +172,10 @@ void ClGemmMatrixMultiplyNativeKernel::configure(const CLCompileContext &compile
 
     ARM_COMPUTE_ERROR_THROW_ON(validate_arguments(src0, src1, src2, dst, alpha, beta, lhs_info, rhs_info, gemm_info));
 
-    auto padding_info         = get_padding_info({ src0, dst });
+    // dst tensor auto initialization if not yet initialized
+    auto_init_if_empty(*dst, src0->clone()->set_tensor_shape(misc::shape_calculator::compute_mm_shape(*src0, *src1, gemm_info)));
+
+    auto padding_info         = get_padding_info({ src0, src1, src2, dst });
     _reinterpret_input_as_3d  = gemm_info.reinterpret_input_as_3d;
     _reinterpret_output_as_3d = gemm_info.depth_output_gemm3d != 0;
     _use_dummy_work_items     = preferred_dummy_work_items_support(CLKernelLibrary::get().get_device());