COMPMID-3925: Dispatch CLGEMM with no padding y requirement

- Add has_pad_y flag in GEMMKernelInfo
- Skip reinterpret as 3D in CLGEMMMatrixMultiplyReshapedOnlyRHSKernel if
  has_pad_y = false
- Add test to validate CLGEMMMatrixMultiplyReshapedOnlyRHSkernel with
  had_pad_y = false/true
- Configure two variants of CLGEMMMatrixMultiplyReshapedOnlyRHSKernel to
  run with has_pad_y = false/true in CLGEMM
- Check if the lhs/dst tensors have pad y. If not, run
  CLGEMMMatrixMultiplyReshapedOnlyRHSKernel without padding requirement

Change-Id: I68bb43389789736d676b899ac7c77fd9138babaf
Signed-off-by: Gian Marco Iodice <gianmarco.iodice@arm.com>
Reviewed-on: https://review.mlplatform.org/c/ml/ComputeLibrary/+/4248
Tested-by: Arm Jenkins <bsgcomp@arm.com>
Reviewed-by: Georgios Pinitas <georgios.pinitas@arm.com>
Reviewed-by: Michele Di Giorgio <michele.digiorgio@arm.com>
Comments-Addressed: Arm Jenkins <bsgcomp@arm.com>

1 files changed, 34 insertions, 32 deletions

diff --git a/src/core/CL/kernels/CLGEMMMatrixMultiplyReshapedOnlyRHSKernel.cpp b/src/core/CL/kernels/CLGEMMMatrixMultiplyReshapedOnlyRHSKernel.cpp
index 912c763ed5..68f761b9e7 100644
--- a/src/core/CL/kernels/CLGEMMMatrixMultiplyReshapedOnlyRHSKernel.cpp
+++ b/src/core/CL/kernels/CLGEMMMatrixMultiplyReshapedOnlyRHSKernel.cpp
@@ -133,7 +133,8 @@ std::pair<Status, Window> validate_and_configure_window(ITensorInfo *input0, ITe
 
     // In case both input and output 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)
+    // This approach should only be used when the input/output tensors have pad on the y direction
+    if((reinterpret_input_as_3d == reinterpret_output_as_3d) && gemm_info.has_pad_y)
     {
         reinterpret_output_as_3d = false;
     }
@@ -159,16 +160,6 @@ std::pair<Status, Window> validate_and_configure_window(ITensorInfo *input0, ITe
     win     = calculate_max_window(tmp_info, Steps(num_elems_processed_per_iteration_x, num_elems_processed_per_iteration_y));
     win_out = calculate_max_window(*output, Steps(num_elems_processed_per_iteration_x, num_elems_processed_per_iteration_y));
 
-    AccessWindowStatic input0_access(input0, 0, 0,
-                                     input0->dimension(0),
-                                     input0->dimension(1));
-    AccessWindowStatic input1_access(input1, 0, 0,
-                                     input1->dimension(0),
-                                     input1->dimension(1));
-    AccessWindowStatic output_access(output, 0, 0,
-                                     output->dimension(0),
-                                     output->dimension(1));
-
     if(input2 != nullptr)
     {
         const int bias_processed_per_iteration_x = num_elems_processed_per_iteration_x;
@@ -177,17 +168,9 @@ std::pair<Status, Window> validate_and_configure_window(ITensorInfo *input0, ITe
                                          ceil_to_multiple(input2->dimension(0), bias_processed_per_iteration_x),
                                          input2->dimension(1));
 
-        window_changed = update_window_and_padding(win, input0_access, input1_access, input2_access) || // window used by the execute_window_loop
-                         update_window_and_padding(win_out, output_access);                             // window used to update the padding requirements of output tensor
-    }
-    else
-    {
-        window_changed = update_window_and_padding(win, input0_access, input1_access) || // window used by the execute_window_loop
-                         update_window_and_padding(win_out, output_access);              // window used to update the padding requirements of output tensor
+        window_changed = update_window_and_padding(win, input2_access);
     }
 
-    output_access.set_valid_region(win_out, ValidRegion(Coordinates(), output->tensor_shape()));
-
     // Collapse along the Z direction
     // This collapse needs to be here in order to tune the Z dimension of LWS
     Window             collapsed             = win;
@@ -201,7 +184,7 @@ std::pair<Status, Window> validate_and_configure_window(ITensorInfo *input0, ITe
 
 CLGEMMMatrixMultiplyReshapedOnlyRHSKernel::CLGEMMMatrixMultiplyReshapedOnlyRHSKernel()
     : _input0(nullptr), _input1(nullptr), _input2(nullptr), _output(nullptr), _slide_matrix_b(true), _reinterpret_input_as_3d(false), _reinterpret_output_as_3d(false), _use_dummy_work_items(false),
-      _add_bias(false), _broadcast_bias(false), _export_to_cl_image(false)
+      _add_bias(false), _broadcast_bias(false), _export_to_cl_image(false), _has_pad_y(false)
 {
 }
 
@@ -232,10 +215,13 @@ void CLGEMMMatrixMultiplyReshapedOnlyRHSKernel::configure(const CLCompileContext
     _add_bias                 = _input2 != nullptr;
     _broadcast_bias           = gemm_info.broadcast_bias;
     _export_to_cl_image       = rhs_info.export_to_cl_image;
+    _has_pad_y                = gemm_info.has_pad_y;
+
+    auto padding_info = get_padding_info({ input0, input1, output });
 
     // In case both input and output 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)
+    if((_reinterpret_input_as_3d == _reinterpret_output_as_3d) && _has_pad_y)
     {
         _reinterpret_input_as_3d  = false;
         _reinterpret_output_as_3d = false;
@@ -257,6 +243,7 @@ void CLGEMMMatrixMultiplyReshapedOnlyRHSKernel::configure(const CLCompileContext
     // This means that the actual m used by the kernel is given by output->info()->dimension(1) and not by gemm_info.m
     const unsigned int internal_m = _reinterpret_output_as_3d ? gemm_info.m : output->info()->dimension(1);
 
+    // These variables are used only if gemm_info.has_pad_y == true
     const unsigned int h_gemm_3d = _reinterpret_output_as_3d ? output->info()->dimension(1) : input0->info()->dimension(1);
     const unsigned int d_gemm_3d = _reinterpret_output_as_3d ? output->info()->dimension(2) : input0->info()->dimension(2);
 
@@ -274,11 +261,7 @@ void CLGEMMMatrixMultiplyReshapedOnlyRHSKernel::configure(const CLCompileContext
     build_opts.add_option_if(!(helpers::float_ops::is_one(alpha)), "-DALPHA=" + float_to_string_with_full_precision(alpha));
     build_opts.add_option_if(_input2 != nullptr, "-DBETA=" + float_to_string_with_full_precision(beta));
     build_opts.add_option_if(helpers::float_ops::is_one(beta), "-DUNIT_BETA");
-    build_opts.add_option_if(_reinterpret_input_as_3d, "-DREINTERPRET_INPUT_AS_3D");
-    build_opts.add_option_if(_reinterpret_output_as_3d, "-DREINTERPRET_OUTPUT_AS_3D");
     build_opts.add_option_if(gemm_info.broadcast_bias, "-DBROADCAST_BIAS");
-    build_opts.add_option_if(_reinterpret_input_as_3d || _reinterpret_output_as_3d, "-DHEIGHT_GEMM3D=" + support::cpp11::to_string(h_gemm_3d));
-    build_opts.add_option_if(_reinterpret_input_as_3d || _reinterpret_output_as_3d, "-DDEPTH_GEMM3D=" + support::cpp11::to_string(d_gemm_3d));
     build_opts.add_option_if(!_slide_matrix_b, "-DMATRIX_B_DEPTH=" + support::cpp11::to_string(input1->info()->dimension(2)));
     build_opts.add_option_if(rhs_info.interleave, "-DRHS_INTERLEAVE");
     build_opts.add_option_if(_use_dummy_work_items, "-DDUMMY_WORK_ITEMS");
@@ -296,6 +279,13 @@ void CLGEMMMatrixMultiplyReshapedOnlyRHSKernel::configure(const CLCompileContext
     build_opts.add_option_if(gemm_info.activation_info.enabled(), "-DACTIVATION_TYPE=" + lower_string(string_from_activation_func(gemm_info.activation_info.activation())));
     build_opts.add_option_if(gemm_info.activation_info.enabled(), "-DA_VAL=" + float_to_string_with_full_precision(gemm_info.activation_info.a()));
     build_opts.add_option_if(gemm_info.activation_info.enabled(), "-DB_VAL=" + float_to_string_with_full_precision(gemm_info.activation_info.b()));
+    if(_has_pad_y)
+    {
+        build_opts.add_option_if(_reinterpret_input_as_3d, "-DREINTERPRET_INPUT_AS_3D");
+        build_opts.add_option_if(_reinterpret_output_as_3d, "-DREINTERPRET_OUTPUT_AS_3D");
+        build_opts.add_option_if(_reinterpret_input_as_3d || _reinterpret_output_as_3d, "-DHEIGHT_GEMM3D=" + support::cpp11::to_string(h_gemm_3d));
+        build_opts.add_option_if(_reinterpret_input_as_3d || _reinterpret_output_as_3d, "-DDEPTH_GEMM3D=" + support::cpp11::to_string(d_gemm_3d));
+    }
 
     std::string kernel_name("gemm_mm_reshaped_only_rhs_");
     kernel_name += rhs_info.transpose ? "t" : "nt";
@@ -307,6 +297,7 @@ void CLGEMMMatrixMultiplyReshapedOnlyRHSKernel::configure(const CLCompileContext
     // Set config_id for enabling LWS tuning
     _config_id = kernel_name;
     _config_id += "_";
+    _config_id += (_has_pad_y ? "" : "no_pad_y_");
     _config_id += (_add_bias ? "add_bias_" : "");
     _config_id += (_broadcast_bias ? "broadcast_bias_" : "");
     _config_id += (_reinterpret_input_as_3d ? "3di_" : "");
@@ -331,6 +322,8 @@ void CLGEMMMatrixMultiplyReshapedOnlyRHSKernel::configure(const CLCompileContext
     _config_id += support::cpp11::to_string(rhs_info.h0);
     _config_id += "_";
     _config_id += support::cpp11::to_string(rhs_info.interleave);
+
+    ARM_COMPUTE_ERROR_ON(has_padding_changed(padding_info));
 }
 
 Status CLGEMMMatrixMultiplyReshapedOnlyRHSKernel::validate(const ITensorInfo *input0, const ITensorInfo *input1, const ITensorInfo *input2, const ITensorInfo *output, float alpha, float beta,
@@ -363,15 +356,24 @@ void CLGEMMMatrixMultiplyReshapedOnlyRHSKernel::run(const Window &window, cl::Co
         ARM_COMPUTE_ERROR_ON(_input1->info()->strides_in_bytes()[3] != 0);
     }
 
+    const size_t lhs_idx_batch_size = _reinterpret_input_as_3d && !_has_pad_y? 3u : 2u;
+    const size_t rhs_idx_batch_size = 2u;
+    const size_t bia_idx_batch_size = 2u;
+    const size_t out_idx_batch_size = _reinterpret_output_as_3d && !_has_pad_y? 3u : 2u;
+
     Window slice          = window.first_slice_window_3D();
     Window slice_matrix_b = slice;
 
     slice_matrix_b.set(Window::DimX, Window::Dimension(0, 1, 1));
     slice_matrix_b.set(Window::DimY, Window::Dimension(0, 1, 1));
 
+    // Get cross plane pads
     const unsigned int total_cross_plane_pad_lhs = _input0->info()->padding().top + _input0->info()->padding().bottom;
     const unsigned int total_cross_plane_pad_out = _output->info()->padding().top + _output->info()->padding().bottom;
 
+    // The execution should fail if we try to run with has_pad_y = false but we have padding in either the LHS or DST tensor
+    ARM_COMPUTE_ERROR_ON(!_has_pad_y && ((total_cross_plane_pad_lhs != 0) || (total_cross_plane_pad_out != 0)));
+
     cl::Image2D input1_image2d;
 
     if(_export_to_cl_image)
@@ -414,28 +416,28 @@ void CLGEMMMatrixMultiplyReshapedOnlyRHSKernel::run(const Window &window, cl::Co
         add_2D_tensor_argument(idx, _output, slice);
 
         // LHS stride_z
-        _kernel.setArg<cl_uint>(idx++, static_cast<unsigned int>(_input0->info()->strides_in_bytes()[2]));
+        _kernel.setArg<cl_uint>(idx++, static_cast<unsigned int>(_input0->info()->strides_in_bytes()[lhs_idx_batch_size]));
 
         // RHS stride_z (not used if _export_to_cl_image == true)
-        _kernel.setArg<cl_uint>(idx++, static_cast<unsigned int>(_input1->info()->strides_in_bytes()[2]));
+        _kernel.setArg<cl_uint>(idx++, static_cast<unsigned int>(_input1->info()->strides_in_bytes()[rhs_idx_batch_size]));
 
         // Bias stride_z (if _add_bias == true)
         if(_add_bias)
         {
-            _kernel.setArg<cl_uint>(idx++, static_cast<unsigned int>(_input2->info()->strides_in_bytes()[2]));
+            _kernel.setArg<cl_uint>(idx++, static_cast<unsigned int>(_input2->info()->strides_in_bytes()[bia_idx_batch_size]));
         }
 
         // Output stride_z
-        _kernel.setArg<cl_uint>(idx++, static_cast<unsigned int>(_output->info()->strides_in_bytes()[2]));
+        _kernel.setArg<cl_uint>(idx++, static_cast<unsigned int>(_output->info()->strides_in_bytes()[out_idx_batch_size]));
 
         // Cross-plan padding (if _reinterpret_input_as_3d = true)
-        if(_reinterpret_input_as_3d)
+        if(_reinterpret_input_as_3d && _has_pad_y)
         {
             _kernel.setArg<cl_uint>(idx++, static_cast<unsigned int>(total_cross_plane_pad_lhs));
         }
 
         // Cross-plan padding (if _reinterpret_output_as_3d = true)
-        if(_reinterpret_output_as_3d)
+        if(_reinterpret_output_as_3d && _has_pad_y)
         {
             _kernel.setArg<cl_uint>(idx++, static_cast<unsigned int>(total_cross_plane_pad_out));
         }