COMPMID-1276 - Allow GEMM to work with 3D input tensor

Skipped im2col in CLGEMMConvolutionLayer for 1x1 convolutions with NHWC data layout

Change-Id: I894e6b952ed8605e8f3ffc0ffc25c24730d4664c
Reviewed-on: https://eu-gerrit-1.euhpc.arm.com/141909
Tested-by: Jenkins <bsgcomp@arm.com>
Reviewed-by: Anthony Barbier <anthony.barbier@arm.com>
Reviewed-by: Georgios Pinitas <georgios.pinitas@arm.com>

1 files changed, 54 insertions, 19 deletions

diff --git a/src/core/CL/kernels/CLGEMMInterleave4x4Kernel.cpp b/src/core/CL/kernels/CLGEMMInterleave4x4Kernel.cpp
index 12a40cd7dc..6ea1160c69 100644
--- a/src/core/CL/kernels/CLGEMMInterleave4x4Kernel.cpp
+++ b/src/core/CL/kernels/CLGEMMInterleave4x4Kernel.cpp
@@ -23,6 +23,7 @@
  */
 #include "arm_compute/core/CL/kernels/CLGEMMInterleave4x4Kernel.h"
 
+#include "arm_compute/core/AccessWindowStatic.h"
 #include "arm_compute/core/CL/CLHelpers.h"
 #include "arm_compute/core/CL/CLKernelLibrary.h"
 #include "arm_compute/core/CL/CLValidate.h"
@@ -30,6 +31,7 @@
 #include "arm_compute/core/CL/OpenCL.h"
 #include "arm_compute/core/Error.h"
 #include "arm_compute/core/Helpers.h"
+#include "arm_compute/core/TensorInfo.h"
 #include "arm_compute/core/Types.h"
 #include "arm_compute/core/Utils.h"
 #include "arm_compute/core/Window.h"
@@ -40,7 +42,7 @@ using namespace arm_compute::misc::shape_calculator;
 
 namespace
 {
-Status validate_arguments(const ITensorInfo *input, const ITensorInfo *output, int mult_interleave4x4_height)
+Status validate_arguments(const ITensorInfo *input, const ITensorInfo *output, int mult_interleave4x4_height, bool reinterpret_input_as_3d)
 {
     ARM_COMPUTE_RETURN_ERROR_ON(mult_interleave4x4_height < 1);
     ARM_COMPUTE_RETURN_ERROR_ON_F16_UNSUPPORTED(input);
@@ -50,24 +52,30 @@ Status validate_arguments(const ITensorInfo *input, const ITensorInfo *output, i
 
     if(output->total_size() != 0)
     {
-        ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_DIMENSIONS(output->tensor_shape(), compute_interleaved_shape(*input, mult_interleave4x4_height));
+        ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_DIMENSIONS(output->tensor_shape(), compute_interleaved_shape(*input, mult_interleave4x4_height, reinterpret_input_as_3d));
         ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_DATA_TYPES(input, output);
     }
 
     return Status{};
 }
 
-std::pair<Status, Window> validate_and_configure_window(ITensorInfo *input, ITensorInfo *output, int mult_interleave4x4_height)
+std::pair<Status, Window> validate_and_configure_window(ITensorInfo *input, ITensorInfo *output, int mult_interleave4x4_height, bool reinterpret_input_as_3d)
 {
     constexpr unsigned int num_elems_processed_per_iteration_x = 4;
     constexpr unsigned int num_elems_processed_per_iteration_y = 4;
     const unsigned int     num_elems_written_per_iteration     = num_elems_processed_per_iteration_x * num_elems_processed_per_iteration_y * mult_interleave4x4_height;
     bool                   window_changed                      = false;
 
-    // Configure kernel window
-    Window                win = calculate_max_window(*input, Steps(num_elems_processed_per_iteration_x, num_elems_processed_per_iteration_y));
-    AccessWindowRectangle input_access(input, 0, 0, num_elems_processed_per_iteration_x, num_elems_processed_per_iteration_y);
-    window_changed = window_changed || update_window_and_padding(win, input_access);
+    TensorInfo tmp_info(*input);
+
+    if(reinterpret_input_as_3d)
+    {
+        // Since the input tensor has to be reinterpreted as 3D and the execute window is based on a 2D interleave,
+        // the window needs to be constructed on the 2D collapsed version of the tensor
+        TensorShape tmp_shape(input->tensor_shape());
+        tmp_shape.collapse(2U, 1U);
+        tmp_info.set_tensor_shape(tmp_shape);
+    }
 
     // Output auto inizialitation if not yet initialized
     auto_init_if_empty(*output, input->clone()->set_tensor_shape(compute_interleaved_shape(*input, mult_interleave4x4_height)));
@@ -76,9 +84,22 @@ std::pair<Status, Window> validate_and_configure_window(ITensorInfo *input, ITen
     const float scale_x = 4.0f * static_cast<float>(mult_interleave4x4_height);
     const float scale_y = 1.0f / (scale_x);
 
+    // Note: bottom paddings are calculated manually as the input can be reinterpreted as 3D tensor
+    // The only way to set properly the paddings, it is to set those explicitly through the AccessWindowStatic
+    const int m          = reinterpret_input_as_3d ? input->tensor_shape()[1] * input->tensor_shape()[2] : input->tensor_shape()[1];
+    const int bottom_pad = (num_elems_processed_per_iteration_y - (m % num_elems_processed_per_iteration_y)) % num_elems_processed_per_iteration_y;
+
+    Window win    = calculate_max_window(tmp_info, Steps(num_elems_processed_per_iteration_x, num_elems_processed_per_iteration_y));
+    Window win_in = calculate_max_window(*input, Steps(num_elems_processed_per_iteration_x, num_elems_processed_per_iteration_y));
+
+    AccessWindowStatic input_access(input, 0, 0,
+                                    ceil_to_multiple(input->dimension(0), num_elems_processed_per_iteration_x),
+                                    input->dimension(1) + bottom_pad);
     AccessWindowRectangle output_access(output, 0, 0, num_elems_written_per_iteration, 1, scale_x, scale_y);
-    window_changed = window_changed || update_window_and_padding(win, output_access);
-    output_access.set_valid_region(win, input->valid_region());
+
+    window_changed = update_window_and_padding(win_in, input_access) || // window used by the execute_window_loop
+                     update_window_and_padding(win, output_access);     // window used to update the padding requirements of output tensor
+    output_access.set_valid_region(win, ValidRegion(Coordinates(0, 0), output->tensor_shape()));
 
     // Collapse along the Z direction
     // This collapse needs to be here in order to tune the Z dimension of LWS
@@ -90,26 +111,31 @@ std::pair<Status, Window> validate_and_configure_window(ITensorInfo *input, ITen
 } // namespace
 
 CLGEMMInterleave4x4Kernel::CLGEMMInterleave4x4Kernel()
-    : _input(nullptr), _output(nullptr)
+    : _input(nullptr), _output(nullptr), _reinterpret_input_as_3d(false)
 {
 }
 
-void CLGEMMInterleave4x4Kernel::configure(const ICLTensor *input, ICLTensor *output, int mult_interleave4x4_height)
+void CLGEMMInterleave4x4Kernel::configure(const ICLTensor *input, ICLTensor *output, int mult_interleave4x4_height, bool reinterpret_input_as_3d)
 {
     ARM_COMPUTE_ERROR_ON_NULLPTR(input, output);
 
     // Output auto inizialitation if not yet initialized
-    auto_init_if_empty(*output->info(), input->info()->clone()->set_tensor_shape(compute_interleaved_shape(*input->info(), mult_interleave4x4_height)));
+    auto_init_if_empty(*output->info(), input->info()->clone()->set_tensor_shape(compute_interleaved_shape(*input->info(), mult_interleave4x4_height, reinterpret_input_as_3d)));
 
     // Perform validate step
-    ARM_COMPUTE_ERROR_THROW_ON(validate_arguments(input->info(), output->info(), mult_interleave4x4_height));
+    ARM_COMPUTE_ERROR_THROW_ON(validate_arguments(input->info(), output->info(), mult_interleave4x4_height, reinterpret_input_as_3d));
 
-    _input  = input;
-    _output = output;
+    _input                   = input;
+    _output                  = output;
+    _reinterpret_input_as_3d = reinterpret_input_as_3d;
 
     // Create build options
     CLBuildOptions build_opts;
     build_opts.add_option("-DMULT_INTERLEAVE4X4_HEIGHT=" + support::cpp11::to_string(mult_interleave4x4_height));
+    build_opts.add_option_if(_reinterpret_input_as_3d, "-DREINTERPRET_INPUT_AS_3D");
+    build_opts.add_option_if(_reinterpret_input_as_3d, "-DHEIGHT_GEMM3D=" + support::cpp11::to_string(input->info()->dimension(1)));
+    build_opts.add_option_if(_reinterpret_input_as_3d, "-DDEPTH_GEMM3D=" + support::cpp11::to_string(input->info()->dimension(2)));
+
     switch(input->info()->element_size())
     {
         case 1:
@@ -129,12 +155,13 @@ void CLGEMMInterleave4x4Kernel::configure(const ICLTensor *input, ICLTensor *out
     _kernel = static_cast<cl::Kernel>(CLKernelLibrary::get().create_kernel("gemm_interleave4x4", build_opts.options()));
 
     // Configure kernel window
-    auto win_config = validate_and_configure_window(input->info(), output->info(), mult_interleave4x4_height);
+    auto win_config = validate_and_configure_window(input->info(), output->info(), mult_interleave4x4_height, reinterpret_input_as_3d);
     ARM_COMPUTE_ERROR_THROW_ON(win_config.first);
     ICLKernel::configure(win_config.second);
 
     // Set config_id for enabling LWS tuning
     _config_id = "interleave4x4_";
+    _config_id += (_reinterpret_input_as_3d ? "3d_" : "");
     _config_id += lower_string(string_from_data_type(input->info()->data_type()));
     _config_id += "_";
     _config_id += support::cpp11::to_string(output->info()->dimension(0));
@@ -146,10 +173,10 @@ void CLGEMMInterleave4x4Kernel::configure(const ICLTensor *input, ICLTensor *out
     _config_id += support::cpp11::to_string(output->info()->dimension(3));
 }
 
-Status CLGEMMInterleave4x4Kernel::validate(const ITensorInfo *input, const ITensorInfo *output, int mult_interleave4x4_height)
+Status CLGEMMInterleave4x4Kernel::validate(const ITensorInfo *input, const ITensorInfo *output, int mult_interleave4x4_height, bool reinterpret_input_as_3d)
 {
-    ARM_COMPUTE_RETURN_ON_ERROR(validate_arguments(input, output, mult_interleave4x4_height));
-    ARM_COMPUTE_RETURN_ON_ERROR(validate_and_configure_window(input->clone().get(), output->clone().get(), mult_interleave4x4_height).first);
+    ARM_COMPUTE_RETURN_ON_ERROR(validate_arguments(input, output, mult_interleave4x4_height, reinterpret_input_as_3d));
+    ARM_COMPUTE_RETURN_ON_ERROR(validate_and_configure_window(input->clone().get(), output->clone().get(), mult_interleave4x4_height, reinterpret_input_as_3d).first);
 
     return Status{};
 }
@@ -170,6 +197,14 @@ void CLGEMMInterleave4x4Kernel::run(const Window &window, cl::CommandQueue &queu
      */
     Window slice = window.first_slice_window_3D();
 
+    if(_reinterpret_input_as_3d)
+    {
+        // Pass bottom paddings to the kernel if the input has to be reinterpreted as 3D tensor
+        const unsigned int idx0                  = 2 * num_arguments_per_3D_tensor();
+        const unsigned int total_cross_plane_pad = _input->info()->padding().top + _input->info()->padding().bottom;
+        _kernel.setArg<cl_uint>(idx0, static_cast<unsigned int>(total_cross_plane_pad));
+    }
+
     do
     {
         unsigned int idx = 0;