COMPMID-748 - Integrating optimized SGEMM for bifrost

This patch introduces a new GEMM capable to improve the mac utilisation
of 10% compared to the GEMM without reshape. However this implementation
is not faster in all cases as we need to take into account the time for
reshaping the matrices. For this reason an heuristic solution to select
the optimal GEMM to use has been added to the function. More information
about the heuristic implementation can be found at COMPMID-852.
With this new patch, GoogleNet, MobileNet, VGG16 and SqueezeNet can
improved the performance of 1.5x.
More information about the performance uplift can be found here:
https://confluence.arm.com/display/MLENG/GEMM+FP32+performance%3A+ACL+18.02

Change-Id: I024563c06b9aed02a211a974e452bae5c233b04c
Reviewed-on: https://eu-gerrit-1.euhpc.arm.com/117140
Reviewed-by: Pablo Tello <pablo.tello@arm.com>
Tested-by: Jenkins <bsgcomp@arm.com>
Reviewed-by: Anthony Barbier <anthony.barbier@arm.com>

1 files changed, 22 insertions, 15 deletions

diff --git a/src/core/CL/kernels/CLGEMMTranspose1xWKernel.cpp b/src/core/CL/kernels/CLGEMMTranspose1xWKernel.cpp
index 69a545b76b..63aed6df32 100644
--- a/src/core/CL/kernels/CLGEMMTranspose1xWKernel.cpp
+++ b/src/core/CL/kernels/CLGEMMTranspose1xWKernel.cpp
@@ -1,5 +1,5 @@
 /*
- * Copyright (c) 2017 ARM Limited.
+ * Copyright (c) 2017-2018 ARM Limited.
  *
  * SPDX-License-Identifier: MIT
  *
@@ -42,8 +42,9 @@ using namespace arm_compute::misc::shape_calculator;
 
 namespace
 {
-Status validate_arguments(const ITensorInfo *input, const ITensorInfo *output)
+Status validate_arguments(const ITensorInfo *input, const ITensorInfo *output, int mult_transpose1xW_width)
 {
+    ARM_COMPUTE_RETURN_ERROR_ON(mult_transpose1xW_width < 1);
     ARM_COMPUTE_RETURN_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(input, 1, DataType::QS8, DataType::QASYMM8, DataType::U8, DataType::S8,
                                                          DataType::QS16, DataType::U16, DataType::S16, DataType::U32, DataType::S32,
                                                          DataType::F16, DataType::F32);
@@ -51,7 +52,7 @@ Status validate_arguments(const ITensorInfo *input, const ITensorInfo *output)
     if(output->total_size() != 0)
     {
         ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_DIMENSIONS(output->tensor_shape(),
-                                                           compute_transpose1xW_with_element_size_shape(*input));
+                                                           compute_transpose1xW_with_element_size_shape(*input, mult_transpose1xW_width));
         ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_DATA_TYPES(input, output);
         ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_FIXED_POINT(input, output);
     }
@@ -59,11 +60,11 @@ Status validate_arguments(const ITensorInfo *input, const ITensorInfo *output)
     return Status{};
 }
 
-std::pair<Status, Window> validate_and_configure_window(ITensorInfo *input, ITensorInfo *output, unsigned int &num_elems_processed_per_iteration)
+std::pair<Status, Window> validate_and_configure_window(ITensorInfo *input, ITensorInfo *output, unsigned int &num_elems_processed_per_iteration, int mult_transpose1xW_width)
 {
     num_elems_processed_per_iteration = 16 / input->element_size();
 
-    const int scale_x        = num_elems_processed_per_iteration;
+    const int scale_x        = num_elems_processed_per_iteration * mult_transpose1xW_width;
     bool      window_changed = false;
 
     // Configure kernel window
@@ -90,25 +91,31 @@ std::pair<Status, Window> validate_and_configure_window(ITensorInfo *input, ITen
 }
 } // namespace
 
-void CLGEMMTranspose1xWKernel::configure(const ICLTensor *input, ICLTensor *output)
+void CLGEMMTranspose1xWKernel::configure(const ICLTensor *input, ICLTensor *output, int mult_transpose1xW_width)
 {
     ARM_COMPUTE_ERROR_ON_NULLPTR(input, output);
 
     // Output tensor auto inizialitation if not yet initialized
-    auto_init_if_empty(*output->info(), input->info()->clone()->set_tensor_shape(compute_transpose1xW_with_element_size_shape(*input->info())));
+    auto_init_if_empty(*output->info(), input->info()->clone()->set_tensor_shape(compute_transpose1xW_with_element_size_shape(*input->info(), mult_transpose1xW_width)));
 
     // Perform validate step
-    ARM_COMPUTE_ERROR_THROW_ON(validate_arguments(input->info(), output->info()));
+    ARM_COMPUTE_ERROR_THROW_ON(validate_arguments(input->info(), output->info(), mult_transpose1xW_width));
 
     _input  = input;
     _output = output;
 
     // Configure kernel window
+    // Note: num_elems_processed_per_iteration will be set in validate_and_configure_window()
     unsigned int num_elems_processed_per_iteration = 1;
-    auto         win_config                        = validate_and_configure_window(input->info(), output->info(), num_elems_processed_per_iteration);
+    auto         win_config                        = validate_and_configure_window(input->info(), output->info(), num_elems_processed_per_iteration, mult_transpose1xW_width);
     ARM_COMPUTE_ERROR_THROW_ON(win_config.first);
     ICLKernel::configure(win_config.second);
 
+    // Create build options
+    CLBuildOptions build_opts;
+    build_opts.add_option("-DTRANSPOSE_W=" + support::cpp11::to_string(num_elems_processed_per_iteration));
+    build_opts.add_option("-DMULT_TRANSPOSE1XW_WIDTH=" + support::cpp11::to_string(mult_transpose1xW_width));
+
     /*
      * Following an example of how the transposition1xW works when the input data type is F32
      *
@@ -117,18 +124,18 @@ void CLGEMMTranspose1xWKernel::configure(const ICLTensor *input, ICLTensor *outp
      *         |a20 a21 a22 a23| = | a00 a01 a02 a03 || a10 a11 a12 a13 || a20 a21 a22 a23 || a30 a31 a32 a33 |
      *         |a30 a31 a32 a33|
      *
-     * The output matrix will have the following shape: [ height * W, ceil(width / W) ], where W = (16 / element size of the tensor)
+     * The output matrix will have the following shape: [ height * W, ceil(width / W) ], where W = (16 / element size of the tensor) * mult_transpose1xW_width
      */
     // Create kernel
-    std::string kernel_name = "gemm_transpose1x" + support::cpp11::to_string(num_elems_processed_per_iteration);
-    _kernel                 = static_cast<cl::Kernel>(CLKernelLibrary::get().create_kernel(kernel_name));
+    std::string kernel_name = "gemm_transpose1xW";
+    _kernel                 = static_cast<cl::Kernel>(CLKernelLibrary::get().create_kernel(kernel_name, build_opts.options()));
 }
 
-Status CLGEMMTranspose1xWKernel::validate(const ITensorInfo *input, const ITensorInfo *output)
+Status CLGEMMTranspose1xWKernel::validate(const ITensorInfo *input, const ITensorInfo *output, int mult_transpose1xW_width)
 {
     unsigned int num_elems_processed_per_iteration = 1;
-    ARM_COMPUTE_RETURN_ON_ERROR(validate_arguments(input, output));
-    ARM_COMPUTE_RETURN_ON_ERROR(validate_and_configure_window(input->clone().get(), output->clone().get(), num_elems_processed_per_iteration).first);
+    ARM_COMPUTE_RETURN_ON_ERROR(validate_arguments(input, output, mult_transpose1xW_width));
+    ARM_COMPUTE_RETURN_ON_ERROR(validate_and_configure_window(input->clone().get(), output->clone().get(), num_elems_processed_per_iteration, mult_transpose1xW_width).first);
 
     return Status{};
 }