COMPMID-1086: Optimizing GCGEMM - Add helpers to get target GPU information

This patch moves some of the helper functions in CLHelpers.[h,cpp] to
GPUTarget.[h,cpp] in order to make them avaialable to both CL and GLES backends.

Change-Id: Ib89b05a2084b73eb643b32fac13b8367cc80b6a5
Reviewed-on: https://eu-gerrit-1.euhpc.arm.com/128988
Tested-by: Jenkins <bsgcomp@arm.com>
Reviewed-by: Anthony Barbier <anthony.barbier@arm.com>

3 files changed, 17 insertions, 4 deletions

diff --git a/src/runtime/GLES_COMPUTE/GCScheduler.cpp b/src/runtime/GLES_COMPUTE/GCScheduler.cpp
index fcc855957f..f7812730fc 100644
--- a/src/runtime/GLES_COMPUTE/GCScheduler.cpp
+++ b/src/runtime/GLES_COMPUTE/GCScheduler.cpp
@@ -1,5 +1,5 @@
 /*
- * Copyright (c) 2017 ARM Limited.
+ * Copyright (c) 2017-2018 ARM Limited.
  *
  * SPDX-License-Identifier: MIT
  *
@@ -24,6 +24,7 @@
 
 #include "arm_compute/runtime/GLES_COMPUTE/GCScheduler.h"
 
+#include "arm_compute/core/GLES_COMPUTE/GCHelpers.h"
 #include "arm_compute/core/GLES_COMPUTE/GCKernelLibrary.h"
 
 using namespace arm_compute;
@@ -31,7 +32,7 @@ using namespace arm_compute;
 std::once_flag GCScheduler::_initialize_symbols;
 
 GCScheduler::GCScheduler()
-    : _display(EGL_NO_DISPLAY), _context(EGL_NO_CONTEXT)
+    : _display(EGL_NO_DISPLAY), _context(EGL_NO_CONTEXT), _target(GPUTarget::MIDGARD)
 {
 }
 
@@ -48,11 +49,13 @@ void GCScheduler::default_init()
 {
     setup_context();
 
-    GCKernelLibrary::get().init("./cs_shaders/", _display, _context);
+    init(_display, _context);
 }
 
 void GCScheduler::init(EGLDisplay dpy, EGLContext ctx)
 {
+    _target = get_target_from_device();
+
     GCKernelLibrary::get().init("./cs_shaders/", dpy, ctx);
 }
 
diff --git a/src/runtime/GLES_COMPUTE/functions/GCConvolutionLayer.cpp b/src/runtime/GLES_COMPUTE/functions/GCConvolutionLayer.cpp
index dc73eb85e6..d1ef87d32c 100644
--- a/src/runtime/GLES_COMPUTE/functions/GCConvolutionLayer.cpp
+++ b/src/runtime/GLES_COMPUTE/functions/GCConvolutionLayer.cpp
@@ -109,6 +109,10 @@ void GCConvolutionLayer::configure(const IGCTensor *input, const IGCTensor *weig
 
     const DataType dt = input->info()->data_type();
 
+    // Set the GPU target for im2col and col2im
+    _input_im2col_kernel.set_target(GCScheduler::get().get_target());
+    _output_col2im_kernel.set_target(GCScheduler::get().get_target());
+
     const bool       append_bias   = (biases != nullptr);
     const unsigned   bias_element  = (append_bias) ? 1 : 0;
     const IGCTensor *biases_to_use = (append_bias) ? biases : nullptr;
diff --git a/src/runtime/GLES_COMPUTE/functions/GCGEMM.cpp b/src/runtime/GLES_COMPUTE/functions/GCGEMM.cpp
index 0a75a38c50..79f8f71713 100644
--- a/src/runtime/GLES_COMPUTE/functions/GCGEMM.cpp
+++ b/src/runtime/GLES_COMPUTE/functions/GCGEMM.cpp
@@ -38,7 +38,6 @@
 #include "arm_compute/runtime/ITensorAllocator.h"
 
 using namespace arm_compute;
-using namespace arm_compute::gles_compute;
 
 namespace
 {
@@ -92,6 +91,13 @@ void GCGEMM::configure(const IGCTensor *a, const IGCTensor *b, const IGCTensor *
     const IGCTensor *matrix_a = a;
     const IGCTensor *matrix_b = b;
 
+    // Get the GPU target
+    const GPUTarget gpu_target = GCScheduler::get().get_target();
+
+    // Set the target for the kernels
+    _interleave_kernel.set_target(gpu_target);
+    _mm_kernel.set_target(gpu_target);
+
     // Arguments used by GEMMReshapeInfo
     // If we pass the matrix A and matrix B reshaped to GCGEMMMatrixMultiplyKernel, we need to pass m, n, k, mult_transpose1xW_width and mult_interleave4x4_height to GCGEMMReshapeInfo
     // in order to know how the matrices have been reshaped