COMPMID-759 - CLGEMM optimization for McVail benchmarks

This patch introduces an optimization for CLGEMM on Bifrost
architectures which can bring to 40% of FMA utilization on
config 3 of McVail. The new CLGEMM does not require any reshape of
matrix A and matrix B.

This patch also adds the auto-config in CLConvolutionLayer and CLGEMM
and extends the interface for NEGEMM and CLGEMM.

Change-Id: Ibb354eda45e9ca64b14a99700fb21dff5989dda9
Reviewed-on: https://eu-gerrit-1.euhpc.arm.com/113716
Tested-by: Jenkins <bsgcomp@arm.com>
Reviewed-by: Michalis Spyrou <michalis.spyrou@arm.com>
Reviewed-by: Anthony Barbier <anthony.barbier@arm.com>

1 files changed, 27 insertions, 24 deletions

diff --git a/src/runtime/CL/functions/CLGEMM.cpp b/src/runtime/CL/functions/CLGEMM.cpp
index ca0228fcdb..be2527f4ba 100644
--- a/src/runtime/CL/functions/CLGEMM.cpp
+++ b/src/runtime/CL/functions/CLGEMM.cpp
@@ -39,14 +39,17 @@
 using namespace arm_compute;
 
 CLGEMM::CLGEMM(std::shared_ptr<IMemoryManager> memory_manager)
-    : _memory_group(std::move(memory_manager)), _interleave_kernel(), _transpose_kernel(), _mm_kernel(), _ma_kernel(), _tmp_a(), _tmp_b(), _is_interleaved_transposed(false), _run_addition(false)
+    : _memory_group(std::move(memory_manager)), _interleave_kernel(), _transpose_kernel(), _mm_kernel(), _ma_kernel(), _tmp_a(), _tmp_b(), _is_interleaved_transposed(false), _run_addition(false),
+      _is_first_run(true), _reshape_b_only_on_first_run(false)
 {
 }
 
-void CLGEMM::configure(const ICLTensor *a, const ICLTensor *b, const ICLTensor *c, ICLTensor *output, float alpha, float beta)
+void CLGEMM::configure(const ICLTensor *a, const ICLTensor *b, const ICLTensor *c, ICLTensor *output, float alpha, float beta, const GEMMInfo &gemm_info)
 {
     ARM_COMPUTE_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(a, 1, DataType::QS8, DataType::QS16, DataType::F16, DataType::F32);
     ARM_COMPUTE_ERROR_ON_MISMATCHING_DATA_TYPES(a, b, output);
+    ARM_COMPUTE_ERROR_ON_MSG(gemm_info.is_a_reshaped(), "Matrix A already reshaped is not supported");
+    ARM_COMPUTE_ERROR_ON_MSG(gemm_info.is_b_reshaped(), "Matrix B already reshaped is not supported");
 
     if(c != nullptr)
     {
@@ -60,7 +63,11 @@ void CLGEMM::configure(const ICLTensor *a, const ICLTensor *b, const ICLTensor *
     ARM_COMPUTE_ERROR_ON_MSG(a->info()->dimension(0) != b->info()->dimension(1), "The product AB is defined only if the number of columns in A is equal to the number of rows in B");
 
     // If the input tensor has less than 16 rows, we run a special version of GEMM without reshaping the input tensors
-    _is_interleaved_transposed = a->info()->dimension(1) > 16;
+    // For Bifrost architectures we do not reshape the input matrices
+    _is_interleaved_transposed = (a->info()->dimension(1) > 16 && CLScheduler::get().target() != GPUTarget::BIFROST);
+
+    // Check if we need to reshape the matrix B only on the first run
+    _reshape_b_only_on_first_run = gemm_info.reshape_b_only_on_first_run();
 
     const ICLTensor *matrix_a = a;
     const ICLTensor *matrix_b = b;
@@ -73,31 +80,17 @@ void CLGEMM::configure(const ICLTensor *a, const ICLTensor *b, const ICLTensor *
         matrix_a = &_tmp_a;
         matrix_b = &_tmp_b;
 
-        TensorShape shape_tmp_a = a->info()->tensor_shape();
-        TensorShape shape_tmp_b = b->info()->tensor_shape();
-
-        shape_tmp_a.set(0, a->info()->dimension(0) * 4);
-        shape_tmp_a.set(1, std::ceil(a->info()->dimension(1) / 4.0f));
-
-        const unsigned int transpose_w = max_cl_vector_width / data_size_from_type(b->info()->data_type());
-        shape_tmp_b.set(0, b->info()->dimension(1) * transpose_w);
-        shape_tmp_b.set(1, std::ceil(b->info()->dimension(0) / static_cast<float>(transpose_w)));
-
-        TensorInfo info_a(shape_tmp_a, 1, a->info()->data_type(), a->info()->fixed_point_position());
-        _tmp_a.allocator()->init(info_a);
-
-        TensorInfo info_b(shape_tmp_b, 1, b->info()->data_type(), b->info()->fixed_point_position());
-        _tmp_b.allocator()->init(info_b);
-
-        // Manage intermediate buffers
-        _memory_group.manage(&_tmp_a);
-        _memory_group.manage(&_tmp_b);
+        // _tmp_a and _tmp_n will be auto configured in _interleave_kernel and in _transpose_kernel
 
         // Configure interleave kernel
         _interleave_kernel.configure(a, &_tmp_a);
 
         // Configure transpose kernel
         _transpose_kernel.configure(b, &_tmp_b);
+
+        // Manage intermediate buffers
+        _memory_group.manage(&_tmp_a);
+        _memory_group.manage(&_tmp_b);
     }
 
     _mm_kernel.configure(matrix_a, matrix_b, output, alpha, _is_interleaved_transposed);
@@ -126,8 +119,18 @@ void CLGEMM::run()
         // Run interleave kernel
         CLScheduler::get().enqueue(_interleave_kernel, false);
 
-        // Run transpose kernel
-        CLScheduler::get().enqueue(_transpose_kernel, false);
+        if(_is_first_run)
+        {
+            // Run transpose kernel
+            CLScheduler::get().enqueue(_transpose_kernel, false);
+
+            _is_first_run = false;
+        }
+        else if(!_reshape_b_only_on_first_run)
+        {
+            // Run transpose kernel
+            CLScheduler::get().enqueue(_transpose_kernel, false);
+        }
     }
 
     // Run matrix multiply kernel