COMPMID-1043: Rework GCGEMMMatrixMultiplyKernel interface and allow auto initialization of the tensors

This patch also:
- removes support for already reshaped weights in GCConvolutionLayer
- makes GCConvolutionLayer similar to CLGEMMConvolutionLayer
- enables usage of the GCGEMM function in GCConvolution instead of calling the
  GEMM kernels directly

Change-Id: I3e4a64335555e86e18585d38d8fda4bfdb44e265
Reviewed-on: https://eu-gerrit-1.euhpc.arm.com/127696
Tested-by: Jenkins <bsgcomp@arm.com>
Reviewed-by: Anthony Barbier <anthony.barbier@arm.com>

1 files changed, 70 insertions, 35 deletions

diff --git a/src/runtime/GLES_COMPUTE/functions/GCGEMM.cpp b/src/runtime/GLES_COMPUTE/functions/GCGEMM.cpp
index 9c8568a329..0a75a38c50 100644
--- a/src/runtime/GLES_COMPUTE/functions/GCGEMM.cpp
+++ b/src/runtime/GLES_COMPUTE/functions/GCGEMM.cpp
@@ -40,62 +40,82 @@
 using namespace arm_compute;
 using namespace arm_compute::gles_compute;
 
-GCGEMM::GCGEMM(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)
+namespace
 {
-}
-
-void GCGEMM::configure(const IGCTensor *a, const IGCTensor *b, const IGCTensor *c, IGCTensor *output, float alpha, float beta, const GEMMInfo &gemm_info)
+Status validate_arguments(const ITensorInfo *a, const ITensorInfo *b, const IGCTensor *c, const ITensorInfo *output, const float alpha, const float beta, const GEMMInfo &gemm_info = GEMMInfo())
 {
-    ARM_COMPUTE_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(a, 1, DataType::F32);
+    ARM_COMPUTE_ERROR_ON_NULLPTR(a, b, output);
+
+    ARM_COMPUTE_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(a, 1, 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");
-    ARM_COMPUTE_ERROR_ON_MSG(gemm_info.reshape_b_only_on_first_run(), "Reshape matrix B only on first run is not supported");
-    ARM_COMPUTE_UNUSED(gemm_info);
 
     if(c != nullptr)
     {
-        ARM_COMPUTE_ERROR_ON_MISMATCHING_DATA_TYPES(a, c);
-        ARM_COMPUTE_ERROR_ON_MSG(a->info()->dimension(1) != c->info()->dimension(1), "The C matrix must have the same number of rows as the matrix A");
-        ARM_COMPUTE_ERROR_ON_MSG(b->info()->dimension(0) != c->info()->dimension(0), "The C matrix must have the same number of columns as the matrix C");
-        ARM_COMPUTE_ERROR_ON_MSG(c->info()->dimension(0) != output->info()->dimension(0), "The C matrix must have the same number of rows as the output matrix");
-        ARM_COMPUTE_ERROR_ON_MSG(c->info()->dimension(1) != output->info()->dimension(1), "The C matrix must have the same number of columns as the output matrix");
+        ARM_COMPUTE_ERROR_ON_MISMATCHING_DATA_TYPES(a, c->info());
+        ARM_COMPUTE_ERROR_ON_MSG(a->dimension(1) != c->info()->dimension(1), "The C matrix must have the same number of rows as the matrix A");
+        ARM_COMPUTE_ERROR_ON_MSG(b->dimension(0) != c->info()->dimension(0), "The C matrix must have the same number of columns as the matrix B");
     }
 
-    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(output->total_size() != 0)
+    {
+        ARM_COMPUTE_RETURN_ERROR_ON_MSG(b->dimension(0) != output->dimension(0), "The output matrix must have the same number of columns as the matrix B");
+        ARM_COMPUTE_RETURN_ERROR_ON_MSG(a->dimension(1) != output->dimension(1), "The output matrix must have the same number of rows as the matrix A");
+    }
 
-    // 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;
+    ARM_COMPUTE_RETURN_ERROR_ON_MSG(a->dimension(0) != b->dimension(1), "The product AB is defined only if the number of columns in A is equal to the number of rows in B");
+
+    ARM_COMPUTE_UNUSED(alpha);
+    ARM_COMPUTE_UNUSED(beta);
+    ARM_COMPUTE_UNUSED(gemm_info);
+    return Status{};
+}
+} // namespace
+
+GCGEMM::GCGEMM(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),
+      _is_first_run(true), _reshape_b_only_on_first_run(false)
+{
+}
+
+void GCGEMM::configure(const IGCTensor *a, const IGCTensor *b, const IGCTensor *c, IGCTensor *output, float alpha, float beta, const GEMMInfo &gemm_info)
+{
+    ARM_COMPUTE_ERROR_ON_NULLPTR(a, b, output);
+
+    // Perform validation step
+    ARM_COMPUTE_ERROR_THROW_ON(validate_arguments(a->info(), b->info(), c, output->info(), alpha, beta, gemm_info));
+
+    // 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 IGCTensor *matrix_a = a;
     const IGCTensor *matrix_b = b;
 
+    // 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
+    const int m                         = a->info()->dimension(1);
+    const int n                         = b->info()->dimension(0);
+    const int k                         = a->info()->dimension(0);
+    int       mult_transpose1xW_width   = 1;
+    int       mult_interleave4x4_height = 1;
+
+    // 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;
+
     if(_is_interleaved_transposed)
     {
         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_gc_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);
+        // Manage intermediate buffers
         _memory_group.manage(&_tmp_a);
-
-        TensorInfo info_b(shape_tmp_b, 1, b->info()->data_type(), b->info()->fixed_point_position());
-        _tmp_b.allocator()->init(info_b);
-        if(!gemm_info.reshape_b_only_on_first_run())
+        if(!_reshape_b_only_on_first_run)
         {
             _memory_group.manage(&_tmp_b);
         }
+        // _tmp_a and _tmp_b will be auto configured in _interleave_kernel and in _transpose_kernel
 
         // Configure interleave kernel
         _interleave_kernel.configure(a, &_tmp_a);
@@ -104,7 +124,7 @@ void GCGEMM::configure(const IGCTensor *a, const IGCTensor *b, const IGCTensor *
         _transpose_kernel.configure(b, &_tmp_b);
     }
 
-    _mm_kernel.configure(matrix_a, matrix_b, output, alpha, _is_interleaved_transposed);
+    _mm_kernel.configure(matrix_a, matrix_b, output, alpha, _is_interleaved_transposed, GEMMReshapeInfo(m, n, k, mult_transpose1xW_width, mult_interleave4x4_height));
 
     if(_is_interleaved_transposed)
     {
@@ -121,6 +141,12 @@ void GCGEMM::configure(const IGCTensor *a, const IGCTensor *b, const IGCTensor *
     }
 }
 
+Status GCGEMM::validate(const ITensorInfo *a, const ITensorInfo *b, const IGCTensor *c, const ITensorInfo *output, const float alpha, const float beta, const GEMMInfo &gemm_info)
+{
+    ARM_COMPUTE_RETURN_ON_ERROR(validate_arguments(a, b, c, output, alpha, beta, gemm_info));
+    return Status{};
+}
+
 void GCGEMM::run()
 {
     _memory_group.acquire();
@@ -129,8 +155,17 @@ void GCGEMM::run()
         // Run interleave kernel
         GCScheduler::get().dispatch(_interleave_kernel, false);
 
-        // Run transpose kernel
-        GCScheduler::get().dispatch(_transpose_kernel, false);
+        if(_is_first_run)
+        {
+            // Run transpose kernel
+            GCScheduler::get().dispatch(_transpose_kernel, false);
+            _is_first_run = false;
+        }
+        else if(!_reshape_b_only_on_first_run)
+        {
+            // Run transpose kernel
+            GCScheduler::get().dispatch(_transpose_kernel, false);
+        }
         GCScheduler::get().memory_barrier();
     }