arm_compute v17.09

Change-Id: I4bf8f4e6e5f84ce0d5b6f5ba570d276879f42a81
diff --git a/src/runtime/CL/functions/CLGEMM.cpp b/src/runtime/CL/functions/CLGEMM.cpp
index 7408054..a81d113 100644
--- a/src/runtime/CL/functions/CLGEMM.cpp
+++ b/src/runtime/CL/functions/CLGEMM.cpp
@@ -38,20 +38,18 @@
 
 using namespace arm_compute;
 
-CLGEMM::CLGEMM()
-    : _interleave_kernel(), _transpose_kernel(), _mm_kernel(), _ma_kernel(), _tmp_a(), _tmp_b(), _run_vector_matrix_multiplication(false), _run_addition(false)
+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)
 {
 }
 
 void CLGEMM::configure(const ICLTensor *a, const ICLTensor *b, const ICLTensor *c, ICLTensor *output, float alpha, float beta)
 {
-    ARM_COMPUTE_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(a, 1, DataType::F32, DataType::F16);
-    ARM_COMPUTE_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(b, 1, DataType::F32, DataType::F16);
-    ARM_COMPUTE_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(output, 1, DataType::F32, DataType::F16);
+    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);
 
     if(c != nullptr)
     {
-        ARM_COMPUTE_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(c, 1, DataType::F32, DataType::F16);
         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");
@@ -59,13 +57,18 @@
         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, b, output);
     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");
 
-    // Check if the first input tensor is a vector. If so, all the kernels for reshaping the tensors can be skipped
-    if(a->info()->dimension(1) != 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;
+
+    const ICLTensor *matrix_a = a;
+    const ICLTensor *matrix_b = b;
+
+    if(_is_interleaved_transposed)
     {
-        _run_vector_matrix_multiplication = false;
+        matrix_a = &_tmp_a;
+        matrix_b = &_tmp_b;
 
         TensorShape shape_tmp_a = a->info()->tensor_shape();
         TensorShape shape_tmp_b = b->info()->tensor_shape();
@@ -73,27 +76,20 @@
         shape_tmp_a.set(0, a->info()->dimension(0) * 4);
         shape_tmp_a.set(1, std::ceil(a->info()->dimension(1) / 4.0f));
 
-        if(DataType::F32 == a->info()->data_type())
-        {
-            shape_tmp_b.set(0, b->info()->dimension(1) * 4);
-            shape_tmp_b.set(1, std::ceil(b->info()->dimension(0) / 4.0f));
-        }
-        else if(DataType::F16 == a->info()->data_type())
-        {
-            shape_tmp_b.set(0, b->info()->dimension(1) * 8);
-            shape_tmp_b.set(1, std::ceil(b->info()->dimension(0) / 8.0f));
-        }
-        else
-        {
-            ARM_COMPUTE_ERROR("DataType not supported");
-        }
+        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());
+        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());
+        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);
+
         // Configure interleave kernel
         _interleave_kernel.configure(a, &_tmp_a);
 
@@ -101,19 +97,17 @@
         _transpose_kernel.configure(b, &_tmp_b);
 
         // Configure matrix multiply kernel
-        _mm_kernel.configure(&_tmp_a, &_tmp_b, output, alpha);
+        _mm_kernel.set_target(CLScheduler::get().target());
+    }
 
+    _mm_kernel.configure(matrix_a, matrix_b, output, alpha, _is_interleaved_transposed);
+
+    if(_is_interleaved_transposed)
+    {
         // Allocate intermediate tensors
         _tmp_a.allocator()->allocate();
         _tmp_b.allocator()->allocate();
     }
-    else // The first input tensor is a vector
-    {
-        _run_vector_matrix_multiplication = true;
-
-        // Configure the matrix multiply kernel
-        _mm_kernel.configure(a, b, output, alpha);
-    }
 
     // Configure matrix addition kernel
     if(beta != 0 && c != nullptr)
@@ -125,7 +119,9 @@
 
 void CLGEMM::run()
 {
-    if(!_run_vector_matrix_multiplication)
+    _memory_group.acquire();
+
+    if(_is_interleaved_transposed)
     {
         // Run interleave kernel
         CLScheduler::get().enqueue(_interleave_kernel, false);
@@ -142,4 +138,6 @@
     {
         CLScheduler::get().enqueue(_ma_kernel);
     }
+
+    _memory_group.release();
 }