COMPMID-1052 - Rework validate method in CLGEMM

Change-Id: Iece5bd6478b5fac5164abff30c1e63e8a77291a9
Reviewed-on: https://eu-gerrit-1.euhpc.arm.com/130374
Reviewed-by: Anthony Barbier <anthony.barbier@arm.com>
Tested-by: Jenkins <bsgcomp@arm.com>

4 files changed, 27 insertions, 36 deletions

diff --git a/src/core/CL/kernels/CLGEMMInterleave4x4Kernel.cpp b/src/core/CL/kernels/CLGEMMInterleave4x4Kernel.cpp
index d12255ff24..8f669a9298 100644
--- a/src/core/CL/kernels/CLGEMMInterleave4x4Kernel.cpp
+++ b/src/core/CL/kernels/CLGEMMInterleave4x4Kernel.cpp
@@ -69,16 +69,16 @@ std::pair<Status, Window> validate_and_configure_window(ITensorInfo *input, ITen
     AccessWindowRectangle input_access(input, 0, 0, num_elems_processed_per_iteration_x, num_elems_processed_per_iteration_y);
     window_changed = window_changed || update_window_and_padding(win, input_access);
 
-    // Configure window in case of configured output
-    if(output->total_size() != 0)
-    {
-        const float scale_x = 4.0f * static_cast<float>(mult_interleave4x4_height);
-        const float scale_y = 1.0f / (scale_x);
+    // Output auto inizialitation if not yet initialized
+    auto_init_if_empty(*output, input->clone()->set_tensor_shape(compute_interleaved_shape(*input, mult_interleave4x4_height)));
 
-        AccessWindowRectangle output_access(output, 0, 0, num_elems_written_per_iteration, 1, scale_x, scale_y);
-        window_changed = window_changed || update_window_and_padding(win, output_access);
-        output_access.set_valid_region(win, input->valid_region());
-    }
+    // Configure window
+    const float scale_x = 4.0f * static_cast<float>(mult_interleave4x4_height);
+    const float scale_y = 1.0f / (scale_x);
+
+    AccessWindowRectangle output_access(output, 0, 0, num_elems_written_per_iteration, 1, scale_x, scale_y);
+    window_changed = window_changed || update_window_and_padding(win, output_access);
+    output_access.set_valid_region(win, input->valid_region());
 
     // Collapse along the Z direction
     // This collapse needs to be here in order to tune the Z dimension of LWS
diff --git a/src/core/CL/kernels/CLGEMMMatrixAdditionKernel.cpp b/src/core/CL/kernels/CLGEMMMatrixAdditionKernel.cpp
index 4b4814f15d..4538812baa 100644
--- a/src/core/CL/kernels/CLGEMMMatrixAdditionKernel.cpp
+++ b/src/core/CL/kernels/CLGEMMMatrixAdditionKernel.cpp
@@ -58,16 +58,15 @@ std::pair<Status, Window> validate_and_configure_window(ITensorInfo *input, ITen
 
 namespace
 {
-Status validate_arguments(const ITensorInfo *input, const ITensorInfo *output, const float beta)
+Status validate_arguments(const ITensorInfo *input, const ITensorInfo *output, float beta)
 {
     ARM_COMPUTE_RETURN_ERROR_ON_NULLPTR(input, output);
+    ARM_COMPUTE_UNUSED(input, output, beta);
 
     ARM_COMPUTE_RETURN_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(input, 1, DataType::QS8, DataType::QS16, DataType::F16, DataType::F32);
     ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_DATA_TYPES(input, output);
-    ARM_COMPUTE_RETURN_ERROR_ON(input->dimension(0) != output->dimension(0));
-    ARM_COMPUTE_RETURN_ERROR_ON(input->dimension(1) != output->dimension(1));
+    ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_SHAPES(input, output);
 
-    ARM_COMPUTE_UNUSED(beta);
     return Status{};
 }
 } // namespace
@@ -114,11 +113,10 @@ void CLGEMMMatrixAdditionKernel::configure(const ICLTensor *input, ICLTensor *ou
     ICLKernel::configure(win_config.second);
 }
 
-Status CLGEMMMatrixAdditionKernel::validate(const ITensorInfo *input, const ITensorInfo *output, const float beta)
+Status CLGEMMMatrixAdditionKernel::validate(const ITensorInfo *input, const ITensorInfo *output, float beta)
 {
-    ARM_COMPUTE_ERROR_ON_NULLPTR(input, output);
-    ARM_COMPUTE_RETURN_ERROR_ON(validate_arguments(input, output, beta));
-    ARM_COMPUTE_RETURN_ERROR_ON(validate_and_configure_window(input->clone().get(), output->clone().get()).first);
+    ARM_COMPUTE_RETURN_ON_ERROR(validate_arguments(input, output, beta));
+    ARM_COMPUTE_RETURN_ON_ERROR(validate_and_configure_window(input->clone().get(), output->clone().get()).first);
     return Status{};
 }
 
diff --git a/src/core/CL/kernels/CLGEMMMatrixMultiplyKernel.cpp b/src/core/CL/kernels/CLGEMMMatrixMultiplyKernel.cpp
index 674937eff0..cc9ae27bed 100644
--- a/src/core/CL/kernels/CLGEMMMatrixMultiplyKernel.cpp
+++ b/src/core/CL/kernels/CLGEMMMatrixMultiplyKernel.cpp
@@ -107,7 +107,7 @@ inline Status validate_arguments(const ITensorInfo *input0, const ITensorInfo *i
 }
 
 inline std::pair<Status, Window> validate_and_configure_window(ITensorInfo *input0, ITensorInfo *input1, ITensorInfo *output,
-                                                               bool is_interleaved_transposed, GPUTarget gpu_target,
+                                                               bool is_interleaved_transposed, const GEMMReshapeInfo &reshape_info, GPUTarget gpu_target,
                                                                ElementsProcessed &num_elements_processed)
 {
     bool   window_changed = false;
@@ -117,6 +117,9 @@ inline std::pair<Status, Window> validate_and_configure_window(ITensorInfo *inpu
     unsigned int &num_elems_processed_per_iteration_x = num_elements_processed[0];
     unsigned int &num_elems_processed_per_iteration_y = num_elements_processed[1];
 
+    // Output tensor auto inizialitation if not yet initialized
+    auto_init_if_empty(*output, input0->clone()->set_tensor_shape(compute_mm_shape(*input0, *input1, is_interleaved_transposed, reshape_info)));
+
     if(is_interleaved_transposed)
     {
         // Configure kernel window
@@ -182,13 +185,6 @@ void CLGEMMMatrixMultiplyKernel::configure(const ICLTensor *input0, const ICLTen
 {
     ARM_COMPUTE_ERROR_ON_NULLPTR(input0, input1, output);
 
-    // Output tensor auto inizialitation if not yet initialized
-    TensorShape tensor_shape{ input0->info()->tensor_shape() };
-    tensor_shape.set(0, is_interleaved_transposed ? reshape_info.n() : input1->info()->dimension(0));
-    tensor_shape.set(1, is_interleaved_transposed ? reshape_info.m() : input0->info()->dimension(1));
-
-    auto_init_if_empty(*output->info(), input0->info()->clone()->set_tensor_shape(tensor_shape));
-
     // Perform validate step
     ARM_COMPUTE_ERROR_THROW_ON(validate_arguments(input0->info(), input1->info(), output->info(), is_interleaved_transposed, reshape_info));
 
@@ -246,7 +242,7 @@ void CLGEMMMatrixMultiplyKernel::configure(const ICLTensor *input0, const ICLTen
     ElementsProcessed num_elements_processed{};
 
     // Configure kernel window
-    auto win_config = validate_and_configure_window(input0->info(), input1->info(), output->info(), is_interleaved_transposed, gpu_target, num_elements_processed);
+    auto win_config = validate_and_configure_window(input0->info(), input1->info(), output->info(), is_interleaved_transposed, reshape_info, gpu_target, num_elements_processed);
     ARM_COMPUTE_ERROR_THROW_ON(win_config.first);
     ICLKernel::configure(win_config.second);
 
@@ -354,6 +350,7 @@ Status CLGEMMMatrixMultiplyKernel::validate(const ITensorInfo *input0, const ITe
                                                               input1->clone().get(),
                                                               output->clone().get(),
                                                               is_interleaved_transposed,
+                                                              reshape_info,
                                                               gpu_target,
                                                               num_elements_processed)
                                 .first);
diff --git a/src/core/CL/kernels/CLGEMMTranspose1xWKernel.cpp b/src/core/CL/kernels/CLGEMMTranspose1xWKernel.cpp
index 54abb1cde0..a9618609e6 100644
--- a/src/core/CL/kernels/CLGEMMTranspose1xWKernel.cpp
+++ b/src/core/CL/kernels/CLGEMMTranspose1xWKernel.cpp
@@ -78,17 +78,13 @@ std::pair<Status, Window> validate_and_configure_window(ITensorInfo *input, ITen
 
     AccessWindowHorizontal input_access(input, 0, num_elems_processed_per_iteration);
 
+    // Output tensor auto inizialitation if not yet initialized
+    auto_init_if_empty(*output, input->clone()->set_tensor_shape(compute_transpose1xW_with_element_size_shape(*input, mult_transpose1xW_width)));
+
     // Configure window in case of configured output
-    if(output->total_size() != 0)
-    {
-        AccessWindowStatic output_access(output, 0, 0, ceil_to_multiple(output->dimension(0), scale_x), output->dimension(1));
-        window_changed = window_changed || update_window_and_padding(win, input_access, output_access);
-        output_access.set_valid_region(win, ValidRegion(Coordinates(0, 0), input->tensor_shape()));
-    }
-    else
-    {
-        window_changed = window_changed || update_window_and_padding(win, input_access);
-    }
+    AccessWindowStatic output_access(output, 0, 0, ceil_to_multiple(output->dimension(0), scale_x), output->dimension(1));
+    window_changed = window_changed || update_window_and_padding(win, input_access, output_access);
+    output_access.set_valid_region(win, ValidRegion(Coordinates(0, 0), input->tensor_shape()));
 
     // Collapse along the Z direction
     Window collapsed = win.collapse(win, Window::DimZ);