COMPMID-1607 - (Nightly) CLGEMMLowpMatrixMultiplyCore errors and mismatches
Change-Id: I5f2e6843526cb154176a5b113627d4f36c3a8edd
Reviewed-on: https://eu-gerrit-1.euhpc.arm.com/150967
Reviewed-by: Georgios Pinitas <georgios.pinitas@arm.com>
Tested-by: bsgcomp <bsgcomp@arm.com>
diff --git a/src/core/CL/kernels/CLGEMMLowpMatrixMultiplyKernel.cpp b/src/core/CL/kernels/CLGEMMLowpMatrixMultiplyKernel.cpp
index ee364e5..56f318d 100644
--- a/src/core/CL/kernels/CLGEMMLowpMatrixMultiplyKernel.cpp
+++ b/src/core/CL/kernels/CLGEMMLowpMatrixMultiplyKernel.cpp
@@ -59,6 +59,8 @@
{
ARM_COMPUTE_RETURN_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(input0, 1, DataType::QASYMM8);
ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_DATA_TYPES(input0, input1);
+ ARM_COMPUTE_RETURN_ERROR_ON_MSG(input0->num_dimensions() > 4, "The number of dimensions for the matrix A must be <= 4");
+ ARM_COMPUTE_RETURN_ERROR_ON_MSG(input1->num_dimensions() > 3, "The number of dimensions for the matrix B must be <= 3");
ARM_COMPUTE_RETURN_ERROR_ON_MSG(is_interleaved_transposed && reshape_info.reinterpret_input_as_3d(), "The input tensor cannot be reinterpreted as 3D if is_interleaved_transposed is true");
ARM_COMPUTE_RETURN_ERROR_ON_MSG(input1->num_dimensions() > 2 && reshape_info.reinterpret_input_as_3d(), "The input1 tensor cannot have more than 2 dimensions if input0 has to be reinterpreted as 3D");
@@ -85,7 +87,7 @@
const TensorInfo tensor_info0 = input0->clone()->set_tensor_shape(tensor_shape0);
const TensorInfo tensor_info1 = input1->clone()->set_tensor_shape(tensor_shape1);
- const TensorInfo tensor_info_reshaped0 = input0->clone()->set_tensor_shape(compute_interleaved_shape(tensor_info0, mult_interleave4x4_height));
+ const TensorInfo tensor_info_reshaped0 = input0->clone()->set_tensor_shape(compute_interleaved_shape(tensor_info0, mult_interleave4x4_height, reshape_info.reinterpret_input_as_3d()));
const TensorInfo tensor_info_reshaped1 = input1->clone()->set_tensor_shape(compute_transpose1xW_with_element_size_shape(tensor_info1, mult_transpose1xW_width));
ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_SHAPES(input0, &tensor_info_reshaped0);
@@ -122,8 +124,11 @@
reinterpret_output_as_3d = false;
}
+ GEMMReshapeInfo reshape_info_to_use = GEMMReshapeInfo(reshape_info.m(), reshape_info.n(), reshape_info.k(), reshape_info.mult_transpose1xW_width(), reshape_info.mult_interleave4x4_height(),
+ reinterpret_output_as_3d ? reshape_info.depth_output_gemm3d() : 1, reinterpret_input_as_3d);
+
// 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)));
+ auto_init_if_empty(*output, input0->clone()->set_tensor_shape(compute_mm_shape(*input0, *input1, is_interleaved_transposed, reshape_info_to_use)));
TensorInfo tmp_info(*output);
@@ -140,7 +145,7 @@
if(is_interleaved_transposed)
{
// reinterpret_input_as_3d is not supported if is_interleaved_transposed is set
- ARM_COMPUTE_ERROR_ON(reshape_info.reinterpret_input_as_3d());
+ ARM_COMPUTE_ERROR_ON(reinterpret_input_as_3d);
// Configure kernel window
num_elems_processed_per_iteration_x = 4;
@@ -216,13 +221,6 @@
{
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(), tensor_shape, 1, DataType::S32, QuantizationInfo());
-
ARM_COMPUTE_ERROR_THROW_ON(validate_arguments(input0->info(), input1->info(), output->info(), is_interleaved_transposed, reshape_info));
_input0 = input0;
diff --git a/src/core/CL/kernels/CLGEMMMatrixMultiplyKernel.cpp b/src/core/CL/kernels/CLGEMMMatrixMultiplyKernel.cpp
index 715edae..c8bcb37 100644
--- a/src/core/CL/kernels/CLGEMMMatrixMultiplyKernel.cpp
+++ b/src/core/CL/kernels/CLGEMMMatrixMultiplyKernel.cpp
@@ -120,8 +120,12 @@
reinterpret_output_as_3d = false;
}
+ const GEMMReshapeInfo reshape_info_to_use(reshape_info.m(), reshape_info.n(), reshape_info.k(), reshape_info.mult_transpose1xW_width(),
+ reshape_info.mult_interleave4x4_height(), reinterpret_output_as_3d ? reshape_info.depth_output_gemm3d() : 1, reinterpret_input_as_3d);
+
// 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)));
+ auto_init_if_empty(*output, input0->clone()->set_tensor_shape(compute_mm_shape(*input0, *input1, is_interleaved_transposed,
+ reshape_info_to_use)));
TensorInfo tmp_info(*output);
@@ -137,7 +141,7 @@
if(is_interleaved_transposed)
{
// reinterpret_input_as_3d is not supported if is_interleaved_transposed is set
- ARM_COMPUTE_ERROR_ON(reshape_info.reinterpret_input_as_3d());
+ ARM_COMPUTE_ERROR_ON(reinterpret_input_as_3d);
// Configure kernel window
num_elems_processed_per_iteration_x = max_cl_vector_width / data_size_from_type(data_type);
diff --git a/src/runtime/CL/functions/CLGEMM.cpp b/src/runtime/CL/functions/CLGEMM.cpp
index 821464e..0c82e6d 100644
--- a/src/runtime/CL/functions/CLGEMM.cpp
+++ b/src/runtime/CL/functions/CLGEMM.cpp
@@ -200,7 +200,8 @@
// If we pass the matrix A and matrix B reshaped to CLGEMMMatrixMultiplyKernel, we need to pass m, n, k, mult_transpose1xW_width and mult_interleave4x4_height to CLGEMMReshapeInfo
// in order to know how the matrices have been reshaped
bool reinterpret_input_as_3d = gemm_info.reinterpret_input_as_3d();
- const int m = reinterpret_input_as_3d ? (a->dimension(1) * a->dimension(2)) : a->dimension(1);
+ bool reinterpret_output_as_3d = (gemm_info.depth_output_gemm3d() != 1);
+ const int m = a->dimension(1);
const int n = b->dimension(0);
const int k = a->dimension(0);
int mult_transpose1xW_width = 1;
@@ -216,13 +217,21 @@
// Check if we need to reshape the matrix A and matrix B
const bool run_interleave_transpose = is_interleaved_transposed(m, n, k, a->data_type(), reshape_b_only_on_first_run, gpu_target);
+ // In case both input and output have to be reinterpreted as 3D tensors,
+ // force reinterpret_input_as_3d and reinterpret_output_as_3d to be false.
+ if(reinterpret_input_as_3d == reinterpret_output_as_3d)
+ {
+ reinterpret_input_as_3d = false;
+ reinterpret_output_as_3d = false;
+ }
+
// if _is_interleaved_transposed is set, force reinterpret_input_as_3d to be false as the output of CLGEMMInterleaveKernel will be 2D
if(run_interleave_transpose)
{
reinterpret_input_as_3d = false;
}
- const GEMMReshapeInfo reshape_info = GEMMReshapeInfo(m, n, k, mult_transpose1xW_width, mult_interleave4x4_height, depth_output_gemm3d, reinterpret_input_as_3d);
+ const GEMMReshapeInfo reshape_info = GEMMReshapeInfo(m, n, k, mult_transpose1xW_width, mult_interleave4x4_height, reinterpret_output_as_3d ? depth_output_gemm3d : 1, reinterpret_input_as_3d);
if(run_interleave_transpose)
{
@@ -230,8 +239,8 @@
matrix_b_info = &tmp_b_info;
// Validate interleave kernel
- auto_init_if_empty(tmp_a_info, a->clone()->set_tensor_shape(compute_interleaved_shape(*a, mult_interleave4x4_height, gemm_info.reinterpret_input_as_3d())));
- ARM_COMPUTE_RETURN_ON_ERROR(CLGEMMInterleave4x4Kernel::validate(a, &tmp_a_info, mult_interleave4x4_height, gemm_info.reinterpret_input_as_3d()));
+ auto_init_if_empty(tmp_a_info, a->clone()->set_tensor_shape(compute_interleaved_shape(*a, mult_interleave4x4_height, reinterpret_input_as_3d)));
+ ARM_COMPUTE_RETURN_ON_ERROR(CLGEMMInterleave4x4Kernel::validate(a, &tmp_a_info, mult_interleave4x4_height, reinterpret_input_as_3d));
// Validate transpose kernel
auto_init_if_empty(tmp_b_info, b->clone()->set_tensor_shape(compute_transpose1xW_with_element_size_shape(*b, mult_transpose1xW_width)));
diff --git a/src/runtime/CL/functions/CLGEMMLowpMatrixMultiplyCore.cpp b/src/runtime/CL/functions/CLGEMMLowpMatrixMultiplyCore.cpp
index 62e7ee7..7aeade1 100644
--- a/src/runtime/CL/functions/CLGEMMLowpMatrixMultiplyCore.cpp
+++ b/src/runtime/CL/functions/CLGEMMLowpMatrixMultiplyCore.cpp
@@ -108,24 +108,31 @@
// If we pass the matrix A and matrix B reshaped to CLGEMMMatrixMultiplyKernel, we need to pass m, n, k, mult_transpose1xW_width and mult_interleave4x4_height to CLGEMMReshapeInfo
// in order to know how the matrices have been reshaped
bool reinterpret_input_as_3d = gemm_info.reinterpret_input_as_3d();
- const int m = reinterpret_input_as_3d ? (a->info()->dimension(1) * a->info()->dimension(2)) : a->info()->dimension(1);
+ bool reinterpret_output_as_3d = (gemm_info.depth_output_gemm3d() != 1);
+ int m = reinterpret_input_as_3d ? (a->info()->dimension(1) * a->info()->dimension(2)) : a->info()->dimension(1);
const int n = b->info()->dimension(0);
const int k = a->info()->dimension(0);
const int depth_output_gemm3d = gemm_info.depth_output_gemm3d();
constexpr int mult_transpose1xW_width = 1;
constexpr int mult_interleave4x4_height = 1;
+ // In case both input and output have to be reinterpreted as 3D tensors,
+ // force reinterpret_input_as_3d and reinterpret_output_as_3d to be false.
+ if(reinterpret_input_as_3d == reinterpret_output_as_3d)
+ {
+ reinterpret_input_as_3d = false;
+ reinterpret_output_as_3d = false;
+ }
+
// Check if we need to reshape the matrix A and matrix B
_is_interleaved_transposed = is_interleaved_transposed(m, n, k, _reshape_b_only_on_first_run, gpu_target);
- // if _is_interleaved_transposed is set, force reinterpret_input_as_3d to be false as the output of CLGEMMInterleaveKernel will be 2D
if(_is_interleaved_transposed)
{
+ // if _is_interleaved_transposed is set, force reinterpret_input_as_3d to be false as the output of CLGEMMInterleaveKernel will be 2D
reinterpret_input_as_3d = false;
- }
+ m = a->info()->dimension(1);
- if(_is_interleaved_transposed)
- {
matrix_a = &_tmp_a;
matrix_b = &_tmp_b;
@@ -136,7 +143,7 @@
}
// Configure interleave kernel
- _mtx_a_reshape_kernel.configure(a, &_tmp_a, mult_interleave4x4_height, gemm_info.reinterpret_input_as_3d());
+ _mtx_a_reshape_kernel.configure(a, &_tmp_a, mult_interleave4x4_height, reinterpret_input_as_3d);
// Configure transpose kernel
_mtx_b_reshape_kernel.configure(b, &_tmp_b, mult_transpose1xW_width);
@@ -144,7 +151,7 @@
// Configure matrix multiply kernel
_mm_kernel.configure(matrix_a, matrix_b, output, _is_interleaved_transposed, GEMMReshapeInfo(m, n, k,
mult_transpose1xW_width, mult_interleave4x4_height,
- depth_output_gemm3d, reinterpret_input_as_3d));
+ reinterpret_output_as_3d ? depth_output_gemm3d : 1, reinterpret_input_as_3d));
// Initialize matrix B reduction kernel only if _a_offset is not equal to 0
if(_a_offset != 0)
@@ -207,13 +214,22 @@
int32_t b_offset = b->quantization_info().offset;
bool reinterpret_input_as_3d = gemm_info.reinterpret_input_as_3d();
- const int m = reinterpret_input_as_3d ? (a->dimension(1) * a->dimension(2)) : a->dimension(1);
+ bool reinterpret_output_as_3d = (gemm_info.depth_output_gemm3d() != 1);
+ int m = a->dimension(1);
const int n = b->dimension(0);
const int k = a->dimension(0);
constexpr int mult_transpose1xW_width = 1;
constexpr int mult_interleave4x4_height = 1;
const int depth_output_gemm3d = gemm_info.depth_output_gemm3d();
+ // In case both input and output have to be reinterpreted as 3D tensors,
+ // force reinterpret_input_as_3d and reinterpret_output_as_3d to be false.
+ if(reinterpret_input_as_3d == reinterpret_output_as_3d)
+ {
+ reinterpret_input_as_3d = false;
+ reinterpret_output_as_3d = false;
+ }
+
bool reshape_matrices = is_interleaved_transposed(m, n, k, gemm_info.reshape_b_only_on_first_run(), CLScheduler::get().target());
// if reshape_matrices is set, force reinterpret_input_as_3d to be false as the output of CLGEMMInterleaveKernel will be 2D
@@ -222,14 +238,14 @@
reinterpret_input_as_3d = false;
}
- const GEMMReshapeInfo reshape_info = GEMMReshapeInfo(m, n, k, mult_transpose1xW_width, mult_interleave4x4_height, depth_output_gemm3d, reinterpret_input_as_3d);
+ const GEMMReshapeInfo reshape_info = GEMMReshapeInfo(m, n, k, mult_transpose1xW_width, mult_interleave4x4_height, reinterpret_output_as_3d ? depth_output_gemm3d : 1, reinterpret_input_as_3d);
if(reshape_matrices)
{
- TensorInfo info_a(compute_interleaved_shape(*a, mult_interleave4x4_height, gemm_info.reinterpret_input_as_3d()), 1, a->data_type());
+ TensorInfo info_a(compute_interleaved_shape(*a, mult_interleave4x4_height, reinterpret_input_as_3d), 1, a->data_type());
TensorInfo info_b(compute_transpose1xW_with_element_size_shape(*b, mult_transpose1xW_width), 1, b->data_type());
- ARM_COMPUTE_RETURN_ON_ERROR(CLGEMMInterleave4x4Kernel::validate(a, &info_a, mult_interleave4x4_height, gemm_info.reinterpret_input_as_3d()));
+ ARM_COMPUTE_RETURN_ON_ERROR(CLGEMMInterleave4x4Kernel::validate(a, &info_a, mult_interleave4x4_height, reinterpret_input_as_3d));
ARM_COMPUTE_RETURN_ON_ERROR(CLGEMMTranspose1xWKernel::validate(b, &info_b, mult_transpose1xW_width));
ARM_COMPUTE_RETURN_ON_ERROR(CLGEMMLowpMatrixMultiplyKernel::validate(&info_a, &info_b, output, reshape_matrices, reshape_info));
}