COMPMID-434 - Port CLGEMM to support 16 bit fixed point
Change-Id: I30aef3c7ecd1ee740c2a7f2ce65a63c7dcd66e49
Reviewed-on: http://mpd-gerrit.cambridge.arm.com/79630
Reviewed-by: Anthony Barbier <anthony.barbier@arm.com>
Tested-by: Kaizen <jeremy.johnson+kaizengerrit@arm.com>
diff --git a/src/core/CL/kernels/CLGEMMInterleave4x4Kernel.cpp b/src/core/CL/kernels/CLGEMMInterleave4x4Kernel.cpp
index 3850c4d..5b6e0ec 100644
--- a/src/core/CL/kernels/CLGEMMInterleave4x4Kernel.cpp
+++ b/src/core/CL/kernels/CLGEMMInterleave4x4Kernel.cpp
@@ -43,7 +43,8 @@
void CLGEMMInterleave4x4Kernel::configure(const ICLTensor *input, ICLTensor *output)
{
- ARM_COMPUTE_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(input, 1, DataType::U8, DataType::S8, DataType::QS8, DataType::U16, DataType::S16, DataType::U32, DataType::S32, DataType::F16, DataType::F32);
+ ARM_COMPUTE_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(input, 1, DataType::U8, DataType::S8, DataType::QS8, DataType::U16, DataType::S16, DataType::QS16, DataType::U32, DataType::S32, DataType::F16,
+ DataType::F32);
ARM_COMPUTE_ERROR_ON_NULLPTR(output);
TensorShape output_shape = input->info()->tensor_shape();
diff --git a/src/core/CL/kernels/CLGEMMMatrixAdditionKernel.cpp b/src/core/CL/kernels/CLGEMMMatrixAdditionKernel.cpp
index 5883dd6..d1cdd7d 100644
--- a/src/core/CL/kernels/CLGEMMMatrixAdditionKernel.cpp
+++ b/src/core/CL/kernels/CLGEMMMatrixAdditionKernel.cpp
@@ -43,7 +43,7 @@
void CLGEMMMatrixAdditionKernel::configure(const ICLTensor *input, ICLTensor *output, float beta)
{
- ARM_COMPUTE_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(input, 1, DataType::QS8, DataType::F16, DataType::F32);
+ ARM_COMPUTE_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(input, 1, DataType::QS8, DataType::QS16, DataType::F16, DataType::F32);
ARM_COMPUTE_ERROR_ON_MISMATCHING_DATA_TYPES(input, output);
ARM_COMPUTE_ERROR_ON(input->info()->dimension(0) != output->info()->dimension(0));
ARM_COMPUTE_ERROR_ON(input->info()->dimension(1) != output->info()->dimension(1));
@@ -53,8 +53,19 @@
const unsigned int num_elems_processed_per_iteration = max_cl_vector_width / data_size_from_type(input->info()->data_type());
std::ostringstream ma_arguments;
- ma_arguments << "-DBETA=" << (input->info()->data_type() == DataType::QS8 ? scvt_qs8_f32(beta, input->info()->fixed_point_position()) : beta) << " ";
- ma_arguments << "-DFIXED_POINT_POSITION=" << input->info()->fixed_point_position();
+ if(is_data_type_fixed_point(input->info()->data_type()))
+ {
+ ma_arguments << "-DBETA=" << (input->info()->data_type() == DataType::QS8 ?
+ scvt_qs8_f32(beta, input->info()->fixed_point_position()) :
+ scvt_qs16_f32(beta, input->info()->fixed_point_position()))
+ << " ";
+ ma_arguments << "-DFIXED_POINT_POSITION=" << input->info()->fixed_point_position();
+ }
+ else
+ {
+ ma_arguments << "-DBETA=" << beta;
+ }
+
std::set<std::string> build_opts;
build_opts.emplace(ma_arguments.str());
diff --git a/src/core/CL/kernels/CLGEMMMatrixMultiplyKernel.cpp b/src/core/CL/kernels/CLGEMMMatrixMultiplyKernel.cpp
index 7c5b3d7..2d6b83a 100644
--- a/src/core/CL/kernels/CLGEMMMatrixMultiplyKernel.cpp
+++ b/src/core/CL/kernels/CLGEMMMatrixMultiplyKernel.cpp
@@ -50,7 +50,7 @@
void CLGEMMMatrixMultiplyKernel::configure(const ICLTensor *input0, const ICLTensor *input1, ICLTensor *output, float alpha)
{
- ARM_COMPUTE_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(input0, 1, DataType::QS8, DataType::F16, DataType::F32);
+ ARM_COMPUTE_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(input0, 1, DataType::QS8, DataType::QS16, DataType::F16, DataType::F32);
ARM_COMPUTE_ERROR_ON_MISMATCHING_DATA_TYPES(input0, input1, output);
ARM_COMPUTE_ERROR_ON_MISMATCHING_FIXED_POINT(input0, input1, output);
@@ -74,8 +74,18 @@
std::ostringstream mm_arguments;
mm_arguments << "-DWIDTH_MATRIX_B=" << input1->info()->dimension(0) << " ";
- mm_arguments << "-DALPHA=" << (input0->info()->data_type() == DataType::QS8 ? scvt_qs8_f32(alpha, input0->info()->fixed_point_position()) : alpha) << " ";
- mm_arguments << "-DFIXED_POINT_POSITION=" << input0->info()->fixed_point_position() << " ";
+ if(is_data_type_fixed_point(input0->info()->data_type()))
+ {
+ mm_arguments << "-DALPHA=" << (input0->info()->data_type() == DataType::QS8 ?
+ scvt_qs8_f32(alpha, input0->info()->fixed_point_position()) :
+ scvt_qs16_f32(alpha, input0->info()->fixed_point_position()))
+ << " ";
+ mm_arguments << "-DFIXED_POINT_POSITION=" << input0->info()->fixed_point_position() << " ";
+ }
+ else
+ {
+ mm_arguments << "-DALPHA=" << alpha << " ";
+ }
std::set<std::string> build_opts;
// Check if the output tensor is a vector. If so,the kernel runs the vector-matrix multiplication
diff --git a/src/core/CL/kernels/CLGEMMTranspose1xWKernel.cpp b/src/core/CL/kernels/CLGEMMTranspose1xWKernel.cpp
index ecef7e1..73c8429 100644
--- a/src/core/CL/kernels/CLGEMMTranspose1xWKernel.cpp
+++ b/src/core/CL/kernels/CLGEMMTranspose1xWKernel.cpp
@@ -40,7 +40,8 @@
void CLGEMMTranspose1xWKernel::configure(const ICLTensor *input, ICLTensor *output)
{
- ARM_COMPUTE_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(input, 1, DataType::U8, DataType::S8, DataType::QS8, DataType::U16, DataType::S16, DataType::U32, DataType::S32, DataType::F16, DataType::F32);
+ ARM_COMPUTE_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(input, 1, DataType::U8, DataType::S8, DataType::QS8, DataType::U16, DataType::S16, DataType::QS16, DataType::U32, DataType::S32, DataType::F16,
+ DataType::F32);
ARM_COMPUTE_ERROR_ON_NULLPTR(output);
TensorShape output_shape{ input->info()->tensor_shape() };