commit 578ab61332aede4e11b7d5d92dbd72993d807a08
author Gian Marco Iodice <gianmarco.iodice@arm.com> Fri Jun 23 09:34:33 2017 +0100
committer Anthony Barbier <anthony.barbier@arm.com> Mon Sep 17 14:14:20 2018 +0100
tree 67e4588a570277a42e8a36c66eca24092e89c362
parent ec8b45e7333006ae06d834e300c11a4fd1b4c067

COMPMID-414 - Port CLConvolutionLayer to support 8 bit fixed point - CLGEMMMatrixAccumulateBiasesKernel

Change-Id: Idba13b578dc564b8003ce2fa3392eea2af3ce806
Reviewed-on: http://mpd-gerrit.cambridge.arm.com/78664
Tested-by: Kaizen <jeremy.johnson+kaizengerrit@arm.com>
Reviewed-by: Georgios Pinitas <georgios.pinitas@arm.com>

src/core/CL/kernels/CLGEMMMatrixAccumulateBiasesKernel.cpp

diff --git a/src/core/CL/kernels/CLGEMMMatrixAccumulateBiasesKernel.cpp b/src/core/CL/kernels/CLGEMMMatrixAccumulateBiasesKernel.cpp
index 289873c..75c1a6e 100644
--- a/src/core/CL/kernels/CLGEMMMatrixAccumulateBiasesKernel.cpp
+++ b/src/core/CL/kernels/CLGEMMMatrixAccumulateBiasesKernel.cpp
@@ -43,20 +43,22 @@
 
 void CLGEMMMatrixAccumulateBiasesKernel::configure(ICLTensor *accum, const ICLTensor *biases)
 {
-    ARM_COMPUTE_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(accum, 1, DataType::F16, DataType::F32);
-    ARM_COMPUTE_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(biases, 1, DataType::F16, DataType::F32);
+    ARM_COMPUTE_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(accum, 1, DataType::QS8, DataType::F16, DataType::F32);
     ARM_COMPUTE_ERROR_ON_MISMATCHING_DATA_TYPES(biases, accum);
+    ARM_COMPUTE_ERROR_ON_MISMATCHING_FIXED_POINT(biases, accum);
     ARM_COMPUTE_ERROR_ON(biases->info()->num_dimensions() != 1);
 
     _biases = biases;
     _accum  = accum;
 
+    std::set<std::string> build_opts;
+    build_opts.insert(("-DDATA_TYPE=" + get_cl_type_from_data_type(accum->info()->data_type())));
+
     // Create kernel
-    std::string data_type_name = lower_string(string_from_data_type(accum->info()->data_type()));
-    _kernel                    = static_cast<cl::Kernel>(CLKernelLibrary::get().create_kernel("gemm_accumulate_biases_" + data_type_name));
+    _kernel = static_cast<cl::Kernel>(CLKernelLibrary::get().create_kernel("gemm_accumulate_biases", build_opts));
 
     // Configure kernel window
-    const unsigned int num_elems_processed_per_iteration = max_cl_vector_width / data_size_from_type(accum->info()->data_type());
+    const unsigned int num_elems_processed_per_iteration = 16;
 
     Window win = calculate_max_window(*_accum->info(), Steps(num_elems_processed_per_iteration));