diff options
Diffstat (limited to 'src/runtime')
| -rw-r--r-- | src/runtime/CL/functions/CLFullyConnectedLayer.cpp | 4 | ||||
| -rw-r--r-- | src/runtime/gpu/cl/operators/ClGemm.cpp | 12 |
2 files changed, 12 insertions, 4 deletions
diff --git a/src/runtime/CL/functions/CLFullyConnectedLayer.cpp b/src/runtime/CL/functions/CLFullyConnectedLayer.cpp index 50a145f9ca..31c8908270 100644 --- a/src/runtime/CL/functions/CLFullyConnectedLayer.cpp +++ b/src/runtime/CL/functions/CLFullyConnectedLayer.cpp @@ -157,7 +157,8 @@ void CLFullyConnectedLayer::configure_mm(const CLCompileContext &compile_context gemmlowp_output_stage, // gemmlowp_output_stage fc_info.fp_mixed_precision, // fp_mixed_precision true, // broadcast_bias - fc_info.activation_info); // activation_info + fc_info.activation_info, // activation_info + fc_info.constant_weights); // constant_weights if(_is_quantized) { @@ -325,6 +326,7 @@ Status CLFullyConnectedLayer::validate(const ITensorInfo *input, const ITensorIn ARM_COMPUTE_RETURN_ERROR_ON(weights->num_dimensions() > 2); ARM_COMPUTE_RETURN_ERROR_ON(fc_info.activation_info.enabled() && is_data_type_quantized(input->data_type()) && fc_info.activation_info.activation() != ActivationLayerInfo::ActivationFunction::RELU && fc_info.activation_info.activation() != ActivationLayerInfo::ActivationFunction::BOUNDED_RELU && fc_info.activation_info.activation() != ActivationLayerInfo::ActivationFunction::LU_BOUNDED_RELU); + ARM_COMPUTE_RETURN_ERROR_ON(!fc_info.constant_weights && (!fc_info.are_weights_reshaped || fc_info.transpose_weights)); bool weights_reshaped = fc_info.transpose_weights ? fc_info.are_weights_reshaped : true; bool is_fc_after_conv = true; diff --git a/src/runtime/gpu/cl/operators/ClGemm.cpp b/src/runtime/gpu/cl/operators/ClGemm.cpp index fcbc6d5fba..a80375447d 100644 --- a/src/runtime/gpu/cl/operators/ClGemm.cpp +++ b/src/runtime/gpu/cl/operators/ClGemm.cpp @@ -78,8 +78,13 @@ inline bool validate_gemm_kernel(CLGEMMKernelType kernel_type) } } //Automatically select between mlgo (prioritized) and default heuristics for gemm kernel type -inline CLGEMMKernelType auto_select_gemm_kernel(auto_heuristics::CommonQuery query, bool reshape_b_only_on_first_run) +inline CLGEMMKernelType auto_select_gemm_kernel(auto_heuristics::CommonQuery query, bool reshape_b_only_on_first_run, bool constant_weights) { + if(!constant_weights) + { + return CLGEMMKernelType::NATIVE_V1; + } + auto gemm_kernel = auto_heuristics::select_mlgo_gemm_kernel(query, reshape_b_only_on_first_run); if(bool(gemm_kernel)) { @@ -564,7 +569,8 @@ void ClGemm::configure(const CLCompileContext &compile_context, ITensorInfo *a, const unsigned int batch_size = reinterpret_input_as_3d ? a->dimension(3) : a->dimension(2); // Select GEMMType - _gemm_kernel_type = auto_select_gemm_kernel(auto_heuristics::CommonQuery{ CLScheduler::get().target(), a->data_type(), m, n, k, batch_size }, _reshape_b_only_on_first_run); + _gemm_kernel_type = auto_select_gemm_kernel(auto_heuristics::CommonQuery{ CLScheduler::get().target(), a->data_type(), m, n, k, batch_size }, _reshape_b_only_on_first_run, + gemm_info.constant_weights()); const bool fuse_add_c = (!(helpers::float_ops::is_zero(beta)) && c != nullptr); @@ -613,7 +619,7 @@ Status ClGemm::validate(const ITensorInfo *a, const ITensorInfo *b, const ITenso { CLScheduler::get().target(), a->data_type(), m, n, k, batch_size, }, - gemm_info.reshape_b_only_on_first_run()); + gemm_info.reshape_b_only_on_first_run(), gemm_info.constant_weights()); const bool fuse_add_c = (!(helpers::float_ops::is_zero(beta)) && c != nullptr); |