diff options
Diffstat (limited to 'src/core/CL/kernels')
3 files changed, 18 insertions, 7 deletions
diff --git a/src/core/CL/kernels/CLGEMMLowpMatrixMultiplyReshapedKernel.cpp b/src/core/CL/kernels/CLGEMMLowpMatrixMultiplyReshapedKernel.cpp index e9be1a6dfc..a8c1704d91 100644 --- a/src/core/CL/kernels/CLGEMMLowpMatrixMultiplyReshapedKernel.cpp +++ b/src/core/CL/kernels/CLGEMMLowpMatrixMultiplyReshapedKernel.cpp @@ -165,7 +165,7 @@ std::pair<Status, Window> validate_and_configure_window(ITensorInfo *input0, ITe } // namespace CLGEMMLowpMatrixMultiplyReshapedKernel::CLGEMMLowpMatrixMultiplyReshapedKernel() - : _input0(nullptr), _input1(nullptr), _output(nullptr), _slide_matrix_b(true), _reinterpret_output_as_3d(false), _k(1) + : _input0(nullptr), _input1(nullptr), _output(nullptr), _slide_matrix_b(true), _reinterpret_output_as_3d(false), _k(1), _use_dummy_work_items(false) { } @@ -181,6 +181,7 @@ void CLGEMMLowpMatrixMultiplyReshapedKernel::configure(const ICLTensor *input0, _output = output; _reinterpret_output_as_3d = (gemm_info.depth_output_gemm3d() != 0); _k = gemm_info.k(); + _use_dummy_work_items = preferred_dummy_work_items_support(CLKernelLibrary::get().get_device()); // Check if we need to slide the matrix B const unsigned int num_dimensions_input0 = _input0->info()->num_dimensions(); @@ -201,6 +202,9 @@ void CLGEMMLowpMatrixMultiplyReshapedKernel::configure(const ICLTensor *input0, build_opts.add_option_if(!_slide_matrix_b, "-DMATRIX_B_DEPTH=" + support::cpp11::to_string(input1->info()->dimension(2))); build_opts.add_option_if(lhs_info.interleave, "-DLHS_INTERLEAVE"); build_opts.add_option_if(rhs_info.interleave, "-DRHS_INTERLEAVE"); + build_opts.add_option_if(_use_dummy_work_items, "-DDUMMY_WORK_ITEMS"); + build_opts.add_option("-DM=" + support::cpp11::to_string(gemm_info.m())); + build_opts.add_option("-DN=" + support::cpp11::to_string(gemm_info.n())); build_opts.add_option("-DM0=" + support::cpp11::to_string(lhs_info.m0)); build_opts.add_option("-DN0=" + support::cpp11::to_string(rhs_info.n0)); build_opts.add_option("-DK0=" + support::cpp11::to_string(lhs_info.k0)); @@ -302,7 +306,7 @@ void CLGEMMLowpMatrixMultiplyReshapedKernel::run(const Window &window, cl::Comma _kernel.setArg<cl_uint>(idx++, static_cast<unsigned int>(_input0->info()->strides_in_bytes()[2])); _kernel.setArg<cl_uint>(idx++, static_cast<unsigned int>(_input1->info()->strides_in_bytes()[2])); _kernel.setArg<cl_uint>(idx++, static_cast<unsigned int>(_output->info()->strides_in_bytes()[2])); - enqueue(queue, *this, slice, lws_hint()); + enqueue(queue, *this, slice, lws_hint(), _use_dummy_work_items); } while(window.slide_window_slice_3D(slice)); }
\ No newline at end of file diff --git a/src/core/CL/kernels/CLGEMMMatrixMultiplyReshapedKernel.cpp b/src/core/CL/kernels/CLGEMMMatrixMultiplyReshapedKernel.cpp index b6816ac5c4..8969124d71 100644 --- a/src/core/CL/kernels/CLGEMMMatrixMultiplyReshapedKernel.cpp +++ b/src/core/CL/kernels/CLGEMMMatrixMultiplyReshapedKernel.cpp @@ -168,7 +168,7 @@ std::pair<Status, Window> validate_and_configure_window(ITensorInfo *input0, ITe } // namespace CLGEMMMatrixMultiplyReshapedKernel::CLGEMMMatrixMultiplyReshapedKernel() - : _input0(nullptr), _input1(nullptr), _output(nullptr), _slide_matrix_b(true), _reinterpret_output_as_3d(false), _k(1) + : _input0(nullptr), _input1(nullptr), _output(nullptr), _slide_matrix_b(true), _reinterpret_output_as_3d(false), _k(1), _use_dummy_work_items(false) { } @@ -184,6 +184,7 @@ void CLGEMMMatrixMultiplyReshapedKernel::configure(const ICLTensor *input0, cons _output = output; _reinterpret_output_as_3d = (gemm_info.depth_output_gemm3d() != 0); _k = gemm_info.k(); + _use_dummy_work_items = preferred_dummy_work_items_support(CLKernelLibrary::get().get_device()); // Check if we need to slide the matrix B const unsigned int num_dimensions_input0 = _input0->info()->num_dimensions(); @@ -206,6 +207,9 @@ void CLGEMMMatrixMultiplyReshapedKernel::configure(const ICLTensor *input0, cons build_opts.add_option_if(!_slide_matrix_b, "-DMATRIX_B_DEPTH=" + support::cpp11::to_string(input1->info()->dimension(2))); build_opts.add_option_if(lhs_info.interleave, "-DLHS_INTERLEAVE"); build_opts.add_option_if(rhs_info.interleave, "-DRHS_INTERLEAVE"); + build_opts.add_option_if(_use_dummy_work_items, "-DDUMMY_WORK_ITEMS"); + build_opts.add_option("-DM=" + support::cpp11::to_string(gemm_info.m())); + build_opts.add_option("-DN=" + support::cpp11::to_string(gemm_info.n())); build_opts.add_option("-DM0=" + support::cpp11::to_string(lhs_info.m0)); build_opts.add_option("-DN0=" + support::cpp11::to_string(rhs_info.n0)); build_opts.add_option("-DK0=" + support::cpp11::to_string(lhs_info.k0)); @@ -308,7 +312,7 @@ void CLGEMMMatrixMultiplyReshapedKernel::run(const Window &window, cl::CommandQu _kernel.setArg<cl_uint>(idx++, static_cast<unsigned int>(_input0->info()->strides_in_bytes()[2])); _kernel.setArg<cl_uint>(idx++, static_cast<unsigned int>(_input1->info()->strides_in_bytes()[2])); _kernel.setArg<cl_uint>(idx++, static_cast<unsigned int>(_output->info()->strides_in_bytes()[2])); - enqueue(queue, *this, slice, lws_hint()); + enqueue(queue, *this, slice, lws_hint(), _use_dummy_work_items); } while(window.slide_window_slice_3D(slice)); }
\ No newline at end of file diff --git a/src/core/CL/kernels/CLGEMMMatrixMultiplyReshapedOnlyRHSKernel.cpp b/src/core/CL/kernels/CLGEMMMatrixMultiplyReshapedOnlyRHSKernel.cpp index 6dc2a9e610..af06fecd00 100644 --- a/src/core/CL/kernels/CLGEMMMatrixMultiplyReshapedOnlyRHSKernel.cpp +++ b/src/core/CL/kernels/CLGEMMMatrixMultiplyReshapedOnlyRHSKernel.cpp @@ -161,7 +161,7 @@ std::pair<Status, Window> validate_and_configure_window(ITensorInfo *input0, ITe } // namespace CLGEMMMatrixMultiplyReshapedOnlyRHSKernel::CLGEMMMatrixMultiplyReshapedOnlyRHSKernel() - : _input0(nullptr), _input1(nullptr), _output(nullptr), _slide_matrix_b(true), _reinterpret_input_as_3d(false), _reinterpret_output_as_3d(false) + : _input0(nullptr), _input1(nullptr), _output(nullptr), _slide_matrix_b(true), _reinterpret_input_as_3d(false), _reinterpret_output_as_3d(false), _use_dummy_work_items(false) { } @@ -177,6 +177,7 @@ void CLGEMMMatrixMultiplyReshapedOnlyRHSKernel::configure(const ICLTensor *input _output = output; _reinterpret_input_as_3d = gemm_info.reinterpret_input_as_3d(); _reinterpret_output_as_3d = (gemm_info.depth_output_gemm3d() != 0); + _use_dummy_work_items = preferred_dummy_work_items_support(CLKernelLibrary::get().get_device()); // 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. @@ -202,13 +203,15 @@ void CLGEMMMatrixMultiplyReshapedOnlyRHSKernel::configure(const ICLTensor *input build_opts.add_option_if(_reinterpret_input_as_3d || _reinterpret_output_as_3d, "-DDEPTH_GEMM3D=" + support::cpp11::to_string(output->info()->dimension(2))); build_opts.add_option_if(!_slide_matrix_b, "-DMATRIX_B_DEPTH=" + support::cpp11::to_string(input1->info()->dimension(2))); build_opts.add_option_if(rhs_info.interleave, "-DRHS_INTERLEAVE"); + build_opts.add_option_if(_use_dummy_work_items, "-DDUMMY_WORK_ITEMS"); + build_opts.add_option("-DM=" + support::cpp11::to_string(gemm_info.m())); + build_opts.add_option("-DN=" + support::cpp11::to_string(gemm_info.n())); build_opts.add_option("-DK=" + support::cpp11::to_string(gemm_info.k())); build_opts.add_option("-DM0=" + support::cpp11::to_string(lhs_info.m0)); build_opts.add_option("-DN0=" + support::cpp11::to_string(rhs_info.n0)); build_opts.add_option("-DK0=" + support::cpp11::to_string(rhs_info.k0)); build_opts.add_option("-DH0=" + support::cpp11::to_string(rhs_info.h0)); - std::string kernel_name("gemm_mm_reshaped_only_rhs_"); kernel_name += rhs_info.transpose ? "t" : "nt"; @@ -308,7 +311,7 @@ void CLGEMMMatrixMultiplyReshapedOnlyRHSKernel::run(const Window &window, cl::Co _kernel.setArg<cl_uint>(idx++, static_cast<unsigned int>(_input0->info()->strides_in_bytes()[2])); _kernel.setArg<cl_uint>(idx++, static_cast<unsigned int>(_input1->info()->strides_in_bytes()[2])); _kernel.setArg<cl_uint>(idx++, static_cast<unsigned int>(_output->info()->strides_in_bytes()[2])); - enqueue(queue, *this, slice, lws_hint()); + enqueue(queue, *this, slice, lws_hint(), _use_dummy_work_items); } while(window.slide_window_slice_3D(slice)); } |