diff options
author | SiCong Li <sicong.li@arm.com> | 2023-05-16 16:46:34 +0100 |
---|---|---|
committer | SiCong Li <sicong.li@arm.com> | 2023-07-04 09:06:41 +0000 |
commit | 19844f605f5e5b71d05164711dee13f8652adafe (patch) | |
tree | 99aa1242a37aefda01e4284bc48c63f93f625258 /src/dynamic_fusion/runtime/gpu/cl/ClKernelRuntime.cpp | |
parent | 2f0ef00cc092b5eff901debee0c3f5a38114f40a (diff) | |
download | ComputeLibrary-19844f605f5e5b71d05164711dee13f8652adafe.tar.gz |
Add Kernel Writer driver code to dynamic fusion
* Partially port ElementwiseBinary component to ckw (broadcast not
supported yet)
* Port Store component to ckw
* Move KernelArgumentsHelpers to ckw_driver/ as it's only used by the
driver
ckw_driver is a middle layer between dynamic fusion and Compute Kernel
Writer (CKW). It consumes the fused kernel component stream produced by
Dynamic Fusion and uses CKW to write the kernel code complete with all
meta info needed by the runtime to enqueue the kernel.
It consists of two parts:
* Kernel writing: This resides in dynamic_fusion/sketch
* Runtime utilities: This resides in dynamic_fusion/runtime
The integration (separation between DF and CKW) occurs in two places:
* Inside GpuCKWDriver
global driver that coordinates how the
final fused kernel code is assembled together alongwith other meta
info needed by runtime.
* Inside each instantiated IGpuCKWComponentDriver
component driver that drives CKW to write component-specific code
or do component-specific configurations
Partially resolves: COMPMID-5792 COMPMID-6282 COMPMID-6260 COMPMID-6266
Signed-off-by: SiCong Li <sicong.li@arm.com>
Change-Id: Ib57a080a65fe8cfee1a8df1529fe572005a6d2f2
Reviewed-on: https://review.mlplatform.org/c/ml/ComputeLibrary/+/9847
Reviewed-by: Gunes Bayir <gunes.bayir@arm.com>
Tested-by: Arm Jenkins <bsgcomp@arm.com>
Comments-Addressed: Arm Jenkins <bsgcomp@arm.com>
Benchmark: Arm Jenkins <bsgcomp@arm.com>
Diffstat (limited to 'src/dynamic_fusion/runtime/gpu/cl/ClKernelRuntime.cpp')
-rw-r--r-- | src/dynamic_fusion/runtime/gpu/cl/ClKernelRuntime.cpp | 14 |
1 files changed, 14 insertions, 0 deletions
diff --git a/src/dynamic_fusion/runtime/gpu/cl/ClKernelRuntime.cpp b/src/dynamic_fusion/runtime/gpu/cl/ClKernelRuntime.cpp index b3ec39362c..6a57482bb2 100644 --- a/src/dynamic_fusion/runtime/gpu/cl/ClKernelRuntime.cpp +++ b/src/dynamic_fusion/runtime/gpu/cl/ClKernelRuntime.cpp @@ -59,6 +59,8 @@ void ClKernelRuntime::configure(const ClCompileContext &compile_ctx, const GpuKe inline void ClKernelRuntime::add_tensor_argument(unsigned int &idx, const GpuKernelArgumentInfo &arg, const ICLTensor *tensor, const Window &arg_slice, std::vector<cl::Image2D> &cl_images) { + ARM_COMPUTE_ERROR_ON_NULLPTR(tensor); + switch(arg.type) { case GpuKernelArgumentInfo::Type::Scalar: @@ -142,6 +144,18 @@ inline void ClKernelRuntime::add_tensor_argument(unsigned int &idx, const GpuKer add_4d_tensor_nhwc_argument(idx, tensor); break; } + case GpuKernelArgumentInfo::Type::Tensor_Special_0: + { + const ITensorInfo *info = tensor->info(); + const Strides &strides = info->strides_in_bytes(); + + _kernel.setArg(idx++, tensor->cl_buffer()); + const size_t dim1xdim2 = info->tensor_shape()[1] * info->tensor_shape()[2]; + _kernel.setArg<cl_int>(idx++, static_cast<int32_t>(dim1xdim2)); + const size_t stride1 = strides[1]; + _kernel.setArg<cl_int>(idx++, static_cast<int32_t>(stride1)); + break; + } default: { ARM_COMPUTE_ERROR("Unsupported"); |