Add GpuKernelArgumentBinding for runtime argument setting
* Add flexible runtime argument setting that accept argument bindings
exported from ckw.
* Introduce internal build flag ACL_INTERNAL_TEST_CKW_IN_DF. If set to
true, ckw will be tested in dynamic fusion validation tests. Otherwise
it will not be tested and the dynamic fusion will keep using
ClTemplateWriter instead.
* Fix CKW sampler for elementwise binary to deal with tile sizes > 1
in both dimensions
Resolves: COMPMID-6282
Partially resolves: COMPMID-6260
Signed-off-by: SiCong Li <sicong.li@arm.com>
Change-Id: I0ab225a4484eb2119643d900a4e72806558626ee
Reviewed-on: https://review.mlplatform.org/c/ml/ComputeLibrary/+/9917
Tested-by: Arm Jenkins <bsgcomp@arm.com>
Reviewed-by: Jakub Sujak <jakub.sujak@arm.com>
Reviewed-by: Anitha Raj <Anitha.Raj@arm.com>
Reviewed-by: Viet-Hoa Do <viet-hoa.do@arm.com>
Comments-Addressed: Arm Jenkins <bsgcomp@arm.com>
Benchmark: Arm Jenkins <bsgcomp@arm.com>
diff --git a/src/dynamic_fusion/runtime/gpu/cl/ClKernelRuntime.cpp b/src/dynamic_fusion/runtime/gpu/cl/ClKernelRuntime.cpp
index 6a57482..92ca855 100644
--- a/src/dynamic_fusion/runtime/gpu/cl/ClKernelRuntime.cpp
+++ b/src/dynamic_fusion/runtime/gpu/cl/ClKernelRuntime.cpp
@@ -24,6 +24,9 @@
#include "ClKernelRuntime.h"
#include "arm_compute/core/CL/ICLTensor.h"
#include "src/core/CL/CLUtils.h"
+#ifdef ACL_INTERNAL_TEST_CKW_IN_DF
+#include "src/dynamic_fusion/runtime/gpu/cl/ckw_driver/GpuCkwKernelArgumentsHelpers.h"
+#endif // ACL_INTERNAL_TEST_CKW_IN_DF
#include "src/dynamic_fusion/sketch/gpu/GpuKernelSourceCode.h"
#include "src/gpu/cl/ClKernelLibrary.h"
@@ -57,6 +60,8 @@
_arguments = code.arguments();
}
+#ifndef ACL_INTERNAL_TEST_CKW_IN_DF
+
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);
@@ -163,21 +168,65 @@
}
}
+#else // ACL_INTERNAL_TEST_CKW_IN_DF
+inline void ClKernelRuntime::add_kernel_argument(unsigned int &idx, const GpuKernelArgumentBinding &arg, const ICLTensor *tensor, std::vector<cl::Image2D> &cl_images)
+{
+ switch(arg.type())
+ {
+ case GpuKernelArgumentBinding::Type::TensorStorage:
+ {
+ switch(arg.tensor_storage_type())
+ {
+ case TensorStorageType::ClBufferUint8Ptr:
+ {
+ cl_add_buffer_argument(_kernel, idx, tensor->cl_buffer());
+ break;
+ }
+ case TensorStorageType::ClImage2dReadOnly:
+ {
+ cl::Image2D tensor_image2d = create_image2d_from_tensor(tensor, CLImage2DType::ReadOnly);
+ cl_images.push_back(tensor_image2d);
+ cl_add_texture_argument(_kernel, idx, tensor_image2d);
+ break;
+ }
+ case TensorStorageType::ClImage2dWriteOnly:
+ {
+ cl::Image2D tensor_image2d = create_image2d_from_tensor(tensor, CLImage2DType::WriteOnly);
+ cl_images.push_back(tensor_image2d);
+ cl_add_texture_argument(_kernel, idx, tensor_image2d);
+ break;
+ }
+ default:
+ {
+ ARM_COMPUTE_ERROR("Do not accept other TensorStorageType");
+ break;
+ }
+ }
+ break;
+ }
+ case GpuKernelArgumentBinding::Type::TensorComponent:
+ {
+ cl_add_tensor_component_argument(_kernel, idx, tensor, arg.tensor_component_type());
+ break;
+ }
+ default:
+ {
+ ARM_COMPUTE_ERROR("Do not accept other types of kernel arguments");
+ break;
+ }
+ }
+}
+
+#endif // ACL_INTERNAL_TEST_CKW_IN_DF
void ClKernelRuntime::run_op(ITensorPack &tensors, const Window &window, cl::CommandQueue &queue)
{
ARM_COMPUTE_ERROR_ON_UNCONFIGURED_KERNEL(this);
ARM_COMPUTE_ERROR_ON_INVALID_SUBWINDOW(ICLKernel::window(), window);
Window slice = window.first_slice_window_3D();
- // Don't slice matrix along the z dimension if matrix has just 2 dimensions and matrix A more than 2
- // This scenario can happen when the matrix multiplication is used to perform a convolution operation
- Window slice_fixed_z = slice;
- slice_fixed_z.set(Window::DimX, Window::Dimension(0, 1, 1));
- slice_fixed_z.set(Window::DimY, Window::Dimension(0, 1, 1));
/// NOTE: Parameters extracted from old kernels. So far they seem to be constant
/// but we may need to make them into another configuration passed from GpuWorkloadSourceCode if needed in the future
- constexpr bool slide_along_dimz = true;
constexpr bool skip_sliding_window = false;
constexpr bool use_dummy_work_items = false;
@@ -185,24 +234,28 @@
do
{
// Set kernel arguments
- Window arg_slice = slice;
// CLImages created from tensor arguments. Need to be retained until enqueue
std::vector<cl::Image2D> cl_images;
+#ifndef ACL_INTERNAL_TEST_CKW_IN_DF
for(auto id_arg : _arguments)
{
const auto arg = id_arg.second;
auto tensor = utils::cast::polymorphic_downcast<ICLTensor *>(tensors.get_tensor(id_arg.first));
ARM_COMPUTE_ERROR_ON_NULLPTR(tensor);
ARM_COMPUTE_ERROR_ON_NULLPTR(tensor->info());
- if(!slide_along_dimz)
- {
- // The stride_z for matrix must be zero if we do not slice
- ARM_COMPUTE_ERROR_ON(tensor->info()->strides_in_bytes()[3] != 0);
- arg_slice = slice_fixed_z;
- }
- add_tensor_argument(idx, *arg.kernel_argument_info(), tensor, arg_slice, cl_images);
+ add_tensor_argument(idx, *arg.kernel_argument_info(), tensor, slice, cl_images);
}
+#else // ACL_INTERNAL_TEST_CKW_IN_DF
+ for(const auto &arg : _arguments)
+ {
+ auto tensor = utils::cast::polymorphic_downcast<ICLTensor *>(tensors.get_tensor(arg.id()));
+ ARM_COMPUTE_ERROR_ON_NULLPTR(tensor);
+ ARM_COMPUTE_ERROR_ON_NULLPTR(tensor->info());
+ add_kernel_argument(idx, arg, tensor, cl_images);
+ }
+#endif // ACL_INTERNAL_TEST_CKW_IN_DF
+
// Dispatch kernel
enqueue(queue, *this, slice, lws_hint(), use_dummy_work_items);
}