diff options
Diffstat (limited to 'src/gpu/cl/operators')
| -rw-r--r-- | src/gpu/cl/operators/ClScatter.cpp | 57 | ||||
| -rw-r--r-- | src/gpu/cl/operators/ClScatter.h | 8 |
2 files changed, 53 insertions, 12 deletions
diff --git a/src/gpu/cl/operators/ClScatter.cpp b/src/gpu/cl/operators/ClScatter.cpp index af5fbb86f3..a11ecd7e6a 100644 --- a/src/gpu/cl/operators/ClScatter.cpp +++ b/src/gpu/cl/operators/ClScatter.cpp @@ -27,6 +27,7 @@ #include "arm_compute/runtime/CL/CLScheduler.h" #include "src/common/utils/Log.h" +#include "src/gpu/cl/kernels/ClCopyKernel.h" #include "src/gpu/cl/kernels/ClFillKernel.h" #include "src/gpu/cl/kernels/ClScatterKernel.h" @@ -47,9 +48,19 @@ Status ClScatter::validate(const ITensorInfo *src, const ScatterInfo &info) { ARM_COMPUTE_RETURN_ERROR_ON_NULLPTR(updates, indices, dst); - ARM_COMPUTE_RETURN_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(indices, 1, DataType::U32); + if (src != nullptr) + { + // Check dst/src are same shape and datatype. + ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_DIMENSIONS(src->tensor_shape(), dst->tensor_shape()); + ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_DATA_TYPES(src, updates, dst); + ARM_COMPUTE_RETURN_ON_ERROR(kernels::ClCopyKernel::validate(src, dst)); // Validate Copy kernel + } + if (src != dst) + { + ARM_COMPUTE_RETURN_ON_ERROR(kernels::ClFillKernel::validate(dst, PixelValue(0.0f))); // Validate Fill kernel. + } - return kernels::ClScatterKernel::validate(src, updates, indices, dst, info); + return kernels::ClScatterKernel::validate(updates, indices, dst, info); } void ClScatter::configure(const CLCompileContext &compile_context, @@ -61,11 +72,6 @@ void ClScatter::configure(const CLCompileContext &compile_context, { ARM_COMPUTE_ERROR_ON_NULLPTR(updates, indices, dst); ARM_COMPUTE_LOG_PARAMS(src, indices, dst, info); - ARM_COMPUTE_UNUSED(src); - ARM_COMPUTE_UNUSED(updates); - ARM_COMPUTE_UNUSED(indices); - ARM_COMPUTE_UNUSED(dst); - ARM_COMPUTE_UNUSED(info); // Perform validation step ARM_COMPUTE_ERROR_THROW_ON(validate(src, updates, indices, dst, info)); @@ -74,19 +80,50 @@ void ClScatter::configure(const CLCompileContext &compile_context, // If necessary, create fill kernel to fill dst tensor. if (_fill_zero) { - _fill_kernel = std::make_unique<kernels::ClFillKernel>(); + auto f = std::make_unique<kernels::ClFillKernel>(); + f->configure(compile_context, dst, PixelValue(0.0f)); + _fill_kernel = std::move(f); + } + else if (src != dst) // Check whether copying is necessary + { + // Fill dst with src copy here. + auto j = std::make_unique<kernels::ClCopyKernel>(); + j->configure(compile_context, src, dst); + _copy_kernel = std::move(j); + _run_copy = true; } // Configure ClScatterKernel auto k = std::make_unique<kernels::ClScatterKernel>(); k->set_target(CLScheduler::get().target()); - k->configure(compile_context, src, updates, indices, dst, info); + k->configure(compile_context, updates, indices, dst, info); _scatter_kernel = std::move(k); } void ClScatter::run(ITensorPack &tensors) { - ARM_COMPUTE_UNUSED(tensors); + // Get tensors. + auto src = tensors.get_const_tensor(ACL_SRC_0); + auto updates = tensors.get_const_tensor(ACL_SRC_1); + auto indices = tensors.get_const_tensor(ACL_SRC_2); + auto dst = tensors.get_tensor(ACL_DST); + + if (_fill_zero) + { + // Fill destination tensor with 0 values if zero init. + ITensorPack fill_pack{{ACL_SRC, dst}}; + CLScheduler::get().enqueue_op(*_fill_kernel, fill_pack, false); + } + + if (_run_copy) + { + // copy src to dst before scatter op. + ITensorPack copy_pack{{ACL_SRC, src}, {ACL_DST, dst}}; + CLScheduler::get().enqueue_op(*_copy_kernel, copy_pack, false); + } + + ITensorPack scatter_pack{{ACL_SRC_0, updates}, {ACL_SRC_1, indices}, {ACL_DST, dst}}; + CLScheduler::get().enqueue_op(*_scatter_kernel, scatter_pack, false); } } // namespace opencl diff --git a/src/gpu/cl/operators/ClScatter.h b/src/gpu/cl/operators/ClScatter.h index 433f7ca3a4..a1b32fed45 100644 --- a/src/gpu/cl/operators/ClScatter.h +++ b/src/gpu/cl/operators/ClScatter.h @@ -39,6 +39,7 @@ namespace opencl // Forward declaration class ClFillKernel; class ClScatterKernel; +class ClCopyKernel; /** Basic operator to execute Scatter on OpenCL. This operator calls the following OpenCL kernels: * @@ -56,13 +57,14 @@ public: * Valid data layouts: * - All * - * @note indices must always be U32 + * @note indices must always be S32. + * @note Negative indices are treated as out of bounds. * @note src, updates and dst tensors must be same datatype. * * @param[in] compile_context The compile context to be used. * @param[in] src Source input tensor info. Can be nullptr when using "Add" Scatter Function with zero initialization. * @param[in] updates Tensor info for tensor storing update values to use for scatter function. Data types supported: same as @p src. - * @param[in] indices Tensor info for tensor storing indices to use for scatter function. Data types supported: U32 only. + * @param[in] indices Tensor info for tensor storing indices to use for scatter function. Data types supported: S32 only. * @param[out] dst Output tensor to store the result of the Scatter Function. Data types supported: same as @p src and @p updates. * @param[in] Scatter_info Contains Scatter operation information described in @ref ScatterInfo. */ @@ -89,7 +91,9 @@ public: private: std::unique_ptr<opencl::IClKernel> _scatter_kernel{nullptr}; std::unique_ptr<opencl::IClKernel> _fill_kernel{nullptr}; + std::unique_ptr<opencl::IClKernel> _copy_kernel{nullptr}; bool _fill_zero{false}; + bool _run_copy{false}; }; } // namespace opencl } // namespace arm_compute |