diff options
Diffstat (limited to 'src/runtime/CL/CLTensorAllocator.cpp')
| -rw-r--r-- | src/runtime/CL/CLTensorAllocator.cpp | 54 |
1 files changed, 43 insertions, 11 deletions
diff --git a/src/runtime/CL/CLTensorAllocator.cpp b/src/runtime/CL/CLTensorAllocator.cpp index 2b5fbb8241..eaf46d42ca 100644 --- a/src/runtime/CL/CLTensorAllocator.cpp +++ b/src/runtime/CL/CLTensorAllocator.cpp @@ -25,6 +25,7 @@ #include "arm_compute/core/Error.h" #include "arm_compute/core/TensorInfo.h" +#include "arm_compute/runtime/CL/CLRuntimeContext.h" #include "arm_compute/runtime/CL/CLScheduler.h" namespace arm_compute @@ -41,10 +42,10 @@ namespace * * @return A wrapped memory region */ -std::unique_ptr<ICLMemoryRegion> allocate_region(const cl::Context &context, size_t size, cl_uint alignment) +std::unique_ptr<ICLMemoryRegion> allocate_region(CLCoreRuntimeContext *ctx, size_t size, cl_uint alignment) { // Try fine-grain SVM - std::unique_ptr<ICLMemoryRegion> region = support::cpp14::make_unique<CLFineSVMMemoryRegion>(context, + std::unique_ptr<ICLMemoryRegion> region = support::cpp14::make_unique<CLFineSVMMemoryRegion>(ctx, CL_MEM_READ_WRITE | CL_MEM_SVM_FINE_GRAIN_BUFFER, size, alignment); @@ -52,12 +53,12 @@ std::unique_ptr<ICLMemoryRegion> allocate_region(const cl::Context &context, siz // Try coarse-grain SVM in case of failure if(region != nullptr && region->ptr() == nullptr) { - region = support::cpp14::make_unique<CLCoarseSVMMemoryRegion>(context, CL_MEM_READ_WRITE, size, alignment); + region = support::cpp14::make_unique<CLCoarseSVMMemoryRegion>(ctx, CL_MEM_READ_WRITE, size, alignment); } // Try legacy buffer memory in case of failure if(region != nullptr && region->ptr() == nullptr) { - region = support::cpp14::make_unique<CLBufferMemoryRegion>(context, CL_MEM_ALLOC_HOST_PTR | CL_MEM_READ_WRITE, size); + region = support::cpp14::make_unique<CLBufferMemoryRegion>(ctx, CL_MEM_ALLOC_HOST_PTR | CL_MEM_READ_WRITE, size); } return region; } @@ -103,8 +104,8 @@ void populate_quantization_info(CLFloatArray &scale, CLInt32Array &offset, const } } // namespace -CLTensorAllocator::CLTensorAllocator(IMemoryManageable *owner) - : _owner(owner), _associated_memory_group(nullptr), _memory(), _mapping(nullptr), _scale(), _offset() +CLTensorAllocator::CLTensorAllocator(IMemoryManageable *owner, CLRuntimeContext *ctx) + : _ctx(ctx), _owner(owner), _associated_memory_group(nullptr), _memory(), _mapping(nullptr), _scale(), _offset() { } @@ -129,7 +130,15 @@ void CLTensorAllocator::allocate() if(_associated_memory_group == nullptr) { // Perform memory allocation - _memory.set_owned_region(allocate_region(CLScheduler::get().context(), info().total_size(), 0)); + if(_ctx == nullptr) + { + auto legacy_ctx = CLCoreRuntimeContext(nullptr, CLScheduler::get().context(), CLScheduler::get().queue()); + _memory.set_owned_region(allocate_region(&legacy_ctx, info().total_size(), 0)); + } + else + { + _memory.set_owned_region(allocate_region(_ctx->core_runtime_context(), info().total_size(), 0)); + } } else { @@ -162,9 +171,17 @@ Status CLTensorAllocator::import_memory(cl::Buffer buffer) ARM_COMPUTE_RETURN_ERROR_ON(buffer.getInfo<CL_MEM_CONTEXT>().get() != CLScheduler::get().context().get()); ARM_COMPUTE_RETURN_ERROR_ON(_associated_memory_group != nullptr); - _memory.set_owned_region(support::cpp14::make_unique<CLBufferMemoryRegion>(buffer)); - info().set_is_resizable(false); + if(_ctx == nullptr) + { + auto legacy_ctx = CLCoreRuntimeContext(nullptr, CLScheduler::get().context(), CLScheduler::get().queue()); + _memory.set_owned_region(support::cpp14::make_unique<CLBufferMemoryRegion>(buffer, &legacy_ctx)); + } + else + { + _memory.set_owned_region(support::cpp14::make_unique<CLBufferMemoryRegion>(buffer, _ctx->core_runtime_context())); + } + info().set_is_resizable(false); return Status{}; } @@ -179,13 +196,28 @@ void CLTensorAllocator::set_associated_memory_group(IMemoryGroup *associated_mem uint8_t *CLTensorAllocator::lock() { - return map(CLScheduler::get().queue(), true); + if(_ctx) + { + return map(_ctx->gpu_scheduler()->queue(), true); + } + else + { + return map(CLScheduler::get().queue(), true); + } } void CLTensorAllocator::unlock() { ARM_COMPUTE_ERROR_ON(_memory.region() == nullptr); - unmap(CLScheduler::get().queue(), reinterpret_cast<uint8_t *>(_memory.region()->buffer())); + if(_ctx) + { + unmap(_ctx->gpu_scheduler()->queue(), reinterpret_cast<uint8_t *>(_memory.region()->buffer())); + } + else + { + //Legacy singleton api + unmap(CLScheduler::get().queue(), reinterpret_cast<uint8_t *>(_memory.region()->buffer())); + } } uint8_t *CLTensorAllocator::map(cl::CommandQueue &q, bool blocking) |