From 62a00d336ef85ad18c11d77b1453b6fbbcd9fbeb Mon Sep 17 00:00:00 2001 From: foreman Date: Fri, 23 Mar 2018 14:12:49 -0400 Subject: [PATCH] P4 to Git Change 1531530 by gandryey@gera-w8 on 2018/03/23 13:45:11 SWDEV-79445 - OCL generic changes and code clean-up - Add explicit type specialization for AQL copy types to avoid memcpy() calls in the most of cases - Increase size of sysmem copy for constants buffers to account the internal arguments Affected files ... ... //depot/stg/opencl/drivers/opencl/runtime/device/pal/palconstbuf.cpp#5 edit ... //depot/stg/opencl/drivers/opencl/runtime/device/pal/palkernel.cpp#46 edit --- rocclr/runtime/device/pal/palconstbuf.cpp | 5 +- rocclr/runtime/device/pal/palkernel.cpp | 59 +++++++++++++++++------ 2 files changed, 47 insertions(+), 17 deletions(-) diff --git a/rocclr/runtime/device/pal/palconstbuf.cpp b/rocclr/runtime/device/pal/palconstbuf.cpp index 3a0aa1a1eb..5257ce3477 100644 --- a/rocclr/runtime/device/pal/palconstbuf.cpp +++ b/rocclr/runtime/device/pal/palconstbuf.cpp @@ -33,9 +33,10 @@ ManagedBuffer::~ManagedBuffer() { // ================================================================================================ bool ManagedBuffer::create(Resource::MemoryType type, bool constbuf) { if (constbuf) { - // Create sysmem copy for the constant buffer + // Create sysmem copy for the constant buffer. + // Allocate extra memory to account the internal arguments sysMemCopy_ = reinterpret_cast
(amd::AlignedMemory::allocate( - gpu_.dev().info().maxParameterSize_, 256)); + 2 * gpu_.dev().info().maxParameterSize_, 256)); if (sysMemCopy_ == nullptr) { LogPrintfError("We couldn't allocate sysmem copy for constant buffer, size(%d)!", size_); return false; diff --git a/rocclr/runtime/device/pal/palkernel.cpp b/rocclr/runtime/device/pal/palkernel.cpp index 05c22fe6f6..4381baf16a 100644 --- a/rocclr/runtime/device/pal/palkernel.cpp +++ b/rocclr/runtime/device/pal/palkernel.cpp @@ -874,21 +874,42 @@ void HSAILKernel::findLocalWorkSize(size_t workDim, const amd::NDRange& gblWorkS } } +template inline static void WriteAqlArg( unsigned char** dst, //!< The write pointer to the buffer - const void* src, //!< The source pointer + const T* src, //!< The source pointer uint size, //!< The size in bytes to copy - uint alignment = 0 //!< The alignment to follow while writing to the buffer - ) { - if (alignment == 0) { - *dst = amd::alignUp(*dst, size); - } else { - *dst = amd::alignUp(*dst, alignment); - } + uint alignment //!< The alignment to follow while writing to the buffer +) { + *dst = amd::alignUp(*dst, alignment); memcpy(*dst, src, size); *dst += size; } +template <> +inline static void WriteAqlArg( + unsigned char** dst, //!< The write pointer to the buffer + const uint32_t* src, //!< The source pointer + uint size, //!< The size in bytes to copy + uint alignment //!< The alignment to follow while writing to the buffer +) { + *dst = amd::alignUp(*dst, alignment); + *(reinterpret_cast(*dst)) = *src; + *dst += size; +} + +template <> +inline static void WriteAqlArg( + unsigned char** dst, //!< The write pointer to the buffer + const uint64_t* src, //!< The source pointer + uint size, //!< The size in bytes to copy + uint alignment //!< The alignment to follow while writing to the buffer +) { + *dst = amd::alignUp(*dst, alignment); + *(reinterpret_cast(*dst)) = *src; + *dst += size; +} + const uint16_t kDispatchPacketHeader = (HSA_PACKET_TYPE_KERNEL_DISPATCH << HSA_PACKET_HEADER_TYPE) | (1 << HSA_PACKET_HEADER_BARRIER) | (HSA_FENCE_SCOPE_SYSTEM << HSA_PACKET_HEADER_ACQUIRE_FENCE_SCOPE) | @@ -989,7 +1010,7 @@ hsa_kernel_dispatch_packet_t* HSAILKernel::loadArguments( amd::Memory* mem = nullptr; if (kernelParams.boundToSvmPointer(dev(), parameters, arg->index_)) { - WriteAqlArg(&aqlArgBuf, paramaddr, sizeof(paramaddr)); + WriteAqlArg(&aqlArgBuf, paramaddr, sizeof(paramaddr), sizeof(paramaddr)); mem = amd::SvmManager::FindSvmBuffer(*reinterpret_cast(paramaddr)); if (mem != nullptr) { gpuMem = dev().getGpuMemory(mem); @@ -1049,12 +1070,20 @@ hsa_kernel_dispatch_packet_t* HSAILKernel::loadArguments( cb->uploadDataToHw(arg->size_); // Then use a pointer in aqlArgBuffer to CB1 size_t gpuPtr = static_cast(cb->vmAddress() + cb->wrtOffset()); - WriteAqlArg(&aqlArgBuf, &gpuPtr, sizeof(size_t)); + WriteAqlArg(&aqlArgBuf, &gpuPtr, sizeof(size_t), sizeof(size_t)); gpu.addVmMemory(cb->activeMemory()); break; } case HSAIL_ARGTYPE_VALUE: - WriteAqlArg(&aqlArgBuf, paramaddr, arg->size_, arg->alignment_); + if (arg->size_ == sizeof(uint32_t)) { + WriteAqlArg(&aqlArgBuf, reinterpret_cast(paramaddr), + sizeof(uint32_t), arg->alignment_); + } else if (arg->size_ == sizeof(uint64_t)) { + WriteAqlArg(&aqlArgBuf, reinterpret_cast(paramaddr), + sizeof(uint64_t), arg->alignment_); + } else { + WriteAqlArg(&aqlArgBuf, paramaddr, arg->size_, arg->alignment_); + } break; case HSAIL_ARGTYPE_IMAGE: { Image* image = nullptr; @@ -1084,11 +1113,11 @@ hsa_kernel_dispatch_packet_t* HSAILKernel::loadArguments( cb->uploadDataToHw(HsaImageObjectSize); // Then use a pointer in aqlArgBuffer to CB1 uint64_t srd = cb->vmAddress() + cb->wrtOffset(); - WriteAqlArg(&aqlArgBuf, &srd, sizeof(srd)); + WriteAqlArg(&aqlArgBuf, &srd, sizeof(srd), sizeof(srd)); gpu.addVmMemory(cb->activeMemory()); } else { uint64_t srd = image->hwSrd(); - WriteAqlArg(&aqlArgBuf, &srd, sizeof(srd)); + WriteAqlArg(&aqlArgBuf, &srd, sizeof(srd), sizeof(srd)); srdResource = true; } @@ -1108,7 +1137,7 @@ hsa_kernel_dispatch_packet_t* HSAILKernel::loadArguments( const amd::Sampler* sampler = *reinterpret_cast(paramaddr); const Sampler* gpuSampler = static_cast(sampler->getDeviceSampler(dev())); uint64_t srd = gpuSampler->hwSrd(); - WriteAqlArg(&aqlArgBuf, &srd, sizeof(srd)); + WriteAqlArg(&aqlArgBuf, &srd, sizeof(srd), sizeof(srd)); srdResource = true; break; } @@ -1125,7 +1154,7 @@ hsa_kernel_dispatch_packet_t* HSAILKernel::loadArguments( } vmQueue = gpu.vQueue()->vmAddress(); } - WriteAqlArg(&aqlArgBuf, &vmQueue, sizeof(vmQueue)); + WriteAqlArg(&aqlArgBuf, &vmQueue, sizeof(vmQueue), sizeof(vmQueue)); break; } default: