AI/rocm-systems
2
0
Garpu 0
Kode Masalah Tarik Permintaan Actions Packages Projects Rilis Wiki Kegiatan

SWDEV-455254 - Reduce blit kernels signature

Telusuri Sumber 
Remove offset from blit kernels, since it can be applied in setup.

Change-Id: I06b585068d68a0ee8e125ddf46a36fccb372f30d
This commit is contained in:
German Andryeyev
2024-04-09 14:27:04 -04:00
orang tua 35c80dd482
melakukan 7de7da4016
3 mengubah file dengan 47 tambahan dan 114 penghapusan
Download Patch File Download Diff File Expand all files Collapse all files
rocclr/device/pal/palblit.cpp
+17 -51
Melihat File
@@ -2194,37 +2194,7 @@ bool KernelBlitManager::fillBuffer(device::Memory& memory, const void* pattern,
// Program kernels arguments for the fill operation
Memory* mem = &gpuMem(memory);
if (alignment == 2 * sizeof(uint64_t)) {
setArgument(kernels_[kFillType], 0, sizeof(cl_mem), nullptr);
setArgument(kernels_[kFillType], 1, sizeof(cl_mem), nullptr);
setArgument(kernels_[kFillType], 2, sizeof(cl_mem), nullptr);
setArgument(kernels_[kFillType], 3, sizeof(cl_mem), nullptr);
setArgument(kernels_[kFillType], 4, sizeof(cl_mem), &mem);
} else if (alignment == sizeof(uint64_t)) {
setArgument(kernels_[kFillType], 0, sizeof(cl_mem), nullptr);
setArgument(kernels_[kFillType], 1, sizeof(cl_mem), nullptr);
setArgument(kernels_[kFillType], 2, sizeof(cl_mem), nullptr);
setArgument(kernels_[kFillType], 3, sizeof(cl_mem), &mem);
setArgument(kernels_[kFillType], 4, sizeof(cl_mem), nullptr);
} else if (alignment == sizeof(uint32_t)) {
setArgument(kernels_[kFillType], 0, sizeof(cl_mem), nullptr);
setArgument(kernels_[kFillType], 1, sizeof(cl_mem), nullptr);
setArgument(kernels_[kFillType], 2, sizeof(cl_mem), &mem);
setArgument(kernels_[kFillType], 3, sizeof(cl_mem), nullptr);
setArgument(kernels_[kFillType], 4, sizeof(cl_mem), nullptr);
} else if (alignment == sizeof(uint16_t)) {
setArgument(kernels_[kFillType], 0, sizeof(cl_mem), nullptr);
setArgument(kernels_[kFillType], 1, sizeof(cl_mem), &mem);
setArgument(kernels_[kFillType], 2, sizeof(cl_mem), nullptr);
setArgument(kernels_[kFillType], 3, sizeof(cl_mem), nullptr);
setArgument(kernels_[kFillType], 4, sizeof(cl_mem), nullptr);
} else {
setArgument(kernels_[kFillType], 0, sizeof(cl_mem), &mem);
setArgument(kernels_[kFillType], 1, sizeof(cl_mem), nullptr);
setArgument(kernels_[kFillType], 2, sizeof(cl_mem), nullptr);
setArgument(kernels_[kFillType], 3, sizeof(cl_mem), nullptr);
setArgument(kernels_[kFillType], 4, sizeof(cl_mem), nullptr);
}
setArgument(kernels_[kFillType], 0, sizeof(cl_mem), &mem, koffset);
const size_t localWorkSize = 256;
size_t globalWorkSize =
std::min(dev().settings().limit_blit_wg_ * localWorkSize, kfill_size);
@@ -2240,20 +2210,20 @@ bool KernelBlitManager::fillBuffer(device::Memory& memory, const void* pattern,
}
gpuCB.unmap(&gpu());
Memory* pGpuCB = &gpuCB;
setArgument(kernels_[kFillType], 5, sizeof(cl_mem), &pGpuCB);
setArgument(kernels_[kFillType], 1, sizeof(cl_mem), &pGpuCB);
uint64_t offset = origin[0];
// Adjust the pattern size in the copy type size
kpattern_size /= alignment;
setArgument(kernels_[kFillType], 6, sizeof(uint32_t), &kpattern_size);
koffset /= alignment;
setArgument(kernels_[kFillType], 7, sizeof(koffset), &koffset);
setArgument(kernels_[kFillType], 2, sizeof(uint32_t), &kpattern_size);
setArgument(kernels_[kFillType], 3, sizeof(alignment), &alignment);
// Calculate max id
uint64_t end_ptr = memory.virtualAddress() +
(koffset + kfill_size * kpattern_size) * alignment;
setArgument(kernels_[kFillType], 8, sizeof(end_ptr), &end_ptr);
uint64_t end_ptr = memory.virtualAddress() + koffset +
kfill_size * kpattern_size * alignment;
setArgument(kernels_[kFillType], 4, sizeof(end_ptr), &end_ptr);
uint32_t next_chunk = globalWorkSize * kpattern_size;
setArgument(kernels_[kFillType], 9, sizeof(uint32_t), &next_chunk);
setArgument(kernels_[kFillType], 5, sizeof(uint32_t), &next_chunk);
// Create ND range object for the kernel's execution
amd::NDRangeContainer ndrange(1, globalWorkOffset, &globalWorkSize, &localWorkSize);
@@ -2297,30 +2267,26 @@ bool KernelBlitManager::copyBuffer(device::Memory& srcMemory, device::Memory& ds
// Program kernels arguments for the blit operation
Memory* mem = &gpuMem(srcMemory);
setArgument(kernels_[kBlitType], 0, sizeof(cl_mem), &mem);
mem = &gpuMem(dstMemory);
setArgument(kernels_[kBlitType], 1, sizeof(cl_mem), &mem);
// Program source origin
uint64_t srcOffset = srcOrigin[0];
setArgument(kernels_[kBlitType], 2, sizeof(srcOffset), &srcOffset);
setArgument(kernels_[kBlitType], 0, sizeof(cl_mem), &mem, srcOffset);
mem = &gpuMem(dstMemory);
// Program destinaiton origin
uint64_t dstOffset = dstOrigin[0];
setArgument(kernels_[kBlitType], 3, sizeof(dstOffset), &dstOffset);
setArgument(kernels_[kBlitType], 1, sizeof(cl_mem), &mem, dstOffset);
uint64_t copySize = sizeIn[0];
setArgument(kernels_[kBlitType], 4, sizeof(copySize), &copySize);
setArgument(kernels_[kBlitType], 2, sizeof(copySize), &copySize);
setArgument(kernels_[kBlitType], 5, sizeof(remainder), &remainder);
setArgument(kernels_[kBlitType], 6, sizeof(aligned_size), &aligned_size);
setArgument(kernels_[kBlitType], 3, sizeof(remainder), &remainder);
setArgument(kernels_[kBlitType], 4, sizeof(aligned_size), &aligned_size);
// End pointer is the aligned copy size and destination offset
uint64_t end_ptr = dstMemory.virtualAddress() + dstOffset + sizeIn[0] - remainder;
setArgument(kernels_[kBlitType], 7, sizeof(end_ptr), &end_ptr);
setArgument(kernels_[kBlitType], 5, sizeof(end_ptr), &end_ptr);
uint32_t next_chunk = globalWorkSize;
setArgument(kernels_[kBlitType], 8, sizeof(next_chunk), &next_chunk);
setArgument(kernels_[kBlitType], 6, sizeof(next_chunk), &next_chunk);
// Create ND range object for the kernel's execution
amd::NDRangeContainer ndrange(1, nullptr, &globalWorkSize, &localWorkSize);