AI/rocm-systems
2
0
Tükrözés 0
Kód Hibajegyek Egyesítési kérések Actions Packages Projektek Kiadások Wiki Tevékenység

SWDEV-297448 - Add 64bit and 16bit write support

Forráskód böngészése 
For the fillBuffer shader, if there are two 32bit writes to a MMIO
register, it can get dropped. It has to be a single 64bit write.
Add optimization to fillBuffer to write 64bit and 16bit writes.

Change-Id: I3aa78e027898f8ae01e9c8f09004615673720c2b


[ROCm/clr commit: 21ba34d0fe]
This commit is contained in:
Saleel Kudchadker
2021-08-30 12:49:30 -07:00
szülő 5fd0dc4062
commit 36ec8c8871
7 fájl változott, egészen pontosan 125 új sor hozzáadva és 56 régi sor törölve
Download Patch File Download Diff File Expand all files Collapse all files
projects/clr/rocclr/device/gpu/gpublit.cpp
+32 -13
Fájl megtekintése
@@ -2034,21 +2034,40 @@ bool KernelBlitManager::fillBuffer(device::Memory& memory, const void* pattern,
synchronize();
return result;
} else {
uint fillType = FillBuffer;
uint fillType = FillBufferAligned;
size_t globalWorkOffset[3] = {0, 0, 0};
uint64_t fillSize = size[0] / patternSize;
size_t globalWorkSize = amd::alignUp(fillSize, 256);
size_t localWorkSize = 256;
bool dwordAligned = ((patternSize % sizeof(uint32_t)) == 0) ? true : false;
uint32_t alignment = (patternSize & 0x7) == 0 ?
sizeof(uint64_t) :
(patternSize & 0x3) == 0 ?
sizeof(uint32_t) :
(patternSize & 0x1) == 0 ?
sizeof(uint16_t) : sizeof(uint8_t);
// Program kernels arguments for the fill operation
Memory* mem = &gpuMem(memory);
if (dwordAligned) {
setArgument(kernels_[fillType], 0, sizeof(cl_mem), NULL);
if (alignment == sizeof(uint64_t)) {
setArgument(kernels_[fillType], 0, sizeof(cl_mem), nullptr);
setArgument(kernels_[fillType], 1, sizeof(cl_mem), nullptr);
setArgument(kernels_[fillType], 2, sizeof(cl_mem), nullptr);
setArgument(kernels_[fillType], 3, sizeof(cl_mem), &mem);
} else if (alignment == sizeof(uint32_t)) {
setArgument(kernels_[fillType], 0, sizeof(cl_mem), nullptr);
setArgument(kernels_[fillType], 1, sizeof(cl_mem), nullptr);
setArgument(kernels_[fillType], 2, sizeof(cl_mem), &mem);
setArgument(kernels_[fillType], 3, sizeof(cl_mem), nullptr);
} else if (alignment == sizeof(uint16_t)) {
setArgument(kernels_[fillType], 0, sizeof(cl_mem), nullptr);
setArgument(kernels_[fillType], 1, sizeof(cl_mem), &mem);
setArgument(kernels_[fillType], 2, sizeof(cl_mem), nullptr);
setArgument(kernels_[fillType], 3, sizeof(cl_mem), nullptr);
} else {
setArgument(kernels_[fillType], 0, sizeof(cl_mem), &mem);
setArgument(kernels_[fillType], 1, sizeof(cl_mem), NULL);
setArgument(kernels_[fillType], 1, sizeof(cl_mem), nullptr);
setArgument(kernels_[fillType], 2, sizeof(cl_mem), nullptr);
setArgument(kernels_[fillType], 3, sizeof(cl_mem), nullptr);
}
Memory* gpuCB = dev().getGpuMemory(constantBuffer_);
if (gpuCB == NULL) {
@@ -2057,15 +2076,15 @@ bool KernelBlitManager::fillBuffer(device::Memory& memory, const void* pattern,
void* constBuf = gpuCB->map(&gpu(), Resource::WriteOnly);
memcpy(constBuf, pattern, patternSize);
gpuCB->unmap(&gpu());
setArgument(kernels_[fillType], 2, sizeof(cl_mem), &gpuCB);
setArgument(kernels_[fillType], 4, sizeof(cl_mem), &gpuCB);
uint64_t offset = origin[0];
if (dwordAligned) {
patternSize /= sizeof(uint32_t);
offset /= sizeof(uint32_t);
}
setArgument(kernels_[fillType], 3, sizeof(uint32_t), &patternSize);
setArgument(kernels_[fillType], 4, sizeof(offset), &offset);
setArgument(kernels_[fillType], 5, sizeof(fillSize), &fillSize);
patternSize/= alignment;
offset /= alignment;
setArgument(kernels_[fillType], 5, sizeof(uint32_t), &patternSize);
setArgument(kernels_[fillType], 6, sizeof(offset), &offset);
setArgument(kernels_[fillType], 7, sizeof(fillSize), &fillSize);
// Create ND range object for the kernel's execution
amd::NDRangeContainer ndrange(1, globalWorkOffset, &globalWorkSize, &localWorkSize);