AI/rocm-systems
2
0
Fork 0
Code Issues Pull Requests Actions Packages Projects Releases Wiki Activity

SWDEV-312822 - Revert "SWDEV-310187 - Change flag to keep track of aligned sizes instead of expanded patterns."

Browse Source 
This reverts commit 8307886644.

Change-Id: I022c2a8375f9929e9723cec66e1e0b960263fc39
This commit is contained in:
kjayapra-amd
2021-11-24 19:18:18 -08:00
committed by Maneesh Gupta
parent 2c23a1091f
commit 2e9bc8f793
3 changed files with 11 additions and 12 deletions
Download Patch File Download Diff File Expand all files Collapse all files
rocclr/device/rocm/rocblit.cpp
+7 -8
View File
@@ -1966,18 +1966,17 @@ bool KernelBlitManager::fillBuffer(device::Memory& memory, const void* pattern,
FillBufferInfo::PackInfo(memory, size[0], origin[0], pattern, patternSize, packed_vector);
size_t overall_offset = origin[0];
uint fillType = FillBufferAligned;
for (auto& packed_obj: packed_vector) {
uint fillType = FillBufferAligned;
size_t globalWorkOffset[3] = {0, 0, 0};
size_t globalWorkSize = amd::alignUp(packed_obj.fill_size_, 256);
size_t localWorkSize = 256;
uint32_t kpattern_size32 = (packed_obj.pattern_aligned_) ? sizeof(size_t) : patternSize;
uint32_t kpattern_size32 = (packed_obj.pattern_expanded_) ? sizeof(size_t) : patternSize;
size_t kfill_size = packed_obj.fill_size_/kpattern_size32;
size_t koffset = overall_offset;
overall_offset += packed_obj.fill_size_;
size_t globalWorkOffset[3] = {0, 0, 0};
size_t globalWorkSize = amd::alignUp(kfill_size, 256);
size_t localWorkSize = 256;
uint32_t alignment = (kpattern_size32 & 0x7) == 0 ?
sizeof(uint64_t) :
(kpattern_size32 & 0x3) == 0 ?
@@ -2019,7 +2018,7 @@ bool KernelBlitManager::fillBuffer(device::Memory& memory, const void* pattern,
auto constBuf = reinterpret_cast<address>(constantBuffer_->getHostMem()) + constBufOffset;
// If pattern has been expanded, use the expanded pattern, otherwise use the default pattern.
if (packed_obj.pattern_aligned_) {
if (packed_obj.pattern_expanded_) {
memcpy(constBuf, &packed_obj.expanded_pattern_, kpattern_size32);
} else {
memcpy(constBuf, pattern, kpattern_size32);
@@ -2028,8 +2027,8 @@ bool KernelBlitManager::fillBuffer(device::Memory& memory, const void* pattern,
mem = as_cl<amd::Memory>(gpuCB->owner());
setArgument(kernels_[fillType], 4, sizeof(cl_mem), &mem, constBufOffset);
kpattern_size32 /= alignment;
koffset /= alignment;
kpattern_size32 /= alignment;
setArgument(kernels_[fillType], 5, sizeof(uint32_t), &kpattern_size32);
setArgument(kernels_[fillType], 6, sizeof(koffset), &koffset);