AI/rocm-systems
2
0
Förgrening 0
Kod Ärenden Pull-förfrågningar Actions Packages Projekt Släpp Wiki Aktiviteter

SWDEV-373173 - Correct max VGPRs and VGPR Granularity in Occupancy calculation

Bläddra i källkod 
Change-Id: I36010f3991f809c61f0c4388b8daa8d3e9b8cc9d


[ROCm/clr commit: 27c5189bca]
This commit is contained in:
Anusha GodavarthySurya
2023-01-17 04:18:24 +00:00
committad av Anusha Godavarthy Surya
förälder a2a134968c
incheckning a7a887ab05
1 ändrade filer med 8 tillägg och 63 borttagningar
Download Patch File Download Diff File Expand all files Collapse all files
projects/clr/hipamd/src/hip_platform.cpp
+8 -63
Visa fil
@@ -333,62 +333,14 @@ hipError_t ihipOccupancyMaxActiveBlocksPerMultiprocessor(
MaxWavesPerSimd = 16;
}
size_t VgprWaves = MaxWavesPerSimd;
uint32_t VgprGranularity = 0;
size_t maxVGPRs = 0;
uint32_t VgprGranularity = device.info().vgprAllocGranularity_;
size_t maxVGPRs = device.info().vgprsPerSimd_;
size_t wavefrontSize = wrkGrpInfo->wavefrontSize_;
switch (device.isa().versionMajor()) {
case (11):
if (device.isa().versionMinor() == 0) {
switch (device.isa().versionStepping()) {
case (0):
case (1):
VgprGranularity = (wavefrontSize == 32) ? 24 : 12;
maxVGPRs = (wavefrontSize == 32) ? 1536 : 768;
break;
case (2):
case (3):
VgprGranularity = (wavefrontSize == 32) ? 16 : 8;
maxVGPRs = (wavefrontSize == 32) ? 1024 : 512;
break;
default:
VgprGranularity = 8;
maxVGPRs = 1024;
break;
}
}
break;
case (10):
switch (device.isa().versionMinor()) {
case (0):
case (1):
VgprGranularity = (wavefrontSize == 32) ? 8 : 4;
maxVGPRs = (wavefrontSize == 32) ? 1024 : 512;
break;
case (3):
VgprGranularity = (wavefrontSize == 32) ? 16 : 8;
maxVGPRs = (wavefrontSize == 32) ? 1024 : 512;
break;
default:
VgprGranularity = 8;
maxVGPRs = 1024;
break;
}
break;
case (9):
if ((device.isa().versionMinor() == 0 && device.isa().versionStepping() == 10) ||
(device.isa().versionMinor() == 4 && device.isa().versionStepping() == 0)) {
VgprGranularity = 8;
maxVGPRs = 512;
} else {
VgprGranularity = 4;
maxVGPRs = 256;
}
break;
default:
// For gfx<=8
VgprGranularity = 4;
maxVGPRs = 256;
break;
if (device.isa().versionMajor() >= 10) {
if (wavefrontSize == 64) {
maxVGPRs = maxVGPRs >> 1;
VgprGranularity = VgprGranularity >> 1;
}
}
if (wrkGrpInfo->usedSGPRs_ > 0) {
VgprWaves = maxVGPRs / amd::alignUp(wrkGrpInfo->usedVGPRs_, VgprGranularity);
@@ -396,14 +348,7 @@ hipError_t ihipOccupancyMaxActiveBlocksPerMultiprocessor(
size_t GprWaves = VgprWaves;
if (wrkGrpInfo->usedSGPRs_ > 0) {
size_t maxSGPRs;
if (device.isa().versionMajor() < 8) {
maxSGPRs = 512;
} else if (device.isa().versionMajor() < 10) {
maxSGPRs = 800;
} else {
maxSGPRs = SIZE_MAX; // gfx10+ does not share SGPRs between waves
}
size_t maxSGPRs = device.info().sgprsPerSimd_;
const size_t SgprWaves = maxSGPRs / amd::alignUp(wrkGrpInfo->usedSGPRs_, 16);
GprWaves = std::min(VgprWaves, SgprWaves);
}