Skip to content

TDM Reduce Copy bug fix #349

New issue

Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.

Sign up for GitHub

By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.

Already on GitHub? Sign in to your account

Open
AtlantaPepsi wants to merge 7 commits into
base: candidate-1.70
Choose a base branch
from
+497 −56
Open

TDM Reduce Copy bug fix #349

Show file tree
Hide file tree
Changes from all commits
Commits
Show all changes
7 commits
Select commit Hold shift + click to select a range
c9922ed
TDM reduceCopy scaffolding
AtlantaPepsi Jul 29, 2026
4734f9a
small addition for TdmSweep config
AtlantaPepsi Jul 29, 2026
48bde52
scratch reduceCopy
AtlantaPepsi Jul 29, 2026
e684860
fix reduce offset error; datatype template; support no src/dst
AtlantaPepsi Aug 6, 2026
d6e346d
temporary fix for LDS access in reduction loop
AtlantaPepsi Aug 10, 2026
ecb3b52
fix0.5
AtlantaPepsi Aug 12, 2026
eb2af9d
fix
AtlantaPepsi Aug 12, 2026
File filter

Filter by extension

Filter by extension
Conversations
Failed to load comments.
Loading
Jump to
Jump to file
Failed to load files.
Loading
Diff view
Diff view
8 changes: 8 additions & 0 deletions src/client/EnvVars.hpp
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -129,6 +129,7 @@ class EnvVars
// TDM options
int tdmBlockOrder; // How threadblocks for multiple Transfers are ordered 0=sequential 1=interleaved
int tdmBlockSize; // Size of each threadblock for TDM kernels (must be multiple of 32)
int tdmKernel; // TDM Kernel to use (-1=auto, 0=copy-only, 1=reduce)
int tdmLdsBytes; // Size of LDS (shared memory) bytes per threadblock for TDM kernels (0 = use device max)

// Developer features
Expand Down Expand Up @@ -184,6 +185,7 @@ class EnvVars
sweepMinPow2 = GetEnvVar("SWEEP_MIN_POW2" , 10);
tdmBlockOrder = GetEnvVar("TDM_BLOCK_ORDER" , 0);
tdmBlockSize = GetEnvVar("TDM_BLOCK_SIZE" , 256);
tdmKernel = GetEnvVar("TDM_KERNEL" , 0);
tdmLdsBytes = GetEnvVar("TDM_LDS_BYTES" , 0);
useHipEvents = GetEnvVar("USE_HIP_EVENTS" , 1);
useHsaDma = GetEnvVar("USE_HSA_DMA" , 0);
Expand Down Expand Up @@ -406,6 +408,7 @@ class EnvVars
printf(" SWEEP_MIN_POW2 - When 0 is specified for data size, this is the starting power of two exponent\n");
printf(" TDM_BLOCK_ORDER - How blocks for TDM transfers are ordered. 0=sequential, 1=interleaved\n");
printf(" TDM_BLOCK_SIZE - # of threads per threadblock for TDM (async tensor) kernels (Must be multiple of 32)\n");
printf(" TDM_KERNEL - -1=auto, 0=force TdmCopyKernel, 1=force TdmReduceKernel (may error if ineligible)\n");
printf(" TDM_LDS_BYTES - Amount of LDS bytes to allocate per workgroup for TDM kernels (0 = device max; K/M/G suffixes accepted)\n");
printf(" USE_HIP_EVENTS - Use HIP events for GFX executor timing\n");
printf(" USE_HIP_EVENTS - Use HIP events for GFX/DMA/TDM executor timing (0=CPU wall-clock)\n");
Expand Down Expand Up @@ -554,6 +557,10 @@ class EnvVars
Print("TDM_BLOCK_ORDER", tdmBlockOrder,
"TDM Thread block ordering: %s", tdmBlockOrder == 0 ? "Sequential" : "Interleaved");
Print("TDM_BLOCK_SIZE", tdmBlockSize, "TDM threadblock size of %d", tdmBlockSize);
Print("TDM_KERNEL", tdmKernel,
"%s", tdmKernel == -1 ? "auto" :
tdmKernel == 0 ? "force TdmCopyKernel" :
tdmKernel == 1 ? "force TdmReduceKernel" : "unknown");
Print("TDM_LDS_BYTES", tdmLdsBytes, "%s",
tdmLdsBytes == 0 ? "Using device max LDS bytes per workgroup"
: (std::string("Setting LDS to ") + std::to_string(tdmLdsBytes) + " bytes per workgroup").c_str());
Expand Down Expand Up @@ -766,6 +773,7 @@ class EnvVars

cfg.tdm.blockOrder = tdmBlockOrder;
cfg.tdm.blockSize = tdmBlockSize;
cfg.tdm.tdmKernel = tdmKernel;
cfg.tdm.ldsBytes = tdmLdsBytes;

return cfg;
Expand Down
97 changes: 55 additions & 42 deletions src/client/Presets/TdmSweep.hpp
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -57,6 +57,7 @@ int TdmSweepPreset(EnvVars& ev,
// Collect environment variables for this preset
vector<int> blockList = EnvVars::GetEnvVarArray("BLOCKSIZES", {64,128,256,512,1024});
vector<int> blockOrders = EnvVars::GetEnvVarArray("BLOCK_ORDERS", {0});
vector<int> byteOffsets = EnvVars::GetEnvVarArray("BYTE_OFFSETS", {0});
vector<int> ldsList = EnvVars::GetEnvVarArray("LDS_BYTES", {8192,16384,32768,65536,0});
vector<int> numSesList = EnvVars::GetEnvVarArray("NUM_SUB_EXECS", {2,4,8,16,32,64});
int numTransfers = EnvVars::GetEnvVar( "NUM_TRANSFERS", 1);
Expand All @@ -71,6 +72,7 @@ int TdmSweepPreset(EnvVars& ev,
Utils::Print("[TDM Sweep Related]\n");
ev.Print("BLOCKSIZES", blockList.size(), EnvVars::ToStr(blockList).c_str());
ev.Print("BLOCK_ORDERS", blockOrders.size(), "%s (0=sequential 1=interleaved 2=random)", EnvVars::ToStr(blockOrders).c_str());
ev.Print("BYTE_OFFSETS", byteOffsets.size(), "%s (src/dst alloc offset; drives the TDM vector head/tail path)", EnvVars::ToStr(byteOffsets).c_str());
ev.Print("LDS_BYTES", ldsList.size(), "%s (0 = device max LDS per block)", EnvVars::ToStr(ldsList).c_str());
ev.Print("NUM_SUB_EXECS", numSesList.size(), EnvVars::ToStr(numSesList).c_str());
ev.Print("NUM_TRANSFERS", numTransfers, "Number of Transfers specified in TDM_TRANSFER");
Expand Down Expand Up @@ -109,6 +111,13 @@ int TdmSweepPreset(EnvVars& ev,
return ERR_FATAL;
}
}
for (int off : byteOffsets) {
if (off < 0 || off % (int)sizeof(float)) {
Utils::Print("[ERROR] BYTE_OFFSETS value %d is invalid (must be a non-negative multiple of %lu)\n",
off, sizeof(float));
return ERR_FATAL;
}
}

std::vector<Transfer> transfers;
Utils::CheckForError(ParseTransfers(std::to_string(numTransfers) + " 1 " + transferStr, transfers));
Expand Down Expand Up @@ -173,68 +182,71 @@ int TdmSweepPreset(EnvVars& ev,

// Print header
char sep = ev.outputToCsv ? ',' : ' ';
Utils::Print(" BlkO %c BlkS %c LDSBytes ", sep, sep);
Utils::Print(" BlkO %c BlkS %c LDSBytes %c ByteOff ", sep, sep, sep);
for (int numSubExec : numSesList)
Utils::Print("%c SE %03d", sep, numSubExec);
Utils::Print("\n");

int bestSe = -1;
double overallBestBw = 0;
vector<double> bestBw(numSesList.size(), 0.0);
// best[s] = {blockOrder, blockSize, ldsBytes, numSubExec}
vector<vector<int>> best(numSesList.size(), vector<int>(4));
// best[s] = {blockOrder, blockSize, ldsBytes, byteOffset, numSubExec}
vector<vector<int>> best(numSesList.size(), vector<int>(5));

// Loop over all combinations
for (int blockOrder : blockOrders) { cfg.tdm.blockOrder = blockOrder;
for (int blockSize : blockList) { cfg.tdm.blockSize = blockSize;
for (int ldsBytes : ldsList) { cfg.tdm.ldsBytes = ldsBytes;
Utils::Print(" %1d %c %4d %c %8d ", blockOrder, sep, blockSize, sep, ldsBytes);
fflush(stdout);
for (auto s = 0; s < numSesList.size(); s++) {
int numSubExec = numSesList[s];
for (Transfer& t : transfers) t.numSubExecs = numSubExec;
for (int byteOffset : byteOffsets) { cfg.data.byteOffset = byteOffset;
Utils::Print(" %1d %c %4d %c %8d %c %7d ",
blockOrder, sep, blockSize, sep, ldsBytes, sep, byteOffset);
fflush(stdout);
for (auto s = 0; s < numSesList.size(); s++) {
int numSubExec = numSesList[s];
for (Transfer& t : transfers) t.numSubExecs = numSubExec;

TestResults result;
// A given combination may be rejected by the library (e.g. LDS window
// larger than the device max). Treat that as a skipped cell (N/A) and
// keep sweeping instead of aborting the whole matrix.
if (RunTransfers(cfg, transfers, result)) {
double bw = 0.0;
switch (timingMode) {
case 0: bw = result.avgTotalBandwidthGbPerSec; break;
case 1:
for (auto const& e : result.exeResults) {
bw = std::max(bw, e.second.avgBandwidthGbPerSec);
}
break;
case 2: default:
for (auto const& t : result.tfrResults) {
bw = std::max(bw, t.avgBandwidthGbPerSec);
TestResults result;
// A given combination may be rejected by the library (e.g. LDS window
// larger than the device max). Treat that as a skipped cell (N/A) and
// keep sweeping instead of aborting the whole matrix.
if (RunTransfers(cfg, transfers, result)) {
double bw = 0.0;
switch (timingMode) {
case 0: bw = result.avgTotalBandwidthGbPerSec; break;
case 1:
for (auto const& e : result.exeResults) {
bw = std::max(bw, e.second.avgBandwidthGbPerSec);
}
break;
case 2: default:
for (auto const& t : result.tfrResults) {
bw = std::max(bw, t.avgBandwidthGbPerSec);
}
break;
}
break;
}

if (bw > bestBw[s]) {
bestBw[s] = bw;
best[s] = {blockOrder, blockSize, ldsBytes, numSubExec};
if (bw > overallBestBw) {
overallBestBw = bw;
bestSe = s;
if (bw > bestBw[s]) {
bestBw[s] = bw;
best[s] = {blockOrder, blockSize, ldsBytes, byteOffset, numSubExec};
if (bw > overallBestBw) {
overallBestBw = bw;
bestSe = s;
}
}
Utils::Print("%c%8.2f", sep, bw);
} else {
Utils::Print("%c%8s", sep, "N/A");
}
Utils::Print("%c%8.2f", sep, bw);
} else {
Utils::Print("%c%8s", sep, "N/A");
fflush(stdout);
}
Utils::Print("\n");
fflush(stdout);
}
Utils::Print("\n");
fflush(stdout);
}
}
}

Utils::Print(" BlkO %c BlkS %c LDSBytes ", sep, sep);
Utils::Print(" BlkO %c BlkS %c LDSBytes %c ByteOff ", sep, sep, sep);
for (auto s = 0; s < numSesList.size(); s++) {
Utils::Print("%c%8.2f", sep, bestBw[s]);
}
Expand All @@ -254,10 +266,11 @@ int TdmSweepPreset(EnvVars& ev,
Utils::Print(" BlockOrder : %7d [TDM_BLOCK_ORDER=%d]\n", best[bestSe][0], best[bestSe][0]);
Utils::Print(" BlockSize : %7d [TDM_BLOCK_SIZE=%d]\n", best[bestSe][1], best[bestSe][1]);
Utils::Print(" LDS Bytes : %7d [TDM_LDS_BYTES=%d]\n", best[bestSe][2], best[bestSe][2]);
Utils::Print(" NumSubExec : %7d\n", best[bestSe][3]);
Utils::Print(" Byte Offset : %7d [BYTE_OFFSET=%d]\n", best[bestSe][3], best[bestSe][3]);
Utils::Print(" NumSubExec : %7d\n", best[bestSe][4]);
Utils::Print("Command to run best result:\n");
Utils::Print("TDM_BLOCK_ORDER=%d TDM_BLOCK_SIZE=%d TDM_LDS_BYTES=%d ./TransferBench cmdline %lu \"%d %d %s\"\n",
best[bestSe][0], best[bestSe][1], best[bestSe][2],
numBytesPerTransfer, numTransfers, best[bestSe][3], transferStr.c_str());
Utils::Print("TDM_BLOCK_ORDER=%d TDM_BLOCK_SIZE=%d TDM_LDS_BYTES=%d BYTE_OFFSET=%d ./TransferBench cmdline %lu \"%d %d %s\"\n",
best[bestSe][0], best[bestSe][1], best[bestSe][2], best[bestSe][3],
numBytesPerTransfer, numTransfers, best[bestSe][4], transferStr.c_str());
return ERR_NONE;
}
Loading