Skip to content

DCPerf v2-beta : Rebatch benchmark memcpy may be optimized out with Clan #828

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
excelle08 wants to merge 1 commit into
base: v2-beta
Choose a base branch
from
Open

DCPerf v2-beta : Rebatch benchmark memcpy may be optimized out with Clan #828

Changes from all commits
Commits
Show all changes
1 commit
Select commit
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
12 changes: 7 additions & 5 deletions packages/ai_wdl/rebatch/rebatchBench.cpp
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -287,8 +287,8 @@ class TensorBatch {
* prevent cache effects
* 3. Multi-threading - Supports configurable number of worker threads
* 4. Prefetching - Optional memory prefetching for improved performance
* 5. Anti-optimization - Uses checksums to prevent compiler from optimizing
* away operations
* 5. Anti-optimization - Uses a compiler barrier to prevent the compiler from
* optimizing away the memcpy operations
*/
class RebatchBenchmark {
public:
Expand Down Expand Up @@ -360,9 +360,6 @@ class RebatchBenchmark {
* 5. Tracks performance statistics
*/
void workerThread(size_t thread_id) {
// Checksum to prevent compiler from optimizing away memcpy operations
volatile uint64_t checksum = 0;

while (!stop_flag_) {
// Allocate a fresh output tensor for this batch
char* output = new char[OUTPUT_TENSOR_SIZE];
Expand Down Expand Up @@ -438,6 +435,11 @@ class RebatchBenchmark {
}
}

// The output buffer is never read elsewhere, so force the compiler to
// treat it as observed; otherwise the memcpy operations above can be
// elided as dead stores.
asm volatile("" : : "r,m"(output) : "memory");

// Update statistics
bytes_processed_[thread_id] += totalProcessedBytes;
batches_processed_[thread_id]++;
Expand Down