Skip to content

@FIR-2148 - Fix GGML Perf Summary Double Counting for Reused Graph Nodes #151

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

Merged
akapoor3518 merged 2 commits into from
Aug 5, 2026
Merged

@FIR-2148 - Fix GGML Perf Summary Double Counting for Reused Graph Nodes #151

Show file tree
Hide file tree
Changes from all commits
Commits
Show all changes
2 commits
Select commit Hold shift + click to select a range
90e7eb8
@FIR-2148 - Fix GGML Perf Summary Double Counting for Reused Graph Nodes
Aug 5, 2026
de3a229
Addressed Qodo PR review comments
Aug 5, 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
31 changes: 19 additions & 12 deletions ggml/src/ggml.c
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -7400,25 +7400,36 @@ void ggml_perf_accumulate(struct ggml_perf_totals totals[GGML_OP_COUNT], struct

if (op >= GGML_OP_COUNT) continue;

const int64_t node_perf_time_us = node->perf_time_us;
const int64_t node_perf_runs = node->perf_runs;
const int64_t node_tsi_kernel_runs = node->tsi_kernel_runs;

if (node_perf_runs == 0 && node_perf_time_us == 0 && node_tsi_kernel_runs == 0) {
continue;
}

totals[op].op_name = ggml_op_name(op);
totals[op].total_us += node->perf_time_us;
totals[op].runs += node->perf_runs;
totals[op].total_us += node_perf_time_us;
totals[op].runs += node_perf_runs;
totals[op].op_count++;

// Count backend runs
enum ggml_compute_backend_type be = node->ggml_compute_backend;
if (be >= GGML_COMPUTE_BACKEND_CPU && be < GGML_COMPUTE_BACKEND_COUNT) {
totals[op].backend_subtotals[be].total_us += node->perf_time_us;
totals[op].backend_subtotals[be].runs += node->perf_runs;
totals[op].backend_subtotals[be].tsi_kernel_count += node->tsi_kernel_runs;
totals[op].backend_subtotals[be].total_us += node_perf_time_us;
totals[op].backend_subtotals[be].runs += node_perf_runs;
totals[op].backend_subtotals[be].tsi_kernel_count += node_tsi_kernel_runs;
}

if (op == GGML_OP_UNARY) {
enum ggml_unary_op subop = ggml_get_unary_op(node);
totals[op].unary_subtotals[subop].total_us += node->perf_time_us;
totals[op].unary_subtotals[subop].runs += node->perf_runs;
totals[op].unary_subtotals[subop].tsi_kernel_count += node->tsi_kernel_runs;
totals[op].unary_subtotals[subop].total_us += node_perf_time_us;
totals[op].unary_subtotals[subop].runs += node_perf_runs;
totals[op].unary_subtotals[subop].tsi_kernel_count += node_tsi_kernel_runs;
}
node->perf_time_us = 0;
Comment thread
akapoor3518 marked this conversation as resolved.
Resolved
node->perf_runs = 0;
node->tsi_kernel_runs = 0;
}
}
#endif /* GML_PERF-related flags */
Expand All @@ -7438,7 +7449,6 @@ FILE * ggml_perf_log_open(const char *filename) {
return fp;
}


void ggml_perf_write_detailed_csv(struct ggml_cgraph * cgraph, FILE *fp) {
if (!fp || !cgraph) return;

Expand Down Expand Up @@ -7739,7 +7749,4 @@ void ggml_perf_write_detailed_csv(struct ggml_cgraph * cgraph, FILE *fp) {
free(aggs);
}




#endif /* GGML_PERF_DETAIL */
2 changes: 1 addition & 1 deletion src/llama-context.cpp
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -1102,12 +1102,12 @@ int llama_context::decode(const llama_batch & batch_inp) {
}
#elif defined(GGML_PERF_DETAIL)
if (res) {
ggml_perf_accumulate(perf_totals, res->get_gf());
if (!perf_all_shape_written_once && perf_all_shape_fp) {
ggml_perf_write_detailed_csv(res->get_gf(), perf_all_shape_fp);
fflush(perf_all_shape_fp);
perf_all_shape_written_once = true;
}
ggml_perf_accumulate(perf_totals, res->get_gf());
}
#endif /* GML_PERF-related flags */

Expand Down