Skip to content

Qwen3.8-Flash-Next on any consumer hardware: one-click install for Windows / Linux. Strata inference engine, OpenAI/Anthropic API on localhost, optional image input.

C++

[Engine] SYCL: native_expert_bench reports microseconds, not milliseconds

Niko12216124c35e39e6
diff --git a/sycl/src/kernels/native_expert_bench.cpp b/sycl/src/kernels/native_expert_bench.cpp
index c75726c91..a32d81160 100644
--- a/sycl/src/kernels/native_expert_bench.cpp
+++ b/sycl/src/kernels/native_expert_bench.cpp
@@ -171,7 +171,12 @@ int main(int argc, char** argv) {
sync("timing loop");
const double ms =
std::chrono::duration<double, std::milli>(std::chrono::steady_clock::now() - t0).count();
- return ms / iters;
+ // SYCL port: the printout below and the GB/s it derives both assume MICROseconds ("all %7.1f us",
+ // bytes / us / 1e3), but this returned milliseconds - so every published number was 1000x too small in
+ // time and 1000x too large in bandwidth (0.1 "us" and 148,713 GB/s for a launch that takes ~100 us and
+ // moves 19.7 MB at ~198 GB/s). The wall clock always agreed with the real work: 200 iterations 618 ms,
+ // 2,000 iterations 3,566 ms. Return microseconds.
+ return ms * 1e3 / iters;
};
const double gu_b = (double) G * 2 * f.up_off, d_b = (double) G * dsz;
double us[2][3] = {{1e30, 1e30, 1e30}, {1e30, 1e30, 1e30}};