perf_matrix

perf_matrix

Phase 14 SIMD gate benchmark. Builds the same int8 MobileNetV2-shaped block (QConv2D 3x3 + QDense) under several TINYMIND_ENABLE_SIMD_* gate combinations and emits a single CSV per binary:

active_backend,conv_iters,conv_total_us,conv_us_per_call,
dense_iters,dense_total_us,dense_us_per_call,
conv_output_checksum,dense_output_checksum

The *_output_checksum columns are the same across backends when Phase 14's bit-exactness invariant holds — any disagreement is a backend bug.

Usage

make            # builds scalar / avx2 / avx512f / avx512_vnni
make report     # runs every binary and prints the combined CSV
make clean

make report writes output/perf_report.csv with one row per backend.

Cross-arch backends

The default Makefile only builds the x86 backends that the host compiler can target. To bench an Arm gate, add a target with the matching cross-compiler, e.g.:

neon_dotprod:
	$(MKDIR)
	aarch64-linux-gnu-g++ $(OPT) $(WARN) -march=armv8.2-a+dotprod \
	  -o ./output/perf_matrix_neon_dotprod $(SOURCES) $(INCLUDES) \
	  $(DEFINES) -DTINYMIND_ENABLE_SIMD_NEON=1 \
	  -DTINYMIND_ENABLE_SIMD_NEON_DOTPROD=1

Run the resulting binary on the target hardware (or under qemu-aarch64 for correctness checks).

Notes on the checksum

output_checksum mixes a running sum with an FNV-style XOR-and-shift hash of every output byte. It is order-sensitive in the hash term and value-sensitive in both terms, so a one-byte off-by-one in any backend flips the value visibly.

The reference (scalar row) is the source of truth.