SuperNPUBench

SuperNPUBench is a C++ tile-kernel benchmark repository for superscalar NPU compiler and model development. It contains direct compiler-validation cases, one-level kernel manifests, instruction microbenchmarks, and the standalone programming manual published from docs/.

Repository Structure

SuperNPUBench/
|-- benchmark/
|   |-- two-level-arch/   # promoted compiler and direct-boot cases
|   `-- one-level-arch/   # source-backed tile-kernel catalog
|-- microbenchmark/       # instruction-family microbenchmarks
|-- docs/                 # standalone MkDocs manual, generators, and checks
`-- compile_all.sh

Build output remains under architecture-local output/ directories and is not committed.

Current Integration Baseline

Use this repository through the linx-isa superproject. Its submodule pins select the matching v0.57 ISA state, LLVM 23 compiler sources, QEMU, and LinxCoreModel. Do not combine an arbitrary compiler binary with unrelated ISA or model revisions.

git clone --recurse-submodules https://github.com/LinxISA/linx-isa.git
cd linx-isa
git submodule sync --recursive
git submodule update --init --recursive
export LINX_ROOT="$PWD"

Configure and build the in-repository compiler:

cmake -S compiler/llvm/llvm -B compiler/llvm/build-linxisa-clang -G Ninja \
  -DCMAKE_BUILD_TYPE=Release \
  -DLLVM_ENABLE_PROJECTS="clang;lld" \
  -DLLVM_EXPERIMENTAL_TARGETS_TO_BUILD=LinxISA \
  -DLLVM_TARGETS_TO_BUILD=""

cmake --build compiler/llvm/build-linxisa-clang \
  --target clang lld llvm-objdump llvm-objcopy llvm-readobj -j 12

export COMPILER_DIR="$LINX_ROOT/compiler/llvm/build-linxisa-clang/bin"
"$COMPILER_DIR/clang" --version

The superproject pin is the source of truth for the exact compiler revision; rebuild after that pin changes.

Quick Start

Compile the promoted memory and arithmetic TileOP cases:

cd "$LINX_ROOT/workloads/SuperNPUBench/benchmark/two-level-arch/test/tileop_api"
make TESTCASE=TAdd PLAT=linx COMPILER_DIR="$COMPILER_DIR"
make TESTCASE=TLoad PLAT=linx COMPILER_DIR="$COMPILER_DIR"
make TESTCASE=TStore PLAT=linx COMPILER_DIR="$COMPILER_DIR"

Run one case through the canonical compiler, QEMU, and cycle-model sequence:

cd "$LINX_ROOT"
export QEMU="$(tools/bringup/run_qemu_build_clean.sh)"
python3 tools/bringup/run_ai_workload_flow.py \
  --profile smoke \
  --kind supernpu \
  --case '=supernpu-tileop_api-TAdd' \
  --run-id supernpu-tadd

The runner stops on the first failed stage. A compiler-passing ELF executes in QEMU before the same artifact is promoted to plain gfsim -f <elf>.

Benchmark catalog

The active one-level manifests contain 87 build variants. Every name below has a source-backed page with its build command and tile intrinsic surface in the website's Benchmark Reference section. Catalog presence records source inventory; it does not imply promotion on the current compiler and model flow.

broadcast (4 names, 6 variants)

broadcast, broadcast_vec_019, broadcast_vec_039, broadcast_vec_07

concat (2 names, 4 variants)

concat_gather, concat_scatter

control (1 name, 6 variants)

hashtable_lookup_simd

deepseek (23 names, 23 variants)

aux_fi, batched_transpose, cast_back_per_channel, cast_back_per_token, engram_hash_layer, expand_to_fused, expand_to_mhc_bwd, expand_to_mhc_fwd, fn_normw_merge_fwd, fused_weight, get_fused_mapping, group_count, inplace_unique_group_indices, mask_indices_by_tp, multilayer_recompute, normalize_weight, per_channel_cast, per_token_cast, reduce_fused, rms_norm, sinkhorn_fwd, swiglu_forward_and_per_token_cast, topk_gate

element_wise/gelu (1 name, 1 variant)

gelu

fa (4 names, 13 variants)

fa_2d_unroll, fa_HIF4_HIF4, fa_softmax_pto, sfa

flashMLA (1 name, 2 variants)

flashMLA

gather (1 name, 1 variant)

gather

matmul (1 name, 16 variants)

matmul

multi_thread/fa (1 name, 1 variant)

fa_2d_unroll_gmma

multi_thread/matmul (2 names, 2 variants)

matmul, matmul_partial

multi_thread/vec (2 names, 2 variants)

tadd, trowsum

reduction/reducemax_col (1 name, 1 variant)

reducemax_col

reduction/reducemax_row (1 name, 1 variant)

reducemax_row

reduction/reducesum_col (1 name, 2 variants)

reducesum_col

reduction/reducesum_row (1 name, 1 variant)

reducesum_row

sort (1 name, 1 variant)

topk

transpose (1 name, 4 variants)

transpose

The public catalog covers active one-level compile.all manifests. Those commands are source inventory, not current LLVM 23 promotion evidence: the one-level make lane still contains a retired compiler option and an unresolved include-root dependency. Use make -n to inspect a catalog command and use the canonical runner for current compile/model validation.

Microbenchmarks

microbenchmark/ organizes cases by matrix, vector, memory, and scalar instruction families. See the microbenchmark README for its generated case inventory and lane-specific commands.

Documentation

The website source is self-contained under docs/. Build and verify it with:

python3 -m pip install -r docs/requirements.txt
python3 docs/scripts/generate_benchmark_manual.py
python3 docs/scripts/generate_deepseek_manifest.py
python3 docs/scripts/generate_deepseek_manual.py
python3 docs/scripts/sync_golden_manual.py --check --linx-root "$LINX_ROOT"
docs/build.sh

Start with:

• toolchain setup
• build, run, and disassembly workflow
• C++ tile kernel programming guide
• intrinsic reference
• benchmark reference

Adding or Updating a Benchmark

1. Add the kernel and its test entrypoint under the appropriate architecture.
2. Add every supported variant to the local compile.all manifest.
3. Regenerate the benchmark and DeepSeek pages when their source inventories change.
4. Compile a bounded promoted case with the pinned compiler.
5. Run docs/build.sh before publication.