Add PowerPC support (POWER8 VSX + Power Mac G5 AltiVec) #404
Conversation
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
Port Microsoft BitNet to IBM POWER8 (ppc64le). Adds scalar fallback implementations for all 5 I2_S dot product kernel functions and the quantize_i2_s function. Also adds PowerPC defines to gemm-config.h. Benchmarks on POWER8 S824 (16c/128t, 512GB RAM, scalar-only): - BitNet Large 700M: 21.5 t/s pp128, 11.2 t/s tg32 - BitNet 2B: 8.0 t/s pp128, 4.1 t/s tg32 - Llama3-8B BitNet: 2.6 t/s pp128, 1.6 t/s tg32 Files changed: - gemm-config.h: Add PARALLEL_SIZE/ROW_BLOCK_SIZE for __VSX__/__powerpc64__ - ggml-bitnet-mad.cpp: Scalar fallbacks for quantize_i2_s, ggml_vec_dot_i2_i8_s_1x1, _1x4_32W, _1xN, _Nx1 - bitnet-lut-kernels.h: Stub header for POWER8 (LUT kernels are x86/ARM) Next: VSX-optimized kernels using vec_perm for 10-16x speedup. Co-Authored-By: Claude Opus 4.5 <noreply@anthropic.com>
Replace scalar fallback with POWER8 AltiVec/VSX optimized kernels for all 5 I2_S functions using vec_msum (vmsummbm) instruction. Key optimizations: - vec_msum: 16 signed*unsigned byte multiply-accumulate per cycle - dcbt prefetch hints for weight/activation data - i2s_vsx_half() helper processes 16-byte blocks in vectorized form - All 5 kernels: quantize, 1x1, 1x4_32W, 1xN, Nx1 Benchmark results (POWER8 S824, 64 threads): 700M: pp128 21.48 -> 211.48 t/s (9.8x) 2B: pp128 8.04 -> 73.03 t/s (9.1x) 8B: pp128 2.60 -> 27.39 t/s (10.5x) 8B: tg32 1.61 -> 4.90 t/s (3.0x) Co-Authored-By: Claude Opus 4.5 <noreply@anthropic.com>
Use dcbt with TH=0x10 (L3 resident hint) for weight block prefetch instead of transient dcbt. Keeps weight data sticky in L3 cache between token generation steps, avoiding re-fetch from DRAM. Key design: only prefetch next block (not bulk scan) to avoid overhead on prompt processing. Bulk prefetch hurts pp because BitNet I2_S blocks are tiny (32 bytes) and vec_msum is so fast that prefetch overhead dominates. Results (vs VSX-only baseline): 700M tg32: 22.77 -> 24.02 t/s (+5.5%) 2B tg32: 10.93 -> 11.99 t/s (+9.7%) 8B tg32: 4.90 -> 5.63 t/s (+14.9%) pp unchanged (within noise) Full speedup from scalar baseline: 8B pp128: 2.60 -> 26.98 t/s (10.4x) 8B tg32: 1.61 -> 5.63 t/s (3.5x) Co-Authored-By: Claude Opus 4.5 <noreply@anthropic.com>
Adds POWER8/PowerPC section with: - Build instructions for ppc64le - Three optimization levels explained (scalar, VSX, dcbt) - Full benchmark tables for 700M, 2B, and 8B models - Scalar-to-VSX speedup comparison (9-10x) - Key technical details (vec_msum, dcbt resident, NUMA) - Model sources and conversion notes Co-Authored-By: Claude Opus 4.5 <noreply@anthropic.com>
Add patches and build infrastructure for running BitNet on PowerPC 970 (Power Mac G5) big-endian systems. GGUF format is always little-endian on disk; this adds byte-swap support for all multi-byte scalar reads and tensor data. Key changes: - g5-big-endian.patch: gguf_fread_val() byte-swap function for GGUF reader, tensor data byte-swap for F32/F16/I2_S at load time, sizeof(bool)==4 fix for PowerPC GCC - regex-ppc.h: POSIX regex wrapper replacing broken std::regex on PPC big-endian - build_g5.sh: Build script with G5-safe compiler flags (-Os) Tested on Power Mac G5 Dual 2.0 GHz, Mac OS X 10.5, GCC 10.5.0. Produces coherent text at 4.31 t/s prompt eval, 1.61 t/s generation with bitnet_b1_58-large (728M). Co-Authored-By: Claude Opus 4.5 <noreply@anthropic.com>
Port POWER8 VSX SIMD kernels to G5 AltiVec using 4 compatibility macros that abstract the ISA differences. One code path, works on both targets. Compatibility macros: - I2S_VEC_LD_UC/I2S_VEC_LD_SC: vec_vsx_ld (POWER8) vs vec_ld (G5) - I2S_DCBT_*: TH-hint dcbt (POWER8) vs basic dcbt (G5) Key changes across all 4 kernel functions (1x1, 1x4_32W, 1xN, Nx1): - vec_vsx_ld → I2S_VEC_LD_UC / I2S_VEC_LD_SC (22 sites) - static const vector arrays → vec_splat_u8() macros (avoids Mach-O alignment issues on old Darwin, generates vspltisb instruction) - hsum_i32_4_vsx → hsum_i32_4_ppc with G5 branch using vec_sums - POWER8-specific dcbt TH hints → G5-safe basic dcbt fallback - Architecture guard extended with __powerpc__ and __ppc__ - Build script updated: -Os → -O3 (safe now that vector constants are in-register), added llama-bench target Scalar baseline on G5 Dual 2.0GHz: pp5 = 4.31 t/s, tg = 1.61 t/s Target with AltiVec: 12-20 t/s (3-5x speedup via vmsummbm) No endian changes needed: vec_msum accumulates all 16 bytes into 4 int32 lanes then hsum reduces all 4 - total is identical regardless of lane assignment order on BE vs LE. Co-Authored-By: Claude Opus 4.5 <noreply@anthropic.com>
- Change build_g5.sh from -O3 to -Os: higher optimization levels cause Bus errors on G5 due to Mach-O ABI stack alignment issues with aggressive vector register spills from GCC 10. - Add __attribute__((always_inline)) to i2s_ppc_half and hsum_i32_4_ppc: without this, the Mach-O ABI generates VRsave save/restore sequences (mfspr/mtspr, ~20 cycles each) on every function call, devastating performance in the inner dot product loop. - Recommend -t 1 for G5 inference: single thread is faster because ggml_barrier() overhead on 870 graph nodes per token exceeds the benefit of 2-thread parallelism. - Remove llama-bench from G5 build (C++ compat issues with GCC 10). G5 AltiVec kernel microbenchmark: 16.1x raw speedup (5.84 vs 0.36 GMAC/s). End-to-end limited by Amdahl's law: matmul is 12-24% of total inference. Co-Authored-By: Claude Opus 4.5 <noreply@anthropic.com>
…tize_row_i8_s Two changes applied via g5-altivec-framework.patch: 1. ggml.c: Add #elif defined(__ALTIVEC__) GGML_SIMD block after POWER9 - Vectorizes all ggml_vec_* functions (scale, dot, mad, add, mul) - Uses vec_ld/vec_st (aligned) instead of VSX vec_xl/vec_xst - Uses vec_madd(a,b,zero) instead of vec_mul (no vmulfp on G5) - F16 via scalar FP16<->FP32 conversion (same approach as WASM backend) 2. ggml-quants.c: Add AltiVec path in quantize_row_i8_s (168 calls/token) - Pass 1: vec_abs + vec_max for finding max absolute value - Pass 2: vec_madd + vec_round + vec_cts + vec_packs for int8 quantize - vec_sums result in element 3 (big-endian PPC) Build script updated: -O3 for C (safe), -Os for C++ (GCC 10.5 miscompile workaround), -lm for roundf(). Benchmark (bitnet_b1_58-large I2_S, G5 dual 2.0GHz, -t 1): AltiVec+framework: pp6=5.14 t/s, tg=1.51 t/s Scalar baseline: pp6=4.78 t/s, tg=1.84 t/s (+7.5% prompt processing, -18% generation due to -Os C++ constraint) Co-Authored-By: Claude Opus 4.5 <noreply@anthropic.com>
MK_CFLAGS/MK_CXXFLAGS on the make command line override Makefile internal += appends, so -fopenmp added by the GGML_NO_OPENMP= path was silently discarded. The code compiled with GGML_USE_OPENMP defined (via MK_CPPFLAGS, which was not overridden), causing it to take the OpenMP codepath — but without -fopenmp, the compiler treated #pragma omp parallel as a no-op. Only thread 0 ever ran. Fix: add -fopenmp explicitly to MK_CFLAGS and MK_CXXFLAGS. Benchmarks (Dual 2.0 GHz G5, BitNet 700M I2_S, -n 20): -t 1: 721 ms/token, MUL_MAT 3186 us/call -t 2: 498 ms/token, MUL_MAT 2189 us/call (1.45x speedup) Co-Authored-By: Claude Opus 4.5 <noreply@anthropic.com>
Replace the scalar scale correction loops in ggml_compute_forward_mul_mat (both GEMM and GEMV I2_S paths) with AltiVec vectorized equivalents. Algebraic refactoring eliminates per-element division: Original: (x - act_sums) / act_scales * scale (sub + div + mul) Optimized: x * (scale/act_scales) + (-(act_sums * scale/act_scales)) This maps to a single vec_madd (vmaddfp) processing 4 floats per cycle. Precomputing factor and neg_offset once per column replaces thousands of divisions with two scalar divides total. Wrapped in #if defined(__ALTIVEC__) || defined(__VSX__) with scalar fallback, so non-PPC builds are unaffected. Benchmarks (Dual 2.0 GHz G5, BitNet 700M I2_S, -t 2): Scale correction: 91,588 us → 18,395 us (5.0x faster) Per-token total: 498 ms → 490 ms (~1.6% end-to-end) Co-Authored-By: Claude Opus 4.5 <noreply@anthropic.com>
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
1 participant
Add this suggestion to a batch that can be applied as a single commit.
This suggestion is invalid because no changes were made to the code.
Suggestions cannot be applied while the pull request is closed.
Suggestions cannot be applied while viewing a subset of changes.
Only one suggestion per line can be applied in a batch.
Add this suggestion to a batch that can be applied as a single commit.
Applying suggestions on deleted lines is not supported.
You must change the existing code in this line in order to create a valid suggestion.
Outdated suggestions cannot be applied.
This suggestion has been applied or marked resolved.
Suggestions cannot be applied from pending reviews.
Suggestions cannot be applied on multi-line comments.
Suggestions cannot be applied while the pull request is queued to merge.
Suggestion cannot be applied right now. Please check back later.
Summary
#ifdefguardsWhat's included
Core kernel (
src/ggml-bitnet-mad.cpp)ggml_gemv_i2_i8_s/ggml_gemm_i2_i8_swith AltiVecvec_msum(16 multiply-accumulates per cycle)#elif defined(__VSX__) || defined(__ALTIVEC__)blocks — zero impact on x86/ARMBig-endian support (
patches/g5-big-endian.patch)patches/regex-ppc.h)Framework vectorization (
patches/g5-altivec-framework.patch)#elif defined(__ALTIVEC__)to ggml.c's GGML_SIMD macro chainggml_vec_*functions (scale, dot, add, mul, mad) via AltiVecquantize_row_i8_swithvec_abs/vec_packs/vec_ctsScale correction vectorization (
patches/g5-altivec-scale.patch)vec_madd(x-A)/B*C→x*(C/B)+(-(A*C/B))= single fused multiply-addBuild script (
patches/build_g5.sh)-Osfor C++,-O3for C)-fopenmpin compiler flagsBenchmarks
Power Mac G5 Dual 2.0 GHz (PowerPC 970, 6GB DDR400, Mac OS X 10.5):
IBM POWER8 S824 (16-core, 512GB RAM, ppc64le):
Files changed
src/ggml-bitnet-mad.cppinclude/bitnet-lut-kernels.hinclude/gemm-config.hpatches/build_g5.shpatches/g5-big-endian.patchpatches/g5-altivec-framework.patchpatches/g5-altivec-scale.patchpatches/regex-ppc.hREADME.mdTest plan
#ifdefguards, no changes to x86/ARM paths🤖 Generated with Claude Code