From 0b6f6becb4e916a24fcaf2966647381a21d1f084 Mon Sep 17 00:00:00 2001 From: Francis Couture-Harpin Date: Wed, 11 Jun 2025 15:29:58 -0400 Subject: [PATCH] ggml-cpu : reorder SVE FMA for consistency with other SIMD arches --- ggml/src/ggml-cpu/ops.cpp | 2 +- ggml/src/ggml-cpu/simd-mappings.h | 2 +- ggml/src/ggml-cpu/vec.cpp | 18 +++++++++--------- ggml/src/ggml-cpu/vec.h | 18 +++++++++--------- 4 files changed, 20 insertions(+), 20 deletions(-) diff --git a/ggml/src/ggml-cpu/ops.cpp b/ggml/src/ggml-cpu/ops.cpp index 11d4819c86..711d8abcc5 100644 --- a/ggml/src/ggml-cpu/ops.cpp +++ b/ggml/src/ggml-cpu/ops.cpp @@ -7771,7 +7771,7 @@ static void ggml_compute_forward_ssm_scan_f32( t1 = exp_ps_sve(svptrue_b32(), t1); svfloat32_t t2 = GGML_F32_VEC_MUL(vx_dt, vB); - vs0 = GGML_F32_VEC_FMA(vs0, t1, t2); + vs0 = GGML_F32_VEC_FMA(t2, vs0, t1); r1_vector = GGML_F32_VEC_ADD(GGML_F32_VEC_MUL(vs0, vC), r1_vector); GGML_F32_VEC_STORE(&s[ii*nc + k], vs0); diff --git a/ggml/src/ggml-cpu/simd-mappings.h b/ggml/src/ggml-cpu/simd-mappings.h index 2e3669c018..91bb867bf5 100644 --- a/ggml/src/ggml-cpu/simd-mappings.h +++ b/ggml/src/ggml-cpu/simd-mappings.h @@ -32,7 +32,7 @@ #define GGML_F32xt_LOAD(...) GGML_F32xt_LOAD_IMPL(DEFAULT_PG, __VA_ARGS__) #define GGML_F32xt_STORE_IMPL(pg,a,b) svst1_f32(pg, a, b) #define GGML_F32xt_STORE(...) GGML_F32xt_STORE_IMPL(DEFAULT_PG, __VA_ARGS__) -#define GGML_F32xt_FMA_IMPL(pg, a, b, c) svmad_f32_m(pg, a, b, c) +#define GGML_F32xt_FMA_IMPL(pg, a, b, c) svmad_f32_m(pg, b, c, a) #define GGML_F32xt_FMA(...) GGML_F32xt_FMA_IMPL(DEFAULT_PG, __VA_ARGS__) #define GGML_F32xt_ADD_IMPL(pg, a, b) svadd_f32_m(pg, a, b) #define GGML_F32xt_ADD(...) GGML_F32xt_ADD_IMPL(DEFAULT_PG, __VA_ARGS__) diff --git a/ggml/src/ggml-cpu/vec.cpp b/ggml/src/ggml-cpu/vec.cpp index f7614568ea..7e61b5bf96 100644 --- a/ggml/src/ggml-cpu/vec.cpp +++ b/ggml/src/ggml-cpu/vec.cpp @@ -37,35 +37,35 @@ void ggml_vec_dot_f32(int n, float * GGML_RESTRICT s, size_t bs, const float * G for (int i = 0; i < np; i += ggml_f32_step) { ax1 = GGML_F32_VEC_LOAD(x + i); ay1 = GGML_F32_VEC_LOAD(y + i); - sum1 = GGML_F32_VEC_FMA(ax1, ay1, sum1); + sum1 = GGML_F32_VEC_FMA(sum1, ax1, ay1); ax2 = GGML_F32_VEC_LOAD(x + i + 1*ggml_f32_epr); ay2 = GGML_F32_VEC_LOAD(y + i + 1*ggml_f32_epr); - sum2 = GGML_F32_VEC_FMA(ax2, ay2, sum2); + sum2 = GGML_F32_VEC_FMA(sum2, ax2, ay2); ax3 = GGML_F32_VEC_LOAD(x + i + 2*ggml_f32_epr); ay3 = GGML_F32_VEC_LOAD(y + i + 2*ggml_f32_epr); - sum3 = GGML_F32_VEC_FMA(ax3, ay3, sum3); + sum3 = GGML_F32_VEC_FMA(sum3, ax3, ay3); ax4 = GGML_F32_VEC_LOAD(x + i + 3*ggml_f32_epr); ay4 = GGML_F32_VEC_LOAD(y + i + 3*ggml_f32_epr); - sum4 = GGML_F32_VEC_FMA(ax4, ay4, sum4); + sum4 = GGML_F32_VEC_FMA(sum4, ax4, ay4); ax5 = GGML_F32_VEC_LOAD(x + i + 4*ggml_f32_epr); ay5 = GGML_F32_VEC_LOAD(y + i + 4*ggml_f32_epr); - sum5 = GGML_F32_VEC_FMA(ax5, ay5, sum5); + sum5 = GGML_F32_VEC_FMA(sum5, ax5, ay5); ax6 = GGML_F32_VEC_LOAD(x + i + 5*ggml_f32_epr); ay6 = GGML_F32_VEC_LOAD(y + i + 5*ggml_f32_epr); - sum6 = GGML_F32_VEC_FMA(ax6, ay6, sum6); + sum6 = GGML_F32_VEC_FMA(sum6, ax6, ay6); ax7 = GGML_F32_VEC_LOAD(x + i + 6*ggml_f32_epr); ay7 = GGML_F32_VEC_LOAD(y + i + 6*ggml_f32_epr); - sum7 = GGML_F32_VEC_FMA(ax7, ay7, sum7); + sum7 = GGML_F32_VEC_FMA(sum7, ax7, ay7); ax8 = GGML_F32_VEC_LOAD(x + i + 7*ggml_f32_epr); ay8 = GGML_F32_VEC_LOAD(y + i + 7*ggml_f32_epr); - sum8 = GGML_F32_VEC_FMA(ax8, ay8, sum8); + sum8 = GGML_F32_VEC_FMA(sum8, ax8, ay8); } // leftovers // Since 8 unrolls are done in above loop, leftovers lie in range [0, ggml_f32_step] which is handled in below loop @@ -73,7 +73,7 @@ void ggml_vec_dot_f32(int n, float * GGML_RESTRICT s, size_t bs, const float * G for (int i = np; i < np2; i += ggml_f32_epr) { ax1 = GGML_F32_VEC_LOAD(x + i); ay1 = GGML_F32_VEC_LOAD(y + i); - sum1 = GGML_F32_VEC_FMA(ax1, ay1, sum1); + sum1 = GGML_F32_VEC_FMA(sum1, ax1, ay1); } // maximum number of leftover elements will be less that ggml_f32_epr. Apply predicated svmad on available elements only if (np2 < n) { diff --git a/ggml/src/ggml-cpu/vec.h b/ggml/src/ggml-cpu/vec.h index 09dbade217..a144259800 100644 --- a/ggml/src/ggml-cpu/vec.h +++ b/ggml/src/ggml-cpu/vec.h @@ -163,49 +163,49 @@ inline static void ggml_vec_mad_f32(const int n, float * GGML_RESTRICT y, const ax1 = GGML_F32_VEC_LOAD(x + i); ay1 = GGML_F32_VEC_LOAD(y + i); - ay1 = GGML_F32_VEC_FMA(ax1, vx, ay1); + ay1 = GGML_F32_VEC_FMA(ay1, ax1, vx); GGML_F32_VEC_STORE(y + i, ay1); ax2 = GGML_F32_VEC_LOAD(x + i + 1*ggml_f32_epr); ay2 = GGML_F32_VEC_LOAD(y + i + 1*ggml_f32_epr); - ay2 = GGML_F32_VEC_FMA(ax2, vx, ay2); + ay2 = GGML_F32_VEC_FMA(ay2, ax2, vx); GGML_F32_VEC_STORE(y + i + 1*ggml_f32_epr, ay2); ax3 = GGML_F32_VEC_LOAD(x + i + 2*ggml_f32_epr); ay3 = GGML_F32_VEC_LOAD(y + i + 2*ggml_f32_epr); - ay3 = GGML_F32_VEC_FMA(ax3, vx, ay3); + ay3 = GGML_F32_VEC_FMA(ay3, ax3, vx); GGML_F32_VEC_STORE(y + i + 2*ggml_f32_epr, ay3); ax4 = GGML_F32_VEC_LOAD(x + i + 3*ggml_f32_epr); ay4 = GGML_F32_VEC_LOAD(y + i + 3*ggml_f32_epr); - ay4 = GGML_F32_VEC_FMA(ax4, vx, ay4); + ay4 = GGML_F32_VEC_FMA(ay4, ax4, vx); GGML_F32_VEC_STORE(y + i + 3*ggml_f32_epr, ay4); ax5 = GGML_F32_VEC_LOAD(x + i + 4*ggml_f32_epr); ay5 = GGML_F32_VEC_LOAD(y + i + 4*ggml_f32_epr); - ay5 = GGML_F32_VEC_FMA(ax5, vx, ay5); + ay5 = GGML_F32_VEC_FMA(ay5, ax5, vx); GGML_F32_VEC_STORE(y + i + 4*ggml_f32_epr, ay5); ax6 = GGML_F32_VEC_LOAD(x + i + 5*ggml_f32_epr); ay6 = GGML_F32_VEC_LOAD(y + i + 5*ggml_f32_epr); - ay6 = GGML_F32_VEC_FMA(ax6, vx, ay6); + ay6 = GGML_F32_VEC_FMA(ay6, ax6, vx); GGML_F32_VEC_STORE(y + i + 5*ggml_f32_epr, ay6); ax7 = GGML_F32_VEC_LOAD(x + i + 6*ggml_f32_epr); ay7 = GGML_F32_VEC_LOAD(y + i + 6*ggml_f32_epr); - ay7 = GGML_F32_VEC_FMA(ax7, vx, ay7); + ay7 = GGML_F32_VEC_FMA(ay7, ax7, vx); GGML_F32_VEC_STORE(y + i + 6*ggml_f32_epr, ay7); ax8 = GGML_F32_VEC_LOAD(x + i + 7*ggml_f32_epr); ay8 = GGML_F32_VEC_LOAD(y + i + 7*ggml_f32_epr); - ay8 = GGML_F32_VEC_FMA(ax8, vx, ay8); + ay8 = GGML_F32_VEC_FMA(ay8, ax8, vx); GGML_F32_VEC_STORE(y + i + 7*ggml_f32_epr, ay8); } @@ -215,7 +215,7 @@ inline static void ggml_vec_mad_f32(const int n, float * GGML_RESTRICT y, const for (int i = np; i < np2; i += ggml_f32_epr) { ax1 = GGML_F32_VEC_LOAD(x + i); ay1 = GGML_F32_VEC_LOAD(y + i); - ay1 = GGML_F32_VEC_FMA(ax1, vx, ay1); + ay1 = GGML_F32_VEC_FMA(ay1, ax1, vx); GGML_F32_VEC_STORE(y + i, ay1); }