From: Aaron Teo Date: Fri, 22 Aug 2025 08:11:04 +0000 (+0800) Subject: ggml-cpu: Support Q5_0 and Q5_1 on s390x (#15486) X-Git-Tag: upstream/0.0.6527~283 X-Git-Url: https://git.djapps.eu/?a=commitdiff_plain;h=ad5c975c2d0297124fad210776ef8eed6b90d578;p=pkg%2Fggml%2Fsources%2Fllama.cpp ggml-cpu: Support Q5_0 and Q5_1 on s390x (#15486) * ggml-cpu: initial q5_0 impl for s390x Signed-off-by: Aaron Teo * ggml-cpu: updated q5_0 code for better performance Signed-off-by: Aaron Teo * ggml-cpu: use optimised hsum for better performance Signed-off-by: Aaron Teo * ggml-cpu: introduce q5_1 simd + refactor q5_0 Signed-off-by: Aaron Teo * ggml-cpu: fix incorrect return type vec_hsum Signed-off-by: Aaron Teo * ggml-cpu: q5_0 incomplete refactor + table_b2b_0 activation Signed-off-by: Aaron Teo * ggml-cpu: refactor q5_1 Signed-off-by: Aaron Teo * ggml-cpu: q5_1 update loop unroll to 4 Signed-off-by: Aaron Teo * ggml-cpu: update q5_0 unroll to 4 Signed-off-by: Aaron Teo * ggml-cpu: update build-s390x docs Signed-off-by: Aaron Teo * ggml-cpu: update unused variables q5_0 Signed-off-by: Aaron Teo * docs: update the last update date Signed-off-by: Aaron Teo --------- Signed-off-by: Aaron Teo --- diff --git a/docs/build-s390x.md b/docs/build-s390x.md index b36a1998..f3cdd63b 100644 --- a/docs/build-s390x.md +++ b/docs/build-s390x.md @@ -265,8 +265,9 @@ IBM VXE/VXE2 SIMD acceleration depends on the BLAS implementation. It is strongl | BF16 | 🚫 | 🚫 | ❓ | ❓ | | Q4_0 | ✅ | ✅ | ❓ | ❓ | | Q4_1 | ✅ | ✅ | ❓ | ❓ | -| Q5_0 | 🚫 | 🚫 | ❓ | ❓ | -| Q5_1 | 🚫 | 🚫 | ❓ | ❓ | +| MXFP4 | 🚫 | 🚫 | ❓ | ❓ | +| Q5_0 | ✅ | ✅ | ❓ | ❓ | +| Q5_1 | ✅ | ✅ | ❓ | ❓ | | Q8_0 | ✅ | ✅ | ❓ | ❓ | | Q2_K | 🚫 | 🚫 | ❓ | ❓ | | Q3_K | ✅ | ✅ | ❓ | ❓ | @@ -291,4 +292,4 @@ IBM VXE/VXE2 SIMD acceleration depends on the BLAS implementation. It is strongl - 🚫 - acceleration unavailable, will still run using scalar implementation - ❓ - acceleration unknown, please contribute if you can test it yourself -Last Updated by **Aaron Teo (aaron.teo1@ibm.com)** on July 31, 2025. +Last Updated by **Aaron Teo (aaron.teo1@ibm.com)** on Aug 22, 2025. diff --git a/ggml/src/ggml-cpu/arch-fallback.h b/ggml/src/ggml-cpu/arch-fallback.h index 0bfb92df..373408a9 100644 --- a/ggml/src/ggml-cpu/arch-fallback.h +++ b/ggml/src/ggml-cpu/arch-fallback.h @@ -150,8 +150,6 @@ #elif defined(__s390x__) // quants.c #define quantize_row_q8_K_generic quantize_row_q8_K -#define ggml_vec_dot_q5_0_q8_0_generic ggml_vec_dot_q5_0_q8_0 -#define ggml_vec_dot_q5_1_q8_1_generic ggml_vec_dot_q5_1_q8_1 #define ggml_vec_dot_tq1_0_q8_K_generic ggml_vec_dot_tq1_0_q8_K #define ggml_vec_dot_tq2_0_q8_K_generic ggml_vec_dot_tq2_0_q8_K #define ggml_vec_dot_q2_K_q8_K_generic ggml_vec_dot_q2_K_q8_K diff --git a/ggml/src/ggml-cpu/arch/s390/quants.c b/ggml/src/ggml-cpu/arch/s390/quants.c index 7e4229d0..1c8176fb 100644 --- a/ggml/src/ggml-cpu/arch/s390/quants.c +++ b/ggml/src/ggml-cpu/arch/s390/quants.c @@ -23,6 +23,27 @@ #define UNUSED GGML_UNUSED +#if defined(__VXE__) || defined(__VXE2__) +#define B1(c,s,n) 0x ## n ## c , 0x ## n ## s +#define B2(c,s,n) B1(c,s,n ## c), B1(c,s,n ## s) +#define B3(c,s,n) B2(c,s,n ## c), B2(c,s,n ## s) +#define B4(c,s,n) B3(c,s,n ## c), B3(c,s,n ## s) +#define B5(c,s,n) B4(c,s,n ## c), B4(c,s,n ## s) +#define B6(c,s,n) B5(c,s,n ## c), B5(c,s,n ## s) +#define B7(c,s,n) B6(c,s,n ## c), B6(c,s,n ## s) +#define B8(c,s ) B7(c,s, c), B7(c,s, s) + +// precomputed tables for expanding 8bits to 8 bytes: +static const __attribute__((aligned(16))) uint64_t table_b2b_0[1 << 8] = { B8(00, 10) }; // ( b ) << 4 +static const __attribute__((aligned(16))) uint64_t table_b2b_1[1 << 8] = { B8(10, 00) }; // (!b) << 4 + +// permute mask for byteswapping +static const uint8x16_t v_kperm = (const uint8x16_t){ + 7, 6, 5, 4, 3, 2, 1, 0, + 15, 14, 13, 12, 11, 10, 9, 8 +}; +#endif + void quantize_row_q8_0(const float * GGML_RESTRICT x, void * GGML_RESTRICT vy, int64_t k) { assert(QK8_0 == 32); assert(k % QK8_0 == 0); @@ -241,6 +262,301 @@ void ggml_vec_dot_q4_1_q8_1(int n, float * GGML_RESTRICT s, size_t bs, const voi #endif } +void ggml_vec_dot_q5_0_q8_0(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, size_t bx, const void * GGML_RESTRICT vy, size_t by, int nrc) { + const int qk = QK8_0; + const int nb = n / qk; + + assert(n % qk == 0); + assert(qk == QK5_0); + assert(nrc == 1); + UNUSED(nrc); + UNUSED(bx); + UNUSED(by); + UNUSED(bs); + + const block_q5_0 * GGML_RESTRICT x = vx; + const block_q8_0 * GGML_RESTRICT y = vy; + + int ib = 0; + float sumf = 0.0f; + +#if defined(__VXE__) || defined(__VXE2__) + float32x4_t v_sum0 = vec_splats(0.0f); + float32x4_t v_sum1 = vec_splats(0.0f); + + uint32_t qh0, qh1; + uint64_t tmp0[4], tmp1[4]; + + const uint8x16_t v_m = vec_splats((uint8_t)0x0F); + + #pragma GCC unroll 4 + for (; ib + 1 < nb; ib += 2) { + const block_q5_0 * GGML_RESTRICT x0 = &x[ib + 0]; + const block_q5_0 * GGML_RESTRICT x1 = &x[ib + 1]; + const block_q8_0 * GGML_RESTRICT y0 = &y[ib + 0]; + const block_q8_0 * GGML_RESTRICT y1 = &y[ib + 1]; + + memcpy(&qh0, x0->qh, sizeof(qh0)); + memcpy(&qh1, x1->qh, sizeof(qh1)); + + tmp0[0] = table_b2b_1[(qh0 >> 0) & 0xFF]; + tmp0[1] = table_b2b_1[(qh0 >> 8) & 0xFF]; + tmp0[2] = table_b2b_1[(qh0 >> 16) & 0xFF]; + tmp0[3] = table_b2b_1[(qh0 >> 24) ]; + + tmp1[0] = table_b2b_1[(qh1 >> 0) & 0xFF]; + tmp1[1] = table_b2b_1[(qh1 >> 8) & 0xFF]; + tmp1[2] = table_b2b_1[(qh1 >> 16) & 0xFF]; + tmp1[3] = table_b2b_1[(qh1 >> 24) ]; + + int8x16_t v_qh0l = vec_xl(0, (const int8_t *)(tmp0 + 0)); + int8x16_t v_qh0h = vec_xl(0, (const int8_t *)(tmp0 + 2)); + int8x16_t v_qh1l = vec_xl(0, (const int8_t *)(tmp1 + 0)); + int8x16_t v_qh1h = vec_xl(0, (const int8_t *)(tmp1 + 2)); + + // required for fixing the byteorder + v_qh0l = vec_perm(v_qh0l, v_qh0l, v_kperm); + v_qh0h = vec_perm(v_qh0h, v_qh0h, v_kperm); + v_qh1l = vec_perm(v_qh1l, v_qh1l, v_kperm); + v_qh1h = vec_perm(v_qh1h, v_qh1h, v_kperm); + + const uint8x16_t v_x0 = vec_xl(0, (const uint8_t *)x0->qs); + const uint8x16_t v_x1 = vec_xl(0, (const uint8_t *)x1->qs); + + int8x16_t v_x0l = (int8x16_t)vec_and(v_x0, v_m); + int8x16_t v_x0h = (int8x16_t)vec_sr(v_x0, 4); + int8x16_t v_x1l = (int8x16_t)vec_and(v_x1, v_m); + int8x16_t v_x1h = (int8x16_t)vec_sr(v_x1, 4); + + const int8x16_t v_x0lf = vec_sub(v_x0l, v_qh0l); + const int8x16_t v_x0hf = vec_sub(v_x0h, v_qh0h); + const int8x16_t v_x1lf = vec_sub(v_x1l, v_qh1l); + const int8x16_t v_x1hf = vec_sub(v_x1h, v_qh1h); + + const int8x16_t v_y0l = vec_xl(0, (const int8_t *)y0->qs); + const int8x16_t v_y0h = vec_xl(QK8_0/2, (const int8_t *)y0->qs); + const int8x16_t v_y1l = vec_xl(0, (const int8_t *)y1->qs); + const int8x16_t v_y1h = vec_xl(QK8_0/2, (const int8_t *)y1->qs); + + const int32x4_t v_xy0 = ggml_vec_dot(ggml_vec_dot(vec_splats(0), v_x0lf, v_y0l), v_x0hf, v_y0h); + const int32x4_t v_xy1 = ggml_vec_dot(ggml_vec_dot(vec_splats(0), v_x1lf, v_y1l), v_x1hf, v_y1h); + + const float32x4_t v_xy0f = vec_float(v_xy0); + const float32x4_t v_xy1f = vec_float(v_xy1); + + const float32x4_t v_d0 = vec_splats(GGML_CPU_FP16_TO_FP32(x0->d) * GGML_CPU_FP16_TO_FP32(y0->d)); + const float32x4_t v_d1 = vec_splats(GGML_CPU_FP16_TO_FP32(x1->d) * GGML_CPU_FP16_TO_FP32(y1->d)); + + v_sum0 = vec_madd(v_xy0f, v_d0, v_sum0); + v_sum1 = vec_madd(v_xy1f, v_d1, v_sum1); + } + + sumf += vec_hsum(v_sum0) + vec_hsum(v_sum1); + + #pragma GCC unroll 4 + for (; ib < nb; ++ib) { + const block_q5_0 * GGML_RESTRICT x0 = &x[ib]; + const block_q8_0 * GGML_RESTRICT y0 = &y[ib]; + + uint32_t qh; + memcpy(&qh, x0->qh, sizeof(qh)); + + uint64_t tmp[4]; + tmp[0] = table_b2b_1[(qh >> 0) & 0xFF]; + tmp[1] = table_b2b_1[(qh >> 8) & 0xFF]; + tmp[2] = table_b2b_1[(qh >> 16) & 0xFF]; + tmp[3] = table_b2b_1[(qh >> 24) ]; + + int8x16_t v_qhl = vec_xl(0, (const int8_t *)(tmp + 0)); + int8x16_t v_qhh = vec_xl(0, (const int8_t *)(tmp + 2)); + + // required for fixing the byteorder + v_qhl = vec_perm(v_qhl, v_qhl, v_kperm); + v_qhh = vec_perm(v_qhh, v_qhh, v_kperm); + + const uint8x16_t v_x = vec_xl(0, (const uint8_t *)x0->qs); + int8x16_t v_xl = (int8x16_t)vec_and(v_x, v_m); + int8x16_t v_xh = (int8x16_t)vec_sr(v_x, 4); + + const int8x16_t v_xlf = vec_sub(v_xl, v_qhl); + const int8x16_t v_xhf = vec_sub(v_xh, v_qhh); + + const int8x16_t v_yl = vec_xl(0, (const int8_t *)y0->qs); + const int8x16_t v_yh = vec_xl(QK8_0/2, (const int8_t *)y0->qs); + + const int32x4_t v_xy = ggml_vec_dot(ggml_vec_dot(vec_splats(0), v_xlf, v_yl), v_xhf, v_yh); + const float32x4_t v_xyf = vec_float(v_xy); + + const float32x4_t v_d = vec_splats(GGML_CPU_FP16_TO_FP32(x0->d) * GGML_CPU_FP16_TO_FP32(y0->d)); + const float32x4_t v_acc = vec_madd(v_xyf, v_d, vec_splats(0.0f)); + + sumf += vec_hsum(v_acc); + } + + *s = sumf; +#else + UNUSED(nb); + UNUSED(x); + UNUSED(y); + UNUSED(ib); + UNUSED(sumf); + ggml_vec_dot_q5_0_q8_0_generic(n, s, bs, vx, bx, vy, by, nrc); +#endif +} + +void ggml_vec_dot_q5_1_q8_1(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, size_t bx, const void * GGML_RESTRICT vy, size_t by, int nrc) { + const int qk = QK8_1; + const int nb = n / qk; + + assert(n % qk == 0); + assert(qk == QK5_1); + assert(nrc == 1); + UNUSED(nrc); + UNUSED(bx); + UNUSED(by); + UNUSED(bs); + + const block_q5_1 * GGML_RESTRICT x = vx; + const block_q8_1 * GGML_RESTRICT y = vy; + + int ib = 0; + float sumf = 0.0f; + +#if defined(__VXE__) || defined(__VXE2__) + float32x4_t v_sum0 = vec_splats(0.0f); + float32x4_t v_sum1 = vec_splats(0.0f); + + float summs0 = 0.0f; + float summs1 = 0.0f; + + uint32_t qh0; + uint32_t qh1; + + uint64_t tmp0[4]; + uint64_t tmp1[4]; + + const uint8x16_t v_m = vec_splats((uint8_t)0x0F); + + #pragma GCC unroll 4 + for (; ib + 1 < nb; ib += 2) { + const block_q5_1 * GGML_RESTRICT x0 = &x[ib + 0]; + const block_q5_1 * GGML_RESTRICT x1 = &x[ib + 1]; + const block_q8_1 * GGML_RESTRICT y0 = &y[ib + 0]; + const block_q8_1 * GGML_RESTRICT y1 = &y[ib + 1]; + + summs0 += GGML_CPU_FP16_TO_FP32(x0->m) * GGML_CPU_FP16_TO_FP32(y0->s); + summs1 += GGML_CPU_FP16_TO_FP32(x1->m) * GGML_CPU_FP16_TO_FP32(y1->s); + + memcpy(&qh0, x0->qh, sizeof(qh0)); + memcpy(&qh1, x1->qh, sizeof(qh1)); + + tmp0[0] = table_b2b_0[(qh0 >> 0) & 0xFF]; + tmp0[1] = table_b2b_0[(qh0 >> 8) & 0xFF]; + tmp0[2] = table_b2b_0[(qh0 >> 16) & 0xFF]; + tmp0[3] = table_b2b_0[(qh0 >> 24) ]; + + tmp1[0] = table_b2b_0[(qh1 >> 0) & 0xFF]; + tmp1[1] = table_b2b_0[(qh1 >> 8) & 0xFF]; + tmp1[2] = table_b2b_0[(qh1 >> 16) & 0xFF]; + tmp1[3] = table_b2b_0[(qh1 >> 24) ]; + + int8x16_t v_qh0l = vec_xl(0, (const int8_t *)(tmp0 + 0)); + int8x16_t v_qh0h = vec_xl(0, (const int8_t *)(tmp0 + 2)); + int8x16_t v_qh1l = vec_xl(0, (const int8_t *)(tmp1 + 0)); + int8x16_t v_qh1h = vec_xl(0, (const int8_t *)(tmp1 + 2)); + + // required for fixing the byteorder + v_qh0l = vec_perm(v_qh0l, v_qh0l, v_kperm); + v_qh0h = vec_perm(v_qh0h, v_qh0h, v_kperm); + v_qh1l = vec_perm(v_qh1l, v_qh1l, v_kperm); + v_qh1h = vec_perm(v_qh1h, v_qh1h, v_kperm); + + const uint8x16_t v_x0 = vec_xl(0, x0->qs); + const uint8x16_t v_x1 = vec_xl(0, x1->qs); + + const int8x16_t v_x0l = (int8x16_t)vec_and(v_x0, v_m); + const int8x16_t v_x0h = (int8x16_t)vec_sr(v_x0, 4); + const int8x16_t v_x1l = (int8x16_t)vec_and(v_x1, v_m); + const int8x16_t v_x1h = (int8x16_t)vec_sr(v_x1, 4); + + const int8x16_t v_x0lf = vec_or(v_x0l, v_qh0l); + const int8x16_t v_x0hf = vec_or(v_x0h, v_qh0h); + const int8x16_t v_x1lf = vec_or(v_x1l, v_qh1l); + const int8x16_t v_x1hf = vec_or(v_x1h, v_qh1h); + + const int8x16_t v_y0l = vec_xl(0 , y0->qs); + const int8x16_t v_y0h = vec_xl(QK8_1/2, y0->qs); + const int8x16_t v_y1l = vec_xl(0 , y1->qs); + const int8x16_t v_y1h = vec_xl(QK8_1/2, y1->qs); + + const int32x4_t v_xy0 = ggml_vec_dot(ggml_vec_dot(vec_splats(0), v_x0lf, v_y0l), v_x0hf, v_y0h); + const int32x4_t v_xy1 = ggml_vec_dot(ggml_vec_dot(vec_splats(0), v_x1lf, v_y1l), v_x1hf, v_y1h); + + const float32x4_t v_xy0f = vec_float(v_xy0); + const float32x4_t v_xy1f = vec_float(v_xy1); + + const float32x4_t v_d0 = vec_splats(GGML_CPU_FP16_TO_FP32(x0->d) * GGML_CPU_FP16_TO_FP32(y0->d)); + const float32x4_t v_d1 = vec_splats(GGML_CPU_FP16_TO_FP32(x1->d) * GGML_CPU_FP16_TO_FP32(y1->d)); + + v_sum0 = vec_madd(v_xy0f, v_d0, v_sum0); + v_sum1 = vec_madd(v_xy1f, v_d1, v_sum1); + } + + sumf += vec_hsum(v_sum0) + vec_hsum(v_sum1) + summs0 + summs1; + + #pragma GCC unroll 4 + for (; ib < nb; ++ib) { + const block_q5_1 * GGML_RESTRICT x0 = &x[ib]; + const block_q8_1 * GGML_RESTRICT y0 = &y[ib]; + + float summs = GGML_CPU_FP16_TO_FP32(x0->m) * GGML_CPU_FP16_TO_FP32(y0->s); + + uint32_t qh; + memcpy(&qh, x0->qh, sizeof(qh)); + + uint64_t tmp[4]; + tmp[0] = table_b2b_0[(qh >> 0) & 0xFF]; + tmp[1] = table_b2b_0[(qh >> 8) & 0xFF]; + tmp[2] = table_b2b_0[(qh >> 16) & 0xFF]; + tmp[3] = table_b2b_0[(qh >> 24) ]; + + int8x16_t v_qhl = vec_xl(0, (const int8_t *)(tmp + 0)); + int8x16_t v_qhh = vec_xl(0, (const int8_t *)(tmp + 2)); + + // required for fixing the byteorder + v_qhl = vec_perm(v_qhl, v_qhl, v_kperm); + v_qhh = vec_perm(v_qhh, v_qhh, v_kperm); + + const uint8x16_t v_x = vec_xl(0, x0->qs); + const int8x16_t v_xl = (int8x16_t)vec_and(v_x, v_m); + const int8x16_t v_xh = (int8x16_t)vec_sr(v_x, 4); + + const int8x16_t v_xlf = vec_or(v_xl, v_qhl); + const int8x16_t v_xhf = vec_or(v_xh, v_qhh); + + const int8x16_t v_yl = vec_xl(0 , y0->qs); + const int8x16_t v_yh = vec_xl(QK8_1/2, y0->qs); + + const int32x4_t v_xy = ggml_vec_dot(ggml_vec_dot(vec_splats(0), v_xlf, v_yl), v_xhf, v_yh); + const float32x4_t v_xyf = vec_float(v_xy); + + const float32x4_t v_d = vec_splats(GGML_CPU_FP16_TO_FP32(x0->d) * GGML_CPU_FP16_TO_FP32(y0->d)); + const float32x4_t v_acc = vec_madd(v_xyf, v_d, v_acc); + + sumf += vec_hsum(v_acc) + summs; + } + + *s = sumf; +#else + UNUSED(nb); + UNUSED(x); + UNUSED(y); + UNUSED(ib); + UNUSED(sumf); + ggml_vec_dot_q5_1_q8_1_generic(n, s, bs, vx, bx, vy, by, nrc); +#endif +} + void ggml_vec_dot_q8_0_q8_0(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, size_t bx, const void * GGML_RESTRICT vy, size_t by, int nrc) { const int qk = QK8_0; const int nb = n / qk; diff --git a/ggml/src/ggml-cpu/ggml-cpu-impl.h b/ggml/src/ggml-cpu/ggml-cpu-impl.h index d839cf5c..1f6844e1 100644 --- a/ggml/src/ggml-cpu/ggml-cpu-impl.h +++ b/ggml/src/ggml-cpu/ggml-cpu-impl.h @@ -486,6 +486,14 @@ inline static int16x8_t vec_padd_s16(int16x8_t a, int16x8_t b) { return v_abo + v_abe; } +/** + * @see https://github.com/ggml-org/llama.cpp/pull/14037 + */ +inline float vec_hsum(float32x4_t v) { + float32x4_t v_temp = v + vec_reve(v); + return v_temp[0] + v_temp[1]; +} + inline static int32x4_t ggml_vec_dot(int32x4_t acc, int8x16_t a, int8x16_t b) { const int16x8_t p = vec_mule(a, b) + vec_mulo(a, b); return acc + (vec_unpackh(p) + vec_unpackl(p));