From 0d26d8ccd8caebab75af697c0275f599075fdacf Mon Sep 17 00:00:00 2001 From: Hong Bo PENG Date: Sun, 12 May 2024 17:17:18 +0800 Subject: [PATCH] ggml : optimize for ppc64le using VSX intrinsics (ggml/784) * optimize for ppc64le using VSX intrinsics * 1. code clean up by removing comments about overflow concern. 2. fix typo in suffix of scaling. * Continue to fix typo in suffix of scaling for QK_K <> 256 --------- Co-authored-by: Georgi Gerganov --- ggml-quants.c | 2169 ++++++++++++++++++++++++++++++++++++++++++++++++- 1 file changed, 2167 insertions(+), 2 deletions(-) diff --git a/ggml-quants.c b/ggml-quants.c index 00334c5fe..f711bd013 100644 --- a/ggml-quants.c +++ b/ggml-quants.c @@ -241,7 +241,7 @@ static inline float hsum_float_4x4(const __m128 a, const __m128 b, const __m128 #endif // __AVX__ || __AVX2__ || __AVX512F__ #endif // defined(__AVX__) || defined(__AVX2__) || defined(__AVX512F__) || defined(__SSSE3__) -#if defined(__ARM_NEON) || defined(__wasm_simd128__) +#if defined(__ARM_NEON) || defined(__wasm_simd128__) || defined(__POWER9_VECTOR__) #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) @@ -643,6 +643,38 @@ void quantize_row_q8_0(const float * restrict x, void * restrict vy, int64_t k) // store result __riscv_vse8_v_i8m1(y[i].qs , vs, vl); } +#elif defined(__POWER9_VECTOR__) + for (int i = 0; i < nb; i++) { + vector float srcv [8]; + vector float asrcv[8]; + vector float amaxv[8]; + vector signed int vi[8]; + + for (int j = 0; j < 8; j++) srcv[j] = vec_xl(0, x + i*32 + 4*j); + for (int j = 0; j < 8; j++) asrcv[j] = vec_abs(srcv[j]); + + for (int j = 0; j < 4; j++) amaxv[2*j] = vec_max(asrcv[2*j], asrcv[2*j+1]); + for (int j = 0; j < 2; j++) amaxv[4*j] = vec_max(amaxv[4*j], amaxv[4*j+2]); + for (int j = 0; j < 1; j++) amaxv[8*j] = vec_max(amaxv[8*j], amaxv[8*j+4]); + + const float amax = MAX(MAX(vec_extract(amaxv[0], 0), + vec_extract(amaxv[0], 1)), + MAX(vec_extract(amaxv[0], 2), + vec_extract(amaxv[0], 3))); + + const float d = amax / ((1 << 7) - 1); + const float id = d ? 1.0f/d : 0.0f; + const vector float vid = vec_splats(id); + + y[i].d = GGML_FP32_TO_FP16(d); + + for (int j = 0; j < 8; j++) { + const vector float v = vec_round(vec_mul(srcv[j], vid)); + vi[j] = vec_cts(v, 0); + } + vec_xst(vec_pack(vec_pack(vi[0], vi[1]), vec_pack(vi[2], vi[3])), 0, &y[i].qs[0]); + vec_xst(vec_pack(vec_pack(vi[4], vi[5]), vec_pack(vi[6], vi[7])), 16, &y[i].qs[0]); + } #else GGML_UNUSED(nb); // scalar @@ -898,6 +930,46 @@ void quantize_row_q8_1(const float * restrict x, void * restrict vy, int64_t k) int sum = __riscv_vmv_x_s_i16m1_i16(vwrs); y[i].s = GGML_FP32_TO_FP16(sum*d); } +#elif defined(__POWER9_VECTOR__) + for (int i = 0; i < nb; i++) { + vector float srcv [8]; + vector float asrcv[8]; + vector float amaxv[8]; + vector signed int vi[8]; + + for (int j = 0; j < 8; j++) srcv[j] = vec_xl(0, x + i*32 + 4*j); + for (int j = 0; j < 8; j++) asrcv[j] = vec_abs(srcv[j]); + + for (int j = 0; j < 4; j++) amaxv[2*j] = vec_max(asrcv[2*j], asrcv[2*j+1]); + for (int j = 0; j < 2; j++) amaxv[4*j] = vec_max(amaxv[4*j], amaxv[4*j+2]); + for (int j = 0; j < 1; j++) amaxv[8*j] = vec_max(amaxv[8*j], amaxv[8*j+4]); + + const float amax = MAX(MAX(vec_extract(amaxv[0], 0), + vec_extract(amaxv[0], 1)), + MAX(vec_extract(amaxv[0], 2), + vec_extract(amaxv[0], 3))); + + const float d = amax / ((1 << 7) - 1); + const float id = d ? 1.0f/d : 0.0f; + const vector float vid = vec_splats(id); + + y[i].d = GGML_FP32_TO_FP16(d); + + vector int accv = vec_splats(0); + + for (int j = 0; j < 8; j++) { + const vector float v = vec_round(vec_mul(srcv[j], vid)); + vi[j] = vec_cts(v, 0); + + accv = vec_add(accv, vi[j]); + } + vec_xst(vec_pack(vec_pack(vi[0], vi[1]), vec_pack(vi[2], vi[3])), 0, &y[i].qs[0]); + vec_xst(vec_pack(vec_pack(vi[4], vi[5]), vec_pack(vi[6], vi[7])), 16, &y[i].qs[0]); + + accv = vec_add(accv, vec_sld(accv, accv, 4)); + accv = vec_add(accv, vec_sld(accv, accv, 8)); + y[i].s = GGML_FP32_TO_FP16(d * vec_extract(accv, 0)); + } #else GGML_UNUSED(nb); // scalar @@ -3740,6 +3812,46 @@ void ggml_vec_dot_q4_0_q8_0(int n, float * restrict s, size_t bs, const void * r } *s = sumf; +#elif defined(__POWER9_VECTOR__) + const vector signed char lowMask = vec_splats((signed char)0xF); + const vector unsigned char v4 = vec_splats((unsigned char)0x4); + const vector signed char v8 = vec_splats((signed char)0x8); + + vector float vsumf0 = vec_splats(0.0f); + +#pragma GCC unroll 4 + for (int i = 0; i < nb; i++) { + __builtin_prefetch(x[i].qs, 0, 1); + __builtin_prefetch(y[i].qs, 0, 1); + + vector float vxd = vec_splats(GGML_FP16_TO_FP32(x[i].d)); + vector float vyd = vec_splats(GGML_FP16_TO_FP32(y[i].d)); + vector float vd = vec_mul(vxd, vyd); + + vector signed char qxs = (vector signed char)vec_xl( 0, x[i].qs); + vector signed char q8y0 = vec_xl( 0, y[i].qs); + vector signed char q8y1 = vec_xl(16, y[i].qs); + + vector signed char q4x0 = vec_and(qxs, lowMask); + vector signed char q4x1 = vec_sr(qxs, v4); + + q4x0 = vec_sub(q4x0, v8); + q4x1 = vec_sub(q4x1, v8); + + vector signed short qv0 = vec_add(vec_mule(q4x0, q8y0), vec_mulo(q4x0, q8y0)); + vector signed short qv1 = vec_add(vec_mule(q4x1, q8y1), vec_mulo(q4x1, q8y1)); + + qv0 = vec_add(qv0, qv1); + + vector signed int vsumi0 = vec_add(vec_unpackh(qv0), vec_unpackl(qv0)); + + vsumf0 = vec_madd(vec_ctf(vsumi0, 0), vd, vsumf0); + } + + vsumf0 = vec_add(vsumf0, vec_sld(vsumf0, vsumf0, 4)); + vsumf0 = vec_add(vsumf0, vec_sld(vsumf0, vsumf0, 8)); + + *s = vec_extract(vsumf0, 0); #else // scalar float sumf = 0.0; @@ -3958,6 +4070,46 @@ void ggml_vec_dot_q4_1_q8_1(int n, float * restrict s, size_t bs, const void * r } *s = sumf; +#elif defined(__POWER9_VECTOR__) + const vector signed char lowMask = vec_splats((signed char)0xF); + const vector unsigned char v4 = vec_splats((unsigned char)0x4); + + vector float vsumf0 = vec_splats(0.0f); + +#pragma GCC unroll 4 + for (int i = 0; i < nb; i++) { + __builtin_prefetch(x[i].qs, 0, 1); + __builtin_prefetch(y[i].qs, 0, 1); + + vector float vxd = vec_splats(GGML_FP16_TO_FP32(x[i].d)); + vector float vyd = vec_splats(GGML_FP16_TO_FP32(y[i].d)); + vector float vd = vec_mul(vxd, vyd); + + vector float vxmin = vec_splats(GGML_FP16_TO_FP32(x[i].m)); + vector float vys = {GGML_FP16_TO_FP32(y[i].s), 0.0f, 0.0f, 0.0f}; + vsumf0 = vec_madd(vxmin, vys, vsumf0); + + vector signed char qxs = (vector signed char)vec_xl( 0, x[i].qs); + vector signed char q8y0 = vec_xl( 0, y[i].qs); + vector signed char q8y1 = vec_xl(16, y[i].qs); + + vector signed char q4x0 = vec_and(qxs, lowMask); + vector signed char q4x1 = vec_sr(qxs, v4); + + vector signed short qv0 = vec_add(vec_mule(q4x0, q8y0), vec_mulo(q4x0, q8y0)); + vector signed short qv1 = vec_add(vec_mule(q4x1, q8y1), vec_mulo(q4x1, q8y1)); + + qv0 = vec_add(qv0, qv1); + + vector signed int vsumi0 = vec_add(vec_unpackh(qv0), vec_unpackl(qv0)); + + vsumf0 = vec_madd(vec_ctf(vsumi0, 0), vd, vsumf0); + } + + vsumf0 = vec_add(vsumf0, vec_sld(vsumf0, vsumf0, 4)); + vsumf0 = vec_add(vsumf0, vec_sld(vsumf0, vsumf0, 8)); + + *s = vec_extract(vsumf0, 0); #else // scalar float sumf = 0.0; @@ -4243,6 +4395,49 @@ void ggml_vec_dot_q5_0_q8_0(int n, float * restrict s, size_t bs, const void * r } *s = sumf; +#elif defined(__POWER9_VECTOR__) + const vector signed char lowMask = vec_splats((signed char)0xF); + const vector unsigned char v4 = vec_splats((unsigned char)4); + + vector float vsumf0 = vec_splats(0.0f); + +#pragma GCC unroll 4 + for (int i = 0; i < nb; ++i) { + __builtin_prefetch(x[i].qs, 0, 1); + __builtin_prefetch(y[i].qs, 0, 1); + + vector float vxd = vec_splats(GGML_FP16_TO_FP32(x[i].d)); + vector float vyd = vec_splats(GGML_FP16_TO_FP32(y[i].d)); + vector float vd = vec_mul(vxd, vyd); + + vector signed long long aux64x2_0 = {(uint64_t)(table_b2b_1[x[i].qh[0]]), (uint64_t)(table_b2b_1[x[i].qh[1]])}; + vector signed long long aux64x2_1 = {(uint64_t)(table_b2b_1[x[i].qh[2]]), (uint64_t)(table_b2b_1[x[i].qh[3]])}; + + vector signed char qh0 = (vector signed char)aux64x2_0; + vector signed char qh1 = (vector signed char)aux64x2_1; + + vector signed char qxs = (vector signed char)vec_xl( 0, x[i].qs); + + vector signed char q5x0 = vec_sub(vec_and (qxs, lowMask), qh0); + vector signed char q5x1 = vec_sub(vec_sr(qxs, v4), qh1); + + vector signed char q8y0 = vec_xl( 0, y[i].qs); + vector signed char q8y1 = vec_xl( 16, y[i].qs); + + vector signed short qv0 = vec_add(vec_mule(q5x0, q8y0), vec_mulo(q5x0, q8y0)); + vector signed short qv1 = vec_add(vec_mule(q5x1, q8y1), vec_mulo(q5x1, q8y1)); + + qv0 = vec_add(qv0, qv1); + + vector signed int vsumi0 = vec_add(vec_unpackh(qv0), vec_unpackl(qv0)); + + vsumf0 = vec_madd(vec_ctf(vsumi0, 0), vd, vsumf0); + } + + vsumf0 = vec_add(vsumf0, vec_sld(vsumf0, vsumf0, 4)); + vsumf0 = vec_add(vsumf0, vec_sld(vsumf0, vsumf0, 8)); + + *s = vec_extract(vsumf0, 0); #else // scalar float sumf = 0.0; @@ -4547,6 +4742,53 @@ void ggml_vec_dot_q5_1_q8_1(int n, float * restrict s, size_t bs, const void * r } *s = sumf; +#elif defined(__POWER9_VECTOR__) + const vector signed char lowMask = vec_splats((signed char)0xF); + const vector unsigned char v4 = vec_splats((unsigned char)0x4); + + vector float vsumf0 = vec_splats(0.0f); + +#pragma GCC unroll 4 + for (int i = 0; i < nb; ++i) { + __builtin_prefetch(x[i].qs, 0, 1); + __builtin_prefetch(y[i].qs, 0, 1); + + vector float vxd = vec_splats(GGML_FP16_TO_FP32(x[i].d)); + vector float vyd = vec_splats(GGML_FP16_TO_FP32(y[i].d)); + vector float vd = vec_mul(vxd, vyd); + + vector float vxmin = vec_splats(GGML_FP16_TO_FP32(x[i].m)); + vector float vys = {GGML_FP16_TO_FP32(y[i].s), 0.f, 0.f, 0.f}; + vsumf0 = vec_madd(vxmin, vys, vsumf0); + + vector unsigned long long aux64x2_0 = {(uint64_t)(table_b2b_0[x[i].qh[0]]), (uint64_t)(table_b2b_0[x[i].qh[1]])}; + vector unsigned long long aux64x2_1 = {(uint64_t)(table_b2b_0[x[i].qh[2]]), (uint64_t)(table_b2b_0[x[i].qh[3]])}; + + vector signed char qh0 = (vector signed char)aux64x2_0; + vector signed char qh1 = (vector signed char)aux64x2_1; + + vector signed char qxs = (vector signed char)vec_xl( 0, x[i].qs); + + vector signed char q5x0 = vec_or(vec_and(qxs, lowMask), qh0); + vector signed char q5x1 = vec_or(vec_sr(qxs, v4), qh1); + + vector signed char q8y0 = vec_xl( 0, y[i].qs); + vector signed char q8y1 = vec_xl( 16, y[i].qs); + + vector signed short qv0 = vec_add(vec_mule(q5x0, q8y0), vec_mulo(q5x0, q8y0)); + vector signed short qv1 = vec_add(vec_mule(q5x1, q8y1), vec_mulo(q5x1, q8y1)); + + qv0 = vec_add(qv0, qv1); + + vector signed int vsumi0 = vec_add(vec_unpackh(qv0), vec_unpackl(qv0)); + + vsumf0 = vec_madd(vec_ctf(vsumi0, 0), vd, vsumf0); + } + + vsumf0 = vec_add(vsumf0, vec_sld(vsumf0, vsumf0, 4)); + vsumf0 = vec_add(vsumf0, vec_sld(vsumf0, vsumf0, 8)); + + *s = vec_extract(vsumf0, 0); #else // scalar float sumf = 0.0; @@ -4722,6 +4964,45 @@ void ggml_vec_dot_q8_0_q8_0(int n, float * restrict s, size_t bs, const void * r } *s = sumf; +#elif defined(__POWER9_VECTOR__) + vector float vsumf0 = vec_splats(0.0f); + +#pragma GCC unroll 4 + for (int i = 0; i < nb; i++) { + __builtin_prefetch(x[i].qs, 0, 1); + __builtin_prefetch(y[i].qs, 0, 1); + + vector float vxd = vec_splats(GGML_FP16_TO_FP32(x[i].d)); + vector float vyd = vec_splats(GGML_FP16_TO_FP32(y[i].d)); + vector float vd = vec_mul(vxd, vyd); + + vector signed char q8x0 = vec_xl( 0, x[i].qs); + vector signed char q8x1 = vec_xl(16, x[i].qs); + vector signed char q8y0 = vec_xl( 0, y[i].qs); + vector signed char q8y1 = vec_xl(16, y[i].qs); + + vector signed short qv0 = vec_mule(q8x0, q8y0); + vector signed short qv1 = vec_mulo(q8x0, q8y0); + vector signed short qv2 = vec_mule(q8x1, q8y1); + vector signed short qv3 = vec_mulo(q8x1, q8y1); + + vector signed int vsumi0 = vec_add(vec_unpackh(qv0), vec_unpackh(qv1)); + vector signed int vsumi1 = vec_add(vec_unpackl(qv0), vec_unpackl(qv1)); + vector signed int vsumi2 = vec_add(vec_unpackh(qv2), vec_unpackh(qv3)); + vector signed int vsumi3 = vec_add(vec_unpackl(qv2), vec_unpackl(qv3)); + + vsumi0 = vec_add(vsumi0, vsumi2); + vsumi1 = vec_add(vsumi1, vsumi3); + + vsumi0 = vec_add(vsumi0, vsumi1); + + vsumf0 = vec_madd(vec_ctf(vsumi0, 0), vd, vsumf0); + } + + vsumf0 = vec_add(vsumf0, vec_sld(vsumf0, vsumf0, 4)); + vsumf0 = vec_add(vsumf0, vec_sld(vsumf0, vsumf0, 8)); + + *s = vec_extract(vsumf0, 0); #else // scalar float sumf = 0.0; @@ -5077,6 +5358,147 @@ void ggml_vec_dot_q2_K_q8_K(int n, float * restrict s, size_t bs, const void * r *s = sumf; +#elif defined(__POWER9_VECTOR__) + const vector signed char lowMask = vec_splats((signed char)0x3); + const vector signed char lowScaleMask = vec_splats((signed char)0xF); + const vector unsigned char v2 = vec_splats((unsigned char)0x2); + const vector unsigned char v6 = vec_splats((unsigned char)0x6); + const vector unsigned char v4 = vec_splats((unsigned char)0x4); + + vector float vsumf0 = vec_splats(0.0f); + vector float vsumf1 = vec_splats(0.0f); + vector float vsumf2 = vec_splats(0.0f); + vector float vsumf3 = vec_splats(0.0f); + + for (int i = 0; i < nb; ++i) { + vector float vxd = vec_splats(GGML_FP16_TO_FP32(x[i].d)); + vector float vyd = vec_splats(y[i].d); + vector float vd = vec_mul(vxd, vyd); + + vector float vxmin = vec_splats(GGML_FP16_TO_FP32(x[i].dmin)); + vector float vdmin = vec_mul(vxmin, vyd); + + vector signed short q8ysums0 = vec_xl( 0, y[i].bsums); + vector signed short q8ysums1 = vec_xl(16, y[i].bsums); + + vector signed char q2xmins = (vector signed char)vec_xl( 0, x[i].scales); + vector signed char vscales = vec_and(q2xmins, lowScaleMask); + + q2xmins = vec_sr(q2xmins, v4); + vector signed short q2xmins0 = vec_unpackh(q2xmins); + vector signed short q2xmins1 = vec_unpackl(q2xmins); + + vector signed int prod0 = vec_mule(q2xmins0, q8ysums0); + vector signed int prod1 = vec_mulo(q2xmins0, q8ysums0); + vector signed int prod2 = vec_mule(q2xmins1, q8ysums1); + vector signed int prod3 = vec_mulo(q2xmins1, q8ysums1); + + vsumf0 = vec_nmsub(vec_ctf(prod0, 0), vdmin, vsumf0); + vsumf1 = vec_nmsub(vec_ctf(prod1, 0), vdmin, vsumf1); + vsumf2 = vec_nmsub(vec_ctf(prod2, 0), vdmin, vsumf2); + vsumf3 = vec_nmsub(vec_ctf(prod3, 0), vdmin, vsumf3); + + vector signed int vsumi0 = vec_splats((int32_t)0); + vector signed int vsumi1 = vec_splats((int32_t)0); + vector signed int vsumi2 = vec_splats((int32_t)0); + vector signed int vsumi3 = vec_splats((int32_t)0); + vector signed int vsumi4 = vec_splats((int32_t)0); + vector signed int vsumi5 = vec_splats((int32_t)0); + vector signed int vsumi6 = vec_splats((int32_t)0); + vector signed int vsumi7 = vec_splats((int32_t)0); + + const uint8_t * restrict q2 = x[i].qs; + const int8_t * restrict q8 = y[i].qs; + + for (int j = 0; j < QK_K/128; ++j) { + __builtin_prefetch(q2, 0, 1); + __builtin_prefetch(q8, 0, 1); + + vector signed char qxs0 = (vector signed char)vec_xl( 0, q2); + vector signed char qxs1 = (vector signed char)vec_xl(16, q2); + q2 += 32; + + vector signed char q2x00 = vec_and(qxs0, lowMask); + vector signed char q2x01 = vec_and(vec_sr(qxs0, v2), lowMask); + vector signed char q2x02 = vec_and(vec_sr(qxs0, v4), lowMask); + vector signed char q2x03 = vec_and(vec_sr(qxs0, v6), lowMask); + vector signed char q2x10 = vec_and(qxs1, lowMask); + vector signed char q2x11 = vec_and(vec_sr(qxs1, v2), lowMask); + vector signed char q2x12 = vec_and(vec_sr(qxs1, v4), lowMask); + vector signed char q2x13 = vec_and(vec_sr(qxs1, v6), lowMask); + + vector signed char q8y00 = vec_xl( 0, q8); + vector signed char q8y10 = vec_xl( 16, q8); + vector signed char q8y01 = vec_xl( 32, q8); + vector signed char q8y11 = vec_xl( 48, q8); + vector signed char q8y02 = vec_xl( 64, q8); + vector signed char q8y12 = vec_xl( 80, q8); + vector signed char q8y03 = vec_xl( 96, q8); + vector signed char q8y13 = vec_xl(112, q8); + q8 += 128; + + vector signed short qv0 = vec_add(vec_mule(q2x00, q8y00), vec_mulo(q2x00, q8y00)); + vector signed short qv1 = vec_add(vec_mule(q2x01, q8y01), vec_mulo(q2x01, q8y01)); + vector signed short qv2 = vec_add(vec_mule(q2x02, q8y02), vec_mulo(q2x02, q8y02)); + vector signed short qv3 = vec_add(vec_mule(q2x03, q8y03), vec_mulo(q2x03, q8y03)); + vector signed short qv4 = vec_add(vec_mule(q2x10, q8y10), vec_mulo(q2x10, q8y10)); + vector signed short qv5 = vec_add(vec_mule(q2x11, q8y11), vec_mulo(q2x11, q8y11)); + vector signed short qv6 = vec_add(vec_mule(q2x12, q8y12), vec_mulo(q2x12, q8y12)); + vector signed short qv7 = vec_add(vec_mule(q2x13, q8y13), vec_mulo(q2x13, q8y13)); + + vector signed short vscales_h = vec_unpackh(vscales); + vector signed short vs0 = vec_splat(vscales_h, 0); + vector signed short vs1 = vec_splat(vscales_h, 1); + vector signed short vs2 = vec_splat(vscales_h, 2); + vector signed short vs3 = vec_splat(vscales_h, 3); + vector signed short vs4 = vec_splat(vscales_h, 4); + vector signed short vs5 = vec_splat(vscales_h, 5); + vector signed short vs6 = vec_splat(vscales_h, 6); + vector signed short vs7 = vec_splat(vscales_h, 7); + vscales = vec_sld(vscales, vscales, 8); + + qv0 = vec_mul(qv0, vs0); + qv1 = vec_mul(qv1, vs2); + qv2 = vec_mul(qv2, vs4); + qv3 = vec_mul(qv3, vs6); + + qv0 = vec_madd(qv4, vs1, qv0); + qv1 = vec_madd(qv5, vs3, qv1); + qv2 = vec_madd(qv6, vs5, qv2); + qv3 = vec_madd(qv7, vs7, qv3); + + vsumi0 = vec_add(vec_unpackh(qv0), vsumi0); + vsumi1 = vec_add(vec_unpackh(qv1), vsumi1); + vsumi2 = vec_add(vec_unpackh(qv2), vsumi2); + vsumi3 = vec_add(vec_unpackh(qv3), vsumi3); + + vsumi4 = vec_add(vec_unpackl(qv0), vsumi4); + vsumi5 = vec_add(vec_unpackl(qv1), vsumi5); + vsumi6 = vec_add(vec_unpackl(qv2), vsumi6); + vsumi7 = vec_add(vec_unpackl(qv3), vsumi7); + } + + vsumi0 = vec_add(vsumi0, vsumi4); + vsumi1 = vec_add(vsumi1, vsumi5); + vsumi2 = vec_add(vsumi2, vsumi6); + vsumi3 = vec_add(vsumi3, vsumi7); + + vsumf0 = vec_madd(vec_ctf(vsumi0, 0), vd, vsumf0); + vsumf1 = vec_madd(vec_ctf(vsumi1, 0), vd, vsumf1); + vsumf2 = vec_madd(vec_ctf(vsumi2, 0), vd, vsumf2); + vsumf3 = vec_madd(vec_ctf(vsumi3, 0), vd, vsumf3); + } + + vsumf0 = vec_add(vsumf0, vsumf2); + vsumf1 = vec_add(vsumf1, vsumf3); + + vsumf0 = vec_add(vsumf0, vsumf1); + + vsumf0 = vec_add(vsumf0, vec_sld(vsumf0, vsumf0, 4)); + vsumf0 = vec_add(vsumf0, vec_sld(vsumf0, vsumf0, 8)); + + *s = vec_extract(vsumf0, 0); + #else float sumf = 0; @@ -5347,6 +5769,87 @@ void ggml_vec_dot_q2_K_q8_K(int n, float * restrict s, size_t bs, const void * r *s = sumf; +#elif defined(__POWER9_VECTOR__) + const vector signed char lowMask = vec_splats((signed char)0x3); + const vector signed char lowScaleMask = vec_splats((signed char)0xF); + const vector unsigned char v2 = vec_splats((unsigned char)0x2); + const vector unsigned char v4 = vec_splats((unsigned char)0x4); + const vector unsigned char v6 = vec_splats((unsigned char)0x6); + + vector float vsumf0 = vec_splats(0.0f); + vector float vsumf1 = vec_splats(0.0f); + vector float vsumf2 = vec_splats(0.0f); + vector float vsumf3 = vec_splats(0.0f); + +#pragma GCC unroll 2 + for (int i = 0; i < nb; ++i) { + __builtin_prefetch(x[i].qs, 0, 1); + __builtin_prefetch(y[i].qs, 0, 1); + + vector float vxd = vec_splats(GGML_FP16_TO_FP32(x[i].d)); + vector float vyd = vec_splats(y[i].d); + vector float vd = vec_mul(vxd, vyd); + + vector float vxmin = vec_splats(GGML_FP16_TO_FP32(x[i].dmin)); + vector float vdmin = vec_mul(vxmin, vyd); + + vector signed short q8ysums0 = vec_xl_len(y[i].bsums, 8); + + vector signed char q2xmins = (vector signed char)vec_xl_len(x[i].scales, 4); + vector signed char vscales = vec_and(q2xmins, lowScaleMask); + + q2xmins = vec_sr(q2xmins, v4); + vector signed short q2xmins0 = vec_unpackh((vector signed char)q2xmins); + + vector signed int prod0 = vec_mule(q2xmins0, q8ysums0); + vector signed int prod1 = vec_mulo(q2xmins0, q8ysums0); + + vsumf0 = vec_nmsub(vec_ctf(prod0, 0), vdmin, vsumf0); + vsumf1 = vec_nmsub(vec_ctf(prod1, 0), vdmin, vsumf1); + + vector signed char qxs0 = (vector signed char)vec_xl( 0, x[i].qs); + vector signed char q2x00 = vec_and(qxs0, lowMask); + vector signed char q2x01 = vec_and(vec_sr(qxs0, v2), lowMask); + vector signed char q2x02 = vec_and(vec_sr(qxs0, v4), lowMask); + vector signed char q2x03 = vec_and(vec_sr(qxs0, v6), lowMask); + + vector signed char q8y00 = vec_xl( 0, y[i].qs); + vector signed char q8y01 = vec_xl( 16, y[i].qs); + vector signed char q8y02 = vec_xl( 32, y[i].qs); + vector signed char q8y03 = vec_xl( 48, y[i].qs); + + vector signed short qv0 = vec_add(vec_mule(q2x00, q8y00), vec_mulo(q2x00, q8y00)); + vector signed short qv1 = vec_add(vec_mule(q2x01, q8y01), vec_mulo(q2x01, q8y01)); + vector signed short qv2 = vec_add(vec_mule(q2x02, q8y02), vec_mulo(q2x02, q8y02)); + vector signed short qv3 = vec_add(vec_mule(q2x03, q8y03), vec_mulo(q2x03, q8y03)); + + vector signed short vscales_h = vec_unpackh(vscales); + vector signed short vs0 = vec_splat(vscales_h, 0); + vector signed short vs1 = vec_splat(vscales_h, 1); + vector signed short vs2 = vec_splat(vscales_h, 2); + vector signed short vs3 = vec_splat(vscales_h, 3); + + vector signed int vsumi0 = vec_add(vec_mule(qv0, vs0), vec_mulo(qv0, vs0)); + vector signed int vsumi1 = vec_add(vec_mule(qv1, vs1), vec_mulo(qv1, vs1)); + vector signed int vsumi2 = vec_add(vec_mule(qv2, vs2), vec_mulo(qv2, vs2)); + vector signed int vsumi3 = vec_add(vec_mule(qv3, vs3), vec_mulo(qv3, vs3)); + + vsumf0 = vec_madd(vec_ctf(vsumi0, 0), vd, vsumf0); + vsumf1 = vec_madd(vec_ctf(vsumi1, 0), vd, vsumf1); + vsumf2 = vec_madd(vec_ctf(vsumi2, 0), vd, vsumf2); + vsumf3 = vec_madd(vec_ctf(vsumi3, 0), vd, vsumf3); + } + + vsumf0 = vec_add(vsumf0, vsumf2); + vsumf1 = vec_add(vsumf1, vsumf3); + + vsumf0 = vec_add(vsumf0, vsumf1); + + vsumf0 = vec_add(vsumf0, vec_sld(vsumf0, vsumf0, 4)); + vsumf0 = vec_add(vsumf0, vec_sld(vsumf0, vsumf0, 8)); + + *s = vec_extract(vsumf0, 0); + #else float sumf = 0; @@ -5841,6 +6344,160 @@ void ggml_vec_dot_q3_K_q8_K(int n, float * restrict s, size_t bs, const void * r *s = sumf; +#elif defined(__POWER9_VECTOR__) + const vector signed char lowMask = vec_splats((signed char)0x3); + const vector signed char v1 = vec_splats((signed char)0x1); + const vector unsigned char v2 = vec_splats((unsigned char)0x2); + const vector unsigned char v3 = vec_splats((unsigned char)0x3); + const vector unsigned char v4 = vec_splats((unsigned char)0x4); + const vector unsigned char v6 = vec_splats((unsigned char)0x6); + const vector signed char off = vec_splats((signed char)0x20); + + vector float vsumf0 = vec_splats(0.0f); + vector float vsumf1 = vec_splats(0.0f); + vector float vsumf2 = vec_splats(0.0f); + vector float vsumf3 = vec_splats(0.0f); + + for (int i = 0; i < nb; ++i) { + vector float vxd = vec_splats(GGML_FP16_TO_FP32(x[i].d)); + vector float vyd = vec_splats(y[i].d); + vector float vd = vec_mul(vxd, vyd); + + uint32_t aux[3]; + uint32_t utmp[4]; + + memcpy(aux, x[i].scales, 12); + utmp[3] = ((aux[1] >> 4) & kmask2) | (((aux[2] >> 6) & kmask1) << 4); + utmp[2] = ((aux[0] >> 4) & kmask2) | (((aux[2] >> 4) & kmask1) << 4); + utmp[1] = (aux[1] & kmask2) | (((aux[2] >> 2) & kmask1) << 4); + utmp[0] = (aux[0] & kmask2) | (((aux[2] >> 0) & kmask1) << 4); + + vector signed char vscales = (vector signed char)vec_xl( 0, utmp); + vector signed char qxhs0 = (vector signed char)vec_xl( 0, x[i].hmask); + vector signed char qxhs1 = (vector signed char)vec_xl(16, x[i].hmask); + + vscales = vec_sub(vscales, off); + + vector signed int vsumi0 = vec_splats((int32_t)0); + vector signed int vsumi1 = vec_splats((int32_t)0); + vector signed int vsumi2 = vec_splats((int32_t)0); + vector signed int vsumi3 = vec_splats((int32_t)0); + vector signed int vsumi4 = vec_splats((int32_t)0); + vector signed int vsumi5 = vec_splats((int32_t)0); + vector signed int vsumi6 = vec_splats((int32_t)0); + vector signed int vsumi7 = vec_splats((int32_t)0); + + const uint8_t * restrict q3 = x[i].qs; + const int8_t * restrict q8 = y[i].qs; + + for (int j = 0; j < QK_K/128; ++j) { + __builtin_prefetch(q3, 0, 1); + __builtin_prefetch(q8, 0, 1); + + vector signed char qxs0 = (vector signed char)vec_xl( 0, q3); + vector signed char qxs1 = (vector signed char)vec_xl(16, q3); + q3 += 32; + + //the low 2 bits + vector signed char qxs00 = vec_and(qxs0, lowMask); + vector signed char qxs01 = vec_and(vec_sr(qxs0, v2), lowMask); + vector signed char qxs02 = vec_and(vec_sr(qxs0, v4), lowMask); + vector signed char qxs03 = vec_and(vec_sr(qxs0, v6), lowMask); + vector signed char qxs10 = vec_and(qxs1, lowMask); + vector signed char qxs11 = vec_and(vec_sr(qxs1, v2), lowMask); + vector signed char qxs12 = vec_and(vec_sr(qxs1, v4), lowMask); + vector signed char qxs13 = vec_and(vec_sr(qxs1, v6), lowMask); + + //the 3rd bit + vector signed char qxh00 = vec_sl(vec_andc(v1, qxhs0), v2); + vector signed char qxh01 = vec_sl(vec_andc(v1, vec_sr(qxhs0, (vector unsigned char)v1)), v2); + vector signed char qxh02 = vec_sl(vec_andc(v1, vec_sr(qxhs0, v2)), v2); + vector signed char qxh03 = vec_sl(vec_andc(v1, vec_sr(qxhs0, v3)), v2); + vector signed char qxh10 = vec_sl(vec_andc(v1, qxhs1), v2); + vector signed char qxh11 = vec_sl(vec_andc(v1, vec_sr(qxhs1, (vector unsigned char)v1)), v2); + vector signed char qxh12 = vec_sl(vec_andc(v1, vec_sr(qxhs1, v2)), v2); + vector signed char qxh13 = vec_sl(vec_andc(v1, vec_sr(qxhs1, v3)), v2); + qxhs0 = vec_sr(qxhs0, v4); + qxhs1 = vec_sr(qxhs1, v4); + + vector signed char q3x00 = vec_sub(qxs00, qxh00); + vector signed char q3x01 = vec_sub(qxs01, qxh01); + vector signed char q3x02 = vec_sub(qxs02, qxh02); + vector signed char q3x03 = vec_sub(qxs03, qxh03); + vector signed char q3x10 = vec_sub(qxs10, qxh10); + vector signed char q3x11 = vec_sub(qxs11, qxh11); + vector signed char q3x12 = vec_sub(qxs12, qxh12); + vector signed char q3x13 = vec_sub(qxs13, qxh13); + + vector signed char q8y00 = vec_xl( 0, q8); + vector signed char q8y10 = vec_xl( 16, q8); + vector signed char q8y01 = vec_xl( 32, q8); + vector signed char q8y11 = vec_xl( 48, q8); + vector signed char q8y02 = vec_xl( 64, q8); + vector signed char q8y12 = vec_xl( 80, q8); + vector signed char q8y03 = vec_xl( 96, q8); + vector signed char q8y13 = vec_xl(112, q8); + q8 += 128; + + vector signed short vscales_h = vec_unpackh(vscales); + vector signed short vs0 = vec_splat(vscales_h, 0); + vector signed short vs1 = vec_splat(vscales_h, 1); + vector signed short vs2 = vec_splat(vscales_h, 2); + vector signed short vs3 = vec_splat(vscales_h, 3); + vector signed short vs4 = vec_splat(vscales_h, 4); + vector signed short vs5 = vec_splat(vscales_h, 5); + vector signed short vs6 = vec_splat(vscales_h, 6); + vector signed short vs7 = vec_splat(vscales_h, 7); + vscales = vec_sld(vscales, vscales, 8); + + vector signed short qv00 = vec_add(vec_mule(q3x00, q8y00), vec_mulo(q3x00, q8y00)); + vector signed short qv01 = vec_add(vec_mule(q3x01, q8y01), vec_mulo(q3x01, q8y01)); + vector signed short qv02 = vec_add(vec_mule(q3x02, q8y02), vec_mulo(q3x02, q8y02)); + vector signed short qv03 = vec_add(vec_mule(q3x03, q8y03), vec_mulo(q3x03, q8y03)); + vector signed short qv10 = vec_add(vec_mule(q3x10, q8y10), vec_mulo(q3x10, q8y10)); + vector signed short qv11 = vec_add(vec_mule(q3x11, q8y11), vec_mulo(q3x11, q8y11)); + vector signed short qv12 = vec_add(vec_mule(q3x12, q8y12), vec_mulo(q3x12, q8y12)); + vector signed short qv13 = vec_add(vec_mule(q3x13, q8y13), vec_mulo(q3x13, q8y13)); + + vector signed int vsum0 = vec_add(vec_mule(qv00, vs0), vec_mulo(qv00, vs0)); + vector signed int vsum1 = vec_add(vec_mule(qv01, vs2), vec_mulo(qv01, vs2)); + vector signed int vsum2 = vec_add(vec_mule(qv02, vs4), vec_mulo(qv02, vs4)); + vector signed int vsum3 = vec_add(vec_mule(qv03, vs6), vec_mulo(qv03, vs6)); + vector signed int vsum4 = vec_add(vec_mule(qv10, vs1), vec_mulo(qv10, vs1)); + vector signed int vsum5 = vec_add(vec_mule(qv11, vs3), vec_mulo(qv11, vs3)); + vector signed int vsum6 = vec_add(vec_mule(qv12, vs5), vec_mulo(qv12, vs5)); + vector signed int vsum7 = vec_add(vec_mule(qv13, vs7), vec_mulo(qv13, vs7)); + + vsumi0 = vec_add(vsum0, vsumi0); + vsumi1 = vec_add(vsum1, vsumi1); + vsumi2 = vec_add(vsum2, vsumi2); + vsumi3 = vec_add(vsum3, vsumi3); + vsumi4 = vec_add(vsum4, vsumi4); + vsumi5 = vec_add(vsum5, vsumi5); + vsumi6 = vec_add(vsum6, vsumi6); + vsumi7 = vec_add(vsum7, vsumi7); + } + + vsumi0 = vec_add(vsumi0, vsumi4); + vsumi1 = vec_add(vsumi1, vsumi5); + vsumi2 = vec_add(vsumi2, vsumi6); + vsumi3 = vec_add(vsumi3, vsumi7); + + vsumf0 = vec_madd(vec_ctf(vsumi0, 0), vd, vsumf0); + vsumf1 = vec_madd(vec_ctf(vsumi1, 0), vd, vsumf1); + vsumf2 = vec_madd(vec_ctf(vsumi2, 0), vd, vsumf2); + vsumf3 = vec_madd(vec_ctf(vsumi3, 0), vd, vsumf3); + } + + vsumf0 = vec_add(vsumf0, vsumf2); + vsumf1 = vec_add(vsumf1, vsumf3); + + vsumf0 = vec_add(vsumf0, vsumf1); + + vsumf0 = vec_add(vsumf0, vec_sld(vsumf0, vsumf0, 4)); + vsumf0 = vec_add(vsumf0, vec_sld(vsumf0, vsumf0, 8)); + + *s = vec_extract(vsumf0, 0); #else // scalar version // This function is written like this so the compiler can manage to vectorize most of it @@ -6207,6 +6864,95 @@ void ggml_vec_dot_q3_K_q8_K(int n, float * restrict s, size_t bs, const void * r *s = sumf; +#elif defined(__POWER9_VECTOR__) + const vector signed char lowMask = vec_splats((signed char)0x3); + const vector signed char v1 = vec_splats((signed char)0x1); + const vector unsigned char v2 = vec_splats((unsigned char)0x2); + const vector unsigned char v4 = vec_splats((unsigned char)0x4); + const vector unsigned char v6 = vec_splats((unsigned char)0x6); + const vector signed char off = vec_splats((signed char)0x8); + + vector float vsumf0 = vec_splats(0.0f); + vector float vsumf1 = vec_splats(0.0f); + vector float vsumf2 = vec_splats(0.0f); + vector float vsumf3 = vec_splats(0.0f); + +#pragma GCC unroll 2 + for (int i = 0; i < nb; ++i) { + __builtin_prefetch(x[i].qs, 0, 1); + __builtin_prefetch(y[i].qs, 0, 1); + + vector float vxd = vec_splats(GGML_FP16_TO_FP32(x[i].d)); + vector float vyd = vec_splats(y[i].d); + vector float vd = vec_mul(vxd, vyd); + + uint16_t aux16[2]; + int8_t * scales = (int8_t *)aux16; + + const uint16_t a = *(const uint16_t *)x[i].scales; + aux16[0] = a & 0x0f0f; + aux16[1] = (a >> 4) & 0x0f0f; + + vector signed char vscales = (vector signed char)vec_xl_len(scales, 8); + vector signed char qxhs0 = (vector signed char)vec_xl_len(x[i].hmask, 8); + qxhs0 = vec_or(qxhs0, vec_sr(vec_sld(qxhs0, qxhs0, 8), (vector unsigned char)v1)); + + vscales = vec_sub(vscales, off); + + vector signed char qxs0 = (vector signed char)vec_xl( 0, x[i].qs); + vector signed char qxs00 = vec_and(qxs0, lowMask); + vector signed char qxs01 = vec_and(vec_sr(qxs0, v2), lowMask); + vector signed char qxs10 = vec_and(vec_sr(qxs0, v4), lowMask); + vector signed char qxs11 = vec_and(vec_sr(qxs0, v6), lowMask); + + //the 3rd bit + vector signed char qxh00 = vec_sl(vec_andc(v1, qxhs0), v2); + vector signed char qxh01 = vec_sl(vec_andc(v1, vec_sr(qxhs0, v2)), v2); + vector signed char qxh02 = vec_sl(vec_andc(v1, vec_sr(qxhs0, v4)), v2); + vector signed char qxh03 = vec_sl(vec_andc(v1, vec_sr(qxhs0, v6)), v2); + qxhs0 = vec_sr(qxhs0, v4); + + vector signed char q3x00 = vec_sub(qxs00, qxh00); + vector signed char q3x01 = vec_sub(qxs01, qxh01); + vector signed char q3x10 = vec_sub(qxs10, qxh02); + vector signed char q3x11 = vec_sub(qxs11, qxh03); + + vector signed char q8y00 = vec_xl( 0, y[i].qs); + vector signed char q8y01 = vec_xl( 16, y[i].qs); + vector signed char q8y10 = vec_xl( 32, y[i].qs); + vector signed char q8y11 = vec_xl( 48, y[i].qs); + + vector signed short vscales_h = vec_unpackh(vscales); + vector signed short vs0 = vec_splat(vscales_h, 0); + vector signed short vs1 = vec_splat(vscales_h, 1); + vector signed short vs2 = vec_splat(vscales_h, 2); + vector signed short vs3 = vec_splat(vscales_h, 3); + + vector signed short qv00 = vec_add(vec_mule(q3x00, q8y00), vec_mulo(q3x00, q8y00)); + vector signed short qv10 = vec_add(vec_mule(q3x10, q8y10), vec_mulo(q3x10, q8y10)); + vector signed short qv01 = vec_add(vec_mule(q3x01, q8y01), vec_mulo(q3x01, q8y01)); + vector signed short qv11 = vec_add(vec_mule(q3x11, q8y11), vec_mulo(q3x11, q8y11)); + + vector signed int vsumi0 = vec_add(vec_mule(qv00, vs0), vec_mulo(qv00, vs0)); + vector signed int vsumi1 = vec_add(vec_mule(qv10, vs1), vec_mulo(qv10, vs1)); + vector signed int vsumi2 = vec_add(vec_mule(qv01, vs2), vec_mulo(qv01, vs2)); + vector signed int vsumi3 = vec_add(vec_mule(qv11, vs3), vec_mulo(qv11, vs3)); + + vsumf0 = vec_madd(vec_ctf(vsumi0, 0), vd, vsumf0); + vsumf1 = vec_madd(vec_ctf(vsumi1, 0), vd, vsumf1); + vsumf2 = vec_madd(vec_ctf(vsumi2, 0), vd, vsumf2); + vsumf3 = vec_madd(vec_ctf(vsumi3, 0), vd, vsumf3); + } + + vsumf0 = vec_add(vsumf0, vsumf2); + vsumf1 = vec_add(vsumf1, vsumf3); + + vsumf0 = vec_add(vsumf0, vsumf1); + + vsumf0 = vec_add(vsumf0, vec_sld(vsumf0, vsumf0, 4)); + vsumf0 = vec_add(vsumf0, vec_sld(vsumf0, vsumf0, 8)); + + *s = vec_extract(vsumf0, 0); #else int8_t aux8[QK_K]; @@ -6559,6 +7305,142 @@ void ggml_vec_dot_q4_K_q8_K(int n, float * restrict s, size_t bs, const void * r *s = sumf; +#elif defined(__POWER9_VECTOR__) + const vector signed char lowMask = vec_splats((signed char)0xF); + const vector unsigned char v4 = vec_splats((unsigned char)0x4); + + vector float vsumf0 = vec_splats(0.0f); + vector float vsumf1 = vec_splats(0.0f); + vector float vsumf2 = vec_splats(0.0f); + vector float vsumf3 = vec_splats(0.0f); + + for (int i = 0; i < nb; ++i) { + vector float vxd = vec_splats(GGML_FP16_TO_FP32(x[i].d)); + vector float vyd = vec_splats(y[i].d); + vector float vd = vec_mul(vxd, vyd); + + vector float vxmin = vec_splats(GGML_FP16_TO_FP32(x[i].dmin)); + vector float vdmin = vec_mul(vxmin, vyd); + + vector signed short q8ysums0 = vec_xl( 0, y[i].bsums); + vector signed short q8ysums1 = vec_xl(16, y[i].bsums); + + memcpy(utmp, x[i].scales, 12); + + utmp[3] = ((utmp[2] >> 4) & kmask2) | (((utmp[1] >> 6) & kmask3) << 4); + const uint32_t uaux = utmp[1] & kmask1; + utmp[1] = (utmp[2] & kmask2) | (((utmp[0] >> 6) & kmask3) << 4); + utmp[2] = uaux; + utmp[0] &= kmask1; + + vector signed char utmps = (vector signed char)vec_xl( 0, utmp); + vector signed short vscales = vec_unpackh(utmps); + vector signed short q4xmins = vec_unpackl(utmps); + vector signed short q4xmins0 = vec_mergeh(q4xmins, q4xmins); + vector signed short q4xmins1 = vec_mergel(q4xmins, q4xmins); + + vector signed int prod0 = vec_mule(q4xmins0, q8ysums0); + vector signed int prod1 = vec_mule(q4xmins1, q8ysums1); + vector signed int prod2 = vec_mulo(q4xmins0, q8ysums0); + vector signed int prod3 = vec_mulo(q4xmins1, q8ysums1); + + vsumf0 = vec_nmsub(vec_ctf(prod0, 0), vdmin, vsumf0); + vsumf1 = vec_nmsub(vec_ctf(prod1, 0), vdmin, vsumf1); + vsumf2 = vec_nmsub(vec_ctf(prod2, 0), vdmin, vsumf2); + vsumf3 = vec_nmsub(vec_ctf(prod3, 0), vdmin, vsumf3); + + vector signed int vsumi0 = vec_splats((int32_t)0); + vector signed int vsumi1 = vec_splats((int32_t)0); + vector signed int vsumi2 = vec_splats((int32_t)0); + vector signed int vsumi3 = vec_splats((int32_t)0); + vector signed int vsumi4 = vec_splats((int32_t)0); + vector signed int vsumi5 = vec_splats((int32_t)0); + vector signed int vsumi6 = vec_splats((int32_t)0); + vector signed int vsumi7 = vec_splats((int32_t)0); + + const uint8_t * restrict q4 = x[i].qs; + const int8_t * restrict q8 = y[i].qs; + + for (int j = 0; j < QK_K/64; j+=2) { + __builtin_prefetch(q4, 0, 1); + __builtin_prefetch(q8, 0, 1); + + vector signed char qxs0 = (vector signed char)vec_xl( 0, q4); + vector signed char qxs1 = (vector signed char)vec_xl(16, q4); + vector signed char qxs2 = (vector signed char)vec_xl(32, q4); + vector signed char qxs3 = (vector signed char)vec_xl(48, q4); + q4 += 64; + + vector signed char q4x00 = vec_and(qxs0, lowMask); + vector signed char q4x01 = vec_sr(qxs0, v4); + vector signed char q4x10 = vec_and(qxs1, lowMask); + vector signed char q4x11 = vec_sr(qxs1, v4); + vector signed char q4x20 = vec_and(qxs2, lowMask); + vector signed char q4x21 = vec_sr(qxs2, v4); + vector signed char q4x30 = vec_and(qxs3, lowMask); + vector signed char q4x31 = vec_sr(qxs3, v4); + + vector signed char q8y00 = vec_xl( 0, q8); + vector signed char q8y10 = vec_xl( 16, q8); + vector signed char q8y01 = vec_xl( 32, q8); + vector signed char q8y11 = vec_xl( 48, q8); + vector signed char q8y20 = vec_xl( 64, q8); + vector signed char q8y30 = vec_xl( 80, q8); + vector signed char q8y21 = vec_xl( 96, q8); + vector signed char q8y31 = vec_xl(112, q8); + q8 += 128; + + vector signed short qv00 = vec_add(vec_mule(q4x00, q8y00), vec_mulo(q4x00, q8y00)); + vector signed short qv01 = vec_add(vec_mule(q4x01, q8y01), vec_mulo(q4x01, q8y01)); + vector signed short qv10 = vec_add(vec_mule(q4x10, q8y10), vec_mulo(q4x10, q8y10)); + vector signed short qv11 = vec_add(vec_mule(q4x11, q8y11), vec_mulo(q4x11, q8y11)); + vector signed short qv20 = vec_add(vec_mule(q4x20, q8y20), vec_mulo(q4x20, q8y20)); + vector signed short qv21 = vec_add(vec_mule(q4x21, q8y21), vec_mulo(q4x21, q8y21)); + vector signed short qv30 = vec_add(vec_mule(q4x30, q8y30), vec_mulo(q4x30, q8y30)); + vector signed short qv31 = vec_add(vec_mule(q4x31, q8y31), vec_mulo(q4x31, q8y31)); + + vector signed short vs0 = vec_splat(vscales, 0); + vector signed short vs1 = vec_splat(vscales, 1); + vector signed short vs2 = vec_splat(vscales, 2); + vector signed short vs3 = vec_splat(vscales, 3); + vscales = vec_sld(vscales, vscales, 8); + + qv00 = vec_add(qv00, qv10); + qv10 = vec_add(qv01, qv11); + qv20 = vec_add(qv20, qv30); + qv30 = vec_add(qv21, qv31); + + vsumi0 = vec_add(vec_mule(qv00, vs0), vsumi0); + vsumi1 = vec_add(vec_mulo(qv00, vs0), vsumi1); + vsumi2 = vec_add(vec_mule(qv10, vs1), vsumi2); + vsumi3 = vec_add(vec_mulo(qv10, vs1), vsumi3); + vsumi4 = vec_add(vec_mule(qv20, vs2), vsumi4); + vsumi5 = vec_add(vec_mulo(qv20, vs2), vsumi5); + vsumi6 = vec_add(vec_mule(qv30, vs3), vsumi6); + vsumi7 = vec_add(vec_mulo(qv30, vs3), vsumi7); + } + + vsumi0 = vec_add(vsumi0, vsumi4); + vsumi1 = vec_add(vsumi1, vsumi5); + vsumi2 = vec_add(vsumi2, vsumi6); + vsumi3 = vec_add(vsumi3, vsumi7); + + vsumf0 = vec_madd(vec_ctf(vsumi0, 0), vd, vsumf0); + vsumf1 = vec_madd(vec_ctf(vsumi1, 0), vd, vsumf1); + vsumf2 = vec_madd(vec_ctf(vsumi2, 0), vd, vsumf2); + vsumf3 = vec_madd(vec_ctf(vsumi3, 0), vd, vsumf3); + } + + vsumf0 = vec_add(vsumf0, vsumf2); + vsumf1 = vec_add(vsumf1, vsumf3); + + vsumf0 = vec_add(vsumf0, vsumf1); + + vsumf0 = vec_add(vsumf0, vec_sld(vsumf0, vsumf0, 4)); + vsumf0 = vec_add(vsumf0, vec_sld(vsumf0, vsumf0, 8)); + + *s = vec_extract(vsumf0, 0); + #else @@ -6825,6 +7707,87 @@ void ggml_vec_dot_q4_K_q8_K(int n, float * restrict s, size_t bs, const void * r *s = sumf; +#elif defined(__POWER9_VECTOR__) + const vector signed char lowMask = vec_splats((signed char)0xF); + const vector unsigned char v4 = vec_splats((unsigned char)0x4); + + vector float vsumf0 = vec_splats(0.0f); + vector float vsumf1 = vec_splats(0.0f); + vector float vsumf2 = vec_splats(0.0f); + vector float vsumf3 = vec_splats(0.0f); + +#pragma GCC unroll 2 + for (int i = 0; i < nb; ++i) { + __builtin_prefetch(x[i].qs, 0, 1); + __builtin_prefetch(y[i].qs, 0, 1); + + vector float vxd = vec_splats(GGML_FP16_TO_FP32(x[i].d[1])); + vector float vyd = vec_splats(y[i].d); + vector float vd= vec_mul(vxd, vyd); + + uint16_t s16[2]; + const uint8_t * scales = (const uint8_t *)s16; + + const uint16_t * restrict b = (const uint16_t *)x[i].scales; + s16[0] = b[0] & 0x0f0f; + s16[1] = (b[0] >> 4) & 0x0f0f; + + vector signed char utmps = (vector signed char)vec_xl_len(scales, 4); + vector signed short vscales = (vector signed short)vec_unpackh(utmps); + vector signed short q4xmins0 = vec_mergeh(vscales, vscales); + q4xmins0 = vec_sld(q4xmins0, q4xmins0, 8); + + vector signed short q8ysums0 = vec_xl_len((const int16_t *)(y[i].bsums), 8); + + vector signed int prod0 = vec_mule(q4xmins0, q8ysums0); + vector signed int prod1 = vec_mulo(q4xmins0, q8ysums0); + + vsumf0 = vec_nmsub(vec_ctf(prod0, 0), vd, vsumf0); + vsumf1 = vec_nmsub(vec_ctf(prod1, 0), vd, vsumf1); + + vd = vec_mul(vyd, vec_splats(GGML_FP16_TO_FP32(x[i].d[0]))); + + vector signed char qxs0 = (vector signed char)vec_xl( 0, x[i].qs); + vector signed char qxs1 = (vector signed char)vec_xl(16, x[i].qs); + vector signed char q4x00 = vec_and(qxs0, lowMask); + vector signed char q4x01 = vec_sr(qxs0, v4); + vector signed char q4x10 = vec_and(qxs1, lowMask); + vector signed char q4x11 = vec_sr(qxs1, v4); + + vector signed char q8y00 = vec_xl( 0, y[i].qs); + vector signed char q8y10 = vec_xl(16, y[i].qs); + vector signed char q8y01 = vec_xl(32, y[i].qs); + vector signed char q8y11 = vec_xl(48, y[i].qs); + + vector signed short qv00 = vec_add(vec_mule(q4x00, q8y00), vec_mulo(q4x00, q8y00)); + vector signed short qv01 = vec_add(vec_mule(q4x01, q8y01), vec_mulo(q4x01, q8y01)); + vector signed short qv10 = vec_add(vec_mule(q4x10, q8y10), vec_mulo(q4x10, q8y10)); + vector signed short qv11 = vec_add(vec_mule(q4x11, q8y11), vec_mulo(q4x11, q8y11)); + + vector signed short vs0 = vec_splat(vscales, 0); + vector signed short vs1 = vec_splat(vscales, 1); + + vector signed int vsumi0 = vec_add(vec_mule(qv00, vs0), vec_mulo(qv00, vs0)); + vector signed int vsumi1 = vec_add(vec_mule(qv10, vs0), vec_mulo(qv10, vs0)); + vector signed int vsumi2 = vec_add(vec_mule(qv01, vs1), vec_mulo(qv01, vs1)); + vector signed int vsumi3 = vec_add(vec_mule(qv11, vs1), vec_mulo(qv11, vs1)); + + vsumf0 = vec_madd(vec_ctf(vsumi0, 0), vd, vsumf0); + vsumf1 = vec_madd(vec_ctf(vsumi1, 0), vd, vsumf1); + vsumf2 = vec_madd(vec_ctf(vsumi2, 0), vd, vsumf2); + vsumf3 = vec_madd(vec_ctf(vsumi3, 0), vd, vsumf3); + } + + vsumf0 = vec_add(vsumf0, vsumf2); + vsumf1 = vec_add(vsumf1, vsumf3); + + vsumf0 = vec_add(vsumf0, vsumf1); + + vsumf0 = vec_add(vsumf0, vec_sld(vsumf0, vsumf0, 4)); + vsumf0 = vec_add(vsumf0, vec_sld(vsumf0, vsumf0, 8)); + + *s = vec_extract(vsumf0, 0); + #else uint8_t aux8[QK_K]; @@ -7226,6 +8189,130 @@ void ggml_vec_dot_q5_K_q8_K(int n, float * restrict s, size_t bs, const void * r *s = sumf+sums; +#elif defined(__POWER9_VECTOR__) + const vector signed char lowMask = vec_splats((signed char)0xF); + const vector unsigned char v1 = vec_splats((unsigned char)0x1); + const vector unsigned char v2 = vec_splats((unsigned char)0x2); + const vector unsigned char v3 = vec_splats((unsigned char)0x3); + const vector unsigned char v4 = vec_splats((unsigned char)0x4); + + vector float vsumf0 = vec_splats(0.0f); + vector float vsumf1 = vec_splats(0.0f); + vector float vsumf2 = vec_splats(0.0f); + vector float vsumf3 = vec_splats(0.0f); + + for (int i = 0; i < nb; ++i) { + vector float vxd = vec_splats(GGML_FP16_TO_FP32(x[i].d)); + vector float vyd = vec_splats(y[i].d); + vector float vd = vec_mul(vxd, vyd); + + vector float vxmin = vec_splats(GGML_FP16_TO_FP32(x[i].dmin)); + vector float vdmin = vec_mul(vxmin, vyd); + + memcpy(utmp, x[i].scales, 12); + + utmp[3] = ((utmp[2] >> 4) & kmask2) | (((utmp[1] >> 6) & kmask3) << 4); + const uint32_t uaux = utmp[1] & kmask1; + utmp[1] = (utmp[2] & kmask2) | (((utmp[0] >> 6) & kmask3) << 4); + utmp[2] = uaux; + utmp[0] &= kmask1; + + vector signed short q8ysums0 = vec_xl( 0, y[i].bsums); + vector signed short q8ysums1 = vec_xl(16, y[i].bsums); + + vector signed char utmps = (vector signed char)vec_xl( 0, utmp); + vector signed short vscales = vec_unpackh(utmps); + + vector signed short q5xmins = vec_unpackl(utmps); + vector signed short q5xmins0 = vec_mergeh(q5xmins, q5xmins); + vector signed short q5xmins1 = vec_mergel(q5xmins, q5xmins); + + vector signed int prod0 = vec_mule(q5xmins0, q8ysums0); + vector signed int prod1 = vec_mule(q5xmins1, q8ysums1); + vector signed int prod2 = vec_mulo(q5xmins0, q8ysums0); + vector signed int prod3 = vec_mulo(q5xmins1, q8ysums1); + + vsumf0 = vec_nmsub(vec_ctf(prod0, 0), vdmin, vsumf0); + vsumf1 = vec_nmsub(vec_ctf(prod1, 0), vdmin, vsumf1); + vsumf2 = vec_nmsub(vec_ctf(prod2, 0), vdmin, vsumf2); + vsumf3 = vec_nmsub(vec_ctf(prod3, 0), vdmin, vsumf3); + + vector signed char qxhs0 = (vector signed char)vec_xl( 0, x[i].qh); + vector signed char qxhs1 = (vector signed char)vec_xl(16, x[i].qh); + + vector signed int vsumi0 = vec_splats((int32_t)0); + vector signed int vsumi1 = vec_splats((int32_t)0); + vector signed int vsumi2 = vec_splats((int32_t)0); + vector signed int vsumi3 = vec_splats((int32_t)0); + + const uint8_t * restrict q5 = x[i].qs; + const int8_t * restrict q8 = y[i].qs; + + for (int j = 0; j < QK_K/64; ++j) { + __builtin_prefetch(q5, 0, 1); + __builtin_prefetch(q8, 0, 1); + + vector signed char qxs0 = (vector signed char)vec_xl( 0, q5); + vector signed char qxs1 = (vector signed char)vec_xl(16, q5); + q5 += 32; + + vector signed char qxs00 = vec_and(qxs0, lowMask); + vector signed char qxs01 = vec_sr(qxs0, v4); + vector signed char qxs10 = vec_and(qxs1, lowMask); + vector signed char qxs11 = vec_sr(qxs1, v4); + + vector signed char q5h00 = vec_sl(vec_and((vector signed char)v1, qxhs0), v4); + vector signed char q5h01 = vec_sl(vec_and((vector signed char)v2, qxhs0), v3); + vector signed char q5h10 = vec_sl(vec_and((vector signed char)v1, qxhs1), v4); + vector signed char q5h11 = vec_sl(vec_and((vector signed char)v2, qxhs1), v3); + qxhs0 = vec_sr(qxhs0, v2); + qxhs1 = vec_sr(qxhs1, v2); + + vector signed char q5x00 = vec_or(q5h00, qxs00); + vector signed char q5x01 = vec_or(q5h01, qxs01); + vector signed char q5x10 = vec_or(q5h10, qxs10); + vector signed char q5x11 = vec_or(q5h11, qxs11); + + vector signed char q8y00 = vec_xl( 0, q8); + vector signed char q8y10 = vec_xl(16, q8); + vector signed char q8y01 = vec_xl(32, q8); + vector signed char q8y11 = vec_xl(48, q8); + q8 += 64; + + vector signed short qv00 = vec_add(vec_mule(q5x00, q8y00), vec_mulo(q5x00, q8y00)); + vector signed short qv01 = vec_add(vec_mule(q5x01, q8y01), vec_mulo(q5x01, q8y01)); + vector signed short qv10 = vec_add(vec_mule(q5x10, q8y10), vec_mulo(q5x10, q8y10)); + vector signed short qv11 = vec_add(vec_mule(q5x11, q8y11), vec_mulo(q5x11, q8y11)); + + vector signed short vs0 = vec_splat(vscales, 0); + vector signed short vs1 = vec_splat(vscales, 1); + vscales = vec_sld(vscales, vscales, 12); + + qv00 = vec_add(qv00, qv10); + qv01 = vec_add(qv01, qv11); + + vsumi0 = vec_add(vec_mule(qv00, vs0), vsumi0); + vsumi1 = vec_add(vec_mulo(qv00, vs0), vsumi1); + vsumi2 = vec_add(vec_mule(qv01, vs1), vsumi2); + vsumi3 = vec_add(vec_mulo(qv01, vs1), vsumi3); + } + + vsumf0 = vec_madd(vec_ctf(vsumi0, 0), vd, vsumf0); + vsumf1 = vec_madd(vec_ctf(vsumi1, 0), vd, vsumf1); + vsumf2 = vec_madd(vec_ctf(vsumi2, 0), vd, vsumf2); + vsumf3 = vec_madd(vec_ctf(vsumi3, 0), vd, vsumf3); + } + + vsumf0 = vec_add(vsumf0, vsumf2); + vsumf1 = vec_add(vsumf1, vsumf3); + + vsumf0 = vec_add(vsumf0, vsumf1); + + vsumf0 = vec_add(vsumf0, vec_sld(vsumf0, vsumf0, 4)); + vsumf0 = vec_add(vsumf0, vec_sld(vsumf0, vsumf0, 8)); + + *s = vec_extract(vsumf0, 0); + #else const uint8_t * scales = (const uint8_t*)&utmp[0]; @@ -7523,6 +8610,83 @@ void ggml_vec_dot_q5_K_q8_K(int n, float * restrict s, size_t bs, const void * r *s = sumf; +#elif defined(__POWER9_VECTOR__) + const vector signed char lowMask = vec_splats((signed char)0xF); + const vector unsigned char v1 = vec_splats((unsigned char)0x1); + const vector unsigned char v2 = vec_splats((unsigned char)0x2); + const vector unsigned char v4 = vec_splats((unsigned char)0x4); + + vector float vsumf0 = vec_splats(0.0f); + vector float vsumf1 = vec_splats(0.0f); + vector float vsumf2 = vec_splats(0.0f); + vector float vsumf3 = vec_splats(0.0f); + +#pragma GCC unroll 2 + for (int i = 0; i < nb; ++i) { + __builtin_prefetch(x[i].qs, 0, 1); + __builtin_prefetch(y[i].qs, 0, 1); + + vector float vxd = vec_splats(GGML_FP16_TO_FP32(x[i].d)); + vector float vyd = vec_splats(y[i].d); + vector float vd= vec_mul(vxd, vyd); + + vector signed char qxs0 = (vector signed char)vec_xl( 0, x[i].qs); + vector signed char qxs1 = (vector signed char)vec_xl(16, x[i].qs); + vector signed char qxs00 = (vector signed char)vec_and(qxs0, lowMask); + vector signed char qxs01 = (vector signed char)vec_sr(qxs0, v4); + vector signed char qxs10 = (vector signed char)vec_and(qxs1, lowMask); + vector signed char qxs11 = (vector signed char)vec_sr(qxs1, v4); + + vector signed char qxhs = (vector signed char)vec_xl_len(x[i].qh, 8); + vector signed char qxhs0 = vec_or(qxhs, vec_sr(vec_sld(qxhs, qxhs, 8), v1)); + vector signed char qxhs1 = vec_sr(qxhs0, v2); + vector signed char qxh00 = vec_sl(vec_andc((vector signed char)v1, qxhs0), v4); + vector signed char qxh10 = vec_sl(vec_andc((vector signed char)v1, qxhs1), v4); + vector signed char qxh01 = vec_sl(vec_andc((vector signed char)v1, vec_sr(qxhs0, v4)), v4); + vector signed char qxh11 = vec_sl(vec_andc((vector signed char)v1, vec_sr(qxhs1, v4)), v4); + + vector signed char q5x00 = vec_sub(qxs00, qxh00); + vector signed char q5x10 = vec_sub(qxs10, qxh10); + vector signed char q5x01 = vec_sub(qxs01, qxh01); + vector signed char q5x11 = vec_sub(qxs11, qxh11); + + vector signed char q8y00 = vec_xl( 0, y[i].qs); + vector signed char q8y10 = vec_xl(16, y[i].qs); + vector signed char q8y01 = vec_xl(32, y[i].qs); + vector signed char q8y11 = vec_xl(48, y[i].qs); + + vector signed short qv00 = vec_add(vec_mule(q5x00, q8y00), vec_mulo(q5x00, q8y00)); + vector signed short qv01 = vec_add(vec_mule(q5x01, q8y01), vec_mulo(q5x01, q8y01)); + vector signed short qv10 = vec_add(vec_mule(q5x10, q8y10), vec_mulo(q5x10, q8y10)); + vector signed short qv11 = vec_add(vec_mule(q5x11, q8y11), vec_mulo(q5x11, q8y11)); + + vector signed short vs = (vector signed short)vec_unpackh(vec_xl_len(x[i].scales, 4)); + vector signed short vs0 = vec_splat(vs, 0); + vector signed short vs1 = vec_splat(vs, 1); + vector signed short vs2 = vec_splat(vs, 2); + vector signed short vs3 = vec_splat(vs, 3); + + vector signed int vsumi0 = vec_add(vec_mule(qv00, vs0), vec_mulo(qv00, vs0)); + vector signed int vsumi1 = vec_add(vec_mule(qv10, vs1), vec_mulo(qv10, vs1)); + vector signed int vsumi2 = vec_add(vec_mule(qv01, vs2), vec_mulo(qv01, vs2)); + vector signed int vsumi3 = vec_add(vec_mule(qv11, vs3), vec_mulo(qv11, vs3)); + + vsumf0 = vec_madd(vec_ctf(vsumi0, 0), vd, vsumf0); + vsumf1 = vec_madd(vec_ctf(vsumi1, 0), vd, vsumf1); + vsumf2 = vec_madd(vec_ctf(vsumi2, 0), vd, vsumf2); + vsumf3 = vec_madd(vec_ctf(vsumi3, 0), vd, vsumf3); + } + + vsumf0 = vec_add(vsumf0, vsumf2); + vsumf1 = vec_add(vsumf1, vsumf3); + + vsumf0 = vec_add(vsumf0, vsumf1); + + vsumf0 = vec_add(vsumf0, vec_sld(vsumf0, vsumf0, 4)); + vsumf0 = vec_add(vsumf0, vec_sld(vsumf0, vsumf0, 8)); + + *s = vec_extract(vsumf0, 0); + #else int8_t aux8[QK_K]; @@ -7953,6 +9117,151 @@ void ggml_vec_dot_q6_K_q8_K(int n, float * restrict s, size_t bs, const void * r *s = sumf; +#elif defined(__POWER9_VECTOR__) + const vector signed char lowMask = vec_splats((signed char)0xF); + const vector unsigned char v2 = vec_splats((unsigned char)0x2); + const vector unsigned char v3 = vec_splats((unsigned char)0x3); + const vector unsigned char v4 = vec_splats((unsigned char)0x4); + const vector unsigned char v6 = vec_splats((unsigned char)0x6); + const vector signed char off = vec_splats((signed char)0x20); + + vector float vsumf0 = vec_splats(0.0f); + vector float vsumf1 = vec_splats(0.0f); + vector float vsumf2 = vec_splats(0.0f); + vector float vsumf3 = vec_splats(0.0f); + + for (int i = 0; i < nb; ++i) { + vector float vxd = vec_splats(GGML_FP16_TO_FP32(x[i].d)); + vector float vyd = vec_splats(y[i].d); + vector float vd = vec_mul(vxd, vyd); + + vector signed int vsumi0 = vec_splats((int32_t)0); + vector signed int vsumi1 = vec_splats((int32_t)0); + vector signed int vsumi2 = vec_splats((int32_t)0); + vector signed int vsumi3 = vec_splats((int32_t)0); + vector signed int vsumi4 = vec_splats((int32_t)0); + vector signed int vsumi5 = vec_splats((int32_t)0); + vector signed int vsumi6 = vec_splats((int32_t)0); + vector signed int vsumi7 = vec_splats((int32_t)0); + + const uint8_t * restrict q6 = x[i].ql; + const uint8_t * restrict qh = x[i].qh; + const int8_t * restrict qs = x[i].scales; + const int8_t * restrict q8 = y[i].qs; + + for (int j = 0; j < QK_K/128; ++j) { + __builtin_prefetch(q6, 0, 0); + __builtin_prefetch(qh, 0, 0); + __builtin_prefetch(q8, 0, 0); + + vector signed char qxs0 = (vector signed char)vec_xl( 0, q6); + vector signed char qxs1 = (vector signed char)vec_xl(16, q6); + vector signed char qxs2 = (vector signed char)vec_xl(32, q6); + vector signed char qxs3 = (vector signed char)vec_xl(48, q6); + q6 += 64; + + vector signed char qxs00 = vec_and(qxs0, lowMask); + vector signed char qxs01 = vec_sr(qxs0, v4); + vector signed char qxs10 = vec_and(qxs1, lowMask); + vector signed char qxs11 = vec_sr(qxs1, v4); + vector signed char qxs20 = vec_and(qxs2, lowMask); + vector signed char qxs21 = vec_sr(qxs2, v4); + vector signed char qxs30 = vec_and(qxs3, lowMask); + vector signed char qxs31 = vec_sr(qxs3, v4); + + vector signed char qxhs0 = (vector signed char)vec_xl( 0, qh); + vector signed char qxhs1 = (vector signed char)vec_xl(16, qh); + qh += 32; + + vector signed char qxh00 = vec_sl(vec_and((vector signed char)v3, qxhs0), v4); + vector signed char qxh01 = vec_sl(vec_and((vector signed char)v3, vec_sr(qxhs0, v4)), v4); + vector signed char qxh10 = vec_sl(vec_and((vector signed char)v3, qxhs1), v4); + vector signed char qxh11 = vec_sl(vec_and((vector signed char)v3, vec_sr(qxhs1, v4)), v4); + vector signed char qxh20 = vec_sl(vec_and((vector signed char)v3, vec_sr(qxhs0, v2)), v4); + vector signed char qxh21 = vec_sl(vec_and((vector signed char)v3, vec_sr(qxhs0, v6)), v4); + vector signed char qxh30 = vec_sl(vec_and((vector signed char)v3, vec_sr(qxhs1, v2)), v4); + vector signed char qxh31 = vec_sl(vec_and((vector signed char)v3, vec_sr(qxhs1, v6)), v4); + + vector signed char q6x00 = vec_sub(vec_or(qxh00, qxs00), off); + vector signed char q6x01 = vec_sub(vec_or(qxh01, qxs01), off); + vector signed char q6x10 = vec_sub(vec_or(qxh10, qxs10), off); + vector signed char q6x11 = vec_sub(vec_or(qxh11, qxs11), off); + vector signed char q6x20 = vec_sub(vec_or(qxh20, qxs20), off); + vector signed char q6x21 = vec_sub(vec_or(qxh21, qxs21), off); + vector signed char q6x30 = vec_sub(vec_or(qxh30, qxs30), off); + vector signed char q6x31 = vec_sub(vec_or(qxh31, qxs31), off); + + vector signed char q8y00 = vec_xl( 0, q8); + vector signed char q8y10 = vec_xl( 16, q8); + vector signed char q8y20 = vec_xl( 32, q8); + vector signed char q8y30 = vec_xl( 48, q8); + vector signed char q8y01 = vec_xl( 64, q8); + vector signed char q8y11 = vec_xl( 80, q8); + vector signed char q8y21 = vec_xl( 96, q8); + vector signed char q8y31 = vec_xl(112, q8); + q8 += 128; + + vector signed short qv00 = vec_add(vec_mule(q6x00, q8y00), vec_mulo(q6x00, q8y00)); + vector signed short qv10 = vec_add(vec_mule(q6x10, q8y10), vec_mulo(q6x10, q8y10)); + vector signed short qv20 = vec_add(vec_mule(q6x20, q8y20), vec_mulo(q6x20, q8y20)); + vector signed short qv30 = vec_add(vec_mule(q6x30, q8y30), vec_mulo(q6x30, q8y30)); + vector signed short qv01 = vec_add(vec_mule(q6x01, q8y01), vec_mulo(q6x01, q8y01)); + vector signed short qv11 = vec_add(vec_mule(q6x11, q8y11), vec_mulo(q6x11, q8y11)); + vector signed short qv21 = vec_add(vec_mule(q6x21, q8y21), vec_mulo(q6x21, q8y21)); + vector signed short qv31 = vec_add(vec_mule(q6x31, q8y31), vec_mulo(q6x31, q8y31)); + + vector signed short vscales = vec_unpackh(vec_xl_len(qs, 8)); + qs += 8; + + vector signed short vs0 = vec_splat(vscales, 0); + vector signed short vs1 = vec_splat(vscales, 1); + vector signed short vs2 = vec_splat(vscales, 2); + vector signed short vs3 = vec_splat(vscales, 3); + vector signed short vs4 = vec_splat(vscales, 4); + vector signed short vs5 = vec_splat(vscales, 5); + vector signed short vs6 = vec_splat(vscales, 6); + vector signed short vs7 = vec_splat(vscales, 7); + + vsumi0 = vec_add(vec_mule(qv00, vs0), vsumi0); + vsumi1 = vec_add(vec_mulo(qv00, vs0), vsumi1); + vsumi2 = vec_add(vec_mule(qv01, vs4), vsumi2); + vsumi3 = vec_add(vec_mulo(qv01, vs4), vsumi3); + vsumi4 = vec_add(vec_mule(qv10, vs1), vsumi4); + vsumi5 = vec_add(vec_mulo(qv10, vs1), vsumi5); + vsumi6 = vec_add(vec_mule(qv11, vs5), vsumi6); + vsumi7 = vec_add(vec_mulo(qv11, vs5), vsumi7); + + vsumi0 = vec_add(vec_mule(qv20, vs2), vsumi0); + vsumi1 = vec_add(vec_mulo(qv20, vs2), vsumi1); + vsumi2 = vec_add(vec_mule(qv21, vs6), vsumi2); + vsumi3 = vec_add(vec_mulo(qv21, vs6), vsumi3); + vsumi4 = vec_add(vec_mule(qv30, vs3), vsumi4); + vsumi5 = vec_add(vec_mulo(qv30, vs3), vsumi5); + vsumi6 = vec_add(vec_mule(qv31, vs7), vsumi6); + vsumi7 = vec_add(vec_mulo(qv31, vs7), vsumi7); + } + + vsumi0 = vec_add(vsumi0, vsumi4); + vsumi1 = vec_add(vsumi1, vsumi5); + vsumi2 = vec_add(vsumi2, vsumi6); + vsumi3 = vec_add(vsumi3, vsumi7); + + vsumf0 = vec_madd(vec_ctf(vsumi0, 0), vd, vsumf0); + vsumf1 = vec_madd(vec_ctf(vsumi1, 0), vd, vsumf1); + vsumf2 = vec_madd(vec_ctf(vsumi2, 0), vd, vsumf2); + vsumf3 = vec_madd(vec_ctf(vsumi3, 0), vd, vsumf3); + } + + vsumf0 = vec_add(vsumf0, vsumf2); + vsumf1 = vec_add(vsumf1, vsumf3); + + vsumf0 = vec_add(vsumf0, vsumf1); + + vsumf0 = vec_add(vsumf0, vec_sld(vsumf0, vsumf0, 4)); + vsumf0 = vec_add(vsumf0, vec_sld(vsumf0, vsumf0, 8)); + + *s = vec_extract(vsumf0, 0); + #else int8_t aux8[QK_K]; @@ -8259,6 +9568,85 @@ void ggml_vec_dot_q6_K_q8_K(int n, float * restrict s, size_t bs, const void * r *s = sumf; +#elif defined(__POWER9_VECTOR__) + const vector signed char lowMask = vec_splats((signed char)0xF); + const vector unsigned char v2 = vec_splats((unsigned char)0x2); + const vector unsigned char v3 = vec_splats((unsigned char)0x3); + const vector unsigned char v4 = vec_splats((unsigned char)0x4); + const vector unsigned char v6 = vec_splats((unsigned char)0x6); + const vector signed char off = vec_splats((signed char)0x20); + + vector float vsumf0 = vec_splats(0.0f); + vector float vsumf1 = vec_splats(0.0f); + vector float vsumf2 = vec_splats(0.0f); + vector float vsumf3 = vec_splats(0.0f); + +#pragma GCC unroll 2 + for (int i = 0; i < nb; ++i) { + __builtin_prefetch(x[i].ql, 0, 1); + __builtin_prefetch(x[i].qh, 0, 1); + __builtin_prefetch(y[i].qs, 0, 1); + + vector float vxd = vec_splats(GGML_FP16_TO_FP32(x[i].d)); + vector float vyd = vec_splats(y[i].d); + vector float vd= vec_mul(vxd, vyd); + + vector signed char qxs0 = (vector signed char)vec_xl( 0, x[i].ql); + vector signed char qxs1 = (vector signed char)vec_xl(16, x[i].ql); + vector signed char qxs00 = vec_and(qxs0, lowMask); + vector signed char qxs01 = vec_sr(qxs0, v4); + vector signed char qxs10 = vec_and(qxs1, lowMask); + vector signed char qxs11 = vec_sr(qxs1, v4); + + vector signed char qxhs0 = (vector signed char)vec_xl( 0, x[i].qh); + + vector signed char qxh00 = vec_sl(vec_and((vector signed char)v3, qxhs0), v4); + vector signed char qxh01 = vec_sl(vec_and((vector signed char)v3, vec_sr(qxhs0, v4)), v4); + vector signed char qxh10 = vec_sl(vec_and((vector signed char)v3, vec_sr(qxhs0, v2)), v4); + vector signed char qxh11 = vec_sl(vec_and((vector signed char)v3, vec_sr(qxhs0, v6)), v4); + + vector signed char q6x00 = vec_sub(vec_or(qxh00, qxs00), off); + vector signed char q6x01 = vec_sub(vec_or(qxh01, qxs01), off); + vector signed char q6x10 = vec_sub(vec_or(qxh10, qxs10), off); + vector signed char q6x11 = vec_sub(vec_or(qxh11, qxs11), off); + + vector signed char q8y00 = vec_xl( 0, y[i].qs); + vector signed char q8y10 = vec_xl(16, y[i].qs); + vector signed char q8y01 = vec_xl(32, y[i].qs); + vector signed char q8y11 = vec_xl(48, y[i].qs); + + vector signed short qv00 = vec_add(vec_mule(q6x00, q8y00), vec_mulo(q6x00, q8y00)); + vector signed short qv10 = vec_add(vec_mule(q6x10, q8y10), vec_mulo(q6x10, q8y10)); + vector signed short qv01 = vec_add(vec_mule(q6x01, q8y01), vec_mulo(q6x01, q8y01)); + vector signed short qv11 = vec_add(vec_mule(q6x11, q8y11), vec_mulo(q6x11, q8y11)); + + vector signed short vs = (vector signed short)vec_unpackh(vec_xl_len(x[i].scales, 4)); + vector signed short vs0 = vec_splat(vs, 0); + vector signed short vs1 = vec_splat(vs, 1); + vector signed short vs2 = vec_splat(vs, 2); + vector signed short vs3 = vec_splat(vs, 3); + + vector signed int vsumi0 = vec_add(vec_mule(qv00, vs0), vec_mulo(qv00, vs0)); + vector signed int vsumi1 = vec_add(vec_mule(qv10, vs1), vec_mulo(qv10, vs1)); + vector signed int vsumi2 = vec_add(vec_mule(qv01, vs2), vec_mulo(qv01, vs2)); + vector signed int vsumi3 = vec_add(vec_mule(qv11, vs3), vec_mulo(qv11, vs3)); + + vsumf0 = vec_madd(vec_ctf(vsumi0, 0), vd, vsumf0); + vsumf1 = vec_madd(vec_ctf(vsumi1, 0), vd, vsumf1); + vsumf2 = vec_madd(vec_ctf(vsumi2, 0), vd, vsumf2); + vsumf3 = vec_madd(vec_ctf(vsumi3, 0), vd, vsumf3); + } + + vsumf0 = vec_add(vsumf0, vsumf2); + vsumf1 = vec_add(vsumf1, vsumf3); + + vsumf0 = vec_add(vsumf0, vsumf1); + + vsumf0 = vec_add(vsumf0, vec_sld(vsumf0, vsumf0, 4)); + vsumf0 = vec_add(vsumf0, vec_sld(vsumf0, vsumf0, 8)); + + *s = vec_extract(vsumf0, 0); + #else int8_t aux8[QK_K]; @@ -8300,7 +9688,7 @@ void ggml_vec_dot_q6_K_q8_K(int n, float * restrict s, size_t bs, const void * r #endif -#if defined (__AVX2__) || defined (__ARM_NEON) +#if defined (__AVX2__) || defined (__ARM_NEON) || defined (__POWER9_VECTOR__) static const int8_t keven_signs_q2xs[1024] = { 1, 1, 1, 1, 1, 1, 1, 1, -1, 1, 1, 1, 1, 1, 1, -1, 1, -1, 1, 1, 1, 1, 1, -1, -1, -1, 1, 1, 1, 1, 1, 1, 1, 1, -1, 1, 1, 1, 1, -1, -1, 1, -1, 1, 1, 1, 1, 1, 1, -1, -1, 1, 1, 1, 1, 1, -1, -1, -1, 1, 1, 1, 1, -1, @@ -8433,6 +9821,103 @@ void ggml_vec_dot_iq2_xxs_q8_K(int n, float * restrict s, size_t bs, const void *s = 0.125f * hsum_float_8(accumf); +#elif defined(__POWER9_VECTOR__) + vector float vsumf0 = vec_splats(0.0f); + vector float vsumf1 = vec_splats(0.0f); + vector float vsumf2 = vec_splats(0.0f); + vector float vsumf3 = vec_splats(0.0f); + + const uint64_t * signs64 = (const uint64_t *)keven_signs_q2xs; + + for (int i = 0; i < nb; ++i) { + vector float vxd = vec_splats(GGML_FP16_TO_FP32(x[i].d)); + vector float vyd = vec_splats(y[i].d); + vector float vd = vec_mul(vxd, vyd); + + vector signed int vsumi0 = vec_splats((int32_t)0); + vector signed int vsumi1 = vec_splats((int32_t)0); + vector signed int vsumi2 = vec_splats((int32_t)0); + vector signed int vsumi3 = vec_splats((int32_t)0); + vector signed int vsumi4 = vec_splats((int32_t)0); + vector signed int vsumi5 = vec_splats((int32_t)0); + vector signed int vsumi6 = vec_splats((int32_t)0); + vector signed int vsumi7 = vec_splats((int32_t)0); + + const uint16_t * restrict q2 = x[i].qs; + const int8_t * restrict q8 = y[i].qs; + + for (int j = 0; j < QK_K/32; j += 2) { + __builtin_prefetch(q2, 0, 1); + __builtin_prefetch(q8, 0, 1); + + uint32_t aux32[4]; + const uint8_t * aux8 = (const uint8_t *)aux32; + + memcpy(aux32, q2, 4*sizeof(uint32_t)); + q2 += 8; + + vector signed long long aux64x2_0 = {*(const int64_t *)(iq2xxs_grid + aux8[ 0]), *(const int64_t *)(iq2xxs_grid + aux8[ 1])}; + vector signed long long aux64x2_1 = {*(const int64_t *)(iq2xxs_grid + aux8[ 2]), *(const int64_t *)(iq2xxs_grid + aux8[ 3])}; + vector signed long long aux64x2_2 = {*(const int64_t *)(iq2xxs_grid + aux8[ 8]), *(const int64_t *)(iq2xxs_grid + aux8[ 9])}; + vector signed long long aux64x2_3 = {*(const int64_t *)(iq2xxs_grid + aux8[10]), *(const int64_t *)(iq2xxs_grid + aux8[11])}; + + vector signed long long vsigns0 = {*(const int64_t *)(signs64 + ((aux32[1] >> 0) & 127)), *(const int64_t *)(signs64 + ((aux32[1] >> 7) & 127))}; + vector signed long long vsigns1 = {*(const int64_t *)(signs64 + ((aux32[1] >> 14) & 127)), *(const int64_t *)(signs64 + ((aux32[1] >> 21) & 127))}; + vector signed long long vsigns2 = {*(const int64_t *)(signs64 + ((aux32[3] >> 0) & 127)), *(const int64_t *)(signs64 + ((aux32[3] >> 7) & 127))}; + vector signed long long vsigns3 = {*(const int64_t *)(signs64 + ((aux32[3] >> 14) & 127)), *(const int64_t *)(signs64 + ((aux32[3] >> 21) & 127))}; + + vector signed char q2x0 = (vector signed char)vec_mul((vector signed char)vsigns0, (vector signed char)aux64x2_0); + vector signed char q2x1 = (vector signed char)vec_mul((vector signed char)vsigns1, (vector signed char)aux64x2_1); + vector signed char q2x2 = (vector signed char)vec_mul((vector signed char)vsigns2, (vector signed char)aux64x2_2); + vector signed char q2x3 = (vector signed char)vec_mul((vector signed char)vsigns3, (vector signed char)aux64x2_3); + + vector signed char q8y0 = vec_xl( 0, q8); + vector signed char q8y1 = vec_xl(16, q8); + vector signed char q8y2 = vec_xl(32, q8); + vector signed char q8y3 = vec_xl(48, q8); + q8 += 64; + + vector signed short qv0 = vec_add(vec_mule(q2x0, q8y0), vec_mulo(q2x0, q8y0)); + vector signed short qv1 = vec_add(vec_mule(q2x1, q8y1), vec_mulo(q2x1, q8y1)); + vector signed short qv2 = vec_add(vec_mule(q2x2, q8y2), vec_mulo(q2x2, q8y2)); + vector signed short qv3 = vec_add(vec_mule(q2x3, q8y3), vec_mulo(q2x3, q8y3)); + + const uint16_t ls0 = aux32[1] >> 28; + const uint16_t ls1 = aux32[3] >> 28; + + vector signed short vscales01 = vec_splats((int16_t)(2*ls0+1)); + vector signed short vscales23 = vec_splats((int16_t)(2*ls1+1)); + + vsumi0 = vec_add(vec_mule(qv0, vscales01), vsumi0); + vsumi1 = vec_add(vec_mule(qv1, vscales01), vsumi1); + vsumi2 = vec_add(vec_mule(qv2, vscales23), vsumi2); + vsumi3 = vec_add(vec_mule(qv3, vscales23), vsumi3); + vsumi4 = vec_add(vec_mulo(qv0, vscales01), vsumi4); + vsumi5 = vec_add(vec_mulo(qv1, vscales01), vsumi5); + vsumi6 = vec_add(vec_mulo(qv2, vscales23), vsumi6); + vsumi7 = vec_add(vec_mulo(qv3, vscales23), vsumi7); + } + + vsumi0 = vec_add(vsumi0, vsumi4); + vsumi1 = vec_add(vsumi1, vsumi5); + vsumi2 = vec_add(vsumi2, vsumi6); + vsumi3 = vec_add(vsumi3, vsumi7); + + vsumf0 = vec_madd(vec_ctf(vsumi0, 0), vd, vsumf0); + vsumf1 = vec_madd(vec_ctf(vsumi1, 0), vd, vsumf1); + vsumf2 = vec_madd(vec_ctf(vsumi2, 0), vd, vsumf2); + vsumf3 = vec_madd(vec_ctf(vsumi3, 0), vd, vsumf3); + } + + vsumf0 = vec_add(vsumf0, vsumf2); + vsumf1 = vec_add(vsumf1, vsumf3); + + vsumf0 = vec_add(vsumf0, vsumf1); + + vsumf0 = vec_add(vsumf0, vec_sld(vsumf0, vsumf0, 4)); + vsumf0 = vec_add(vsumf0, vec_sld(vsumf0, vsumf0, 8)); + + *s = 0.125f * vec_extract(vsumf0, 0); #else uint32_t aux32[2]; @@ -8708,6 +10193,104 @@ void ggml_vec_dot_iq2_xs_q8_K(int n, float * restrict s, size_t bs, const void * *s = 0.125f * hsum_float_8(accumf); #endif +#elif defined(__POWER9_VECTOR__) + vector float vsumf0 = vec_splats(0.0f); + vector float vsumf1 = vec_splats(0.0f); + vector float vsumf2 = vec_splats(0.0f); + vector float vsumf3 = vec_splats(0.0f); + + const uint64_t * signs64 = (const uint64_t *)keven_signs_q2xs; + + for (int i = 0; i < nb; ++i) { + vector float vxd = vec_splats(GGML_FP16_TO_FP32(x[i].d)); + vector float vyd = vec_splats(y[i].d); + vector float vd = vec_mul(vxd, vyd); + + vector signed int vsumi0 = vec_splats((int32_t)0); + vector signed int vsumi1 = vec_splats((int32_t)0); + vector signed int vsumi2 = vec_splats((int32_t)0); + vector signed int vsumi3 = vec_splats((int32_t)0); + vector signed int vsumi4 = vec_splats((int32_t)0); + vector signed int vsumi5 = vec_splats((int32_t)0); + vector signed int vsumi6 = vec_splats((int32_t)0); + vector signed int vsumi7 = vec_splats((int32_t)0); + + const uint16_t * restrict q2 = x[i].qs; + const uint8_t * restrict sc = x[i].scales; + const int8_t * restrict q8 = y[i].qs; + + for (int j = 0; j < QK_K/64; ++j) { + __builtin_prefetch(q2, 0, 1); + __builtin_prefetch(q8, 0, 1); + + vector signed long long aux64x2_0 = {*(const int64_t *)(iq2xs_grid + (q2[0] & 511)), *(const int64_t *)(iq2xs_grid + (q2[1] & 511))}; + vector signed long long aux64x2_1 = {*(const int64_t *)(iq2xs_grid + (q2[2] & 511)), *(const int64_t *)(iq2xs_grid + (q2[3] & 511))}; + vector signed long long aux64x2_2 = {*(const int64_t *)(iq2xs_grid + (q2[4] & 511)), *(const int64_t *)(iq2xs_grid + (q2[5] & 511))}; + vector signed long long aux64x2_3 = {*(const int64_t *)(iq2xs_grid + (q2[6] & 511)), *(const int64_t *)(iq2xs_grid + (q2[7] & 511))}; + + vector signed long long vsigns0 = {*(const int64_t *)(signs64 + ((q2[0] >> 9))), *(const int64_t *)(signs64 + ((q2[1] >> 9)))}; + vector signed long long vsigns1 = {*(const int64_t *)(signs64 + ((q2[2] >> 9))), *(const int64_t *)(signs64 + ((q2[3] >> 9)))}; + vector signed long long vsigns2 = {*(const int64_t *)(signs64 + ((q2[4] >> 9))), *(const int64_t *)(signs64 + ((q2[5] >> 9)))}; + vector signed long long vsigns3 = {*(const int64_t *)(signs64 + ((q2[6] >> 9))), *(const int64_t *)(signs64 + ((q2[7] >> 9)))}; + q2 += 8; + + vector signed char q2x0 = (vector signed char)vec_mul((vector signed char)vsigns0, (vector signed char)aux64x2_0); + vector signed char q2x1 = (vector signed char)vec_mul((vector signed char)vsigns1, (vector signed char)aux64x2_1); + vector signed char q2x2 = (vector signed char)vec_mul((vector signed char)vsigns2, (vector signed char)aux64x2_2); + vector signed char q2x3 = (vector signed char)vec_mul((vector signed char)vsigns3, (vector signed char)aux64x2_3); + + vector signed char q8y0 = vec_xl( 0, q8); + vector signed char q8y1 = vec_xl(16, q8); + vector signed char q8y2 = vec_xl(32, q8); + vector signed char q8y3 = vec_xl(48, q8); + q8 += 64; + + vector signed short qv0 = vec_add(vec_mule(q2x0, q8y0), vec_mulo(q2x0, q8y0)); + vector signed short qv1 = vec_add(vec_mule(q2x1, q8y1), vec_mulo(q2x1, q8y1)); + vector signed short qv2 = vec_add(vec_mule(q2x2, q8y2), vec_mulo(q2x2, q8y2)); + vector signed short qv3 = vec_add(vec_mule(q2x3, q8y3), vec_mulo(q2x3, q8y3)); + + const uint16_t ls0 = (uint16_t)(sc[0] & 0xf); + const uint16_t ls1 = (uint16_t)(sc[0] >> 4); + const uint16_t ls2 = (uint16_t)(sc[1] & 0xf); + const uint16_t ls3 = (uint16_t)(sc[1] >> 4); + sc += 2; + + vector signed short vscales0 = vec_splats((int16_t)(2*ls0+1)); + vector signed short vscales1 = vec_splats((int16_t)(2*ls1+1)); + vector signed short vscales2 = vec_splats((int16_t)(2*ls2+1)); + vector signed short vscales3 = vec_splats((int16_t)(2*ls3+1)); + + vsumi0 = vec_add(vec_mule(qv0, vscales0), vsumi0); + vsumi1 = vec_add(vec_mule(qv1, vscales1), vsumi1); + vsumi2 = vec_add(vec_mule(qv2, vscales2), vsumi2); + vsumi3 = vec_add(vec_mule(qv3, vscales3), vsumi3); + vsumi4 = vec_add(vec_mulo(qv0, vscales0), vsumi4); + vsumi5 = vec_add(vec_mulo(qv1, vscales1), vsumi5); + vsumi6 = vec_add(vec_mulo(qv2, vscales2), vsumi6); + vsumi7 = vec_add(vec_mulo(qv3, vscales3), vsumi7); + } + + vsumi0 = vec_add(vsumi0, vsumi4); + vsumi1 = vec_add(vsumi1, vsumi5); + vsumi2 = vec_add(vsumi2, vsumi6); + vsumi3 = vec_add(vsumi3, vsumi7); + + vsumf0 = vec_madd(vec_ctf(vsumi0, 0), vd, vsumf0); + vsumf1 = vec_madd(vec_ctf(vsumi1, 0), vd, vsumf1); + vsumf2 = vec_madd(vec_ctf(vsumi2, 0), vd, vsumf2); + vsumf3 = vec_madd(vec_ctf(vsumi3, 0), vd, vsumf3); + } + + vsumf0 = vec_add(vsumf0, vsumf2); + vsumf1 = vec_add(vsumf1, vsumf3); + + vsumf0 = vec_add(vsumf0, vsumf1); + + vsumf0 = vec_add(vsumf0, vec_sld(vsumf0, vsumf0, 4)); + vsumf0 = vec_add(vsumf0, vec_sld(vsumf0, vsumf0, 8)); + + *s = 0.125f * vec_extract(vsumf0, 0); #else float sumf = 0.f; @@ -8908,6 +10491,124 @@ void ggml_vec_dot_iq2_s_q8_K(int n, float * restrict s, size_t bs, const void * *s = 0.125f * hsum_float_8(accumf); +#elif defined(__POWER9_VECTOR__) + static const uint8_t k_mask1[32] = {0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x01, 0x01, 0x01, 0x01, 0x01, 0x01, 0x01, 0x01, + 0x02, 0x02, 0x02, 0x02, 0x02, 0x02, 0x02, 0x02, 0x03, 0x03, 0x03, 0x03, 0x03, 0x03, 0x03, 0x03 + }; + + static const uint8_t k_mask2[16] = {0x01, 0x02, 0x04, 0x08, 0x10, 0x20, 0x40, 0x80, 0x01, 0x02, 0x04, 0x08, 0x10, 0x20, 0x40, 0x80,}; + + vector float vsumf0 = vec_splats(0.0f); + vector float vsumf1 = vec_splats(0.0f); + vector float vsumf2 = vec_splats(0.0f); + vector float vsumf3 = vec_splats(0.0f); + + const vector unsigned char mask0 = vec_xl( 0, k_mask1); + const vector unsigned char mask1 = vec_xl(16, k_mask1); + const vector signed char mask2 = (vector signed char)vec_xl( 0, k_mask2); + + for (int i = 0; i < nb; ++i) { + vector float vxd = vec_splats(GGML_FP16_TO_FP32(x[i].d)); + vector float vyd = vec_splats(y[i].d); + vector float vd = vec_mul(vxd, vyd); + + vector signed int vsumi0 = vec_splats((int32_t)0); + vector signed int vsumi1 = vec_splats((int32_t)0); + vector signed int vsumi2 = vec_splats((int32_t)0); + vector signed int vsumi3 = vec_splats((int32_t)0); + vector signed int vsumi4 = vec_splats((int32_t)0); + vector signed int vsumi5 = vec_splats((int32_t)0); + vector signed int vsumi6 = vec_splats((int32_t)0); + vector signed int vsumi7 = vec_splats((int32_t)0); + + const uint8_t * restrict q2 = x[i].qs; + const uint8_t * restrict qh = x[i].qh; + const uint16_t * restrict signs = (const uint16_t *)(x[i].qs + QK_K/8); + const uint8_t * restrict sc = x[i].scales; + const int8_t * restrict q8 = y[i].qs; + + for (int j = 0; j < QK_K/32; j += 2) { + __builtin_prefetch(q2, 0, 1); + __builtin_prefetch(q8, 0, 1); + + vector signed long long aux64x2_0 = {*(const int64_t *)(iq2s_grid + (q2[0] | ((qh[0] << 8) & 0x300))), *(const int64_t *)(iq2s_grid + (q2[1] | ((qh[0] << 6) & 0x300)))}; + vector signed long long aux64x2_1 = {*(const int64_t *)(iq2s_grid + (q2[2] | ((qh[0] << 4) & 0x300))), *(const int64_t *)(iq2s_grid + (q2[3] | ((qh[0] << 2) & 0x300)))}; + vector signed long long aux64x2_2 = {*(const int64_t *)(iq2s_grid + (q2[4] | ((qh[1] << 8) & 0x300))), *(const int64_t *)(iq2s_grid + (q2[5] | ((qh[1] << 6) & 0x300)))}; + vector signed long long aux64x2_3 = {*(const int64_t *)(iq2s_grid + (q2[6] | ((qh[1] << 4) & 0x300))), *(const int64_t *)(iq2s_grid + (q2[7] | ((qh[1] << 2) & 0x300)))}; + q2 += 8; + qh += 2; + + vector signed char vsigns01 = (vector signed char)vec_splats(*(const uint32_t *)&signs[0]); + vector signed char vsigns23 = (vector signed char)vec_splats(*(const uint32_t *)&signs[2]); + signs += 4; + + vector signed char vsigns0 = vec_perm(vsigns01, vsigns01, mask0); + vector signed char vsigns1 = vec_perm(vsigns01, vsigns01, mask1); + vector signed char vsigns2 = vec_perm(vsigns23, vsigns23, mask0); + vector signed char vsigns3 = vec_perm(vsigns23, vsigns23, mask1); + + vsigns0 = (vector signed char)vec_cmpeq(vec_and(vsigns0, mask2), mask2); + vsigns1 = (vector signed char)vec_cmpeq(vec_and(vsigns1, mask2), mask2); + vsigns2 = (vector signed char)vec_cmpeq(vec_and(vsigns2, mask2), mask2); + vsigns3 = (vector signed char)vec_cmpeq(vec_and(vsigns3, mask2), mask2); + + vector signed char q2x0 = vec_sub(vec_xor(vsigns0, (vector signed char)aux64x2_0), vsigns0); + vector signed char q2x1 = vec_sub(vec_xor(vsigns1, (vector signed char)aux64x2_1), vsigns1); + vector signed char q2x2 = vec_sub(vec_xor(vsigns2, (vector signed char)aux64x2_2), vsigns2); + vector signed char q2x3 = vec_sub(vec_xor(vsigns3, (vector signed char)aux64x2_3), vsigns3); + + vector signed char q8y0 = vec_xl( 0, q8); + vector signed char q8y1 = vec_xl(16, q8); + vector signed char q8y2 = vec_xl(32, q8); + vector signed char q8y3 = vec_xl(48, q8); + q8 += 64; + + vector signed short qv0 = vec_add(vec_mule(q2x0, q8y0), vec_mulo(q2x0, q8y0)); + vector signed short qv1 = vec_add(vec_mule(q2x1, q8y1), vec_mulo(q2x1, q8y1)); + vector signed short qv2 = vec_add(vec_mule(q2x2, q8y2), vec_mulo(q2x2, q8y2)); + vector signed short qv3 = vec_add(vec_mule(q2x3, q8y3), vec_mulo(q2x3, q8y3)); + + const uint16_t ls0 = (uint16_t)(sc[0] & 0xf); + const uint16_t ls1 = (uint16_t)(sc[0] >> 4); + const uint16_t ls2 = (uint16_t)(sc[1] & 0xf); + const uint16_t ls3 = (uint16_t)(sc[1] >> 4); + sc += 2; + + vector signed short vscales0 = vec_splats((int16_t)(2*ls0+1)); + vector signed short vscales1 = vec_splats((int16_t)(2*ls1+1)); + vector signed short vscales2 = vec_splats((int16_t)(2*ls2+1)); + vector signed short vscales3 = vec_splats((int16_t)(2*ls3+1)); + + vsumi0 = vec_add(vec_mule(qv0, vscales0), vsumi0); + vsumi1 = vec_add(vec_mule(qv1, vscales1), vsumi1); + vsumi2 = vec_add(vec_mule(qv2, vscales2), vsumi2); + vsumi3 = vec_add(vec_mule(qv3, vscales3), vsumi3); + vsumi4 = vec_add(vec_mulo(qv0, vscales0), vsumi4); + vsumi5 = vec_add(vec_mulo(qv1, vscales1), vsumi5); + vsumi6 = vec_add(vec_mulo(qv2, vscales2), vsumi6); + vsumi7 = vec_add(vec_mulo(qv3, vscales3), vsumi7); + } + + vsumi0 = vec_add(vsumi0, vsumi4); + vsumi1 = vec_add(vsumi1, vsumi5); + vsumi2 = vec_add(vsumi2, vsumi6); + vsumi3 = vec_add(vsumi3, vsumi7); + + vsumf0 = vec_madd(vec_ctf(vsumi0, 0), vd, vsumf0); + vsumf1 = vec_madd(vec_ctf(vsumi1, 0), vd, vsumf1); + vsumf2 = vec_madd(vec_ctf(vsumi2, 0), vd, vsumf2); + vsumf3 = vec_madd(vec_ctf(vsumi3, 0), vd, vsumf3); + } + + vsumf0 = vec_add(vsumf0, vsumf2); + vsumf1 = vec_add(vsumf1, vsumf3); + + vsumf0 = vec_add(vsumf0, vsumf1); + + vsumf0 = vec_add(vsumf0, vec_sld(vsumf0, vsumf0, 4)); + vsumf0 = vec_add(vsumf0, vec_sld(vsumf0, vsumf0, 8)); + + *s = 0.125f * vec_extract(vsumf0, 0); #else float sumf = 0; @@ -9052,6 +10753,101 @@ void ggml_vec_dot_iq3_xxs_q8_K(int n, float * restrict s, size_t bs, const void *s = 0.25f * hsum_float_8(accumf); +#elif defined(__POWER9_VECTOR__) + const uint64_t * signs64 = (const uint64_t *)keven_signs_q2xs; + + vector float vsumf0 = vec_splats(0.0f); + vector float vsumf1 = vec_splats(0.0f); + vector float vsumf2 = vec_splats(0.0f); + vector float vsumf3 = vec_splats(0.0f); + + for (int i = 0; i < nb; ++i) { + vector float vxd = vec_splats(GGML_FP16_TO_FP32(x[i].d)); + vector float vyd = vec_splats(y[i].d); + vector float vd = vec_mul(vxd, vyd); + + vector signed int vsumi0 = vec_splats((int32_t)0); + vector signed int vsumi1 = vec_splats((int32_t)0); + vector signed int vsumi2 = vec_splats((int32_t)0); + vector signed int vsumi3 = vec_splats((int32_t)0); + vector signed int vsumi4 = vec_splats((int32_t)0); + vector signed int vsumi5 = vec_splats((int32_t)0); + vector signed int vsumi6 = vec_splats((int32_t)0); + vector signed int vsumi7 = vec_splats((int32_t)0); + + const uint8_t * restrict q3 = x[i].qs; + const uint32_t * restrict signs = (const uint32_t *)(x[i].qs + QK_K/4); + const int8_t * restrict q8 = y[i].qs; + +#pragma GCC unroll 1 + for (int j = 0; j < QK_K/32; j += 2) { + __builtin_prefetch(q3, 0, 1); + __builtin_prefetch(q8, 0, 1); + + vector unsigned int aux32x4_0 = {iq3xxs_grid[q3[ 0]], iq3xxs_grid[q3[ 1]], iq3xxs_grid[q3[ 2]], iq3xxs_grid[q3[ 3]]}; + vector unsigned int aux32x4_1 = {iq3xxs_grid[q3[ 4]], iq3xxs_grid[q3[ 5]], iq3xxs_grid[q3[ 6]], iq3xxs_grid[q3[ 7]]}; + vector unsigned int aux32x4_2 = {iq3xxs_grid[q3[ 8]], iq3xxs_grid[q3[ 9]], iq3xxs_grid[q3[10]], iq3xxs_grid[q3[11]]}; + vector unsigned int aux32x4_3 = {iq3xxs_grid[q3[12]], iq3xxs_grid[q3[13]], iq3xxs_grid[q3[14]], iq3xxs_grid[q3[15]]}; + q3 += 16; + + vector unsigned long long aux64x2_0 = {(uint64_t)(signs64[(signs[0] >> 0) & 127]), (uint64_t)(signs64[(signs[0] >> 7) & 127])}; + vector unsigned long long aux64x2_1 = {(uint64_t)(signs64[(signs[0] >> 14) & 127]), (uint64_t)(signs64[(signs[0] >> 21) & 127])}; + vector unsigned long long aux64x2_2 = {(uint64_t)(signs64[(signs[1] >> 0) & 127]), (uint64_t)(signs64[(signs[1] >> 7) & 127])}; + vector unsigned long long aux64x2_3 = {(uint64_t)(signs64[(signs[1] >> 14) & 127]), (uint64_t)(signs64[(signs[1] >> 21) & 127])}; + + vector signed char q3x0 = vec_mul((vector signed char)aux64x2_0, (vector signed char)aux32x4_0); + vector signed char q3x1 = vec_mul((vector signed char)aux64x2_1, (vector signed char)aux32x4_1); + vector signed char q3x2 = vec_mul((vector signed char)aux64x2_2, (vector signed char)aux32x4_2); + vector signed char q3x3 = vec_mul((vector signed char)aux64x2_3, (vector signed char)aux32x4_3); + + vector signed char q8y0 = vec_xl( 0, q8); + vector signed char q8y1 = vec_xl(16, q8); + vector signed char q8y2 = vec_xl(32, q8); + vector signed char q8y3 = vec_xl(48, q8); + q8 += 64; + + vector signed short qv0 = vec_add(vec_mule(q3x0, q8y0), vec_mulo(q3x0, q8y0)); + vector signed short qv1 = vec_add(vec_mule(q3x1, q8y1), vec_mulo(q3x1, q8y1)); + vector signed short qv2 = vec_add(vec_mule(q3x2, q8y2), vec_mulo(q3x2, q8y2)); + vector signed short qv3 = vec_add(vec_mule(q3x3, q8y3), vec_mulo(q3x3, q8y3)); + + const uint16_t ls0 = (uint16_t)(signs[0] >> 28); + const uint16_t ls1 = (uint16_t)(signs[1] >> 28); + signs += 2; + + vector signed short vscales01 = (vector signed short)vec_splats((uint16_t)(2*ls0+1)); + vector signed short vscales23 = (vector signed short)vec_splats((uint16_t)(2*ls1+1)); + + vsumi0 = vec_add(vec_mule(qv0, vscales01), vsumi0); + vsumi1 = vec_add(vec_mule(qv1, vscales01), vsumi1); + vsumi2 = vec_add(vec_mule(qv2, vscales23), vsumi2); + vsumi3 = vec_add(vec_mule(qv3, vscales23), vsumi3); + vsumi4 = vec_add(vec_mulo(qv0, vscales01), vsumi4); + vsumi5 = vec_add(vec_mulo(qv1, vscales01), vsumi5); + vsumi6 = vec_add(vec_mulo(qv2, vscales23), vsumi6); + vsumi7 = vec_add(vec_mulo(qv3, vscales23), vsumi7); + } + + vsumi0 = vec_add(vsumi0, vsumi4); + vsumi1 = vec_add(vsumi1, vsumi5); + vsumi2 = vec_add(vsumi2, vsumi6); + vsumi3 = vec_add(vsumi3, vsumi7); + + vsumf0 = vec_madd(vec_ctf(vsumi0, 0), vd, vsumf0); + vsumf1 = vec_madd(vec_ctf(vsumi1, 0), vd, vsumf1); + vsumf2 = vec_madd(vec_ctf(vsumi2, 0), vd, vsumf2); + vsumf3 = vec_madd(vec_ctf(vsumi3, 0), vd, vsumf3); + } + + vsumf0 = vec_add(vsumf0, vsumf2); + vsumf1 = vec_add(vsumf1, vsumf3); + + vsumf0 = vec_add(vsumf0, vsumf1); + + vsumf0 = vec_add(vsumf0, vec_sld(vsumf0, vsumf0, 4)); + vsumf0 = vec_add(vsumf0, vec_sld(vsumf0, vsumf0, 8)); + + *s = 0.25f * vec_extract(vsumf0, 0); #else uint32_t aux32; @@ -9279,6 +11075,124 @@ void ggml_vec_dot_iq3_s_q8_K (int n, float * restrict s, size_t bs, const void * *s = hsum_float_8(accumf); +#elif defined(__POWER9_VECTOR__) + static const uint8_t k_mask1[32] = {0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x01, 0x01, 0x01, 0x01, 0x01, 0x01, 0x01, 0x01, + 0x02, 0x02, 0x02, 0x02, 0x02, 0x02, 0x02, 0x02, 0x03, 0x03, 0x03, 0x03, 0x03, 0x03, 0x03, 0x03 + }; + + static const uint8_t k_mask2[16] = {0x01, 0x02, 0x04, 0x08, 0x10, 0x20, 0x40, 0x80, 0x01, 0x02, 0x04, 0x08, 0x10, 0x20, 0x40, 0x80,}; + + vector float vsumf0 = vec_splats(0.0f); + vector float vsumf1 = vec_splats(0.0f); + vector float vsumf2 = vec_splats(0.0f); + vector float vsumf3 = vec_splats(0.0f); + + const vector unsigned char mask0 = vec_xl( 0, k_mask1); + const vector unsigned char mask1 = vec_xl(16, k_mask1); + const vector signed char mask2 = (vector signed char)vec_xl( 0, k_mask2); + + for (int i = 0; i < nb; ++i) { + vector float vxd = vec_splats(GGML_FP16_TO_FP32(x[i].d)); + vector float vyd = vec_splats(y[i].d); + vector float vd = vec_mul(vxd, vyd); + + const uint8_t * restrict q3 = x[i].qs; + const uint8_t * restrict qh = x[i].qh; + const uint16_t * restrict signs = (const uint16_t *)(x[i].signs); + const uint8_t * restrict sc = x[i].scales; + const int8_t * restrict q8 = y[i].qs; + + vector signed int vsumi0 = vec_splats((int32_t)0); + vector signed int vsumi1 = vec_splats((int32_t)0); + vector signed int vsumi2 = vec_splats((int32_t)0); + vector signed int vsumi3 = vec_splats((int32_t)0); + vector signed int vsumi4 = vec_splats((int32_t)0); + vector signed int vsumi5 = vec_splats((int32_t)0); + vector signed int vsumi6 = vec_splats((int32_t)0); + vector signed int vsumi7 = vec_splats((int32_t)0); + + for (int j = 0; j < QK_K/32; j += 2) { + __builtin_prefetch(q3, 0, 1); + __builtin_prefetch(q8, 0, 1); + + vector unsigned int aux32x4_0 = {iq3s_grid[q3[ 0] | ((qh[0] << 8) & 256)], iq3s_grid[q3[ 1] | ((qh[0] << 7) & 256)], + iq3s_grid[q3[ 2] | ((qh[0] << 6) & 256)], iq3s_grid[q3[ 3] | ((qh[0] << 5) & 256)]}; + vector unsigned int aux32x4_1 = {iq3s_grid[q3[ 4] | ((qh[0] << 4) & 256)], iq3s_grid[q3[ 5] | ((qh[0] << 3) & 256)], + iq3s_grid[q3[ 6] | ((qh[0] << 2) & 256)], iq3s_grid[q3[ 7] | ((qh[0] << 1) & 256)]}; + vector unsigned int aux32x4_2 = {iq3s_grid[q3[ 8] | ((qh[1] << 8) & 256)], iq3s_grid[q3[ 9] | ((qh[1] << 7) & 256)], + iq3s_grid[q3[10] | ((qh[1] << 6) & 256)], iq3s_grid[q3[11] | ((qh[1] << 5) & 256)]}; + vector unsigned int aux32x4_3 = {iq3s_grid[q3[12] | ((qh[1] << 4) & 256)], iq3s_grid[q3[13] | ((qh[1] << 3) & 256)], + iq3s_grid[q3[14] | ((qh[1] << 2) & 256)], iq3s_grid[q3[15] | ((qh[1] << 1) & 256)]}; + q3 += 16; + qh += 2; + + vector signed char vsigns01 = (vector signed char)vec_splats(*(const uint32_t *)&signs[0]); + vector signed char vsigns02 = (vector signed char)vec_splats(*(const uint32_t *)&signs[2]); + signs += 4; + + vector signed char vsigns0 = vec_perm(vsigns01, vsigns01, mask0); + vector signed char vsigns1 = vec_perm(vsigns01, vsigns01, mask1); + vector signed char vsigns2 = vec_perm(vsigns02, vsigns02, mask0); + vector signed char vsigns3 = vec_perm(vsigns02, vsigns02, mask1); + + vsigns0 = (vector signed char)vec_cmpeq(vec_and(vsigns0, mask2), mask2); + vsigns1 = (vector signed char)vec_cmpeq(vec_and(vsigns1, mask2), mask2); + vsigns2 = (vector signed char)vec_cmpeq(vec_and(vsigns2, mask2), mask2); + vsigns3 = (vector signed char)vec_cmpeq(vec_and(vsigns3, mask2), mask2); + + vector signed char q3x0 = vec_sub(vec_xor(vsigns0, (vector signed char)aux32x4_0), vsigns0); + vector signed char q3x1 = vec_sub(vec_xor(vsigns1, (vector signed char)aux32x4_1), vsigns1); + vector signed char q3x2 = vec_sub(vec_xor(vsigns2, (vector signed char)aux32x4_2), vsigns2); + vector signed char q3x3 = vec_sub(vec_xor(vsigns3, (vector signed char)aux32x4_3), vsigns3); + + vector signed char q8y0 = vec_xl( 0, q8); + vector signed char q8y1 = vec_xl(16, q8); + vector signed char q8y2 = vec_xl(32, q8); + vector signed char q8y3 = vec_xl(48, q8); + q8 += 64; + + vector signed short qv0 = vec_add(vec_mule(q3x0, q8y0), vec_mulo(q3x0, q8y0)); + vector signed short qv1 = vec_add(vec_mule(q3x1, q8y1), vec_mulo(q3x1, q8y1)); + vector signed short qv2 = vec_add(vec_mule(q3x2, q8y2), vec_mulo(q3x2, q8y2)); + vector signed short qv3 = vec_add(vec_mule(q3x3, q8y3), vec_mulo(q3x3, q8y3)); + + const uint16_t ls0 = (uint16_t)(sc[0] & 0xf); + const uint16_t ls1 = (uint16_t)(sc[0] >> 4); + sc ++; + + vector signed short vscales01 = (vector signed short)vec_splats((uint16_t)(2*ls0+1)); + vector signed short vscales23 = (vector signed short)vec_splats((uint16_t)(2*ls1+1)); + + vsumi0 = vec_add(vec_mule(qv0, vscales01), vsumi0); + vsumi1 = vec_add(vec_mule(qv1, vscales01), vsumi1); + vsumi2 = vec_add(vec_mule(qv2, vscales23), vsumi2); + vsumi3 = vec_add(vec_mule(qv3, vscales23), vsumi3); + vsumi4 = vec_add(vec_mulo(qv0, vscales01), vsumi4); + vsumi5 = vec_add(vec_mulo(qv1, vscales01), vsumi5); + vsumi6 = vec_add(vec_mulo(qv2, vscales23), vsumi6); + vsumi7 = vec_add(vec_mulo(qv3, vscales23), vsumi7); + } + + vsumi0 = vec_add(vsumi0, vsumi4); + vsumi1 = vec_add(vsumi1, vsumi5); + vsumi2 = vec_add(vsumi2, vsumi6); + vsumi3 = vec_add(vsumi3, vsumi7); + + vsumf0 = vec_madd(vec_ctf(vsumi0, 0), vd, vsumf0); + vsumf1 = vec_madd(vec_ctf(vsumi1, 0), vd, vsumf1); + vsumf2 = vec_madd(vec_ctf(vsumi2, 0), vd, vsumf2); + vsumf3 = vec_madd(vec_ctf(vsumi3, 0), vd, vsumf3); + } + + vsumf0 = vec_add(vsumf0, vsumf2); + vsumf1 = vec_add(vsumf1, vsumf3); + + vsumf0 = vec_add(vsumf0, vsumf1); + + vsumf0 = vec_add(vsumf0, vec_sld(vsumf0, vsumf0, 4)); + vsumf0 = vec_add(vsumf0, vec_sld(vsumf0, vsumf0, 8)); + + *s = vec_extract(vsumf0, 0); #else float sumf = 0.f; @@ -9433,6 +11347,113 @@ void ggml_vec_dot_iq1_s_q8_K (int n, float * restrict s, size_t bs, const void *s = hsum_float_8(accum) + IQ1S_DELTA * accum1; +#elif defined(__POWER9_VECTOR__) + const vector unsigned char v0 = vec_splats((unsigned char)0x0); + const vector unsigned short vsign = vec_splats((unsigned short)0x8000); + + vector float vsumf0 = vec_splats(0.0f); + vector float vsumf1 = vec_splats(0.0f); + vector float vsumf2 = vec_splats(0.0f); + vector float vsumf3 = vec_splats(0.0f); + + for (int i = 0; i < nb; ++i) { + vector float vxd = vec_splats(GGML_FP16_TO_FP32(x[i].d)); + vector float vyd = vec_splats(y[i].d); + vector float vd = vec_mul(vxd, vyd); + + vector signed int vsumi0 = vec_splats((int32_t)0); + vector signed int vsumi1 = vec_splats((int32_t)0); + vector signed int vsumi2 = vec_splats((int32_t)0); + vector signed int vsumi3 = vec_splats((int32_t)0); + vector signed int vsumi4 = vec_splats((int32_t)0); + vector signed int vsumi5 = vec_splats((int32_t)0); + vector signed int vsumi6 = vec_splats((int32_t)0); + vector signed int vsumi7 = vec_splats((int32_t)0); + vector signed int vsumi8 = vec_splats((int32_t)0); + + const uint8_t * restrict q1 = x[i].qs; + const uint16_t * restrict qh = x[i].qh; + const int8_t * restrict q8 = y[i].qs; + const int16_t * restrict qs = y[i].bsums; + + for (int j = 0; j < QK_K/32; j += 2) { + __builtin_prefetch(q1, 0, 1); + __builtin_prefetch(qh, 0, 1); + __builtin_prefetch(q8, 0, 1); + + vector signed long long aux64x2_0 = {*(const int64_t *)(iq1s_grid + (q1[0] | ((qh[0] << 8) & 0x700))), *(const int64_t *)(iq1s_grid + (q1[1] | ((qh[0] << 5) & 0x700)))}; + vector signed long long aux64x2_1 = {*(const int64_t *)(iq1s_grid + (q1[2] | ((qh[0] << 2) & 0x700))), *(const int64_t *)(iq1s_grid + (q1[3] | ((qh[0] >> 1) & 0x700)))}; + vector signed long long aux64x2_2 = {*(const int64_t *)(iq1s_grid + (q1[4] | ((qh[1] << 8) & 0x700))), *(const int64_t *)(iq1s_grid + (q1[5] | ((qh[1] << 5) & 0x700)))}; + vector signed long long aux64x2_3 = {*(const int64_t *)(iq1s_grid + (q1[6] | ((qh[1] << 2) & 0x700))), *(const int64_t *)(iq1s_grid + (q1[7] | ((qh[1] >> 1) & 0x700)))}; + q1 += 8; + + vector signed char q1x0 = (vector signed char)aux64x2_0; + vector signed char q1x1 = (vector signed char)aux64x2_1; + vector signed char q1x2 = (vector signed char)aux64x2_2; + vector signed char q1x3 = (vector signed char)aux64x2_3; + + vector signed char q8y0 = vec_xl( 0, q8); + vector signed char q8y1 = vec_xl(16, q8); + vector signed char q8y2 = vec_xl(32, q8); + vector signed char q8y3 = vec_xl(48, q8); + q8 += 64; + + vector signed short qv0 = vec_add(vec_mule(q1x0, q8y0), vec_mulo(q1x0, q8y0)); + vector signed short qv1 = vec_add(vec_mule(q1x1, q8y1), vec_mulo(q1x1, q8y1)); + vector signed short qv2 = vec_add(vec_mule(q1x2, q8y2), vec_mulo(q1x2, q8y2)); + vector signed short qv3 = vec_add(vec_mule(q1x3, q8y3), vec_mulo(q1x3, q8y3)); + + const uint16_t ls0 = (uint16_t)((qh[0] >> 12) & 7); + const uint16_t ls1 = (uint16_t)((qh[1] >> 12) & 7); + + vector signed short vscales01 = (vector signed short)vec_splats((uint16_t)(2*ls0+1)); + vector signed short vscales23 = (vector signed short)vec_splats((uint16_t)(2*ls1+1)); + vector signed short vscales = vec_sld(vscales23, vscales01, 8); + + vsumi0 = vec_add(vec_mule(qv0, vscales01), vsumi0); + vsumi1 = vec_add(vec_mule(qv1, vscales01), vsumi1); + vsumi2 = vec_add(vec_mule(qv2, vscales23), vsumi2); + vsumi3 = vec_add(vec_mule(qv3, vscales23), vsumi3); + vsumi4 = vec_add(vec_mulo(qv0, vscales01), vsumi4); + vsumi5 = vec_add(vec_mulo(qv1, vscales01), vsumi5); + vsumi6 = vec_add(vec_mulo(qv2, vscales23), vsumi6); + vsumi7 = vec_add(vec_mulo(qv3, vscales23), vsumi7); + + vector signed short q8ysums = vec_xl_len(qs, 8); + qs += 4; + q8ysums = vec_mergeh(q8ysums, (vector signed short)v0); + + vector signed short qxh = (vector signed short)vec_sld(vec_splats(qh[1]), vec_splats(qh[0]), 8); + qh += 2; + vector bool short vsel = vec_cmpge(qxh, (vector signed short)v0); + + vector signed short q8ysum = vec_sel((vector signed short)vec_xor((vector unsigned short)q8ysums, vsign), q8ysums, vsel); + + vsumi8 = vec_add(vec_mule(q8ysum, vscales), vsumi8); + } + + vsumi0 = vec_add(vsumi0, vsumi4); + vsumi1 = vec_add(vsumi1, vsumi5); + vsumi2 = vec_add(vsumi2, vsumi6); + vsumi3 = vec_add(vsumi3, vsumi7); + + vsumf0 = vec_madd(vec_ctf(vsumi0, 0), vd, vsumf0); + vsumf1 = vec_madd(vec_ctf(vsumi1, 0), vd, vsumf1); + vsumf2 = vec_madd(vec_ctf(vsumi2, 0), vd, vsumf2); + vsumf3 = vec_madd(vec_ctf(vsumi3, 0), vd, vsumf3); + + vsumf0 = vec_madd(vec_ctf(vsumi8, 0), vec_mul(vd, vec_splats(IQ1S_DELTA)), vsumf0); + } + + vsumf0 = vec_add(vsumf0, vsumf2); + vsumf1 = vec_add(vsumf1, vsumf3); + + vsumf0 = vec_add(vsumf0, vsumf1); + + vsumf0 = vec_add(vsumf0, vec_sld(vsumf0, vsumf0, 4)); + vsumf0 = vec_add(vsumf0, vec_sld(vsumf0, vsumf0, 8)); + + *s = vec_extract(vsumf0, 0); #else float sumf = 0; @@ -9789,6 +11810,51 @@ void ggml_vec_dot_iq4_nl_q8_0(int n, float * restrict s, size_t bs, const void * *s = hsum_float_8(_mm256_add_ps(accum1, accum2)); +#elif defined(__POWER9_VECTOR__) + const vector signed char lowMask = vec_splats((signed char)0xF); + const vector unsigned char v4 = vec_splats((unsigned char)0x4); + + vector float vsumf0 = vec_splats(0.0f); + vector float vsumf1 = vec_splats(0.0f); + + const vector signed char values = vec_xl( 0, kvalues_iq4nl); + +#pragma GCC unroll 4 + for (int ib = 0; ib < nb; ++ib) { + __builtin_prefetch(x[ib].qs, 0, 1); + __builtin_prefetch(y[ib].qs, 0, 1); + + + vector float vxd = vec_splats(GGML_FP16_TO_FP32(x[ib].d)); + vector float vyd = vec_splats(GGML_FP16_TO_FP32(y[ib].d)); + vector float vd = vec_mul(vxd, vyd); + + vector signed char qxs = (vector signed char)vec_xl( 0, x[ib].qs); + vector signed char q4x0 = vec_and(qxs, lowMask); + vector signed char q4x1 = vec_sr(qxs, v4); + + q4x0 = vec_perm(values, values, (vector unsigned char)q4x0); + q4x1 = vec_perm(values, values, (vector unsigned char)q4x1); + + vector signed char q8y0 = vec_xl( 0, y[ib].qs); + vector signed char q8y1 = vec_xl(16, y[ib].qs); + + vector signed short qv0 = vec_add(vec_mule(q4x0, q8y0), vec_mulo(q4x0, q8y0)); + vector signed short qv1 = vec_add(vec_mule(q4x1, q8y1), vec_mulo(q4x1, q8y1)); + + vector signed int vsumi0 = vec_add(vec_unpackh(qv0), vec_unpackl(qv0)); + vector signed int vsumi1 = vec_add(vec_unpackh(qv1), vec_unpackl(qv1)); + + vsumf0 = vec_madd(vec_ctf(vsumi0, 0), vd, vsumf0); + vsumf1 = vec_madd(vec_ctf(vsumi1, 0), vd, vsumf1); + } + + vsumf0 = vec_add(vsumf0, vsumf1); + + vsumf0 = vec_add(vsumf0, vec_sld(vsumf0, vsumf0, 4)); + vsumf0 = vec_add(vsumf0, vec_sld(vsumf0, vsumf0, 8)); + + *s = vec_extract(vsumf0, 0); #else float sumf = 0; for (int ib = 0; ib < nb; ++ib) { @@ -9900,6 +11966,105 @@ void ggml_vec_dot_iq4_xs_q8_K(int n, float * restrict s, size_t bs, const void * *s = hsum_float_8(accum); +#elif defined(__POWER9_VECTOR__) + const vector signed char lowMask = vec_splats((signed char)0xF); + const vector unsigned char v4 = vec_splats((unsigned char)0x4); + + vector float vsumf0 = vec_splats(0.0f); + vector float vsumf1 = vec_splats(0.0f); + vector float vsumf2 = vec_splats(0.0f); + vector float vsumf3 = vec_splats(0.0f); + + const vector signed char values = vec_xl( 0, kvalues_iq4nl); + + for (int ibl = 0; ibl < nb; ++ibl) { + + vector float vxd = vec_splats(GGML_FP16_TO_FP32(x[ibl].d)); + vector float vyd = vec_splats(y[ibl].d); + vector float vd = vec_mul(vxd, vyd); + + vector signed int vsumi0 = vec_splats((int32_t)0); + vector signed int vsumi1 = vec_splats((int32_t)0); + vector signed int vsumi2 = vec_splats((int32_t)0); + vector signed int vsumi3 = vec_splats((int32_t)0); + vector signed int vsumi4 = vec_splats((int32_t)0); + vector signed int vsumi5 = vec_splats((int32_t)0); + vector signed int vsumi6 = vec_splats((int32_t)0); + vector signed int vsumi7 = vec_splats((int32_t)0); + + uint16_t h = x[ibl].scales_h; + + const uint8_t * restrict q4 = x[ibl].qs; + const uint8_t * restrict sc = x[ibl].scales_l; + const int8_t * restrict q8 = y[ibl].qs; + + for (int ib = 0; ib < QK_K/64; ib ++ ) { + __builtin_prefetch(q4, 0, 1); + __builtin_prefetch(q8, 0, 1); + + vector signed char qxs0 = (vector signed char)vec_xl( 0, q4); + vector signed char qxs1 = (vector signed char)vec_xl(16, q4); + q4 += 32; + + vector signed char q4x00 = (vector signed char)vec_and(qxs0, lowMask); + vector signed char q4x01 = (vector signed char)vec_sr(qxs0, v4); + vector signed char q4x10 = (vector signed char)vec_and(qxs1, lowMask); + vector signed char q4x11 = (vector signed char)vec_sr(qxs1, v4); + + q4x00 = vec_perm(values, values, (vector unsigned char)q4x00); + q4x01 = vec_perm(values, values, (vector unsigned char)q4x01); + q4x10 = vec_perm(values, values, (vector unsigned char)q4x10); + q4x11 = vec_perm(values, values, (vector unsigned char)q4x11); + + vector signed char q8y0 = vec_xl( 0, q8); + vector signed char q8y1 = vec_xl(16, q8); + vector signed char q8y2 = vec_xl(32, q8); + vector signed char q8y3 = vec_xl(48, q8); + q8 += 64; + + vector signed short qv0 = vec_add(vec_mule(q4x00, q8y0), vec_mulo(q4x00, q8y0)); + vector signed short qv1 = vec_add(vec_mule(q4x01, q8y1), vec_mulo(q4x01, q8y1)); + vector signed short qv2 = vec_add(vec_mule(q4x10, q8y2), vec_mulo(q4x10, q8y2)); + vector signed short qv3 = vec_add(vec_mule(q4x11, q8y3), vec_mulo(q4x11, q8y3)); + + const uint16_t ls0 = (uint16_t)(((sc[0] & 0xf) | ((h << 4) & 0x30)) - 32); + const uint16_t ls1 = (uint16_t)(((sc[0] >> 4) | ((h << 2) & 0x30)) - 32); + h >>= 4; + sc ++; + + vector signed short vscales01 = vec_splats((int16_t)ls0); + vector signed short vscales23 = vec_splats((int16_t)ls1); + + vsumi0 = vec_add(vec_mule(qv0, vscales01), vsumi0); + vsumi1 = vec_add(vec_mule(qv1, vscales01), vsumi1); + vsumi2 = vec_add(vec_mule(qv2, vscales23), vsumi2); + vsumi3 = vec_add(vec_mule(qv3, vscales23), vsumi3); + vsumi4 = vec_add(vec_mulo(qv0, vscales01), vsumi4); + vsumi5 = vec_add(vec_mulo(qv1, vscales01), vsumi5); + vsumi6 = vec_add(vec_mulo(qv2, vscales23), vsumi6); + vsumi7 = vec_add(vec_mulo(qv3, vscales23), vsumi7); + } + + vsumi0 = vec_add(vsumi0, vsumi4); + vsumi1 = vec_add(vsumi1, vsumi5); + vsumi2 = vec_add(vsumi2, vsumi6); + vsumi3 = vec_add(vsumi3, vsumi7); + + vsumf0 = vec_madd(vec_ctf(vsumi0, 0), vd, vsumf0); + vsumf1 = vec_madd(vec_ctf(vsumi1, 0), vd, vsumf1); + vsumf2 = vec_madd(vec_ctf(vsumi2, 0), vd, vsumf2); + vsumf3 = vec_madd(vec_ctf(vsumi3, 0), vd, vsumf3); + } + + vsumf0 = vec_add(vsumf0, vsumf2); + vsumf1 = vec_add(vsumf1, vsumf3); + + vsumf0 = vec_add(vsumf0, vsumf1); + + vsumf0 = vec_add(vsumf0, vec_sld(vsumf0, vsumf0, 4)); + vsumf0 = vec_add(vsumf0, vec_sld(vsumf0, vsumf0, 8)); + + *s = vec_extract(vsumf0, 0); #else float sumf = 0; for (int ibl = 0; ibl < nb; ++ibl) {