From 5db246a547b004eef139f37ddbec37c8c875c58a Mon Sep 17 00:00:00 2001 From: Georgi Gerganov Date: Fri, 21 Apr 2023 20:12:42 +0300 Subject: [PATCH 1/2] ggml : prefer vzip to vuzp This way we always use the same type of instruction across all quantizations --- ggml.c | 12 ++++++------ 1 file changed, 6 insertions(+), 6 deletions(-) diff --git a/ggml.c b/ggml.c index 6cea937c8d735..6f915517dc9cf 100644 --- a/ggml.c +++ b/ggml.c @@ -2632,18 +2632,18 @@ static void ggml_vec_dot_q4_1_q8_0(const int n, float * restrict s, const void * const int8x16_t v0_1l = vreinterpretq_s8_u8(vandq_u8 (v0_1, m4b)); const int8x16_t v0_1h = vreinterpretq_s8_u8(vshrq_n_u8(v0_1, 4)); + // interleave + const int8x16_t v0_0lz = vzip1q_s8(v0_0l, v0_0h); + const int8x16_t v0_0hz = vzip2q_s8(v0_0l, v0_0h); + const int8x16_t v0_1lz = vzip1q_s8(v0_1l, v0_1h); + const int8x16_t v0_1hz = vzip2q_s8(v0_1l, v0_1h); + // load y const int8x16_t v1_0l = vld1q_s8(y0->qs); const int8x16_t v1_0h = vld1q_s8(y0->qs + 16); const int8x16_t v1_1l = vld1q_s8(y1->qs); const int8x16_t v1_1h = vld1q_s8(y1->qs + 16); - // interleave - const int8x16_t v1_0ls = vuzp1q_s8(v1_0l, v1_0h); - const int8x16_t v1_0hs = vuzp2q_s8(v1_0l, v1_0h); - const int8x16_t v1_1ls = vuzp1q_s8(v1_1l, v1_1h); - const int8x16_t v1_1hs = vuzp2q_s8(v1_1l, v1_1h); - #if defined(__ARM_FEATURE_DOTPROD) // dot product into int32x4_t const int32x4_t p_0 = vdotq_s32(vdotq_s32(vdupq_n_s32(0), v0_0l, v1_0ls), v0_0h, v1_0hs); From 266bb63f685bbc9fd272fdc91db112585d09d3f7 Mon Sep 17 00:00:00 2001 From: Georgi Gerganov Date: Fri, 21 Apr 2023 20:13:12 +0300 Subject: [PATCH 2/2] ggml : alternative Q4_3 format + implementation This one achieves 50m / token on M1 Pro --- ggml.c | 150 ++++++++++++++++++++++++++++++--------------------------- 1 file changed, 78 insertions(+), 72 deletions(-) diff --git a/ggml.c b/ggml.c index 6f915517dc9cf..d8b42b3580b6e 100644 --- a/ggml.c +++ b/ggml.c @@ -646,13 +646,14 @@ typedef struct { } block_q4_2; static_assert(sizeof(block_q4_2) == sizeof(ggml_fp16_t) + QK4_2 / 2, "wrong q4_2 block size/padding"); -#define QK4_3 16 +#define QK4_3 32 typedef struct { - ggml_fp16_t d; // delta + ggml_fp16_t d0; // delta + ggml_fp16_t d1; // delta ggml_fp16_t m; // min uint8_t qs[QK4_3 / 2]; // nibbles / quants } block_q4_3; -static_assert(sizeof(block_q4_3) == 2 * sizeof(ggml_fp16_t) + QK4_3 / 2, "wrong q4_3 block size/padding"); +static_assert(sizeof(block_q4_3) == 3 * sizeof(ggml_fp16_t) + QK4_3 / 2, "wrong q4_3 block size/padding"); #define QK8_0 32 typedef struct { @@ -1244,32 +1245,46 @@ static void quantize_row_q4_3_reference(const float * restrict x, block_q4_3 * r const int nb = k / QK4_3; for (int i = 0; i < nb; i++) { - float min = FLT_MAX; - float max = -FLT_MAX; + float min0 = FLT_MAX; + float max0 = -FLT_MAX; + float min1 = FLT_MAX; + float max1 = -FLT_MAX; - for (int l = 0; l < QK4_3; l++) { - const float v = x[i*QK4_3 + l]; - if (v < min) min = v; - if (v > max) max = v; + for (int l = 0; l < QK4_3/2; l++) { + const float v0 = x[i*QK4_3 + l]; + const float v1 = x[i*QK4_3 + l + QK4_3/2]; + if (v0 < min0) min0 = v0; + if (v0 > max0) max0 = v0; + if (v1 < min1) min1 = v1; + if (v1 > max1) max1 = v1; } - const float d = (max - min) / ((1 << 4) - 1); - const float id = d ? 1.0f/d : 0.0f; + const float min = MIN(min0, min1); - y[i].d = GGML_FP32_TO_FP16(d); + const float d0 = (max0 - min) / ((1 << 4) - 1); + const float d1 = (max1 - min) / ((1 << 4) - 1); + const float id0 = d0 ? 1.0f/d0 : 0.0f; + const float id1 = d1 ? 1.0f/d1 : 0.0f; + + y[i].d0 = GGML_FP32_TO_FP16(d0); + y[i].d1 = GGML_FP32_TO_FP16(d1); y[i].m = GGML_FP32_TO_FP16(min); - for (int l = 0; l < QK4_3; l += 2) { - const float v0 = (x[i*QK4_3 + l + 0] - min)*id; - const float v1 = (x[i*QK4_3 + l + 1] - min)*id; + for (int l = 0; l < QK4_3/2; l += 2) { + const float v0_0 = (x[i*QK4_3 + l + 0] - min)*id0; + const float v0_1 = (x[i*QK4_3 + l + 1] - min)*id0; - const uint8_t vi0 = (int) (v0 + 0.5f); - const uint8_t vi1 = (int) (v1 + 0.5f); + const float v1_0 = (x[i*QK4_3 + l + QK4_3/2 + 0] - min)*id1; + const float v1_1 = (x[i*QK4_3 + l + QK4_3/2 + 1] - min)*id1; - assert(vi0 < 16); - assert(vi1 < 16); + const uint8_t vi0_0 = (int) (v0_0 + 0.5f); + const uint8_t vi0_1 = (int) (v0_1 + 0.5f); - y[i].qs[l/2] = vi0 | (vi1 << 4); + const uint8_t vi1_0 = (int) (v1_0 + 0.5f); + const uint8_t vi1_1 = (int) (v1_1 + 0.5f); + + y[i].qs[l/2] = vi0_0 | (vi0_1 << 4); + y[i].qs[l/2 + QK4_3/4] = vi1_0 | (vi1_1 << 4); } } } @@ -1737,25 +1752,31 @@ static void dequantize_row_q4_3(const void * restrict vx, float * restrict y, in const block_q4_3 * restrict x = vx; for (int i = 0; i < nb; i++) { - const float d = GGML_FP16_TO_FP32(x[i].d); - const float m = GGML_FP16_TO_FP32(x[i].m); + const float d0 = GGML_FP16_TO_FP32(x[i].d0); + const float d1 = GGML_FP16_TO_FP32(x[i].d1); + const float m = GGML_FP16_TO_FP32(x[i].m); - const uint8_t * restrict pp = x[i].qs; + for (int l = 0; l < QK4_3/2; l += 2) { + const uint8_t vi0 = x[i].qs[ l/2]; + const uint8_t vi1 = x[i].qs[QK4_3/4 + l/2]; - for (int l = 0; l < QK4_3; l += 2) { - const uint8_t vi = pp[l/2]; + const int8_t vi0_0 = vi0 & 0xf; + const int8_t vi0_1 = vi0 >> 4; - const int8_t vi0 = vi & 0xf; - const int8_t vi1 = vi >> 4; + const int8_t vi1_0 = vi1 & 0xf; + const int8_t vi1_1 = vi1 >> 4; - const float v0 = vi0*d + m; - const float v1 = vi1*d + m; + const float v0_0 = vi0_0*d0 + m; + const float v0_1 = vi0_1*d0 + m; + + const float v1_0 = vi1_0*d1 + m; + const float v1_1 = vi1_1*d1 + m; - y[i*QK4_3 + l + 0] = v0; - y[i*QK4_3 + l + 1] = v1; + y[i*QK4_3 + l + 0] = v0_0; + y[i*QK4_3 + l + 1] = v0_1; - assert(!isnan(y[i*QK4_3 + l + 0])); - assert(!isnan(y[i*QK4_3 + l + 1])); + y[i*QK4_3 + QK4_3/2 + l + 0] = v1_0; + y[i*QK4_3 + QK4_3/2 + l + 1] = v1_1; } } } @@ -2646,21 +2667,21 @@ static void ggml_vec_dot_q4_1_q8_0(const int n, float * restrict s, const void * #if defined(__ARM_FEATURE_DOTPROD) // dot product into int32x4_t - const int32x4_t p_0 = vdotq_s32(vdotq_s32(vdupq_n_s32(0), v0_0l, v1_0ls), v0_0h, v1_0hs); - const int32x4_t p_1 = vdotq_s32(vdotq_s32(vdupq_n_s32(0), v0_1l, v1_1ls), v0_1h, v1_1hs); + const int32x4_t p_0 = vdotq_s32(vdotq_s32(vdupq_n_s32(0), v0_0lz, v1_0l), v0_0hz, v1_0h); + const int32x4_t p_1 = vdotq_s32(vdotq_s32(vdupq_n_s32(0), v0_1lz, v1_1l), v0_1hz, v1_1h); sumv0 = vmlaq_n_f32(sumv0, vcvtq_f32_s32(p_0), x0->d*y0->d); sumv1 = vmlaq_n_f32(sumv1, vcvtq_f32_s32(p_1), x1->d*y1->d); #else - const int16x8_t pl0l = vmull_s8(vget_low_s8 (v0_0l), vget_low_s8 (v1_0ls)); - const int16x8_t pl0h = vmull_s8(vget_high_s8(v0_0l), vget_high_s8(v1_0ls)); - const int16x8_t ph0l = vmull_s8(vget_low_s8 (v0_0h), vget_low_s8 (v1_0hs)); - const int16x8_t ph0h = vmull_s8(vget_high_s8(v0_0h), vget_high_s8(v1_0hs)); + const int16x8_t pl0l = vmull_s8(vget_low_s8 (v0_0lz), vget_low_s8 (v1_0l)); + const int16x8_t pl0h = vmull_s8(vget_high_s8(v0_0lz), vget_high_s8(v1_0l)); + const int16x8_t ph0l = vmull_s8(vget_low_s8 (v0_0hz), vget_low_s8 (v1_0h)); + const int16x8_t ph0h = vmull_s8(vget_high_s8(v0_0hz), vget_high_s8(v1_0h)); - const int16x8_t pl1l = vmull_s8(vget_low_s8 (v0_1l), vget_low_s8 (v1_1ls)); - const int16x8_t pl1h = vmull_s8(vget_high_s8(v0_1l), vget_high_s8(v1_1ls)); - const int16x8_t ph1l = vmull_s8(vget_low_s8 (v0_1h), vget_low_s8 (v1_1hs)); - const int16x8_t ph1h = vmull_s8(vget_high_s8(v0_1h), vget_high_s8(v1_1hs)); + const int16x8_t pl1l = vmull_s8(vget_low_s8 (v0_1lz), vget_low_s8 (v1_1l)); + const int16x8_t pl1h = vmull_s8(vget_high_s8(v0_1lz), vget_high_s8(v1_1l)); + const int16x8_t ph1l = vmull_s8(vget_low_s8 (v0_1hz), vget_low_s8 (v1_1h)); + const int16x8_t ph1h = vmull_s8(vget_high_s8(v0_1hz), vget_high_s8(v1_1h)); const int32x4_t pl0 = vaddq_s32(vpaddlq_s16(pl0l), vpaddlq_s16(pl0h)); const int32x4_t ph0 = vaddq_s32(vpaddlq_s16(ph0l), vpaddlq_s16(ph0h)); @@ -2931,7 +2952,7 @@ static void ggml_vec_dot_q4_3_q8_0(const int n, float * restrict s, const void * assert(n % QK8_0 == 0); assert(nb % 2 == 0); - assert(QK8_0 == 2*QK4_2); + assert(QK8_0 == QK4_3); const block_q4_3 * restrict x = vx; const block_q8_0 * restrict y = vy; @@ -2942,29 +2963,25 @@ static void ggml_vec_dot_q4_3_q8_0(const int n, float * restrict s, const void * float32x4_t sumv0 = vdupq_n_f32(0.0f); float32x4_t sumv1 = vdupq_n_f32(0.0f); - for (int i = 0; i < nb; i += 2) { - const block_q4_3 * restrict x0_0 = &x[2*(i + 0) + 0]; - const block_q4_3 * restrict x0_1 = &x[2*(i + 0) + 1]; - const block_q4_3 * restrict x1_0 = &x[2*(i + 1) + 0]; - const block_q4_3 * restrict x1_1 = &x[2*(i + 1) + 1]; + float summs = 0; + for (int i = 0; i < nb; i += 2) { + const block_q4_3 * restrict x0 = &x[i + 0]; + const block_q4_3 * restrict x1 = &x[i + 1]; const block_q8_0 * restrict y0 = &y[i + 0]; const block_q8_0 * restrict y1 = &y[i + 1]; - const uint8x16_t m4b = vdupq_n_u8(0xf); + summs += GGML_FP16_TO_FP32(x0->m) * y0->s + GGML_FP16_TO_FP32(x1->m) * y1->s; - const float x0_0d = GGML_FP16_TO_FP32(x0_0->d); - const float x0_1d = GGML_FP16_TO_FP32(x0_1->d); - const float x1_0d = GGML_FP16_TO_FP32(x1_0->d); - const float x1_1d = GGML_FP16_TO_FP32(x1_1->d); + const uint8x16_t m4b = vdupq_n_u8(0xf); - const float x0_0m = GGML_FP16_TO_FP32(x0_0->m); - const float x0_1m = GGML_FP16_TO_FP32(x0_1->m); - const float x1_0m = GGML_FP16_TO_FP32(x1_0->m); - const float x1_1m = GGML_FP16_TO_FP32(x1_1->m); + const float x0_0d = GGML_FP16_TO_FP32(x0->d0); + const float x0_1d = GGML_FP16_TO_FP32(x0->d1); + const float x1_0d = GGML_FP16_TO_FP32(x1->d0); + const float x1_1d = GGML_FP16_TO_FP32(x1->d1); - const uint8x16_t v0_0 = vcombine_u8(vld1_u8(x0_0->qs), vld1_u8(x0_1->qs)); - const uint8x16_t v0_1 = vcombine_u8(vld1_u8(x1_0->qs), vld1_u8(x1_1->qs)); + const uint8x16_t v0_0 = vld1q_u8(x0->qs); + const uint8x16_t v0_1 = vld1q_u8(x1->qs); // 4-bit -> 8-bit const int8x16_t v0_0l = vreinterpretq_s8_u8(vandq_u8 (v0_0, m4b)); @@ -2984,17 +3001,6 @@ static void ggml_vec_dot_q4_3_q8_0(const int n, float * restrict s, const void * const int8x16_t v1_1l = vld1q_s8(y1->qs); const int8x16_t v1_1h = vld1q_s8(y1->qs + 16); - const int16x8_t sy0_0 = vaddq_s16(vmovl_s8(vget_low_s8(v1_0l)), vmovl_s8(vget_high_s8(v1_0l))); - const int16x8_t sy0_1 = vaddq_s16(vmovl_s8(vget_low_s8(v1_0h)), vmovl_s8(vget_high_s8(v1_0h))); - - const int16x8_t sy1_0 = vaddq_s16(vmovl_s8(vget_low_s8(v1_1l)), vmovl_s8(vget_high_s8(v1_1l))); - const int16x8_t sy1_1 = vaddq_s16(vmovl_s8(vget_low_s8(v1_1h)), vmovl_s8(vget_high_s8(v1_1h))); - - sumv0 = vmlaq_n_f32(sumv0, vcvtq_f32_s32(vaddl_s16(vget_low_s16(sy0_0), vget_high_s16(sy0_0))), x0_0m*y0->d); - sumv0 = vmlaq_n_f32(sumv0, vcvtq_f32_s32(vaddl_s16(vget_low_s16(sy0_1), vget_high_s16(sy0_1))), x0_1m*y0->d); - sumv1 = vmlaq_n_f32(sumv1, vcvtq_f32_s32(vaddl_s16(vget_low_s16(sy1_0), vget_high_s16(sy1_0))), x1_0m*y1->d); - sumv1 = vmlaq_n_f32(sumv1, vcvtq_f32_s32(vaddl_s16(vget_low_s16(sy1_1), vget_high_s16(sy1_1))), x1_1m*y1->d); - #if defined(__ARM_FEATURE_DOTPROD) sumv0 = vmlaq_n_f32(sumv0, vcvtq_f32_s32(vdotq_s32(vdupq_n_s32(0), v0_0lz, v1_0l)), x0_0d*y0->d); sumv0 = vmlaq_n_f32(sumv0, vcvtq_f32_s32(vdotq_s32(vdupq_n_s32(0), v0_0hz, v1_0h)), x0_1d*y0->d); @@ -3023,7 +3029,7 @@ static void ggml_vec_dot_q4_3_q8_0(const int n, float * restrict s, const void * #endif } - sumf = vaddvq_f32(sumv0) + vaddvq_f32(sumv1); + sumf = vaddvq_f32(vaddq_f32(sumv0, sumv1)) + summs; #else // scalar for (int i = 0; i < nb; i++) {