const int K = 1280;
const int QK = 64;
-const int QB = 7;
+#define QB 7
-#define gq_t uint64_t
+//#define GGML_GQ_USE_FP16_SCALE
+
+#if defined(GGML_GQ_USE_FP16_SCALE)
+#define gq_scale_t ggml_fp16_t
+#define GGML_FP32_TO_GQ(x) ggml_fp32_to_fp16(x)
+#define GGML_GQ_TO_FP32(x) ggml_fp16_to_fp32(x)
+#else
+#define gq_scale_t float
+#define GGML_FP32_TO_GQ(x) (x)
+#define GGML_GQ_TO_FP32(x) (x)
+#endif
+
+#define gq_quant_t uint64_t
#define gq_t_bits 64
uint64_t get_time_us() {
// naive implementation
//
-void mul_mat_vec_f32_naive(
+void mul_mat_f32_naive(
const float * restrict src0, // M x K
const float * restrict src1, // N x K (transposed)
float * dst,
void quantize_1(const float * src, void * dst, int n, int k) {
char * p0 = dst;
- gq_t pp[QB];
+ gq_quant_t pp[QB];
for (int j = 0; j < n; j++) {
for (int i = 0; i < k/QK; i++) {
const uint8_t q = (v - min)*id;
for (int b = 0; b < QB; b++) {
- pp[b] |= q & (1 << b) ? (1LL << l) : 0;
+ pp[b] |= q & (1 << b) ? (1ULL << l) : 0;
}
}
for (int b = 0; b < QB; b++) {
- memcpy(p0, &pp[b], sizeof(gq_t)); p0 += sizeof(gq_t);
+ memcpy(p0, &pp[b], sizeof(gq_quant_t)); p0 += sizeof(gq_quant_t);
}
}
}
}
}
-void mul_mat_vec_gq_1(
+void mul_mat_gq_1(
const void * src0,
const void * src1,
float * dst,
float s0[QB + 1];
float s1[QB + 1];
- gq_t m0[QB + 1];
- gq_t m1[QB + 1];
+ gq_quant_t m0[QB + 1];
+ gq_quant_t m1[QB + 1];
for (int ir0 = 0; ir0 < m; ir0++) {
for (int ir1 = 0; ir1 < n; ir1++) {
float sumf = 0.0;
- const char * restrict pp0 = p0 + ir0*((2*sizeof(float) + (QK/gq_t_bits)*QB*sizeof(gq_t))*(k/QK));
- const char * restrict pp1 = p1 + ir1*((2*sizeof(float) + (QK/gq_t_bits)*QB*sizeof(gq_t))*(k/QK));
+ const char * restrict pp0 = p0 + ir0*((2*sizeof(float) + (QK/gq_t_bits)*QB*sizeof(gq_quant_t))*(k/QK));
+ const char * restrict pp1 = p1 + ir1*((2*sizeof(float) + (QK/gq_t_bits)*QB*sizeof(gq_quant_t))*(k/QK));
for (int i = 0; i < kp/QK; i++) {
float min0, d0;
s1[b + 1] = d1*(1 << b);
}
- m0[0] = -1LL;
- m1[0] = -1LL;
+ m0[0] = -1ULL;
+ m1[0] = -1ULL;
for (int s = 0; s < QK/gq_t_bits; ++s) {
for (int b = 0; b < QB; b++) {
- memcpy(&m0[b + 1], pp0, sizeof(gq_t)); pp0 += sizeof(gq_t);
- memcpy(&m1[b + 1], pp1, sizeof(gq_t)); pp1 += sizeof(gq_t);
+ memcpy(&m0[b + 1], pp0, sizeof(gq_quant_t)); pp0 += sizeof(gq_quant_t);
+ memcpy(&m1[b + 1], pp1, sizeof(gq_quant_t)); pp1 += sizeof(gq_quant_t);
}
for (int q0 = 0; q0 < QB + 1; q0++) {
// method 2
//
-void quantize_2(const float * src, void * dst, int n, int k) {
- char * p0 = dst;
+static inline int quantize_2_blocks_per_row(int k) {
+ return k/QK;
+}
- gq_t pp[QB];
+static inline int quantize_2_quants_per_block() {
+ return QK/gq_t_bits;
+}
- for (int j = 0; j < n; j++) {
- for (int i = 0; i < k/QK; i++) {
- float min = FLT_MAX;
- float max = -FLT_MAX;
+static inline int quantize_2_row_size(int k) {
+ const int nb = quantize_2_blocks_per_row(k);
+ const int nq = quantize_2_quants_per_block();
- // find min/max
-#ifdef __ARM_NEON
- {
- float32x4_t minv = vdupq_n_f32(FLT_MAX);
- float32x4_t maxv = vdupq_n_f32(-FLT_MAX);
+ return nb*(2*sizeof(gq_scale_t) + nq*QB*sizeof(gq_quant_t));
+}
- for (int l = 0; l < QK; l += 4) {
- float32x4_t v = vld1q_f32(src + j*k + i*QK + l);
- minv = vminq_f32(minv, v);
- maxv = vmaxq_f32(maxv, v);
- }
+void quantize_2_row(const float * restrict src, void * restrict dst, int k) {
+ assert(k % QK == 0);
- float32x2_t minv32 = vpmin_f32(vget_low_f32(minv), vget_high_f32(minv));
- float32x2_t maxv32 = vpmax_f32(vget_low_f32(maxv), vget_high_f32(maxv));
+ const int nb = quantize_2_blocks_per_row(k);
+ const int nq = quantize_2_quants_per_block();
- min = MIN(vget_lane_f32(minv32, 0), vget_lane_f32(minv32, 1));
- max = MAX(vget_lane_f32(maxv32, 0), vget_lane_f32(maxv32, 1));
+ gq_scale_t * restrict pm = (gq_scale_t *) (dst);
+ gq_scale_t * restrict pd = (gq_scale_t *) (pm + nb);
+ gq_quant_t * restrict pb = (gq_quant_t *) (pd + nb);
- //printf("SIMD min/max: %f %f\n", min, max);
- }
-#else
- {
- for (int l = 0; l < QK; l++) {
- const float v = src[j*k + i*QK + l];
- if (v < min) min = v;
- if (v > max) max = v;
- }
+ gq_quant_t pp[QB];
- //printf("NORM min/max: %f %f\n", min, max);
- }
-#endif
+ for (int i = 0; i < nb; i++) {
+ float min = FLT_MAX;
+ float max = -FLT_MAX;
- const float d = (max - min) / ((1 << QB) - 1);
- const float id = d ? 1.0/d : 0.0;
-
- memcpy(p0, &min, sizeof(float)); p0 += sizeof(float);
- memcpy(p0, &d, sizeof(float)); p0 += sizeof(float);
+ for (int l = 0; l < QK; l++) {
+ const float v = src[i*QK + l];
+ if (v < min) min = v;
+ if (v > max) max = v;
+ }
- //printf("min/max/d/id: %f %f %f %f\n", min, max, d, id);
+ const float d = (max - min) / ((1 << QB) - 1);
+ const float id = d ? 1.0/d : 0.0;
- for (int s = 0; s < QK/gq_t_bits; ++s) {
- memset(pp, 0, sizeof(pp));
+ pm[i] = GGML_FP32_TO_GQ(min);
+ pd[i] = GGML_FP32_TO_GQ(d);
- for (int l = 0; l < gq_t_bits; l++) {
- const float v = src[j*k + i*QK + s*gq_t_bits + l];
- const uint8_t q = (v - min)*id;
+ for (int s = 0; s < nq; ++s) {
+ memset(pp, 0, sizeof(pp));
- for (int b = 0; b < QB; b++) {
- pp[b] |= q & (1 << b) ? (1LL << l) : 0;
- }
- }
+ for (int l = 0; l < gq_t_bits; l++) {
+ const float v = src[i*QK + s*gq_t_bits + l];
+ const uint8_t q = (v - min)*id;
for (int b = 0; b < QB; b++) {
- memcpy(p0, &pp[b], sizeof(gq_t)); p0 += sizeof(gq_t);
+ pp[b] |= q & (1 << b) ? (1ULL << l) : 0;
}
}
+
+ for (int b = 0; b < QB; b++) {
+ pb[i*nq*QB + s*QB + b] = pp[b];
+ }
}
}
}
-void mul_mat_vec_gq_2(
- const void * src0,
- const void * src1,
- float * dst,
- int m, int n, int k) {
- const int kp = k & ~(gq_t_bits - 1);
-
- const char * restrict p0 = src0;
- const char * restrict p1 = src1;
+// reimplementation of quantize_2 using quantize_2_row
+void quantize_2(const float * restrict src, char * restrict dst, int n, int k) {
+ assert(k % QK == 0);
- float s0[QB + 1];
- float s1[QB + 1];
+ for (int j = 0; j < n; j++) {
+ quantize_2_row(src + j*k, dst, k);
+ dst = (char *) dst + quantize_2_row_size(k);
+ }
+}
- gq_t m0[QB + 1];
- gq_t m1[QB + 1];
+void vec_dot_gq_2(const int n, float * restrict s, const void * restrict x, const void * restrict y) {
+ float sumf[(QB + 1)*(QB + 1)];
+ memset(sumf, 0, sizeof(sumf));
- for (int ir0 = 0; ir0 < m; ir0++) {
- for (int ir1 = 0; ir1 < n; ir1++) {
- float sumf = 0.0;
+ const int nb = quantize_2_blocks_per_row(n);
+ const int nq = quantize_2_quants_per_block();
- const char * restrict pp0 = p0 + ir0*((2*sizeof(float) + (QK/gq_t_bits)*QB*sizeof(gq_t))*(k/QK));
- const char * restrict pp1 = p1 + ir1*((2*sizeof(float) + (QK/gq_t_bits)*QB*sizeof(gq_t))*(k/QK));
+ const gq_scale_t * restrict pm0 = (const gq_scale_t *) x;
+ const gq_scale_t * restrict pm1 = (const gq_scale_t *) y;
- for (int i = 0; i < kp/QK; i++) {
- float min0, d0;
- memcpy(&min0, pp0, sizeof(float)); pp0 += sizeof(float);
- memcpy(&d0, pp0, sizeof(float)); pp0 += sizeof(float);
+ const gq_scale_t * restrict pd0 = pm0 + nb;
+ const gq_scale_t * restrict pd1 = pm1 + nb;
- float min1, d1;
- memcpy(&min1, pp1, sizeof(float)); pp1 += sizeof(float);
- memcpy(&d1, pp1, sizeof(float)); pp1 += sizeof(float);
-
- //printf("min0/d0 = %f %f | min1/d1 = %f %f\n", min0, d0, min1, d1);
+ const gq_quant_t * restrict pb0 = (const gq_quant_t *) (pd0 + nb);
+ const gq_quant_t * restrict pb1 = (const gq_quant_t *) (pd1 + nb);
#if 1
- // >>> General case for any QB
+ float s0[QB + 1];
+ float s1[QB + 1];
- s0[0] = min0;
- s1[0] = min1;
+ for (int i = 0; i < nb; i++) {
+ const float m0 = GGML_GQ_TO_FP32(pm0[i]);
+ const float d0 = GGML_GQ_TO_FP32(pd0[i]);
- for (int b = 0; b < QB; b++) {
- s0[b + 1] = d0*(1 << b);
- s1[b + 1] = d1*(1 << b);
- }
+ const float m1 = GGML_GQ_TO_FP32(pm1[i]);
+ const float d1 = GGML_GQ_TO_FP32(pd1[i]);
- m0[0] = -1LL;
- m1[0] = -1LL;
+ s0[0] = m0;
+ s1[0] = m1;
- for (int s = 0; s < QK/gq_t_bits; ++s) {
- for (int b = 0; b < QB; b++) {
- memcpy(&m0[b + 1], pp0, sizeof(gq_t)); pp0 += sizeof(gq_t);
- memcpy(&m1[b + 1], pp1, sizeof(gq_t)); pp1 += sizeof(gq_t);
- }
+ for (int b = 0; b < QB; b++) {
+ s0[b + 1] = d0*(1 << b);
+ s1[b + 1] = d1*(1 << b);
+ }
- for (int q0 = 0; q0 < QB + 1; q0++) {
- for (int q1 = 0; q1 < QB + 1; q1++) {
- sumf += s0[q0]*s1[q1]*__builtin_popcountll(m0[q0] & m1[q1]);
- }
- }
+ for (int s = 0; s < nq; ++s) {
+ for (int q0 = 0; q0 < QB + 1; q0++) {
+ const gq_quant_t mm0 = q0 ? pb0[i*nq*QB + s*QB + q0 - 1] : -1ULL;
+ for (int q1 = 0; q1 < QB + 1; q1++) {
+ const gq_quant_t mm1 = q1 ? pb1[i*nq*QB + s*QB + q1 - 1] : -1ULL;
+ sumf[q0*(QB + 1) + q1] += s0[q0]*s1[q1]*__builtin_popcountll(mm0 & mm1);
}
+ }
+ }
+ }
+#else
+ // SIMD-ify with the assumptions:
+ // - nb is a multiple of 4
+ // - gq_scale_t is float
+ // - gq_quant_t is uint64_t
+ // - QB == 7
+ assert(nb % 4 == 0);
+
+#ifdef __ARM_NEON
#else
+ // TODO
#endif
- }
- dst[ir0*n + ir1] = sumf;
+#endif
+
+ for (int q0 = 0; q0 < QB + 1; q0++) {
+ for (int q1 = 1; q1 < QB + 1; q1++) {
+ sumf[q0*(QB + 1)] += sumf[q0*(QB + 1) + q1];
}
}
+
+ *s = sumf[0];
+ for (int q0 = 1; q0 < QB + 1; q0++) {
+ *s += sumf[q0*(QB + 1)];
+ }
+}
+
+// use vec_dot_gq_2 to compute the dot product of two rows
+void mul_mat_gq_2(
+ const void * src0,
+ const void * src1, // transposed
+ float * dst,
+ int m, int n, int k) {
+ assert(k % QK == 0);
+
+ const int nb = quantize_2_blocks_per_row(k);
+ const int nq = quantize_2_quants_per_block();
+
+ for (int ir0 = 0; ir0 < m; ir0++) {
+ for (int ir1 = 0; ir1 < n; ir1++) {
+ vec_dot_gq_2(k, dst + ir1, src0, src1);
+ src1 = (const char *) src1 + quantize_2_row_size(k);
+ }
+ src0 = (const char *) src0 + quantize_2_row_size(k);
+ src1 = (const char *) src1 - n*quantize_2_row_size(k);
+
+ dst = (float *) dst + n;
+ }
}
int main(int argc, const char ** argv) {
- assert(sizeof(gq_t)*8 == gq_t_bits);
+ assert(sizeof(gq_quant_t)*8 == gq_t_bits);
float * src0 = (float *)malloc(sizeof(float)*M*K);
float * src1 = (float *)malloc(sizeof(float)*N*K);
src1[i] = rand() / (float)RAND_MAX;
}
- void * src0_gq = calloc(1, (2*sizeof(float) + (QK/gq_t_bits)*QB*sizeof(gq_t))*(K/QK)*M);
- void * src1_gq = calloc(1, (2*sizeof(float) + (QK/gq_t_bits)*QB*sizeof(gq_t))*(K/QK)*N);
+ void * src0_gq = calloc(1, quantize_2_row_size(K)*M);
+ void * src1_gq = calloc(1, quantize_2_row_size(K)*N);
const size_t sizef16 = sizeof(ggml_fp16_t)*M*K + sizeof(ggml_fp16_t)*N*K;
- const size_t sizegq = (2*sizeof(float) + (QK/gq_t_bits)*QB*sizeof(gq_t))*(K/QK)*M +
- (2*sizeof(float) + (QK/gq_t_bits)*QB*sizeof(gq_t))*(K/QK)*N;
+ const size_t sizegq = quantize_2_row_size(K)*M + quantize_2_row_size(K)*N;
printf("compression: %f\n", (float)sizegq/sizef16);
double sum = 0.0f;
for (int i = 0; i < nIter; i++) {
if (method == 0) {
- mul_mat_vec_f32_naive(src0, src1, dst, M, N, K);
+ mul_mat_f32_naive(src0, src1, dst, M, N, K);
}
if (method == 1) {
- mul_mat_vec_gq_1(src0_gq, src1_gq, dst, M, N, K);
+ mul_mat_gq_1(src0_gq, src1_gq, dst, M, N, K);
}
if (method == 2) {
- mul_mat_vec_gq_1(src0_gq, src1_gq, dst, M, N, K);
+ mul_mat_gq_2(src0_gq, src1_gq, dst, M, N, K);
}
}
overview / pkg / ggml / sources / ggml / commitdiff