vrr/koboldcpp
2
0
Fork 0
mirror of https://github.com/LostRuins/koboldcpp.git synced 2025-09-11 01:24:36 +00:00
Code Issues Projects Releases Packages Wiki Activity Actions

Merge branch 'master' into concedo_experimental

Browse source 
# Conflicts:
#	.devops/nix/scope.nix
#	.github/workflows/nix-ci-aarch64.yml
#	.github/workflows/nix-ci.yml
#	README.md
#	scripts/sync-ggml.last
This commit is contained in:
Concedo 2024-02-24 18:30:51 +08:00
commit 359a14d3c2
10 changed files with 251 additions and 72 deletions
Download patch file Download diff file Expand all files Collapse all files

65
ggml-quants.c
View file

@ -440,6 +440,30 @@ inline static ggml_int8x16x4_t ggml_vld1q_s8_x4(const int8_t * ptr) {
return res;
}
// NOTE: not tested
inline static int8x16_t ggml_vqtbl1q_s8(int8x16_t a, uint8x16_t b) {
int8x16_t res;
res[ 0] = a[b[ 0]];
res[ 1] = a[b[ 1]];
res[ 2] = a[b[ 2]];
res[ 3] = a[b[ 3]];
res[ 4] = a[b[ 4]];
res[ 5] = a[b[ 5]];
res[ 6] = a[b[ 6]];
res[ 7] = a[b[ 7]];
res[ 8] = a[b[ 8]];
res[ 9] = a[b[ 9]];
res[10] = a[b[10]];
res[11] = a[b[11]];
res[12] = a[b[12]];
res[13] = a[b[13]];
res[14] = a[b[14]];
res[15] = a[b[15]];
return res;
}
#else
#define ggml_int16x8x2_t int16x8x2_t
@ -453,6 +477,7 @@ inline static ggml_int8x16x4_t ggml_vld1q_s8_x4(const int8_t * ptr) {
#define ggml_vld1q_u8_x4 vld1q_u8_x4
#define ggml_vld1q_s8_x2 vld1q_s8_x2
#define ggml_vld1q_s8_x4 vld1q_s8_x4
#define ggml_vqtbl1q_s8 vqtbl1q_s8
#endif
@ -5631,8 +5656,8 @@ void ggml_vec_dot_q2_K_q8_K(int n, float * restrict s, size_t bs, const void * r
for (int i = 0; i < nb; ++i) {
const float d = y[i].d * (float)x[i].d;
const float dmin = -y[i].d * (float)x[i].dmin;
const float d = y[i].d * GGML_FP16_TO_FP32(x[i].d);
const float dmin = -y[i].d * GGML_FP16_TO_FP32(x[i].dmin);
const uint8_t * restrict q2 = x[i].qs;
const int8_t * restrict q8 = y[i].qs;
@ -5781,8 +5806,8 @@ void ggml_vec_dot_q2_K_q8_K(int n, float * restrict s, size_t bs, const void * r
for (int i = 0; i < nb; ++i) {
const float d = y[i].d * (float)x[i].d;
const float dmin = -y[i].d * (float)x[i].dmin;
const float d = y[i].d * GGML_FP16_TO_FP32(x[i].d);
const float dmin = -y[i].d * GGML_FP16_TO_FP32(x[i].dmin);
const uint8_t * restrict q2 = x[i].qs;
const int8_t * restrict q8 = y[i].qs;
@ -6435,7 +6460,7 @@ void ggml_vec_dot_q3_K_q8_K(int n, float * restrict s, size_t bs, const void * r
int32_t isum = -4*(scales[0] * y[i].bsums[0] + scales[2] * y[i].bsums[1] + scales[1] * y[i].bsums[2] + scales[3] * y[i].bsums[3]);
const float d = y[i].d * (float)x[i].d;
const float d = y[i].d * GGML_FP16_TO_FP32(x[i].d);
const uint8x16_t htmp = vcombine_u8(hbits, vshr_n_u8(hbits, 1));
q3h.val[0] = vandq_u8(mh, vshlq_n_u8(htmp, 2));
@ -6637,7 +6662,7 @@ void ggml_vec_dot_q3_K_q8_K(int n, float * restrict s, size_t bs, const void * r
int32_t isum = -4*(scales[0] * y[i].bsums[0] + scales[2] * y[i].bsums[1] + scales[1] * y[i].bsums[2] + scales[3] * y[i].bsums[3]);
const float d = y[i].d * (float)x[i].d;
const float d = y[i].d * GGML_FP16_TO_FP32(x[i].d);
vint32m1_t vzero = __riscv_vmv_v_x_i32m1(0, 1);
@ -7140,9 +7165,9 @@ void ggml_vec_dot_q4_K_q8_K(int n, float * restrict s, size_t bs, const void * r
aux16[1] = (a[0] >> 4) & 0x0f0f;
const int32_t summi = scales[2] * (y[i].bsums[0] + y[i].bsums[1]) + scales[3] * (y[i].bsums[2] + y[i].bsums[3]);
sum_mins += y[i].d * (float)x[i].d[1] * summi;
sum_mins += y[i].d * GGML_FP16_TO_FP32(x[i].d[1]) * summi;
const float d = y[i].d * (float)x[i].d[0];
const float d = y[i].d * GGML_FP16_TO_FP32(x[i].d[0]);
const ggml_uint8x16x2_t q4bits = ggml_vld1q_u8_x2(q4);
@ -7800,7 +7825,7 @@ void ggml_vec_dot_q5_K_q8_K(int n, float * restrict s, size_t bs, const void * r
for (int i = 0; i < nb; ++i) {
const float d = y[i].d * (float)x[i].d;
const float d = y[i].d * GGML_FP16_TO_FP32(x[i].d);
const int8_t * sc = x[i].scales;
const uint8_t * restrict q5 = x[i].qs;
@ -7942,7 +7967,7 @@ void ggml_vec_dot_q5_K_q8_K(int n, float * restrict s, size_t bs, const void * r
for (int i = 0; i < nb; ++i) {
const float d = y[i].d * (float)x[i].d;
const float d = y[i].d * GGML_FP16_TO_FP32(x[i].d);
const int8_t * sc = x[i].scales;
const uint8_t * restrict q5 = x[i].qs;
@ -8510,7 +8535,7 @@ void ggml_vec_dot_q6_K_q8_K(int n, float * restrict s, size_t bs, const void * r
for (int i = 0; i < nb; ++i) {
const float d_all = (float)x[i].d;
const float d_all = GGML_FP16_TO_FP32(x[i].d);
const uint8_t * restrict q6 = x[i].ql;
const uint8_t * restrict qh = x[i].qh;
@ -8681,7 +8706,7 @@ void ggml_vec_dot_q6_K_q8_K(int n, float * restrict s, size_t bs, const void * r
for (int i = 0; i < nb; ++i) {
const float d_all = (float)x[i].d;
const float d_all = GGML_FP16_TO_FP32(x[i].d);
const uint8_t * restrict q6 = x[i].ql;
const uint8_t * restrict qh = x[i].qh;
@ -9335,7 +9360,7 @@ void ggml_vec_dot_iq1_s_q8_K (int n, float * GGML_RESTRICT s, size_t bs, const
uint16_t gindex[8];
uint16x8x2_t vindex;
int8x16x4_t q1b;
int8x16x4_t q8b;
ggml_int8x16x4_t q8b;
uint16x8x4_t scales;
int32x4x2_t sumi;
int32x4x2_t dotq;
@ -9500,7 +9525,6 @@ void ggml_vec_dot_iq4_nl_q8_0(int n, float * restrict s, size_t bs, const void *
float sumf = 0;
for (int ib = 0; ib < nb; ib += 2) {
q4bits.val[0] = vld1q_u8(x[ib+0].qs);
q4bits.val[1] = vld1q_u8(x[ib+1].qs);
q8b.val[0] = vld1q_s8(y[ib+0].qs);
@ -9508,16 +9532,17 @@ void ggml_vec_dot_iq4_nl_q8_0(int n, float * restrict s, size_t bs, const void *
q8b.val[2] = vld1q_s8(y[ib+1].qs);
q8b.val[3] = vld1q_s8(y[ib+1].qs + 16);
q4b.val[0] = vqtbl1q_s8(values, vandq_u8(q4bits.val[0], m4b));
q4b.val[1] = vqtbl1q_s8(values, vshrq_n_u8(q4bits.val[0], 4));
q4b.val[2] = vqtbl1q_s8(values, vandq_u8(q4bits.val[1], m4b));
q4b.val[3] = vqtbl1q_s8(values, vshrq_n_u8(q4bits.val[1], 4));
q4b.val[0] = ggml_vqtbl1q_s8(values, vandq_u8 (q4bits.val[0], m4b));
q4b.val[1] = ggml_vqtbl1q_s8(values, vshrq_n_u8(q4bits.val[0], 4));
q4b.val[2] = ggml_vqtbl1q_s8(values, vandq_u8 (q4bits.val[1], m4b));
q4b.val[3] = ggml_vqtbl1q_s8(values, vshrq_n_u8(q4bits.val[1], 4));
prod_1 = ggml_vdotq_s32(ggml_vdotq_s32(vdupq_n_s32(0), q4b.val[0], q8b.val[0]), q4b.val[1], q8b.val[1]);
prod_2 = ggml_vdotq_s32(ggml_vdotq_s32(vdupq_n_s32(0), q4b.val[2], q8b.val[2]), q4b.val[3], q8b.val[3]);
sumf += (float)x[ib+0].d * (float)y[ib+0].d * vaddvq_s32(prod_1) + (float)x[ib+1].d * (float)y[ib+1].d * vaddvq_s32(prod_2);
sumf +=
GGML_FP16_TO_FP32(x[ib+0].d) * GGML_FP16_TO_FP32(y[ib+0].d) * vaddvq_s32(prod_1) +
GGML_FP16_TO_FP32(x[ib+1].d) * GGML_FP16_TO_FP32(y[ib+1].d) * vaddvq_s32(prod_2);
}
*s = sumf;