happyz
/
llama.cpp
mirror of https://github.com/ggerganov/llama.cpp.git
1
0
Fork 0
Code Issues Packages Projects Releases Wiki Activity

ggml-cpu: refactor; add rvv checks

This commit is contained in:
taimur-10x 2026-03-18 17:18:56 +05:00
parent 4b12d409ff
commit 05a5425e4e
1 changed files with 38 additions and 12 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

50
ggml/src/ggml-cpu/arch/riscv/quants.c
View File

@ -121,10 +121,10 @@ void quantize_row_q8_1(const float * GGML_RESTRICT x, void * GGML_RESTRICT vy, i
void quantize_row_q8_K(const float * GGML_RESTRICT x, void * GGML_RESTRICT y, int64_t k) {
assert(k % QK_K == 0);
block_q8_K * y_blocks = (block_q8_K *)y;
size_t nb = k / QK_K;
#if defined(__riscv_v_intrinsic)
#if defined __riscv_v_intrinsic
block_q8_K * y_blocks = (block_q8_K *)y;
const size_t vlmax_f32m8 = __riscv_vsetvlmax_e32m8();
for (size_t i = 0; i < nb; i++) {
@ -2058,6 +2058,7 @@ void ggml_vec_dot_q6_K_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const voi
#endif
}
#if defined __riscv_v_intrinsic
static NOINLINE void ggml_vec_dot_iq1_s_q8_K_vl128(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, size_t bx, const void * GGML_RESTRICT vy, size_t by, int nrc) {
assert(n % QK_K == 0);
assert(nrc == 1);
@ -2265,6 +2266,7 @@ static NOINLINE void ggml_vec_dot_iq1_s_q8_K_vl256(int n, float * GGML_RESTRICT
*s = sumf;
}
#endif
void ggml_vec_dot_iq1_s_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, size_t bx, const void * GGML_RESTRICT vy, size_t by, int nrc) {
#if defined __riscv_v_intrinsic
@ -2284,6 +2286,7 @@ void ggml_vec_dot_iq1_s_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const vo
#endif
}
#if defined __riscv_v_intrinsic
static NOINLINE void ggml_vec_dot_iq1_m_q8_K_vl128(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, size_t bx, const void * GGML_RESTRICT vy, size_t by, int nrc) {
assert(n % QK_K == 0);
assert(nrc == 1);
@ -2563,6 +2566,7 @@ static NOINLINE void ggml_vec_dot_iq1_m_q8_K_vl256(int n, float * GGML_RESTRICT
*s = sumf;
}
#endif
void ggml_vec_dot_iq1_m_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, size_t bx, const void * GGML_RESTRICT vy, size_t by, int nrc) {
#if defined __riscv_v_intrinsic
@ -2582,6 +2586,7 @@ void ggml_vec_dot_iq1_m_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const vo
#endif
}
#if defined __riscv_v_intrinsic
static const uint8_t sign_gather_indices_arr[64] = {
0,0,0,0,0,0,0,0, 1,1,1,1,1,1,1,1, 2,2,2,2,2,2,2,2, 3,3,3,3,3,3,3,3,
4,4,4,4,4,4,4,4, 5,5,5,5,5,5,5,5, 6,6,6,6,6,6,6,6, 7,7,7,7,7,7,7,7
@ -2784,6 +2789,7 @@ static NOINLINE void ggml_vec_dot_iq2_s_q8_K_vl256(int n, float * GGML_RESTRICT
}
*s = 0.125f * sumf;
}
#endif
void ggml_vec_dot_iq2_s_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, size_t bx, const void * GGML_RESTRICT vy, size_t by, int nrc) {
#if defined __riscv_v_intrinsic
@ -2803,7 +2809,7 @@ void ggml_vec_dot_iq2_s_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const vo
#endif
}
#if defined(__riscv_v)
#if defined __riscv_v_intrinsic
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,
@ -2838,7 +2844,6 @@ 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,
};
#endif
static NOINLINE void ggml_vec_dot_iq2_xs_q8_K_vl128(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, size_t bx, const void * GGML_RESTRICT vy, size_t by, int nrc) {
assert(n % QK_K == 0);
@ -2991,6 +2996,7 @@ static NOINLINE void ggml_vec_dot_iq2_xs_q8_K_vl256(int n, float * GGML_RESTRICT
}
*s = 0.125f * sumf;
}
#endif
void ggml_vec_dot_iq2_xs_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, size_t bx, const void * GGML_RESTRICT vy, size_t by, int nrc) {
#if defined __riscv_v_intrinsic
@ -3010,6 +3016,7 @@ void ggml_vec_dot_iq2_xs_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const v
#endif
}
#if defined __riscv_v_intrinsic
static NOINLINE void ggml_vec_dot_iq2_xxs_q8_K_vl128(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, size_t bx, const void * GGML_RESTRICT vy, size_t by, int nrc) {
assert(n % QK_K == 0);
assert(nrc == 1);
@ -3194,6 +3201,7 @@ static NOINLINE void ggml_vec_dot_iq2_xxs_q8_K_vl256(int n, float * GGML_RESTRIC
}
*s = 0.125f * sumf;
}
#endif
void ggml_vec_dot_iq2_xxs_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, size_t bx, const void * GGML_RESTRICT vy, size_t by, int nrc) {
#if defined __riscv_v_intrinsic
@ -3201,15 +3209,16 @@ void ggml_vec_dot_iq2_xxs_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const
case 128:
ggml_vec_dot_iq2_xxs_q8_K_vl128(n, s, bs, vx, bx, vy, by, nrc);
break;
default:
default: // 256 and above
ggml_vec_dot_iq2_xxs_q8_K_vl256(n, s, bs, vx, bx, vy, by, nrc);
break;
}
#else
ggml_vec_dot_iq2_xxs_q8_K(n, s, bs, vx, bx, vy, by, nrc);
ggml_vec_dot_iq2_xxs_q8_K_generic(n, s, bs, vx, bx, vy, by, nrc);
#endif
}
#if defined __riscv_v_intrinsic
static NOINLINE void ggml_vec_dot_iq3_s_q8_K_vl128(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, size_t bx, const void * GGML_RESTRICT vy, size_t by, int nrc) {
assert(n % QK_K == 0);
UNUSED(nrc); UNUSED(bx); UNUSED(by); UNUSED(bs);
@ -3399,6 +3408,7 @@ static NOINLINE void ggml_vec_dot_iq3_s_q8_K_vl256(int n, float * GGML_RESTRICT
}
*s = sumf;
}
#endif
void ggml_vec_dot_iq3_s_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, size_t bx, const void * GGML_RESTRICT vy, size_t by, int nrc) {
#if defined __riscv_v_intrinsic
@ -3418,6 +3428,7 @@ void ggml_vec_dot_iq3_s_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const vo
#endif
}
#if defined __riscv_v_intrinsic
static NOINLINE void ggml_vec_dot_iq3_xxs_q8_K_vl128(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, size_t bx, const void * GGML_RESTRICT vy, size_t by, int nrc) {
assert(n % QK_K == 0);
UNUSED(nrc); UNUSED(bx); UNUSED(by); UNUSED(bs);
@ -3603,6 +3614,7 @@ static NOINLINE void ggml_vec_dot_iq3_xxs_q8_K_vl256(int n, float * GGML_RESTRIC
}
*s = 0.25f * sumf;
}
#endif
void ggml_vec_dot_iq3_xxs_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, size_t bx, const void * GGML_RESTRICT vy, size_t by, int nrc) {
#if defined __riscv_v_intrinsic
@ -3622,6 +3634,7 @@ void ggml_vec_dot_iq3_xxs_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const
#endif
}
#if defined __riscv_v_intrinsic
static NOINLINE void ggml_vec_dot_iq4_nl_q8_0_vl128(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, size_t bx, const void * GGML_RESTRICT vy, size_t by, int nrc) {
assert(nrc == 1);
UNUSED(nrc);
@ -3733,6 +3746,7 @@ static NOINLINE void ggml_vec_dot_iq4_nl_q8_0_vl256(int n, float * GGML_RESTRICT
*s = sumf;
}
#endif
void ggml_vec_dot_iq4_nl_q8_0(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, size_t bx, const void * GGML_RESTRICT vy, size_t by, int nrc) {
#if defined __riscv_v_intrinsic
@ -3740,7 +3754,7 @@ void ggml_vec_dot_iq4_nl_q8_0(int n, float * GGML_RESTRICT s, size_t bs, const v
case 128:
ggml_vec_dot_iq4_nl_q8_0_vl128(n, s, bs, vx, bx, vy, by, nrc);
break;
default:
default: // 256 and above
ggml_vec_dot_iq4_nl_q8_0_vl256(n, s, bs, vx, bx, vy, by, nrc);
break;
}
@ -3749,6 +3763,7 @@ void ggml_vec_dot_iq4_nl_q8_0(int n, float * GGML_RESTRICT s, size_t bs, const v
#endif
}
#if defined __riscv_v_intrinsic
static NOINLINE void ggml_vec_dot_iq4_xs_q8_K_vl128(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, size_t bx, const void * GGML_RESTRICT vy, size_t by, int nrc) {
assert(nrc == 1);
UNUSED(nrc);
@ -3894,6 +3909,7 @@ static NOINLINE void ggml_vec_dot_iq4_xs_q8_K_vl256(int n, float * GGML_RESTRICT
*s = sumf;
}
#endif
void ggml_vec_dot_iq4_xs_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, size_t bx, const void * GGML_RESTRICT vy, size_t by, int nrc) {
#if defined __riscv_v_intrinsic
@ -3913,6 +3929,7 @@ void ggml_vec_dot_iq4_xs_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const v
#endif
}
#if defined __riscv_v_intrinsic
static NOINLINE void ggml_vec_dot_tq1_0_q8_K_vl128(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, size_t bx, const void * GGML_RESTRICT vy, size_t by, int nrc) {
assert(nrc == 1);
UNUSED(nrc);
@ -4106,14 +4123,16 @@ static NOINLINE void ggml_vec_dot_tq1_0_q8_K_vl256(int n, float * GGML_RESTRICT
suml3 = __riscv_vmul_vv_i16m1(__riscv_vreinterpret_v_u16m1_i16m1(__riscv_vsub_vx_u16m1(tq0, 1, vl)), q80, vl);
}
vint16m2_t sumb = __riscv_vadd_vv_i16m2(suml1, __riscv_vlmul_ext_v_i16m1_i16m2(__riscv_vadd_vv_i16m1(suml2, suml3, 16)), 16);
vint16m1_t sumb = __riscv_vadd_vv_i16m1(__riscv_vget_v_i16m2_i16m1(suml1, 0), __riscv_vget_v_i16m2_i16m1(suml1, 1), 16);
sumb = __riscv_vadd_vv_i16m1(sumb, __riscv_vadd_vv_i16m1(suml2, suml3, 16), 16);
vint32m1_t sum = __riscv_vwredsum_vs_i16m2_i32m1(sumb, __riscv_vmv_v_x_i32m1(0, 1), 32);
vint32m1_t sum = __riscv_vwredsum_vs_i16m1_i32m1(sumb, __riscv_vmv_v_x_i32m1(0, 1), 16);
sumf += __riscv_vmv_x_s_i32m1_i32(sum) * y[i].d * GGML_CPU_FP16_TO_FP32(x[i].d);
}
*s = sumf;
}
#endif
void ggml_vec_dot_tq1_0_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, size_t bx, const void * GGML_RESTRICT vy, size_t by, int nrc) {
#if defined __riscv_v_intrinsic
@ -4121,15 +4140,19 @@ void ggml_vec_dot_tq1_0_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const vo
case 128:
ggml_vec_dot_tq1_0_q8_K_vl128(n, s, bs, vx, bx, vy, by, nrc);
break;
default:
case 256:
ggml_vec_dot_tq1_0_q8_K_vl256(n, s, bs, vx, bx, vy, by, nrc);
break;
default:
ggml_vec_dot_tq1_0_q8_K_generic(n, s, bs, vx, bx, vy, by, nrc);
break;
}
#else
ggml_vec_dot_tq1_0_q8_K_generic(n, s, bs, vx, bx, vy, by, nrc);
#endif
}
#if defined __riscv_v_intrinsic
static NOINLINE void ggml_vec_dot_tq2_0_q8_K_vl128(const int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, size_t bx, const void * GGML_RESTRICT vy, size_t by, int nrc) {
assert(n % QK_K == 0);
assert(nrc == 1);
@ -4282,6 +4305,7 @@ static NOINLINE void ggml_vec_dot_tq2_0_q8_K_vl256(int n, float * GGML_RESTRICT
*s = sumf;
}
#endif
void ggml_vec_dot_tq2_0_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, size_t bx, const void * GGML_RESTRICT vy, size_t by, int nrc) {
#if defined __riscv_v_intrinsic
@ -4301,6 +4325,7 @@ void ggml_vec_dot_tq2_0_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const vo
#endif
}
#if defined __riscv_v_intrinsic
static NOINLINE void ggml_vec_dot_mxfp4_q8_0_vl128(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, size_t bx, const void * GGML_RESTRICT vy, size_t by, int nrc) {
assert(nrc == 1);
UNUSED(nrc);
@ -4412,6 +4437,7 @@ static NOINLINE void ggml_vec_dot_mxfp4_q8_0_vl256(int n, float * GGML_RESTRICT
*s = sumf;
}
#endif
void ggml_vec_dot_mxfp4_q8_0(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, size_t bx, const void * GGML_RESTRICT vy, size_t by, int nrc) {
#if defined __riscv_v_intrinsic
@ -4419,11 +4445,11 @@ void ggml_vec_dot_mxfp4_q8_0(int n, float * GGML_RESTRICT s, size_t bs, const vo
case 128:
ggml_vec_dot_mxfp4_q8_0_vl128(n, s, bs, vx, bx, vy, by, nrc);
break;
default:
default: // 256 and above
ggml_vec_dot_mxfp4_q8_0_vl256(n, s, bs, vx, bx, vy, by, nrc);
break;
}
#else
return ggml_vec_dot_mxfp4_q8_0_generic(n, s, bs, vx, bx, vy, by, nrc);
ggml_vec_dot_mxfp4_q8_0_generic(n, s, bs, vx, bx, vy, by, nrc);
#endif
}