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: add vlen switch for rvv vec_dot

This commit is contained in:
taimur-10x 2026-01-20 21:46:04 +05:00
parent b582ec427d
commit d0876893d1
1 changed files with 65 additions and 33 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

98
ggml/src/ggml-cpu/arch/riscv/quants.c
View File

@ -2175,7 +2175,7 @@ void ggml_vec_dot_iq2_s_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const vo
#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) {
static void ggml_vec_dot_iq3_s_q8_K_vl256(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);
@ -2187,7 +2187,6 @@ void ggml_vec_dot_iq3_s_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const vo
const int nb = n / QK_K;
#if defined __riscv_v_intrinsic
const uint64_t * grid64 = (const uint64_t *)iq3s_grid;
// --- Pre-load Constants ---
@ -2269,16 +2268,24 @@ void ggml_vec_dot_iq3_s_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const vo
sumf += sum_block * combined_scale;
}
*s = 0.125f * sumf;
#else
UNUSED(x);
UNUSED(y);
UNUSED(nb);
}
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
switch (__riscv_vlenb() * 8) {
case 256:
ggml_vec_dot_iq3_s_q8_K_vl256(n, s, bs, vx, bx, vy, by, nrc);
break;
default:
ggml_vec_dot_iq3_s_q8_K_generic(n, s, bs, vx, bx, vy, by, nrc);
break;
}
#else
ggml_vec_dot_iq3_s_q8_K_generic(n, s, bs, vx, bx, vy, by, nrc);
#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) {
static void ggml_vec_dot_tq1_0_q8_K_vl256(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);
UNUSED(bx);
@ -2290,7 +2297,6 @@ void ggml_vec_dot_tq1_0_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const vo
const int nb = n / QK_K;
#if defined __riscv_v_intrinsic
float sumf = 0.0f;
uint8_t pow[16] = {1, 1, 1, 1, 3, 3, 3, 3, 9, 9, 9, 9, 27, 27, 27, 27};
@ -2383,17 +2389,24 @@ void ggml_vec_dot_tq1_0_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const vo
}
*s = sumf;
}
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
switch (__riscv_vlenb() * 8) {
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
UNUSED(x);
UNUSED(y);
UNUSED(nb);
ggml_vec_dot_tq1_0_q8_K_generic(n, s, bs, vx, bx, vy, by, nrc);
#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) {
static void ggml_vec_dot_tq2_0_q8_K_vl256(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);
UNUSED(nrc);
@ -2406,7 +2419,6 @@ void ggml_vec_dot_tq2_0_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const vo
const int nb = n / QK_K;
#if defined __riscv_v_intrinsic
float sumf = 0.0f;
for (int i = 0; i < nb; ++i) {
int32_t sumi = 0;
@ -2463,17 +2475,24 @@ void ggml_vec_dot_tq2_0_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const vo
}
*s = sumf;
#else
UNUSED(x);
UNUSED(y);
UNUSED(nb);
ggml_vec_dot_tq2_0_q8_K_generic(n, s, bs, vx, bx, vy, by, nrc);
#endif // __riscv_vector
}
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) {
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
switch (__riscv_vlenb() * 8) {
case 256:
ggml_vec_dot_tq2_0_q8_K_vl256(n, s, bs, vx, bx, vy, by, nrc);
break;
default:
ggml_vec_dot_tq2_0_q8_K_generic(n, s, bs, vx, bx, vy, by, nrc);
break;
}
#else
ggml_vec_dot_tq2_0_q8_K_generic(n, s, bs, vx, bx, vy, by, nrc);
#endif
}
static void ggml_vec_dot_iq1_s_q8_K_vl256(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);
UNUSED(nrc);
@ -2486,7 +2505,6 @@ void ggml_vec_dot_iq1_s_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const vo
const int nb = n / QK_K;
#if defined __riscv_v_intrinsic
float sumf = 0;
for (int i = 0; i < nb; ++i) {
// Load qh once for the entire superblock.
@ -2568,16 +2586,24 @@ void ggml_vec_dot_iq1_s_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const vo
}
*s = sumf;
}
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
switch (__riscv_vlenb() * 8) {
case 256:
ggml_vec_dot_iq1_s_q8_K_vl256(n, s, bs, vx, bx, vy, by, nrc);
break;
default:
ggml_vec_dot_iq1_s_q8_K_generic(n, s, bs, vx, bx, vy, by, nrc);
break;
}
#else
UNUSED(x);
UNUSED(y);
UNUSED(nb);
ggml_vec_dot_iq1_s_q8_K_generic(n, s, bs, vx, bx, vy, by, nrc);
#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) {
static void ggml_vec_dot_iq1_m_q8_K_vl256(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);
UNUSED(nrc);
@ -2590,7 +2616,6 @@ void ggml_vec_dot_iq1_m_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const vo
const int nb = n / QK_K;
#if defined __riscv_v_intrinsic
iq1m_scale_t scale;
float sumf = 0.0f;
for (int i = 0; i < nb; ++i) {
@ -2683,12 +2708,19 @@ void ggml_vec_dot_iq1_m_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const vo
}
*s = sumf;
}
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
switch (__riscv_vlenb() * 8) {
case 256:
ggml_vec_dot_iq1_m_q8_K_vl256(n, s, bs, vx, bx, vy, by, nrc);
break;
default:
ggml_vec_dot_iq1_m_q8_K_generic(n, s, bs, vx, bx, vy, by, nrc);
break;
}
#else
UNUSED(x);
UNUSED(y);
UNUSED(nb);
ggml_vec_dot_iq1_m_q8_K_generic(n, s, bs, vx, bx, vy, by, nrc);
#endif
}