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

fix xtheadvector builds

This commit is contained in:
Xiongchuan Tan 2026-01-19 16:05:51 +08:00
parent 344ce2e184
commit 0e373e245b
5 changed files with 26 additions and 18 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

2
ggml/src/ggml-cpu/arch/riscv/dispatch.cpp
View File

@ -9,7 +9,7 @@ extern "C" {
#include "kernels.inc"
}
#if defined(__riscv_v)
#if defined(__riscv_v) && __riscv_v >= 1000000
// helper macros for runtime kernel dispatch

24
ggml/src/ggml-cpu/arch/riscv/quants.c
View File

@ -32,7 +32,7 @@ void quantize_row_q8_0(const float * GGML_RESTRICT x, void * GGML_RESTRICT vy, i
block_q8_0 * GGML_RESTRICT y = vy;
#if defined(__riscv_v)
#if defined(__riscv_v) && __riscv_v >= 1000000
size_t vl = QK8_0;
@ -72,7 +72,7 @@ void quantize_row_q8_1(const float * GGML_RESTRICT x, void * GGML_RESTRICT vy, i
block_q8_1 * GGML_RESTRICT y = vy;
#if defined(__riscv_v)
#if defined(__riscv_v) && __riscv_v >= 1000000
size_t vl = QK8_1;
@ -118,7 +118,7 @@ void quantize_row_q8_1(const float * GGML_RESTRICT x, void * GGML_RESTRICT vy, i
//===================================== Dot products =================================
void ggml_vec_dot_q4_0_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)
#if defined(__riscv_v) && __riscv_v >= 1000000
const int qk = QK8_0;
const int nb = n / qk;
@ -173,7 +173,7 @@ void ggml_vec_dot_q4_0_q8_0(int n, float * GGML_RESTRICT s, size_t bs, const voi
}
void ggml_vec_dot_q4_1_q8_1(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)
#if defined(__riscv_v) && __riscv_v >= 1000000
const int qk = QK8_1;
const int nb = n / qk;
@ -224,7 +224,7 @@ void ggml_vec_dot_q4_1_q8_1(int n, float * GGML_RESTRICT s, size_t bs, const voi
}
void ggml_vec_dot_q5_0_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)
#if defined(__riscv_v) && __riscv_v >= 1000000
const int qk = QK8_0;
const int nb = n / qk;
@ -278,7 +278,7 @@ void ggml_vec_dot_q5_0_q8_0(int n, float * GGML_RESTRICT s, size_t bs, const voi
}
void ggml_vec_dot_q5_1_q8_1(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)
#if defined(__riscv_v) && __riscv_v >= 1000000
const int qk = QK8_1;
const int nb = n / qk;
@ -347,7 +347,7 @@ void ggml_vec_dot_q8_0_q8_0(int n, float * GGML_RESTRICT s, size_t bs, const voi
int ib = 0;
float sumf = 0;
#if defined(__riscv_v)
#if defined(__riscv_v) && __riscv_v >= 1000000
size_t vl = qk;
for (; ib < nb; ++ib) {
@ -488,7 +488,7 @@ void ggml_vec_dot_q2_K_q8_K_071(int n, float * GGML_RESTRICT s, size_t bs, const
*s = sumf;
}
#elif defined(__riscv_v)
#elif defined(__riscv_v) && __riscv_v >= 1000000
void ggml_vec_dot_q2_K_q8_K_256(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);
@ -885,7 +885,7 @@ void ggml_vec_dot_q3_K_q8_K_071(int n, float * GGML_RESTRICT s, size_t bs, const
*s = sumf;
}
#elif defined(__riscv_v)
#elif defined(__riscv_v) && __riscv_v >= 1000000
void ggml_vec_dot_q3_K_q8_K_256(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);
@ -1291,7 +1291,7 @@ void ggml_vec_dot_q4_K_q8_K_071(int n, float * GGML_RESTRICT s, size_t bs, const
*s = sumf;
}
#elif defined(__riscv_v)
#elif defined(__riscv_v) && __riscv_v >= 1000000
void ggml_vec_dot_q4_K_q8_K_256(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);
@ -1590,7 +1590,7 @@ void ggml_vec_dot_q4_K_q8_K_128(int n, float * GGML_RESTRICT s, size_t bs, const
#endif // ggml_vec_dot_q4_K_q8_K
#if defined __riscv_v
#if defined(__riscv_v) && __riscv_v >= 1000000
void ggml_vec_dot_q5_K_q8_K_128(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);
@ -1793,7 +1793,7 @@ void ggml_vec_dot_q6_K_q8_K_071(int n, float * GGML_RESTRICT s, size_t bs, const
*s = sumf;
}
#elif defined(__riscv_v)
#elif defined(__riscv_v) && __riscv_v >= 1000000
void ggml_vec_dot_q6_K_q8_K_256(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);

4
ggml/src/ggml-cpu/arch/riscv/repack.cpp
View File

@ -43,7 +43,7 @@ void ggml_gemv_q4_0_8x8_q8_0(int n, float * GGML_RESTRICT s, size_t bs, const vo
UNUSED(ncols_interleaved);
UNUSED(blocklen);
#if defined __riscv_v
#if defined(__riscv_v) && __riscv_v >= 1000000
if (__riscv_vlenb() >= QK4_0) {
const size_t vl = QK4_0;
@ -135,7 +135,7 @@ void ggml_gemm_q4_0_8x8_q8_0(int n, float * GGML_RESTRICT s, size_t bs, const vo
UNUSED(ncols_interleaved);
UNUSED(blocklen);
#if defined __riscv_v
#if defined(__riscv_v) && __riscv_v >= 1000000
if (__riscv_vlenb() >= QK4_0) {
const size_t vl = QK4_0;

8
ggml/src/ggml-cpu/llamafile/sgemm.cpp
View File

@ -391,6 +391,8 @@ template <> inline vfloat16m2_t load(const ggml_fp16_t *p) {
template <> inline vfloat16m4_t load(const ggml_fp16_t *p) {
return __riscv_vle16_v_f16m4(reinterpret_cast<const _Float16 *>(p), __riscv_vsetvlmax_e16m4());
}
#endif
#if defined(__riscv_v) && __riscv_v >= 1000000
template <> inline vfloat32m1_t load(const float *p) {
return __riscv_vle32_v_f32m1(p, __riscv_vsetvlmax_e32m1());
}
@ -432,6 +434,8 @@ template <> inline vfloat16m2_t set_zero() {
template <> inline vfloat16m4_t set_zero() {
return __riscv_vfmv_v_f_f16m4(0, __riscv_vsetvlmax_e16m4());
}
#endif
#if defined(__riscv_v) && __riscv_v >= 1000000
template <> inline vfloat32m1_t set_zero() {
return __riscv_vfmv_v_f_f32m1(0.0f, __riscv_vsetvlmax_e32m1());
}
@ -446,7 +450,7 @@ template <> inline vfloat32m8_t set_zero() {
}
#endif
#if defined(__riscv_v_intrinsic)
#if defined(__riscv_v) && __riscv_v >= 1000000
template <typename T> size_t vlmax() {
if constexpr (std::is_same_v<T, vfloat16mf2_t>) { return __riscv_vsetvlmax_e16mf2(); }
else if constexpr (std::is_same_v<T, vfloat16m1_t>) { return __riscv_vsetvlmax_e16m1(); }
@ -633,7 +637,7 @@ class tinyBLAS {
const int64_t ldc;
};
#if defined(__riscv_v_intrinsic)
#if defined(__riscv_v) && __riscv_v >= 1000000
template <typename D, typename V, typename TA, typename TB, typename TC>
class tinyBLAS_RVV {
public:

6
ggml/src/ggml-cpu/simd-mappings.h
View File

@ -97,6 +97,9 @@ extern "C" {
return r;
}
#elif defined(__riscv) && defined(__riscv_zfhmin)
// suppress _Float16 warnings
#pragma GCC diagnostic push
#pragma GCC diagnostic ignored "-Wpedantic"
static inline float riscv_compute_fp16_to_fp32(ggml_fp16_t h) {
_Float16 hf;
memcpy(&hf, &h, sizeof(ggml_fp16_t));
@ -109,6 +112,7 @@ extern "C" {
memcpy(&res, &hf, sizeof(ggml_fp16_t));
return res;
}
#pragma GCC diagnostic pop
#define GGML_CPU_COMPUTE_FP16_TO_FP32(x) riscv_compute_fp16_to_fp32(x)
#define GGML_CPU_COMPUTE_FP32_TO_FP16(x) riscv_compute_fp32_to_fp16(x)
@ -1171,7 +1175,7 @@ static inline void __lzs_f16cx4_store(ggml_fp16_t * x, float32x4_t v_y) {
#define GGML_F16_VEC_MUL GGML_F32x4_MUL
#define GGML_F16_VEC_REDUCE GGML_F32x4_REDUCE
#elif defined(__riscv_v_intrinsic)
#elif defined(__riscv_v) && __riscv_v >= 1000000
// compatible with vlen >= 128