CS348Project/llama.cpp
5
0
Fork 0
mirror of https://github.com/ggml-org/llama.cpp.git synced 2025-10-31 08:51:55 +00:00
Code Issues Packages Projects Releases Wiki Activity

llamafile : support s390x SIMD instruction set (#14273)

Browse Source
This commit is contained in:
Aaron Teo
2025-06-19 17:48:54 +08:00
committed by GitHub
parent 10bb545c5b
commit faed5a5f5d
2 changed files with 59 additions and 1 deletions
Download Patch File Download Diff File Expand all files Collapse all files

55
ggml/src/ggml-cpu/llamafile/sgemm.cpp
View File

@@ -62,7 +62,7 @@
#define NOINLINE __attribute__((__noinline__))
#endif
#if defined(__ARM_NEON) || defined(__AVX512F__)
#if defined(__ARM_NEON) || defined(__AVX512F__) || defined(__VXE__) || defined(__VXE2__)
#define VECTOR_REGISTERS 32
#else
#define VECTOR_REGISTERS 16
@@ -109,6 +109,12 @@ inline float16x8_t sub(float16x8_t x, float16x8_t y) { return vsubq_f16(x, y); }
inline float16x8_t mul(float16x8_t x, float16x8_t y) { return vmulq_f16(x, y); }
#endif // __ARM_FEATURE_FP16_VECTOR_ARITHMETIC
#if defined(__VXE__) || defined(__VXE2__)
inline float32x4_t add(float32x4_t x, float32x4_t y) { return vec_add(x, y); }
inline float32x4_t sub(float32x4_t x, float32x4_t y) { return vec_sub(x, y); }
inline float32x4_t mul(float32x4_t x, float32x4_t y) { return vec_mul(x, y); }
#endif
#if defined(__MMA__)
typedef vector unsigned char vec_t;
typedef __vector_quad acc_t;
@@ -162,6 +168,13 @@ inline float16x8_t madd(float16x8_t a, float16x8_t b, float16x8_t c) {
#endif
#endif
#if defined(__VXE__) || defined(__VXE2__)
template <>
inline float32x4_t madd(float32x4_t a, float32x4_t b, float32x4_t c) {
return vec_madd(a, b, c);
}
#endif
////////////////////////////////////////////////////////////////////////////////////////////////////
// VECTORIZED HORIZONTAL SUM
@@ -178,6 +191,13 @@ inline float hsum(float16x8_t x) {
}
#endif // __ARM_FEATURE_FP16_VECTOR_ARITHMETIC
#if defined(__VXE__) || defined(__VXE2__)
inline float hsum(float32x4_t x) {
float32x4_t tmp = x + vec_reve(x);
return tmp[0] + tmp[1];
}
#endif
#if defined(__SSE__) || defined(__AVX__) || defined(__AVX2__) || defined(__AVX512F__)
inline float hsum(__m128 x) {
#if defined(__AVX__) || defined(__AVX2__) || defined(__AVX512F__)
@@ -227,6 +247,21 @@ template <> inline float32x4_t load(const ggml_fp16_t *p) {
#endif // _MSC_VER
#endif // __ARM_NEON
#if defined(__VXE__) || defined(__VXE2__)
template <> inline float32x4_t load(const ggml_fp16_t * p) {
float tmp[4];
for (int i = 0; i < 4; i++) {
tmp[i] = GGML_FP16_TO_FP32(p[i]);
}
return vec_xl(0, (const float *)(tmp));
}
template <> inline float32x4_t load(const float * p) {
return vec_xl(0, p);
}
#endif
#if defined(__SSE__) || defined(__AVX__) || defined(__AVX2__) || defined(__AVX512F__)
template <> inline __m128 load(const float *p) {
return _mm_loadu_ps(p);
@@ -3319,6 +3354,14 @@ bool llamafile_sgemm(const struct ggml_compute_params * params, int64_t m, int64
(const float *)B, ldb,
(float *)C, ldc};
return tb.matmul(m, n);
#elif defined(__VXE__) || defined(__VXE2__)
if (n < 4)
return false;
tinyBLAS<4, float32x4_t, float32x4_t, float, float, float> tb{ params,
k, (const float *)A, lda,
(const float *)B, ldb,
(float *)C, ldc};
return tb.matmul(m, n);
#elif defined(__MMA__)
if (k % 8)
return false;
@@ -3410,6 +3453,16 @@ bool llamafile_sgemm(const struct ggml_compute_params * params, int64_t m, int64
(float *)C, ldc};
return tb.matmul(m, n);
}
#elif defined(__VXE__) || defined(__VXE2__)
if (n < 4)
return false;
if (Btype == GGML_TYPE_F16) {
tinyBLAS<4, float32x4_t, float32x4_t, ggml_fp16_t, ggml_fp16_t, float> tb{ params,
k, (const ggml_fp16_t *)A, lda,
(const ggml_fp16_t *)B, ldb,
(float *)C, ldc};
return tb.matmul(m, n);
}
#endif
return false;
}

5
ggml/src/ggml-cpu/llamafile/sgemm.h
View File

@@ -1,6 +1,11 @@
#pragma once
#include <stdint.h>
#include <stdbool.h>
#if defined(__VXE__) || defined(__VXE2__)
#include <vecintrin.h>
#endif
#ifdef __cplusplus
extern "C" {
#endif