vbatts/llama.cpp
1
0
Fork 0
Code Issues Pull requests Projects Releases Packages Wiki Activity Actions

k-quants : fix build on armv7 (android only) (#2920)

Browse source 
* k-quants : fix build on armv7

* ggml : cleanup unused arm32 specific impl

* k-quants : avoid some unused vzero / mzero define

* ggml-alloc : use 4g for MEASURE_MAX_SIZE in 32-bit arm
This commit is contained in:
Jhen-Jie Hong 2023-09-02 20:23:45 +08:00 • committed by GitHub
parent 571083f508
commit 21f3d1be86
No known key found for this signature in database
GPG key ID: 4AEE18F83AFDEB23
3 changed files with 42 additions and 51 deletions
Download patch file Download diff file Expand all files Collapse all files

40
k_quants.c
View file

@ -13,6 +13,26 @@
//
#include <arm_neon.h>
#if !defined(__aarch64__)
inline static int32_t vaddvq_s16(int16x8_t v) {
return
(int32_t)vgetq_lane_s16(v, 0) + (int32_t)vgetq_lane_s16(v, 1) +
(int32_t)vgetq_lane_s16(v, 2) + (int32_t)vgetq_lane_s16(v, 3) +
(int32_t)vgetq_lane_s16(v, 4) + (int32_t)vgetq_lane_s16(v, 5) +
(int32_t)vgetq_lane_s16(v, 6) + (int32_t)vgetq_lane_s16(v, 7);
}
inline static int16x8_t vpaddq_s16(int16x8_t a, int16x8_t b) {
int16x4_t a0 = vpadd_s16(vget_low_s16(a), vget_high_s16(a));
int16x4_t b0 = vpadd_s16(vget_low_s16(b), vget_high_s16(b));
return vcombine_s16(a0, b0);
}
inline static int32_t vaddvq_s32(int32x4_t v) {
return vgetq_lane_s32(v, 0) + vgetq_lane_s32(v, 1) + vgetq_lane_s32(v, 2) + vgetq_lane_s32(v, 3);
}
#endif
#else
#ifdef __wasm_simd128__
@ -1302,7 +1322,9 @@ void ggml_vec_dot_q2_K_q8_K(const int n, float * restrict s, const void * restri
const uint8x16_t m3 = vdupq_n_u8(0x3);
const uint8x16_t m4 = vdupq_n_u8(0xF);
#if defined(__ARM_FEATURE_DOTPROD)
const int32x4_t vzero = vdupq_n_s32(0);
#endif
int8x16x2_t q2bytes;
uint8_t aux[16];
@ -1608,7 +1630,9 @@ void ggml_vec_dot_q2_K_q8_K(const int n, float * restrict s, const void * restri
#ifdef __ARM_NEON
const uint8x16_t m3 = vdupq_n_u8(0x3);
#if defined(__ARM_FEATURE_DOTPROD)
const int32x4_t vzero = vdupq_n_s32(0);
#endif
int8x16x4_t q2bytes;
@ -2592,8 +2616,6 @@ void ggml_vec_dot_q4_K_q8_K(const int n, float * restrict s, const void * restri
const uint8_t * restrict q4 = x[i].qs;
const int8_t * restrict q8 = y[i].qs;
//int32x4_t isum = mzero;
int32_t sumi1 = 0;
int32_t sumi2 = 0;
@ -3092,9 +3114,11 @@ void ggml_vec_dot_q5_K_q8_K(const int n, float * restrict s, const void * restri
#ifdef __ARM_NEON
const uint8x16_t m4b = vdupq_n_u8(0xf);
const int32x4_t mzero = vdupq_n_s32(0);
const uint8x16_t mone = vdupq_n_u8(1);
const uint8x16_t mtwo = vdupq_n_u8(2);
#if defined(__ARM_FEATURE_DOTPROD)
const int32x4_t mzero = vdupq_n_s32(0);
#endif
int8x16x4_t q5bytes;
@ -3437,8 +3461,10 @@ void ggml_vec_dot_q5_K_q8_K(const int n, float * restrict s, const void * restri
#ifdef __ARM_NEON
const uint8x16_t m4b = vdupq_n_u8(0xf);
const int32x4_t mzero = vdupq_n_s32(0);
const uint8x16_t mh = vdupq_n_u8(16);
#if defined(__ARM_FEATURE_DOTPROD)
const int32x4_t mzero = vdupq_n_s32(0);
#endif
int8x16x4_t q5bytes;
uint8x16x4_t q5h;
@ -3656,7 +3682,9 @@ void ggml_vec_dot_q6_K_q8_K(const int n, float * restrict s, const void * restri
float sum = 0;
const uint8x16_t m4b = vdupq_n_u8(0xF);
#if defined(__ARM_FEATURE_DOTPROD)
const int32x4_t vzero = vdupq_n_s32(0);
#endif
//const int8x16_t m32s = vdupq_n_s8(32);
const uint8x16_t mone = vdupq_n_u8(3);
@ -4045,8 +4073,10 @@ void ggml_vec_dot_q6_K_q8_K(const int n, float * restrict s, const void * restri
float sum = 0;
const uint8x16_t m4b = vdupq_n_u8(0xF);
const int32x4_t vzero = vdupq_n_s32(0);
const int8x16_t m32s = vdupq_n_s8(32);
#if defined(__ARM_FEATURE_DOTPROD)
const int32x4_t vzero = vdupq_n_s32(0);
#endif
const uint8x16_t mone = vdupq_n_u8(3);