ExternalVendorCode/whisper.cpp
1
0
Fork 0
mirror of https://github.com/ggerganov/whisper.cpp.git synced 2025-02-20 09:06:18 +00:00
Code Issues Actions Packages Projects Releases Wiki Activity

ggml : fix 32-bit ARM build + quantization

Browse Source
This commit is contained in:
Georgi Gerganov 2023-05-02 21:47:12 +03:00
parent 0bcb64b184
commit 5974c8facd
2 changed files with 72 additions and 13 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

13
Makefile
View File

@ -177,7 +177,7 @@ ifdef WHISPER_GPROF
endif
ifneq ($(filter aarch64%,$(UNAME_M)),)
CFLAGS += -mcpu=native
CFLAGS += -mcpu=native
CXXFLAGS += -mcpu=native
endif
@ -188,15 +188,18 @@ endif
ifneq ($(filter armv7%,$(UNAME_M)),)
# 32-bit ARM, for example on Armbian or possibly raspbian
CFLAGS += -mfpu=neon -mfp16-format=ieee -mno-unaligned-access -funsafe-math-optimizations
#CFLAGS += -mfpu=neon -mfp16-format=ieee -funsafe-math-optimizations -mno-unaligned-access
#CXXFLAGS += -mfpu=neon -mfp16-format=ieee -funsafe-math-optimizations -mno-unaligned-access
# 64-bit ARM, use these (TODO: auto-detect 64-bit)
# CFLAGS += -mfpu=neon-fp-armv8 -mfp16-format=ieee -mno-unaligned-access -funsafe-math-optimizations
# 64-bit ARM on 32-bit OS, use these (TODO: auto-detect 64-bit)
CFLAGS += -mfpu=neon-fp-armv8 -mfp16-format=ieee -funsafe-math-optimizations -mno-unaligned-access
CXXFLAGS += -mfpu=neon-fp-armv8 -mfp16-format=ieee -funsafe-math-optimizations -mno-unaligned-access
endif
ifneq ($(filter armv8%,$(UNAME_M)),)
# Raspberry Pi 4
CFLAGS += -mfp16-format=ieee -mno-unaligned-access
CFLAGS += -mfpu=neon-fp-armv8 -mfp16-format=ieee -funsafe-math-optimizations -mno-unaligned-access
CXXFLAGS += -mfpu=neon-fp-armv8 -mfp16-format=ieee -funsafe-math-optimizations -mno-unaligned-access
endif
#

72
ggml.c
View File

@ -671,35 +671,91 @@ float vmaxvq_f32(float32x4_t v) {
}
int8x8_t vzip1_s8(int8x8_t a, int8x8_t b) {
return vget_low_s8(vcombine_s8(a, b));
int8x8_t res;
res[0] = a[0]; res[1] = b[0];
res[2] = a[1]; res[3] = b[1];
res[4] = a[2]; res[5] = b[2];
res[6] = a[3]; res[7] = b[3];
return res;
}
int8x8_t vzip2_s8(int8x8_t a, int8x8_t b) {
return vget_high_s8(vcombine_s8(a, b));
int8x8_t res;
res[0] = a[4]; res[1] = b[4];
res[2] = a[5]; res[3] = b[5];
res[4] = a[6]; res[5] = b[6];
res[6] = a[7]; res[7] = b[7];
return res;
}
uint8x8_t vzip1_u8(uint8x8_t a, uint8x8_t b) {
return vget_low_u8(vcombine_u8(a, b));
uint8x8_t res;
res[0] = a[0]; res[1] = b[0];
res[2] = a[1]; res[3] = b[1];
res[4] = a[2]; res[5] = b[2];
res[6] = a[3]; res[7] = b[3];
return res;
}
uint8x8_t vzip2_u8(uint8x8_t a, uint8x8_t b) {
return vget_high_u8(vcombine_u8(a, b));
uint8x8_t res;
res[0] = a[4]; res[1] = b[4];
res[2] = a[5]; res[3] = b[5];
res[4] = a[6]; res[5] = b[6];
res[6] = a[7]; res[7] = b[7];
return res;
}
int8x16_t vzip1q_s8(int8x16_t a, int8x16_t b) {
return vcombine_s8(vget_low_s8(a), vget_low_s8(b));
int8x16_t res;
res[0] = a[0]; res[1] = b[0]; res[2] = a[1]; res[3] = b[1];
res[4] = a[2]; res[5] = b[2]; res[6] = a[3]; res[7] = b[3];
res[8] = a[4]; res[9] = b[4]; res[10] = a[5]; res[11] = b[5];
res[12] = a[6]; res[13] = b[6]; res[14] = a[7]; res[15] = b[7];
return res;
}
int8x16_t vzip2q_s8(int8x16_t a, int8x16_t b) {
return vcombine_s8(vget_high_s8(a), vget_high_s8(b));
int8x16_t res;
res[0] = a[8]; res[1] = b[8]; res[2] = a[9]; res[3] = b[9];
res[4] = a[10]; res[5] = b[10]; res[6] = a[11]; res[7] = b[11];
res[8] = a[12]; res[9] = b[12]; res[10] = a[13]; res[11] = b[13];
res[12] = a[14]; res[13] = b[14]; res[14] = a[15]; res[15] = b[15];
return res;
}
uint8x16_t vzip1q_u8(uint8x16_t a, uint8x16_t b) {
return vcombine_u8(vget_low_u8(a), vget_low_u8(b));
uint8x16_t res;
res[0] = a[0]; res[1] = b[0]; res[2] = a[1]; res[3] = b[1];
res[4] = a[2]; res[5] = b[2]; res[6] = a[3]; res[7] = b[3];
res[8] = a[4]; res[9] = b[4]; res[10] = a[5]; res[11] = b[5];
res[12] = a[6]; res[13] = b[6]; res[14] = a[7]; res[15] = b[7];
return res;
}
uint8x16_t vzip2q_u8(uint8x16_t a, uint8x16_t b) {
return vcombine_u8(vget_high_u8(a), vget_high_u8(b));
uint8x16_t res;
res[0] = a[8]; res[1] = b[8]; res[2] = a[9]; res[3] = b[9];
res[4] = a[10]; res[5] = b[10]; res[6] = a[11]; res[7] = b[11];
res[8] = a[12]; res[9] = b[12]; res[10] = a[13]; res[11] = b[13];
res[12] = a[14]; res[13] = b[14]; res[14] = a[15]; res[15] = b[15];
return res;
}
int32x4_t vcvtnq_s32_f32(float32x4_t v) {