root/llama.cpp
1
0
Fork 0
mirror of https://github.com/ggerganov/llama.cpp.git synced 2024-12-25 02:44:36 +00:00
Code Issues Actions 9 Packages Projects Releases Wiki Activity

ggml-quants : provide ggml_vqtbl1q_u8 for 64bit compatibility (#5711)

Browse Source 
* [ggml-quants] Provide ggml_vqtbl1q_u8 for 64bit compatibility

vqtbl1q_u8 is not part of arm v7 neon library

* [android-example] Remove abi filter after arm v7a fix

* [github-workflows] Do not skip Android armeabi-v7a build
This commit is contained in:
Radosław Gryta 2024-02-25 19:43:00 +01:00 committed by GitHub
parent f1a98c5254
commit abbabc5e51
No known key found for this signature in database
GPG Key ID: B5690EEEBB952194
3 changed files with 32 additions and 12 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

3
.github/workflows/build.yml vendored
View File

@ -669,8 +669,7 @@ jobs:
run: | run: |
cd examples/llama.android cd examples/llama.android
# Skip armeabi-v7a for now (https://github.com/llvm/llvm-project/issues/65820). ./gradlew build --no-daemon
./gradlew build --no-daemon -Pskip-armeabi-v7a
# freeBSD-latest: # freeBSD-latest:
# runs-on: macos-12 # runs-on: macos-12

8
examples/llama.android/app/build.gradle.kts
View File

@ -21,12 +21,8 @@ android {
useSupportLibrary = true useSupportLibrary = true
} }
ndk { ndk {
// Workaround for https://github.com/llvm/llvm-project/issues/65820 // Add NDK properties if wanted, e.g.
// affecting armeabi-v7a. Skip armeabi-v7a when invoked with // abiFilters += listOf("arm64-v8a")
// -Pskip-armeabi-v7a (e.g., ./gradlew build -Pskip-armeabi-v7a).
if (project.hasProperty("skip-armeabi-v7a")) {
abiFilters += listOf("arm64-v8a", "x86_64", "x86")
}
} }
externalNativeBuild { externalNativeBuild {
cmake { cmake {

33
ggml-quants.c
View File

@ -462,6 +462,30 @@ inline static int8x16_t ggml_vqtbl1q_s8(int8x16_t a, uint8x16_t b) {
return res; return res;
} }
// NOTE: not tested
inline static int8x16_t ggml_vqtbl1q_u8(uint8x16_t a, uint8x16_t b) {
int8x16_t res;
res[ 0] = a[b[ 0]];
res[ 1] = a[b[ 1]];
res[ 2] = a[b[ 2]];
res[ 3] = a[b[ 3]];
res[ 4] = a[b[ 4]];
res[ 5] = a[b[ 5]];
res[ 6] = a[b[ 6]];
res[ 7] = a[b[ 7]];
res[ 8] = a[b[ 8]];
res[ 9] = a[b[ 9]];
res[10] = a[b[10]];
res[11] = a[b[11]];
res[12] = a[b[12]];
res[13] = a[b[13]];
res[14] = a[b[14]];
res[15] = a[b[15]];
return res;
}
#else #else
#define ggml_int16x8x2_t int16x8x2_t #define ggml_int16x8x2_t int16x8x2_t
@ -476,6 +500,7 @@ inline static int8x16_t ggml_vqtbl1q_s8(int8x16_t a, uint8x16_t b) {
#define ggml_vld1q_s8_x2 vld1q_s8_x2 #define ggml_vld1q_s8_x2 vld1q_s8_x2
#define ggml_vld1q_s8_x4 vld1q_s8_x4 #define ggml_vld1q_s8_x4 vld1q_s8_x4
#define ggml_vqtbl1q_s8 vqtbl1q_s8 #define ggml_vqtbl1q_s8 vqtbl1q_s8
#define ggml_vqtbl1q_u8 vqtbl1q_u8
#endif #endif
@ -9488,8 +9513,8 @@ void ggml_vec_dot_iq3_s_q8_K (int n, float * GGML_RESTRICT s, size_t bs, const v
qs += 16; qs += 16;
vs.val[0] = vreinterpretq_u8_u32(vdupq_n_u32(signs[0] | (signs[1] << 16))); vs.val[0] = vreinterpretq_u8_u32(vdupq_n_u32(signs[0] | (signs[1] << 16)));
vs.val[1] = vandq_u8(vqtbl1q_u8(vs.val[0], mask1.val[1]), mask2); vs.val[1] = vandq_u8(ggml_vqtbl1q_u8(vs.val[0], mask1.val[1]), mask2);
vs.val[0] = vandq_u8(vqtbl1q_u8(vs.val[0], mask1.val[0]), mask2); vs.val[0] = vandq_u8(ggml_vqtbl1q_u8(vs.val[0], mask1.val[0]), mask2);
vs.val[0] = vceqq_u8(vs.val[0], mask2); vs.val[0] = vceqq_u8(vs.val[0], mask2);
vs.val[1] = vceqq_u8(vs.val[1], mask2); vs.val[1] = vceqq_u8(vs.val[1], mask2);
@ -9497,8 +9522,8 @@ void ggml_vec_dot_iq3_s_q8_K (int n, float * GGML_RESTRICT s, size_t bs, const v
q3s.val[1] = vsubq_s8(vreinterpretq_s8_u8(veorq_u8(vs.val[1], vreinterpretq_u8_u32(aux32x4_1))), vreinterpretq_s8_u8(vs.val[1])); q3s.val[1] = vsubq_s8(vreinterpretq_s8_u8(veorq_u8(vs.val[1], vreinterpretq_u8_u32(aux32x4_1))), vreinterpretq_s8_u8(vs.val[1]));
vs.val[0] = vreinterpretq_u8_u32(vdupq_n_u32(signs[2] | (signs[3] << 16))); vs.val[0] = vreinterpretq_u8_u32(vdupq_n_u32(signs[2] | (signs[3] << 16)));
vs.val[1] = vandq_u8(vqtbl1q_u8(vs.val[0], mask1.val[1]), mask2); vs.val[1] = vandq_u8(ggml_vqtbl1q_u8(vs.val[0], mask1.val[1]), mask2);
vs.val[0] = vandq_u8(vqtbl1q_u8(vs.val[0], mask1.val[0]), mask2); vs.val[0] = vandq_u8(ggml_vqtbl1q_u8(vs.val[0], mask1.val[0]), mask2);
vs.val[0] = vceqq_u8(vs.val[0], mask2); vs.val[0] = vceqq_u8(vs.val[0], mask2);
vs.val[1] = vceqq_u8(vs.val[1], mask2); vs.val[1] = vceqq_u8(vs.val[1], mask2);