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

Added sve for dequantized_q8_0

This commit is contained in:
vithulep 2025-07-31 09:44:18 +05:30
parent a6b0726eec
commit 89bbd50599
1 changed files with 32 additions and 62 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

94
ggml/src/ggml-cpu/arch/arm/quants.c
View File

@ -47,73 +47,43 @@ void quantize_row_q8_0(const float * GGML_RESTRICT x, void * GGML_RESTRICT vy, i
block_q8_0 * GGML_RESTRICT y = vy;
// #if defined(__ARM_FEATURE_SVE)
// // const int sve_register_length = svcntb() * 8; //get the vector length
// // const int ggml_f32_epr = sve_register_length / 32;
// const svfloat32_t inactive1 = svdup_n_f32(0.0f);
// const svbool_t pg = svptrue_b32();
// svfloat32_t zero = svdup_f32(0.0f);
// svfloat32_t half = svdup_f32(0.5f);
// const svint32_t inactive2 = svdup_n_s32(0);
#if defined(__ARM_FEATURE_SVE)
const int sve_register_length = svcntb() * 8;
const int ggml_f32_epr = sve_register_length / 32;
const svfloat32_t inactive1 = svdup_n_f32(0.0f);
const svbool_t pg = svptrue_b32();
svfloat32_t zero = svdup_f32(0.0f);
svfloat32_t half = svdup_f32(0.5f);
// for (int i = 0; i < nb; i+=1) {
// svfloat32_t srcv1, srcv2, srcv3, srcv4;
// svfloat32_t asrcv1, asrcv2, asrcv3, asrcv4;
// float32_t amax1 = 0.0;
for (int i = 0; i < nb; i+=1) {
svfloat32_t srcv1, asrcv1;
svfloat32_t sv_max = svdup_n_f32(0.0f);
float32_t amax = 0.0;
// srcv1 = svld1_f32(pg, x + i*32);
// asrcv1 = svabs_f32_m(inactive1, pg, srcv1);
for (int j = 0; j < QK8_0; j+=ggml_f32_epr) {
srcv1 = svld1_f32(pg, x + i*32 + j);
asrcv1 = svabs_f32_m(inactive1, pg, srcv1);
sv_max = svmax_f32_m(pg, sv_max, asrcv1);
}
amax = svmaxv_f32(pg, sv_max);
float32_t d = amax / ((1 << 7) - 1);
float32_t id = d ? 1.0f/d : 0.0f;
y[i].d = GGML_FP32_TO_FP16(d);
for (int j = 0; j < QK8_0; j+=ggml_f32_epr) {
srcv1 = svld1_f32(pg, x + i*32 + j);
const svfloat32_t v1 = svmul_n_f32_m(pg, srcv1, id);
// srcv2 = svld1_f32(pg, x + i*32 + 8);
// asrcv2 = svabs_f32_m(inactive1, pg, srcv2);
svbool_t ge_zero = svcmpge_f32(pg, v1, zero);
svfloat32_t v_pos = svadd_f32_m(pg, v1, half);
svfloat32_t v_neg = svsub_f32_m(pg, v1, half);
// srcv3 = svld1_f32(pg, x + i*32 + 16);
// asrcv3 = svabs_f32_m(inactive1, pg, srcv3);
svfloat32_t v_rounded = svsel_f32(ge_zero, v_pos, v_neg);
svint32_t result = svcvt_s32_f32_x(pg, v_rounded);
svst1b_s32(pg, &y[i].qs[j], result);
}
}
// srcv4 = svld1_f32(pg, x + i*32 + 24);
// asrcv4 = svabs_f32_m(inactive1, pg, srcv4);
// svfloat32_t max1 = svmax_f32_m(pg, asrcv2, asrcv1);
// svfloat32_t max2 = svmax_f32_m(pg, asrcv4, asrcv3);
// svfloat32_t max3 = svmax_f32_m(pg, max2, max1);
// amax1 = svmaxv_f32(pg, max3);
// float32_t d1 = amax1 / ((1 << 7) - 1);
// float32_t id1 = d1 ? 1.0f/d1 : 0.0f;
// y[i].d = GGML_FP32_TO_FP16(d1);
// const svfloat32_t v1 = svmul_n_f32_m(pg, srcv1, id1);
// const svfloat32_t v2 = svmul_n_f32_m(pg, srcv2, id1);
// const svfloat32_t v3 = svmul_n_f32_m(pg, srcv3, id1);
// const svfloat32_t v4 = svmul_n_f32_m(pg, srcv4, id1);
// svbool_t ge_zero = svcmpge_f32(pg, v1, zero);
// svfloat32_t v_rounded = svsel_f32(ge_zero, svadd_f32_m(pg, v1, half), svsub_f32_m(pg, v1, half));
// // svint32_t v_rounded = svcvt_s32_f32_m(inactive2, pg, v1);
// svbool_t ge_zero_2 = svcmpge_f32(pg, v2, zero);
// svfloat32_t v_rounded_2 = svsel_f32(ge_zero_2, svadd_f32_m(pg, v2, half), svsub_f32_m(pg, v2, half));
// svbool_t ge_zero_3 = svcmpge_f32(pg, v3, zero);
// svfloat32_t v_rounded_3 = svsel_f32(ge_zero_3, svadd_f32_m(pg, v3, half), svsub_f32_m(pg, v3, half));
// svbool_t ge_zero_4 = svcmpge_f32(pg, v4, zero);
// svfloat32_t v_rounded_4 = svsel_f32(ge_zero_4, svadd_f32_m(pg, v4, half), svsub_f32_m(pg, v4, half));
// svint32_t result = svcvt_s32_f32_x(pg, v_rounded);
// svst1b_s32(pg, &y[i].qs[0], result);
// svint32_t result_2 = svcvt_s32_f32_x(pg, v_rounded_2);
// svst1b_s32(pg, &y[i].qs[8], result_2);
// svint32_t result_3 = svcvt_s32_f32_x(pg, v_rounded_3);
// svst1b_s32(pg, &y[i].qs[16], result_3);
// svint32_t result_4 = svcvt_s32_f32_x(pg, v_rounded_4);
// svst1b_s32(pg, &y[i].qs[24], result_4);
// }
#if defined(__ARM_NEON)
#elif defined(__ARM_NEON)
for (int i = 0; i < nb; i++) {
float32x4_t srcv [8];
float32x4_t asrcv[8];