Skip to content

Commit 4806498

Browse files
Vithuleppvnameggerganov
authored
ggml: aarch64: implement SVE kernels for q3_K_q8_K vector dot (#11917)
* Added SVE Implementation for Q3_K Kernel in ggml-cpu-quants.c file * Improved Formating of code in ggml-cpu-quants.c file * style : minor fixes * style : less whitespaces * style : ptr spaceing --------- Co-authored-by: vithulep <[email protected]> Co-authored-by: Georgi Gerganov <[email protected]>
1 parent 0d55958 commit 4806498

File tree

1 file changed

+176
-1
lines changed

1 file changed

+176
-1
lines changed

‎ggml/src/ggml-cpu/ggml-cpu-quants.c

Lines changed: 176 additions & 1 deletion
Original file line numberDiff line numberDiff line change
@@ -5112,7 +5112,182 @@ void ggml_vec_dot_q3_K_q8_K(int n, float * restrict s, size_t bs, const void * r
51125112

51135113
const int nb = n / QK_K;
51145114

5115-
#ifdef __ARM_NEON
5115+
#if defined(__ARM_FEATURE_SVE)
5116+
5117+
uint32_t utmp[4];
5118+
5119+
const int8_t m32 = 32;
5120+
const int vector_length = svcntb()*8;
5121+
const svuint8_t m3b_sv = svdup_n_u8(0x3);
5122+
const svint32_t vzero_sv = svdup_n_s32(0);
5123+
5124+
const svuint8_t m0_sv = svdup_n_u8(1);
5125+
const svuint8_t m1_sv = svlsl_n_u8_x(svptrue_b8(), m0_sv, 1);
5126+
const svuint8_t m2_sv = svlsl_n_u8_x(svptrue_b8(), m0_sv, 2);
5127+
const svuint8_t m3_sv = svlsl_n_u8_x(svptrue_b8(), m0_sv, 3);
5128+
svbool_t pred_s32 = svnot_b_z (svptrue_b32(), svptrue_pat_b32(SV_VL4));
5129+
5130+
float sum = 0;
5131+
5132+
for (int i = 0; i < nb; ++i) {
5133+
5134+
const float d = y[i].d * GGML_FP16_TO_FP32(x[i].d);
5135+
5136+
const uint8_t * restrict q3_sv = x[i].qs;
5137+
const uint8_t * restrict qh_sv = x[i].hmask;
5138+
const int8_t * restrict q8_sv = y[i].qs;
5139+
5140+
// Set up scales
5141+
uint32_t * aux = &x[i].scales;
5142+
utmp[3] = ((aux[1] >> 4) & kmask2) | (((aux[2] >> 6) & kmask1) << 4);
5143+
utmp[2] = ((aux[0] >> 4) & kmask2) | (((aux[2] >> 4) & kmask1) << 4);
5144+
utmp[1] = (aux[1] & kmask2) | (((aux[2] >> 2) & kmask1) << 4);
5145+
utmp[0] = (aux[0] & kmask2) | (((aux[2] >> 0) & kmask1) << 4);
5146+
5147+
int8_t * scale = (int8_t *)utmp;
5148+
5149+
for (int j = 0; j < 16; ++j) scale[j] -= m32;
5150+
5151+
switch (vector_length) {
5152+
case 128:
5153+
{
5154+
svuint8_t qhbits_sv_1 = svld1_u8(svptrue_b8(), qh_sv);
5155+
svuint8_t qhbits_sv_2 = svld1_u8(svptrue_b8(), qh_sv+16);
5156+
svuint8_t q3h_sv;
5157+
5158+
svint32_t sumi1_1 = svdup_n_s32(0);
5159+
svint8_t q3bytes_sv;
5160+
5161+
for (int j = 0; j < QK_K/128; ++j) {
5162+
5163+
const svuint8_t q3bits_sv = svld1_u8(svptrue_b8(), q3_sv); q3_sv += 16;
5164+
const svuint8_t q3bits_sv_1 = svld1_u8(svptrue_b8(), q3_sv); q3_sv += 16;
5165+
svint8_t q8bytes_1_sv_1 = svld1_s8(svptrue_b8(), q8_sv); q8_sv += 16;
5166+
svint8_t q8bytes_1_sv_2 = svld1_s8(svptrue_b8(), q8_sv); q8_sv += 16;
5167+
5168+
q3h_sv = svlsl_n_u8_x(svptrue_b8(), svbic_u8_x(svptrue_b8(), m0_sv, qhbits_sv_1), 2);
5169+
q3bytes_sv = svsub_s8_x(svptrue_b8(), svreinterpret_s8_u8(svand_u8_m(svptrue_b8(), q3bits_sv, m3b_sv)), svreinterpret_s8_u8(q3h_sv));
5170+
5171+
sumi1_1 = svmla_s32_m(svptrue_b32(), sumi1_1, svdot_s32(vzero_sv, q3bytes_sv, q8bytes_1_sv_1), svdup_n_s32((int32_t)scale[0]));
5172+
5173+
q3h_sv = svlsl_n_u8_x(svptrue_b8(), svbic_u8_x(svptrue_b8(), m0_sv, qhbits_sv_2), 2);
5174+
q3bytes_sv = svsub_s8_x(svptrue_b8(), svreinterpret_s8_u8(svand_u8_m(svptrue_b8(), q3bits_sv_1, m3b_sv)), svreinterpret_s8_u8(q3h_sv));
5175+
5176+
sumi1_1 = svmla_s32_m(svptrue_b32(), sumi1_1, svdot_s32(vzero_sv, q3bytes_sv, q8bytes_1_sv_2), svdup_n_s32((int32_t)scale[1]));
5177+
5178+
q8bytes_1_sv_1 = svld1_s8(svptrue_b8(), q8_sv); q8_sv += 16;
5179+
q8bytes_1_sv_2 = svld1_s8(svptrue_b8(), q8_sv); q8_sv += 16;
5180+
5181+
q3h_sv = svlsl_n_u8_x(svptrue_b8(), svbic_u8_x(svptrue_b8(), m1_sv, qhbits_sv_1), 1);
5182+
q3bytes_sv = svsub_s8_x(svptrue_b8(), svreinterpret_s8_u8(svand_u8_m(svptrue_b8(), svlsr_n_u8_x(svptrue_b8(), q3bits_sv, 2), m3b_sv)), svreinterpret_s8_u8(q3h_sv));
5183+
5184+
sumi1_1 = svmla_s32_m(svptrue_b32(), sumi1_1, svdot_s32(vzero_sv, q3bytes_sv, q8bytes_1_sv_1), svdup_n_s32((int32_t)scale[2]));
5185+
5186+
q3h_sv = svlsl_n_u8_x(svptrue_b8(), svbic_u8_x(svptrue_b8(), m1_sv, qhbits_sv_2), 1);
5187+
q3bytes_sv = svsub_s8_x(svptrue_b8(), svreinterpret_s8_u8(svand_u8_m(svptrue_b8(), svlsr_n_u8_x(svptrue_b8(), q3bits_sv_1, 2), m3b_sv)), svreinterpret_s8_u8(q3h_sv));
5188+
5189+
sumi1_1 = svmla_s32_m(svptrue_b32(), sumi1_1, svdot_s32(vzero_sv, q3bytes_sv, q8bytes_1_sv_2), svdup_n_s32((int32_t)scale[3]));
5190+
5191+
5192+
scale += 4;
5193+
q8bytes_1_sv_1 = svld1_s8(svptrue_b8(), q8_sv); q8_sv += 16;
5194+
q8bytes_1_sv_2 = svld1_s8(svptrue_b8(), q8_sv); q8_sv += 16;
5195+
5196+
q3h_sv = svbic_u8_x(svptrue_b8(), m2_sv, qhbits_sv_1);
5197+
q3bytes_sv = svsub_s8_x(svptrue_b8(), svreinterpret_s8_u8(svand_u8_m(svptrue_b8(), svlsr_n_u8_x(svptrue_b8(), q3bits_sv, 4), m3b_sv)), svreinterpret_s8_u8(q3h_sv));
5198+
5199+
sumi1_1 = svmla_s32_m(svptrue_b32(), sumi1_1, svdot_s32(vzero_sv, q3bytes_sv, q8bytes_1_sv_1), svdup_n_s32((int32_t)scale[0]));
5200+
5201+
q3h_sv = svbic_u8_x(svptrue_b8(), m2_sv, qhbits_sv_2);
5202+
q3bytes_sv = svsub_s8_x(svptrue_b8(), svreinterpret_s8_u8(svand_u8_m(svptrue_b8(), svlsr_n_u8_x(svptrue_b8(), q3bits_sv_1, 4), m3b_sv)), svreinterpret_s8_u8(q3h_sv));
5203+
5204+
sumi1_1 = svmla_s32_m(svptrue_b32(), sumi1_1, svdot_s32(vzero_sv, q3bytes_sv, q8bytes_1_sv_2), svdup_n_s32((int32_t)scale[1]));
5205+
5206+
5207+
q8bytes_1_sv_1 = svld1_s8(svptrue_b8(), q8_sv); q8_sv += 16;
5208+
q8bytes_1_sv_2 = svld1_s8(svptrue_b8(), q8_sv); q8_sv += 16;
5209+
5210+
q3h_sv = svlsr_n_u8_x(svptrue_b8(), svbic_u8_x(svptrue_b8(), m3_sv, qhbits_sv_1), 1);
5211+
q3bytes_sv = svsub_s8_x(svptrue_b8(), svreinterpret_s8_u8(svand_u8_m(svptrue_b8(), svlsr_n_u8_x(svptrue_b8(), q3bits_sv, 6), m3b_sv)), svreinterpret_s8_u8(q3h_sv));
5212+
5213+
sumi1_1 = svmla_s32_m(svptrue_b32(), sumi1_1, svdot_s32(vzero_sv, q3bytes_sv, q8bytes_1_sv_1), svdup_n_s32((int32_t)scale[2]));
5214+
5215+
q3h_sv = svlsr_n_u8_x(svptrue_b8(), svbic_u8_x(svptrue_b8(), m3_sv, qhbits_sv_2), 1);
5216+
q3bytes_sv = svsub_s8_x(svptrue_b8(), svreinterpret_s8_u8(svand_u8_m(svptrue_b8(), svlsr_n_u8_x(svptrue_b8(), q3bits_sv_1, 6), m3b_sv)), svreinterpret_s8_u8(q3h_sv));
5217+
5218+
sumi1_1 = svmla_s32_m(svptrue_b32(), sumi1_1, svdot_s32(vzero_sv, q3bytes_sv, q8bytes_1_sv_2), svdup_n_s32((int32_t)scale[3]));
5219+
5220+
if (j == 0) {
5221+
qhbits_sv_1 = svlsr_n_u8_x(svptrue_b8(), qhbits_sv_1, 4);
5222+
qhbits_sv_2 = svlsr_n_u8_x(svptrue_b8(), qhbits_sv_2, 4);
5223+
}
5224+
5225+
scale += 4;
5226+
}
5227+
5228+
sum += d * (svaddv_s32(svptrue_b32(), sumi1_1));
5229+
} break;
5230+
case 256:
5231+
case 512:
5232+
{
5233+
svuint8_t qhbits_sv = svld1_u8(svptrue_pat_b8(SV_VL32), qh_sv);
5234+
svuint8_t q3h_sv;
5235+
5236+
svint32_t sumi1_1 = svdup_n_s32(0);
5237+
svint8_t q3bytes_sv;
5238+
5239+
for (int j = 0; j < QK_K/128; ++j) {
5240+
5241+
const svuint8_t q3bits_sv = svld1_u8(svptrue_pat_b8(SV_VL32), q3_sv); q3_sv += 32;
5242+
svint8_t q8bytes_1_sv_1 = svld1_s8(svptrue_pat_b8(SV_VL32), q8_sv); q8_sv += 32;
5243+
svint8_t q8bytes_1_sv_2 = svld1_s8(svptrue_pat_b8(SV_VL32), q8_sv); q8_sv += 32;
5244+
5245+
q3h_sv = svlsl_n_u8_x(svptrue_pat_b8(SV_VL32), svbic_u8_x(svptrue_pat_b8(SV_VL32), m0_sv, qhbits_sv), 2);
5246+
q3bytes_sv = svsub_s8_x(svptrue_pat_b8(SV_VL32), svreinterpret_s8_u8(svand_u8_m(svptrue_pat_b8(SV_VL32), q3bits_sv, m3b_sv)), svreinterpret_s8_u8(q3h_sv));
5247+
5248+
5249+
svint32_t scale_1 = svsel_s32(svptrue_pat_b32(SV_VL4), svdup_n_s32((int32_t)scale[0]), svdup_n_s32((int32_t)scale[1]));
5250+
sumi1_1 = svmla_s32_m(svptrue_pat_b32(SV_VL8), sumi1_1, svdot_s32(vzero_sv, q3bytes_sv, q8bytes_1_sv_1), scale_1);
5251+
5252+
q3h_sv = svlsl_n_u8_x(svptrue_pat_b8(SV_VL32), svbic_u8_x(svptrue_pat_b8(SV_VL32), m1_sv, qhbits_sv), 1);
5253+
q3bytes_sv = svsub_s8_x(svptrue_pat_b8(SV_VL32), svreinterpret_s8_u8(svand_u8_m(svptrue_pat_b8(SV_VL32), svlsr_n_u8_x(svptrue_pat_b8(SV_VL32), q3bits_sv, 2), m3b_sv)), svreinterpret_s8_u8(q3h_sv));
5254+
5255+
scale_1 = svsel_s32(svptrue_pat_b32(SV_VL4), svdup_n_s32((int32_t)scale[2]), svdup_n_s32((int32_t)scale[3]));
5256+
sumi1_1 = svmla_s32_m(svptrue_pat_b32(SV_VL8), sumi1_1, svdot_s32(vzero_sv, q3bytes_sv, q8bytes_1_sv_2), scale_1);
5257+
5258+
scale += 4;
5259+
q8bytes_1_sv_1 = svld1_s8(svptrue_pat_b8(SV_VL32), q8_sv); q8_sv += 32;
5260+
q8bytes_1_sv_2 = svld1_s8(svptrue_pat_b8(SV_VL32), q8_sv); q8_sv += 32;
5261+
5262+
q3h_sv = svbic_u8_x(svptrue_pat_b8(SV_VL32), m2_sv, qhbits_sv);
5263+
q3bytes_sv = svsub_s8_x(svptrue_pat_b8(SV_VL32), svreinterpret_s8_u8(svand_u8_m(svptrue_pat_b8(SV_VL32), svlsr_n_u8_x(svptrue_pat_b8(SV_VL32), q3bits_sv, 4), m3b_sv)), svreinterpret_s8_u8(q3h_sv));
5264+
5265+
scale_1 = svsel_s32(svptrue_pat_b32(SV_VL4), svdup_n_s32((int32_t)scale[0]), svdup_n_s32((int32_t)scale[1]));
5266+
sumi1_1 = svmla_s32_m(svptrue_pat_b32(SV_VL8), sumi1_1, svdot_s32(vzero_sv, q3bytes_sv, q8bytes_1_sv_1), scale_1);
5267+
5268+
q3h_sv = svlsr_n_u8_x(svptrue_pat_b8(SV_VL32), svbic_u8_x(svptrue_pat_b8(SV_VL32), m3_sv, qhbits_sv), 1);
5269+
q3bytes_sv = svsub_s8_x(svptrue_pat_b8(SV_VL32), svreinterpret_s8_u8(svand_u8_m(svptrue_pat_b8(SV_VL32), svlsr_n_u8_x(svptrue_pat_b8(SV_VL32), q3bits_sv, 6), m3b_sv)), svreinterpret_s8_u8(q3h_sv));
5270+
5271+
scale_1 = svsel_s32(svptrue_pat_b32(SV_VL4), svdup_n_s32((int32_t)scale[2]), svdup_n_s32((int32_t)scale[3]));
5272+
sumi1_1 = svmla_s32_m(svptrue_pat_b32(SV_VL8), sumi1_1, svdot_s32(vzero_sv, q3bytes_sv, q8bytes_1_sv_2), scale_1);
5273+
5274+
if (j == 0) {
5275+
qhbits_sv = svlsr_n_u8_x(svptrue_pat_b8(SV_VL32), qhbits_sv, 4);
5276+
}
5277+
5278+
scale += 4;
5279+
}
5280+
5281+
sum += d * (svaddv_s32(svptrue_pat_b32(SV_VL8), sumi1_1));
5282+
} break;
5283+
default:
5284+
assert(false && "Unsupported vector length");
5285+
break;
5286+
}
5287+
}
5288+
*s = sum;
5289+
5290+
#elif __ARM_NEON
51165291

51175292
uint32_t aux[3];
51185293
uint32_t utmp[4];

0 commit comments

Comments
 (0)