8000 arm64: optimize q4_k_q8_k kernel with i8mm (#13886) · ggml-org/llama.cpp@54a2c7a · GitHub
[go: up one dir, main page]
Skip to content

Commit 54a2c7a

Browse files
cyb70289cyb70289
authored
arm64: optimize q4_k_q8_k kernel with i8mm (#13886)
This PR improves q4_k_q8_k gemm kernel with arm64 i8mm instruction. Tested on neoverse-n2 with llama3 8b q4_k_m quantization model. - 34% ~ 50% S_PP uplift for all batch sizes - 12% ~ 37% S_TG uplift for batch size 4 and above Perplexity doesn't change with this PR. ``` // tested on neoverse-n2 $ llama-batched-bench \ -m Meta-Llama-3-8B-Instruct-Q4_K_M.gguf \ --no-mmap -fa \ -c 8192 -b 4096 -ub 512 -npp 128 -ntg 128 \ -npl 1,2,4,8,16,32 \ -t 64 --------------------------------------------------------------------- | PP | TG | B | S_PP t/s | S_TG t/s | | | | | original | this pr | original | this pr | |-------|--------|------|----------|----------|----------|----------| | 128 | 128 | 1 | 110.12 | 147.83 | 24.36 | 24.28 | | 128 | 128 | 2 | 121.16 | 172.42 | 46.36 | 47.93 | | 128 | 128 | 4 | 120.15 | 169.75 | 74.68 | 84.00 | | 128 | 128 | 8 | 130.97 | 196.81 | 91.04 | 114.74 | | 128 | 128 | 16 | 131.01 | 196.88 | 101.43 | 135.79 | | 128 | 128 | 32 | 130.85 | 196.51 | 106.97 | 147.29 | --------------------------------------------------------------------- ```
1 parent 21fcc21 commit 54a2c7a

File tree

2 files changed

+148
-0
lines changed

2 files changed

+148
-0
lines changed

‎ggml/src/ggml-cpu/ggml-cpu-quants.c

Lines changed: 144 additions & 0 deletions
Original file line numberDiff line numberDiff line change
@@ -6995,7 +6995,11 @@ void ggml_vec_dot_q3_K_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const voi
69956995

69966996
void ggml_vec_dot_q4_K_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, size_t bx, const void * GGML_RESTRICT vy, size_t by, int nrc) {
69976997
assert(n % QK_K == 0);
6998+
#ifdef __ARM_FEATURE_MATMUL_INT8
6999+
assert((nrc == 2) || (nrc == 1));
7000+
#else
69987001
assert(nrc == 1);
7002+
#endif
69997003
UNUSED(nrc);
70007004
UNUSED(bx);
70017005
UNUSED(by);
@@ -7012,6 +7016,146 @@ void ggml_vec_dot_q4_K_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const voi
70127016

70137017
uint32_t utmp[4];
70147018

7019+
#if defined(__ARM_FEATURE_MATMUL_INT8)
7020+
if (nrc == 2) {
7021+
const block_q4_K * GGML_RESTRICT x0 = x;
7022+
const block_q4_K * GGML_RESTRICT x1 = (const block_q4_K *) ((const uint8_t *)vx + bx);
7023+
const block_q8_K * GGML_RESTRICT y0 = y;
7024+
const block_q8_K * GGML_RESTRICT y1 = (const block_q8_K *) ((const uint8_t *)vy + by);
7025+
7026+
const uint8x16_t m4b = vdupq_n_u8(0x0f);
7027+
7028+
float32x4_t vfsum = vdupq_n_f32(0.0f);
7029+
7030+
for (int i = 0; i < nb; ++i, ++x0, ++x1, ++y0, ++y1) {
7031+
const uint8_t * GGML_RESTRICT qx0 = x0->qs;
7032+
const uint8_t * GGML_RESTRICT qx1 = x1->qs;
7033+
const int8_t * GGML_RESTRICT qy0 = y0->qs;
7034+
const int8_t * GGML_RESTRICT qy1 = y1->qs;
7035+
7036+
// decode scales and mins
7037+
int8_t x0_scales[8], x1_scales[8];
7038+
int16x8_t x0_mins, x1_mins;
7039+
{
7040+
uint32_t scales_mins[3];
7041+
memcpy(scales_mins, x0->scales, 12);
7042+
const uint32_t mins_0_3 = scales_mins[1] & kmask1;
7043+
const uint32_t mins_4_7 = ((scales_mins[2] >> 4) & kmask2) | (((scales_mins[1] >> 6) & kmask3) << 4);
7044+
const uint32x2_t mins = {mins_0_3, mins_4_7};
7045+
x0_mins = vreinterpretq_s16_u16(vmovl_u8(vreinterpret_u8_u32(mins)));
7046+
uint32_t scales[2];
7047+
scales[0] = scales_mins[0] & kmask1; // scales 0~3
7048+
scales[1] = (scales_mins[2] & kmask2) | (((scales_mins[0] >> 6) & kmask3) << 4); // scales 4~7
7049+
memcpy(x0_scales, scales, 8);
7050+
}
7051+
{
7052+
uint32_t scales_mins[3];
7053+
memcpy(scales_mins, x1->scales, 12);
7054+
const uint32_t mins_0_3 = scales_mins[1] & kmask1;
7055+
const uint32_t mins_4_7 = ((scales_mins[2] >> 4) & kmask2) | (((scales_mins[1] >> 6) & kmask3) << 4);
7056+
const uint32x2_t mins = {mins_0_3, mins_4_7};
7057+
x1_mins = vreinterpretq_s16_u16(vmovl_u8(vreinterpret_u8_u32(mins)));
7058+
uint32_t scales[2];
7059+
scales[0] = scales_mins[0] & kmask1; // scales 0~3
7060+
scales[1] = (scales_mins[2] & kmask2) | (((scales_mins[0] >> 6) & kmask3) << 4); // scales 4~7
7061+
memcpy(x1_scales, scales, 8);
7062+
}
7063+
7064+
int32x4_t visum = {0};
7065+
7066+
// process 64 data points per iteration, totally 256 data points
7067+
for (int j = 0; j < QK_K / 64; ++j, qx0 += 32, qx1 += 32, qy0 += 64, qy1 += 64) {
7068+
const int8x16x4_t vy0 = vld1q_s8_x4(qy0);
7069+
const int8x16x4_t vy1 = vld1q_s8_x4(qy1);
7070+
7071+
int8x16_t vx0[4], vx1[4];
7072+
{
7073+
const uint8x16x2_t vv = vld1q_u8_x2(qx0);
7074+
vx0[0] = vreinterpretq_s8_u8(vandq_u8(vv.val[0], m4b));
7075+
vx0[1] = vreinterpretq_s8_u8(vandq_u8(vv.val[1], m4b));
7076+
vx0[2] = vreinterpretq_s8_u8(vshrq_n_u8(vv.val[0], 4));
7077+
vx0[3] = vreinterpretq_s8_u8(vshrq_n_u8(vv.val[1], 4));
7078+
}
7079+
{
7080+
const uint8x16x2_t vv = vld1q_u8_x2(qx1);
7081+
vx1[0] = vreinterpretq_s8_u8(vandq_u8(vv.val[0], m4b));
7082+
vx1[1] = vreinterpretq_s8_u8(vandq_u8(vv.val[1], m4b));
7083+
vx1[2] = vreinterpretq_s8_u8(vshrq_n_u8(vv.val[0], 4));
7084+
vx1[3] = vreinterpretq_s8_u8(vshrq_n_u8(vv.val[1], 4));
7085+
}
7086+
7087+
// process 32 data points (share same block scale) per iteration
7088+
for (int k = 0; k < 2; ++k) {
7089+
const int blk = j * 2 + k;
7090+
const int32x4_t block_scale = {
7091+
x0_scales[blk],
7092+
x0_scales[blk],
7093+
x1_scales[blk],
7094+
x1_scales[blk],
7095+
};
7096+
7097+
int32x4_t vr = {0};
7098+
for (int l = 0; l < 2; ++l) {
7099+
const int idx = k * 2 + l;
7100+
const int64x2_t vx0_s64 = vreinterpretq_s64_s8(vx0[idx]);
7101+
const int64x2_t vx1_s64 = vreinterpretq_s64_s8(vx1[idx]);
7102+
const int64x2_t vy0_s64 = vreinterpretq_s64_s8(vy0.val[idx]);
7103+
const int64x2_t vy1_s64 = vreinterpretq_s64_s8(vy1.val[idx]);
7104+
const int8x16_t vx_l = vreinterpretq_s8_s64(vzip1q_s64(vx0_s64, vx1_s64));
7105+
const int8x16_t vx_h = vreinterpretq_s8_s64(vzip2q_s64(vx0_s64, vx1_s64));
7106+
const int8x16_t vy_l = vreinterpretq_s8_s64(vzip1q_s64(vy0_s64, vy1_s64));
7107+
const int8x16_t vy_h = vreinterpretq_s8_s64(vzip2q_s64(vy0_s64, vy1_s64));
7108+
vr = vmmlaq_s32(vr, vx_l, vy_l);
7109+
vr = vmmlaq_s32(vr, vx_h, vy_h);
7110+
}
7111+
// apply block scale, will NOT overflow
7112+
// block_scale * sum_256(int4*int8) <= 2^(8+8+4+8) = 28 bits
7113+
visum = vmlaq_s32(visum, vr, block_scale);
7114+
}
7115+
}
7116+
7117+
// adjust bias, apply superblock scale
7118+
{
7119+
int32_t bias[4];
7120+
// no obvious uplift from sve sdot-16, just use neon mul add
7121+
const int16x8_t y0_sums = vpaddq_s16(vld1q_s16(y0->bsums), vld1q_s16(y0->bsums+8));
7122+
const int16x8_t y1_sums = vpaddq_s16(vld1q_s16(y1->bsums), vld1q_s16(y1->bsums+8));
7123+
bias[0] = vaddvq_s32(vaddq_s32(vmull_s16(vget_low_s16(y0_sums), vget_low_s16(x0_mins)),
7124+
vmull_s16(vget_high_s16(y0_sums), vget_high_s16(x0_mins))));
7125+
bias[1] = vaddvq_s32(vaddq_s32(vmull_s16(vget_low_s16(y1_sums), vget_low_s16(x0_mins)),
7126+
vmull_s16(vget_high_s16(y1_sums), vget_high_s16(x0_mins))));
7127+
bias[2] = vaddvq_s32(vaddq_s32(vmull_s16(vget_low_s16(y0_sums), vget_low_s16(x1_mins)),
7128+
vmull_s16(vget_high_s16(y0_sums), vget_high_s16(x1_mins))));
7129+
bias[3] = vaddvq_s32(vaddq_s32(vmull_s16(vget_low_s16(y1_sums), vget_low_s16(x1_mins)),
7130+
vmull_s16(vget_high_s16(y1_sums), vget_high_s16(x1_mins))));
7131+
const float32x4_t dmins = {
7132+
GGML_FP16_TO_FP32(x0->dmin) * y0->d,
7133+
GGML_FP16_TO_FP32(x0->dmin) * y1->d,
7134+
GGML_FP16_TO_FP32(x1->dmin) * y0->d,
7135+
GGML_FP16_TO_FP32(x1->dmin) * y1->d,
7136+
};
7137+
vfsum = vmlsq_f32(vfsum, vcvtq_f32_s32(vld1q_s32(bias)), dmins);
7138+
7139+
const float32x4_t superblock_scale = {
7140+
GGML_FP16_TO_FP32(x0->d) * y0->d,
7141+
GGML_FP16_TO_FP32(x0->d) * y1->d,
7142+
GGML_FP16_TO_FP32(x1->d) * y0->d,
7143+
GGML_FP16_TO_FP32(x1->d) * y1->d,
7144+
};
7145+
vfsum = vmlaq_f32(vfsum, vcvtq_f32_s32(visum), superblock_scale);
7146+
}
7147+
}
7148+
7149+
// vfsum = ABCD -> ACBD
7150+
// AC -> s, BD -> (s+bs)
7151+
vfsum = vzip1q_f32(vfsum, vextq_f32(vfsum, vfsum, 2));
7152+
vst1_f32(s, vget_low_f32 (vfsum));
7153+
vst1_f32(s + bs, vget_high_f32(vfsum));
7154+
7155+
return;
7156+
}
7157+
#endif
7158+
70157159
#ifdef __ARM_FEATURE_SVE
70167160
float sumf = 0;
70177161
for (int i = 0; i < nb; ++i) {

‎ggml/src/ggml-cpu/ggml-cpu.c

Lines changed: 4 additions & 0 deletions
Original file line numberDiff line numberDiff line change
@@ -270,7 +270,11 @@ static const struct ggml_type_traits_cpu type_traits_cpu[GGML_TYPE_COUNT] = {
270270
.from_float = quantize_row_q4_K,
271271
.vec_dot = ggml_vec_dot_q4_K_q8_K,
272272
.vec_dot_type = GGML_TYPE_Q8_K,
273+
#if defined (__ARM_FEATURE_MATMUL_INT8)
274+
.nrows = 2,
275+
#else
273276
.nrows = 1,
277+
#endif
274278
},
275279
[GGML_TYPE_Q5_K] = {
276280
.from_float = quantize_row_q5_K,

0 commit comments

Comments
 (0)
0