ikawrakow/ik_llama.cpp
1
0
Fork 0
mirror of https://github.com/ikawrakow/ik_llama.cpp.git synced 2026-04-30 19:31:48 +00:00
Code Issues Packages Projects Releases Wiki Activity

q6_k_r4: slightly faster NEON

Browse Source 
PP-512(LLaMA-3.1-8B) = 83.25 t/s
This commit is contained in:
Iwan Kawrakow
2024-12-10 10:14:24 +01:00
parent 99d79116ee
commit 8de44659ca
1 changed files with 37 additions and 30 deletions
Download Patch File Download Diff File Expand all files Collapse all files

67
ggml/src/iqk/iqk_mul_mat.cpp
View File

@@ -163,11 +163,14 @@ struct MulMat {
static bool prepare(int typeA, int typeB, int ne00, MulMat& mm, int Ny); static bool prepare(int typeA, int typeB, int ne00, MulMat& mm, int Ny);
static inline int num_rows(ggml_type type) { static inline int num_rows(ggml_type type) {
switch (type) { switch (type) {
case GGML_TYPE_Q4_K_R4:
case GGML_TYPE_Q6_K_R4:
case GGML_TYPE_Q4_0_R4: case GGML_TYPE_Q4_0_R4:
case GGML_TYPE_Q5_0_R4: case GGML_TYPE_Q5_0_R4:
case GGML_TYPE_Q6_0_R4: case GGML_TYPE_Q6_0_R4:
case GGML_TYPE_Q8_0_R4: case GGML_TYPE_Q8_0_R4:
case GGML_TYPE_IQ4_NL_R4: case GGML_TYPE_IQ4_NL_R4:
case GGML_TYPE_IQ4_XS_R4:
case GGML_TYPE_IQ2_BN_R4: return 4; case GGML_TYPE_IQ2_BN_R4: return 4;
default: return 1; default: return 1;
} }
@@ -7968,6 +7971,15 @@ IQK_ALWAYS_INLINE int32x4x2_t interleaved_dotq_b16(const int8x16_t * qx, const i
return sumi; return sumi;
} }
IQK_ALWAYS_INLINE int32x4_t interleaved_dotq(const int8x16_t * qx, const int8x16_t& y) {
auto sumi = vdupq_n_s32(0);
sumi = vdotq_laneq_s32(sumi, qx[0], y, 0);
sumi = vdotq_laneq_s32(sumi, qx[1], y, 1);
sumi = vdotq_laneq_s32(sumi, qx[2], y, 2);
sumi = vdotq_laneq_s32(sumi, qx[3], y, 3);
return sumi;
}
IQK_ALWAYS_INLINE void prepare_iq4_nl_quants(const int8x16_t& values, const uint8x16_t& m4, const uint8x16x4_t& bits, int8x16_t * qx) { IQK_ALWAYS_INLINE void prepare_iq4_nl_quants(const int8x16_t& values, const uint8x16_t& m4, const uint8x16x4_t& bits, int8x16_t * qx) {
qx[0] = vqtbl1q_s8(values, vandq_u8(bits.val[0], m4)); // 0...3 from the 4 rows qx[0] = vqtbl1q_s8(values, vandq_u8(bits.val[0], m4)); // 0...3 from the 4 rows
qx[1] = vqtbl1q_s8(values, vandq_u8(bits.val[1], m4)); // 16..19 qx[1] = vqtbl1q_s8(values, vandq_u8(bits.val[1], m4)); // 16..19
@@ -8121,48 +8133,43 @@ void mul_mat_q6_k_r4_q8_k(int n, const void * vx, size_t bx, const DataInfo& inf
auto mf = vdupq_n_u8(0x0f); auto mf = vdupq_n_u8(0x0f);
auto m3 = vdupq_n_u8(0x30); auto m3 = vdupq_n_u8(0x30);
auto m32 = vdupq_n_s8(-32); auto m32 = vdupq_n_s8(-32);
//auto mx = vdupq_n_u8(0xaa);
int nbl = n / QK_K; int nbl = n / QK_K;
int8x16_t qx[8]; int8x16_t qx[4];
float32x4x2_t scales; float32x4x2_t scales;
float32x4_t acc[nrc_y] = {}; float32x4_t acc[nrc_y] = {};
float32x4_t d4[nrc_y] = {};
for (int ix = 0; ix < nrc_x; ix += 4) { for (int ix = 0; ix < nrc_x; ix += 4) {
const block_q6_k_r4 * iq6 = (const block_q6_k_r4 *)((const char *)vx + ix*bx); const block_q6_k_r4 * iq6 = (const block_q6_k_r4 *)((const char *)vx + ix*bx);
for (int ibl = 0; ibl < nbl; ++ibl) { for (int ibl = 0; ibl < nbl; ++ibl) {
auto d4 = vcvt_f32_f16(vld1_f16((const float16_t *)iq6[ibl].d)); auto dtmp = vcvt_f32_f16(vld1_f16((const float16_t *)iq6[ibl].d));
if constexpr (nrc_y == 1) { for (int iy = 0; iy < nrc_y; ++iy) {
d4 = vmulq_f32(d4, vdupq_n_f32(q8.scale(0, ibl))); d4[iy] = vmulq_f32(dtmp, vdupq_n_f32(q8.scale(iy, ibl)));
} }
for (int is = 0; is < 2; ++is) { for (int is = 0; is < 2; ++is) {
for (int ib = 0; ib < 4; ++ib) { for (int ib = 0; ib < 4; ++ib) {
auto lbits = vld1q_u8_x4(iq6[ibl].ql + 256*is + 64*ib); auto lbits = vld1q_u8_x4(iq6[ibl].ql + 256*is + 64*ib);
auto hbits = vld1q_u8_x2(iq6[ibl].qh + 128*is + 32*ib); auto hbits = vld1q_u8(iq6[ibl].qh + 128*is + 32*ib);
auto iscales = vmovl_s8(vld1_s8(iq6[ibl].scales + 32*is + 8*ib)); auto iscales = vmovl_s8(vld1_s8(iq6[ibl].scales + 32*is + 8*ib));
scales.val[0] = vmulq_f32(d4, vcvtq_f32_s32(vmovl_s16(vget_low_s16(iscales)))); scales.val[0] = vcvtq_f32_s32(vmovl_s16(vget_low_s16(iscales)));
scales.val[1] = vmulq_f32(d4, vcvtq_f32_s32(vmovl_s16(vget_high_s16(iscales)))); scales.val[1] = vcvtq_f32_s32(vmovl_s16(vget_high_s16(iscales)));
//hbits.val[0] = veorq_u8(hbits.val[0], mx); qx[0] = vaddq_s8(m32, vorrq_u8(vandq_u8 (lbits.val[0], mf), vandq_u8(m3, vshlq_n_u8(hbits, 4))));
//hbits.val[1] = veorq_u8(hbits.val[1], mx); qx[1] = vaddq_s8(m32, vorrq_u8(vandq_u8 (lbits.val[2], mf), vandq_u8(m3, hbits)));
//qx[0] = vorrq_u8(vandq_u8(vshlq_n_u8(hbits.val[0], 6), m3), vandq_u8(vshlq_n_u8(lbits.val[0], 4), mf)); qx[2] = vaddq_s8(m32, vorrq_u8(vshrq_n_u8(lbits.val[0], 4), vandq_u8(m3, vshlq_n_u8(hbits, 2))));
prepare_q4_k_quants(mf, lbits, qx); qx[3] = vaddq_s8(m32, vorrq_u8(vshrq_n_u8(lbits.val[2], 4), vandq_u8(m3, vshrq_n_u8(hbits, 2))));
qx[0] = vaddq_s8(m32, vorrq_u8(qx[0], vandq_u8(m3, vshlq_n_u8(hbits.val[0], 4))));
qx[1] = vaddq_s8(m32, vorrq_u8(qx[1], vandq_u8(m3, vshlq_n_u8(hbits.val[1], 4))));
qx[2] = vaddq_s8(m32, vorrq_u8(qx[2], vandq_u8(m3, hbits.val[0])));
qx[3] = vaddq_s8(m32, vorrq_u8(qx[3], vandq_u8(m3, hbits.val[1])));
qx[4] = vaddq_s8(m32, vorrq_u8(qx[4], vandq_u8(m3, vshlq_n_u8(hbits.val[0], 2))));
qx[5] = vaddq_s8(m32, vorrq_u8(qx[5], vandq_u8(m3, vshlq_n_u8(hbits.val[1], 2))));
qx[6] = vaddq_s8(m32, vorrq_u8(qx[6], vandq_u8(m3, vshrq_n_u8(hbits.val[0], 2))));
qx[7] = vaddq_s8(m32, vorrq_u8(qx[7], vandq_u8(m3, vshrq_n_u8(hbits.val[1], 2))));
for (int iy = 0; iy < nrc_y; ++iy) { for (int iy = 0; iy < nrc_y; ++iy) {
auto y = vld1q_s8_x2(q8.y[iy][ibl].qs+128*is+32*ib); auto y = vld1q_s8(q8.y[iy][ibl].qs+128*is+32*ib);
auto sumi = interleaved_dotq_b16(qx, y); auto sumi = interleaved_dotq(qx, y);
if constexpr (nrc_y == 1) { acc[iy] = vfmaq_f32(acc[iy], vmulq_f32(scales.val[0], d4[iy]), vcvtq_f32_s32(sumi));
acc[iy] = vfmaq_f32(acc[iy], scales.val[0], vcvtq_f32_s32(sumi.val[0])); }
acc[iy] = vfmaq_f32(acc[iy], scales.val[1], vcvtq_f32_s32(sumi.val[1])); hbits = vld1q_u8(iq6[ibl].qh + 128*is + 32*ib + 16);
} else { qx[0] = vaddq_s8(m32, vorrq_u8(vandq_u8 (lbits.val[1], mf), vandq_u8(m3, vshlq_n_u8(hbits, 4))));
auto d8 = vdupq_n_f32(q8.scale(iy, ibl)); qx[1] = vaddq_s8(m32, vorrq_u8(vandq_u8 (lbits.val[3], mf), vandq_u8(m3, hbits)));
acc[iy] = vfmaq_f32(acc[iy], vmulq_f32(scales.val[0], d8), vcvtq_f32_s32(sumi.val[0])); qx[2] = vaddq_s8(m32, vorrq_u8(vshrq_n_u8(lbits.val[1], 4), vandq_u8(m3, vshlq_n_u8(hbits, 2))));
acc[iy] = vfmaq_f32(acc[iy], vmulq_f32(scales.val[1], d8), vcvtq_f32_s32(sumi.val[1])); qx[3] = vaddq_s8(m32, vorrq_u8(vshrq_n_u8(lbits.val[3], 4), vandq_u8(m3, vshrq_n_u8(hbits, 2))));
} for (int iy = 0; iy < nrc_y; ++iy) {
auto y = vld1q_s8(q8.y[iy][ibl].qs+128*is+32*ib+16);
auto sumi = interleaved_dotq(qx, y);
acc[iy] = vfmaq_f32(acc[iy], vmulq_f32(scales.val[1], d4[iy]), vcvtq_f32_s32(sumi));
} }
} }
} }