ikawrakow/ik_llama.cpp
1
0
Fork 0
mirror of https://github.com/ikawrakow/ik_llama.cpp.git synced 2026-04-23 16:09:18 +00:00
Code Issues Packages Projects Releases Wiki Activity

q8_k_r16: iq5_ks, iq5_k, and iq6_k now use q8_k_r16 on Zen4+

Browse Source
This commit is contained in:
Iwan Kawrakow
2025-08-20 14:28:15 +03:00
parent ab90830d25
commit be2d5e5fed
1 changed files with 68 additions and 23 deletions
Download Patch File Download Diff File Expand all files Collapse all files

91
ggml/src/iqk/iqk_gemm_iqk_quants.cpp
View File

@@ -2944,14 +2944,21 @@ void iqk_convert_iq4_k_q8_k_r8(int n, const void * vx, size_t bx, void * vy, int
}
void iqk_convert_iq5_ks_q8_k_r8(int n, const void * vx, size_t bx, void * vy, int nrc_x) {
#ifdef HAVE_FANCY_SIMD
constexpr int k_nr = 16;
using block_q8_k_r = block_q8_k_r16;
#else
constexpr int k_nr = 8;
using block_q8_k_r = block_q8_k_r8;
#endif
GGML_ASSERT(n%QK_K == 0);
GGML_ASSERT(nrc_x%8 == 0);
GGML_ASSERT(nrc_x%k_nr == 0);
int nb = n/QK_K;
const block_iq5_ks * x8[8];
const block_iq5_ks * x8[k_nr];
block_q8_k_r8 * y = (block_q8_k_r8 *)vy;
block_q8_k_r * y = (block_q8_k_r *)vy;
__m256i values[2];
{
@@ -2961,8 +2968,8 @@ void iqk_convert_iq5_ks_q8_k_r8(int n, const void * vx, size_t bx, void * vy, in
values[1] = MM256_SET_M128I(v2, v2);
}
float drow[8];
float dnew[8];
float drow[k_nr];
float dnew[k_nr];
int16_t ls[16];
__m256i xv[8];
@@ -2970,15 +2977,19 @@ void iqk_convert_iq5_ks_q8_k_r8(int n, const void * vx, size_t bx, void * vy, in
auto mh = _mm256_set1_epi8(-128); // to avoid stupid warning about 0x80 overflowing
for (int ix = 0; ix < nrc_x; ix += 8) {
for (int k = 0; k < 8; ++k) {
for (int ix = 0; ix < nrc_x; ix += k_nr) {
for (int k = 0; k < k_nr; ++k) {
const float * dptr = (const float *)((const char *)vx + (ix + k)*bx);
drow[k] = dptr[0];
x8[k] = (const block_iq5_ks *)(dptr + 1);
}
#ifdef HAVE_FANCY_SIMD
auto vd = _mm512_loadu_ps(drow);
#else
auto vd = _mm256_loadu_ps(drow);
#endif
for (int i = 0; i < nb; ++i) {
for (int k = 0; k < 8; ++k) {
for (int k = 0; k < k_nr; ++k) {
auto hbits = _mm256_loadu_si256((const __m256i *)x8[k][i].qh);
for (int ib64 = 0; ib64 < 4; ++ib64) {
ls[4*ib64+0] = (x8[k][i].scales[2*ib64+0] & 254) - 127;
@@ -3002,23 +3013,38 @@ void iqk_convert_iq5_ks_q8_k_r8(int n, const void * vx, size_t bx, void * vy, in
xv[2*ib64+1] = _mm256_add_epi8(xv[2*ib64+1], shift2);
hbits = _mm256_srli_epi16(hbits, 2);
}
dnew[k] = convert_to_q8_k_r8(k, 1.f/127, xv, ls, block, y[i].qs);
dnew[k] = convert_to_q8_k_r8<k_nr>(k, 1.f/127, xv, ls, block, y[i].qs);
}
#ifdef HAVE_FANCY_SIMD
_mm256_storeu_si256((__m256i *)y[i].d, _mm512_cvtps_ph(_mm512_mul_ps(vd, _mm512_loadu_ps(dnew)), _MM_ROUND_NEAREST));
for (int l = 0; l < 64; ++l) {
auto v = _mm512_xor_si512(_mm512_loadu_si512((const __m512i *)y[i].qs + l), _mm512_set1_epi8(-128));
_mm512_storeu_si512((__m512i *)y[i].qs + l, v);
}
#else
_mm_storeu_si128((__m128i *)y[i].d, _mm256_cvtps_ph(_mm256_mul_ps(vd, _mm256_loadu_ps(dnew)), _MM_ROUND_NEAREST));
#endif
}
y += nb;
}
}
void iqk_convert_iq5_k_q8_k_r8(int n, const void * vx, size_t bx, void * vy, int nrc_x) {
#ifdef HAVE_FANCY_SIMD
constexpr int k_nr = 16;
using block_q8_k_r = block_q8_k_r16;
#else
constexpr int k_nr = 8;
using block_q8_k_r = block_q8_k_r8;
#endif
GGML_ASSERT(n%QK_K == 0);
GGML_ASSERT(nrc_x%8 == 0);
GGML_ASSERT(nrc_x%k_nr == 0);
int nb = n/QK_K;
const block_iq5_k * x8[8];
const block_iq5_k * x8[k_nr];
block_q8_k_r8 * y = (block_q8_k_r8 *)vy;
block_q8_k_r * y = (block_q8_k_r *)vy;
__m256i values[2];
{
@@ -3034,10 +3060,10 @@ void iqk_convert_iq5_k_q8_k_r8(int n, const void * vx, size_t bx, void * vy, int
auto mh = _mm256_set1_epi8(-128); // to avoid stupid warning about 0x80 overflowing
for (int ix = 0; ix < nrc_x; ix += 8) {
for (int k = 0; k < 8; ++k) x8[k] = (const block_iq5_k *)((const char *)vx + (ix+k)*bx);
for (int ix = 0; ix < nrc_x; ix += k_nr) {
for (int k = 0; k < k_nr; ++k) x8[k] = (const block_iq5_k *)((const char *)vx + (ix+k)*bx);
for (int i = 0; i < nb; ++i) {
for (int k = 0; k < 8; ++k) {
for (int k = 0; k < k_nr; ++k) {
float d = GGML_FP16_TO_FP32(x8[k][i].d);
auto extra = x8[k][i].extra;
auto hbits = _mm256_loadu_si256((const __m256i *)x8[k][i].qh);
@@ -3064,9 +3090,15 @@ void iqk_convert_iq5_k_q8_k_r8(int n, const void * vx, size_t bx, void * vy, int
hbits = _mm256_srli_epi16(hbits, 2);
extra >>= 4;
}
float dnew = convert_to_q8_k_r8(k, 1.f/127, xv, ls, block, y[i].qs);
float dnew = convert_to_q8_k_r8<k_nr>(k, 1.f/127, xv, ls, block, y[i].qs);
y[i].d[k] = GGML_FP32_TO_FP16(d*dnew);
}
#ifdef HAVE_FANCY_SIMD
for (int l = 0; l < 64; ++l) {
auto v = _mm512_xor_si512(_mm512_loadu_si512((const __m512i *)y[i].qs + l), _mm512_set1_epi8(-128));
_mm512_storeu_si512((__m512i *)y[i].qs + l, v);
}
#endif
}
y += nb;
}
@@ -3178,14 +3210,21 @@ void iqk_convert_iq5_k_q8_0_r8(int n, const void * vx, size_t bx, void * vy, int
}
void iqk_convert_iq6_k_q8_k_r8(int n, const void * vx, size_t bx, void * vy, int nrc_x) {
#ifdef HAVE_FANCY_SIMD
constexpr int k_nr = 16;
using block_q8_k_r = block_q8_k_r16;
#else
constexpr int k_nr = 8;
using block_q8_k_r = block_q8_k_r8;
#endif
GGML_ASSERT(n%QK_K == 0);
GGML_ASSERT(nrc_x%8 == 0);
GGML_ASSERT(nrc_x%k_nr == 0);
int nb = n/QK_K;
const block_iq6_k * x8[8];
const block_iq6_k * x8[k_nr];
block_q8_k_r8 * y = (block_q8_k_r8 *)vy;
block_q8_k_r * y = (block_q8_k_r *)vy;
__m256i values[4];
for (int k = 0; k < 4; ++k) {
@@ -3214,10 +3253,10 @@ void iqk_convert_iq6_k_q8_k_r8(int n, const void * vx, size_t bx, void * vy, int
_mm256_and_si256(mask4, _mm256_shuffle_epi8(values[3], l))));
};
for (int ix = 0; ix < nrc_x; ix += 8) {
for (int k = 0; k < 8; ++k) x8[k] = (const block_iq6_k *)((const char *)vx + (ix+k)*bx);
for (int ix = 0; ix < nrc_x; ix += k_nr) {
for (int k = 0; k < k_nr; ++k) x8[k] = (const block_iq6_k *)((const char *)vx + (ix+k)*bx);
for (int i = 0; i < nb; ++i) {
for (int k = 0; k < 8; ++k) {
for (int k = 0; k < k_nr; ++k) {
float d = GGML_FP16_TO_FP32(x8[k][i].d);
helper.vec = _mm256_cvtepi8_epi16(_mm_loadu_si128((const __m128i*)x8[k][i].scales));
auto extra = x8[k][i].extra;
@@ -3243,9 +3282,15 @@ void iqk_convert_iq6_k_q8_k_r8(int n, const void * vx, size_t bx, void * vy, int
xv[4*i128+3] = _mm256_add_epi8(xv[4*i128+3], shift4);
extra >>= 8;
}
float dnew = convert_to_q8_k_r8(k, 1.f/127, xv, helper.val, block, y[i].qs);
float dnew = convert_to_q8_k_r8<k_nr>(k, 1.f/127, xv, helper.val, block, y[i].qs);
y[i].d[k] = GGML_FP32_TO_FP16(d*dnew);
}
#ifdef HAVE_FANCY_SIMD
for (int l = 0; l < 64; ++l) {
auto v = _mm512_xor_si512(_mm512_loadu_si512((const __m512i *)y[i].qs + l), _mm512_set1_epi8(-128));
_mm512_storeu_si512((__m512i *)y[i].qs + l, v);
}
#endif
}
y += nb;
}