ikawrakow/ik_llama.cpp
1
0
Fork 0
mirror of https://github.com/ikawrakow/ik_llama.cpp.git synced 2026-01-27 01:29:51 +00:00
Code Issues Packages Projects Releases Wiki Activity
Files
7cb77d7a67a85ae59040bb3572f808e0847eb094
ik_llama.cpp/iqk-quantize.cpp
Kawrakow b0967ffa79 bitnet: fix scalar dot product 
I had forgotten to adjust for the change to q8_K64.
On the M2 I'm getting 10.8 t/s with the scalar version!
2024-06-22 12:02:51 +03:00

436 lines
17 KiB
C++
Raw Blame History

#include "ggml-quants.h"
#include "ggml-impl.h"
#define GGML_COMMON_IMPL_C
#include "ggml-common.h"
#include <vector>
#include <utility>
#include <cstdint>
#include <cmath>
#include <array>
#include <algorithm>
#include <cstring>
#include <mutex>
namespace {
inline int nearest_int(float fval) {
assert(fval <= 4194303.f);
float val = fval + 12582912.f;
int i; memcpy(&i, &val, sizeof(int));
return (i & 0x007fffff) - 0x00400000;
}
struct IQ1BNData {
IQ1BNData();
std::vector<std::pair<int16_t, bool>> map;
std::vector<uint16_t> rmap;
};
const IQ1BNData& get_iq1bn_data() {
static std::mutex mutex;
std::lock_guard<std::mutex> lock(mutex);
static IQ1BNData iq1bn;
return iq1bn;
}
IQ1BNData::IQ1BNData() {
map.resize(1 << 16, {int16_t(-1), false});
uint64_t aux64;
uint8_t * aux8 = (uint8_t *)&aux64;
std::vector<uint64_t> values;
values.reserve(6561);
rmap.reserve(6561);
for (int i = 0; i < (1 << 16); ++i) {
bool is_good = true;
for (int j = 0; j < 8; ++j) {
aux8[j] = (i >> 2*j) & 3;
if (aux8[j] == 3u) { is_good = false; break; }
}
if (!is_good) continue;
auto orig = aux64;
for (int j = 0; j < 8; ++j) aux8[j] = 2 - aux8[j];
int k = 0;
for (; k < int(values.size()); ++k) {
if (values[k] == aux64) break;
}
if (k < int(values.size())) {
map[i] = {k, true};
} else {
map[i].first = values.size();
values.push_back(orig);
rmap.push_back(i);
}
}
printf("==================== %s: initialized %d grid points\n", __func__, int(rmap.size()));
}
struct IQ1BNQuantizer {
typedef union {
float f;
uint32_t i;
} scale_t;
constexpr static int block_size = QK_IQ1BN;
int8_t L[QK_IQ1BN];
void quantize_one_row(const float * src, block_iq1_bn * y, int n_per_row, const float * imatrix);
};
void IQ1BNQuantizer::quantize_one_row(const float * src, block_iq1_bn * y, int n_per_row, const float * imatrix) {
(void)imatrix;
constexpr int Nk = block_size/8;
const int nblock = n_per_row/QK_IQ1BN;
const auto& iq1bn = get_iq1bn_data();
float max_in_row = 0;
for (int j = 0; j < n_per_row; ++j) {
float ax = fabsf(src[j]);
max_in_row = std::max(max_in_row, ax);
}
max_in_row *= 1.03125f; // i.e., round to nearest in our fp8 representation
scale_t s;
uint8_t u = 0;
if (max_in_row > 1.9074e-06f && max_in_row < 0.12109f) {
s.f = max_in_row;
u = ((((s.i >> 23) + 132) & 0xf) << 4) | ((s.i >> 19) & 0xf);
s.i = ((((u >> 4) | 0xf0) - 132) << 23) | ((u & 0x0f) << 19);
} else {
// outside the allowed range. Small values we can habdle via quants set to zero, so we only warn about too large values
if (max_in_row >= 0.12109f) {
u = 255;
fprintf(stderr, "%s: found scale %g, which is outside the range of out fp8 representation\n", __func__, max_in_row);
} else{
u = 0;
}
}
for (int ib = 0; ib < nblock; ++ib) {
std::memset(&y[ib], 0, sizeof(block_iq1_bn));
auto xb = src + QK_IQ1BN*ib;
for (int j = 0; j < QK_IQ1BN; ++j) {
L[j] = fabsf(xb[j]) < 1e-6f ? 1 : xb[j] < 0 ? 0 : 2;
}
auto ql = y[ib].ql;
auto qh = y[ib].qh;
uint16_t extra = 0;
for (int k = 0; k < Nk; ++k) {
auto Lk = L + 8*k;
uint16_t u = 0;
for (int j = 0; j < 8; ++j) u |= (Lk[j] << 2*j);
auto& val = iq1bn.map[u];
GGML_ASSERT(val.first >= 0);
ql[k] = val.first & 255;
qh[k/2] |= (val.first >> 8) << 4*(k%2);
if (val.second) extra |= (1 << k);
}
y[ib].extra = u | (extra << 8);
}
}
}
void iq1bn_init_impl(void) {
get_iq1bn_data();
}
size_t quantize_iq1_bn(const float * src, void * dst, int64_t nrows, int64_t n_per_row, const float * imatrix) {
IQ1BNQuantizer iq1bn;
int nblock = n_per_row/QK_IQ1BN;
block_iq1_bn * y = (block_iq1_bn *)dst;
for (int row = 0; row < nrows; ++row) {
iq1bn.quantize_one_row(src + row*n_per_row, y, n_per_row, imatrix);
y += nblock;
}
return sizeof(block_iq1_bn)*nblock*nrows;
}
void quantize_row_iq1_bn_reference(const float * x, block_iq1_bn * y, int64_t k) {
quantize_iq1_bn(x, y, 1, k, nullptr);
}
void quantize_row_iq1_bn(const float * x, void * y, int64_t k) {
quantize_iq1_bn(x, y, 1, k, nullptr);
}
void dequantize_row_iq1_bn(const block_iq1_bn * x, float * y, int64_t k) {
assert(k%QK_IQ1BN == 0);
int nblock = k / QK_IQ1BN;
IQ1BNQuantizer::scale_t s;
for (int i = 0; i < nblock; ++i) {
uint16_t u = x[i].extra & 0xff;
s.i = ((((u >> 4) | 0xf0) - 132) << 23) | ((u & 0x0f) << 19);
float d = s.f;
uint8_t extra = x[i].extra >> 8;
auto qh = x[i].qh;
auto ql = x[i].ql;
for (int k = 0; k < QK_IQ1BN/8; ++k) {
uint16_t idx = ql[k] | ((qh[k/2] << (8 - 4*(k%2))) & 0x0f00);
uint16_t val = iq1bn_grid_u16[idx];
float dls = extra & (1 << k) ? -d : d;
for (int j = 0; j < 8; ++j) y[j] = dls * (((val >> 2*j) & 3) - 1);
y += 8;
}
}
}
#if __AVX__ || __AVX2__ || __AVX512F__
// horizontally add 8 floats
static inline float hsum_float_8(const __m256 x) {
__m128 res = _mm256_extractf128_ps(x, 1);
res = _mm_add_ps(res, _mm256_castps256_ps128(x));
res = _mm_add_ps(res, _mm_movehl_ps(res, res));
res = _mm_add_ss(res, _mm_movehdup_ps(res));
return _mm_cvtss_f32(res);
}
#endif
void ggml_vec_dot_iq1_bn_q8_0 (int n, float * s, size_t bs, const void * vx, size_t bx, const void * vy, size_t by, int nrc) {
GGML_UNUSED(bs);
GGML_UNUSED(bx);
GGML_UNUSED(by);
GGML_UNUSED(nrc);
static_assert(QK_IQ1BN == 64, "This dot product implementation for iq1_bn requires a block size of 64");
const block_iq1_bn * x = (const block_iq1_bn *)vx;
const block_q8_0 * y = (const block_q8_0 *)vy;
int nblock = n / QK_IQ1BN;
float sumf = 0;
IQ1BNQuantizer::scale_t scale;
#if defined __AVX2__
const auto m1_8 = _mm256_set1_epi8(1);
const auto shuff1 = _mm256_set_epi64x(0x0808080808080808, 0x0000000000000000, 0x0808080808080808, 0x0000000000000000);
const auto shuff2 = _mm256_add_epi8(shuff1, m1_8);
const auto shuff3 = _mm256_set_epi64x(0x0303030303030303, 0x0202020202020202, 0x0101010101010101, 0x0000000000000000);
const auto shuff4 = _mm256_set_epi64x(0x0707070707070707, 0x0606060606060606, 0x0505050505050505, 0x0404040404040404);
const auto mask1 = _mm256_set1_epi64x(0x8040201008040201);
#if !(defined __AVX512VNNI__ && defined __AVX512VL__)
const auto m1_16 = _mm256_set1_epi16(1);
#endif
__m256 acc1 = _mm256_setzero_ps();
__m256 acc2 = _mm256_setzero_ps();
// All scales are the same in BitNet!
uint16_t u = x[0].extra & 0xff;
scale.i = ((((u >> 4) | 0xf0) - 132) << 23) | ((u & 0x0f) << 19);
for (int i = 0; i < nblock; ++i) {
// We would uncomment this if we wanted to use this implementation for a model that has per block scales
//uint16_t u = x[i].extra & 0xff;
//scale.i = ((((u >> 4) | 0xf0) - 132) << 23) | ((u & 0x0f) << 19);
auto signs = _mm256_set1_epi8(x[i].extra >> 8);
// signs for groups of 8 ordered as 0, 1, 2, 3, 4, 5, 6, 7, 0, 1, 2, 3, 4, 5, 6, 7, ...
// To use these to sign the q8 values we need
// 0 0 0 0 0 0 0 0 1 1 1 1 1 1 1 1 2 2 2 2 2 2 2 2 3 3 3 3 3 3 3 3 amd the same for 4...7
signs = _mm256_or_si256(_mm256_cmpeq_epi8(_mm256_and_si256(signs, mask1), mask1), m1_8);
auto q8_1 = _mm256_sign_epi8(_mm256_loadu_si256((const __m256i *)y[2*i+0].qs), _mm256_shuffle_epi8(signs, shuff3));
auto q8_2 = _mm256_sign_epi8(_mm256_loadu_si256((const __m256i *)y[2*i+1].qs), _mm256_shuffle_epi8(signs, shuff4));
auto ql = x[i].ql;
auto qh = x[i].qh;
auto aux1 = _mm256_set_epi64x(iq1bn_grid_xxx[ql[3] | ((qh[1] << 4) & 0x0f00)], iq1bn_grid_xxx[ql[2] | ((qh[1] << 8) & 0x0f00)],
iq1bn_grid_xxx[ql[1] | ((qh[0] << 4) & 0x0f00)], iq1bn_grid_xxx[ql[0] | ((qh[0] << 8) & 0x0f00)]);
auto aux2 = _mm256_set_epi64x(iq1bn_grid_xxx[ql[7] | ((qh[3] << 4) & 0x0f00)], iq1bn_grid_xxx[ql[6] | ((qh[3] << 8) & 0x0f00)],
iq1bn_grid_xxx[ql[5] | ((qh[2] << 4) & 0x0f00)], iq1bn_grid_xxx[ql[4] | ((qh[2] << 8) & 0x0f00)]);
auto v1_p = _mm256_cmpeq_epi8(_mm256_and_si256(_mm256_shuffle_epi8(aux1, shuff1), mask1), mask1);
auto v1_m = _mm256_cmpeq_epi8(_mm256_and_si256(_mm256_shuffle_epi8(aux1, shuff2), mask1), mask1);
auto v2_p = _mm256_cmpeq_epi8(_mm256_and_si256(_mm256_shuffle_epi8(aux2, shuff1), mask1), mask1);
auto v2_m = _mm256_cmpeq_epi8(_mm256_and_si256(_mm256_shuffle_epi8(aux2, shuff2), mask1), mask1);
auto dot1 = _mm256_sub_epi8(_mm256_sign_epi8(q8_1, v1_m), _mm256_sign_epi8(q8_1, v1_p));
auto dot2 = _mm256_sub_epi8(_mm256_sign_epi8(q8_2, v2_m), _mm256_sign_epi8(q8_2, v2_p));
#if defined __AVX512VNNI__ && defined __AVX512VL__
dot1 = _mm256_dpbusd_epi32(_mm256_setzero_si256(), m1_8, dot1);
dot2 = _mm256_dpbusd_epi32(_mm256_setzero_si256(), m1_8, dot2);
#else
dot1 = _mm256_madd_epi16(m1_16, _mm256_maddubs_epi16(m1_8, dot1));
dot2 = _mm256_madd_epi16(m1_16, _mm256_maddubs_epi16(m1_8, dot2));
#endif
// We would uncomment this if we wanted to use this implementation for a model that has per block scales
//acc1 = _mm256_fmadd_ps(_mm256_set1_ps(scale.f*GGML_FP16_TO_FP32(y[2*i+0].d)), _mm256_cvtepi32_ps(dot1), acc1);
//acc2 = _mm256_fmadd_ps(_mm256_set1_ps(scale.f*GGML_FP16_TO_FP32(y[2*i+1].d)), _mm256_cvtepi32_ps(dot2), acc2);
// All scales are the same for BitNet!
// This is slower
//uint32_t aux32 = y[2*i+0].d | (y[2*i+1].d << 16);
//auto d8 = _mm256_cvtph_ps(_mm_set1_epi32(aux32));
//acc1 = _mm256_fmadd_ps(_mm256_permute_ps(d8, 0x00), _mm256_cvtepi32_ps(dot1), acc1);
//acc2 = _mm256_fmadd_ps(_mm256_permute_ps(d8, 0x55), _mm256_cvtepi32_ps(dot2), acc2);
acc1 = _mm256_fmadd_ps(_mm256_set1_ps(GGML_FP16_TO_FP32(y[2*i+0].d)), _mm256_cvtepi32_ps(dot1), acc1);
acc2 = _mm256_fmadd_ps(_mm256_set1_ps(GGML_FP16_TO_FP32(y[2*i+1].d)), _mm256_cvtepi32_ps(dot2), acc2);
}
//sumf = hsum_float_8(_mm256_add_ps(acc1, acc2));
sumf = scale.f * hsum_float_8(_mm256_add_ps(acc1, acc2));
#else
for (int i = 0; i < nblock; ++i) {
uint16_t u = x[i].extra & 0xff;
scale.i = ((((u >> 4) | 0xf0) - 132) << 23) | ((u & 0x0f) << 19);
uint8_t extra = x[i].extra >> 8;
auto qh = x[i].qh;
auto ql = x[i].ql;
auto q8 = y[2*i+0].qs;
int16_t sumi1 = 0;
for (int k = 0; k < 4; ++k) {
uint16_t idx = ql[k] | ((qh[k/2] << (8 - 4*(k%2))) & 0x0f00);
uint16_t val = iq1bn_grid_u16[idx];
int16_t sl = 0;
for (int j = 0; j < 8; ++j) sl += q8[j] * (((val >> 2*j) & 3) - 1);
sumi1 += extra & (1 << k) ? -sl : sl;
q8 += 8;
}
q8 = y[2*i+1].qs;
int16_t sumi2 = 0;
for (int k = 4; k < 8; ++k) {
uint16_t idx = ql[k] | ((qh[k/2] << (8 - 4*(k%2))) & 0x0f00);
uint16_t val = iq1bn_grid_u16[idx];
int16_t sl = 0;
for (int j = 0; j < 8; ++j) sl += q8[j] * (((val >> 2*j) & 3) - 1);
sumi2 += extra & (1 << k) ? -sl : sl;
q8 += 8;
}
sumf += scale.f * (GGML_FP16_TO_FP32(y[2*i+0].d) * sumi1 + GGML_FP16_TO_FP32(y[2*i+1].d) * sumi2);
}
#endif
*s = sumf;
}
void ggml_vec_dot_iq1_bn_q8_K64(int n, float * s, size_t bs, const void * vx, size_t bx, const void * vy, size_t by, int nrc) {
GGML_UNUSED(bs);
GGML_UNUSED(bx);
GGML_UNUSED(by);
GGML_UNUSED(nrc);
static_assert(QK_IQ1BN == 64, "This dot product implementation for iq1_bn requires a block size of 64");
const block_iq1_bn * x = (const block_iq1_bn *)vx;
const block_q8_K64 * y = (const block_q8_K64 *)vy;
int nblock = n / QK_IQ1BN;
float sumf = 0;
IQ1BNQuantizer::scale_t scale;
#if defined __AVX2__
const auto m1_8 = _mm256_set1_epi8(1);
const auto shuff1 = _mm256_set_epi64x(0x0808080808080808, 0x0000000000000000, 0x0808080808080808, 0x0000000000000000);
const auto shuff2 = _mm256_add_epi8(shuff1, m1_8);
const auto shuff3 = _mm256_set_epi64x(0x0303030303030303, 0x0202020202020202, 0x0101010101010101, 0x0000000000000000);
const auto shuff4 = _mm256_set_epi64x(0x0707070707070707, 0x0606060606060606, 0x0505050505050505, 0x0404040404040404);
const auto mask1 = _mm256_set1_epi64x(0x8040201008040201);
#if !(defined __AVX512VNNI__ && defined __AVX512VL__)
const auto m1_16 = _mm256_set1_epi16(1);
#endif
__m256 acc = _mm256_setzero_ps();
// All scales are the same in BitNet!
uint16_t u = x[0].extra & 0xff;
scale.i = ((((u >> 4) | 0xf0) - 132) << 23) | ((u & 0x0f) << 19);
for (int i = 0; i < nblock; ++i) {
// We would uncomment this if we wanted to use this implementation for a model that has per block scales
//uint16_t u = x[i].extra & 0xff;
//scale.i = ((((u >> 4) | 0xf0) - 132) << 23) | ((u & 0x0f) << 19);
auto signs = _mm256_set1_epi8(x[i].extra >> 8);
// signs for groups of 8 ordered as 0, 1, 2, 3, 4, 5, 6, 7, 0, 1, 2, 3, 4, 5, 6, 7, ...
// To use these to sign the q8 values we need
// 0 0 0 0 0 0 0 0 1 1 1 1 1 1 1 1 2 2 2 2 2 2 2 2 3 3 3 3 3 3 3 3 amd the same for 4...7
signs = _mm256_or_si256(_mm256_cmpeq_epi8(_mm256_and_si256(signs, mask1), mask1), m1_8);
auto q8_1 = _mm256_sign_epi8(_mm256_loadu_si256((const __m256i *)y[i].qs+0), _mm256_shuffle_epi8(signs, shuff3));
auto q8_2 = _mm256_sign_epi8(_mm256_loadu_si256((const __m256i *)y[i].qs+1), _mm256_shuffle_epi8(signs, shuff4));
auto ql = x[i].ql;
auto qh = x[i].qh;
auto aux1 = _mm256_set_epi64x(iq1bn_grid_xxx[ql[3] | ((qh[1] << 4) & 0x0f00)], iq1bn_grid_xxx[ql[2] | ((qh[1] << 8) & 0x0f00)],
iq1bn_grid_xxx[ql[1] | ((qh[0] << 4) & 0x0f00)], iq1bn_grid_xxx[ql[0] | ((qh[0] << 8) & 0x0f00)]);
auto aux2 = _mm256_set_epi64x(iq1bn_grid_xxx[ql[7] | ((qh[3] << 4) & 0x0f00)], iq1bn_grid_xxx[ql[6] | ((qh[3] << 8) & 0x0f00)],
iq1bn_grid_xxx[ql[5] | ((qh[2] << 4) & 0x0f00)], iq1bn_grid_xxx[ql[4] | ((qh[2] << 8) & 0x0f00)]);
auto v1_p = _mm256_cmpeq_epi8(_mm256_and_si256(_mm256_shuffle_epi8(aux1, shuff1), mask1), mask1);
auto v1_m = _mm256_cmpeq_epi8(_mm256_and_si256(_mm256_shuffle_epi8(aux1, shuff2), mask1), mask1);
auto v2_p = _mm256_cmpeq_epi8(_mm256_and_si256(_mm256_shuffle_epi8(aux2, shuff1), mask1), mask1);
auto v2_m = _mm256_cmpeq_epi8(_mm256_and_si256(_mm256_shuffle_epi8(aux2, shuff2), mask1), mask1);
auto dot1 = _mm256_sub_epi8(_mm256_sign_epi8(q8_1, v1_m), _mm256_sign_epi8(q8_1, v1_p));
auto dot2 = _mm256_sub_epi8(_mm256_sign_epi8(q8_2, v2_m), _mm256_sign_epi8(q8_2, v2_p));
#if defined __AVX512VNNI__ && defined __AVX512VL__
dot1 = _mm256_dpbusd_epi32(_mm256_setzero_si256(), m1_8, dot1);
dot2 = _mm256_dpbusd_epi32(_mm256_setzero_si256(), m1_8, dot2);
#else
dot1 = _mm256_madd_epi16(m1_16, _mm256_maddubs_epi16(m1_8, dot1));
dot2 = _mm256_madd_epi16(m1_16, _mm256_maddubs_epi16(m1_8, dot2));
#endif
// We would uncomment this if we wanted to use this implementation for a model that has per block scales
//acc1 = _mm256_fmadd_ps(_mm256_set1_ps(scale.f*GGML_FP16_TO_FP32(y[2*i+0].d)), _mm256_cvtepi32_ps(dot1), acc1);
//acc2 = _mm256_fmadd_ps(_mm256_set1_ps(scale.f*GGML_FP16_TO_FP32(y[2*i+1].d)), _mm256_cvtepi32_ps(dot2), acc2);
// All scales are the same for BitNet!
// This is slower
//uint32_t aux32 = y[2*i+0].d | (y[2*i+1].d << 16);
//auto d8 = _mm256_cvtph_ps(_mm_set1_epi32(aux32));
//acc1 = _mm256_fmadd_ps(_mm256_permute_ps(d8, 0x00), _mm256_cvtepi32_ps(dot1), acc1);
//acc2 = _mm256_fmadd_ps(_mm256_permute_ps(d8, 0x55), _mm256_cvtepi32_ps(dot2), acc2);
acc = _mm256_fmadd_ps(_mm256_set1_ps(y[i].d), _mm256_cvtepi32_ps(_mm256_add_epi32(dot1, dot2)), acc);
}
sumf = scale.f * hsum_float_8(acc);
#else
uint16_t u = x[0].extra & 0xff;
scale.i = ((((u >> 4) | 0xf0) - 132) << 23) | ((u & 0x0f) << 19);
for (int i = 0; i < nblock; ++i) {
uint8_t extra = x[i].extra >> 8;
auto qh = x[i].qh;
auto ql = x[i].ql;
auto q8 = y[i].qs;
int sumi = 0;
for (int k = 0; k < 4; ++k) {
uint16_t idx = ql[k] | ((qh[k/2] << (8 - 4*(k%2))) & 0x0f00);
uint16_t val = iq1bn_grid_u16[idx];
int16_t sl = 0;
for (int j = 0; j < 8; ++j) sl += q8[j] * (((val >> 2*j) & 3) - 1);
sumi += extra & (1 << k) ? -sl : sl;
q8 += 8;
}
for (int k = 4; k < 8; ++k) {
uint16_t idx = ql[k] | ((qh[k/2] << (8 - 4*(k%2))) & 0x0f00);
uint16_t val = iq1bn_grid_u16[idx];
int16_t sl = 0;
for (int j = 0; j < 8; ++j) sl += q8[j] * (((val >> 2*j) & 3) - 1);
sumi += extra & (1 << k) ? -sl : sl;
q8 += 8;
}
sumf += scale.f * (y[i].d) * sumi;
}
#endif
*s = sumf;
}
Reference in New Issue View Git Blame Copy Permalink