ikawrakow/ik_llama.cpp
1
0
Fork 0
mirror of https://github.com/ikawrakow/ik_llama.cpp.git synced 2026-03-08 04:50:13 +00:00
Code Issues Packages Projects Releases Wiki Activity

Fix Compile error (C2668) (#508)

Browse Source 
* cmake: force MSVC compiler charset to utf-8

* build: apply MSVC /bigobj option to c/cpp files only

* Update CMakeLists.txt

* Fix Compile error (C2668)

* revert hsum_float_8x8
This commit is contained in:
Gaolingx
2025-06-10 13:30:17 +08:00
committed by GitHub
parent fdc60d0aae
commit b2a4df3590
2 changed files with 18 additions and 48 deletions
Download Patch File Download Diff File Expand all files Collapse all files

21
ggml/src/iqk/iqk_common.h
View File

@@ -225,9 +225,15 @@ static inline int hsum_i32_8(const __m256i a) {
const __m128i sum128 = _mm_add_epi32(_mm256_castsi256_si128(a), _mm256_extractf128_si256(a, 1));
const __m128i hi64 = _mm_unpackhi_epi64(sum128, sum128);
const __m128i sum64 = _mm_add_epi32(hi64, sum128);
const __m128i hi32 = _mm_shuffle_epi32(sum64, _MM_SHUFFLE(2, 3, 0, 1));
const __m128i hi32 = _mm_shuffle_epi32(sum64, _MM_SHUFFLE(2, 3, 0, 1));
return _mm_cvtsi128_si32(_mm_add_epi32(sum64, hi32));
}
static inline float hmax_f32_8(__m256 x) {
__m128 max4 = _mm_max_ps(_mm256_extractf128_ps(x, 1), _mm256_castps256_ps128(x));
max4 = _mm_max_ps(max4, _mm_movehl_ps(max4, max4));
max4 = _mm_max_ss(max4, _mm_movehdup_ps(max4));
return _mm_cvtss_f32(max4);
}
static inline float hmax_float_8(__m256 x) {
__m128 max4 = _mm_max_ps(_mm256_extractf128_ps(x, 1), _mm256_castps256_ps128(x));
max4 = _mm_max_ps( max4, _mm_movehl_ps(max4, max4));
@@ -235,13 +241,22 @@ static inline float hmax_float_8(__m256 x) {
return _mm_cvtss_f32(max4);
}
static inline __m128 hsum_float_4x4(__m128 * accm) {
accm[0] = _mm_add_ps(_mm_unpacklo_ps(accm[0], accm[2]), _mm_unpackhi_ps(accm[0], accm[2]));
accm[1] = _mm_add_ps(_mm_unpacklo_ps(accm[1], accm[3]), _mm_unpackhi_ps(accm[1], accm[3]));
return _mm_add_ps(_mm_unpacklo_ps(accm[0], accm[1]), _mm_unpackhi_ps(accm[0], accm[1]));
}
static inline __m256 hsum_float_8x8(__m256 * accm) {
for (int i = 0; i < 4; ++i) {
accm[i] = _mm256_add_ps(_mm256_permute2f128_ps(accm[i], accm[i+4], 0x20), _mm256_permute2f128_ps(accm[i], accm[i+4], 0x31));
accm[i] = _mm256_add_ps(_mm256_permute2f128_ps(accm[i], accm[i + 4], 0x20), _mm256_permute2f128_ps(accm[i], accm[i + 4], 0x31));
//accm[i] = _mm256_set_m128(_mm_add_ps(_mm256_castps256_ps128(accm[i+4]), _mm256_extractf128_ps(accm[i+4], 1)),
// _mm_add_ps(_mm256_castps256_ps128(accm[i+0]), _mm256_extractf128_ps(accm[i+0], 1)));
}
for (int i = 0; i < 2; ++i) accm[i] = _mm256_add_ps(_mm256_unpacklo_ps(accm[i], accm[i+2]), _mm256_unpackhi_ps(accm[i], accm[i+2]));
for (int i = 0; i < 2; ++i) accm[i] = _mm256_add_ps(_mm256_unpacklo_ps(accm[i], accm[i + 2]), _mm256_unpackhi_ps(accm[i], accm[i + 2]));
return _mm256_add_ps(_mm256_unpacklo_ps(accm[0], accm[1]), _mm256_unpackhi_ps(accm[0], accm[1]));
}
static inline __m256 hsum_float_4x8(__m256 * accm) {
for (int i = 0; i < 2; ++i) accm[i] = _mm256_add_ps(_mm256_unpacklo_ps(accm[i], accm[i + 2]), _mm256_unpackhi_ps(accm[i], accm[i + 2]));
return _mm256_add_ps(_mm256_unpacklo_ps(accm[0], accm[1]), _mm256_unpackhi_ps(accm[0], accm[1]));
}

45
ggml/src/iqk/iqk_quantize.cpp
View File

@@ -557,32 +557,6 @@ void quantize_row_q8_K64(const float * x, void * y, int64_t k) {
quantize_row_q8_K64_ref(x, (block_q8_K64 *)y, k);
}
#ifdef __AVX2__
namespace {
inline float hsum_float_4(__m128 x) {
x = _mm_add_ps(x, _mm_movehl_ps(x, x));
x = _mm_add_ss(x, _mm_movehdup_ps(x));
return _mm_cvtss_f32(x);
}
inline float hsum_float_8(__m256 x) {
return hsum_float_4(_mm_add_ps(_mm256_castps256_ps128(x), _mm256_extractf128_ps(x, 1)));
}
inline int hsum_i32_8(const __m256i a) {
const __m128i sum128 = _mm_add_epi32(_mm256_castsi256_si128(a), _mm256_extractf128_si256(a, 1));
const __m128i hi64 = _mm_unpackhi_epi64(sum128, sum128);
const __m128i sum64 = _mm_add_epi32(hi64, sum128);
const __m128i hi32 = _mm_shuffle_epi32(sum64, _MM_SHUFFLE(2, 3, 0, 1));
return _mm_cvtsi128_si32(_mm_add_epi32(sum64, hi32));
}
inline float hmax_f32_8(__m256 x) {
__m128 max4 = _mm_max_ps(_mm256_extractf128_ps(x, 1), _mm256_castps256_ps128(x));
max4 = _mm_max_ps( max4, _mm_movehl_ps(max4, max4));
max4 = _mm_max_ss( max4, _mm_movehdup_ps( max4));
return _mm_cvtss_f32(max4);
}
}
#endif
void quantize_row_q8_K16(const float * x, void * vy, int64_t nk) {
float * dptr = (float *)vy;
int8_t * qy = (int8_t *)(dptr + 5);
@@ -7413,25 +7387,6 @@ void dequantize_row_ms_i2s(const void * vx, float * y, int64_t k) {
}
namespace {
#ifdef __AVX2__
__m128 hsum_float_4x4(__m128 * accm) {
accm[0] = _mm_add_ps(_mm_unpacklo_ps(accm[0], accm[2]), _mm_unpackhi_ps(accm[0], accm[2]));
accm[1] = _mm_add_ps(_mm_unpacklo_ps(accm[1], accm[3]), _mm_unpackhi_ps(accm[1], accm[3]));
return _mm_add_ps(_mm_unpacklo_ps(accm[0], accm[1]), _mm_unpackhi_ps(accm[0], accm[1]));
}
__m256 hsum_float_8x8(__m256 * accm) {
for (int i = 0; i < 4; ++i) {
accm[i] = _mm256_set_m128(_mm_add_ps(_mm256_castps256_ps128(accm[i+4]), _mm256_extractf128_ps(accm[i+4], 1)),
_mm_add_ps(_mm256_castps256_ps128(accm[i+0]), _mm256_extractf128_ps(accm[i+0], 1)));
}
for (int i = 0; i < 2; ++i) accm[i] = _mm256_add_ps(_mm256_unpacklo_ps(accm[i], accm[i+2]), _mm256_unpackhi_ps(accm[i], accm[i+2]));
return _mm256_add_ps(_mm256_unpacklo_ps(accm[0], accm[1]), _mm256_unpackhi_ps(accm[0], accm[1]));
}
__m256 hsum_float_4x8(__m256 * accm) {
for (int i = 0; i < 2; ++i) accm[i] = _mm256_add_ps(_mm256_unpacklo_ps(accm[i], accm[i+2]), _mm256_unpackhi_ps(accm[i], accm[i+2]));
return _mm256_add_ps(_mm256_unpacklo_ps(accm[0], accm[1]), _mm256_unpackhi_ps(accm[0], accm[1]));
}
#endif
template <int block_size, int group_size, int num_bits, bool is_abs = false>
class QuantizerIQKT {
static_assert(group_size == 8 || group_size == 4);