ikawrakow/ik_llama.cpp
1
0
Fork 0
mirror of https://github.com/ikawrakow/ik_llama.cpp.git synced 2026-04-30 11:21:56 +00:00
Code Issues Packages Projects Releases Wiki Activity

Fix MSVC compilation (#448)

Browse Source 
* Fix MSVC compilation

* MSVC cannot capture constexpr in lambdas

* Arghhh

---------

Co-authored-by: Iwan Kawrakow <iwan.kawrakow@gmail.com>
This commit is contained in:
Kawrakow
2025-05-23 16:46:27 +03:00
committed by GitHub
parent 7f2edd1a85
commit 6b12c2e7e8
2 changed files with 11 additions and 33 deletions
Download Patch File Download Diff File Expand all files Collapse all files

12
examples/quantize-stats/quantize-stats.cpp
View File

@@ -550,6 +550,10 @@ static void analyze_x_v2(const char * name, int nrows, int n_per_row, const floa
int counter = 0; int counter = 0;
float mse = 0, mse_q = 0; float mse = 0, mse_q = 0;
auto compute = [&mutex, &counter, &mse, &mse_q, values, nrows, n_per_row, chunk] () { auto compute = [&mutex, &counter, &mse, &mse_q, values, nrows, n_per_row, chunk] () {
constexpr int kNumVal = 1 << 15;
constexpr int kBlockSize = 32;
constexpr int kGroupSize = 8;
constexpr int kNg = kBlockSize/kGroupSize;
double lmse = 0, lmse_q = 0; double lmse = 0, lmse_q = 0;
std::vector<float> scales(n_per_row/kBlockSize); std::vector<float> scales(n_per_row/kBlockSize);
std::vector<int> best_idx(n_per_row/kGroupSize); std::vector<int> best_idx(n_per_row/kGroupSize);
@@ -689,9 +693,8 @@ static void analyze_x_v2(const char * name, int nrows, int n_per_row, const floa
} }
} }
}; };
std::vector<std::thread> workers(nthread-1); std::vector<std::thread> workers(nthread);
for (auto& w : workers) w = std::thread(compute); for (auto& w : workers) w = std::thread(compute);
compute();
for (auto& w : workers) w.join(); for (auto& w : workers) w.join();
tot_mse += mse; tot_mse += mse;
tot_mse_q += mse_q; tot_mse_q += mse_q;
@@ -718,6 +721,8 @@ static void analyze_x(const char * name, int nrows, int n_per_row, const float *
int counter = 0; int counter = 0;
float mse = 0, mse_q = 0; float mse = 0, mse_q = 0;
auto compute = [&mutex, &counter, &mse, &mse_q, &codes, &sumq2i, values, nrows, n_per_row, chunk] () { auto compute = [&mutex, &counter, &mse, &mse_q, &codes, &sumq2i, values, nrows, n_per_row, chunk] () {
constexpr int kBlockSize = 8;
constexpr int kNumVal = 1 << 12;
float lmse = 0, lmse_q = 0; float lmse = 0, lmse_q = 0;
std::vector<float> scales(n_per_row/kBlockSize); std::vector<float> scales(n_per_row/kBlockSize);
std::vector<int> best_idx(n_per_row/kBlockSize); std::vector<int> best_idx(n_per_row/kBlockSize);
@@ -816,9 +821,8 @@ static void analyze_x(const char * name, int nrows, int n_per_row, const float *
} }
} }
}; };
std::vector<std::thread> workers(nthread-1); std::vector<std::thread> workers(nthread);
for (auto& w : workers) w = std::thread(compute); for (auto& w : workers) w = std::thread(compute);
compute();
for (auto& w : workers) w.join(); for (auto& w : workers) w.join();
tot_mse += mse; tot_mse += mse;
tot_mse_q += mse_q; tot_mse_q += mse_q;

32
ggml/src/iqk/iqk_gemm_ktquants.cpp
View File

@@ -21,32 +21,6 @@ static inline uint32_t trellis_next(uint32_t& val) {
return (val & kmask) ^ km32; return (val & kmask) ^ km32;
} }
static inline __m256i trellis_next8(uint32_t val) {
constexpr uint32_t kmask = 0x8fff8fff;
constexpr uint32_t km32 = 0x3b603b60;
constexpr uint32_t ka = 89226354;
constexpr uint32_t kb = 64248484;
constexpr uint32_t ka1 = ka*ka;
constexpr uint32_t kb1 = kb*ka+kb;
constexpr uint32_t ka2 = ka1*ka;
constexpr uint32_t kb2 = kb1*ka+kb;
constexpr uint32_t ka3 = ka2*ka;
constexpr uint32_t kb3 = kb2*ka+kb;
constexpr uint32_t ka4 = ka3*ka;
constexpr uint32_t kb4 = kb3*ka+kb;
constexpr uint32_t ka5 = ka4*ka;
constexpr uint32_t kb5 = kb4*ka+kb;
constexpr uint32_t ka6 = ka5*ka;
constexpr uint32_t kb6 = kb5*ka+kb;
constexpr uint32_t ka7 = ka6*ka;
constexpr uint32_t kb7 = kb6*ka+kb;
__m256i mka = _mm256_setr_epi32(ka, ka1, ka2, ka3, ka4, ka5, ka6, ka7);
__m256i mkb = _mm256_setr_epi32(kb, kb1, kb2, kb3, kb4, kb5, kb6, kb7);
__m256i mval = _mm256_set1_epi32(val);
__m256i mres = _mm256_add_epi32(_mm256_mullo_epi32(mval, mka), mkb);
return _mm256_and_si256(mres, _mm256_set1_epi32(kmask)) ^ _mm256_set1_epi32(km32);
}
static inline float trellis_gen(uint32_t& val, uint32_t* s) { static inline float trellis_gen(uint32_t& val, uint32_t* s) {
const ggml_fp16_t * h = (const ggml_fp16_t *)s; const ggml_fp16_t * h = (const ggml_fp16_t *)s;
s[0] = trellis_next(val); s[0] = trellis_next(val);
@@ -80,7 +54,7 @@ struct Trellis1 {
inline __m256i next8(uint32_t val) const { inline __m256i next8(uint32_t val) const {
auto mval = _mm256_set1_epi32(val); auto mval = _mm256_set1_epi32(val);
auto mres = _mm256_add_epi32(_mm256_mullo_epi32(mval, mka), mkb); auto mres = _mm256_add_epi32(_mm256_mullo_epi32(mval, mka), mkb);
return _mm256_and_si256(mres, mask1) ^ mask2; return _mm256_xor_si256(_mm256_and_si256(mres, mask1), mask2);
} }
}; };
@@ -117,7 +91,7 @@ struct Trellis2 {
inline __m256i next8(uint32_t val1, uint32_t val2) { inline __m256i next8(uint32_t val1, uint32_t val2) {
__m256i mval = _mm256_setr_epi32(val1, val1, val1, val1, val2, val2, val2, val2); __m256i mval = _mm256_setr_epi32(val1, val1, val1, val1, val2, val2, val2, val2);
__m256i mres = _mm256_add_epi32(_mm256_mullo_epi32(mval, mka), mkb); __m256i mres = _mm256_add_epi32(_mm256_mullo_epi32(mval, mka), mkb);
return _mm256_and_si256(mres, _mm256_set1_epi32(kmask)) ^ _mm256_set1_epi32(km32); return _mm256_xor_si256(_mm256_and_si256(mres, _mm256_set1_epi32(kmask)), _mm256_set1_epi32(km32));
} }
}; };
@@ -400,4 +374,4 @@ bool iqk_set_kernels_ktquants(int ne00, int typeA, int typeB, std::array<mul_mat
#endif #endif
#endif #endif