public_git
/
whisper.cpp
mirror of https://github.com/ggml-org/whisper.cpp.git
1
0
Fork 0
Code Issues Packages Projects Releases Wiki Activity

ggml-cpu: fix _pdep_u64 usage on Linux x86 32-bit 

_pdep_u64 is a BMI2 intrinsic only available in 64-bit (x86_64) mode.
On 32-bit i386 with BMI2, only _pdep_u32 exists. The previous guard
'#ifdef __BMI2__' was insufficient and produced wrong results on Linux 32-bit.

Fix both occurrences in ggml_vec_dot_iq1_s_q8_K and
ggml_vec_dot_iq1_m_q8_K to use:
    #if defined(__BMI2__) && defined(__x86_64__)

This ensures 32-bit builds use the scalar fallback paths.

Fixes: https://github.com/ggml-org/whisper.cpp/issues/3758
This commit is contained in:
Kabir08 2026-04-19 11:47:17 +05:30
parent 166c20b473
commit 6e209b9bde
1 changed files with 4 additions and 2 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

6
ggml/src/ggml-cpu/arch/x86/quants.c
View File

@ -3329,7 +3329,8 @@ void ggml_vec_dot_iq1_s_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const vo
__m256i sumi = _mm256_setzero_si256();
int sumi1 = 0;
for (int ib = 0; ib < QK_K/32; ib += 2) {
#ifdef __BMI2__
// _pdep_u64 is only available on x86_64 (64-bit); use scalar fallback on 32-bit
#if defined(__BMI2__) && defined(__x86_64__)
const uint64_t packed_idx1 = _pdep_u64(*(const uint32_t *)qs, 0x00ff00ff00ff00ffULL) | _pdep_u64(qh[ib], 0x700070007000700ULL);
const uint64_t packed_idx2 = _pdep_u64(*(const uint32_t *)(qs + 4), 0x00ff00ff00ff00ffULL) | _pdep_u64(qh[ib + 1], 0x700070007000700ULL);
const uint16_t *idx1 = (const uint16_t *)(&packed_idx1);
@ -3468,7 +3469,8 @@ void ggml_vec_dot_iq1_m_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const vo
__m256i sumi1 = _mm256_setzero_si256();
__m256i sumi2 = _mm256_setzero_si256();
for (int ib = 0; ib < QK_K/32; ib += 2) {
#ifdef __BMI2__
// _pdep_u64 is only available on x86_64 (64-bit); use scalar fallback on 32-bit
#if defined(__BMI2__) && defined(__x86_64__)
const uint64_t packed_idx1 = _pdep_u64(*(const uint32_t *)qs, 0x00ff00ff00ff00ffULL)
| _pdep_u64(*(const uint16_t*)(qh) & 0x7777, 0xf000f000f000f00ULL);
const uint64_t packed_idx2 = _pdep_u64(*(const uint32_t *)(qs + 4), 0x00ff00ff00ff00ffULL)