Home
last modified time | relevance | path

Searched refs:_mm_cvtps_ph (Results 1 – 25 of 35) sorted by relevance

12

/external/XNNPACK/src/f32-f16-vcvt/gen/
Dvcvt-f16c-x8.c47 _mm_storel_epi64((__m128i*) o, _mm_cvtps_ph(vf_lo, _MM_FROUND_NO_EXC)); in xnn_f32_f16_vcvt_ukernel__f16c_x8()
51 __m128i vh = _mm_cvtps_ph(vf_lo, _MM_FROUND_NO_EXC); in xnn_f32_f16_vcvt_ukernel__f16c_x8()
Dvcvt-f16c-x16.c56 _mm_storel_epi64((__m128i*) o, _mm_cvtps_ph(vf_lo, _MM_FROUND_NO_EXC)); in xnn_f32_f16_vcvt_ukernel__f16c_x16()
60 __m128i vh = _mm_cvtps_ph(vf_lo, _MM_FROUND_NO_EXC); in xnn_f32_f16_vcvt_ukernel__f16c_x16()
/external/XNNPACK/src/f32-f16-vcvt/
Df16c.c.in58 _mm_storel_epi64((__m128i*) o, _mm_cvtps_ph(vf_lo, _MM_FROUND_NO_EXC));
62 __m128i vh = _mm_cvtps_ph(vf_lo, _MM_FROUND_NO_EXC);
/external/clang/test/CodeGen/
Df16c-builtins.c49 return _mm_cvtps_ph(a, 0); in test_mm_cvtps_ph()
/external/clang/lib/Headers/
Df16cintrin.h102 #define _mm_cvtps_ph(a, imm) \ macro
/external/XNNPACK/src/f16-rmax/
Df16c.c65 *((uint16_t*) output) = (uint16_t) _mm_extract_epi16(_mm_cvtps_ph(vmax_lo, _MM_FROUND_NO_EXC), 0); in xnn_f16_rmax_ukernel__f16c()
/external/rust/crates/half/src/binary16/
Dconvert.rs634 __m128, __m128i, __m256, _mm256_cvtph_ps, _mm256_cvtps_ph, _mm_cvtph_ps, _mm_cvtps_ph,
654 let retval = _mm_cvtps_ph(vec.assume_init(), _MM_FROUND_TO_NEAREST_INT); in f32_to_f16_x86_f16c()
672 let retval = _mm_cvtps_ph(vec.assume_init(), _MM_FROUND_TO_NEAREST_INT);
/external/FP16/bench/
Dieee-element.cc185 _mm_cvtsi128_si32(_mm_cvtps_ph(_mm_set_ss(fp32), _MM_FROUND_CUR_DIRECTION))); in fp16_ieee_from_fp32_hardware()
Dto-ieee-array.cc73 _mm_cvtps_ph(_mm_loadu_ps(&input[i]), _MM_FROUND_CUR_DIRECTION)); in hardware_mm_cvtps_ph()
/external/XNNPACK/src/f16-raddstoreexpminusmax/gen/
Davx2-rr1-p2-x32-acc2.c175 *((uint16_t*) sum) = (uint16_t) _mm_extract_epi16(_mm_cvtps_ph(vacc_lo, _MM_FROUND_NO_EXC), 0); in xnn_f16_raddstoreexpminusmax_ukernel__avx2_rr1_p2_x32_acc2()
Davx2-rr1-p2-x32-acc4.c179 *((uint16_t*) sum) = (uint16_t) _mm_extract_epi16(_mm_cvtps_ph(vacc_lo, _MM_FROUND_NO_EXC), 0); in xnn_f16_raddstoreexpminusmax_ukernel__avx2_rr1_p2_x32_acc4()
Davx2-rr1-p2-x32.c173 *((uint16_t*) sum) = (uint16_t) _mm_extract_epi16(_mm_cvtps_ph(vacc_lo, _MM_FROUND_NO_EXC), 0); in xnn_f16_raddstoreexpminusmax_ukernel__avx2_rr1_p2_x32()
Davx2-rr1-p2-x40.c185 *((uint16_t*) sum) = (uint16_t) _mm_extract_epi16(_mm_cvtps_ph(vacc_lo, _MM_FROUND_NO_EXC), 0); in xnn_f16_raddstoreexpminusmax_ukernel__avx2_rr1_p2_x40()
Davx2-rr1-p2-x40-acc2.c187 *((uint16_t*) sum) = (uint16_t) _mm_extract_epi16(_mm_cvtps_ph(vacc_lo, _MM_FROUND_NO_EXC), 0); in xnn_f16_raddstoreexpminusmax_ukernel__avx2_rr1_p2_x40_acc2()
Davx2-rr1-p2-x40-acc5.c193 *((uint16_t*) sum) = (uint16_t) _mm_extract_epi16(_mm_cvtps_ph(vacc_lo, _MM_FROUND_NO_EXC), 0); in xnn_f16_raddstoreexpminusmax_ukernel__avx2_rr1_p2_x40_acc5()
Davx2-rr1-p2-x48-acc2.c199 *((uint16_t*) sum) = (uint16_t) _mm_extract_epi16(_mm_cvtps_ph(vacc_lo, _MM_FROUND_NO_EXC), 0); in xnn_f16_raddstoreexpminusmax_ukernel__avx2_rr1_p2_x48_acc2()
Davx2-rr1-p2-x48-acc3.c201 *((uint16_t*) sum) = (uint16_t) _mm_extract_epi16(_mm_cvtps_ph(vacc_lo, _MM_FROUND_NO_EXC), 0); in xnn_f16_raddstoreexpminusmax_ukernel__avx2_rr1_p2_x48_acc3()
Davx2-rr1-p2-x48.c197 *((uint16_t*) sum) = (uint16_t) _mm_extract_epi16(_mm_cvtps_ph(vacc_lo, _MM_FROUND_NO_EXC), 0); in xnn_f16_raddstoreexpminusmax_ukernel__avx2_rr1_p2_x48()
Davx2-rr1-p2-x64-acc2.c223 *((uint16_t*) sum) = (uint16_t) _mm_extract_epi16(_mm_cvtps_ph(vacc_lo, _MM_FROUND_NO_EXC), 0); in xnn_f16_raddstoreexpminusmax_ukernel__avx2_rr1_p2_x64_acc2()
Davx2-rr1-p2-x64-acc4.c227 *((uint16_t*) sum) = (uint16_t) _mm_extract_epi16(_mm_cvtps_ph(vacc_lo, _MM_FROUND_NO_EXC), 0); in xnn_f16_raddstoreexpminusmax_ukernel__avx2_rr1_p2_x64_acc4()
Davx2-rr1-p2-x64.c221 *((uint16_t*) sum) = (uint16_t) _mm_extract_epi16(_mm_cvtps_ph(vacc_lo, _MM_FROUND_NO_EXC), 0); in xnn_f16_raddstoreexpminusmax_ukernel__avx2_rr1_p2_x64()
Davx2-rr1-p2-x72.c233 *((uint16_t*) sum) = (uint16_t) _mm_extract_epi16(_mm_cvtps_ph(vacc_lo, _MM_FROUND_NO_EXC), 0); in xnn_f16_raddstoreexpminusmax_ukernel__avx2_rr1_p2_x72()
Davx2-rr1-p2-x72-acc3.c237 *((uint16_t*) sum) = (uint16_t) _mm_extract_epi16(_mm_cvtps_ph(vacc_lo, _MM_FROUND_NO_EXC), 0); in xnn_f16_raddstoreexpminusmax_ukernel__avx2_rr1_p2_x72_acc3()
Davx2-rr1-p2-x80.c245 *((uint16_t*) sum) = (uint16_t) _mm_extract_epi16(_mm_cvtps_ph(vacc_lo, _MM_FROUND_NO_EXC), 0); in xnn_f16_raddstoreexpminusmax_ukernel__avx2_rr1_p2_x80()
/external/XNNPACK/src/f16-raddstoreexpminusmax/
Davx2-rr1-p2.c.in159 *((uint16_t*) sum) = (uint16_t) _mm_extract_epi16(_mm_cvtps_ph(vacc_lo, _MM_FROUND_NO_EXC), 0);

12