| /external/XNNPACK/src/x8-lut/gen/ |
| D | lut-scalar-x8.c | 41 const uint32_t vt4 = (uint32_t) t[vx4]; in xnn_x8_lut_ukernel__scalar_x8() local
|
| D | lut-scalar-x16.c | 49 const uint32_t vt4 = (uint32_t) t[vx4]; in xnn_x8_lut_ukernel__scalar_x16() local
|
| D | lut-ssse3-x16.c | 33 const __m128i vt4 = _mm_load_si128((const __m128i*) (t + 64)); in xnn_x8_lut_ukernel__ssse3_x16() local
|
| D | lut-avx512skx-vpshufb-x64.c | 33 const __m512i vt4 = _mm512_broadcast_i32x4(_mm_load_si128((const __m128i*) (t + 64))); in xnn_x8_lut_ukernel__avx512skx_vpshufb_x64() local
|
| D | lut-avx-x16.c | 33 const __m128i vt4 = _mm_load_si128((const __m128i*) (t + 64)); in xnn_x8_lut_ukernel__avx_x16() local
|
| D | lut-avx2-x32.c | 33 const __m256i vt4 = _mm256_broadcastsi128_si256(_mm_load_si128((const __m128i*) (t + 64))); in xnn_x8_lut_ukernel__avx2_x32() local
|
| D | lut-avx2-x64.c | 33 const __m256i vt4 = _mm256_broadcastsi128_si256(_mm_load_si128((const __m128i*) (t + 64))); in xnn_x8_lut_ukernel__avx2_x64() local
|
| D | lut-avx-x32.c | 33 const __m128i vt4 = _mm_load_si128((const __m128i*) (t + 64)); in xnn_x8_lut_ukernel__avx_x32() local
|
| D | lut-ssse3-x32.c | 33 const __m128i vt4 = _mm_load_si128((const __m128i*) (t + 64)); in xnn_x8_lut_ukernel__ssse3_x32() local
|
| D | lut-avx512skx-vpshufb-x128.c | 33 const __m512i vt4 = _mm512_broadcast_i32x4(_mm_load_si128((const __m128i*) (t + 64))); in xnn_x8_lut_ukernel__avx512skx_vpshufb_x128() local
|
| D | lut-avx-x48.c | 33 const __m128i vt4 = _mm_load_si128((const __m128i*) (t + 64)); in xnn_x8_lut_ukernel__avx_x48() local
|
| D | lut-avx2-x96.c | 33 const __m256i vt4 = _mm256_broadcastsi128_si256(_mm_load_si128((const __m128i*) (t + 64))); in xnn_x8_lut_ukernel__avx2_x96() local
|
| D | lut-avx512skx-vpshufb-x192.c | 33 const __m512i vt4 = _mm512_broadcast_i32x4(_mm_load_si128((const __m128i*) (t + 64))); in xnn_x8_lut_ukernel__avx512skx_vpshufb_x192() local
|
| D | lut-avx2-x128.c | 33 const __m256i vt4 = _mm256_broadcastsi128_si256(_mm_load_si128((const __m128i*) (t + 64))); in xnn_x8_lut_ukernel__avx2_x128() local
|
| D | lut-avx-x64.c | 33 const __m128i vt4 = _mm_load_si128((const __m128i*) (t + 64)); in xnn_x8_lut_ukernel__avx_x64() local
|
| D | lut-avx512skx-vpshufb-x256.c | 33 const __m512i vt4 = _mm512_broadcast_i32x4(_mm_load_si128((const __m128i*) (t + 64))); in xnn_x8_lut_ukernel__avx512skx_vpshufb_x256() local
|
| /external/XNNPACK/src/f32-vscaleexpminusmax/gen/ |
| D | avx512f-p5-scalef-x80.c | 70 __m512 vt4 = _mm512_fmadd_ps(vn4, vminus_ln2_hi, vx4); in xnn_f32_vscaleexpminusmax_ukernel__avx512f_p5_scalef_x80() local
|
| D | avx512f-p5-scalef-x96.c | 73 __m512 vt4 = _mm512_fmadd_ps(vn4, vminus_ln2_hi, vx4); in xnn_f32_vscaleexpminusmax_ukernel__avx512f_p5_scalef_x96() local
|
| D | avx2-p5-x40.c | 89 __m256 vt4 = _mm256_fmadd_ps(vn4, vminus_ln2_hi, vx4); in xnn_f32_vscaleexpminusmax_ukernel__avx2_p5_x40() local
|
| /external/XNNPACK/src/f32-vscaleextexp/gen/ |
| D | avx512f-p5-scalef-x80.c | 64 __m512 vt4 = _mm512_fmadd_ps(vn4, vminus_ln2_hi, vx4); in xnn_f32_vscaleextexp_ukernel__avx512f_p5_scalef_x80() local
|
| D | avx512f-p5-scalef-x96.c | 66 __m512 vt4 = _mm512_fmadd_ps(vn4, vminus_ln2_hi, vx4); in xnn_f32_vscaleextexp_ukernel__avx512f_p5_scalef_x96() local
|
| /external/XNNPACK/src/f32-velu/gen/ |
| D | velu-wasm-rr2-p6-x5.c | 76 float vt4 = vn4 * vminus_ln2_hi + vz4; in xnn_f32_velu_ukernel__wasm_rr2_p6_x5() local
|
| D | velu-scalar-rr2-p6-x5.c | 76 float vt4 = vn4 * vminus_ln2_hi + vz4; in xnn_f32_velu_ukernel__scalar_rr2_p6_x5() local
|
| /external/XNNPACK/src/f16-raddstoreexpminusmax/gen/ |
| D | avx2-rr1-p2-x40.c | 75 __m256 vt4 = _mm256_fmadd_ps(vn4, vminus_ln2, vx4); in xnn_f16_raddstoreexpminusmax_ukernel__avx2_rr1_p2_x40() local
|
| /external/XNNPACK/src/f32-vsigmoid/gen/ |
| D | vsigmoid-avx512f-rr1-p5-scalef-div-x80.c | 67 __m512 vt4 = _mm512_fmadd_ps(vn4, vminus_ln2, vz4); in xnn_f32_vsigmoid_ukernel__avx512f_rr1_p5_scalef_div_x80() local
|