/external/XNNPACK/src/f32-velu/gen/ |
D | velu-avx-rr2-lut4-p4-perm-x8.c | 57 …const __m128 ven_hi = _mm_castsi128_ps(_mm_slli_epi32(_mm_castps_si128(_mm256_extractf128_ps(ven, … in xnn_f32_velu_ukernel__avx_rr2_lut4_p4_perm_x8() local 93 …const __m128 ven_hi = _mm_castsi128_ps(_mm_slli_epi32(_mm_castps_si128(_mm256_extractf128_ps(ven, … in xnn_f32_velu_ukernel__avx_rr2_lut4_p4_perm_x8() local
|
D | velu-avx-rr2-lut4-p4-perm-x16.c | 119 …const __m128 ven_hi = _mm_castsi128_ps(_mm_slli_epi32(_mm_castps_si128(_mm256_extractf128_ps(ven, … in xnn_f32_velu_ukernel__avx_rr2_lut4_p4_perm_x16() local 155 …const __m128 ven_hi = _mm_castsi128_ps(_mm_slli_epi32(_mm_castps_si128(_mm256_extractf128_ps(ven, … in xnn_f32_velu_ukernel__avx_rr2_lut4_p4_perm_x16() local
|
D | velu-avx-rr2-lut16-p3-x8.c | 81 const __m128i ven_hi = _mm_slli_epi32(_mm_castps_si128(_mm256_extractf128_ps(vn, 1)), 19); in xnn_f32_velu_ukernel__avx_rr2_lut16_p3_x8() local 147 const __m128i ven_hi = _mm_slli_epi32(_mm_castps_si128(_mm256_extractf128_ps(vn, 1)), 19); in xnn_f32_velu_ukernel__avx_rr2_lut16_p3_x8() local
|
D | velu-avx-rr2-lut4-p4-perm-x24.c | 141 …const __m128 ven_hi = _mm_castsi128_ps(_mm_slli_epi32(_mm_castps_si128(_mm256_extractf128_ps(ven, … in xnn_f32_velu_ukernel__avx_rr2_lut4_p4_perm_x24() local 177 …const __m128 ven_hi = _mm_castsi128_ps(_mm_slli_epi32(_mm_castps_si128(_mm256_extractf128_ps(ven, … in xnn_f32_velu_ukernel__avx_rr2_lut4_p4_perm_x24() local
|
D | velu-avx-rr2-lut4-p4-perm-x32.c | 163 …const __m128 ven_hi = _mm_castsi128_ps(_mm_slli_epi32(_mm_castps_si128(_mm256_extractf128_ps(ven, … in xnn_f32_velu_ukernel__avx_rr2_lut4_p4_perm_x32() local 199 …const __m128 ven_hi = _mm_castsi128_ps(_mm_slli_epi32(_mm_castps_si128(_mm256_extractf128_ps(ven, … in xnn_f32_velu_ukernel__avx_rr2_lut4_p4_perm_x32() local
|
D | velu-avx-rr2-lut4-p4-perm-x40.c | 185 …const __m128 ven_hi = _mm_castsi128_ps(_mm_slli_epi32(_mm_castps_si128(_mm256_extractf128_ps(ven, … in xnn_f32_velu_ukernel__avx_rr2_lut4_p4_perm_x40() local 221 …const __m128 ven_hi = _mm_castsi128_ps(_mm_slli_epi32(_mm_castps_si128(_mm256_extractf128_ps(ven, … in xnn_f32_velu_ukernel__avx_rr2_lut4_p4_perm_x40() local
|
D | velu-avx-rr2-lut4-p4-perm-x48.c | 207 …const __m128 ven_hi = _mm_castsi128_ps(_mm_slli_epi32(_mm_castps_si128(_mm256_extractf128_ps(ven, … in xnn_f32_velu_ukernel__avx_rr2_lut4_p4_perm_x48() local 243 …const __m128 ven_hi = _mm_castsi128_ps(_mm_slli_epi32(_mm_castps_si128(_mm256_extractf128_ps(ven, … in xnn_f32_velu_ukernel__avx_rr2_lut4_p4_perm_x48() local
|
D | velu-avx-rr2-lut16-p3-x16.c | 196 const __m128i ven_hi = _mm_slli_epi32(_mm_castps_si128(_mm256_extractf128_ps(vn, 1)), 19); in xnn_f32_velu_ukernel__avx_rr2_lut16_p3_x16() local 262 const __m128i ven_hi = _mm_slli_epi32(_mm_castps_si128(_mm256_extractf128_ps(vn, 1)), 19); in xnn_f32_velu_ukernel__avx_rr2_lut16_p3_x16() local
|
D | velu-avx-rr2-lut16-p3-x24.c | 245 const __m128i ven_hi = _mm_slli_epi32(_mm_castps_si128(_mm256_extractf128_ps(vn, 1)), 19); in xnn_f32_velu_ukernel__avx_rr2_lut16_p3_x24() local 311 const __m128i ven_hi = _mm_slli_epi32(_mm_castps_si128(_mm256_extractf128_ps(vn, 1)), 19); in xnn_f32_velu_ukernel__avx_rr2_lut16_p3_x24() local
|
D | velu-avx-rr2-lut16-p3-x32.c | 294 const __m128i ven_hi = _mm_slli_epi32(_mm_castps_si128(_mm256_extractf128_ps(vn, 1)), 19); in xnn_f32_velu_ukernel__avx_rr2_lut16_p3_x32() local 360 const __m128i ven_hi = _mm_slli_epi32(_mm_castps_si128(_mm256_extractf128_ps(vn, 1)), 19); in xnn_f32_velu_ukernel__avx_rr2_lut16_p3_x32() local
|
D | velu-avx-rr2-lut16-p3-x40.c | 343 const __m128i ven_hi = _mm_slli_epi32(_mm_castps_si128(_mm256_extractf128_ps(vn, 1)), 19); in xnn_f32_velu_ukernel__avx_rr2_lut16_p3_x40() local 409 const __m128i ven_hi = _mm_slli_epi32(_mm_castps_si128(_mm256_extractf128_ps(vn, 1)), 19); in xnn_f32_velu_ukernel__avx_rr2_lut16_p3_x40() local
|
D | velu-avx-rr2-lut16-p3-x48.c | 392 const __m128i ven_hi = _mm_slli_epi32(_mm_castps_si128(_mm256_extractf128_ps(vn, 1)), 19); in xnn_f32_velu_ukernel__avx_rr2_lut16_p3_x48() local 458 const __m128i ven_hi = _mm_slli_epi32(_mm_castps_si128(_mm256_extractf128_ps(vn, 1)), 19); in xnn_f32_velu_ukernel__avx_rr2_lut16_p3_x48() local
|
/external/XNNPACK/src/f32-velu/ |
D | avx-rr2-lut4-p4-perm.c.in | 114 …const __m128 ven_hi = _mm_castsi128_ps(_mm_slli_epi32(_mm_castps_si128(_mm256_extractf128_ps(ven, … variable 150 …const __m128 ven_hi = _mm_castsi128_ps(_mm_slli_epi32(_mm_castps_si128(_mm256_extractf128_ps(ven, … variable
|
D | avx-rr2-lut16-p3.c.in | 163 const __m128i ven_hi = _mm_slli_epi32(_mm_castps_si128(_mm256_extractf128_ps(vn, 1)), 19); variable 229 const __m128i ven_hi = _mm_slli_epi32(_mm_castps_si128(_mm256_extractf128_ps(vn, 1)), 19); variable
|
/external/XNNPACK/src/math/ |
D | expm1minus-avx-rr2-lut4-p4-perm.c | 72 …const __m128 ven_hi = _mm_castsi128_ps(_mm_slli_epi32(_mm_castps_si128(_mm256_extractf128_ps(ven, … in xnn_math_f32_expm1minus__avx_rr2_lut4_p4_perm() local
|
D | exp-avx-rr2-p5.c | 63 __m128i ven_hi = _mm_max_epi16(veo_hi, vmin_exponent); in xnn_math_f32_exp__avx_rr2_p5() local
|
D | expm1minus-avx-rr2-lut16-p3.c | 69 const __m128i ven_hi = _mm_slli_epi32(_mm_castps_si128(_mm256_extractf128_ps(vn, 1)), 19); in xnn_math_f32_expm1minus__avx_rr2_lut16_p3() local
|