/external/XNNPACK/src/f32-sigmoid/gen/ |
D | avx512f-rr1-lut16-p3-perm-scalef-div-x112.c | 66 const __m512 vl0 = _mm512_permutexvar_ps(_mm512_castps_si512(vn0), vtable); in xnn_f32_sigmoid_ukernel__avx512f_rr1_lut16_p3_perm_scalef_div_x112() 67 const __m512 vl1 = _mm512_permutexvar_ps(_mm512_castps_si512(vn1), vtable); in xnn_f32_sigmoid_ukernel__avx512f_rr1_lut16_p3_perm_scalef_div_x112() 68 const __m512 vl2 = _mm512_permutexvar_ps(_mm512_castps_si512(vn2), vtable); in xnn_f32_sigmoid_ukernel__avx512f_rr1_lut16_p3_perm_scalef_div_x112() 69 const __m512 vl3 = _mm512_permutexvar_ps(_mm512_castps_si512(vn3), vtable); in xnn_f32_sigmoid_ukernel__avx512f_rr1_lut16_p3_perm_scalef_div_x112() 70 const __m512 vl4 = _mm512_permutexvar_ps(_mm512_castps_si512(vn4), vtable); in xnn_f32_sigmoid_ukernel__avx512f_rr1_lut16_p3_perm_scalef_div_x112() 71 const __m512 vl5 = _mm512_permutexvar_ps(_mm512_castps_si512(vn5), vtable); in xnn_f32_sigmoid_ukernel__avx512f_rr1_lut16_p3_perm_scalef_div_x112() 72 const __m512 vl6 = _mm512_permutexvar_ps(_mm512_castps_si512(vn6), vtable); in xnn_f32_sigmoid_ukernel__avx512f_rr1_lut16_p3_perm_scalef_div_x112() 170 const __m512 vl = _mm512_permutexvar_ps(_mm512_castps_si512(vn), vtable); in xnn_f32_sigmoid_ukernel__avx512f_rr1_lut16_p3_perm_scalef_div_x112() 202 const __m512 vl = _mm512_permutexvar_ps(_mm512_castps_si512(vn), vtable); in xnn_f32_sigmoid_ukernel__avx512f_rr1_lut16_p3_perm_scalef_div_x112()
|
D | avx512f-rr1-lut16-p3-perm-scalef-div-x128.c | 69 const __m512 vl0 = _mm512_permutexvar_ps(_mm512_castps_si512(vn0), vtable); in xnn_f32_sigmoid_ukernel__avx512f_rr1_lut16_p3_perm_scalef_div_x128() 70 const __m512 vl1 = _mm512_permutexvar_ps(_mm512_castps_si512(vn1), vtable); in xnn_f32_sigmoid_ukernel__avx512f_rr1_lut16_p3_perm_scalef_div_x128() 71 const __m512 vl2 = _mm512_permutexvar_ps(_mm512_castps_si512(vn2), vtable); in xnn_f32_sigmoid_ukernel__avx512f_rr1_lut16_p3_perm_scalef_div_x128() 72 const __m512 vl3 = _mm512_permutexvar_ps(_mm512_castps_si512(vn3), vtable); in xnn_f32_sigmoid_ukernel__avx512f_rr1_lut16_p3_perm_scalef_div_x128() 73 const __m512 vl4 = _mm512_permutexvar_ps(_mm512_castps_si512(vn4), vtable); in xnn_f32_sigmoid_ukernel__avx512f_rr1_lut16_p3_perm_scalef_div_x128() 74 const __m512 vl5 = _mm512_permutexvar_ps(_mm512_castps_si512(vn5), vtable); in xnn_f32_sigmoid_ukernel__avx512f_rr1_lut16_p3_perm_scalef_div_x128() 75 const __m512 vl6 = _mm512_permutexvar_ps(_mm512_castps_si512(vn6), vtable); in xnn_f32_sigmoid_ukernel__avx512f_rr1_lut16_p3_perm_scalef_div_x128() 76 const __m512 vl7 = _mm512_permutexvar_ps(_mm512_castps_si512(vn7), vtable); in xnn_f32_sigmoid_ukernel__avx512f_rr1_lut16_p3_perm_scalef_div_x128() 185 const __m512 vl = _mm512_permutexvar_ps(_mm512_castps_si512(vn), vtable); in xnn_f32_sigmoid_ukernel__avx512f_rr1_lut16_p3_perm_scalef_div_x128() 217 const __m512 vl = _mm512_permutexvar_ps(_mm512_castps_si512(vn), vtable); in xnn_f32_sigmoid_ukernel__avx512f_rr1_lut16_p3_perm_scalef_div_x128()
|
D | avx512f-rr1-lut16-p3-perm-scalef-div-x96.c | 63 const __m512 vl0 = _mm512_permutexvar_ps(_mm512_castps_si512(vn0), vtable); in xnn_f32_sigmoid_ukernel__avx512f_rr1_lut16_p3_perm_scalef_div_x96() 64 const __m512 vl1 = _mm512_permutexvar_ps(_mm512_castps_si512(vn1), vtable); in xnn_f32_sigmoid_ukernel__avx512f_rr1_lut16_p3_perm_scalef_div_x96() 65 const __m512 vl2 = _mm512_permutexvar_ps(_mm512_castps_si512(vn2), vtable); in xnn_f32_sigmoid_ukernel__avx512f_rr1_lut16_p3_perm_scalef_div_x96() 66 const __m512 vl3 = _mm512_permutexvar_ps(_mm512_castps_si512(vn3), vtable); in xnn_f32_sigmoid_ukernel__avx512f_rr1_lut16_p3_perm_scalef_div_x96() 67 const __m512 vl4 = _mm512_permutexvar_ps(_mm512_castps_si512(vn4), vtable); in xnn_f32_sigmoid_ukernel__avx512f_rr1_lut16_p3_perm_scalef_div_x96() 68 const __m512 vl5 = _mm512_permutexvar_ps(_mm512_castps_si512(vn5), vtable); in xnn_f32_sigmoid_ukernel__avx512f_rr1_lut16_p3_perm_scalef_div_x96() 155 const __m512 vl = _mm512_permutexvar_ps(_mm512_castps_si512(vn), vtable); in xnn_f32_sigmoid_ukernel__avx512f_rr1_lut16_p3_perm_scalef_div_x96() 187 const __m512 vl = _mm512_permutexvar_ps(_mm512_castps_si512(vn), vtable); in xnn_f32_sigmoid_ukernel__avx512f_rr1_lut16_p3_perm_scalef_div_x96()
|
D | avx512f-rr1-lut16-p3-perm-scalef-nr1fma-x128.c | 69 const __m512 vl0 = _mm512_permutexvar_ps(_mm512_castps_si512(vn0), vtable); in xnn_f32_sigmoid_ukernel__avx512f_rr1_lut16_p3_perm_scalef_nr1fma_x128() 70 const __m512 vl1 = _mm512_permutexvar_ps(_mm512_castps_si512(vn1), vtable); in xnn_f32_sigmoid_ukernel__avx512f_rr1_lut16_p3_perm_scalef_nr1fma_x128() 71 const __m512 vl2 = _mm512_permutexvar_ps(_mm512_castps_si512(vn2), vtable); in xnn_f32_sigmoid_ukernel__avx512f_rr1_lut16_p3_perm_scalef_nr1fma_x128() 72 const __m512 vl3 = _mm512_permutexvar_ps(_mm512_castps_si512(vn3), vtable); in xnn_f32_sigmoid_ukernel__avx512f_rr1_lut16_p3_perm_scalef_nr1fma_x128() 73 const __m512 vl4 = _mm512_permutexvar_ps(_mm512_castps_si512(vn4), vtable); in xnn_f32_sigmoid_ukernel__avx512f_rr1_lut16_p3_perm_scalef_nr1fma_x128() 74 const __m512 vl5 = _mm512_permutexvar_ps(_mm512_castps_si512(vn5), vtable); in xnn_f32_sigmoid_ukernel__avx512f_rr1_lut16_p3_perm_scalef_nr1fma_x128() 75 const __m512 vl6 = _mm512_permutexvar_ps(_mm512_castps_si512(vn6), vtable); in xnn_f32_sigmoid_ukernel__avx512f_rr1_lut16_p3_perm_scalef_nr1fma_x128() 76 const __m512 vl7 = _mm512_permutexvar_ps(_mm512_castps_si512(vn7), vtable); in xnn_f32_sigmoid_ukernel__avx512f_rr1_lut16_p3_perm_scalef_nr1fma_x128() 204 const __m512 vl = _mm512_permutexvar_ps(_mm512_castps_si512(vn), vtable); in xnn_f32_sigmoid_ukernel__avx512f_rr1_lut16_p3_perm_scalef_nr1fma_x128() 239 const __m512 vl = _mm512_permutexvar_ps(_mm512_castps_si512(vn), vtable); in xnn_f32_sigmoid_ukernel__avx512f_rr1_lut16_p3_perm_scalef_nr1fma_x128()
|
D | avx512f-rr1-lut16-p3-perm-scalef-div-x80.c | 60 const __m512 vl0 = _mm512_permutexvar_ps(_mm512_castps_si512(vn0), vtable); in xnn_f32_sigmoid_ukernel__avx512f_rr1_lut16_p3_perm_scalef_div_x80() 61 const __m512 vl1 = _mm512_permutexvar_ps(_mm512_castps_si512(vn1), vtable); in xnn_f32_sigmoid_ukernel__avx512f_rr1_lut16_p3_perm_scalef_div_x80() 62 const __m512 vl2 = _mm512_permutexvar_ps(_mm512_castps_si512(vn2), vtable); in xnn_f32_sigmoid_ukernel__avx512f_rr1_lut16_p3_perm_scalef_div_x80() 63 const __m512 vl3 = _mm512_permutexvar_ps(_mm512_castps_si512(vn3), vtable); in xnn_f32_sigmoid_ukernel__avx512f_rr1_lut16_p3_perm_scalef_div_x80() 64 const __m512 vl4 = _mm512_permutexvar_ps(_mm512_castps_si512(vn4), vtable); in xnn_f32_sigmoid_ukernel__avx512f_rr1_lut16_p3_perm_scalef_div_x80() 140 const __m512 vl = _mm512_permutexvar_ps(_mm512_castps_si512(vn), vtable); in xnn_f32_sigmoid_ukernel__avx512f_rr1_lut16_p3_perm_scalef_div_x80() 172 const __m512 vl = _mm512_permutexvar_ps(_mm512_castps_si512(vn), vtable); in xnn_f32_sigmoid_ukernel__avx512f_rr1_lut16_p3_perm_scalef_div_x80()
|
D | avx512f-rr1-lut16-p3-perm-scalef-nr1fma-x112.c | 66 const __m512 vl0 = _mm512_permutexvar_ps(_mm512_castps_si512(vn0), vtable); in xnn_f32_sigmoid_ukernel__avx512f_rr1_lut16_p3_perm_scalef_nr1fma_x112() 67 const __m512 vl1 = _mm512_permutexvar_ps(_mm512_castps_si512(vn1), vtable); in xnn_f32_sigmoid_ukernel__avx512f_rr1_lut16_p3_perm_scalef_nr1fma_x112() 68 const __m512 vl2 = _mm512_permutexvar_ps(_mm512_castps_si512(vn2), vtable); in xnn_f32_sigmoid_ukernel__avx512f_rr1_lut16_p3_perm_scalef_nr1fma_x112() 69 const __m512 vl3 = _mm512_permutexvar_ps(_mm512_castps_si512(vn3), vtable); in xnn_f32_sigmoid_ukernel__avx512f_rr1_lut16_p3_perm_scalef_nr1fma_x112() 70 const __m512 vl4 = _mm512_permutexvar_ps(_mm512_castps_si512(vn4), vtable); in xnn_f32_sigmoid_ukernel__avx512f_rr1_lut16_p3_perm_scalef_nr1fma_x112() 71 const __m512 vl5 = _mm512_permutexvar_ps(_mm512_castps_si512(vn5), vtable); in xnn_f32_sigmoid_ukernel__avx512f_rr1_lut16_p3_perm_scalef_nr1fma_x112() 72 const __m512 vl6 = _mm512_permutexvar_ps(_mm512_castps_si512(vn6), vtable); in xnn_f32_sigmoid_ukernel__avx512f_rr1_lut16_p3_perm_scalef_nr1fma_x112() 187 const __m512 vl = _mm512_permutexvar_ps(_mm512_castps_si512(vn), vtable); in xnn_f32_sigmoid_ukernel__avx512f_rr1_lut16_p3_perm_scalef_nr1fma_x112() 222 const __m512 vl = _mm512_permutexvar_ps(_mm512_castps_si512(vn), vtable); in xnn_f32_sigmoid_ukernel__avx512f_rr1_lut16_p3_perm_scalef_nr1fma_x112()
|
D | avx512f-rr1-lut16-p3-perm-scalef-div-x64.c | 57 const __m512 vl0 = _mm512_permutexvar_ps(_mm512_castps_si512(vn0), vtable); in xnn_f32_sigmoid_ukernel__avx512f_rr1_lut16_p3_perm_scalef_div_x64() 58 const __m512 vl1 = _mm512_permutexvar_ps(_mm512_castps_si512(vn1), vtable); in xnn_f32_sigmoid_ukernel__avx512f_rr1_lut16_p3_perm_scalef_div_x64() 59 const __m512 vl2 = _mm512_permutexvar_ps(_mm512_castps_si512(vn2), vtable); in xnn_f32_sigmoid_ukernel__avx512f_rr1_lut16_p3_perm_scalef_div_x64() 60 const __m512 vl3 = _mm512_permutexvar_ps(_mm512_castps_si512(vn3), vtable); in xnn_f32_sigmoid_ukernel__avx512f_rr1_lut16_p3_perm_scalef_div_x64() 125 const __m512 vl = _mm512_permutexvar_ps(_mm512_castps_si512(vn), vtable); in xnn_f32_sigmoid_ukernel__avx512f_rr1_lut16_p3_perm_scalef_div_x64() 157 const __m512 vl = _mm512_permutexvar_ps(_mm512_castps_si512(vn), vtable); in xnn_f32_sigmoid_ukernel__avx512f_rr1_lut16_p3_perm_scalef_div_x64()
|
D | avx512f-rr1-lut16-p3-perm-scalef-nr1fma-x80.c | 60 const __m512 vl0 = _mm512_permutexvar_ps(_mm512_castps_si512(vn0), vtable); in xnn_f32_sigmoid_ukernel__avx512f_rr1_lut16_p3_perm_scalef_nr1fma_x80() 61 const __m512 vl1 = _mm512_permutexvar_ps(_mm512_castps_si512(vn1), vtable); in xnn_f32_sigmoid_ukernel__avx512f_rr1_lut16_p3_perm_scalef_nr1fma_x80() 62 const __m512 vl2 = _mm512_permutexvar_ps(_mm512_castps_si512(vn2), vtable); in xnn_f32_sigmoid_ukernel__avx512f_rr1_lut16_p3_perm_scalef_nr1fma_x80() 63 const __m512 vl3 = _mm512_permutexvar_ps(_mm512_castps_si512(vn3), vtable); in xnn_f32_sigmoid_ukernel__avx512f_rr1_lut16_p3_perm_scalef_nr1fma_x80() 64 const __m512 vl4 = _mm512_permutexvar_ps(_mm512_castps_si512(vn4), vtable); in xnn_f32_sigmoid_ukernel__avx512f_rr1_lut16_p3_perm_scalef_nr1fma_x80() 153 const __m512 vl = _mm512_permutexvar_ps(_mm512_castps_si512(vn), vtable); in xnn_f32_sigmoid_ukernel__avx512f_rr1_lut16_p3_perm_scalef_nr1fma_x80() 188 const __m512 vl = _mm512_permutexvar_ps(_mm512_castps_si512(vn), vtable); in xnn_f32_sigmoid_ukernel__avx512f_rr1_lut16_p3_perm_scalef_nr1fma_x80()
|
D | avx512f-rr1-lut16-p3-perm-scalef-nr1fma-x96.c | 63 const __m512 vl0 = _mm512_permutexvar_ps(_mm512_castps_si512(vn0), vtable); in xnn_f32_sigmoid_ukernel__avx512f_rr1_lut16_p3_perm_scalef_nr1fma_x96() 64 const __m512 vl1 = _mm512_permutexvar_ps(_mm512_castps_si512(vn1), vtable); in xnn_f32_sigmoid_ukernel__avx512f_rr1_lut16_p3_perm_scalef_nr1fma_x96() 65 const __m512 vl2 = _mm512_permutexvar_ps(_mm512_castps_si512(vn2), vtable); in xnn_f32_sigmoid_ukernel__avx512f_rr1_lut16_p3_perm_scalef_nr1fma_x96() 66 const __m512 vl3 = _mm512_permutexvar_ps(_mm512_castps_si512(vn3), vtable); in xnn_f32_sigmoid_ukernel__avx512f_rr1_lut16_p3_perm_scalef_nr1fma_x96() 67 const __m512 vl4 = _mm512_permutexvar_ps(_mm512_castps_si512(vn4), vtable); in xnn_f32_sigmoid_ukernel__avx512f_rr1_lut16_p3_perm_scalef_nr1fma_x96() 68 const __m512 vl5 = _mm512_permutexvar_ps(_mm512_castps_si512(vn5), vtable); in xnn_f32_sigmoid_ukernel__avx512f_rr1_lut16_p3_perm_scalef_nr1fma_x96() 170 const __m512 vl = _mm512_permutexvar_ps(_mm512_castps_si512(vn), vtable); in xnn_f32_sigmoid_ukernel__avx512f_rr1_lut16_p3_perm_scalef_nr1fma_x96() 205 const __m512 vl = _mm512_permutexvar_ps(_mm512_castps_si512(vn), vtable); in xnn_f32_sigmoid_ukernel__avx512f_rr1_lut16_p3_perm_scalef_nr1fma_x96()
|
D | avx512f-rr1-lut16-p3-perm-scalef-div-x48.c | 54 const __m512 vl0 = _mm512_permutexvar_ps(_mm512_castps_si512(vn0), vtable); in xnn_f32_sigmoid_ukernel__avx512f_rr1_lut16_p3_perm_scalef_div_x48() 55 const __m512 vl1 = _mm512_permutexvar_ps(_mm512_castps_si512(vn1), vtable); in xnn_f32_sigmoid_ukernel__avx512f_rr1_lut16_p3_perm_scalef_div_x48() 56 const __m512 vl2 = _mm512_permutexvar_ps(_mm512_castps_si512(vn2), vtable); in xnn_f32_sigmoid_ukernel__avx512f_rr1_lut16_p3_perm_scalef_div_x48() 110 const __m512 vl = _mm512_permutexvar_ps(_mm512_castps_si512(vn), vtable); in xnn_f32_sigmoid_ukernel__avx512f_rr1_lut16_p3_perm_scalef_div_x48() 142 const __m512 vl = _mm512_permutexvar_ps(_mm512_castps_si512(vn), vtable); in xnn_f32_sigmoid_ukernel__avx512f_rr1_lut16_p3_perm_scalef_div_x48()
|
D | avx512f-rr1-lut16-p3-perm-scalef-nr1fma-x64.c | 57 const __m512 vl0 = _mm512_permutexvar_ps(_mm512_castps_si512(vn0), vtable); in xnn_f32_sigmoid_ukernel__avx512f_rr1_lut16_p3_perm_scalef_nr1fma_x64() 58 const __m512 vl1 = _mm512_permutexvar_ps(_mm512_castps_si512(vn1), vtable); in xnn_f32_sigmoid_ukernel__avx512f_rr1_lut16_p3_perm_scalef_nr1fma_x64() 59 const __m512 vl2 = _mm512_permutexvar_ps(_mm512_castps_si512(vn2), vtable); in xnn_f32_sigmoid_ukernel__avx512f_rr1_lut16_p3_perm_scalef_nr1fma_x64() 60 const __m512 vl3 = _mm512_permutexvar_ps(_mm512_castps_si512(vn3), vtable); in xnn_f32_sigmoid_ukernel__avx512f_rr1_lut16_p3_perm_scalef_nr1fma_x64() 136 const __m512 vl = _mm512_permutexvar_ps(_mm512_castps_si512(vn), vtable); in xnn_f32_sigmoid_ukernel__avx512f_rr1_lut16_p3_perm_scalef_nr1fma_x64() 171 const __m512 vl = _mm512_permutexvar_ps(_mm512_castps_si512(vn), vtable); in xnn_f32_sigmoid_ukernel__avx512f_rr1_lut16_p3_perm_scalef_nr1fma_x64()
|
D | avx512f-rr1-lut16-p3-perm-scalef-nr1fma-x48.c | 54 const __m512 vl0 = _mm512_permutexvar_ps(_mm512_castps_si512(vn0), vtable); in xnn_f32_sigmoid_ukernel__avx512f_rr1_lut16_p3_perm_scalef_nr1fma_x48() 55 const __m512 vl1 = _mm512_permutexvar_ps(_mm512_castps_si512(vn1), vtable); in xnn_f32_sigmoid_ukernel__avx512f_rr1_lut16_p3_perm_scalef_nr1fma_x48() 56 const __m512 vl2 = _mm512_permutexvar_ps(_mm512_castps_si512(vn2), vtable); in xnn_f32_sigmoid_ukernel__avx512f_rr1_lut16_p3_perm_scalef_nr1fma_x48() 119 const __m512 vl = _mm512_permutexvar_ps(_mm512_castps_si512(vn), vtable); in xnn_f32_sigmoid_ukernel__avx512f_rr1_lut16_p3_perm_scalef_nr1fma_x48() 154 const __m512 vl = _mm512_permutexvar_ps(_mm512_castps_si512(vn), vtable); in xnn_f32_sigmoid_ukernel__avx512f_rr1_lut16_p3_perm_scalef_nr1fma_x48()
|
D | avx512f-rr1-lut16-p3-perm-scalef-div-x32.c | 51 const __m512 vl0 = _mm512_permutexvar_ps(_mm512_castps_si512(vn0), vtable); in xnn_f32_sigmoid_ukernel__avx512f_rr1_lut16_p3_perm_scalef_div_x32() 52 const __m512 vl1 = _mm512_permutexvar_ps(_mm512_castps_si512(vn1), vtable); in xnn_f32_sigmoid_ukernel__avx512f_rr1_lut16_p3_perm_scalef_div_x32() 95 const __m512 vl = _mm512_permutexvar_ps(_mm512_castps_si512(vn), vtable); in xnn_f32_sigmoid_ukernel__avx512f_rr1_lut16_p3_perm_scalef_div_x32() 127 const __m512 vl = _mm512_permutexvar_ps(_mm512_castps_si512(vn), vtable); in xnn_f32_sigmoid_ukernel__avx512f_rr1_lut16_p3_perm_scalef_div_x32()
|
D | avx512f-rr1-lut16-p3-perm-scalef-nr1fma-x32.c | 51 const __m512 vl0 = _mm512_permutexvar_ps(_mm512_castps_si512(vn0), vtable); in xnn_f32_sigmoid_ukernel__avx512f_rr1_lut16_p3_perm_scalef_nr1fma_x32() 52 const __m512 vl1 = _mm512_permutexvar_ps(_mm512_castps_si512(vn1), vtable); in xnn_f32_sigmoid_ukernel__avx512f_rr1_lut16_p3_perm_scalef_nr1fma_x32() 102 const __m512 vl = _mm512_permutexvar_ps(_mm512_castps_si512(vn), vtable); in xnn_f32_sigmoid_ukernel__avx512f_rr1_lut16_p3_perm_scalef_nr1fma_x32() 137 const __m512 vl = _mm512_permutexvar_ps(_mm512_castps_si512(vn), vtable); in xnn_f32_sigmoid_ukernel__avx512f_rr1_lut16_p3_perm_scalef_nr1fma_x32()
|
D | avx512f-rr1-lut16-p3-perm-scalef-div-x16.c | 47 const __m512 vl = _mm512_permutexvar_ps(_mm512_castps_si512(vn), vtable); in xnn_f32_sigmoid_ukernel__avx512f_rr1_lut16_p3_perm_scalef_div_x16() 79 const __m512 vl = _mm512_permutexvar_ps(_mm512_castps_si512(vn), vtable); in xnn_f32_sigmoid_ukernel__avx512f_rr1_lut16_p3_perm_scalef_div_x16()
|
D | avx512f-rr1-lut16-p3-perm-scalef-nr1fma-x16.c | 47 const __m512 vl = _mm512_permutexvar_ps(_mm512_castps_si512(vn), vtable); in xnn_f32_sigmoid_ukernel__avx512f_rr1_lut16_p3_perm_scalef_nr1fma_x16() 82 const __m512 vl = _mm512_permutexvar_ps(_mm512_castps_si512(vn), vtable); in xnn_f32_sigmoid_ukernel__avx512f_rr1_lut16_p3_perm_scalef_nr1fma_x16()
|
/external/XNNPACK/src/math/ |
D | exp-avx512f-rr2-lut16-p3-perm-scalef.c | 46 const __m512 vl = _mm512_permutexvar_ps(_mm512_castps_si512(vn), vtable); in xnn_math_f32_exp__avx512f_rr2_lut16_p3_perm_scalef()
|
D | sigmoid-avx512f-rr1-lut16-p3-perm-scalef-div.c | 61 const __m512 vl = _mm512_permutexvar_ps(_mm512_castps_si512(vn), vtable); in xnn_math_f32_sigmoid__avx512f_rr1_lut16_p3_perm_scalef_div()
|
D | sigmoid-avx512f-rr2-lut16-p3-perm-scalef-div.c | 62 const __m512 vl = _mm512_permutexvar_ps(_mm512_castps_si512(vn), vtable); in xnn_math_f32_sigmoid__avx512f_rr2_lut16_p3_perm_scalef_div()
|
D | sigmoid-avx512f-rr1-lut16-p3-perm-scalef-nr1fma.c | 61 const __m512 vl = _mm512_permutexvar_ps(_mm512_castps_si512(vn), vtable); in xnn_math_f32_sigmoid__avx512f_rr1_lut16_p3_perm_scalef_nr1fma()
|
D | sigmoid-avx512f-rr1-lut16-p3-perm-scalef-nr1fma1adj.c | 61 const __m512 vl = _mm512_permutexvar_ps(_mm512_castps_si512(vn), vtable); in xnn_math_f32_sigmoid__avx512f_rr1_lut16_p3_perm_scalef_nr1fma1adj()
|
D | sigmoid-avx512f-rr2-lut16-p3-perm-scalef-nr1fma.c | 62 const __m512 vl = _mm512_permutexvar_ps(_mm512_castps_si512(vn), vtable); in xnn_math_f32_sigmoid__avx512f_rr2_lut16_p3_perm_scalef_nr1fma()
|
D | sigmoid-avx512f-rr2-lut16-p3-perm-scalef-nr1fma1adj.c | 62 const __m512 vl = _mm512_permutexvar_ps(_mm512_castps_si512(vn), vtable); in xnn_math_f32_sigmoid__avx512f_rr2_lut16_p3_perm_scalef_nr1fma1adj()
|
D | exp-avx512f-rr2-lut16-p3-perm.c | 74 const __m512 vl = _mm512_permutexvar_ps(_mm512_castps_si512(vn), vtable); in xnn_math_f32_exp__avx512f_rr2_lut16_p3_perm()
|
/external/XNNPACK/src/f32-sigmoid/ |
D | avx512f-lut16-p3-perm-scalef.c.in | 60 const __m512 vl${ABC[N]} = _mm512_permutexvar_ps(_mm512_castps_si512(vn${ABC[N]}), vtable); 125 const __m512 vl = _mm512_permutexvar_ps(_mm512_castps_si512(vn), vtable); 169 const __m512 vl = _mm512_permutexvar_ps(_mm512_castps_si512(vn), vtable);
|