/external/XNNPACK/src/f32-prelu/gen/ |
D | avx512f-2x32.c | 87 const __mmask16 vsign1 = _mm512_cmp_ps_mask(vi1, vzero, _CMP_LT_OQ); in xnn_f32_prelu_ukernel__avx512f_2x32() local 88 const __m512 vacc1 = _mm512_mask_mul_ps(vi1, vsign1, vi1, vw); in xnn_f32_prelu_ukernel__avx512f_2x32() 110 const __mmask16 vsign1 = _mm512_cmp_ps_mask(vi1, vzero, _CMP_LT_OQ); in xnn_f32_prelu_ukernel__avx512f_2x32() local 111 const __m512 vacc1 = _mm512_mask_mul_ps(vi1, vsign1, vi1, vw); in xnn_f32_prelu_ukernel__avx512f_2x32()
|
D | avx512f-2x16.c | 82 const __mmask16 vsign1 = _mm512_cmp_ps_mask(vi1, vzero, _CMP_LT_OQ); in xnn_f32_prelu_ukernel__avx512f_2x16() local 83 const __m512 vacc1 = _mm512_mask_mul_ps(vi1, vsign1, vi1, vw); in xnn_f32_prelu_ukernel__avx512f_2x16()
|
/external/XNNPACK/src/f32-velu/gen/ |
D | velu-avx512f-rr1-lut16-p3-perm-x32.c | 84 const __mmask16 vsign1 = _mm512_cmp_ps_mask(vx1, vzero, _CMP_NLT_US); in xnn_f32_velu_ukernel__avx512f_rr1_lut16_p3_perm_x32() local 87 vy1 = _mm512_mask_mul_ps(vy1, vsign1, vx1, vbeta); in xnn_f32_velu_ukernel__avx512f_rr1_lut16_p3_perm_x32()
|
D | velu-avx512f-rr1-p6-x32.c | 88 const __mmask16 vsign1 = _mm512_cmp_ps_mask(vx1, vzero, _CMP_NLT_US); in xnn_f32_velu_ukernel__avx512f_rr1_p6_x32() local 91 vy1 = _mm512_mask_mul_ps(vy1, vsign1, vx1, vbeta); in xnn_f32_velu_ukernel__avx512f_rr1_p6_x32()
|
D | velu-avx512f-rr1-lut16-p3-perm-x48.c | 97 const __mmask16 vsign1 = _mm512_cmp_ps_mask(vx1, vzero, _CMP_NLT_US); in xnn_f32_velu_ukernel__avx512f_rr1_lut16_p3_perm_x48() local 102 vy1 = _mm512_mask_mul_ps(vy1, vsign1, vx1, vbeta); in xnn_f32_velu_ukernel__avx512f_rr1_lut16_p3_perm_x48()
|
D | velu-avx512f-rr1-p6-x48.c | 102 const __mmask16 vsign1 = _mm512_cmp_ps_mask(vx1, vzero, _CMP_NLT_US); in xnn_f32_velu_ukernel__avx512f_rr1_p6_x48() local 107 vy1 = _mm512_mask_mul_ps(vy1, vsign1, vx1, vbeta); in xnn_f32_velu_ukernel__avx512f_rr1_p6_x48()
|
D | velu-avx512f-rr1-p6-x64.c | 116 const __mmask16 vsign1 = _mm512_cmp_ps_mask(vx1, vzero, _CMP_NLT_US); in xnn_f32_velu_ukernel__avx512f_rr1_p6_x64() local 123 vy1 = _mm512_mask_mul_ps(vy1, vsign1, vx1, vbeta); in xnn_f32_velu_ukernel__avx512f_rr1_p6_x64()
|
D | velu-avx512f-rr1-lut16-p3-perm-x64.c | 110 const __mmask16 vsign1 = _mm512_cmp_ps_mask(vx1, vzero, _CMP_NLT_US); in xnn_f32_velu_ukernel__avx512f_rr1_lut16_p3_perm_x64() local 117 vy1 = _mm512_mask_mul_ps(vy1, vsign1, vx1, vbeta); in xnn_f32_velu_ukernel__avx512f_rr1_lut16_p3_perm_x64()
|
D | velu-avx512f-rr1-lut16-p3-perm-x80.c | 123 const __mmask16 vsign1 = _mm512_cmp_ps_mask(vx1, vzero, _CMP_NLT_US); in xnn_f32_velu_ukernel__avx512f_rr1_lut16_p3_perm_x80() local 132 vy1 = _mm512_mask_mul_ps(vy1, vsign1, vx1, vbeta); in xnn_f32_velu_ukernel__avx512f_rr1_lut16_p3_perm_x80()
|
D | velu-avx512f-rr1-p6-x80.c | 130 const __mmask16 vsign1 = _mm512_cmp_ps_mask(vx1, vzero, _CMP_NLT_US); in xnn_f32_velu_ukernel__avx512f_rr1_p6_x80() local 139 vy1 = _mm512_mask_mul_ps(vy1, vsign1, vx1, vbeta); in xnn_f32_velu_ukernel__avx512f_rr1_p6_x80()
|
D | velu-avx512f-rr1-lut16-p3-perm-x96.c | 136 const __mmask16 vsign1 = _mm512_cmp_ps_mask(vx1, vzero, _CMP_NLT_US); in xnn_f32_velu_ukernel__avx512f_rr1_lut16_p3_perm_x96() local 147 vy1 = _mm512_mask_mul_ps(vy1, vsign1, vx1, vbeta); in xnn_f32_velu_ukernel__avx512f_rr1_lut16_p3_perm_x96()
|
D | velu-avx512f-rr1-p6-x96.c | 144 const __mmask16 vsign1 = _mm512_cmp_ps_mask(vx1, vzero, _CMP_NLT_US); in xnn_f32_velu_ukernel__avx512f_rr1_p6_x96() local 155 vy1 = _mm512_mask_mul_ps(vy1, vsign1, vx1, vbeta); in xnn_f32_velu_ukernel__avx512f_rr1_p6_x96()
|
D | velu-avx512f-rr1-lut16-p3-perm-x112.c | 149 const __mmask16 vsign1 = _mm512_cmp_ps_mask(vx1, vzero, _CMP_NLT_US); in xnn_f32_velu_ukernel__avx512f_rr1_lut16_p3_perm_x112() local 162 vy1 = _mm512_mask_mul_ps(vy1, vsign1, vx1, vbeta); in xnn_f32_velu_ukernel__avx512f_rr1_lut16_p3_perm_x112()
|
D | velu-avx512f-rr1-p6-x112.c | 158 const __mmask16 vsign1 = _mm512_cmp_ps_mask(vx1, vzero, _CMP_NLT_US); in xnn_f32_velu_ukernel__avx512f_rr1_p6_x112() local 171 vy1 = _mm512_mask_mul_ps(vy1, vsign1, vx1, vbeta); in xnn_f32_velu_ukernel__avx512f_rr1_p6_x112()
|
D | velu-avx512f-rr1-lut16-p3-perm-x128.c | 162 const __mmask16 vsign1 = _mm512_cmp_ps_mask(vx1, vzero, _CMP_NLT_US); in xnn_f32_velu_ukernel__avx512f_rr1_lut16_p3_perm_x128() local 177 vy1 = _mm512_mask_mul_ps(vy1, vsign1, vx1, vbeta); in xnn_f32_velu_ukernel__avx512f_rr1_lut16_p3_perm_x128()
|
D | velu-avx512f-rr1-p6-x128.c | 172 const __mmask16 vsign1 = _mm512_cmp_ps_mask(vx1, vzero, _CMP_NLT_US); in xnn_f32_velu_ukernel__avx512f_rr1_p6_x128() local 187 vy1 = _mm512_mask_mul_ps(vy1, vsign1, vx1, vbeta); in xnn_f32_velu_ukernel__avx512f_rr1_p6_x128()
|