Home
last modified time | relevance | path

Searched refs:_mm512_slli_epi32 (Results 1 – 25 of 30) sorted by relevance

12

/external/XNNPACK/src/f32-velu/gen/
Dvelu-avx512f-rr1-lut16-p3-perm-x128.c71 const __m512i ven0 = _mm512_slli_epi32(_mm512_castps_si512(vn0), 19); in xnn_f32_velu_ukernel__avx512f_rr1_lut16_p3_perm_x128()
73 const __m512i ven1 = _mm512_slli_epi32(_mm512_castps_si512(vn1), 19); in xnn_f32_velu_ukernel__avx512f_rr1_lut16_p3_perm_x128()
75 const __m512i ven2 = _mm512_slli_epi32(_mm512_castps_si512(vn2), 19); in xnn_f32_velu_ukernel__avx512f_rr1_lut16_p3_perm_x128()
77 const __m512i ven3 = _mm512_slli_epi32(_mm512_castps_si512(vn3), 19); in xnn_f32_velu_ukernel__avx512f_rr1_lut16_p3_perm_x128()
79 const __m512i ven4 = _mm512_slli_epi32(_mm512_castps_si512(vn4), 19); in xnn_f32_velu_ukernel__avx512f_rr1_lut16_p3_perm_x128()
81 const __m512i ven5 = _mm512_slli_epi32(_mm512_castps_si512(vn5), 19); in xnn_f32_velu_ukernel__avx512f_rr1_lut16_p3_perm_x128()
83 const __m512i ven6 = _mm512_slli_epi32(_mm512_castps_si512(vn6), 19); in xnn_f32_velu_ukernel__avx512f_rr1_lut16_p3_perm_x128()
85 const __m512i ven7 = _mm512_slli_epi32(_mm512_castps_si512(vn7), 19); in xnn_f32_velu_ukernel__avx512f_rr1_lut16_p3_perm_x128()
203 const __m512i ven = _mm512_slli_epi32(_mm512_castps_si512(vn), 19); in xnn_f32_velu_ukernel__avx512f_rr1_lut16_p3_perm_x128()
236 const __m512i ven = _mm512_slli_epi32(_mm512_castps_si512(vn), 19); in xnn_f32_velu_ukernel__avx512f_rr1_lut16_p3_perm_x128()
Dvelu-avx512f-rr1-lut16-p3-perm-x112.c68 const __m512i ven0 = _mm512_slli_epi32(_mm512_castps_si512(vn0), 19); in xnn_f32_velu_ukernel__avx512f_rr1_lut16_p3_perm_x112()
70 const __m512i ven1 = _mm512_slli_epi32(_mm512_castps_si512(vn1), 19); in xnn_f32_velu_ukernel__avx512f_rr1_lut16_p3_perm_x112()
72 const __m512i ven2 = _mm512_slli_epi32(_mm512_castps_si512(vn2), 19); in xnn_f32_velu_ukernel__avx512f_rr1_lut16_p3_perm_x112()
74 const __m512i ven3 = _mm512_slli_epi32(_mm512_castps_si512(vn3), 19); in xnn_f32_velu_ukernel__avx512f_rr1_lut16_p3_perm_x112()
76 const __m512i ven4 = _mm512_slli_epi32(_mm512_castps_si512(vn4), 19); in xnn_f32_velu_ukernel__avx512f_rr1_lut16_p3_perm_x112()
78 const __m512i ven5 = _mm512_slli_epi32(_mm512_castps_si512(vn5), 19); in xnn_f32_velu_ukernel__avx512f_rr1_lut16_p3_perm_x112()
80 const __m512i ven6 = _mm512_slli_epi32(_mm512_castps_si512(vn6), 19); in xnn_f32_velu_ukernel__avx512f_rr1_lut16_p3_perm_x112()
186 const __m512i ven = _mm512_slli_epi32(_mm512_castps_si512(vn), 19); in xnn_f32_velu_ukernel__avx512f_rr1_lut16_p3_perm_x112()
219 const __m512i ven = _mm512_slli_epi32(_mm512_castps_si512(vn), 19); in xnn_f32_velu_ukernel__avx512f_rr1_lut16_p3_perm_x112()
Dvelu-avx512f-rr1-p6-x128.c71 __m512 vs0 = _mm512_castsi512_ps(_mm512_slli_epi32(_mm512_castps_si512(vn0), 23)); in xnn_f32_velu_ukernel__avx512f_rr1_p6_x128()
73 __m512 vs1 = _mm512_castsi512_ps(_mm512_slli_epi32(_mm512_castps_si512(vn1), 23)); in xnn_f32_velu_ukernel__avx512f_rr1_p6_x128()
75 __m512 vs2 = _mm512_castsi512_ps(_mm512_slli_epi32(_mm512_castps_si512(vn2), 23)); in xnn_f32_velu_ukernel__avx512f_rr1_p6_x128()
77 __m512 vs3 = _mm512_castsi512_ps(_mm512_slli_epi32(_mm512_castps_si512(vn3), 23)); in xnn_f32_velu_ukernel__avx512f_rr1_p6_x128()
79 __m512 vs4 = _mm512_castsi512_ps(_mm512_slli_epi32(_mm512_castps_si512(vn4), 23)); in xnn_f32_velu_ukernel__avx512f_rr1_p6_x128()
81 __m512 vs5 = _mm512_castsi512_ps(_mm512_slli_epi32(_mm512_castps_si512(vn5), 23)); in xnn_f32_velu_ukernel__avx512f_rr1_p6_x128()
83 __m512 vs6 = _mm512_castsi512_ps(_mm512_slli_epi32(_mm512_castps_si512(vn6), 23)); in xnn_f32_velu_ukernel__avx512f_rr1_p6_x128()
85 __m512 vs7 = _mm512_castsi512_ps(_mm512_slli_epi32(_mm512_castps_si512(vn7), 23)); in xnn_f32_velu_ukernel__avx512f_rr1_p6_x128()
213 __m512 vs = _mm512_castsi512_ps(_mm512_slli_epi32(_mm512_castps_si512(vn), 23)); in xnn_f32_velu_ukernel__avx512f_rr1_p6_x128()
247 __m512 vs = _mm512_castsi512_ps(_mm512_slli_epi32(_mm512_castps_si512(vn), 23)); in xnn_f32_velu_ukernel__avx512f_rr1_p6_x128()
Dvelu-avx512f-rr1-lut16-p3-perm-x80.c62 const __m512i ven0 = _mm512_slli_epi32(_mm512_castps_si512(vn0), 19); in xnn_f32_velu_ukernel__avx512f_rr1_lut16_p3_perm_x80()
64 const __m512i ven1 = _mm512_slli_epi32(_mm512_castps_si512(vn1), 19); in xnn_f32_velu_ukernel__avx512f_rr1_lut16_p3_perm_x80()
66 const __m512i ven2 = _mm512_slli_epi32(_mm512_castps_si512(vn2), 19); in xnn_f32_velu_ukernel__avx512f_rr1_lut16_p3_perm_x80()
68 const __m512i ven3 = _mm512_slli_epi32(_mm512_castps_si512(vn3), 19); in xnn_f32_velu_ukernel__avx512f_rr1_lut16_p3_perm_x80()
70 const __m512i ven4 = _mm512_slli_epi32(_mm512_castps_si512(vn4), 19); in xnn_f32_velu_ukernel__avx512f_rr1_lut16_p3_perm_x80()
152 const __m512i ven = _mm512_slli_epi32(_mm512_castps_si512(vn), 19); in xnn_f32_velu_ukernel__avx512f_rr1_lut16_p3_perm_x80()
185 const __m512i ven = _mm512_slli_epi32(_mm512_castps_si512(vn), 19); in xnn_f32_velu_ukernel__avx512f_rr1_lut16_p3_perm_x80()
Dvelu-avx512f-rr1-lut16-p3-perm-x96.c65 const __m512i ven0 = _mm512_slli_epi32(_mm512_castps_si512(vn0), 19); in xnn_f32_velu_ukernel__avx512f_rr1_lut16_p3_perm_x96()
67 const __m512i ven1 = _mm512_slli_epi32(_mm512_castps_si512(vn1), 19); in xnn_f32_velu_ukernel__avx512f_rr1_lut16_p3_perm_x96()
69 const __m512i ven2 = _mm512_slli_epi32(_mm512_castps_si512(vn2), 19); in xnn_f32_velu_ukernel__avx512f_rr1_lut16_p3_perm_x96()
71 const __m512i ven3 = _mm512_slli_epi32(_mm512_castps_si512(vn3), 19); in xnn_f32_velu_ukernel__avx512f_rr1_lut16_p3_perm_x96()
73 const __m512i ven4 = _mm512_slli_epi32(_mm512_castps_si512(vn4), 19); in xnn_f32_velu_ukernel__avx512f_rr1_lut16_p3_perm_x96()
75 const __m512i ven5 = _mm512_slli_epi32(_mm512_castps_si512(vn5), 19); in xnn_f32_velu_ukernel__avx512f_rr1_lut16_p3_perm_x96()
169 const __m512i ven = _mm512_slli_epi32(_mm512_castps_si512(vn), 19); in xnn_f32_velu_ukernel__avx512f_rr1_lut16_p3_perm_x96()
202 const __m512i ven = _mm512_slli_epi32(_mm512_castps_si512(vn), 19); in xnn_f32_velu_ukernel__avx512f_rr1_lut16_p3_perm_x96()
Dvelu-avx512f-rr1-p6-x96.c65 __m512 vs0 = _mm512_castsi512_ps(_mm512_slli_epi32(_mm512_castps_si512(vn0), 23)); in xnn_f32_velu_ukernel__avx512f_rr1_p6_x96()
67 __m512 vs1 = _mm512_castsi512_ps(_mm512_slli_epi32(_mm512_castps_si512(vn1), 23)); in xnn_f32_velu_ukernel__avx512f_rr1_p6_x96()
69 __m512 vs2 = _mm512_castsi512_ps(_mm512_slli_epi32(_mm512_castps_si512(vn2), 23)); in xnn_f32_velu_ukernel__avx512f_rr1_p6_x96()
71 __m512 vs3 = _mm512_castsi512_ps(_mm512_slli_epi32(_mm512_castps_si512(vn3), 23)); in xnn_f32_velu_ukernel__avx512f_rr1_p6_x96()
73 __m512 vs4 = _mm512_castsi512_ps(_mm512_slli_epi32(_mm512_castps_si512(vn4), 23)); in xnn_f32_velu_ukernel__avx512f_rr1_p6_x96()
75 __m512 vs5 = _mm512_castsi512_ps(_mm512_slli_epi32(_mm512_castps_si512(vn5), 23)); in xnn_f32_velu_ukernel__avx512f_rr1_p6_x96()
177 __m512 vs = _mm512_castsi512_ps(_mm512_slli_epi32(_mm512_castps_si512(vn), 23)); in xnn_f32_velu_ukernel__avx512f_rr1_p6_x96()
211 __m512 vs = _mm512_castsi512_ps(_mm512_slli_epi32(_mm512_castps_si512(vn), 23)); in xnn_f32_velu_ukernel__avx512f_rr1_p6_x96()
Dvelu-avx512f-rr1-p6-x112.c68 __m512 vs0 = _mm512_castsi512_ps(_mm512_slli_epi32(_mm512_castps_si512(vn0), 23)); in xnn_f32_velu_ukernel__avx512f_rr1_p6_x112()
70 __m512 vs1 = _mm512_castsi512_ps(_mm512_slli_epi32(_mm512_castps_si512(vn1), 23)); in xnn_f32_velu_ukernel__avx512f_rr1_p6_x112()
72 __m512 vs2 = _mm512_castsi512_ps(_mm512_slli_epi32(_mm512_castps_si512(vn2), 23)); in xnn_f32_velu_ukernel__avx512f_rr1_p6_x112()
74 __m512 vs3 = _mm512_castsi512_ps(_mm512_slli_epi32(_mm512_castps_si512(vn3), 23)); in xnn_f32_velu_ukernel__avx512f_rr1_p6_x112()
76 __m512 vs4 = _mm512_castsi512_ps(_mm512_slli_epi32(_mm512_castps_si512(vn4), 23)); in xnn_f32_velu_ukernel__avx512f_rr1_p6_x112()
78 __m512 vs5 = _mm512_castsi512_ps(_mm512_slli_epi32(_mm512_castps_si512(vn5), 23)); in xnn_f32_velu_ukernel__avx512f_rr1_p6_x112()
80 __m512 vs6 = _mm512_castsi512_ps(_mm512_slli_epi32(_mm512_castps_si512(vn6), 23)); in xnn_f32_velu_ukernel__avx512f_rr1_p6_x112()
195 __m512 vs = _mm512_castsi512_ps(_mm512_slli_epi32(_mm512_castps_si512(vn), 23)); in xnn_f32_velu_ukernel__avx512f_rr1_p6_x112()
229 __m512 vs = _mm512_castsi512_ps(_mm512_slli_epi32(_mm512_castps_si512(vn), 23)); in xnn_f32_velu_ukernel__avx512f_rr1_p6_x112()
Dvelu-avx512f-rr1-p6-x80.c62 __m512 vs0 = _mm512_castsi512_ps(_mm512_slli_epi32(_mm512_castps_si512(vn0), 23)); in xnn_f32_velu_ukernel__avx512f_rr1_p6_x80()
64 __m512 vs1 = _mm512_castsi512_ps(_mm512_slli_epi32(_mm512_castps_si512(vn1), 23)); in xnn_f32_velu_ukernel__avx512f_rr1_p6_x80()
66 __m512 vs2 = _mm512_castsi512_ps(_mm512_slli_epi32(_mm512_castps_si512(vn2), 23)); in xnn_f32_velu_ukernel__avx512f_rr1_p6_x80()
68 __m512 vs3 = _mm512_castsi512_ps(_mm512_slli_epi32(_mm512_castps_si512(vn3), 23)); in xnn_f32_velu_ukernel__avx512f_rr1_p6_x80()
70 __m512 vs4 = _mm512_castsi512_ps(_mm512_slli_epi32(_mm512_castps_si512(vn4), 23)); in xnn_f32_velu_ukernel__avx512f_rr1_p6_x80()
159 __m512 vs = _mm512_castsi512_ps(_mm512_slli_epi32(_mm512_castps_si512(vn), 23)); in xnn_f32_velu_ukernel__avx512f_rr1_p6_x80()
193 __m512 vs = _mm512_castsi512_ps(_mm512_slli_epi32(_mm512_castps_si512(vn), 23)); in xnn_f32_velu_ukernel__avx512f_rr1_p6_x80()
Dvelu-avx512f-rr1-p6-x64.c59 __m512 vs0 = _mm512_castsi512_ps(_mm512_slli_epi32(_mm512_castps_si512(vn0), 23)); in xnn_f32_velu_ukernel__avx512f_rr1_p6_x64()
61 __m512 vs1 = _mm512_castsi512_ps(_mm512_slli_epi32(_mm512_castps_si512(vn1), 23)); in xnn_f32_velu_ukernel__avx512f_rr1_p6_x64()
63 __m512 vs2 = _mm512_castsi512_ps(_mm512_slli_epi32(_mm512_castps_si512(vn2), 23)); in xnn_f32_velu_ukernel__avx512f_rr1_p6_x64()
65 __m512 vs3 = _mm512_castsi512_ps(_mm512_slli_epi32(_mm512_castps_si512(vn3), 23)); in xnn_f32_velu_ukernel__avx512f_rr1_p6_x64()
141 __m512 vs = _mm512_castsi512_ps(_mm512_slli_epi32(_mm512_castps_si512(vn), 23)); in xnn_f32_velu_ukernel__avx512f_rr1_p6_x64()
175 __m512 vs = _mm512_castsi512_ps(_mm512_slli_epi32(_mm512_castps_si512(vn), 23)); in xnn_f32_velu_ukernel__avx512f_rr1_p6_x64()
Dvelu-avx512f-rr1-lut16-p3-perm-x64.c59 const __m512i ven0 = _mm512_slli_epi32(_mm512_castps_si512(vn0), 19); in xnn_f32_velu_ukernel__avx512f_rr1_lut16_p3_perm_x64()
61 const __m512i ven1 = _mm512_slli_epi32(_mm512_castps_si512(vn1), 19); in xnn_f32_velu_ukernel__avx512f_rr1_lut16_p3_perm_x64()
63 const __m512i ven2 = _mm512_slli_epi32(_mm512_castps_si512(vn2), 19); in xnn_f32_velu_ukernel__avx512f_rr1_lut16_p3_perm_x64()
65 const __m512i ven3 = _mm512_slli_epi32(_mm512_castps_si512(vn3), 19); in xnn_f32_velu_ukernel__avx512f_rr1_lut16_p3_perm_x64()
135 const __m512i ven = _mm512_slli_epi32(_mm512_castps_si512(vn), 19); in xnn_f32_velu_ukernel__avx512f_rr1_lut16_p3_perm_x64()
168 const __m512i ven = _mm512_slli_epi32(_mm512_castps_si512(vn), 19); in xnn_f32_velu_ukernel__avx512f_rr1_lut16_p3_perm_x64()
Dvelu-avx512f-rr1-lut16-p3-perm-x48.c56 const __m512i ven0 = _mm512_slli_epi32(_mm512_castps_si512(vn0), 19); in xnn_f32_velu_ukernel__avx512f_rr1_lut16_p3_perm_x48()
58 const __m512i ven1 = _mm512_slli_epi32(_mm512_castps_si512(vn1), 19); in xnn_f32_velu_ukernel__avx512f_rr1_lut16_p3_perm_x48()
60 const __m512i ven2 = _mm512_slli_epi32(_mm512_castps_si512(vn2), 19); in xnn_f32_velu_ukernel__avx512f_rr1_lut16_p3_perm_x48()
118 const __m512i ven = _mm512_slli_epi32(_mm512_castps_si512(vn), 19); in xnn_f32_velu_ukernel__avx512f_rr1_lut16_p3_perm_x48()
151 const __m512i ven = _mm512_slli_epi32(_mm512_castps_si512(vn), 19); in xnn_f32_velu_ukernel__avx512f_rr1_lut16_p3_perm_x48()
Dvelu-avx512f-rr1-p6-x48.c56 __m512 vs0 = _mm512_castsi512_ps(_mm512_slli_epi32(_mm512_castps_si512(vn0), 23)); in xnn_f32_velu_ukernel__avx512f_rr1_p6_x48()
58 __m512 vs1 = _mm512_castsi512_ps(_mm512_slli_epi32(_mm512_castps_si512(vn1), 23)); in xnn_f32_velu_ukernel__avx512f_rr1_p6_x48()
60 __m512 vs2 = _mm512_castsi512_ps(_mm512_slli_epi32(_mm512_castps_si512(vn2), 23)); in xnn_f32_velu_ukernel__avx512f_rr1_p6_x48()
123 __m512 vs = _mm512_castsi512_ps(_mm512_slli_epi32(_mm512_castps_si512(vn), 23)); in xnn_f32_velu_ukernel__avx512f_rr1_p6_x48()
157 __m512 vs = _mm512_castsi512_ps(_mm512_slli_epi32(_mm512_castps_si512(vn), 23)); in xnn_f32_velu_ukernel__avx512f_rr1_p6_x48()
Dvelu-avx512f-rr1-lut16-p3-perm-x32.c53 const __m512i ven0 = _mm512_slli_epi32(_mm512_castps_si512(vn0), 19); in xnn_f32_velu_ukernel__avx512f_rr1_lut16_p3_perm_x32()
55 const __m512i ven1 = _mm512_slli_epi32(_mm512_castps_si512(vn1), 19); in xnn_f32_velu_ukernel__avx512f_rr1_lut16_p3_perm_x32()
101 const __m512i ven = _mm512_slli_epi32(_mm512_castps_si512(vn), 19); in xnn_f32_velu_ukernel__avx512f_rr1_lut16_p3_perm_x32()
134 const __m512i ven = _mm512_slli_epi32(_mm512_castps_si512(vn), 19); in xnn_f32_velu_ukernel__avx512f_rr1_lut16_p3_perm_x32()
Dvelu-avx512f-rr1-p6-x32.c53 __m512 vs0 = _mm512_castsi512_ps(_mm512_slli_epi32(_mm512_castps_si512(vn0), 23)); in xnn_f32_velu_ukernel__avx512f_rr1_p6_x32()
55 __m512 vs1 = _mm512_castsi512_ps(_mm512_slli_epi32(_mm512_castps_si512(vn1), 23)); in xnn_f32_velu_ukernel__avx512f_rr1_p6_x32()
105 __m512 vs = _mm512_castsi512_ps(_mm512_slli_epi32(_mm512_castps_si512(vn), 23)); in xnn_f32_velu_ukernel__avx512f_rr1_p6_x32()
139 __m512 vs = _mm512_castsi512_ps(_mm512_slli_epi32(_mm512_castps_si512(vn), 23)); in xnn_f32_velu_ukernel__avx512f_rr1_p6_x32()
Dvelu-avx512f-rr1-lut16-p3-perm-x16.c50 const __m512i ven = _mm512_slli_epi32(_mm512_castps_si512(vn), 19); in xnn_f32_velu_ukernel__avx512f_rr1_lut16_p3_perm_x16()
83 const __m512i ven = _mm512_slli_epi32(_mm512_castps_si512(vn), 19); in xnn_f32_velu_ukernel__avx512f_rr1_lut16_p3_perm_x16()
Dvelu-avx512f-rr1-p6-x16.c50 __m512 vs = _mm512_castsi512_ps(_mm512_slli_epi32(_mm512_castps_si512(vn), 23)); in xnn_f32_velu_ukernel__avx512f_rr1_p6_x16()
84 __m512 vs = _mm512_castsi512_ps(_mm512_slli_epi32(_mm512_castps_si512(vn), 23)); in xnn_f32_velu_ukernel__avx512f_rr1_p6_x16()
/external/llvm-project/clang/lib/Headers/
Davx512bf16intrin.h237 return _mm512_castsi512_ps((__m512i)_mm512_slli_epi32( in _mm512_cvtpbh_ps()
253 return _mm512_castsi512_ps((__m512i)_mm512_slli_epi32( in _mm512_maskz_cvtpbh_ps()
/external/XNNPACK/src/f32-velu/
Davx512f-rr1-lut16-p3-perm.c.in55 const __m512i ven${N} = _mm512_slli_epi32(_mm512_castps_si512(vn${N}), 19);
99 const __m512i ven = _mm512_slli_epi32(_mm512_castps_si512(vn), 19);
132 const __m512i ven = _mm512_slli_epi32(_mm512_castps_si512(vn), 19);
Davx512f-rr1-p6.c.in55 __m512 vs${N} = _mm512_castsi512_ps(_mm512_slli_epi32(_mm512_castps_si512(vn${N}), 23));
104 __m512 vs = _mm512_castsi512_ps(_mm512_slli_epi32(_mm512_castps_si512(vn), 23));
138 __m512 vs = _mm512_castsi512_ps(_mm512_slli_epi32(_mm512_castps_si512(vn), 23));
/external/XNNPACK/src/math/
Dexpm1minus-avx512f-rr1-lut16-p3-perm.c67 const __m512i ven = _mm512_slli_epi32(_mm512_castps_si512(vn), 19); in xnn_math_f32_expm1minus__avx512f_rr1_lut16_p3_perm()
Dexpm1minus-avx512f-rr1-p6.c60 const __m512 vs = _mm512_castsi512_ps(_mm512_slli_epi32(_mm512_castps_si512(vn), 23)); in xnn_math_f32_expm1minus__avx512f_rr1_p6()
Dexp-avx512f-rr2-lut32-p2-perm2.c71 __m512i veo = _mm512_slli_epi32(_mm512_and_si512(_mm512_castps_si512(vn), vmantissa_mask), 18); in xnn_math_f32_exp__avx512f_rr2_lut32_p2_perm2()
Dexp-avx512f-rr2-p5.c59 __m512i veo = _mm512_slli_epi32(_mm512_castps_si512(vn), 23); in xnn_math_f32_exp__avx512f_rr2_p5()
Dexp-avx512f-rr2-lut16-p3-perm.c66 __m512i veo = _mm512_slli_epi32(_mm512_and_si512(_mm512_castps_si512(vn), vmantissa_mask), 19); in xnn_math_f32_exp__avx512f_rr2_lut16_p3_perm()
/external/tensorflow/tensorflow/core/kernels/
Dsparse_matmul_op.h451 return _mm512_castsi512_ps(_mm512_slli_epi32( in pexpand_bf16_l()
460 return _mm512_castsi512_ps(_mm512_slli_epi32( in pexpand_bf16_u()

12