| /external/XNNPACK/src/f32-velu/gen/ |
| D | velu-avx512f-rr1-lut16-p3-perm-x128.c | 140 vs0 = _mm512_fmsub_ps(vs0, valpha, valpha); in xnn_f32_velu_ukernel__avx512f_rr1_lut16_p3_perm_x128()
141 vs1 = _mm512_fmsub_ps(vs1, valpha, valpha); in xnn_f32_velu_ukernel__avx512f_rr1_lut16_p3_perm_x128()
142 vs2 = _mm512_fmsub_ps(vs2, valpha, valpha); in xnn_f32_velu_ukernel__avx512f_rr1_lut16_p3_perm_x128()
143 vs3 = _mm512_fmsub_ps(vs3, valpha, valpha); in xnn_f32_velu_ukernel__avx512f_rr1_lut16_p3_perm_x128()
144 vs4 = _mm512_fmsub_ps(vs4, valpha, valpha); in xnn_f32_velu_ukernel__avx512f_rr1_lut16_p3_perm_x128()
145 vs5 = _mm512_fmsub_ps(vs5, valpha, valpha); in xnn_f32_velu_ukernel__avx512f_rr1_lut16_p3_perm_x128()
146 vs6 = _mm512_fmsub_ps(vs6, valpha, valpha); in xnn_f32_velu_ukernel__avx512f_rr1_lut16_p3_perm_x128()
147 vs7 = _mm512_fmsub_ps(vs7, valpha, valpha); in xnn_f32_velu_ukernel__avx512f_rr1_lut16_p3_perm_x128()
214 vs = _mm512_fmsub_ps(vs, valpha, valpha); in xnn_f32_velu_ukernel__avx512f_rr1_lut16_p3_perm_x128()
247 vs = _mm512_fmsub_ps(vs, valpha, valpha); in xnn_f32_velu_ukernel__avx512f_rr1_lut16_p3_perm_x128()
|
| D | velu-avx512f-rr1-lut16-p3-perm-x112.c | 129 vs0 = _mm512_fmsub_ps(vs0, valpha, valpha); in xnn_f32_velu_ukernel__avx512f_rr1_lut16_p3_perm_x112()
130 vs1 = _mm512_fmsub_ps(vs1, valpha, valpha); in xnn_f32_velu_ukernel__avx512f_rr1_lut16_p3_perm_x112()
131 vs2 = _mm512_fmsub_ps(vs2, valpha, valpha); in xnn_f32_velu_ukernel__avx512f_rr1_lut16_p3_perm_x112()
132 vs3 = _mm512_fmsub_ps(vs3, valpha, valpha); in xnn_f32_velu_ukernel__avx512f_rr1_lut16_p3_perm_x112()
133 vs4 = _mm512_fmsub_ps(vs4, valpha, valpha); in xnn_f32_velu_ukernel__avx512f_rr1_lut16_p3_perm_x112()
134 vs5 = _mm512_fmsub_ps(vs5, valpha, valpha); in xnn_f32_velu_ukernel__avx512f_rr1_lut16_p3_perm_x112()
135 vs6 = _mm512_fmsub_ps(vs6, valpha, valpha); in xnn_f32_velu_ukernel__avx512f_rr1_lut16_p3_perm_x112()
197 vs = _mm512_fmsub_ps(vs, valpha, valpha); in xnn_f32_velu_ukernel__avx512f_rr1_lut16_p3_perm_x112()
230 vs = _mm512_fmsub_ps(vs, valpha, valpha); in xnn_f32_velu_ukernel__avx512f_rr1_lut16_p3_perm_x112()
|
| D | velu-avx512f-rr1-p6-x128.c | 150 vs0 = _mm512_fmsub_ps(vs0, valpha, valpha); in xnn_f32_velu_ukernel__avx512f_rr1_p6_x128()
151 vs1 = _mm512_fmsub_ps(vs1, valpha, valpha); in xnn_f32_velu_ukernel__avx512f_rr1_p6_x128()
152 vs2 = _mm512_fmsub_ps(vs2, valpha, valpha); in xnn_f32_velu_ukernel__avx512f_rr1_p6_x128()
153 vs3 = _mm512_fmsub_ps(vs3, valpha, valpha); in xnn_f32_velu_ukernel__avx512f_rr1_p6_x128()
154 vs4 = _mm512_fmsub_ps(vs4, valpha, valpha); in xnn_f32_velu_ukernel__avx512f_rr1_p6_x128()
155 vs5 = _mm512_fmsub_ps(vs5, valpha, valpha); in xnn_f32_velu_ukernel__avx512f_rr1_p6_x128()
156 vs6 = _mm512_fmsub_ps(vs6, valpha, valpha); in xnn_f32_velu_ukernel__avx512f_rr1_p6_x128()
157 vs7 = _mm512_fmsub_ps(vs7, valpha, valpha); in xnn_f32_velu_ukernel__avx512f_rr1_p6_x128()
225 vs = _mm512_fmsub_ps(vs, valpha, valpha); in xnn_f32_velu_ukernel__avx512f_rr1_p6_x128()
259 vs = _mm512_fmsub_ps(vs, valpha, valpha); in xnn_f32_velu_ukernel__avx512f_rr1_p6_x128()
|
| D | velu-avx512f-rr1-lut16-p3-perm-x80.c | 107 vs0 = _mm512_fmsub_ps(vs0, valpha, valpha); in xnn_f32_velu_ukernel__avx512f_rr1_lut16_p3_perm_x80()
108 vs1 = _mm512_fmsub_ps(vs1, valpha, valpha); in xnn_f32_velu_ukernel__avx512f_rr1_lut16_p3_perm_x80()
109 vs2 = _mm512_fmsub_ps(vs2, valpha, valpha); in xnn_f32_velu_ukernel__avx512f_rr1_lut16_p3_perm_x80()
110 vs3 = _mm512_fmsub_ps(vs3, valpha, valpha); in xnn_f32_velu_ukernel__avx512f_rr1_lut16_p3_perm_x80()
111 vs4 = _mm512_fmsub_ps(vs4, valpha, valpha); in xnn_f32_velu_ukernel__avx512f_rr1_lut16_p3_perm_x80()
163 vs = _mm512_fmsub_ps(vs, valpha, valpha); in xnn_f32_velu_ukernel__avx512f_rr1_lut16_p3_perm_x80()
196 vs = _mm512_fmsub_ps(vs, valpha, valpha); in xnn_f32_velu_ukernel__avx512f_rr1_lut16_p3_perm_x80()
|
| D | velu-avx512f-rr1-lut16-p3-perm-x96.c | 118 vs0 = _mm512_fmsub_ps(vs0, valpha, valpha); in xnn_f32_velu_ukernel__avx512f_rr1_lut16_p3_perm_x96()
119 vs1 = _mm512_fmsub_ps(vs1, valpha, valpha); in xnn_f32_velu_ukernel__avx512f_rr1_lut16_p3_perm_x96()
120 vs2 = _mm512_fmsub_ps(vs2, valpha, valpha); in xnn_f32_velu_ukernel__avx512f_rr1_lut16_p3_perm_x96()
121 vs3 = _mm512_fmsub_ps(vs3, valpha, valpha); in xnn_f32_velu_ukernel__avx512f_rr1_lut16_p3_perm_x96()
122 vs4 = _mm512_fmsub_ps(vs4, valpha, valpha); in xnn_f32_velu_ukernel__avx512f_rr1_lut16_p3_perm_x96()
123 vs5 = _mm512_fmsub_ps(vs5, valpha, valpha); in xnn_f32_velu_ukernel__avx512f_rr1_lut16_p3_perm_x96()
180 vs = _mm512_fmsub_ps(vs, valpha, valpha); in xnn_f32_velu_ukernel__avx512f_rr1_lut16_p3_perm_x96()
213 vs = _mm512_fmsub_ps(vs, valpha, valpha); in xnn_f32_velu_ukernel__avx512f_rr1_lut16_p3_perm_x96()
|
| D | velu-avx512f-rr1-p6-x96.c | 126 vs0 = _mm512_fmsub_ps(vs0, valpha, valpha); in xnn_f32_velu_ukernel__avx512f_rr1_p6_x96()
127 vs1 = _mm512_fmsub_ps(vs1, valpha, valpha); in xnn_f32_velu_ukernel__avx512f_rr1_p6_x96()
128 vs2 = _mm512_fmsub_ps(vs2, valpha, valpha); in xnn_f32_velu_ukernel__avx512f_rr1_p6_x96()
129 vs3 = _mm512_fmsub_ps(vs3, valpha, valpha); in xnn_f32_velu_ukernel__avx512f_rr1_p6_x96()
130 vs4 = _mm512_fmsub_ps(vs4, valpha, valpha); in xnn_f32_velu_ukernel__avx512f_rr1_p6_x96()
131 vs5 = _mm512_fmsub_ps(vs5, valpha, valpha); in xnn_f32_velu_ukernel__avx512f_rr1_p6_x96()
189 vs = _mm512_fmsub_ps(vs, valpha, valpha); in xnn_f32_velu_ukernel__avx512f_rr1_p6_x96()
223 vs = _mm512_fmsub_ps(vs, valpha, valpha); in xnn_f32_velu_ukernel__avx512f_rr1_p6_x96()
|
| D | velu-avx512f-rr1-p6-x112.c | 138 vs0 = _mm512_fmsub_ps(vs0, valpha, valpha); in xnn_f32_velu_ukernel__avx512f_rr1_p6_x112()
139 vs1 = _mm512_fmsub_ps(vs1, valpha, valpha); in xnn_f32_velu_ukernel__avx512f_rr1_p6_x112()
140 vs2 = _mm512_fmsub_ps(vs2, valpha, valpha); in xnn_f32_velu_ukernel__avx512f_rr1_p6_x112()
141 vs3 = _mm512_fmsub_ps(vs3, valpha, valpha); in xnn_f32_velu_ukernel__avx512f_rr1_p6_x112()
142 vs4 = _mm512_fmsub_ps(vs4, valpha, valpha); in xnn_f32_velu_ukernel__avx512f_rr1_p6_x112()
143 vs5 = _mm512_fmsub_ps(vs5, valpha, valpha); in xnn_f32_velu_ukernel__avx512f_rr1_p6_x112()
144 vs6 = _mm512_fmsub_ps(vs6, valpha, valpha); in xnn_f32_velu_ukernel__avx512f_rr1_p6_x112()
207 vs = _mm512_fmsub_ps(vs, valpha, valpha); in xnn_f32_velu_ukernel__avx512f_rr1_p6_x112()
241 vs = _mm512_fmsub_ps(vs, valpha, valpha); in xnn_f32_velu_ukernel__avx512f_rr1_p6_x112()
|
| D | velu-avx512f-rr1-p6-x80.c | 114 vs0 = _mm512_fmsub_ps(vs0, valpha, valpha); in xnn_f32_velu_ukernel__avx512f_rr1_p6_x80()
115 vs1 = _mm512_fmsub_ps(vs1, valpha, valpha); in xnn_f32_velu_ukernel__avx512f_rr1_p6_x80()
116 vs2 = _mm512_fmsub_ps(vs2, valpha, valpha); in xnn_f32_velu_ukernel__avx512f_rr1_p6_x80()
117 vs3 = _mm512_fmsub_ps(vs3, valpha, valpha); in xnn_f32_velu_ukernel__avx512f_rr1_p6_x80()
118 vs4 = _mm512_fmsub_ps(vs4, valpha, valpha); in xnn_f32_velu_ukernel__avx512f_rr1_p6_x80()
171 vs = _mm512_fmsub_ps(vs, valpha, valpha); in xnn_f32_velu_ukernel__avx512f_rr1_p6_x80()
205 vs = _mm512_fmsub_ps(vs, valpha, valpha); in xnn_f32_velu_ukernel__avx512f_rr1_p6_x80()
|
| D | velu-avx512f-rr1-p6-x64.c | 102 vs0 = _mm512_fmsub_ps(vs0, valpha, valpha); in xnn_f32_velu_ukernel__avx512f_rr1_p6_x64()
103 vs1 = _mm512_fmsub_ps(vs1, valpha, valpha); in xnn_f32_velu_ukernel__avx512f_rr1_p6_x64()
104 vs2 = _mm512_fmsub_ps(vs2, valpha, valpha); in xnn_f32_velu_ukernel__avx512f_rr1_p6_x64()
105 vs3 = _mm512_fmsub_ps(vs3, valpha, valpha); in xnn_f32_velu_ukernel__avx512f_rr1_p6_x64()
153 vs = _mm512_fmsub_ps(vs, valpha, valpha); in xnn_f32_velu_ukernel__avx512f_rr1_p6_x64()
187 vs = _mm512_fmsub_ps(vs, valpha, valpha); in xnn_f32_velu_ukernel__avx512f_rr1_p6_x64()
|
| D | velu-avx512f-rr1-lut16-p3-perm-x64.c | 96 vs0 = _mm512_fmsub_ps(vs0, valpha, valpha); in xnn_f32_velu_ukernel__avx512f_rr1_lut16_p3_perm_x64()
97 vs1 = _mm512_fmsub_ps(vs1, valpha, valpha); in xnn_f32_velu_ukernel__avx512f_rr1_lut16_p3_perm_x64()
98 vs2 = _mm512_fmsub_ps(vs2, valpha, valpha); in xnn_f32_velu_ukernel__avx512f_rr1_lut16_p3_perm_x64()
99 vs3 = _mm512_fmsub_ps(vs3, valpha, valpha); in xnn_f32_velu_ukernel__avx512f_rr1_lut16_p3_perm_x64()
146 vs = _mm512_fmsub_ps(vs, valpha, valpha); in xnn_f32_velu_ukernel__avx512f_rr1_lut16_p3_perm_x64()
179 vs = _mm512_fmsub_ps(vs, valpha, valpha); in xnn_f32_velu_ukernel__avx512f_rr1_lut16_p3_perm_x64()
|
| D | velu-avx512f-rr1-lut16-p3-perm-x48.c | 85 vs0 = _mm512_fmsub_ps(vs0, valpha, valpha); in xnn_f32_velu_ukernel__avx512f_rr1_lut16_p3_perm_x48()
86 vs1 = _mm512_fmsub_ps(vs1, valpha, valpha); in xnn_f32_velu_ukernel__avx512f_rr1_lut16_p3_perm_x48()
87 vs2 = _mm512_fmsub_ps(vs2, valpha, valpha); in xnn_f32_velu_ukernel__avx512f_rr1_lut16_p3_perm_x48()
129 vs = _mm512_fmsub_ps(vs, valpha, valpha); in xnn_f32_velu_ukernel__avx512f_rr1_lut16_p3_perm_x48()
162 vs = _mm512_fmsub_ps(vs, valpha, valpha); in xnn_f32_velu_ukernel__avx512f_rr1_lut16_p3_perm_x48()
|
| D | velu-avx512f-rr1-p6-x48.c | 90 vs0 = _mm512_fmsub_ps(vs0, valpha, valpha); in xnn_f32_velu_ukernel__avx512f_rr1_p6_x48()
91 vs1 = _mm512_fmsub_ps(vs1, valpha, valpha); in xnn_f32_velu_ukernel__avx512f_rr1_p6_x48()
92 vs2 = _mm512_fmsub_ps(vs2, valpha, valpha); in xnn_f32_velu_ukernel__avx512f_rr1_p6_x48()
135 vs = _mm512_fmsub_ps(vs, valpha, valpha); in xnn_f32_velu_ukernel__avx512f_rr1_p6_x48()
169 vs = _mm512_fmsub_ps(vs, valpha, valpha); in xnn_f32_velu_ukernel__avx512f_rr1_p6_x48()
|
| D | velu-avx512f-rr1-lut16-p3-perm-x32.c | 74 vs0 = _mm512_fmsub_ps(vs0, valpha, valpha); in xnn_f32_velu_ukernel__avx512f_rr1_lut16_p3_perm_x32()
75 vs1 = _mm512_fmsub_ps(vs1, valpha, valpha); in xnn_f32_velu_ukernel__avx512f_rr1_lut16_p3_perm_x32()
112 vs = _mm512_fmsub_ps(vs, valpha, valpha); in xnn_f32_velu_ukernel__avx512f_rr1_lut16_p3_perm_x32()
145 vs = _mm512_fmsub_ps(vs, valpha, valpha); in xnn_f32_velu_ukernel__avx512f_rr1_lut16_p3_perm_x32()
|
| D | velu-avx512f-rr1-p6-x32.c | 78 vs0 = _mm512_fmsub_ps(vs0, valpha, valpha); in xnn_f32_velu_ukernel__avx512f_rr1_p6_x32()
79 vs1 = _mm512_fmsub_ps(vs1, valpha, valpha); in xnn_f32_velu_ukernel__avx512f_rr1_p6_x32()
117 vs = _mm512_fmsub_ps(vs, valpha, valpha); in xnn_f32_velu_ukernel__avx512f_rr1_p6_x32()
151 vs = _mm512_fmsub_ps(vs, valpha, valpha); in xnn_f32_velu_ukernel__avx512f_rr1_p6_x32()
|
| D | velu-avx512f-rr1-lut16-p3-perm-x16.c | 61 vs = _mm512_fmsub_ps(vs, valpha, valpha); in xnn_f32_velu_ukernel__avx512f_rr1_lut16_p3_perm_x16()
94 vs = _mm512_fmsub_ps(vs, valpha, valpha); in xnn_f32_velu_ukernel__avx512f_rr1_lut16_p3_perm_x16()
|
| D | velu-avx512f-rr1-p6-x16.c | 62 vs = _mm512_fmsub_ps(vs, valpha, valpha); in xnn_f32_velu_ukernel__avx512f_rr1_p6_x16()
96 vs = _mm512_fmsub_ps(vs, valpha, valpha); in xnn_f32_velu_ukernel__avx512f_rr1_p6_x16()
|
| /external/XNNPACK/src/f32-velu/ |
| D | avx512f-rr1-lut16-p3-perm.c.in | 73 vs${N} = _mm512_fmsub_ps(vs${N}, valpha, valpha);
110 vs = _mm512_fmsub_ps(vs, valpha, valpha);
143 vs = _mm512_fmsub_ps(vs, valpha, valpha);
|
| D | avx512f-rr1-p6.c.in | 78 vs${N} = _mm512_fmsub_ps(vs${N}, valpha, valpha);
116 vs = _mm512_fmsub_ps(vs, valpha, valpha);
150 vs = _mm512_fmsub_ps(vs, valpha, valpha);
|
| /external/clang/test/CodeGen/ |
| D | avx512f-builtins.c | 641 return _mm512_fmsub_ps(__A, __B, __C); in test_mm512_fmsub_ps()
|
| /external/llvm-project/clang/test/CodeGen/X86/ |
| D | avx512f-builtins.c | 784 return _mm512_fmsub_ps(__A, __B, __C); in test_mm512_fmsub_ps()
|
| /external/clang/lib/Headers/ |
| D | avx512fintrin.h | 2889 _mm512_fmsub_ps(__m512 __A, __m512 __B, __m512 __C) in _mm512_fmsub_ps() function
|
| /external/llvm-project/clang/lib/Headers/ |
| D | avx512fintrin.h | 2756 _mm512_fmsub_ps(__m512 __A, __m512 __B, __m512 __C) in _mm512_fmsub_ps() function
|