/external/XNNPACK/src/f32-velu/gen/ |
D | velu-wasmsimd-x86-rr2-p6-x4.c | 56 const v128_t vsatm = wasm_f32x4_le(vz, vsat_cutoff); in xnn_f32_velu_ukernel__wasmsimd_x86_rr2_p6_x4() local 58 vs = wasm_v128_andnot(vs, vsatm); in xnn_f32_velu_ukernel__wasmsimd_x86_rr2_p6_x4() 59 vt = wasm_v128_andnot(vt, vsatm); in xnn_f32_velu_ukernel__wasmsimd_x86_rr2_p6_x4() 89 const v128_t vsatm = wasm_f32x4_le(vz, vsat_cutoff); in xnn_f32_velu_ukernel__wasmsimd_x86_rr2_p6_x4() local 91 vs = wasm_v128_andnot(vs, vsatm); in xnn_f32_velu_ukernel__wasmsimd_x86_rr2_p6_x4() 92 vt = wasm_v128_andnot(vt, vsatm); in xnn_f32_velu_ukernel__wasmsimd_x86_rr2_p6_x4()
|
D | velu-wasmsimd-x86-rr2-lut16-p3-x4.c | 67 const v128_t vsatm = wasm_f32x4_le(vz, vsat_cutoff); in xnn_f32_velu_ukernel__wasmsimd_x86_rr2_lut16_p3_x4() local 69 vs = wasm_v128_andnot(vs, vsatm); in xnn_f32_velu_ukernel__wasmsimd_x86_rr2_lut16_p3_x4() 70 vt = wasm_v128_andnot(vt, vsatm); in xnn_f32_velu_ukernel__wasmsimd_x86_rr2_lut16_p3_x4() 108 const v128_t vsatm = wasm_f32x4_le(vz, vsat_cutoff); in xnn_f32_velu_ukernel__wasmsimd_x86_rr2_lut16_p3_x4() local 110 vs = wasm_v128_andnot(vs, vsatm); in xnn_f32_velu_ukernel__wasmsimd_x86_rr2_lut16_p3_x4() 111 vt = wasm_v128_andnot(vt, vsatm); in xnn_f32_velu_ukernel__wasmsimd_x86_rr2_lut16_p3_x4()
|
D | velu-wasmsimd-x86-rr2-p6-x8.c | 123 const v128_t vsatm = wasm_f32x4_le(vz, vsat_cutoff); in xnn_f32_velu_ukernel__wasmsimd_x86_rr2_p6_x8() local 125 vs = wasm_v128_andnot(vs, vsatm); in xnn_f32_velu_ukernel__wasmsimd_x86_rr2_p6_x8() 126 vt = wasm_v128_andnot(vt, vsatm); in xnn_f32_velu_ukernel__wasmsimd_x86_rr2_p6_x8() 156 const v128_t vsatm = wasm_f32x4_le(vz, vsat_cutoff); in xnn_f32_velu_ukernel__wasmsimd_x86_rr2_p6_x8() local 158 vs = wasm_v128_andnot(vs, vsatm); in xnn_f32_velu_ukernel__wasmsimd_x86_rr2_p6_x8() 159 vt = wasm_v128_andnot(vt, vsatm); in xnn_f32_velu_ukernel__wasmsimd_x86_rr2_p6_x8()
|
D | velu-wasmsimd-x86-rr2-lut16-p3-x8.c | 144 const v128_t vsatm = wasm_f32x4_le(vz, vsat_cutoff); in xnn_f32_velu_ukernel__wasmsimd_x86_rr2_lut16_p3_x8() local 146 vs = wasm_v128_andnot(vs, vsatm); in xnn_f32_velu_ukernel__wasmsimd_x86_rr2_lut16_p3_x8() 147 vt = wasm_v128_andnot(vt, vsatm); in xnn_f32_velu_ukernel__wasmsimd_x86_rr2_lut16_p3_x8() 185 const v128_t vsatm = wasm_f32x4_le(vz, vsat_cutoff); in xnn_f32_velu_ukernel__wasmsimd_x86_rr2_lut16_p3_x8() local 187 vs = wasm_v128_andnot(vs, vsatm); in xnn_f32_velu_ukernel__wasmsimd_x86_rr2_lut16_p3_x8() 188 vt = wasm_v128_andnot(vt, vsatm); in xnn_f32_velu_ukernel__wasmsimd_x86_rr2_lut16_p3_x8()
|
D | velu-wasmsimd-x86-rr2-p6-x12.c | 146 const v128_t vsatm = wasm_f32x4_le(vz, vsat_cutoff); in xnn_f32_velu_ukernel__wasmsimd_x86_rr2_p6_x12() local 148 vs = wasm_v128_andnot(vs, vsatm); in xnn_f32_velu_ukernel__wasmsimd_x86_rr2_p6_x12() 149 vt = wasm_v128_andnot(vt, vsatm); in xnn_f32_velu_ukernel__wasmsimd_x86_rr2_p6_x12() 179 const v128_t vsatm = wasm_f32x4_le(vz, vsat_cutoff); in xnn_f32_velu_ukernel__wasmsimd_x86_rr2_p6_x12() local 181 vs = wasm_v128_andnot(vs, vsatm); in xnn_f32_velu_ukernel__wasmsimd_x86_rr2_p6_x12() 182 vt = wasm_v128_andnot(vt, vsatm); in xnn_f32_velu_ukernel__wasmsimd_x86_rr2_p6_x12()
|
D | velu-wasmsimd-x86-rr2-lut16-p3-x12.c | 173 const v128_t vsatm = wasm_f32x4_le(vz, vsat_cutoff); in xnn_f32_velu_ukernel__wasmsimd_x86_rr2_lut16_p3_x12() local 175 vs = wasm_v128_andnot(vs, vsatm); in xnn_f32_velu_ukernel__wasmsimd_x86_rr2_lut16_p3_x12() 176 vt = wasm_v128_andnot(vt, vsatm); in xnn_f32_velu_ukernel__wasmsimd_x86_rr2_lut16_p3_x12() 214 const v128_t vsatm = wasm_f32x4_le(vz, vsat_cutoff); in xnn_f32_velu_ukernel__wasmsimd_x86_rr2_lut16_p3_x12() local 216 vs = wasm_v128_andnot(vs, vsatm); in xnn_f32_velu_ukernel__wasmsimd_x86_rr2_lut16_p3_x12() 217 vt = wasm_v128_andnot(vt, vsatm); in xnn_f32_velu_ukernel__wasmsimd_x86_rr2_lut16_p3_x12()
|
D | velu-wasmsimd-x86-rr2-p6-x16.c | 169 const v128_t vsatm = wasm_f32x4_le(vz, vsat_cutoff); in xnn_f32_velu_ukernel__wasmsimd_x86_rr2_p6_x16() local 171 vs = wasm_v128_andnot(vs, vsatm); in xnn_f32_velu_ukernel__wasmsimd_x86_rr2_p6_x16() 172 vt = wasm_v128_andnot(vt, vsatm); in xnn_f32_velu_ukernel__wasmsimd_x86_rr2_p6_x16() 202 const v128_t vsatm = wasm_f32x4_le(vz, vsat_cutoff); in xnn_f32_velu_ukernel__wasmsimd_x86_rr2_p6_x16() local 204 vs = wasm_v128_andnot(vs, vsatm); in xnn_f32_velu_ukernel__wasmsimd_x86_rr2_p6_x16() 205 vt = wasm_v128_andnot(vt, vsatm); in xnn_f32_velu_ukernel__wasmsimd_x86_rr2_p6_x16()
|
D | velu-wasmsimd-x86-rr2-p6-x20.c | 192 const v128_t vsatm = wasm_f32x4_le(vz, vsat_cutoff); in xnn_f32_velu_ukernel__wasmsimd_x86_rr2_p6_x20() local 194 vs = wasm_v128_andnot(vs, vsatm); in xnn_f32_velu_ukernel__wasmsimd_x86_rr2_p6_x20() 195 vt = wasm_v128_andnot(vt, vsatm); in xnn_f32_velu_ukernel__wasmsimd_x86_rr2_p6_x20() 225 const v128_t vsatm = wasm_f32x4_le(vz, vsat_cutoff); in xnn_f32_velu_ukernel__wasmsimd_x86_rr2_p6_x20() local 227 vs = wasm_v128_andnot(vs, vsatm); in xnn_f32_velu_ukernel__wasmsimd_x86_rr2_p6_x20() 228 vt = wasm_v128_andnot(vt, vsatm); in xnn_f32_velu_ukernel__wasmsimd_x86_rr2_p6_x20()
|
D | velu-wasmsimd-x86-rr2-lut16-p3-x16.c | 202 const v128_t vsatm = wasm_f32x4_le(vz, vsat_cutoff); in xnn_f32_velu_ukernel__wasmsimd_x86_rr2_lut16_p3_x16() local 204 vs = wasm_v128_andnot(vs, vsatm); in xnn_f32_velu_ukernel__wasmsimd_x86_rr2_lut16_p3_x16() 205 vt = wasm_v128_andnot(vt, vsatm); in xnn_f32_velu_ukernel__wasmsimd_x86_rr2_lut16_p3_x16() 243 const v128_t vsatm = wasm_f32x4_le(vz, vsat_cutoff); in xnn_f32_velu_ukernel__wasmsimd_x86_rr2_lut16_p3_x16() local 245 vs = wasm_v128_andnot(vs, vsatm); in xnn_f32_velu_ukernel__wasmsimd_x86_rr2_lut16_p3_x16() 246 vt = wasm_v128_andnot(vt, vsatm); in xnn_f32_velu_ukernel__wasmsimd_x86_rr2_lut16_p3_x16()
|
D | velu-wasmsimd-x86-rr2-p6-x24.c | 215 const v128_t vsatm = wasm_f32x4_le(vz, vsat_cutoff); in xnn_f32_velu_ukernel__wasmsimd_x86_rr2_p6_x24() local 217 vs = wasm_v128_andnot(vs, vsatm); in xnn_f32_velu_ukernel__wasmsimd_x86_rr2_p6_x24() 218 vt = wasm_v128_andnot(vt, vsatm); in xnn_f32_velu_ukernel__wasmsimd_x86_rr2_p6_x24() 248 const v128_t vsatm = wasm_f32x4_le(vz, vsat_cutoff); in xnn_f32_velu_ukernel__wasmsimd_x86_rr2_p6_x24() local 250 vs = wasm_v128_andnot(vs, vsatm); in xnn_f32_velu_ukernel__wasmsimd_x86_rr2_p6_x24() 251 vt = wasm_v128_andnot(vt, vsatm); in xnn_f32_velu_ukernel__wasmsimd_x86_rr2_p6_x24()
|
D | velu-wasmsimd-x86-rr2-lut16-p3-x20.c | 231 const v128_t vsatm = wasm_f32x4_le(vz, vsat_cutoff); in xnn_f32_velu_ukernel__wasmsimd_x86_rr2_lut16_p3_x20() local 233 vs = wasm_v128_andnot(vs, vsatm); in xnn_f32_velu_ukernel__wasmsimd_x86_rr2_lut16_p3_x20() 234 vt = wasm_v128_andnot(vt, vsatm); in xnn_f32_velu_ukernel__wasmsimd_x86_rr2_lut16_p3_x20() 272 const v128_t vsatm = wasm_f32x4_le(vz, vsat_cutoff); in xnn_f32_velu_ukernel__wasmsimd_x86_rr2_lut16_p3_x20() local 274 vs = wasm_v128_andnot(vs, vsatm); in xnn_f32_velu_ukernel__wasmsimd_x86_rr2_lut16_p3_x20() 275 vt = wasm_v128_andnot(vt, vsatm); in xnn_f32_velu_ukernel__wasmsimd_x86_rr2_lut16_p3_x20()
|
D | velu-wasmsimd-x86-rr2-lut16-p3-x24.c | 260 const v128_t vsatm = wasm_f32x4_le(vz, vsat_cutoff); in xnn_f32_velu_ukernel__wasmsimd_x86_rr2_lut16_p3_x24() local 262 vs = wasm_v128_andnot(vs, vsatm); in xnn_f32_velu_ukernel__wasmsimd_x86_rr2_lut16_p3_x24() 263 vt = wasm_v128_andnot(vt, vsatm); in xnn_f32_velu_ukernel__wasmsimd_x86_rr2_lut16_p3_x24() 301 const v128_t vsatm = wasm_f32x4_le(vz, vsat_cutoff); in xnn_f32_velu_ukernel__wasmsimd_x86_rr2_lut16_p3_x24() local 303 vs = wasm_v128_andnot(vs, vsatm); in xnn_f32_velu_ukernel__wasmsimd_x86_rr2_lut16_p3_x24() 304 vt = wasm_v128_andnot(vt, vsatm); in xnn_f32_velu_ukernel__wasmsimd_x86_rr2_lut16_p3_x24()
|
/external/XNNPACK/src/f32-velu/ |
D | wasmsimd-rr2-p6.c.in | 69 const v128_t vsatm${ABC[N:N+4]} = wasm_f32x4_le(vz${ABC[N:N+4]}, vsat_cutoff); 74 vs${ABC[N:N+4]} = wasm_v128_andnot(vs${ABC[N:N+4]}, vsatm${ABC[N:N+4]}); 78 vt${ABC[N:N+4]} = wasm_v128_andnot(vt${ABC[N:N+4]}, vsatm${ABC[N:N+4]}); 130 const v128_t vsatm = wasm_f32x4_le(vz, vsat_cutoff); 133 vs = wasm_v128_andnot(vs, vsatm); 134 vt = wasm_v128_andnot(vt, vsatm); 168 const v128_t vsatm = wasm_f32x4_le(vz, vsat_cutoff); 171 vs = wasm_v128_andnot(vs, vsatm); 172 vt = wasm_v128_andnot(vt, vsatm);
|
D | wasmsimd-rr2-lut16-p3.c.in | 80 const v128_t vsatm${ABC[N:N+4]} = wasm_f32x4_le(vz${ABC[N:N+4]}, vsat_cutoff); 85 vs${ABC[N:N+4]} = wasm_v128_andnot(vs${ABC[N:N+4]}, vsatm${ABC[N:N+4]}); 89 vt${ABC[N:N+4]} = wasm_v128_andnot(vt${ABC[N:N+4]}, vsatm${ABC[N:N+4]}); 143 const v128_t vsatm = wasm_f32x4_le(vz, vsat_cutoff); 146 vs = wasm_v128_andnot(vs, vsatm); 147 vt = wasm_v128_andnot(vt, vsatm); 189 const v128_t vsatm = wasm_f32x4_le(vz, vsat_cutoff); 192 vs = wasm_v128_andnot(vs, vsatm); 193 vt = wasm_v128_andnot(vt, vsatm);
|