/external/XNNPACK/src/f32-dwconv2d-chw/gen/ |
D | 5x5p2-minmax-scalar-3x1.c | 248 float vo2 = math_max_f32(vo2p0, vmin); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__scalar_3x1() local 351 float vo2 = math_max_f32(vo2p0, vmin); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__scalar_3x1() local 416 float vo2 = math_max_f32(vo2p0, vmin); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__scalar_3x1() local
|
D | 3x3p1-minmax-scalar-3x1.c | 134 float vo2 = math_max_f32(vo2p0, vmin); in xnn_f32_dwconv2d_chw_ukernel_3x3p1__scalar_3x1() local 170 float vo2 = math_max_f32(vo2p0, vmin); in xnn_f32_dwconv2d_chw_ukernel_3x3p1__scalar_3x1() local
|
D | 5x5p2-minmax-scalar-3x1-acc2.c | 251 float vo2 = math_max_f32(vo2p0, vmin); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__scalar_3x1_acc2() local 357 float vo2 = math_max_f32(vo2p0, vmin); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__scalar_3x1_acc2() local 425 float vo2 = math_max_f32(vo2p0, vmin); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__scalar_3x1_acc2() local
|
D | 3x3p1-minmax-scalar-4x1.c | 154 float vo2 = math_max_f32(vo2p0, vmin); in xnn_f32_dwconv2d_chw_ukernel_3x3p1__scalar_4x1() local 199 float vo2 = math_max_f32(vo2p0, vmin); in xnn_f32_dwconv2d_chw_ukernel_3x3p1__scalar_4x1() local
|
D | 5x5s2p2-minmax-scalar-3x1.c | 281 float vo2 = math_max_f32(vo2p0, vmin); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__scalar_3x1() local 369 float vo2 = math_max_f32(vo2p0, vmin); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__scalar_3x1() local 430 float vo2 = math_max_f32(vo2p0, vmin); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__scalar_3x1() local
|
D | 3x3p1-minmax-wasmsimd-arm-splat-3x4.c | 155 v128_t vo2 = wasm_f32x4_max(vo2p0, vmin); in xnn_f32_dwconv2d_chw_ukernel_3x3p1__wasmsimd_arm_splat_3x4() local 230 v128_t vo2 = wasm_f32x4_max(vo2p0, vmin); in xnn_f32_dwconv2d_chw_ukernel_3x3p1__wasmsimd_arm_splat_3x4() local
|
D | 3x3p1-minmax-wasmsimd-x86-splat-3x4.c | 155 v128_t vo2 = wasm_f32x4_pmax(vmin, vo2p0); in xnn_f32_dwconv2d_chw_ukernel_3x3p1__wasmsimd_x86_splat_3x4() local 230 v128_t vo2 = wasm_f32x4_pmax(vmin, vo2p0); in xnn_f32_dwconv2d_chw_ukernel_3x3p1__wasmsimd_x86_splat_3x4() local
|
D | 3x3s2p1-minmax-scalar-3x1.c | 158 float vo2 = math_max_f32(vo2p0, vmin); in xnn_f32_dwconv2d_chw_ukernel_3x3s2p1__scalar_3x1() local 202 float vo2 = math_max_f32(vo2p0, vmin); in xnn_f32_dwconv2d_chw_ukernel_3x3s2p1__scalar_3x1() local
|
D | 5x5s2p2-minmax-scalar-3x1-acc2.c | 284 float vo2 = math_max_f32(vo2p0, vmin); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__scalar_3x1_acc2() local 375 float vo2 = math_max_f32(vo2p0, vmin); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__scalar_3x1_acc2() local 439 float vo2 = math_max_f32(vo2p0, vmin); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__scalar_3x1_acc2() local
|
D | 3x3p1-minmax-neon-3x4.c | 154 float32x4_t vo2 = vmaxq_f32(vo2p0, vmin); in xnn_f32_dwconv2d_chw_ukernel_3x3p1__neon_3x4() local 230 float32x4_t vo2 = vmaxq_f32(vo2p0, vmin); in xnn_f32_dwconv2d_chw_ukernel_3x3p1__neon_3x4() local
|
D | 3x3p1-minmax-neonfma-3x4.c | 154 float32x4_t vo2 = vmaxq_f32(vo2p0, vmin); in xnn_f32_dwconv2d_chw_ukernel_3x3p1__neonfma_3x4() local 230 float32x4_t vo2 = vmaxq_f32(vo2p0, vmin); in xnn_f32_dwconv2d_chw_ukernel_3x3p1__neonfma_3x4() local
|
D | 3x3p1-minmax-scalar-5x1.c | 174 float vo2 = math_max_f32(vo2p0, vmin); in xnn_f32_dwconv2d_chw_ukernel_3x3p1__scalar_5x1() local 228 float vo2 = math_max_f32(vo2p0, vmin); in xnn_f32_dwconv2d_chw_ukernel_3x3p1__scalar_5x1() local
|
D | 3x3p1-minmax-ssse3-3x4.c | 161 __m128 vo2 = _mm_max_ps(vo2p0, vmin); in xnn_f32_dwconv2d_chw_ukernel_3x3p1__ssse3_3x4() local 230 __m128 vo2 = _mm_max_ps(vo2p0, vmin); in xnn_f32_dwconv2d_chw_ukernel_3x3p1__ssse3_3x4() local
|
D | 3x3p1-minmax-scalar-6x1.c | 194 float vo2 = math_max_f32(vo2p0, vmin); in xnn_f32_dwconv2d_chw_ukernel_3x3p1__scalar_6x1() local 257 float vo2 = math_max_f32(vo2p0, vmin); in xnn_f32_dwconv2d_chw_ukernel_3x3p1__scalar_6x1() local
|
D | 3x3p1-minmax-wasmsimd-x86-loadsplat-3x4.c | 166 v128_t vo2 = wasm_f32x4_pmax(vmin, vo2p0); in xnn_f32_dwconv2d_chw_ukernel_3x3p1__wasmsimd_x86_loadsplat_3x4() local 234 v128_t vo2 = wasm_f32x4_pmax(vmin, vo2p0); in xnn_f32_dwconv2d_chw_ukernel_3x3p1__wasmsimd_x86_loadsplat_3x4() local
|
D | 3x3p1-minmax-wasmsimd-x86-splat-4x4.c | 178 v128_t vo2 = wasm_f32x4_pmax(vmin, vo2p0); in xnn_f32_dwconv2d_chw_ukernel_3x3p1__wasmsimd_x86_splat_4x4() local 269 v128_t vo2 = wasm_f32x4_pmax(vmin, vo2p0); in xnn_f32_dwconv2d_chw_ukernel_3x3p1__wasmsimd_x86_splat_4x4() local
|
D | 3x3p1-minmax-wasmsimd-arm-splat-4x4.c | 178 v128_t vo2 = wasm_f32x4_max(vo2p0, vmin); in xnn_f32_dwconv2d_chw_ukernel_3x3p1__wasmsimd_arm_splat_4x4() local 269 v128_t vo2 = wasm_f32x4_max(vo2p0, vmin); in xnn_f32_dwconv2d_chw_ukernel_3x3p1__wasmsimd_arm_splat_4x4() local
|
D | 3x3p1-minmax-wasmsimd-arm-loadsplat-3x4.c | 166 v128_t vo2 = wasm_f32x4_max(vo2p0, vmin); in xnn_f32_dwconv2d_chw_ukernel_3x3p1__wasmsimd_arm_loadsplat_3x4() local 234 v128_t vo2 = wasm_f32x4_max(vo2p0, vmin); in xnn_f32_dwconv2d_chw_ukernel_3x3p1__wasmsimd_arm_loadsplat_3x4() local
|
D | 3x3p1-minmax-ssse3-4x4.c | 185 __m128 vo2 = _mm_max_ps(vo2p0, vmin); in xnn_f32_dwconv2d_chw_ukernel_3x3p1__ssse3_4x4() local 270 __m128 vo2 = _mm_max_ps(vo2p0, vmin); in xnn_f32_dwconv2d_chw_ukernel_3x3p1__ssse3_4x4() local
|
D | 3x3p1-minmax-neonfma-4x4.c | 177 float32x4_t vo2 = vmaxq_f32(vo2p0, vmin); in xnn_f32_dwconv2d_chw_ukernel_3x3p1__neonfma_4x4() local 269 float32x4_t vo2 = vmaxq_f32(vo2p0, vmin); in xnn_f32_dwconv2d_chw_ukernel_3x3p1__neonfma_4x4() local
|
D | 3x3s2p1-minmax-neonfma-3x4.c | 158 float32x4_t vo2 = vmaxq_f32(vo2p0, vmin); in xnn_f32_dwconv2d_chw_ukernel_3x3s2p1__neonfma_3x4() local 245 float32x4_t vo2 = vmaxq_f32(vo2p0, vmin); in xnn_f32_dwconv2d_chw_ukernel_3x3s2p1__neonfma_3x4() local
|
D | 3x3p1-minmax-neon-4x4.c | 177 float32x4_t vo2 = vmaxq_f32(vo2p0, vmin); in xnn_f32_dwconv2d_chw_ukernel_3x3p1__neon_4x4() local 269 float32x4_t vo2 = vmaxq_f32(vo2p0, vmin); in xnn_f32_dwconv2d_chw_ukernel_3x3p1__neon_4x4() local
|
D | 3x3s2p1-minmax-neon-3x4.c | 158 float32x4_t vo2 = vmaxq_f32(vo2p0, vmin); in xnn_f32_dwconv2d_chw_ukernel_3x3s2p1__neon_3x4() local 245 float32x4_t vo2 = vmaxq_f32(vo2p0, vmin); in xnn_f32_dwconv2d_chw_ukernel_3x3s2p1__neon_3x4() local
|
D | 3x3s2p1-minmax-scalar-4x1.c | 187 float vo2 = math_max_f32(vo2p0, vmin); in xnn_f32_dwconv2d_chw_ukernel_3x3s2p1__scalar_4x1() local 242 float vo2 = math_max_f32(vo2p0, vmin); in xnn_f32_dwconv2d_chw_ukernel_3x3s2p1__scalar_4x1() local
|
D | 5x5p2-minmax-neon-3x4.c | 255 float32x4_t vo2 = vmaxq_f32(vo2p0, vmin); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neon_3x4() local 436 float32x4_t vo2 = vmaxq_f32(vo2p0, vmin); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neon_3x4() local 599 float32x4_t vo2 = vmaxq_f32(vo2p0, vmin); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neon_3x4() local
|