/external/XNNPACK/src/f32-dwconv2d-chw/gen/ |
D | 3x3p1-minmax-scalar-4x1.c | 155 float vo3 = math_max_f32(vo3p0, vmin); in xnn_f32_dwconv2d_chw_ukernel_3x3p1__scalar_4x1() local 200 float vo3 = math_max_f32(vo3p0, vmin); in xnn_f32_dwconv2d_chw_ukernel_3x3p1__scalar_4x1() local
|
D | 3x3p1-minmax-scalar-5x1.c | 175 float vo3 = math_max_f32(vo3p0, vmin); in xnn_f32_dwconv2d_chw_ukernel_3x3p1__scalar_5x1() local 229 float vo3 = math_max_f32(vo3p0, vmin); in xnn_f32_dwconv2d_chw_ukernel_3x3p1__scalar_5x1() local
|
D | 3x3p1-minmax-scalar-6x1.c | 195 float vo3 = math_max_f32(vo3p0, vmin); in xnn_f32_dwconv2d_chw_ukernel_3x3p1__scalar_6x1() local 258 float vo3 = math_max_f32(vo3p0, vmin); in xnn_f32_dwconv2d_chw_ukernel_3x3p1__scalar_6x1() local
|
D | 3x3p1-minmax-wasmsimd-x86-splat-4x4.c | 179 v128_t vo3 = wasm_f32x4_pmax(vmin, vo3p0); in xnn_f32_dwconv2d_chw_ukernel_3x3p1__wasmsimd_x86_splat_4x4() local 270 v128_t vo3 = wasm_f32x4_pmax(vmin, vo3p0); in xnn_f32_dwconv2d_chw_ukernel_3x3p1__wasmsimd_x86_splat_4x4() local
|
D | 3x3p1-minmax-wasmsimd-arm-splat-4x4.c | 179 v128_t vo3 = wasm_f32x4_max(vo3p0, vmin); in xnn_f32_dwconv2d_chw_ukernel_3x3p1__wasmsimd_arm_splat_4x4() local 270 v128_t vo3 = wasm_f32x4_max(vo3p0, vmin); in xnn_f32_dwconv2d_chw_ukernel_3x3p1__wasmsimd_arm_splat_4x4() local
|
D | 3x3p1-minmax-ssse3-4x4.c | 186 __m128 vo3 = _mm_max_ps(vo3p0, vmin); in xnn_f32_dwconv2d_chw_ukernel_3x3p1__ssse3_4x4() local 271 __m128 vo3 = _mm_max_ps(vo3p0, vmin); in xnn_f32_dwconv2d_chw_ukernel_3x3p1__ssse3_4x4() local
|
D | 3x3p1-minmax-neonfma-4x4.c | 178 float32x4_t vo3 = vmaxq_f32(vo3p0, vmin); in xnn_f32_dwconv2d_chw_ukernel_3x3p1__neonfma_4x4() local 270 float32x4_t vo3 = vmaxq_f32(vo3p0, vmin); in xnn_f32_dwconv2d_chw_ukernel_3x3p1__neonfma_4x4() local
|
D | 3x3p1-minmax-neon-4x4.c | 178 float32x4_t vo3 = vmaxq_f32(vo3p0, vmin); in xnn_f32_dwconv2d_chw_ukernel_3x3p1__neon_4x4() local 270 float32x4_t vo3 = vmaxq_f32(vo3p0, vmin); in xnn_f32_dwconv2d_chw_ukernel_3x3p1__neon_4x4() local
|
D | 3x3s2p1-minmax-scalar-4x1.c | 188 float vo3 = math_max_f32(vo3p0, vmin); in xnn_f32_dwconv2d_chw_ukernel_3x3s2p1__scalar_4x1() local 243 float vo3 = math_max_f32(vo3p0, vmin); in xnn_f32_dwconv2d_chw_ukernel_3x3s2p1__scalar_4x1() local
|
D | 3x3p1-minmax-wasmsimd-arm-splat-5x4.c | 202 v128_t vo3 = wasm_f32x4_max(vo3p0, vmin); in xnn_f32_dwconv2d_chw_ukernel_3x3p1__wasmsimd_arm_splat_5x4() local 309 v128_t vo3 = wasm_f32x4_max(vo3p0, vmin); in xnn_f32_dwconv2d_chw_ukernel_3x3p1__wasmsimd_arm_splat_5x4() local
|
D | 3x3p1-minmax-wasmsimd-x86-loadsplat-4x4.c | 191 v128_t vo3 = wasm_f32x4_pmax(vmin, vo3p0); in xnn_f32_dwconv2d_chw_ukernel_3x3p1__wasmsimd_x86_loadsplat_4x4() local 275 v128_t vo3 = wasm_f32x4_pmax(vmin, vo3p0); in xnn_f32_dwconv2d_chw_ukernel_3x3p1__wasmsimd_x86_loadsplat_4x4() local
|
D | 3x3p1-minmax-wasmsimd-arm-loadsplat-4x4.c | 191 v128_t vo3 = wasm_f32x4_max(vo3p0, vmin); in xnn_f32_dwconv2d_chw_ukernel_3x3p1__wasmsimd_arm_loadsplat_4x4() local 275 v128_t vo3 = wasm_f32x4_max(vo3p0, vmin); in xnn_f32_dwconv2d_chw_ukernel_3x3p1__wasmsimd_arm_loadsplat_4x4() local
|
D | 3x3p1-minmax-wasmsimd-x86-splat-5x4.c | 202 v128_t vo3 = wasm_f32x4_pmax(vmin, vo3p0); in xnn_f32_dwconv2d_chw_ukernel_3x3p1__wasmsimd_x86_splat_5x4() local 309 v128_t vo3 = wasm_f32x4_pmax(vmin, vo3p0); in xnn_f32_dwconv2d_chw_ukernel_3x3p1__wasmsimd_x86_splat_5x4() local
|
D | 3x3p1-minmax-neonfma-5x4.c | 201 float32x4_t vo3 = vmaxq_f32(vo3p0, vmin); in xnn_f32_dwconv2d_chw_ukernel_3x3p1__neonfma_5x4() local 309 float32x4_t vo3 = vmaxq_f32(vo3p0, vmin); in xnn_f32_dwconv2d_chw_ukernel_3x3p1__neonfma_5x4() local
|
D | 3x3p1-minmax-ssse3-5x4.c | 210 __m128 vo3 = _mm_max_ps(vo3p0, vmin); in xnn_f32_dwconv2d_chw_ukernel_3x3p1__ssse3_5x4() local 311 __m128 vo3 = _mm_max_ps(vo3p0, vmin); in xnn_f32_dwconv2d_chw_ukernel_3x3p1__ssse3_5x4() local
|
D | 3x3p1-minmax-wasmsimd-arm-loadsplat-5x4.c | 215 v128_t vo3 = wasm_f32x4_max(vo3p0, vmin); in xnn_f32_dwconv2d_chw_ukernel_3x3p1__wasmsimd_arm_loadsplat_5x4() local 315 v128_t vo3 = wasm_f32x4_max(vo3p0, vmin); in xnn_f32_dwconv2d_chw_ukernel_3x3p1__wasmsimd_arm_loadsplat_5x4() local
|
D | 3x3p1-minmax-wasmsimd-x86-loadsplat-5x4.c | 215 v128_t vo3 = wasm_f32x4_pmax(vmin, vo3p0); in xnn_f32_dwconv2d_chw_ukernel_3x3p1__wasmsimd_x86_loadsplat_5x4() local 315 v128_t vo3 = wasm_f32x4_pmax(vmin, vo3p0); in xnn_f32_dwconv2d_chw_ukernel_3x3p1__wasmsimd_x86_loadsplat_5x4() local
|
D | 3x3p1-minmax-neon-5x4.c | 201 float32x4_t vo3 = vmaxq_f32(vo3p0, vmin); in xnn_f32_dwconv2d_chw_ukernel_3x3p1__neon_5x4() local 309 float32x4_t vo3 = vmaxq_f32(vo3p0, vmin); in xnn_f32_dwconv2d_chw_ukernel_3x3p1__neon_5x4() local
|
D | 5x5p2-minmax-neonfma-4x4.c | 297 float32x4_t vo3 = vmaxq_f32(vo3p0, vmin); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neonfma_4x4() local 515 float32x4_t vo3 = vmaxq_f32(vo3p0, vmin); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neonfma_4x4() local 712 float32x4_t vo3 = vmaxq_f32(vo3p0, vmin); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neonfma_4x4() local
|
D | 5x5p2-minmax-wasmsimd-x86-splat-4x4.c | 298 v128_t vo3 = wasm_f32x4_pmax(vmin, vo3p0); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_x86_splat_4x4() local 515 v128_t vo3 = wasm_f32x4_pmax(vmin, vo3p0); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_x86_splat_4x4() local 711 v128_t vo3 = wasm_f32x4_pmax(vmin, vo3p0); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_x86_splat_4x4() local
|
D | 5x5p2-minmax-wasmsimd-arm-splat-4x4.c | 298 v128_t vo3 = wasm_f32x4_max(vo3p0, vmin); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_arm_splat_4x4() local 515 v128_t vo3 = wasm_f32x4_max(vo3p0, vmin); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_arm_splat_4x4() local 711 v128_t vo3 = wasm_f32x4_max(vo3p0, vmin); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_arm_splat_4x4() local
|
D | 5x5p2-minmax-wasmsimd-x86-splat-4x4-acc2.c | 302 v128_t vo3 = wasm_f32x4_pmax(vmin, vo3p0); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_x86_splat_4x4_acc2() local 523 v128_t vo3 = wasm_f32x4_pmax(vmin, vo3p0); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_x86_splat_4x4_acc2() local 723 v128_t vo3 = wasm_f32x4_pmax(vmin, vo3p0); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_x86_splat_4x4_acc2() local
|
D | 5x5p2-minmax-neon-4x4.c | 297 float32x4_t vo3 = vmaxq_f32(vo3p0, vmin); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neon_4x4() local 515 float32x4_t vo3 = vmaxq_f32(vo3p0, vmin); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neon_4x4() local 712 float32x4_t vo3 = vmaxq_f32(vo3p0, vmin); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neon_4x4() local
|
D | 5x5p2-minmax-wasmsimd-arm-splat-4x4-acc2.c | 302 v128_t vo3 = wasm_f32x4_max(vo3p0, vmin); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_arm_splat_4x4_acc2() local 523 v128_t vo3 = wasm_f32x4_max(vo3p0, vmin); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_arm_splat_4x4_acc2() local 723 v128_t vo3 = wasm_f32x4_max(vo3p0, vmin); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_arm_splat_4x4_acc2() local
|
D | 3x3s2p1-minmax-neonfma-4x4.c | 187 float32x4_t vo3 = vmaxq_f32(vo3p0, vmin); in xnn_f32_dwconv2d_chw_ukernel_3x3s2p1__neonfma_4x4() local 295 float32x4_t vo3 = vmaxq_f32(vo3p0, vmin); in xnn_f32_dwconv2d_chw_ukernel_3x3s2p1__neonfma_4x4() local
|