/external/XNNPACK/src/f32-dwconv2d-chw/gen/ |
D | 3x3p1-minmax-scalar-5x1.c | 176 float vo4 = math_max_f32(vo4p0, vmin); in xnn_f32_dwconv2d_chw_ukernel_3x3p1__scalar_5x1() local 230 float vo4 = math_max_f32(vo4p0, vmin); in xnn_f32_dwconv2d_chw_ukernel_3x3p1__scalar_5x1() local
|
D | 3x3p1-minmax-scalar-6x1.c | 196 float vo4 = math_max_f32(vo4p0, vmin); in xnn_f32_dwconv2d_chw_ukernel_3x3p1__scalar_6x1() local 259 float vo4 = math_max_f32(vo4p0, vmin); in xnn_f32_dwconv2d_chw_ukernel_3x3p1__scalar_6x1() local
|
D | 3x3p1-minmax-wasmsimd-arm-splat-5x4.c | 205 v128_t vo4 = wasm_f32x4_max(vo4p0, vmin); in xnn_f32_dwconv2d_chw_ukernel_3x3p1__wasmsimd_arm_splat_5x4() local 311 v128_t vo4 = wasm_f32x4_max(vo4p0, vmin); in xnn_f32_dwconv2d_chw_ukernel_3x3p1__wasmsimd_arm_splat_5x4() local
|
D | 3x3p1-minmax-wasmsimd-x86-splat-5x4.c | 205 v128_t vo4 = wasm_v128_bitselect(vmin, vo4p0, wasm_f32x4_lt(vo4p0, vmin)); in xnn_f32_dwconv2d_chw_ukernel_3x3p1__wasmsimd_x86_splat_5x4() local 311 v128_t vo4 = wasm_v128_bitselect(vmin, vo4p0, wasm_f32x4_lt(vo4p0, vmin)); in xnn_f32_dwconv2d_chw_ukernel_3x3p1__wasmsimd_x86_splat_5x4() local
|
D | 3x3p1-minmax-wasmsimd-x86-loadsplat-5x4.c | 218 v128_t vo4 = wasm_v128_bitselect(vmin, vo4p0, wasm_f32x4_lt(vo4p0, vmin)); in xnn_f32_dwconv2d_chw_ukernel_3x3p1__wasmsimd_x86_loadsplat_5x4() local 317 v128_t vo4 = wasm_v128_bitselect(vmin, vo4p0, wasm_f32x4_lt(vo4p0, vmin)); in xnn_f32_dwconv2d_chw_ukernel_3x3p1__wasmsimd_x86_loadsplat_5x4() local
|
D | 3x3p1-minmax-ssse3-5x4.c | 211 __m128 vo4 = _mm_max_ps(vo4p0, vmin); in xnn_f32_dwconv2d_chw_ukernel_3x3p1__ssse3_5x4() local 312 __m128 vo4 = _mm_max_ps(vo4p0, vmin); in xnn_f32_dwconv2d_chw_ukernel_3x3p1__ssse3_5x4() local
|
D | 3x3p1-minmax-wasmsimd-arm-loadsplat-5x4.c | 218 v128_t vo4 = wasm_f32x4_max(vo4p0, vmin); in xnn_f32_dwconv2d_chw_ukernel_3x3p1__wasmsimd_arm_loadsplat_5x4() local 317 v128_t vo4 = wasm_f32x4_max(vo4p0, vmin); in xnn_f32_dwconv2d_chw_ukernel_3x3p1__wasmsimd_arm_loadsplat_5x4() local
|
D | 3x3p1-minmax-neon-5x4.c | 202 float32x4_t vo4 = vmaxq_f32(vo4p0, vmin); in xnn_f32_dwconv2d_chw_ukernel_3x3p1__neon_5x4() local 310 float32x4_t vo4 = vmaxq_f32(vo4p0, vmin); in xnn_f32_dwconv2d_chw_ukernel_3x3p1__neon_5x4() local
|
D | 3x3p1-minmax-neonfma-5x4.c | 202 float32x4_t vo4 = vmaxq_f32(vo4p0, vmin); in xnn_f32_dwconv2d_chw_ukernel_3x3p1__neonfma_5x4() local 310 float32x4_t vo4 = vmaxq_f32(vo4p0, vmin); in xnn_f32_dwconv2d_chw_ukernel_3x3p1__neonfma_5x4() local
|
D | 3x3p1-minmax-wasmsimd-arm-splat-6x4.c | 228 v128_t vo4 = wasm_f32x4_max(vo4p0, vmin); in xnn_f32_dwconv2d_chw_ukernel_3x3p1__wasmsimd_arm_splat_6x4() local 350 v128_t vo4 = wasm_f32x4_max(vo4p0, vmin); in xnn_f32_dwconv2d_chw_ukernel_3x3p1__wasmsimd_arm_splat_6x4() local
|
D | 3x3p1-minmax-neon-6x4.c | 225 float32x4_t vo4 = vmaxq_f32(vo4p0, vmin); in xnn_f32_dwconv2d_chw_ukernel_3x3p1__neon_6x4() local 349 float32x4_t vo4 = vmaxq_f32(vo4p0, vmin); in xnn_f32_dwconv2d_chw_ukernel_3x3p1__neon_6x4() local
|
D | 3x3p1-minmax-neonfma-6x4.c | 225 float32x4_t vo4 = vmaxq_f32(vo4p0, vmin); in xnn_f32_dwconv2d_chw_ukernel_3x3p1__neonfma_6x4() local 349 float32x4_t vo4 = vmaxq_f32(vo4p0, vmin); in xnn_f32_dwconv2d_chw_ukernel_3x3p1__neonfma_6x4() local
|
D | 3x3p1-minmax-wasmsimd-x86-splat-6x4.c | 228 v128_t vo4 = wasm_v128_bitselect(vmin, vo4p0, wasm_f32x4_lt(vo4p0, vmin)); in xnn_f32_dwconv2d_chw_ukernel_3x3p1__wasmsimd_x86_splat_6x4() local 350 v128_t vo4 = wasm_v128_bitselect(vmin, vo4p0, wasm_f32x4_lt(vo4p0, vmin)); in xnn_f32_dwconv2d_chw_ukernel_3x3p1__wasmsimd_x86_splat_6x4() local
|
D | 3x3p1-minmax-wasmsimd-x86-loadsplat-6x4.c | 242 v128_t vo4 = wasm_v128_bitselect(vmin, vo4p0, wasm_f32x4_lt(vo4p0, vmin)); in xnn_f32_dwconv2d_chw_ukernel_3x3p1__wasmsimd_x86_loadsplat_6x4() local 357 v128_t vo4 = wasm_v128_bitselect(vmin, vo4p0, wasm_f32x4_lt(vo4p0, vmin)); in xnn_f32_dwconv2d_chw_ukernel_3x3p1__wasmsimd_x86_loadsplat_6x4() local
|
D | 3x3p1-minmax-wasmsimd-arm-loadsplat-6x4.c | 242 v128_t vo4 = wasm_f32x4_max(vo4p0, vmin); in xnn_f32_dwconv2d_chw_ukernel_3x3p1__wasmsimd_arm_loadsplat_6x4() local 357 v128_t vo4 = wasm_f32x4_max(vo4p0, vmin); in xnn_f32_dwconv2d_chw_ukernel_3x3p1__wasmsimd_arm_loadsplat_6x4() local
|
D | 3x3p1-minmax-ssse3-6x4.c | 235 __m128 vo4 = _mm_max_ps(vo4p0, vmin); in xnn_f32_dwconv2d_chw_ukernel_3x3p1__ssse3_6x4() local 352 __m128 vo4 = _mm_max_ps(vo4p0, vmin); in xnn_f32_dwconv2d_chw_ukernel_3x3p1__ssse3_6x4() local
|
D | 5x5p2-minmax-wasmsimd-x86-splat-5x4.c | 342 v128_t vo4 = wasm_v128_bitselect(vmin, vo4p0, wasm_f32x4_lt(vo4p0, vmin)); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_x86_splat_5x4() local 596 v128_t vo4 = wasm_v128_bitselect(vmin, vo4p0, wasm_f32x4_lt(vo4p0, vmin)); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_x86_splat_5x4() local 825 v128_t vo4 = wasm_v128_bitselect(vmin, vo4p0, wasm_f32x4_lt(vo4p0, vmin)); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_x86_splat_5x4() local
|
D | 5x5p2-minmax-wasmsimd-arm-splat-5x4.c | 342 v128_t vo4 = wasm_f32x4_max(vo4p0, vmin); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_arm_splat_5x4() local 596 v128_t vo4 = wasm_f32x4_max(vo4p0, vmin); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_arm_splat_5x4() local 825 v128_t vo4 = wasm_f32x4_max(vo4p0, vmin); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_arm_splat_5x4() local
|
D | 3x3p1-minmax-sse-5x4.c | 269 __m128 vo4 = _mm_max_ps(vo4p0, vmin); in xnn_f32_dwconv2d_chw_ukernel_3x3p1__sse_5x4() local 414 __m128 vo4 = _mm_max_ps(vo4p0, vmin); in xnn_f32_dwconv2d_chw_ukernel_3x3p1__sse_5x4() local
|
D | 5x5p2-minmax-neonfma-5x4.c | 339 float32x4_t vo4 = vmaxq_f32(vo4p0, vmin); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neonfma_5x4() local 594 float32x4_t vo4 = vmaxq_f32(vo4p0, vmin); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neonfma_5x4() local 825 float32x4_t vo4 = vmaxq_f32(vo4p0, vmin); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neonfma_5x4() local
|
D | 5x5p2-minmax-neon-5x4.c | 339 float32x4_t vo4 = vmaxq_f32(vo4p0, vmin); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neon_5x4() local 594 float32x4_t vo4 = vmaxq_f32(vo4p0, vmin); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neon_5x4() local 825 float32x4_t vo4 = vmaxq_f32(vo4p0, vmin); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neon_5x4() local
|
D | 5x5p2-minmax-wasmsimd-x86-loadsplat-5x4.c | 368 v128_t vo4 = wasm_v128_bitselect(vmin, vo4p0, wasm_f32x4_lt(vo4p0, vmin)); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_x86_loadsplat_5x4() local 622 v128_t vo4 = wasm_v128_bitselect(vmin, vo4p0, wasm_f32x4_lt(vo4p0, vmin)); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_x86_loadsplat_5x4() local 851 v128_t vo4 = wasm_v128_bitselect(vmin, vo4p0, wasm_f32x4_lt(vo4p0, vmin)); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_x86_loadsplat_5x4() local
|
D | 5x5p2-minmax-wasmsimd-arm-loadsplat-5x4.c | 368 v128_t vo4 = wasm_f32x4_max(vo4p0, vmin); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_arm_loadsplat_5x4() local 622 v128_t vo4 = wasm_f32x4_max(vo4p0, vmin); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_arm_loadsplat_5x4() local 851 v128_t vo4 = wasm_f32x4_max(vo4p0, vmin); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_arm_loadsplat_5x4() local
|
D | 3x3p1-minmax-sse-6x4.c | 301 __m128 vo4 = _mm_max_ps(vo4p0, vmin); in xnn_f32_dwconv2d_chw_ukernel_3x3p1__sse_6x4() local 468 __m128 vo4 = _mm_max_ps(vo4p0, vmin); in xnn_f32_dwconv2d_chw_ukernel_3x3p1__sse_6x4() local
|
D | 5x5p2-minmax-sse-5x4.c | 370 __m128 vo4 = _mm_max_ps(vo4p0, vmin); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__sse_5x4() local 623 __m128 vo4 = _mm_max_ps(vo4p0, vmin); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__sse_5x4() local 853 __m128 vo4 = _mm_max_ps(vo4p0, vmin); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__sse_5x4() local
|