/aosp_15_r20/external/XNNPACK/src/f32-dwconv2d-chw/gen/ |
H A 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
|
H A D | 5x5p2-minmax-wasmsimd-x86-splat-5x4.c | 340 v128_t vo4 = wasm_f32x4_pmax(vmin, vo4p0); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_x86_splat_5x4() local 594 v128_t vo4 = wasm_f32x4_pmax(vmin, vo4p0); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_x86_splat_5x4() local 824 v128_t vo4 = wasm_f32x4_pmax(vmin, vo4p0); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_x86_splat_5x4() local
|
H A D | 5x5p2-minmax-wasmsimd-arm-splat-5x4.c | 340 v128_t vo4 = wasm_f32x4_max(vo4p0, vmin); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_arm_splat_5x4() local 594 v128_t vo4 = wasm_f32x4_max(vo4p0, vmin); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_arm_splat_5x4() local 824 v128_t vo4 = wasm_f32x4_max(vo4p0, vmin); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_arm_splat_5x4() local
|
H A 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
|
H A 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
|
H A 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
|
H A D | 3x3p1-minmax-wasmsimd-arm-splat-5x4.c | 203 v128_t vo4 = wasm_f32x4_max(vo4p0, vmin); in xnn_f32_dwconv2d_chw_ukernel_3x3p1__wasmsimd_arm_splat_5x4() local 310 v128_t vo4 = wasm_f32x4_max(vo4p0, vmin); in xnn_f32_dwconv2d_chw_ukernel_3x3p1__wasmsimd_arm_splat_5x4() local
|
H A D | 3x3p1-minmax-wasmsimd-x86-splat-5x4.c | 203 v128_t vo4 = wasm_f32x4_pmax(vmin, vo4p0); in xnn_f32_dwconv2d_chw_ukernel_3x3p1__wasmsimd_x86_splat_5x4() local 310 v128_t vo4 = wasm_f32x4_pmax(vmin, vo4p0); in xnn_f32_dwconv2d_chw_ukernel_3x3p1__wasmsimd_x86_splat_5x4() local
|
H A D | 5x5p2-minmax-wasmsimd-arm-loadsplat-5x4.c | 366 v128_t vo4 = wasm_f32x4_max(vo4p0, vmin); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_arm_loadsplat_5x4() local 620 v128_t vo4 = wasm_f32x4_max(vo4p0, vmin); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_arm_loadsplat_5x4() local 850 v128_t vo4 = wasm_f32x4_max(vo4p0, vmin); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_arm_loadsplat_5x4() local
|
H A D | 5x5p2-minmax-wasmsimd-x86-loadsplat-5x4.c | 366 v128_t vo4 = wasm_f32x4_pmax(vmin, vo4p0); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_x86_loadsplat_5x4() local 620 v128_t vo4 = wasm_f32x4_pmax(vmin, vo4p0); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_x86_loadsplat_5x4() local 850 v128_t vo4 = wasm_f32x4_pmax(vmin, vo4p0); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_x86_loadsplat_5x4() local
|
H A D | 3x3p1-minmax-wasmsimd-x86-loadsplat-5x4.c | 216 v128_t vo4 = wasm_f32x4_pmax(vmin, vo4p0); in xnn_f32_dwconv2d_chw_ukernel_3x3p1__wasmsimd_x86_loadsplat_5x4() local 316 v128_t vo4 = wasm_f32x4_pmax(vmin, vo4p0); in xnn_f32_dwconv2d_chw_ukernel_3x3p1__wasmsimd_x86_loadsplat_5x4() local
|
H A 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
|
H A 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
|
H A D | 3x3p1-minmax-wasmsimd-arm-loadsplat-5x4.c | 216 v128_t vo4 = wasm_f32x4_max(vo4p0, vmin); in xnn_f32_dwconv2d_chw_ukernel_3x3p1__wasmsimd_arm_loadsplat_5x4() local 316 v128_t vo4 = wasm_f32x4_max(vo4p0, vmin); in xnn_f32_dwconv2d_chw_ukernel_3x3p1__wasmsimd_arm_loadsplat_5x4() local
|
H A 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
|
H A D | 3x3p1-minmax-wasmsimd-x86-splat-6x4.c | 226 v128_t vo4 = wasm_f32x4_pmax(vmin, vo4p0); in xnn_f32_dwconv2d_chw_ukernel_3x3p1__wasmsimd_x86_splat_6x4() local 349 v128_t vo4 = wasm_f32x4_pmax(vmin, vo4p0); in xnn_f32_dwconv2d_chw_ukernel_3x3p1__wasmsimd_x86_splat_6x4() local
|
H A 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
|
H A D | 3x3p1-minmax-wasmsimd-arm-splat-6x4.c | 226 v128_t vo4 = wasm_f32x4_max(vo4p0, vmin); in xnn_f32_dwconv2d_chw_ukernel_3x3p1__wasmsimd_arm_splat_6x4() local 349 v128_t vo4 = wasm_f32x4_max(vo4p0, vmin); in xnn_f32_dwconv2d_chw_ukernel_3x3p1__wasmsimd_arm_splat_6x4() local
|
H A 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
|
H A 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
|
H A D | 3x3p1-minmax-wasmsimd-x86-loadsplat-6x4.c | 240 v128_t vo4 = wasm_f32x4_pmax(vmin, vo4p0); in xnn_f32_dwconv2d_chw_ukernel_3x3p1__wasmsimd_x86_loadsplat_6x4() local 356 v128_t vo4 = wasm_f32x4_pmax(vmin, vo4p0); in xnn_f32_dwconv2d_chw_ukernel_3x3p1__wasmsimd_x86_loadsplat_6x4() local
|
H A D | 3x3p1-minmax-wasmsimd-arm-loadsplat-6x4.c | 240 v128_t vo4 = wasm_f32x4_max(vo4p0, vmin); in xnn_f32_dwconv2d_chw_ukernel_3x3p1__wasmsimd_arm_loadsplat_6x4() local 356 v128_t vo4 = wasm_f32x4_max(vo4p0, vmin); in xnn_f32_dwconv2d_chw_ukernel_3x3p1__wasmsimd_arm_loadsplat_6x4() local
|
/aosp_15_r20/external/XNNPACK/src/f16-dwconv2d-chw/gen/ |
H A D | 5x5p2-minmax-neonfp16arith-5x4.c | 337 float16x4_t vo4 = vmax_f16(vo4p0, vmin); in xnn_f16_dwconv2d_chw_ukernel_5x5p2__neonfp16arith_5x4() local 592 float16x4_t vo4 = vmax_f16(vo4p0, vmin); in xnn_f16_dwconv2d_chw_ukernel_5x5p2__neonfp16arith_5x4() local 823 float16x4_t vo4 = vmax_f16(vo4p0, vmin); in xnn_f16_dwconv2d_chw_ukernel_5x5p2__neonfp16arith_5x4() local
|
H A D | 3x3p1-minmax-neonfp16arith-5x8.c | 205 float16x8_t vo4 = vmaxq_f16(vo4p0, vmin); in xnn_f16_dwconv2d_chw_ukernel_3x3p1__neonfp16arith_5x8() local 316 float16x8_t vo4 = vmaxq_f16(vo4p0, vmin); in xnn_f16_dwconv2d_chw_ukernel_3x3p1__neonfp16arith_5x8() local
|
H A D | 3x3p1-minmax-neonfp16arith-6x8.c | 228 float16x8_t vo4 = vmaxq_f16(vo4p0, vmin); in xnn_f16_dwconv2d_chw_ukernel_3x3p1__neonfp16arith_6x8() local 355 float16x8_t vo4 = vmaxq_f16(vo4p0, vmin); in xnn_f16_dwconv2d_chw_ukernel_3x3p1__neonfp16arith_6x8() local
|