| /external/XNNPACK/src/f32-dwconv2d-chw/gen/ |
| D | 3x3p1-minmax-scalar-6x1.c | 121 float vo5p0 = vbias + vi5x0 * vk00; in xnn_f32_dwconv2d_chw_ukernel_3x3p1__scalar_6x1() local
221 float vo5p0 = vbias + vi5x0 * vk00; in xnn_f32_dwconv2d_chw_ukernel_3x3p1__scalar_6x1() local
|
| D | 3x3p1-minmax-wasmsimd-x86-splat-6x4.c | 111 v128_t vo5p0 = wasm_v32x4_shuffle(vw0123, vw0123, 0, 0, 0, 0); in xnn_f32_dwconv2d_chw_ukernel_3x3p1__wasmsimd_x86_splat_6x4() local
251 v128_t vo5p0 = wasm_v32x4_shuffle(vw0123, vw0123, 0, 0, 0, 0); in xnn_f32_dwconv2d_chw_ukernel_3x3p1__wasmsimd_x86_splat_6x4() local
|
| D | 3x3p1-minmax-neonfma-6x4.c | 110 float32x4_t vo5p0 = vdupq_lane_f32(vget_low_f32(vw0123), 0); in xnn_f32_dwconv2d_chw_ukernel_3x3p1__neonfma_6x4() local
251 float32x4_t vo5p0 = vdupq_lane_f32(vget_low_f32(vw0123), 0); in xnn_f32_dwconv2d_chw_ukernel_3x3p1__neonfma_6x4() local
|
| D | 3x3p1-minmax-wasmsimd-arm-splat-6x4.c | 111 v128_t vo5p0 = wasm_v32x4_shuffle(vw0123, vw0123, 0, 0, 0, 0); in xnn_f32_dwconv2d_chw_ukernel_3x3p1__wasmsimd_arm_splat_6x4() local
251 v128_t vo5p0 = wasm_v32x4_shuffle(vw0123, vw0123, 0, 0, 0, 0); in xnn_f32_dwconv2d_chw_ukernel_3x3p1__wasmsimd_arm_splat_6x4() local
|
| D | 3x3p1-minmax-neon-6x4.c | 110 float32x4_t vo5p0 = vdupq_lane_f32(vget_low_f32(vw0123), 0); in xnn_f32_dwconv2d_chw_ukernel_3x3p1__neon_6x4() local
251 float32x4_t vo5p0 = vdupq_lane_f32(vget_low_f32(vw0123), 0); in xnn_f32_dwconv2d_chw_ukernel_3x3p1__neon_6x4() local
|
| D | 3x3p1-minmax-ssse3-6x4.c | 142 __m128 vo5p0 = _mm_add_ps(vbias, _mm_mul_ps(vi5x4567, vk01)); in xnn_f32_dwconv2d_chw_ukernel_3x3p1__ssse3_6x4() local
276 __m128 vo5p0 = _mm_add_ps(vbias, _mm_mul_ps(vi5x4567, vk01)); in xnn_f32_dwconv2d_chw_ukernel_3x3p1__ssse3_6x4() local
|
| D | 3x3p1-minmax-wasmsimd-x86-loadsplat-6x4.c | 147 v128_t vo5p0 = wasm_f32x4_add(vbias, wasm_f32x4_mul(vi5x4567, vk01)); in xnn_f32_dwconv2d_chw_ukernel_3x3p1__wasmsimd_x86_loadsplat_6x4() local
280 v128_t vo5p0 = wasm_f32x4_add(vbias, wasm_f32x4_mul(vi5x4567, vk01)); in xnn_f32_dwconv2d_chw_ukernel_3x3p1__wasmsimd_x86_loadsplat_6x4() local
|
| D | 3x3p1-minmax-wasmsimd-arm-loadsplat-6x4.c | 147 v128_t vo5p0 = wasm_f32x4_add(vbias, wasm_f32x4_mul(vi5x4567, vk01)); in xnn_f32_dwconv2d_chw_ukernel_3x3p1__wasmsimd_arm_loadsplat_6x4() local
280 v128_t vo5p0 = wasm_f32x4_add(vbias, wasm_f32x4_mul(vi5x4567, vk01)); in xnn_f32_dwconv2d_chw_ukernel_3x3p1__wasmsimd_arm_loadsplat_6x4() local
|
| D | 3x3p1-minmax-sse-6x4.c | 175 __m128 vo5p0 = _mm_add_ps(vbias, _mm_mul_ps(vi5x4567, vk01)); in xnn_f32_dwconv2d_chw_ukernel_3x3p1__sse_6x4() local
359 __m128 vo5p0 = _mm_add_ps(vbias, _mm_mul_ps(vi5x4567, vk01)); in xnn_f32_dwconv2d_chw_ukernel_3x3p1__sse_6x4() local
|
| D | 3x3s2p1-minmax-sse-6x4.c | 207 __m128 vo5p0 = _mm_add_ps(vbias, _mm_mul_ps(vi10x8ACE, vk01)); in xnn_f32_dwconv2d_chw_ukernel_3x3s2p1__sse_6x4() local
391 __m128 vo5p0 = _mm_add_ps(vbias, _mm_mul_ps(vi10x8ACE, vk01)); in xnn_f32_dwconv2d_chw_ukernel_3x3s2p1__sse_6x4() local
|