| /external/XNNPACK/src/f32-gavgpool-cw/ |
| D | neon-x4.c | 39 float32x4_t vsum3 = vmovq_n_f32(0.0f); in xnn_f32_gavgpool_cw_ukernel__neon_x4() local
50 vsum3 = vaddq_f32(vsum3, vi3); in xnn_f32_gavgpool_cw_ukernel__neon_x4()
68 vsum3 = vaddq_f32(vsum3, vi3); in xnn_f32_gavgpool_cw_ukernel__neon_x4()
75 const float32x4_t vsum23 = vpaddq_f32(vsum2, vsum3); in xnn_f32_gavgpool_cw_ukernel__neon_x4()
81 vadd_f32(vget_low_f32(vsum3), vget_high_f32(vsum3))); in xnn_f32_gavgpool_cw_ukernel__neon_x4()
|
| D | wasmsimd-arm-x4.c | 39 v128_t vsum3 = vsum0; in xnn_f32_gavgpool_cw_ukernel__wasmsimd_arm_x4() local
54 vsum3 = wasm_f32x4_add(vsum3, vi3); in xnn_f32_gavgpool_cw_ukernel__wasmsimd_arm_x4()
71 vsum3 = wasm_f32x4_add(vsum3, vi3); in xnn_f32_gavgpool_cw_ukernel__wasmsimd_arm_x4()
77 … vsum23 = wasm_f32x4_add(wasm_v32x4_shuffle(vsum2, vsum3, 0, 2, 4, 6), wasm_v32x4_shuffle(vsum2, v… in xnn_f32_gavgpool_cw_ukernel__wasmsimd_arm_x4()
|
| D | wasmsimd-x86-x4.c | 39 v128_t vsum3 = vsum0; in xnn_f32_gavgpool_cw_ukernel__wasmsimd_x86_x4() local
54 vsum3 = wasm_f32x4_add(vsum3, vi3); in xnn_f32_gavgpool_cw_ukernel__wasmsimd_x86_x4()
71 vsum3 = wasm_f32x4_add(vsum3, vi3); in xnn_f32_gavgpool_cw_ukernel__wasmsimd_x86_x4()
77 … vsum23 = wasm_f32x4_add(wasm_v32x4_shuffle(vsum2, vsum3, 0, 2, 4, 6), wasm_v32x4_shuffle(vsum2, v… in xnn_f32_gavgpool_cw_ukernel__wasmsimd_x86_x4()
|
| D | sse-x4.c | 39 __m128 vsum3 = _mm_setzero_ps(); in xnn_f32_gavgpool_cw_ukernel__sse_x4() local
54 vsum3 = _mm_add_ps(vsum3, vi3); in xnn_f32_gavgpool_cw_ukernel__sse_x4()
71 vsum3 = _mm_add_ps(vsum3, vi3); in xnn_f32_gavgpool_cw_ukernel__sse_x4()
77 const __m128 vsum23 = _mm_add_ps(_mm_unpacklo_ps(vsum2, vsum3), _mm_unpackhi_ps(vsum2, vsum3)); in xnn_f32_gavgpool_cw_ukernel__sse_x4()
|
| D | scalar-x1.c | 33 float vsum3 = 0.f; in xnn_f32_gavgpool_cw_ukernel__scalar_x1() local
39 vsum3 += i0[3]; in xnn_f32_gavgpool_cw_ukernel__scalar_x1()
50 float vout = ( (vsum0 + vsum1) + (vsum2 + vsum3) ) * vmultiplier; in xnn_f32_gavgpool_cw_ukernel__scalar_x1()
|
| /external/libaom/libaom/av1/encoder/x86/ |
| D | highbd_temporal_filter_avx2.c | 182 __m256i vsum3 = _mm256_add_epi32(vsum1, vsum2); in highbd_apply_temporal_filter() local
183 __m256i vsum = _mm256_add_epi32(vsum3, vsrc[4]); in highbd_apply_temporal_filter()
200 __m256i vsum3 = _mm256_add_epi32(vsum1, vsum2); in highbd_apply_temporal_filter() local
201 __m256i vsum = _mm256_add_epi32(vsum3, vsrc[4]); in highbd_apply_temporal_filter()
228 __m256i vsum3 = _mm256_add_epi32(vsum1, vsum2); in highbd_apply_temporal_filter() local
229 __m256i vsum = _mm256_add_epi32(vsum3, vsrc[4]); in highbd_apply_temporal_filter()
247 __m256i vsum3 = _mm256_add_epi32(vsum1, vsum2); in highbd_apply_temporal_filter() local
248 __m256i vsum = _mm256_add_epi32(vsum3, vsrc[4]); in highbd_apply_temporal_filter()
276 __m256i vsum3 = _mm256_add_epi32(vsum1, vsum2); in highbd_apply_temporal_filter() local
277 __m256i vsum = _mm256_add_epi32(vsum3, vsrc[4]); in highbd_apply_temporal_filter()
[all …]
|
| /external/arm-optimized-routines/networking/aarch64/ |
| D | chksum_simd.c | 81 uint64x2_t vsum3 = { 0, 0 }; in __chksum_aarch64_simd() local
93 vsum3 = vpadalq_u32(vsum3, vtmp3); in __chksum_aarch64_simd()
100 vsum1 = vpadalq_u32(vsum1, vreinterpretq_u32_u64(vsum3)); in __chksum_aarch64_simd()
|
| /external/arm-optimized-routines/networking/arm/ |
| D | chksum_simd.c | 55 uint64x2_t vsum3 = { 0, 0 }; in __chksum_arm_simd() local
66 vsum3 = vpadalq_u32(vsum3, vtmp3); in __chksum_arm_simd()
73 vsum1 = vpadalq_u32(vsum1, vreinterpretq_u32_u64(vsum3)); in __chksum_arm_simd()
|
| /external/XNNPACK/src/amalgam/ |
| D | sse.c | 4191 __m128 vsum3 = _mm_setzero_ps(); in xnn_f32_gavgpool_cw_ukernel__sse_x4() local
4206 vsum3 = _mm_add_ps(vsum3, vi3); in xnn_f32_gavgpool_cw_ukernel__sse_x4()
4223 vsum3 = _mm_add_ps(vsum3, vi3); in xnn_f32_gavgpool_cw_ukernel__sse_x4()
4229 const __m128 vsum23 = _mm_add_ps(_mm_unpacklo_ps(vsum2, vsum3), _mm_unpackhi_ps(vsum2, vsum3)); in xnn_f32_gavgpool_cw_ukernel__sse_x4()
|