| /external/arm-optimized-routines/networking/arm/ |
| D | chksum_simd.c | 52 uint64x2_t vsum0 = { 0, 0 }; in __chksum_arm_simd() local
63 vsum0 = vpadalq_u32(vsum0, vtmp0); in __chksum_arm_simd()
72 vsum0 = vpadalq_u32(vsum0, vreinterpretq_u32_u64(vsum2)); in __chksum_arm_simd()
74 vsum0 = vpadalq_u32(vsum0, vreinterpretq_u32_u64(vsum1)); in __chksum_arm_simd()
80 vsum0 = vpadalq_u32(vsum0, vtmp0); in __chksum_arm_simd()
89 vsum0 = vpaddlq_u32(vreinterpretq_u32_u64(vsum0)); in __chksum_arm_simd()
91 vsum0 = vpaddlq_u32(vreinterpretq_u32_u64(vsum0)); in __chksum_arm_simd()
93 Assert((vgetq_lane_u64(vsum0, 0) >> 32) == 0); in __chksum_arm_simd()
94 Assert((vgetq_lane_u64(vsum0, 1) >> 32) == 0); in __chksum_arm_simd()
95 uint32x2_t vtmp = vmovn_u64(vsum0); in __chksum_arm_simd()
|
| /external/llvm-project/libc/AOR_v20.02/networking/arm/ |
| D | chksum_simd.c | 53 uint64x2_t vsum0 = { 0, 0 }; in __chksum_arm_simd() local
64 vsum0 = vpadalq_u32(vsum0, vtmp0); in __chksum_arm_simd()
73 vsum0 = vpadalq_u32(vsum0, vreinterpretq_u32_u64(vsum2)); in __chksum_arm_simd()
75 vsum0 = vpadalq_u32(vsum0, vreinterpretq_u32_u64(vsum1)); in __chksum_arm_simd()
81 vsum0 = vpadalq_u32(vsum0, vtmp0); in __chksum_arm_simd()
90 vsum0 = vpaddlq_u32(vreinterpretq_u32_u64(vsum0)); in __chksum_arm_simd()
92 vsum0 = vpaddlq_u32(vreinterpretq_u32_u64(vsum0)); in __chksum_arm_simd()
94 Assert((vgetq_lane_u64(vsum0, 0) >> 32) == 0); in __chksum_arm_simd()
95 Assert((vgetq_lane_u64(vsum0, 1) >> 32) == 0); in __chksum_arm_simd()
96 uint32x2_t vtmp = vmovn_u64(vsum0); in __chksum_arm_simd()
|
| /external/arm-optimized-routines/networking/aarch64/ |
| D | chksum_simd.c | 78 uint64x2_t vsum0 = { 0, 0 }; in __chksum_aarch64_simd() local
90 vsum0 = vpadalq_u32(vsum0, vtmp0); in __chksum_aarch64_simd()
99 vsum0 = vpadalq_u32(vsum0, vreinterpretq_u32_u64(vsum2)); in __chksum_aarch64_simd()
107 vsum0 = vpadalq_u32(vsum0, vtmp0); in __chksum_aarch64_simd()
115 vsum0 = vpadalq_u32(vsum0, vreinterpretq_u32_u64(vsum1)); in __chksum_aarch64_simd()
121 vsum0 = vpadalq_u32(vsum0, vtmp); in __chksum_aarch64_simd()
131 vsum0 = vaddw_u32(vsum0, vtmp); in __chksum_aarch64_simd()
137 uint64_t val = vaddlvq_u32(vreinterpretq_u32_u64(vsum0)); in __chksum_aarch64_simd()
|
| /external/llvm-project/libc/AOR_v20.02/networking/aarch64/ |
| D | chksum_simd.c | 79 uint64x2_t vsum0 = { 0, 0 }; in __chksum_aarch64_simd() local
91 vsum0 = vpadalq_u32(vsum0, vtmp0); in __chksum_aarch64_simd()
100 vsum0 = vpadalq_u32(vsum0, vreinterpretq_u32_u64(vsum2)); in __chksum_aarch64_simd()
108 vsum0 = vpadalq_u32(vsum0, vtmp0); in __chksum_aarch64_simd()
116 vsum0 = vpadalq_u32(vsum0, vreinterpretq_u32_u64(vsum1)); in __chksum_aarch64_simd()
122 vsum0 = vpadalq_u32(vsum0, vtmp); in __chksum_aarch64_simd()
132 vsum0 = vaddw_u32(vsum0, vtmp); in __chksum_aarch64_simd()
138 uint64_t val = vaddlvq_u32(vreinterpretq_u32_u64(vsum0)); in __chksum_aarch64_simd()
|
| /external/XNNPACK/src/f32-gavgpool-cw/ |
| D | neon-x4.c | 36 float32x4_t vsum0 = vmovq_n_f32(0.0f); in xnn_f32_gavgpool_cw_ukernel__neon_x4() local
47 vsum0 = vaddq_f32(vsum0, vi0); in xnn_f32_gavgpool_cw_ukernel__neon_x4()
65 vsum0 = vaddq_f32(vsum0, vi0); in xnn_f32_gavgpool_cw_ukernel__neon_x4()
74 const float32x4_t vsum01 = vpaddq_f32(vsum0, vsum1); in xnn_f32_gavgpool_cw_ukernel__neon_x4()
78 const float32x4_t vsum01 = vcombine_f32(vadd_f32(vget_low_f32(vsum0), vget_high_f32(vsum0)), in xnn_f32_gavgpool_cw_ukernel__neon_x4()
100 float32x4_t vsum0 = vmovq_n_f32(0.0f); in xnn_f32_gavgpool_cw_ukernel__neon_x4() local
104 vsum0 = vaddq_f32(vsum0, vi0); in xnn_f32_gavgpool_cw_ukernel__neon_x4()
111 vsum0 = vaddq_f32(vsum0, vi0); in xnn_f32_gavgpool_cw_ukernel__neon_x4()
114 float32x2_t vsum = vadd_f32(vget_low_f32(vsum0), vget_high_f32(vsum0)); in xnn_f32_gavgpool_cw_ukernel__neon_x4()
|
| D | wasmsimd-arm-x4.c | 36 v128_t vsum0 = wasm_f64x2_splat(0.0); in xnn_f32_gavgpool_cw_ukernel__wasmsimd_arm_x4() local
37 v128_t vsum1 = vsum0; in xnn_f32_gavgpool_cw_ukernel__wasmsimd_arm_x4()
38 v128_t vsum2 = vsum0; in xnn_f32_gavgpool_cw_ukernel__wasmsimd_arm_x4()
39 v128_t vsum3 = vsum0; in xnn_f32_gavgpool_cw_ukernel__wasmsimd_arm_x4()
51 vsum0 = wasm_f32x4_add(vsum0, vi0); in xnn_f32_gavgpool_cw_ukernel__wasmsimd_arm_x4()
68 vsum0 = wasm_f32x4_add(vsum0, vi0); in xnn_f32_gavgpool_cw_ukernel__wasmsimd_arm_x4()
76 … v128_t vsum01 = wasm_f32x4_add(wasm_v32x4_shuffle(vsum0, vsum1, 0, 2, 4, 6), wasm_v32x4_shuffle(v… in xnn_f32_gavgpool_cw_ukernel__wasmsimd_arm_x4()
|
| D | wasmsimd-x86-x4.c | 36 v128_t vsum0 = wasm_f64x2_splat(0.0); in xnn_f32_gavgpool_cw_ukernel__wasmsimd_x86_x4() local
37 v128_t vsum1 = vsum0; in xnn_f32_gavgpool_cw_ukernel__wasmsimd_x86_x4()
38 v128_t vsum2 = vsum0; in xnn_f32_gavgpool_cw_ukernel__wasmsimd_x86_x4()
39 v128_t vsum3 = vsum0; in xnn_f32_gavgpool_cw_ukernel__wasmsimd_x86_x4()
51 vsum0 = wasm_f32x4_add(vsum0, vi0); in xnn_f32_gavgpool_cw_ukernel__wasmsimd_x86_x4()
68 vsum0 = wasm_f32x4_add(vsum0, vi0); in xnn_f32_gavgpool_cw_ukernel__wasmsimd_x86_x4()
76 … v128_t vsum01 = wasm_f32x4_add(wasm_v32x4_shuffle(vsum0, vsum1, 0, 2, 4, 6), wasm_v32x4_shuffle(v… in xnn_f32_gavgpool_cw_ukernel__wasmsimd_x86_x4()
|
| D | sse-x4.c | 36 __m128 vsum0 = _mm_setzero_ps(); in xnn_f32_gavgpool_cw_ukernel__sse_x4() local
51 vsum0 = _mm_add_ps(vsum0, vi0); in xnn_f32_gavgpool_cw_ukernel__sse_x4()
68 vsum0 = _mm_add_ps(vsum0, vi0); in xnn_f32_gavgpool_cw_ukernel__sse_x4()
76 const __m128 vsum01 = _mm_add_ps(_mm_unpacklo_ps(vsum0, vsum1), _mm_unpackhi_ps(vsum0, vsum1)); in xnn_f32_gavgpool_cw_ukernel__sse_x4()
|
| D | scalar-x1.c | 30 float vsum0 = 0.f; in xnn_f32_gavgpool_cw_ukernel__scalar_x1() local
36 vsum0 += i0[0]; in xnn_f32_gavgpool_cw_ukernel__scalar_x1()
46 vsum0 += *i0++; in xnn_f32_gavgpool_cw_ukernel__scalar_x1()
50 float vout = ( (vsum0 + vsum1) + (vsum2 + vsum3) ) * vmultiplier; in xnn_f32_gavgpool_cw_ukernel__scalar_x1()
|