/external/XNNPACK/src/f32-pavgpool/ |
D | 9p8x-minmax-sse-c4.c | 106 const __m128 vsum67 = _mm_add_ps(vi6, vi7); in xnn_f32_pavgpool_minmax_ukernel_9p8x__sse_c4() local 109 const __m128 vsum01678 = _mm_add_ps(vsum018, vsum67); in xnn_f32_pavgpool_minmax_ukernel_9p8x__sse_c4() 182 const __m128 vsum67 = _mm_add_ps(vi6, vi7); in xnn_f32_pavgpool_minmax_ukernel_9p8x__sse_c4() local 185 const __m128 vsum0167a = _mm_add_ps(vsum01a, vsum67); in xnn_f32_pavgpool_minmax_ukernel_9p8x__sse_c4() 284 const __m128 vsum67 = _mm_add_ps(vi6, vi7); in xnn_f32_pavgpool_minmax_ukernel_9p8x__sse_c4() local 287 const __m128 vsum0167a = _mm_add_ps(vsum01a, vsum67); in xnn_f32_pavgpool_minmax_ukernel_9p8x__sse_c4() 313 const __m128 vsum67 = _mm_add_ps(vi6, vi7); in xnn_f32_pavgpool_minmax_ukernel_9p8x__sse_c4() local 316 const __m128 vsum0167a = _mm_add_ps(vsum01a, vsum67); in xnn_f32_pavgpool_minmax_ukernel_9p8x__sse_c4()
|
D | 9p8x-minmax-wasmsimd-arm-c4.c | 106 const v128_t vsum67 = wasm_f32x4_add(vi6, vi7); in xnn_f32_pavgpool_minmax_ukernel_9p8x__wasmsimd_arm_c4() local 109 const v128_t vsum01678 = wasm_f32x4_add(vsum018, vsum67); in xnn_f32_pavgpool_minmax_ukernel_9p8x__wasmsimd_arm_c4() 183 const v128_t vsum67 = wasm_f32x4_add(vi6, vi7); in xnn_f32_pavgpool_minmax_ukernel_9p8x__wasmsimd_arm_c4() local 186 const v128_t vsum0167a = wasm_f32x4_add(vsum01a, vsum67); in xnn_f32_pavgpool_minmax_ukernel_9p8x__wasmsimd_arm_c4() 286 const v128_t vsum67 = wasm_f32x4_add(vi6, vi7); in xnn_f32_pavgpool_minmax_ukernel_9p8x__wasmsimd_arm_c4() local 289 const v128_t vsum0167a = wasm_f32x4_add(vsum01a, vsum67); in xnn_f32_pavgpool_minmax_ukernel_9p8x__wasmsimd_arm_c4() 315 const v128_t vsum67 = wasm_f32x4_add(vi6, vi7); in xnn_f32_pavgpool_minmax_ukernel_9p8x__wasmsimd_arm_c4() local 318 const v128_t vsum0167a = wasm_f32x4_add(vsum01a, vsum67); in xnn_f32_pavgpool_minmax_ukernel_9p8x__wasmsimd_arm_c4()
|
D | 9p8x-minmax-neon-c4.c | 97 const float32x4_t vsum67 = vaddq_f32(vi6, vi7); in xnn_f32_pavgpool_minmax_ukernel_9p8x__neon_c4() local 100 const float32x4_t vsum01678 = vaddq_f32(vsum018, vsum67); in xnn_f32_pavgpool_minmax_ukernel_9p8x__neon_c4() 165 const float32x4_t vsum67 = vaddq_f32(vi6, vi7); in xnn_f32_pavgpool_minmax_ukernel_9p8x__neon_c4() local 168 const float32x4_t vsum0167a = vaddq_f32(vsum01a, vsum67); in xnn_f32_pavgpool_minmax_ukernel_9p8x__neon_c4() 257 const float32x4_t vsum67 = vaddq_f32(vi6, vi7); in xnn_f32_pavgpool_minmax_ukernel_9p8x__neon_c4() local 260 const float32x4_t vsum0167a = vaddq_f32(vsum01a, vsum67); in xnn_f32_pavgpool_minmax_ukernel_9p8x__neon_c4() 285 const float32x4_t vsum67 = vaddq_f32(vi6, vi7); in xnn_f32_pavgpool_minmax_ukernel_9p8x__neon_c4() local 288 const float32x4_t vsum0167a = vaddq_f32(vsum01a, vsum67); in xnn_f32_pavgpool_minmax_ukernel_9p8x__neon_c4()
|
D | 9p8x-minmax-wasm-c1.c | 97 const float vsum67 = vi6 + vi7; in xnn_f32_pavgpool_minmax_ukernel_9p8x__wasm_c1() local 100 const float vsum01678 = vsum018 + vsum67; in xnn_f32_pavgpool_minmax_ukernel_9p8x__wasm_c1() 166 const float vsum67 = vi6 + vi7; in xnn_f32_pavgpool_minmax_ukernel_9p8x__wasm_c1() local 169 const float vsum0167a = vsum01a + vsum67; in xnn_f32_pavgpool_minmax_ukernel_9p8x__wasm_c1() 258 const float vsum67 = vi6 + vi7; in xnn_f32_pavgpool_minmax_ukernel_9p8x__wasm_c1() local 261 const float vsum0167a = vsum01a + vsum67; in xnn_f32_pavgpool_minmax_ukernel_9p8x__wasm_c1()
|
D | 9p8x-minmax-scalar-c1.c | 97 const float vsum67 = vi6 + vi7; in xnn_f32_pavgpool_minmax_ukernel_9p8x__scalar_c1() local 100 const float vsum01678 = vsum018 + vsum67; in xnn_f32_pavgpool_minmax_ukernel_9p8x__scalar_c1() 166 const float vsum67 = vi6 + vi7; in xnn_f32_pavgpool_minmax_ukernel_9p8x__scalar_c1() local 169 const float vsum0167a = vsum01a + vsum67; in xnn_f32_pavgpool_minmax_ukernel_9p8x__scalar_c1() 258 const float vsum67 = vi6 + vi7; in xnn_f32_pavgpool_minmax_ukernel_9p8x__scalar_c1() local 261 const float vsum0167a = vsum01a + vsum67; in xnn_f32_pavgpool_minmax_ukernel_9p8x__scalar_c1()
|
D | 9p8x-minmax-wasmsimd-x86-c4.c | 106 const v128_t vsum67 = wasm_f32x4_add(vi6, vi7); in xnn_f32_pavgpool_minmax_ukernel_9p8x__wasmsimd_x86_c4() local 109 const v128_t vsum01678 = wasm_f32x4_add(vsum018, vsum67); in xnn_f32_pavgpool_minmax_ukernel_9p8x__wasmsimd_x86_c4() 183 const v128_t vsum67 = wasm_f32x4_add(vi6, vi7); in xnn_f32_pavgpool_minmax_ukernel_9p8x__wasmsimd_x86_c4() local 186 const v128_t vsum0167a = wasm_f32x4_add(vsum01a, vsum67); in xnn_f32_pavgpool_minmax_ukernel_9p8x__wasmsimd_x86_c4() 286 const v128_t vsum67 = wasm_f32x4_add(vi6, vi7); in xnn_f32_pavgpool_minmax_ukernel_9p8x__wasmsimd_x86_c4() local 289 const v128_t vsum0167a = wasm_f32x4_add(vsum01a, vsum67); in xnn_f32_pavgpool_minmax_ukernel_9p8x__wasmsimd_x86_c4() 317 const v128_t vsum67 = wasm_f32x4_add(vi6, vi7); in xnn_f32_pavgpool_minmax_ukernel_9p8x__wasmsimd_x86_c4() local 320 const v128_t vsum0167a = wasm_f32x4_add(vsum01a, vsum67); in xnn_f32_pavgpool_minmax_ukernel_9p8x__wasmsimd_x86_c4()
|
D | 9x-minmax-neon-c4.c | 123 const float32x4_t vsum67 = vaddq_f32(vi6, vi7); in xnn_f32_pavgpool_minmax_ukernel_9x__neon_c4() local 126 const float32x4_t vsum01678 = vaddq_f32(vsum018, vsum67); in xnn_f32_pavgpool_minmax_ukernel_9x__neon_c4() 151 const float32x4_t vsum67 = vaddq_f32(vi6, vi7); in xnn_f32_pavgpool_minmax_ukernel_9x__neon_c4() local 154 const float32x4_t vsum01678 = vaddq_f32(vsum018, vsum67); in xnn_f32_pavgpool_minmax_ukernel_9x__neon_c4()
|
D | 9x-minmax-sse-c4.c | 133 const __m128 vsum67 = _mm_add_ps(vi6, vi7); in xnn_f32_pavgpool_minmax_ukernel_9x__sse_c4() local 136 const __m128 vsum01678 = _mm_add_ps(vsum018, vsum67); in xnn_f32_pavgpool_minmax_ukernel_9x__sse_c4() 161 const __m128 vsum67 = _mm_add_ps(vi6, vi7); in xnn_f32_pavgpool_minmax_ukernel_9x__sse_c4() local 164 const __m128 vsum01678 = _mm_add_ps(vsum018, vsum67); in xnn_f32_pavgpool_minmax_ukernel_9x__sse_c4()
|
D | 9x-minmax-wasmsimd-arm-c4.c | 133 const v128_t vsum67 = wasm_f32x4_add(vi6, vi7); in xnn_f32_pavgpool_minmax_ukernel_9x__wasmsimd_arm_c4() local 136 const v128_t vsum01678 = wasm_f32x4_add(vsum018, vsum67); in xnn_f32_pavgpool_minmax_ukernel_9x__wasmsimd_arm_c4() 162 const v128_t vsum67 = wasm_f32x4_add(vi6, vi7); in xnn_f32_pavgpool_minmax_ukernel_9x__wasmsimd_arm_c4() local 165 const v128_t vsum01678 = wasm_f32x4_add(vsum018, vsum67); in xnn_f32_pavgpool_minmax_ukernel_9x__wasmsimd_arm_c4()
|
D | 9x-minmax-wasmsimd-x86-c4.c | 133 const v128_t vsum67 = wasm_f32x4_add(vi6, vi7); in xnn_f32_pavgpool_minmax_ukernel_9x__wasmsimd_x86_c4() local 136 const v128_t vsum01678 = wasm_f32x4_add(vsum018, vsum67); in xnn_f32_pavgpool_minmax_ukernel_9x__wasmsimd_x86_c4() 164 const v128_t vsum67 = wasm_f32x4_add(vi6, vi7); in xnn_f32_pavgpool_minmax_ukernel_9x__wasmsimd_x86_c4() local 167 const v128_t vsum01678 = wasm_f32x4_add(vsum018, vsum67); in xnn_f32_pavgpool_minmax_ukernel_9x__wasmsimd_x86_c4()
|
D | 9x-minmax-scalar-c1.c | 122 const float vsum67 = vi6 + vi7; in xnn_f32_pavgpool_minmax_ukernel_9x__scalar_c1() local 125 const float vsum01678 = vsum018 + vsum67; in xnn_f32_pavgpool_minmax_ukernel_9x__scalar_c1()
|
/external/XNNPACK/src/f32-avgpool/ |
D | 9p8x-minmax-neon-c4.c | 97 const float32x4_t vsum67 = vaddq_f32(vi6, vi7); in xnn_f32_avgpool_minmax_ukernel_9p8x__neon_c4() local 100 const float32x4_t vsum01678 = vaddq_f32(vsum018, vsum67); in xnn_f32_avgpool_minmax_ukernel_9p8x__neon_c4() 165 const float32x4_t vsum67 = vaddq_f32(vi6, vi7); in xnn_f32_avgpool_minmax_ukernel_9p8x__neon_c4() local 168 const float32x4_t vsum0167a = vaddq_f32(vsum01a, vsum67); in xnn_f32_avgpool_minmax_ukernel_9p8x__neon_c4() 256 const float32x4_t vsum67 = vaddq_f32(vi6, vi7); in xnn_f32_avgpool_minmax_ukernel_9p8x__neon_c4() local 259 const float32x4_t vsum0167a = vaddq_f32(vsum01a, vsum67); in xnn_f32_avgpool_minmax_ukernel_9p8x__neon_c4() 284 const float32x4_t vsum67 = vaddq_f32(vi6, vi7); in xnn_f32_avgpool_minmax_ukernel_9p8x__neon_c4() local 287 const float32x4_t vsum0167a = vaddq_f32(vsum01a, vsum67); in xnn_f32_avgpool_minmax_ukernel_9p8x__neon_c4()
|
D | 9p8x-minmax-wasmsimd-arm-c4.c | 106 const v128_t vsum67 = wasm_f32x4_add(vi6, vi7); in xnn_f32_avgpool_minmax_ukernel_9p8x__wasmsimd_arm_c4() local 109 const v128_t vsum01678 = wasm_f32x4_add(vsum018, vsum67); in xnn_f32_avgpool_minmax_ukernel_9p8x__wasmsimd_arm_c4() 183 const v128_t vsum67 = wasm_f32x4_add(vi6, vi7); in xnn_f32_avgpool_minmax_ukernel_9p8x__wasmsimd_arm_c4() local 186 const v128_t vsum0167a = wasm_f32x4_add(vsum01a, vsum67); in xnn_f32_avgpool_minmax_ukernel_9p8x__wasmsimd_arm_c4() 283 const v128_t vsum67 = wasm_f32x4_add(vi6, vi7); in xnn_f32_avgpool_minmax_ukernel_9p8x__wasmsimd_arm_c4() local 286 const v128_t vsum0167a = wasm_f32x4_add(vsum01a, vsum67); in xnn_f32_avgpool_minmax_ukernel_9p8x__wasmsimd_arm_c4() 312 const v128_t vsum67 = wasm_f32x4_add(vi6, vi7); in xnn_f32_avgpool_minmax_ukernel_9p8x__wasmsimd_arm_c4() local 315 const v128_t vsum0167a = wasm_f32x4_add(vsum01a, vsum67); in xnn_f32_avgpool_minmax_ukernel_9p8x__wasmsimd_arm_c4()
|
D | 9p8x-minmax-sse-c4.c | 106 const __m128 vsum67 = _mm_add_ps(vi6, vi7); in xnn_f32_avgpool_minmax_ukernel_9p8x__sse_c4() local 109 const __m128 vsum01678 = _mm_add_ps(vsum018, vsum67); in xnn_f32_avgpool_minmax_ukernel_9p8x__sse_c4() 182 const __m128 vsum67 = _mm_add_ps(vi6, vi7); in xnn_f32_avgpool_minmax_ukernel_9p8x__sse_c4() local 185 const __m128 vsum0167a = _mm_add_ps(vsum01a, vsum67); in xnn_f32_avgpool_minmax_ukernel_9p8x__sse_c4() 281 const __m128 vsum67 = _mm_add_ps(vi6, vi7); in xnn_f32_avgpool_minmax_ukernel_9p8x__sse_c4() local 284 const __m128 vsum0167a = _mm_add_ps(vsum01a, vsum67); in xnn_f32_avgpool_minmax_ukernel_9p8x__sse_c4() 310 const __m128 vsum67 = _mm_add_ps(vi6, vi7); in xnn_f32_avgpool_minmax_ukernel_9p8x__sse_c4() local 313 const __m128 vsum0167a = _mm_add_ps(vsum01a, vsum67); in xnn_f32_avgpool_minmax_ukernel_9p8x__sse_c4()
|
D | 9p8x-minmax-wasm-c1.c | 97 const float vsum67 = vi6 + vi7; in xnn_f32_avgpool_minmax_ukernel_9p8x__wasm_c1() local 100 const float vsum01678 = vsum018 + vsum67; in xnn_f32_avgpool_minmax_ukernel_9p8x__wasm_c1() 166 const float vsum67 = vi6 + vi7; in xnn_f32_avgpool_minmax_ukernel_9p8x__wasm_c1() local 169 const float vsum0167a = vsum01a + vsum67; in xnn_f32_avgpool_minmax_ukernel_9p8x__wasm_c1() 256 const float vsum67 = vi6 + vi7; in xnn_f32_avgpool_minmax_ukernel_9p8x__wasm_c1() local 259 const float vsum0167a = vsum01a + vsum67; in xnn_f32_avgpool_minmax_ukernel_9p8x__wasm_c1()
|
D | 9p8x-minmax-scalar-c1.c | 97 const float vsum67 = vi6 + vi7; in xnn_f32_avgpool_minmax_ukernel_9p8x__scalar_c1() local 100 const float vsum01678 = vsum018 + vsum67; in xnn_f32_avgpool_minmax_ukernel_9p8x__scalar_c1() 166 const float vsum67 = vi6 + vi7; in xnn_f32_avgpool_minmax_ukernel_9p8x__scalar_c1() local 169 const float vsum0167a = vsum01a + vsum67; in xnn_f32_avgpool_minmax_ukernel_9p8x__scalar_c1() 256 const float vsum67 = vi6 + vi7; in xnn_f32_avgpool_minmax_ukernel_9p8x__scalar_c1() local 259 const float vsum0167a = vsum01a + vsum67; in xnn_f32_avgpool_minmax_ukernel_9p8x__scalar_c1()
|
D | 9p8x-minmax-wasmsimd-x86-c4.c | 106 const v128_t vsum67 = wasm_f32x4_add(vi6, vi7); in xnn_f32_avgpool_minmax_ukernel_9p8x__wasmsimd_x86_c4() local 109 const v128_t vsum01678 = wasm_f32x4_add(vsum018, vsum67); in xnn_f32_avgpool_minmax_ukernel_9p8x__wasmsimd_x86_c4() 183 const v128_t vsum67 = wasm_f32x4_add(vi6, vi7); in xnn_f32_avgpool_minmax_ukernel_9p8x__wasmsimd_x86_c4() local 186 const v128_t vsum0167a = wasm_f32x4_add(vsum01a, vsum67); in xnn_f32_avgpool_minmax_ukernel_9p8x__wasmsimd_x86_c4() 283 const v128_t vsum67 = wasm_f32x4_add(vi6, vi7); in xnn_f32_avgpool_minmax_ukernel_9p8x__wasmsimd_x86_c4() local 286 const v128_t vsum0167a = wasm_f32x4_add(vsum01a, vsum67); in xnn_f32_avgpool_minmax_ukernel_9p8x__wasmsimd_x86_c4() 314 const v128_t vsum67 = wasm_f32x4_add(vi6, vi7); in xnn_f32_avgpool_minmax_ukernel_9p8x__wasmsimd_x86_c4() local 317 const v128_t vsum0167a = wasm_f32x4_add(vsum01a, vsum67); in xnn_f32_avgpool_minmax_ukernel_9p8x__wasmsimd_x86_c4()
|
D | 9x-minmax-sse-c4.c | 130 const __m128 vsum67 = _mm_add_ps(vi6, vi7); in xnn_f32_avgpool_minmax_ukernel_9x__sse_c4() local 133 const __m128 vsum01678 = _mm_add_ps(vsum018, vsum67); in xnn_f32_avgpool_minmax_ukernel_9x__sse_c4() 158 const __m128 vsum67 = _mm_add_ps(vi6, vi7); in xnn_f32_avgpool_minmax_ukernel_9x__sse_c4() local 161 const __m128 vsum01678 = _mm_add_ps(vsum018, vsum67); in xnn_f32_avgpool_minmax_ukernel_9x__sse_c4()
|
D | 9x-minmax-neon-c4.c | 121 const float32x4_t vsum67 = vaddq_f32(vi6, vi7); in xnn_f32_avgpool_minmax_ukernel_9x__neon_c4() local 124 const float32x4_t vsum01678 = vaddq_f32(vsum018, vsum67); in xnn_f32_avgpool_minmax_ukernel_9x__neon_c4() 149 const float32x4_t vsum67 = vaddq_f32(vi6, vi7); in xnn_f32_avgpool_minmax_ukernel_9x__neon_c4() local 152 const float32x4_t vsum01678 = vaddq_f32(vsum018, vsum67); in xnn_f32_avgpool_minmax_ukernel_9x__neon_c4()
|
D | 9x-minmax-wasmsimd-arm-c4.c | 130 const v128_t vsum67 = wasm_f32x4_add(vi6, vi7); in xnn_f32_avgpool_minmax_ukernel_9x__wasmsimd_arm_c4() local 133 const v128_t vsum01678 = wasm_f32x4_add(vsum018, vsum67); in xnn_f32_avgpool_minmax_ukernel_9x__wasmsimd_arm_c4() 159 const v128_t vsum67 = wasm_f32x4_add(vi6, vi7); in xnn_f32_avgpool_minmax_ukernel_9x__wasmsimd_arm_c4() local 162 const v128_t vsum01678 = wasm_f32x4_add(vsum018, vsum67); in xnn_f32_avgpool_minmax_ukernel_9x__wasmsimd_arm_c4()
|
D | 9x-minmax-wasmsimd-x86-c4.c | 130 const v128_t vsum67 = wasm_f32x4_add(vi6, vi7); in xnn_f32_avgpool_minmax_ukernel_9x__wasmsimd_x86_c4() local 133 const v128_t vsum01678 = wasm_f32x4_add(vsum018, vsum67); in xnn_f32_avgpool_minmax_ukernel_9x__wasmsimd_x86_c4() 161 const v128_t vsum67 = wasm_f32x4_add(vi6, vi7); in xnn_f32_avgpool_minmax_ukernel_9x__wasmsimd_x86_c4() local 164 const v128_t vsum01678 = wasm_f32x4_add(vsum018, vsum67); in xnn_f32_avgpool_minmax_ukernel_9x__wasmsimd_x86_c4()
|
D | 9x-minmax-wasm-c1.c | 120 const float vsum67 = vi6 + vi7; in xnn_f32_avgpool_minmax_ukernel_9x__wasm_c1() local 123 const float vsum01678 = vsum018 + vsum67; in xnn_f32_avgpool_minmax_ukernel_9x__wasm_c1()
|
/external/XNNPACK/src/qu8-avgpool/ |
D | 9p8x-minmax-scalar-c1.c | 104 const uint32_t vsum67 = vi6 + vi7; in xnn_qu8_avgpool_minmax_ukernel_9p8x__scalar_c1() local 107 const uint32_t vsum01678 = vsum018 + vsum67; in xnn_qu8_avgpool_minmax_ukernel_9p8x__scalar_c1() 175 const uint32_t vsum67 = vi6 + vi7; in xnn_qu8_avgpool_minmax_ukernel_9p8x__scalar_c1() local 177 const uint32_t vsum4567 = vsum45 + vsum67; in xnn_qu8_avgpool_minmax_ukernel_9p8x__scalar_c1() 267 const uint32_t vsum67 = vi6 + vi7; in xnn_qu8_avgpool_minmax_ukernel_9p8x__scalar_c1() local 269 const uint32_t vsum4567 = vsum45 + vsum67; in xnn_qu8_avgpool_minmax_ukernel_9p8x__scalar_c1()
|
D | 9p8x-minmax-neon-c8.c | 108 const uint16x8_t vsum67 = vaddl_u8(vi6, vi7); in xnn_qu8_avgpool_minmax_ukernel_9p8x__neon_c8() local 111 const uint16x8_t vsum01678 = vaddq_u16(vsum018, vsum67); in xnn_qu8_avgpool_minmax_ukernel_9p8x__neon_c8() 181 const uint16x8_t vsum67 = vaddl_u8(vi6, vi7); in xnn_qu8_avgpool_minmax_ukernel_9p8x__neon_c8() local 184 const uint16x8_t vsum4567 = vaddq_u16(vsum45, vsum67); in xnn_qu8_avgpool_minmax_ukernel_9p8x__neon_c8() 276 const int16x8_t vsum67 = vreinterpretq_s16_u16(vaddl_u8(vi6, vi7)); in xnn_qu8_avgpool_minmax_ukernel_9p8x__neon_c8() local 279 const int16x8_t vsum4567 = vaddq_s16(vsum45, vsum67); in xnn_qu8_avgpool_minmax_ukernel_9p8x__neon_c8() 350 const int16x8_t vsum67 = vreinterpretq_s16_u16(vaddl_u8(vi6, vi7)); in xnn_qu8_avgpool_minmax_ukernel_9p8x__neon_c8() local 353 const int16x8_t vsum4567 = vaddq_s16(vsum45, vsum67); in xnn_qu8_avgpool_minmax_ukernel_9p8x__neon_c8()
|
D | 9p8x-minmax-sse2-c8.c | 112 const __m128i vsum67 = _mm_add_epi16(vxi6, vxi7); in xnn_qu8_avgpool_minmax_ukernel_9p8x__sse2_c8() local 115 const __m128i vsum01678 = _mm_add_epi16(vsum018, vsum67); in xnn_qu8_avgpool_minmax_ukernel_9p8x__sse2_c8() 195 const __m128i vsum67 = _mm_add_epi16(vxi6, vxi7); in xnn_qu8_avgpool_minmax_ukernel_9p8x__sse2_c8() local 198 const __m128i vsum4567 = _mm_add_epi16(vsum45, vsum67); in xnn_qu8_avgpool_minmax_ukernel_9p8x__sse2_c8() 301 const __m128i vsum67 = _mm_add_epi16(vxi6, vxi7); in xnn_qu8_avgpool_minmax_ukernel_9p8x__sse2_c8() local 304 const __m128i vsum4567 = _mm_add_epi16(vsum45, vsum67); in xnn_qu8_avgpool_minmax_ukernel_9p8x__sse2_c8() 376 const __m128i vsum67 = _mm_add_epi16(vxi6, vxi7); in xnn_qu8_avgpool_minmax_ukernel_9p8x__sse2_c8() local 379 const __m128i vsum4567 = _mm_add_epi16(vsum45, vsum67); in xnn_qu8_avgpool_minmax_ukernel_9p8x__sse2_c8()
|