/external/XNNPACK/src/qs8-gavgpool/gen/ |
D | 7x-minmax-neon-c8-acc2.c | 89 const int32x4_t vsgnacc4567 = vreinterpretq_s32_u32(vcltq_s32(vacc4567, vmovq_n_s32(0))); in xnn_qs8_gavgpool_minmax_ukernel_7x__neon_c8_acc2() local 99 const int64x2_t vadjprod45 = vaddw_s32(vprod45, vget_low_s32(vsgnacc4567)); in xnn_qs8_gavgpool_minmax_ukernel_7x__neon_c8_acc2() 100 const int64x2_t vadjprod67 = vaddw_high_s32(vprod67, vsgnacc4567); in xnn_qs8_gavgpool_minmax_ukernel_7x__neon_c8_acc2() 109 const int64x2_t vadjprod45 = vaddw_s32(vprod45, vget_low_s32(vsgnacc4567)); in xnn_qs8_gavgpool_minmax_ukernel_7x__neon_c8_acc2() 110 const int64x2_t vadjprod67 = vaddw_s32(vprod67, vget_high_s32(vsgnacc4567)); in xnn_qs8_gavgpool_minmax_ukernel_7x__neon_c8_acc2() 166 const int32x4_t vsgnacc4567 = vreinterpretq_s32_u32(vcltq_s32(vacc4567, vmovq_n_s32(0))); in xnn_qs8_gavgpool_minmax_ukernel_7x__neon_c8_acc2() local 176 const int64x2_t vadjprod45 = vaddw_s32(vprod45, vget_low_s32(vsgnacc4567)); in xnn_qs8_gavgpool_minmax_ukernel_7x__neon_c8_acc2() 177 const int64x2_t vadjprod67 = vaddw_high_s32(vprod67, vsgnacc4567); in xnn_qs8_gavgpool_minmax_ukernel_7x__neon_c8_acc2() 186 const int64x2_t vadjprod45 = vaddw_s32(vprod45, vget_low_s32(vsgnacc4567)); in xnn_qs8_gavgpool_minmax_ukernel_7x__neon_c8_acc2() 187 const int64x2_t vadjprod67 = vaddw_s32(vprod67, vget_high_s32(vsgnacc4567)); in xnn_qs8_gavgpool_minmax_ukernel_7x__neon_c8_acc2()
|
D | 7x-minmax-sse2-c8-acc2.c | 99 const __m128i vsgnacc4567 = _mm_cmpgt_epi32(_mm_setzero_si128(), vacc4567); in xnn_qs8_gavgpool_minmax_ukernel_7x__sse2_c8_acc2() local 102 const __m128i vabsacc4567 = _mm_sub_epi32(_mm_xor_si128(vacc4567, vsgnacc4567), vsgnacc4567); in xnn_qs8_gavgpool_minmax_ukernel_7x__sse2_c8_acc2() 126 const __m128i vout4567 = _mm_sub_epi32(_mm_xor_si128(vabsout4567, vsgnacc4567), vsgnacc4567); in xnn_qs8_gavgpool_minmax_ukernel_7x__sse2_c8_acc2() 182 const __m128i vsgnacc4567 = _mm_cmpgt_epi32(_mm_setzero_si128(), vacc4567); in xnn_qs8_gavgpool_minmax_ukernel_7x__sse2_c8_acc2() local 185 const __m128i vabsacc4567 = _mm_sub_epi32(_mm_xor_si128(vacc4567, vsgnacc4567), vsgnacc4567); in xnn_qs8_gavgpool_minmax_ukernel_7x__sse2_c8_acc2() 209 const __m128i vout4567 = _mm_sub_epi32(_mm_xor_si128(vabsout4567, vsgnacc4567), vsgnacc4567); in xnn_qs8_gavgpool_minmax_ukernel_7x__sse2_c8_acc2()
|
D | 7x-minmax-neon-c16-acc2.c | 104 const int32x4_t vsgnacc4567 = vreinterpretq_s32_u32(vcltq_s32(vacc4567, vmovq_n_s32(0))); in xnn_qs8_gavgpool_minmax_ukernel_7x__neon_c16_acc2() local 120 const int64x2_t vadjprod45 = vaddw_s32(vprod45, vget_low_s32(vsgnacc4567)); in xnn_qs8_gavgpool_minmax_ukernel_7x__neon_c16_acc2() 121 const int64x2_t vadjprod67 = vaddw_high_s32(vprod67, vsgnacc4567); in xnn_qs8_gavgpool_minmax_ukernel_7x__neon_c16_acc2() 138 const int64x2_t vadjprod45 = vaddw_s32(vprod45, vget_low_s32(vsgnacc4567)); in xnn_qs8_gavgpool_minmax_ukernel_7x__neon_c16_acc2() 139 const int64x2_t vadjprod67 = vaddw_s32(vprod67, vget_high_s32(vsgnacc4567)); in xnn_qs8_gavgpool_minmax_ukernel_7x__neon_c16_acc2() 209 const int32x4_t vsgnacc4567 = vreinterpretq_s32_u32(vcltq_s32(vacc4567, vmovq_n_s32(0))); in xnn_qs8_gavgpool_minmax_ukernel_7x__neon_c16_acc2() local 219 const int64x2_t vadjprod45 = vaddw_s32(vprod45, vget_low_s32(vsgnacc4567)); in xnn_qs8_gavgpool_minmax_ukernel_7x__neon_c16_acc2() 220 const int64x2_t vadjprod67 = vaddw_high_s32(vprod67, vsgnacc4567); in xnn_qs8_gavgpool_minmax_ukernel_7x__neon_c16_acc2() 229 const int64x2_t vadjprod45 = vaddw_s32(vprod45, vget_low_s32(vsgnacc4567)); in xnn_qs8_gavgpool_minmax_ukernel_7x__neon_c16_acc2() 230 const int64x2_t vadjprod67 = vaddw_s32(vprod67, vget_high_s32(vsgnacc4567)); in xnn_qs8_gavgpool_minmax_ukernel_7x__neon_c16_acc2()
|
D | 7p7x-minmax-neon-c8-acc2.c | 171 const int32x4_t vsgnacc4567 = vreinterpretq_s32_u32(vcltq_s32(vacc4567, vmovq_n_s32(0))); in xnn_qs8_gavgpool_minmax_ukernel_7p7x__neon_c8_acc2() local 181 const int64x2_t vadjprod45 = vaddw_s32(vprod45, vget_low_s32(vsgnacc4567)); in xnn_qs8_gavgpool_minmax_ukernel_7p7x__neon_c8_acc2() 182 const int64x2_t vadjprod67 = vaddw_high_s32(vprod67, vsgnacc4567); in xnn_qs8_gavgpool_minmax_ukernel_7p7x__neon_c8_acc2() 191 const int64x2_t vadjprod45 = vaddw_s32(vprod45, vget_low_s32(vsgnacc4567)); in xnn_qs8_gavgpool_minmax_ukernel_7p7x__neon_c8_acc2() 192 const int64x2_t vadjprod67 = vaddw_s32(vprod67, vget_high_s32(vsgnacc4567)); in xnn_qs8_gavgpool_minmax_ukernel_7p7x__neon_c8_acc2() 251 const int32x4_t vsgnacc4567 = vreinterpretq_s32_u32(vcltq_s32(vacc4567, vmovq_n_s32(0))); in xnn_qs8_gavgpool_minmax_ukernel_7p7x__neon_c8_acc2() local 261 const int64x2_t vadjprod45 = vaddw_s32(vprod45, vget_low_s32(vsgnacc4567)); in xnn_qs8_gavgpool_minmax_ukernel_7p7x__neon_c8_acc2() 262 const int64x2_t vadjprod67 = vaddw_high_s32(vprod67, vsgnacc4567); in xnn_qs8_gavgpool_minmax_ukernel_7p7x__neon_c8_acc2() 271 const int64x2_t vadjprod45 = vaddw_s32(vprod45, vget_low_s32(vsgnacc4567)); in xnn_qs8_gavgpool_minmax_ukernel_7p7x__neon_c8_acc2() 272 const int64x2_t vadjprod67 = vaddw_s32(vprod67, vget_high_s32(vsgnacc4567)); in xnn_qs8_gavgpool_minmax_ukernel_7p7x__neon_c8_acc2()
|
D | 7x-minmax-sse2-c16-acc2.c | 122 const __m128i vsgnacc4567 = _mm_cmpgt_epi32(_mm_setzero_si128(), vacc4567); in xnn_qs8_gavgpool_minmax_ukernel_7x__sse2_c16_acc2() local 127 const __m128i vabsacc4567 = _mm_sub_epi32(_mm_xor_si128(vacc4567, vsgnacc4567), vsgnacc4567); in xnn_qs8_gavgpool_minmax_ukernel_7x__sse2_c16_acc2() 169 const __m128i vout4567 = _mm_sub_epi32(_mm_xor_si128(vabsout4567, vsgnacc4567), vsgnacc4567); in xnn_qs8_gavgpool_minmax_ukernel_7x__sse2_c16_acc2() 229 const __m128i vsgnacc4567 = _mm_cmpgt_epi32(_mm_setzero_si128(), vacc4567); in xnn_qs8_gavgpool_minmax_ukernel_7x__sse2_c16_acc2() local 232 const __m128i vabsacc4567 = _mm_sub_epi32(_mm_xor_si128(vacc4567, vsgnacc4567), vsgnacc4567); in xnn_qs8_gavgpool_minmax_ukernel_7x__sse2_c16_acc2() 256 const __m128i vout4567 = _mm_sub_epi32(_mm_xor_si128(vabsout4567, vsgnacc4567), vsgnacc4567); in xnn_qs8_gavgpool_minmax_ukernel_7x__sse2_c16_acc2()
|
D | 7x-minmax-wasmsimd-c8-acc2.c | 94 const v128_t vsgnacc4567 = wasm_i32x4_gt(vabsacc4567, vacc4567); in xnn_qs8_gavgpool_minmax_ukernel_7x__wasmsimd_c8_acc2() local 115 const v128_t vout4567 = wasm_i32x4_sub(wasm_v128_xor(vabsout4567, vsgnacc4567), vsgnacc4567); in xnn_qs8_gavgpool_minmax_ukernel_7x__wasmsimd_c8_acc2() 163 const v128_t vsgnacc4567 = wasm_i32x4_gt(vabsacc4567, vacc4567); in xnn_qs8_gavgpool_minmax_ukernel_7x__wasmsimd_c8_acc2() local 184 const v128_t vout4567 = wasm_i32x4_sub(wasm_v128_xor(vabsout4567, vsgnacc4567), vsgnacc4567); in xnn_qs8_gavgpool_minmax_ukernel_7x__wasmsimd_c8_acc2()
|
D | 7p7x-minmax-sse2-c8-acc2.c | 210 const __m128i vsgnacc4567 = _mm_cmpgt_epi32(_mm_setzero_si128(), vacc4567); in xnn_qs8_gavgpool_minmax_ukernel_7p7x__sse2_c8_acc2() local 213 const __m128i vabsacc4567 = _mm_sub_epi32(_mm_xor_si128(vacc4567, vsgnacc4567), vsgnacc4567); in xnn_qs8_gavgpool_minmax_ukernel_7p7x__sse2_c8_acc2() 237 const __m128i vout4567 = _mm_sub_epi32(_mm_xor_si128(vabsout4567, vsgnacc4567), vsgnacc4567); in xnn_qs8_gavgpool_minmax_ukernel_7p7x__sse2_c8_acc2() 294 const __m128i vsgnacc4567 = _mm_cmpgt_epi32(_mm_setzero_si128(), vacc4567); in xnn_qs8_gavgpool_minmax_ukernel_7p7x__sse2_c8_acc2() local 297 const __m128i vabsacc4567 = _mm_sub_epi32(_mm_xor_si128(vacc4567, vsgnacc4567), vsgnacc4567); in xnn_qs8_gavgpool_minmax_ukernel_7p7x__sse2_c8_acc2() 321 const __m128i vout4567 = _mm_sub_epi32(_mm_xor_si128(vabsout4567, vsgnacc4567), vsgnacc4567); in xnn_qs8_gavgpool_minmax_ukernel_7p7x__sse2_c8_acc2()
|
D | 7x-minmax-sse2-c24-acc2.c | 145 const __m128i vsgnacc4567 = _mm_cmpgt_epi32(_mm_setzero_si128(), vacc4567); in xnn_qs8_gavgpool_minmax_ukernel_7x__sse2_c24_acc2() local 152 const __m128i vabsacc4567 = _mm_sub_epi32(_mm_xor_si128(vacc4567, vsgnacc4567), vsgnacc4567); in xnn_qs8_gavgpool_minmax_ukernel_7x__sse2_c24_acc2() 212 const __m128i vout4567 = _mm_sub_epi32(_mm_xor_si128(vabsout4567, vsgnacc4567), vsgnacc4567); in xnn_qs8_gavgpool_minmax_ukernel_7x__sse2_c24_acc2() 278 const __m128i vsgnacc4567 = _mm_cmpgt_epi32(_mm_setzero_si128(), vacc4567); in xnn_qs8_gavgpool_minmax_ukernel_7x__sse2_c24_acc2() local 281 const __m128i vabsacc4567 = _mm_sub_epi32(_mm_xor_si128(vacc4567, vsgnacc4567), vsgnacc4567); in xnn_qs8_gavgpool_minmax_ukernel_7x__sse2_c24_acc2() 305 const __m128i vout4567 = _mm_sub_epi32(_mm_xor_si128(vabsout4567, vsgnacc4567), vsgnacc4567); in xnn_qs8_gavgpool_minmax_ukernel_7x__sse2_c24_acc2()
|
D | 7x-minmax-neon-c24-acc2.c | 119 const int32x4_t vsgnacc4567 = vreinterpretq_s32_u32(vcltq_s32(vacc4567, vmovq_n_s32(0))); in xnn_qs8_gavgpool_minmax_ukernel_7x__neon_c24_acc2() local 141 const int64x2_t vadjprod45 = vaddw_s32(vprod45, vget_low_s32(vsgnacc4567)); in xnn_qs8_gavgpool_minmax_ukernel_7x__neon_c24_acc2() 142 const int64x2_t vadjprod67 = vaddw_high_s32(vprod67, vsgnacc4567); in xnn_qs8_gavgpool_minmax_ukernel_7x__neon_c24_acc2() 167 const int64x2_t vadjprod45 = vaddw_s32(vprod45, vget_low_s32(vsgnacc4567)); in xnn_qs8_gavgpool_minmax_ukernel_7x__neon_c24_acc2() 168 const int64x2_t vadjprod67 = vaddw_s32(vprod67, vget_high_s32(vsgnacc4567)); in xnn_qs8_gavgpool_minmax_ukernel_7x__neon_c24_acc2() 257 const int32x4_t vsgnacc4567 = vreinterpretq_s32_u32(vcltq_s32(vacc4567, vmovq_n_s32(0))); in xnn_qs8_gavgpool_minmax_ukernel_7x__neon_c24_acc2() local 267 const int64x2_t vadjprod45 = vaddw_s32(vprod45, vget_low_s32(vsgnacc4567)); in xnn_qs8_gavgpool_minmax_ukernel_7x__neon_c24_acc2() 268 const int64x2_t vadjprod67 = vaddw_high_s32(vprod67, vsgnacc4567); in xnn_qs8_gavgpool_minmax_ukernel_7x__neon_c24_acc2() 277 const int64x2_t vadjprod45 = vaddw_s32(vprod45, vget_low_s32(vsgnacc4567)); in xnn_qs8_gavgpool_minmax_ukernel_7x__neon_c24_acc2() 278 const int64x2_t vadjprod67 = vaddw_s32(vprod67, vget_high_s32(vsgnacc4567)); in xnn_qs8_gavgpool_minmax_ukernel_7x__neon_c24_acc2()
|
D | 7p7x-minmax-neon-c16-acc2.c | 224 const int32x4_t vsgnacc4567 = vreinterpretq_s32_u32(vcltq_s32(vacc4567, vmovq_n_s32(0))); in xnn_qs8_gavgpool_minmax_ukernel_7p7x__neon_c16_acc2() local 240 const int64x2_t vadjprod45 = vaddw_s32(vprod45, vget_low_s32(vsgnacc4567)); in xnn_qs8_gavgpool_minmax_ukernel_7p7x__neon_c16_acc2() 241 const int64x2_t vadjprod67 = vaddw_high_s32(vprod67, vsgnacc4567); in xnn_qs8_gavgpool_minmax_ukernel_7p7x__neon_c16_acc2() 258 const int64x2_t vadjprod45 = vaddw_s32(vprod45, vget_low_s32(vsgnacc4567)); in xnn_qs8_gavgpool_minmax_ukernel_7p7x__neon_c16_acc2() 259 const int64x2_t vadjprod67 = vaddw_s32(vprod67, vget_high_s32(vsgnacc4567)); in xnn_qs8_gavgpool_minmax_ukernel_7p7x__neon_c16_acc2() 332 const int32x4_t vsgnacc4567 = vreinterpretq_s32_u32(vcltq_s32(vacc4567, vmovq_n_s32(0))); in xnn_qs8_gavgpool_minmax_ukernel_7p7x__neon_c16_acc2() local 342 const int64x2_t vadjprod45 = vaddw_s32(vprod45, vget_low_s32(vsgnacc4567)); in xnn_qs8_gavgpool_minmax_ukernel_7p7x__neon_c16_acc2() 343 const int64x2_t vadjprod67 = vaddw_high_s32(vprod67, vsgnacc4567); in xnn_qs8_gavgpool_minmax_ukernel_7p7x__neon_c16_acc2() 352 const int64x2_t vadjprod45 = vaddw_s32(vprod45, vget_low_s32(vsgnacc4567)); in xnn_qs8_gavgpool_minmax_ukernel_7p7x__neon_c16_acc2() 353 const int64x2_t vadjprod67 = vaddw_s32(vprod67, vget_high_s32(vsgnacc4567)); in xnn_qs8_gavgpool_minmax_ukernel_7p7x__neon_c16_acc2()
|
D | 7x-minmax-wasmsimd-c16-acc2.c | 111 const v128_t vsgnacc4567 = wasm_i32x4_gt(vabsacc4567, vacc4567); in xnn_qs8_gavgpool_minmax_ukernel_7x__wasmsimd_c16_acc2() local 148 const v128_t vout4567 = wasm_i32x4_sub(wasm_v128_xor(vabsout4567, vsgnacc4567), vsgnacc4567); in xnn_qs8_gavgpool_minmax_ukernel_7x__wasmsimd_c16_acc2() 199 const v128_t vsgnacc4567 = wasm_i32x4_gt(vabsacc4567, vacc4567); in xnn_qs8_gavgpool_minmax_ukernel_7x__wasmsimd_c16_acc2() local 220 const v128_t vout4567 = wasm_i32x4_sub(wasm_v128_xor(vabsout4567, vsgnacc4567), vsgnacc4567); in xnn_qs8_gavgpool_minmax_ukernel_7x__wasmsimd_c16_acc2()
|
D | 7x-minmax-neon-c32-acc2.c | 134 const int32x4_t vsgnacc4567 = vreinterpretq_s32_u32(vcltq_s32(vacc4567, vmovq_n_s32(0))); in xnn_qs8_gavgpool_minmax_ukernel_7x__neon_c32_acc2() local 162 const int64x2_t vadjprod45 = vaddw_s32(vprod45, vget_low_s32(vsgnacc4567)); in xnn_qs8_gavgpool_minmax_ukernel_7x__neon_c32_acc2() 163 const int64x2_t vadjprod67 = vaddw_high_s32(vprod67, vsgnacc4567); in xnn_qs8_gavgpool_minmax_ukernel_7x__neon_c32_acc2() 196 const int64x2_t vadjprod45 = vaddw_s32(vprod45, vget_low_s32(vsgnacc4567)); in xnn_qs8_gavgpool_minmax_ukernel_7x__neon_c32_acc2() 197 const int64x2_t vadjprod67 = vaddw_s32(vprod67, vget_high_s32(vsgnacc4567)); in xnn_qs8_gavgpool_minmax_ukernel_7x__neon_c32_acc2() 300 const int32x4_t vsgnacc4567 = vreinterpretq_s32_u32(vcltq_s32(vacc4567, vmovq_n_s32(0))); in xnn_qs8_gavgpool_minmax_ukernel_7x__neon_c32_acc2() local 310 const int64x2_t vadjprod45 = vaddw_s32(vprod45, vget_low_s32(vsgnacc4567)); in xnn_qs8_gavgpool_minmax_ukernel_7x__neon_c32_acc2() 311 const int64x2_t vadjprod67 = vaddw_high_s32(vprod67, vsgnacc4567); in xnn_qs8_gavgpool_minmax_ukernel_7x__neon_c32_acc2() 320 const int64x2_t vadjprod45 = vaddw_s32(vprod45, vget_low_s32(vsgnacc4567)); in xnn_qs8_gavgpool_minmax_ukernel_7x__neon_c32_acc2() 321 const int64x2_t vadjprod67 = vaddw_s32(vprod67, vget_high_s32(vsgnacc4567)); in xnn_qs8_gavgpool_minmax_ukernel_7x__neon_c32_acc2()
|
D | 7p7x-minmax-wasmsimd-c8-acc2.c | 187 const v128_t vsgnacc4567 = wasm_i32x4_gt(vabsacc4567, vacc4567); in xnn_qs8_gavgpool_minmax_ukernel_7p7x__wasmsimd_c8_acc2() local 208 const v128_t vout4567 = wasm_i32x4_sub(wasm_v128_xor(vabsout4567, vsgnacc4567), vsgnacc4567); in xnn_qs8_gavgpool_minmax_ukernel_7p7x__wasmsimd_c8_acc2() 257 const v128_t vsgnacc4567 = wasm_i32x4_gt(vabsacc4567, vacc4567); in xnn_qs8_gavgpool_minmax_ukernel_7p7x__wasmsimd_c8_acc2() local 278 const v128_t vout4567 = wasm_i32x4_sub(wasm_v128_xor(vabsout4567, vsgnacc4567), vsgnacc4567); in xnn_qs8_gavgpool_minmax_ukernel_7p7x__wasmsimd_c8_acc2()
|
D | 7x-minmax-wasmsimd-c24-acc2.c | 128 const v128_t vsgnacc4567 = wasm_i32x4_gt(vabsacc4567, vacc4567); in xnn_qs8_gavgpool_minmax_ukernel_7x__wasmsimd_c24_acc2() local 181 const v128_t vout4567 = wasm_i32x4_sub(wasm_v128_xor(vabsout4567, vsgnacc4567), vsgnacc4567); in xnn_qs8_gavgpool_minmax_ukernel_7x__wasmsimd_c24_acc2() 237 const v128_t vsgnacc4567 = wasm_i32x4_gt(vabsacc4567, vacc4567); in xnn_qs8_gavgpool_minmax_ukernel_7x__wasmsimd_c24_acc2() local 258 const v128_t vout4567 = wasm_i32x4_sub(wasm_v128_xor(vabsout4567, vsgnacc4567), vsgnacc4567); in xnn_qs8_gavgpool_minmax_ukernel_7x__wasmsimd_c24_acc2()
|
D | 7p7x-minmax-sse2-c16-acc2.c | 283 const __m128i vsgnacc4567 = _mm_cmpgt_epi32(_mm_setzero_si128(), vacc4567); in xnn_qs8_gavgpool_minmax_ukernel_7p7x__sse2_c16_acc2() local 288 const __m128i vabsacc4567 = _mm_sub_epi32(_mm_xor_si128(vacc4567, vsgnacc4567), vsgnacc4567); in xnn_qs8_gavgpool_minmax_ukernel_7p7x__sse2_c16_acc2() 330 const __m128i vout4567 = _mm_sub_epi32(_mm_xor_si128(vabsout4567, vsgnacc4567), vsgnacc4567); in xnn_qs8_gavgpool_minmax_ukernel_7p7x__sse2_c16_acc2() 391 const __m128i vsgnacc4567 = _mm_cmpgt_epi32(_mm_setzero_si128(), vacc4567); in xnn_qs8_gavgpool_minmax_ukernel_7p7x__sse2_c16_acc2() local 394 const __m128i vabsacc4567 = _mm_sub_epi32(_mm_xor_si128(vacc4567, vsgnacc4567), vsgnacc4567); in xnn_qs8_gavgpool_minmax_ukernel_7p7x__sse2_c16_acc2() 418 const __m128i vout4567 = _mm_sub_epi32(_mm_xor_si128(vabsout4567, vsgnacc4567), vsgnacc4567); in xnn_qs8_gavgpool_minmax_ukernel_7p7x__sse2_c16_acc2()
|
D | 7p7x-minmax-neon-c24-acc2.c | 338 const int32x4_t vsgnacc4567 = vreinterpretq_s32_u32(vcltq_s32(vacc4567, vmovq_n_s32(0))); in xnn_qs8_gavgpool_minmax_ukernel_7p7x__neon_c24_acc2() local 360 const int64x2_t vadjprod45 = vaddw_s32(vprod45, vget_low_s32(vsgnacc4567)); in xnn_qs8_gavgpool_minmax_ukernel_7p7x__neon_c24_acc2() 361 const int64x2_t vadjprod67 = vaddw_high_s32(vprod67, vsgnacc4567); in xnn_qs8_gavgpool_minmax_ukernel_7p7x__neon_c24_acc2() 386 const int64x2_t vadjprod45 = vaddw_s32(vprod45, vget_low_s32(vsgnacc4567)); in xnn_qs8_gavgpool_minmax_ukernel_7p7x__neon_c24_acc2() 387 const int64x2_t vadjprod67 = vaddw_s32(vprod67, vget_high_s32(vsgnacc4567)); in xnn_qs8_gavgpool_minmax_ukernel_7p7x__neon_c24_acc2() 479 const int32x4_t vsgnacc4567 = vreinterpretq_s32_u32(vcltq_s32(vacc4567, vmovq_n_s32(0))); in xnn_qs8_gavgpool_minmax_ukernel_7p7x__neon_c24_acc2() local 489 const int64x2_t vadjprod45 = vaddw_s32(vprod45, vget_low_s32(vsgnacc4567)); in xnn_qs8_gavgpool_minmax_ukernel_7p7x__neon_c24_acc2() 490 const int64x2_t vadjprod67 = vaddw_high_s32(vprod67, vsgnacc4567); in xnn_qs8_gavgpool_minmax_ukernel_7p7x__neon_c24_acc2() 499 const int64x2_t vadjprod45 = vaddw_s32(vprod45, vget_low_s32(vsgnacc4567)); in xnn_qs8_gavgpool_minmax_ukernel_7p7x__neon_c24_acc2() 500 const int64x2_t vadjprod67 = vaddw_s32(vprod67, vget_high_s32(vsgnacc4567)); in xnn_qs8_gavgpool_minmax_ukernel_7p7x__neon_c24_acc2()
|
D | 7p7x-minmax-wasmsimd-c16-acc2.c | 238 const v128_t vsgnacc4567 = wasm_i32x4_gt(vabsacc4567, vacc4567); in xnn_qs8_gavgpool_minmax_ukernel_7p7x__wasmsimd_c16_acc2() local 275 const v128_t vout4567 = wasm_i32x4_sub(wasm_v128_xor(vabsout4567, vsgnacc4567), vsgnacc4567); in xnn_qs8_gavgpool_minmax_ukernel_7p7x__wasmsimd_c16_acc2() 327 const v128_t vsgnacc4567 = wasm_i32x4_gt(vabsacc4567, vacc4567); in xnn_qs8_gavgpool_minmax_ukernel_7p7x__wasmsimd_c16_acc2() local 348 const v128_t vout4567 = wasm_i32x4_sub(wasm_v128_xor(vabsout4567, vsgnacc4567), vsgnacc4567); in xnn_qs8_gavgpool_minmax_ukernel_7p7x__wasmsimd_c16_acc2()
|
D | 7p7x-minmax-sse2-c24-acc2.c | 448 const __m128i vsgnacc4567 = _mm_cmpgt_epi32(_mm_setzero_si128(), vacc4567); in xnn_qs8_gavgpool_minmax_ukernel_7p7x__sse2_c24_acc2() local 455 const __m128i vabsacc4567 = _mm_sub_epi32(_mm_xor_si128(vacc4567, vsgnacc4567), vsgnacc4567); in xnn_qs8_gavgpool_minmax_ukernel_7p7x__sse2_c24_acc2() 515 const __m128i vout4567 = _mm_sub_epi32(_mm_xor_si128(vabsout4567, vsgnacc4567), vsgnacc4567); in xnn_qs8_gavgpool_minmax_ukernel_7p7x__sse2_c24_acc2() 582 const __m128i vsgnacc4567 = _mm_cmpgt_epi32(_mm_setzero_si128(), vacc4567); in xnn_qs8_gavgpool_minmax_ukernel_7p7x__sse2_c24_acc2() local 585 const __m128i vabsacc4567 = _mm_sub_epi32(_mm_xor_si128(vacc4567, vsgnacc4567), vsgnacc4567); in xnn_qs8_gavgpool_minmax_ukernel_7p7x__sse2_c24_acc2() 609 const __m128i vout4567 = _mm_sub_epi32(_mm_xor_si128(vabsout4567, vsgnacc4567), vsgnacc4567); in xnn_qs8_gavgpool_minmax_ukernel_7p7x__sse2_c24_acc2()
|
D | 7p7x-minmax-neon-c32-acc2.c | 391 const int32x4_t vsgnacc4567 = vreinterpretq_s32_u32(vcltq_s32(vacc4567, vmovq_n_s32(0))); in xnn_qs8_gavgpool_minmax_ukernel_7p7x__neon_c32_acc2() local 419 const int64x2_t vadjprod45 = vaddw_s32(vprod45, vget_low_s32(vsgnacc4567)); in xnn_qs8_gavgpool_minmax_ukernel_7p7x__neon_c32_acc2() 420 const int64x2_t vadjprod67 = vaddw_high_s32(vprod67, vsgnacc4567); in xnn_qs8_gavgpool_minmax_ukernel_7p7x__neon_c32_acc2() 453 const int64x2_t vadjprod45 = vaddw_s32(vprod45, vget_low_s32(vsgnacc4567)); in xnn_qs8_gavgpool_minmax_ukernel_7p7x__neon_c32_acc2() 454 const int64x2_t vadjprod67 = vaddw_s32(vprod67, vget_high_s32(vsgnacc4567)); in xnn_qs8_gavgpool_minmax_ukernel_7p7x__neon_c32_acc2() 560 const int32x4_t vsgnacc4567 = vreinterpretq_s32_u32(vcltq_s32(vacc4567, vmovq_n_s32(0))); in xnn_qs8_gavgpool_minmax_ukernel_7p7x__neon_c32_acc2() local 570 const int64x2_t vadjprod45 = vaddw_s32(vprod45, vget_low_s32(vsgnacc4567)); in xnn_qs8_gavgpool_minmax_ukernel_7p7x__neon_c32_acc2() 571 const int64x2_t vadjprod67 = vaddw_high_s32(vprod67, vsgnacc4567); in xnn_qs8_gavgpool_minmax_ukernel_7p7x__neon_c32_acc2() 580 const int64x2_t vadjprod45 = vaddw_s32(vprod45, vget_low_s32(vsgnacc4567)); in xnn_qs8_gavgpool_minmax_ukernel_7p7x__neon_c32_acc2() 581 const int64x2_t vadjprod67 = vaddw_s32(vprod67, vget_high_s32(vsgnacc4567)); in xnn_qs8_gavgpool_minmax_ukernel_7p7x__neon_c32_acc2()
|
D | 7p7x-minmax-wasmsimd-c24-acc2.c | 363 const v128_t vsgnacc4567 = wasm_i32x4_gt(vabsacc4567, vacc4567); in xnn_qs8_gavgpool_minmax_ukernel_7p7x__wasmsimd_c24_acc2() local 416 const v128_t vout4567 = wasm_i32x4_sub(wasm_v128_xor(vabsout4567, vsgnacc4567), vsgnacc4567); in xnn_qs8_gavgpool_minmax_ukernel_7p7x__wasmsimd_c24_acc2() 473 const v128_t vsgnacc4567 = wasm_i32x4_gt(vabsacc4567, vacc4567); in xnn_qs8_gavgpool_minmax_ukernel_7p7x__wasmsimd_c24_acc2() local 494 const v128_t vout4567 = wasm_i32x4_sub(wasm_v128_xor(vabsout4567, vsgnacc4567), vsgnacc4567); in xnn_qs8_gavgpool_minmax_ukernel_7p7x__wasmsimd_c24_acc2()
|