/external/XNNPACK/src/qs8-gavgpool/gen/ |
D | 7x-minmax-wasmsimd-c8-acc2.c | 112 const v128_t vabsout4567 = wasm_v32x4_shuffle(vabsout45, vabsout67, 0, 2, 4, 6); 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() 181 const v128_t vabsout4567 = wasm_v32x4_shuffle(vabsout45, vabsout67, 0, 2, 4, 6); 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 | 7x-minmax-sse41-c8-acc2.c | 107 …const __m128i vabsout4567 = _mm_blend_epi16(vabsout46, _mm_shuffle_epi32(vabsout57, _MM_SHUFFLE(2,… in xnn_qs8_gavgpool_minmax_ukernel_7x__sse41_c8_acc2() local 110 const __m128i vout4567 = _mm_sign_epi32(vabsout4567, vacc4567); in xnn_qs8_gavgpool_minmax_ukernel_7x__sse41_c8_acc2() 174 …const __m128i vabsout4567 = _mm_blend_epi16(vabsout46, _mm_shuffle_epi32(vabsout57, _MM_SHUFFLE(2,… in xnn_qs8_gavgpool_minmax_ukernel_7x__sse41_c8_acc2() local 177 const __m128i vout4567 = _mm_sign_epi32(vabsout4567, vacc4567); in xnn_qs8_gavgpool_minmax_ukernel_7x__sse41_c8_acc2()
|
D | 7x-minmax-ssse3-c8-acc2.c | 120 const __m128i vabsout4567 = _mm_shuffle_epi32(vabsout4657, _MM_SHUFFLE(3, 1, 2, 0)); in xnn_qs8_gavgpool_minmax_ukernel_7x__ssse3_c8_acc2() local 123 const __m128i vout4567 = _mm_sign_epi32(vabsout4567, vacc4567); in xnn_qs8_gavgpool_minmax_ukernel_7x__ssse3_c8_acc2() 200 const __m128i vabsout4567 = _mm_shuffle_epi32(vabsout4657, _MM_SHUFFLE(3, 1, 2, 0)); in xnn_qs8_gavgpool_minmax_ukernel_7x__ssse3_c8_acc2() local 203 const __m128i vout4567 = _mm_sign_epi32(vabsout4567, vacc4567); in xnn_qs8_gavgpool_minmax_ukernel_7x__ssse3_c8_acc2()
|
D | 7x-minmax-sse2-c8-acc2.c | 123 const __m128i vabsout4567 = _mm_shuffle_epi32(vabsout4657, _MM_SHUFFLE(3, 1, 2, 0)); in xnn_qs8_gavgpool_minmax_ukernel_7x__sse2_c8_acc2() local 126 const __m128i vout4567 = _mm_sub_epi32(_mm_xor_si128(vabsout4567, vsgnacc4567), vsgnacc4567); in xnn_qs8_gavgpool_minmax_ukernel_7x__sse2_c8_acc2() 206 const __m128i vabsout4567 = _mm_shuffle_epi32(vabsout4657, _MM_SHUFFLE(3, 1, 2, 0)); in xnn_qs8_gavgpool_minmax_ukernel_7x__sse2_c8_acc2() local 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-sse41-c16-acc2.c | 134 …const __m128i vabsout4567 = _mm_blend_epi16(vabsout46, _mm_shuffle_epi32(vabsout57, _MM_SHUFFLE(2,… in xnn_qs8_gavgpool_minmax_ukernel_7x__sse41_c16_acc2() local 139 const __m128i vout4567 = _mm_sign_epi32(vabsout4567, vacc4567); in xnn_qs8_gavgpool_minmax_ukernel_7x__sse41_c16_acc2() 207 …const __m128i vabsout4567 = _mm_blend_epi16(vabsout46, _mm_shuffle_epi32(vabsout57, _MM_SHUFFLE(2,… in xnn_qs8_gavgpool_minmax_ukernel_7x__sse41_c16_acc2() local 210 const __m128i vout4567 = _mm_sign_epi32(vabsout4567, vacc4567); in xnn_qs8_gavgpool_minmax_ukernel_7x__sse41_c16_acc2()
|
D | 7x-minmax-wasmsimd-c16-acc2.c | 143 const v128_t vabsout4567 = wasm_v32x4_shuffle(vabsout45, vabsout67, 0, 2, 4, 6); 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() 217 const v128_t vabsout4567 = wasm_v32x4_shuffle(vabsout45, vabsout67, 0, 2, 4, 6); 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 | 7p7x-minmax-wasmsimd-c8-acc2.c | 205 const v128_t vabsout4567 = wasm_v32x4_shuffle(vabsout45, vabsout67, 0, 2, 4, 6); 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() 275 const v128_t vabsout4567 = wasm_v32x4_shuffle(vabsout45, vabsout67, 0, 2, 4, 6); 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 | 7p7x-minmax-sse41-c8-acc2.c | 202 …const __m128i vabsout4567 = _mm_blend_epi16(vabsout46, _mm_shuffle_epi32(vabsout57, _MM_SHUFFLE(2,… in xnn_qs8_gavgpool_minmax_ukernel_7p7x__sse41_c8_acc2() local 205 const __m128i vout4567 = _mm_sign_epi32(vabsout4567, vacc4567); in xnn_qs8_gavgpool_minmax_ukernel_7p7x__sse41_c8_acc2() 270 …const __m128i vabsout4567 = _mm_blend_epi16(vabsout46, _mm_shuffle_epi32(vabsout57, _MM_SHUFFLE(2,… in xnn_qs8_gavgpool_minmax_ukernel_7p7x__sse41_c8_acc2() local 273 const __m128i vout4567 = _mm_sign_epi32(vabsout4567, vacc4567); in xnn_qs8_gavgpool_minmax_ukernel_7p7x__sse41_c8_acc2()
|
D | 7x-minmax-ssse3-c16-acc2.c | 159 const __m128i vabsout4567 = _mm_shuffle_epi32(vabsout4657, _MM_SHUFFLE(3, 1, 2, 0)); in xnn_qs8_gavgpool_minmax_ukernel_7x__ssse3_c16_acc2() local 164 const __m128i vout4567 = _mm_sign_epi32(vabsout4567, vacc4567); in xnn_qs8_gavgpool_minmax_ukernel_7x__ssse3_c16_acc2() 245 const __m128i vabsout4567 = _mm_shuffle_epi32(vabsout4657, _MM_SHUFFLE(3, 1, 2, 0)); in xnn_qs8_gavgpool_minmax_ukernel_7x__ssse3_c16_acc2() local 248 const __m128i vout4567 = _mm_sign_epi32(vabsout4567, vacc4567); in xnn_qs8_gavgpool_minmax_ukernel_7x__ssse3_c16_acc2()
|
D | 7x-minmax-sse41-c24-acc2.c | 161 …const __m128i vabsout4567 = _mm_blend_epi16(vabsout46, _mm_shuffle_epi32(vabsout57, _MM_SHUFFLE(2,… in xnn_qs8_gavgpool_minmax_ukernel_7x__sse41_c24_acc2() local 168 const __m128i vout4567 = _mm_sign_epi32(vabsout4567, vacc4567); in xnn_qs8_gavgpool_minmax_ukernel_7x__sse41_c24_acc2() 242 …const __m128i vabsout4567 = _mm_blend_epi16(vabsout46, _mm_shuffle_epi32(vabsout57, _MM_SHUFFLE(2,… in xnn_qs8_gavgpool_minmax_ukernel_7x__sse41_c24_acc2() local 245 const __m128i vout4567 = _mm_sign_epi32(vabsout4567, vacc4567); in xnn_qs8_gavgpool_minmax_ukernel_7x__sse41_c24_acc2()
|
D | 7x-minmax-wasmsimd-c24-acc2.c | 174 const v128_t vabsout4567 = wasm_v32x4_shuffle(vabsout45, vabsout67, 0, 2, 4, 6); 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() 255 const v128_t vabsout4567 = wasm_v32x4_shuffle(vabsout45, vabsout67, 0, 2, 4, 6); 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 | 7x-minmax-sse2-c16-acc2.c | 164 const __m128i vabsout4567 = _mm_shuffle_epi32(vabsout4657, _MM_SHUFFLE(3, 1, 2, 0)); in xnn_qs8_gavgpool_minmax_ukernel_7x__sse2_c16_acc2() local 169 const __m128i vout4567 = _mm_sub_epi32(_mm_xor_si128(vabsout4567, vsgnacc4567), vsgnacc4567); in xnn_qs8_gavgpool_minmax_ukernel_7x__sse2_c16_acc2() 253 const __m128i vabsout4567 = _mm_shuffle_epi32(vabsout4657, _MM_SHUFFLE(3, 1, 2, 0)); in xnn_qs8_gavgpool_minmax_ukernel_7x__sse2_c16_acc2() local 256 const __m128i vout4567 = _mm_sub_epi32(_mm_xor_si128(vabsout4567, vsgnacc4567), vsgnacc4567); in xnn_qs8_gavgpool_minmax_ukernel_7x__sse2_c16_acc2()
|
D | 7p7x-minmax-ssse3-c8-acc2.c | 231 const __m128i vabsout4567 = _mm_shuffle_epi32(vabsout4657, _MM_SHUFFLE(3, 1, 2, 0)); in xnn_qs8_gavgpool_minmax_ukernel_7p7x__ssse3_c8_acc2() local 234 const __m128i vout4567 = _mm_sign_epi32(vabsout4567, vacc4567); in xnn_qs8_gavgpool_minmax_ukernel_7p7x__ssse3_c8_acc2() 312 const __m128i vabsout4567 = _mm_shuffle_epi32(vabsout4657, _MM_SHUFFLE(3, 1, 2, 0)); in xnn_qs8_gavgpool_minmax_ukernel_7p7x__ssse3_c8_acc2() local 315 const __m128i vout4567 = _mm_sign_epi32(vabsout4567, vacc4567); in xnn_qs8_gavgpool_minmax_ukernel_7p7x__ssse3_c8_acc2()
|
D | 7p7x-minmax-sse2-c8-acc2.c | 234 const __m128i vabsout4567 = _mm_shuffle_epi32(vabsout4657, _MM_SHUFFLE(3, 1, 2, 0)); in xnn_qs8_gavgpool_minmax_ukernel_7p7x__sse2_c8_acc2() local 237 const __m128i vout4567 = _mm_sub_epi32(_mm_xor_si128(vabsout4567, vsgnacc4567), vsgnacc4567); in xnn_qs8_gavgpool_minmax_ukernel_7p7x__sse2_c8_acc2() 318 const __m128i vabsout4567 = _mm_shuffle_epi32(vabsout4657, _MM_SHUFFLE(3, 1, 2, 0)); in xnn_qs8_gavgpool_minmax_ukernel_7p7x__sse2_c8_acc2() local 321 const __m128i vout4567 = _mm_sub_epi32(_mm_xor_si128(vabsout4567, vsgnacc4567), vsgnacc4567); in xnn_qs8_gavgpool_minmax_ukernel_7p7x__sse2_c8_acc2()
|
D | 7p7x-minmax-wasmsimd-c16-acc2.c | 270 const v128_t vabsout4567 = wasm_v32x4_shuffle(vabsout45, vabsout67, 0, 2, 4, 6); 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() 345 const v128_t vabsout4567 = wasm_v32x4_shuffle(vabsout45, vabsout67, 0, 2, 4, 6); 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-sse41-c16-acc2.c | 263 …const __m128i vabsout4567 = _mm_blend_epi16(vabsout46, _mm_shuffle_epi32(vabsout57, _MM_SHUFFLE(2,… in xnn_qs8_gavgpool_minmax_ukernel_7p7x__sse41_c16_acc2() local 268 const __m128i vout4567 = _mm_sign_epi32(vabsout4567, vacc4567); in xnn_qs8_gavgpool_minmax_ukernel_7p7x__sse41_c16_acc2() 337 …const __m128i vabsout4567 = _mm_blend_epi16(vabsout46, _mm_shuffle_epi32(vabsout57, _MM_SHUFFLE(2,… in xnn_qs8_gavgpool_minmax_ukernel_7p7x__sse41_c16_acc2() local 340 const __m128i vout4567 = _mm_sign_epi32(vabsout4567, vacc4567); in xnn_qs8_gavgpool_minmax_ukernel_7p7x__sse41_c16_acc2()
|
D | 7x-minmax-sse2-c24-acc2.c | 205 const __m128i vabsout4567 = _mm_shuffle_epi32(vabsout4657, _MM_SHUFFLE(3, 1, 2, 0)); in xnn_qs8_gavgpool_minmax_ukernel_7x__sse2_c24_acc2() local 212 const __m128i vout4567 = _mm_sub_epi32(_mm_xor_si128(vabsout4567, vsgnacc4567), vsgnacc4567); in xnn_qs8_gavgpool_minmax_ukernel_7x__sse2_c24_acc2() 302 const __m128i vabsout4567 = _mm_shuffle_epi32(vabsout4657, _MM_SHUFFLE(3, 1, 2, 0)); in xnn_qs8_gavgpool_minmax_ukernel_7x__sse2_c24_acc2() local 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-ssse3-c24-acc2.c | 198 const __m128i vabsout4567 = _mm_shuffle_epi32(vabsout4657, _MM_SHUFFLE(3, 1, 2, 0)); in xnn_qs8_gavgpool_minmax_ukernel_7x__ssse3_c24_acc2() local 205 const __m128i vout4567 = _mm_sign_epi32(vabsout4567, vacc4567); in xnn_qs8_gavgpool_minmax_ukernel_7x__ssse3_c24_acc2() 292 const __m128i vabsout4567 = _mm_shuffle_epi32(vabsout4657, _MM_SHUFFLE(3, 1, 2, 0)); in xnn_qs8_gavgpool_minmax_ukernel_7x__ssse3_c24_acc2() local 295 const __m128i vout4567 = _mm_sign_epi32(vabsout4567, vacc4567); in xnn_qs8_gavgpool_minmax_ukernel_7x__ssse3_c24_acc2()
|
D | 7p7x-minmax-ssse3-c16-acc2.c | 320 const __m128i vabsout4567 = _mm_shuffle_epi32(vabsout4657, _MM_SHUFFLE(3, 1, 2, 0)); in xnn_qs8_gavgpool_minmax_ukernel_7p7x__ssse3_c16_acc2() local 325 const __m128i vout4567 = _mm_sign_epi32(vabsout4567, vacc4567); in xnn_qs8_gavgpool_minmax_ukernel_7p7x__ssse3_c16_acc2() 407 const __m128i vabsout4567 = _mm_shuffle_epi32(vabsout4657, _MM_SHUFFLE(3, 1, 2, 0)); in xnn_qs8_gavgpool_minmax_ukernel_7p7x__ssse3_c16_acc2() local 410 const __m128i vout4567 = _mm_sign_epi32(vabsout4567, vacc4567); in xnn_qs8_gavgpool_minmax_ukernel_7p7x__ssse3_c16_acc2()
|
D | 7p7x-minmax-sse2-c16-acc2.c | 325 const __m128i vabsout4567 = _mm_shuffle_epi32(vabsout4657, _MM_SHUFFLE(3, 1, 2, 0)); in xnn_qs8_gavgpool_minmax_ukernel_7p7x__sse2_c16_acc2() local 330 const __m128i vout4567 = _mm_sub_epi32(_mm_xor_si128(vabsout4567, vsgnacc4567), vsgnacc4567); in xnn_qs8_gavgpool_minmax_ukernel_7p7x__sse2_c16_acc2() 415 const __m128i vabsout4567 = _mm_shuffle_epi32(vabsout4657, _MM_SHUFFLE(3, 1, 2, 0)); in xnn_qs8_gavgpool_minmax_ukernel_7p7x__sse2_c16_acc2() local 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-wasmsimd-c24-acc2.c | 409 const v128_t vabsout4567 = wasm_v32x4_shuffle(vabsout45, vabsout67, 0, 2, 4, 6); 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() 491 const v128_t vabsout4567 = wasm_v32x4_shuffle(vabsout45, vabsout67, 0, 2, 4, 6); 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()
|
D | 7p7x-minmax-sse41-c24-acc2.c | 400 …const __m128i vabsout4567 = _mm_blend_epi16(vabsout46, _mm_shuffle_epi32(vabsout57, _MM_SHUFFLE(2,… in xnn_qs8_gavgpool_minmax_ukernel_7p7x__sse41_c24_acc2() local 407 const __m128i vout4567 = _mm_sign_epi32(vabsout4567, vacc4567); in xnn_qs8_gavgpool_minmax_ukernel_7p7x__sse41_c24_acc2() 482 …const __m128i vabsout4567 = _mm_blend_epi16(vabsout46, _mm_shuffle_epi32(vabsout57, _MM_SHUFFLE(2,… in xnn_qs8_gavgpool_minmax_ukernel_7p7x__sse41_c24_acc2() local 485 const __m128i vout4567 = _mm_sign_epi32(vabsout4567, vacc4567); in xnn_qs8_gavgpool_minmax_ukernel_7p7x__sse41_c24_acc2()
|
D | 7p7x-minmax-sse2-c24-acc2.c | 508 const __m128i vabsout4567 = _mm_shuffle_epi32(vabsout4657, _MM_SHUFFLE(3, 1, 2, 0)); in xnn_qs8_gavgpool_minmax_ukernel_7p7x__sse2_c24_acc2() local 515 const __m128i vout4567 = _mm_sub_epi32(_mm_xor_si128(vabsout4567, vsgnacc4567), vsgnacc4567); in xnn_qs8_gavgpool_minmax_ukernel_7p7x__sse2_c24_acc2() 606 const __m128i vabsout4567 = _mm_shuffle_epi32(vabsout4657, _MM_SHUFFLE(3, 1, 2, 0)); in xnn_qs8_gavgpool_minmax_ukernel_7p7x__sse2_c24_acc2() local 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-ssse3-c24-acc2.c | 501 const __m128i vabsout4567 = _mm_shuffle_epi32(vabsout4657, _MM_SHUFFLE(3, 1, 2, 0)); in xnn_qs8_gavgpool_minmax_ukernel_7p7x__ssse3_c24_acc2() local 508 const __m128i vout4567 = _mm_sign_epi32(vabsout4567, vacc4567); in xnn_qs8_gavgpool_minmax_ukernel_7p7x__ssse3_c24_acc2() 596 const __m128i vabsout4567 = _mm_shuffle_epi32(vabsout4657, _MM_SHUFFLE(3, 1, 2, 0)); in xnn_qs8_gavgpool_minmax_ukernel_7p7x__ssse3_c24_acc2() local 599 const __m128i vout4567 = _mm_sign_epi32(vabsout4567, vacc4567); in xnn_qs8_gavgpool_minmax_ukernel_7p7x__ssse3_c24_acc2()
|