| /external/XNNPACK/src/qs8-gavgpool/gen/ |
| D | 7x-minmax-wasmsimd-c8-acc2.c | 90 const v128_t vabsacc0123 = wasm_i32x4_abs(vacc0123); in xnn_qs8_gavgpool_minmax_ukernel_7x__wasmsimd_c8_acc2() local
93 const v128_t vsgnacc0123 = wasm_i32x4_gt(vabsacc0123, vacc0123); in xnn_qs8_gavgpool_minmax_ukernel_7x__wasmsimd_c8_acc2()
96 const v128_t vabsacc01 = wasm_v32x4_shuffle(vabsacc0123, vzero, 0, 4, 1, 5); in xnn_qs8_gavgpool_minmax_ukernel_7x__wasmsimd_c8_acc2()
97 const v128_t vabsacc23 = wasm_v32x4_shuffle(vabsacc0123, vzero, 2, 6, 3, 7); in xnn_qs8_gavgpool_minmax_ukernel_7x__wasmsimd_c8_acc2()
159 const v128_t vabsacc0123 = wasm_i32x4_abs(vacc0123); in xnn_qs8_gavgpool_minmax_ukernel_7x__wasmsimd_c8_acc2() local
162 const v128_t vsgnacc0123 = wasm_i32x4_gt(vabsacc0123, vacc0123); in xnn_qs8_gavgpool_minmax_ukernel_7x__wasmsimd_c8_acc2()
165 const v128_t vabsacc01 = wasm_v32x4_shuffle(vabsacc0123, vzero, 0, 4, 1, 5); in xnn_qs8_gavgpool_minmax_ukernel_7x__wasmsimd_c8_acc2()
166 const v128_t vabsacc23 = wasm_v32x4_shuffle(vabsacc0123, vzero, 2, 6, 3, 7); in xnn_qs8_gavgpool_minmax_ukernel_7x__wasmsimd_c8_acc2()
|
| D | 7x-minmax-wasmsimd-c16-acc2.c | 105 const v128_t vabsacc0123 = wasm_i32x4_abs(vacc0123); in xnn_qs8_gavgpool_minmax_ukernel_7x__wasmsimd_c16_acc2() local
110 const v128_t vsgnacc0123 = wasm_i32x4_gt(vabsacc0123, vacc0123); in xnn_qs8_gavgpool_minmax_ukernel_7x__wasmsimd_c16_acc2()
115 const v128_t vabsacc01 = wasm_v32x4_shuffle(vabsacc0123, vzero, 0, 4, 1, 5); in xnn_qs8_gavgpool_minmax_ukernel_7x__wasmsimd_c16_acc2()
116 const v128_t vabsacc23 = wasm_v32x4_shuffle(vabsacc0123, vzero, 2, 6, 3, 7); in xnn_qs8_gavgpool_minmax_ukernel_7x__wasmsimd_c16_acc2()
195 const v128_t vabsacc0123 = wasm_i32x4_abs(vacc0123); in xnn_qs8_gavgpool_minmax_ukernel_7x__wasmsimd_c16_acc2() local
198 const v128_t vsgnacc0123 = wasm_i32x4_gt(vabsacc0123, vacc0123); in xnn_qs8_gavgpool_minmax_ukernel_7x__wasmsimd_c16_acc2()
201 const v128_t vabsacc01 = wasm_v32x4_shuffle(vabsacc0123, vzero, 0, 4, 1, 5); in xnn_qs8_gavgpool_minmax_ukernel_7x__wasmsimd_c16_acc2()
202 const v128_t vabsacc23 = wasm_v32x4_shuffle(vabsacc0123, vzero, 2, 6, 3, 7); in xnn_qs8_gavgpool_minmax_ukernel_7x__wasmsimd_c16_acc2()
|
| D | 7x-minmax-sse41-c8-acc2.c | 90 const __m128i vabsacc0123 = _mm_abs_epi32(vacc0123); in xnn_qs8_gavgpool_minmax_ukernel_7x__sse41_c8_acc2() local
93 const __m128i vabsacc13 = _mm_shuffle_epi32(vabsacc0123, _MM_SHUFFLE(3, 3, 1, 1)); in xnn_qs8_gavgpool_minmax_ukernel_7x__sse41_c8_acc2()
96 const __m128i vabsprod02 = _mm_mul_epu32(vabsacc0123, vmultiplier); in xnn_qs8_gavgpool_minmax_ukernel_7x__sse41_c8_acc2()
157 const __m128i vabsacc0123 = _mm_abs_epi32(vacc0123); in xnn_qs8_gavgpool_minmax_ukernel_7x__sse41_c8_acc2() local
160 const __m128i vabsacc13 = _mm_shuffle_epi32(vabsacc0123, _MM_SHUFFLE(3, 3, 1, 1)); in xnn_qs8_gavgpool_minmax_ukernel_7x__sse41_c8_acc2()
163 const __m128i vabsprod02 = _mm_mul_epu32(vabsacc0123, vmultiplier); in xnn_qs8_gavgpool_minmax_ukernel_7x__sse41_c8_acc2()
|
| D | 7p7x-minmax-wasmsimd-c8-acc2.c | 183 const v128_t vabsacc0123 = wasm_i32x4_abs(vacc0123); in xnn_qs8_gavgpool_minmax_ukernel_7p7x__wasmsimd_c8_acc2() local
186 const v128_t vsgnacc0123 = wasm_i32x4_gt(vabsacc0123, vacc0123); in xnn_qs8_gavgpool_minmax_ukernel_7p7x__wasmsimd_c8_acc2()
189 const v128_t vabsacc01 = wasm_v32x4_shuffle(vabsacc0123, vzero, 0, 4, 1, 5); in xnn_qs8_gavgpool_minmax_ukernel_7p7x__wasmsimd_c8_acc2()
190 const v128_t vabsacc23 = wasm_v32x4_shuffle(vabsacc0123, vzero, 2, 6, 3, 7); in xnn_qs8_gavgpool_minmax_ukernel_7p7x__wasmsimd_c8_acc2()
253 const v128_t vabsacc0123 = wasm_i32x4_abs(vacc0123); in xnn_qs8_gavgpool_minmax_ukernel_7p7x__wasmsimd_c8_acc2() local
256 const v128_t vsgnacc0123 = wasm_i32x4_gt(vabsacc0123, vacc0123); in xnn_qs8_gavgpool_minmax_ukernel_7p7x__wasmsimd_c8_acc2()
259 const v128_t vabsacc01 = wasm_v32x4_shuffle(vabsacc0123, vzero, 0, 4, 1, 5); in xnn_qs8_gavgpool_minmax_ukernel_7p7x__wasmsimd_c8_acc2()
260 const v128_t vabsacc23 = wasm_v32x4_shuffle(vabsacc0123, vzero, 2, 6, 3, 7); in xnn_qs8_gavgpool_minmax_ukernel_7p7x__wasmsimd_c8_acc2()
|
| D | 7x-minmax-wasmsimd-c24-acc2.c | 120 const v128_t vabsacc0123 = wasm_i32x4_abs(vacc0123); in xnn_qs8_gavgpool_minmax_ukernel_7x__wasmsimd_c24_acc2() local
127 const v128_t vsgnacc0123 = wasm_i32x4_gt(vabsacc0123, vacc0123); in xnn_qs8_gavgpool_minmax_ukernel_7x__wasmsimd_c24_acc2()
134 const v128_t vabsacc01 = wasm_v32x4_shuffle(vabsacc0123, vzero, 0, 4, 1, 5); in xnn_qs8_gavgpool_minmax_ukernel_7x__wasmsimd_c24_acc2()
135 const v128_t vabsacc23 = wasm_v32x4_shuffle(vabsacc0123, vzero, 2, 6, 3, 7); in xnn_qs8_gavgpool_minmax_ukernel_7x__wasmsimd_c24_acc2()
233 const v128_t vabsacc0123 = wasm_i32x4_abs(vacc0123); in xnn_qs8_gavgpool_minmax_ukernel_7x__wasmsimd_c24_acc2() local
236 const v128_t vsgnacc0123 = wasm_i32x4_gt(vabsacc0123, vacc0123); in xnn_qs8_gavgpool_minmax_ukernel_7x__wasmsimd_c24_acc2()
239 const v128_t vabsacc01 = wasm_v32x4_shuffle(vabsacc0123, vzero, 0, 4, 1, 5); in xnn_qs8_gavgpool_minmax_ukernel_7x__wasmsimd_c24_acc2()
240 const v128_t vabsacc23 = wasm_v32x4_shuffle(vabsacc0123, vzero, 2, 6, 3, 7); in xnn_qs8_gavgpool_minmax_ukernel_7x__wasmsimd_c24_acc2()
|
| D | 7x-minmax-ssse3-c8-acc2.c | 98 const __m128i vabsacc0123 = _mm_abs_epi32(vacc0123); in xnn_qs8_gavgpool_minmax_ukernel_7x__ssse3_c8_acc2() local
101 const __m128i vabsacc13 = _mm_shuffle_epi32(vabsacc0123, _MM_SHUFFLE(3, 3, 1, 1)); in xnn_qs8_gavgpool_minmax_ukernel_7x__ssse3_c8_acc2()
104 const __m128i vabsprod02 = _mm_mul_epu32(vabsacc0123, vmultiplier); in xnn_qs8_gavgpool_minmax_ukernel_7x__ssse3_c8_acc2()
178 const __m128i vabsacc0123 = _mm_abs_epi32(vacc0123); in xnn_qs8_gavgpool_minmax_ukernel_7x__ssse3_c8_acc2() local
181 const __m128i vabsacc13 = _mm_shuffle_epi32(vabsacc0123, _MM_SHUFFLE(3, 3, 1, 1)); in xnn_qs8_gavgpool_minmax_ukernel_7x__ssse3_c8_acc2()
184 const __m128i vabsprod02 = _mm_mul_epu32(vabsacc0123, vmultiplier); in xnn_qs8_gavgpool_minmax_ukernel_7x__ssse3_c8_acc2()
|
| D | 7x-minmax-sse2-c8-acc2.c | 101 const __m128i vabsacc0123 = _mm_sub_epi32(_mm_xor_si128(vacc0123, vsgnacc0123), vsgnacc0123); in xnn_qs8_gavgpool_minmax_ukernel_7x__sse2_c8_acc2() local
104 const __m128i vabsacc13 = _mm_shuffle_epi32(vabsacc0123, _MM_SHUFFLE(3, 3, 1, 1)); in xnn_qs8_gavgpool_minmax_ukernel_7x__sse2_c8_acc2()
107 const __m128i vabsprod02 = _mm_mul_epu32(vabsacc0123, vmultiplier); in xnn_qs8_gavgpool_minmax_ukernel_7x__sse2_c8_acc2()
184 const __m128i vabsacc0123 = _mm_sub_epi32(_mm_xor_si128(vacc0123, vsgnacc0123), vsgnacc0123); in xnn_qs8_gavgpool_minmax_ukernel_7x__sse2_c8_acc2() local
187 const __m128i vabsacc13 = _mm_shuffle_epi32(vabsacc0123, _MM_SHUFFLE(3, 3, 1, 1)); in xnn_qs8_gavgpool_minmax_ukernel_7x__sse2_c8_acc2()
190 const __m128i vabsprod02 = _mm_mul_epu32(vabsacc0123, vmultiplier); in xnn_qs8_gavgpool_minmax_ukernel_7x__sse2_c8_acc2()
|
| D | 7x-minmax-sse41-c16-acc2.c | 105 const __m128i vabsacc0123 = _mm_abs_epi32(vacc0123); in xnn_qs8_gavgpool_minmax_ukernel_7x__sse41_c16_acc2() local
110 const __m128i vabsacc13 = _mm_shuffle_epi32(vabsacc0123, _MM_SHUFFLE(3, 3, 1, 1)); in xnn_qs8_gavgpool_minmax_ukernel_7x__sse41_c16_acc2()
115 const __m128i vabsprod02 = _mm_mul_epu32(vabsacc0123, vmultiplier); in xnn_qs8_gavgpool_minmax_ukernel_7x__sse41_c16_acc2()
190 const __m128i vabsacc0123 = _mm_abs_epi32(vacc0123); in xnn_qs8_gavgpool_minmax_ukernel_7x__sse41_c16_acc2() local
193 const __m128i vabsacc13 = _mm_shuffle_epi32(vabsacc0123, _MM_SHUFFLE(3, 3, 1, 1)); in xnn_qs8_gavgpool_minmax_ukernel_7x__sse41_c16_acc2()
196 const __m128i vabsprod02 = _mm_mul_epu32(vabsacc0123, vmultiplier); in xnn_qs8_gavgpool_minmax_ukernel_7x__sse41_c16_acc2()
|
| D | 7p7x-minmax-wasmsimd-c16-acc2.c | 232 const v128_t vabsacc0123 = wasm_i32x4_abs(vacc0123); in xnn_qs8_gavgpool_minmax_ukernel_7p7x__wasmsimd_c16_acc2() local
237 const v128_t vsgnacc0123 = wasm_i32x4_gt(vabsacc0123, vacc0123); in xnn_qs8_gavgpool_minmax_ukernel_7p7x__wasmsimd_c16_acc2()
242 const v128_t vabsacc01 = wasm_v32x4_shuffle(vabsacc0123, vzero, 0, 4, 1, 5); in xnn_qs8_gavgpool_minmax_ukernel_7p7x__wasmsimd_c16_acc2()
243 const v128_t vabsacc23 = wasm_v32x4_shuffle(vabsacc0123, vzero, 2, 6, 3, 7); in xnn_qs8_gavgpool_minmax_ukernel_7p7x__wasmsimd_c16_acc2()
323 const v128_t vabsacc0123 = wasm_i32x4_abs(vacc0123); in xnn_qs8_gavgpool_minmax_ukernel_7p7x__wasmsimd_c16_acc2() local
326 const v128_t vsgnacc0123 = wasm_i32x4_gt(vabsacc0123, vacc0123); in xnn_qs8_gavgpool_minmax_ukernel_7p7x__wasmsimd_c16_acc2()
329 const v128_t vabsacc01 = wasm_v32x4_shuffle(vabsacc0123, vzero, 0, 4, 1, 5); in xnn_qs8_gavgpool_minmax_ukernel_7p7x__wasmsimd_c16_acc2()
330 const v128_t vabsacc23 = wasm_v32x4_shuffle(vabsacc0123, vzero, 2, 6, 3, 7); in xnn_qs8_gavgpool_minmax_ukernel_7p7x__wasmsimd_c16_acc2()
|
| D | 7p7x-minmax-sse41-c8-acc2.c | 185 const __m128i vabsacc0123 = _mm_abs_epi32(vacc0123); in xnn_qs8_gavgpool_minmax_ukernel_7p7x__sse41_c8_acc2() local
188 const __m128i vabsacc13 = _mm_shuffle_epi32(vabsacc0123, _MM_SHUFFLE(3, 3, 1, 1)); in xnn_qs8_gavgpool_minmax_ukernel_7p7x__sse41_c8_acc2()
191 const __m128i vabsprod02 = _mm_mul_epu32(vabsacc0123, vmultiplier); in xnn_qs8_gavgpool_minmax_ukernel_7p7x__sse41_c8_acc2()
253 const __m128i vabsacc0123 = _mm_abs_epi32(vacc0123); in xnn_qs8_gavgpool_minmax_ukernel_7p7x__sse41_c8_acc2() local
256 const __m128i vabsacc13 = _mm_shuffle_epi32(vabsacc0123, _MM_SHUFFLE(3, 3, 1, 1)); in xnn_qs8_gavgpool_minmax_ukernel_7p7x__sse41_c8_acc2()
259 const __m128i vabsprod02 = _mm_mul_epu32(vabsacc0123, vmultiplier); in xnn_qs8_gavgpool_minmax_ukernel_7p7x__sse41_c8_acc2()
|
| D | 7x-minmax-ssse3-c16-acc2.c | 121 const __m128i vabsacc0123 = _mm_abs_epi32(vacc0123); in xnn_qs8_gavgpool_minmax_ukernel_7x__ssse3_c16_acc2() local
126 const __m128i vabsacc13 = _mm_shuffle_epi32(vabsacc0123, _MM_SHUFFLE(3, 3, 1, 1)); in xnn_qs8_gavgpool_minmax_ukernel_7x__ssse3_c16_acc2()
131 const __m128i vabsprod02 = _mm_mul_epu32(vabsacc0123, vmultiplier); in xnn_qs8_gavgpool_minmax_ukernel_7x__ssse3_c16_acc2()
223 const __m128i vabsacc0123 = _mm_abs_epi32(vacc0123); in xnn_qs8_gavgpool_minmax_ukernel_7x__ssse3_c16_acc2() local
226 const __m128i vabsacc13 = _mm_shuffle_epi32(vabsacc0123, _MM_SHUFFLE(3, 3, 1, 1)); in xnn_qs8_gavgpool_minmax_ukernel_7x__ssse3_c16_acc2()
229 const __m128i vabsprod02 = _mm_mul_epu32(vabsacc0123, vmultiplier); in xnn_qs8_gavgpool_minmax_ukernel_7x__ssse3_c16_acc2()
|
| D | 7x-minmax-sse41-c24-acc2.c | 120 const __m128i vabsacc0123 = _mm_abs_epi32(vacc0123); in xnn_qs8_gavgpool_minmax_ukernel_7x__sse41_c24_acc2() local
127 const __m128i vabsacc13 = _mm_shuffle_epi32(vabsacc0123, _MM_SHUFFLE(3, 3, 1, 1)); in xnn_qs8_gavgpool_minmax_ukernel_7x__sse41_c24_acc2()
134 const __m128i vabsprod02 = _mm_mul_epu32(vabsacc0123, vmultiplier); in xnn_qs8_gavgpool_minmax_ukernel_7x__sse41_c24_acc2()
225 const __m128i vabsacc0123 = _mm_abs_epi32(vacc0123); in xnn_qs8_gavgpool_minmax_ukernel_7x__sse41_c24_acc2() local
228 const __m128i vabsacc13 = _mm_shuffle_epi32(vabsacc0123, _MM_SHUFFLE(3, 3, 1, 1)); in xnn_qs8_gavgpool_minmax_ukernel_7x__sse41_c24_acc2()
231 const __m128i vabsprod02 = _mm_mul_epu32(vabsacc0123, vmultiplier); in xnn_qs8_gavgpool_minmax_ukernel_7x__sse41_c24_acc2()
|
| D | 7x-minmax-sse2-c16-acc2.c | 126 const __m128i vabsacc0123 = _mm_sub_epi32(_mm_xor_si128(vacc0123, vsgnacc0123), vsgnacc0123); in xnn_qs8_gavgpool_minmax_ukernel_7x__sse2_c16_acc2() local
131 const __m128i vabsacc13 = _mm_shuffle_epi32(vabsacc0123, _MM_SHUFFLE(3, 3, 1, 1)); in xnn_qs8_gavgpool_minmax_ukernel_7x__sse2_c16_acc2()
136 const __m128i vabsprod02 = _mm_mul_epu32(vabsacc0123, vmultiplier); in xnn_qs8_gavgpool_minmax_ukernel_7x__sse2_c16_acc2()
231 const __m128i vabsacc0123 = _mm_sub_epi32(_mm_xor_si128(vacc0123, vsgnacc0123), vsgnacc0123); in xnn_qs8_gavgpool_minmax_ukernel_7x__sse2_c16_acc2() local
234 const __m128i vabsacc13 = _mm_shuffle_epi32(vabsacc0123, _MM_SHUFFLE(3, 3, 1, 1)); in xnn_qs8_gavgpool_minmax_ukernel_7x__sse2_c16_acc2()
237 const __m128i vabsprod02 = _mm_mul_epu32(vabsacc0123, vmultiplier); in xnn_qs8_gavgpool_minmax_ukernel_7x__sse2_c16_acc2()
|
| D | 7p7x-minmax-ssse3-c8-acc2.c | 209 const __m128i vabsacc0123 = _mm_abs_epi32(vacc0123); in xnn_qs8_gavgpool_minmax_ukernel_7p7x__ssse3_c8_acc2() local
212 const __m128i vabsacc13 = _mm_shuffle_epi32(vabsacc0123, _MM_SHUFFLE(3, 3, 1, 1)); in xnn_qs8_gavgpool_minmax_ukernel_7p7x__ssse3_c8_acc2()
215 const __m128i vabsprod02 = _mm_mul_epu32(vabsacc0123, vmultiplier); in xnn_qs8_gavgpool_minmax_ukernel_7p7x__ssse3_c8_acc2()
290 const __m128i vabsacc0123 = _mm_abs_epi32(vacc0123); in xnn_qs8_gavgpool_minmax_ukernel_7p7x__ssse3_c8_acc2() local
293 const __m128i vabsacc13 = _mm_shuffle_epi32(vabsacc0123, _MM_SHUFFLE(3, 3, 1, 1)); in xnn_qs8_gavgpool_minmax_ukernel_7p7x__ssse3_c8_acc2()
296 const __m128i vabsprod02 = _mm_mul_epu32(vabsacc0123, vmultiplier); in xnn_qs8_gavgpool_minmax_ukernel_7p7x__ssse3_c8_acc2()
|
| D | 7p7x-minmax-sse2-c8-acc2.c | 212 const __m128i vabsacc0123 = _mm_sub_epi32(_mm_xor_si128(vacc0123, vsgnacc0123), vsgnacc0123); in xnn_qs8_gavgpool_minmax_ukernel_7p7x__sse2_c8_acc2() local
215 const __m128i vabsacc13 = _mm_shuffle_epi32(vabsacc0123, _MM_SHUFFLE(3, 3, 1, 1)); in xnn_qs8_gavgpool_minmax_ukernel_7p7x__sse2_c8_acc2()
218 const __m128i vabsprod02 = _mm_mul_epu32(vabsacc0123, vmultiplier); in xnn_qs8_gavgpool_minmax_ukernel_7p7x__sse2_c8_acc2()
296 const __m128i vabsacc0123 = _mm_sub_epi32(_mm_xor_si128(vacc0123, vsgnacc0123), vsgnacc0123); in xnn_qs8_gavgpool_minmax_ukernel_7p7x__sse2_c8_acc2() local
299 const __m128i vabsacc13 = _mm_shuffle_epi32(vabsacc0123, _MM_SHUFFLE(3, 3, 1, 1)); in xnn_qs8_gavgpool_minmax_ukernel_7p7x__sse2_c8_acc2()
302 const __m128i vabsprod02 = _mm_mul_epu32(vabsacc0123, vmultiplier); in xnn_qs8_gavgpool_minmax_ukernel_7p7x__sse2_c8_acc2()
|
| D | 7p7x-minmax-sse41-c16-acc2.c | 234 const __m128i vabsacc0123 = _mm_abs_epi32(vacc0123); in xnn_qs8_gavgpool_minmax_ukernel_7p7x__sse41_c16_acc2() local
239 const __m128i vabsacc13 = _mm_shuffle_epi32(vabsacc0123, _MM_SHUFFLE(3, 3, 1, 1)); in xnn_qs8_gavgpool_minmax_ukernel_7p7x__sse41_c16_acc2()
244 const __m128i vabsprod02 = _mm_mul_epu32(vabsacc0123, vmultiplier); in xnn_qs8_gavgpool_minmax_ukernel_7p7x__sse41_c16_acc2()
320 const __m128i vabsacc0123 = _mm_abs_epi32(vacc0123); in xnn_qs8_gavgpool_minmax_ukernel_7p7x__sse41_c16_acc2() local
323 const __m128i vabsacc13 = _mm_shuffle_epi32(vabsacc0123, _MM_SHUFFLE(3, 3, 1, 1)); in xnn_qs8_gavgpool_minmax_ukernel_7p7x__sse41_c16_acc2()
326 const __m128i vabsprod02 = _mm_mul_epu32(vabsacc0123, vmultiplier); in xnn_qs8_gavgpool_minmax_ukernel_7p7x__sse41_c16_acc2()
|
| D | 7x-minmax-sse2-c24-acc2.c | 151 const __m128i vabsacc0123 = _mm_sub_epi32(_mm_xor_si128(vacc0123, vsgnacc0123), vsgnacc0123); in xnn_qs8_gavgpool_minmax_ukernel_7x__sse2_c24_acc2() local
158 const __m128i vabsacc13 = _mm_shuffle_epi32(vabsacc0123, _MM_SHUFFLE(3, 3, 1, 1)); in xnn_qs8_gavgpool_minmax_ukernel_7x__sse2_c24_acc2()
165 const __m128i vabsprod02 = _mm_mul_epu32(vabsacc0123, vmultiplier); in xnn_qs8_gavgpool_minmax_ukernel_7x__sse2_c24_acc2()
280 const __m128i vabsacc0123 = _mm_sub_epi32(_mm_xor_si128(vacc0123, vsgnacc0123), vsgnacc0123); in xnn_qs8_gavgpool_minmax_ukernel_7x__sse2_c24_acc2() local
283 const __m128i vabsacc13 = _mm_shuffle_epi32(vabsacc0123, _MM_SHUFFLE(3, 3, 1, 1)); in xnn_qs8_gavgpool_minmax_ukernel_7x__sse2_c24_acc2()
286 const __m128i vabsprod02 = _mm_mul_epu32(vabsacc0123, vmultiplier); in xnn_qs8_gavgpool_minmax_ukernel_7x__sse2_c24_acc2()
|
| D | 7x-minmax-ssse3-c24-acc2.c | 144 const __m128i vabsacc0123 = _mm_abs_epi32(vacc0123); in xnn_qs8_gavgpool_minmax_ukernel_7x__ssse3_c24_acc2() local
151 const __m128i vabsacc13 = _mm_shuffle_epi32(vabsacc0123, _MM_SHUFFLE(3, 3, 1, 1)); in xnn_qs8_gavgpool_minmax_ukernel_7x__ssse3_c24_acc2()
158 const __m128i vabsprod02 = _mm_mul_epu32(vabsacc0123, vmultiplier); in xnn_qs8_gavgpool_minmax_ukernel_7x__ssse3_c24_acc2()
270 const __m128i vabsacc0123 = _mm_abs_epi32(vacc0123); in xnn_qs8_gavgpool_minmax_ukernel_7x__ssse3_c24_acc2() local
273 const __m128i vabsacc13 = _mm_shuffle_epi32(vabsacc0123, _MM_SHUFFLE(3, 3, 1, 1)); in xnn_qs8_gavgpool_minmax_ukernel_7x__ssse3_c24_acc2()
276 const __m128i vabsprod02 = _mm_mul_epu32(vabsacc0123, vmultiplier); in xnn_qs8_gavgpool_minmax_ukernel_7x__ssse3_c24_acc2()
|
| D | 7p7x-minmax-wasmsimd-c24-acc2.c | 355 const v128_t vabsacc0123 = wasm_i32x4_abs(vacc0123); in xnn_qs8_gavgpool_minmax_ukernel_7p7x__wasmsimd_c24_acc2() local
362 const v128_t vsgnacc0123 = wasm_i32x4_gt(vabsacc0123, vacc0123); in xnn_qs8_gavgpool_minmax_ukernel_7p7x__wasmsimd_c24_acc2()
369 const v128_t vabsacc01 = wasm_v32x4_shuffle(vabsacc0123, vzero, 0, 4, 1, 5); in xnn_qs8_gavgpool_minmax_ukernel_7p7x__wasmsimd_c24_acc2()
370 const v128_t vabsacc23 = wasm_v32x4_shuffle(vabsacc0123, vzero, 2, 6, 3, 7); in xnn_qs8_gavgpool_minmax_ukernel_7p7x__wasmsimd_c24_acc2()
469 const v128_t vabsacc0123 = wasm_i32x4_abs(vacc0123); in xnn_qs8_gavgpool_minmax_ukernel_7p7x__wasmsimd_c24_acc2() local
472 const v128_t vsgnacc0123 = wasm_i32x4_gt(vabsacc0123, vacc0123); in xnn_qs8_gavgpool_minmax_ukernel_7p7x__wasmsimd_c24_acc2()
475 const v128_t vabsacc01 = wasm_v32x4_shuffle(vabsacc0123, vzero, 0, 4, 1, 5); in xnn_qs8_gavgpool_minmax_ukernel_7p7x__wasmsimd_c24_acc2()
476 const v128_t vabsacc23 = wasm_v32x4_shuffle(vabsacc0123, vzero, 2, 6, 3, 7); in xnn_qs8_gavgpool_minmax_ukernel_7p7x__wasmsimd_c24_acc2()
|
| D | 7p7x-minmax-ssse3-c16-acc2.c | 282 const __m128i vabsacc0123 = _mm_abs_epi32(vacc0123); in xnn_qs8_gavgpool_minmax_ukernel_7p7x__ssse3_c16_acc2() local
287 const __m128i vabsacc13 = _mm_shuffle_epi32(vabsacc0123, _MM_SHUFFLE(3, 3, 1, 1)); in xnn_qs8_gavgpool_minmax_ukernel_7p7x__ssse3_c16_acc2()
292 const __m128i vabsprod02 = _mm_mul_epu32(vabsacc0123, vmultiplier); in xnn_qs8_gavgpool_minmax_ukernel_7p7x__ssse3_c16_acc2()
385 const __m128i vabsacc0123 = _mm_abs_epi32(vacc0123); in xnn_qs8_gavgpool_minmax_ukernel_7p7x__ssse3_c16_acc2() local
388 const __m128i vabsacc13 = _mm_shuffle_epi32(vabsacc0123, _MM_SHUFFLE(3, 3, 1, 1)); in xnn_qs8_gavgpool_minmax_ukernel_7p7x__ssse3_c16_acc2()
391 const __m128i vabsprod02 = _mm_mul_epu32(vabsacc0123, vmultiplier); in xnn_qs8_gavgpool_minmax_ukernel_7p7x__ssse3_c16_acc2()
|
| D | 7p7x-minmax-sse2-c16-acc2.c | 287 const __m128i vabsacc0123 = _mm_sub_epi32(_mm_xor_si128(vacc0123, vsgnacc0123), vsgnacc0123); in xnn_qs8_gavgpool_minmax_ukernel_7p7x__sse2_c16_acc2() local
292 const __m128i vabsacc13 = _mm_shuffle_epi32(vabsacc0123, _MM_SHUFFLE(3, 3, 1, 1)); in xnn_qs8_gavgpool_minmax_ukernel_7p7x__sse2_c16_acc2()
297 const __m128i vabsprod02 = _mm_mul_epu32(vabsacc0123, vmultiplier); in xnn_qs8_gavgpool_minmax_ukernel_7p7x__sse2_c16_acc2()
393 const __m128i vabsacc0123 = _mm_sub_epi32(_mm_xor_si128(vacc0123, vsgnacc0123), vsgnacc0123); in xnn_qs8_gavgpool_minmax_ukernel_7p7x__sse2_c16_acc2() local
396 const __m128i vabsacc13 = _mm_shuffle_epi32(vabsacc0123, _MM_SHUFFLE(3, 3, 1, 1)); in xnn_qs8_gavgpool_minmax_ukernel_7p7x__sse2_c16_acc2()
399 const __m128i vabsprod02 = _mm_mul_epu32(vabsacc0123, vmultiplier); in xnn_qs8_gavgpool_minmax_ukernel_7p7x__sse2_c16_acc2()
|
| D | 7p7x-minmax-sse41-c24-acc2.c | 359 const __m128i vabsacc0123 = _mm_abs_epi32(vacc0123); in xnn_qs8_gavgpool_minmax_ukernel_7p7x__sse41_c24_acc2() local
366 const __m128i vabsacc13 = _mm_shuffle_epi32(vabsacc0123, _MM_SHUFFLE(3, 3, 1, 1)); in xnn_qs8_gavgpool_minmax_ukernel_7p7x__sse41_c24_acc2()
373 const __m128i vabsprod02 = _mm_mul_epu32(vabsacc0123, vmultiplier); in xnn_qs8_gavgpool_minmax_ukernel_7p7x__sse41_c24_acc2()
465 const __m128i vabsacc0123 = _mm_abs_epi32(vacc0123); in xnn_qs8_gavgpool_minmax_ukernel_7p7x__sse41_c24_acc2() local
468 const __m128i vabsacc13 = _mm_shuffle_epi32(vabsacc0123, _MM_SHUFFLE(3, 3, 1, 1)); in xnn_qs8_gavgpool_minmax_ukernel_7p7x__sse41_c24_acc2()
471 const __m128i vabsprod02 = _mm_mul_epu32(vabsacc0123, vmultiplier); in xnn_qs8_gavgpool_minmax_ukernel_7p7x__sse41_c24_acc2()
|
| /external/XNNPACK/src/qs8-dwconv/gen/ |
| D | up8x9-minmax-sse2-mul16.c | 212 const __m128i vabsacc0123 = _mm_sub_epi32(_mm_xor_si128(vacc0123, vnmask0123), vnmask0123); in xnn_qs8_dwconv_minmax_ukernel_up8x9__sse2_mul16() local
215 const __m128i vabsacc13 = _mm_shuffle_epi32(vabsacc0123, _MM_SHUFFLE(3, 3, 1, 1)); in xnn_qs8_dwconv_minmax_ukernel_up8x9__sse2_mul16()
216 const __m128i vabsprod02 = _mm_mul_epu32(vabsacc0123, vmultiplier); in xnn_qs8_dwconv_minmax_ukernel_up8x9__sse2_mul16()
391 const __m128i vabsacc0123 = _mm_sub_epi32(_mm_xor_si128(vacc0123, vnmask0123), vnmask0123); in xnn_qs8_dwconv_minmax_ukernel_up8x9__sse2_mul16() local
394 const __m128i vabsacc13 = _mm_shuffle_epi32(vabsacc0123, _MM_SHUFFLE(3, 3, 1, 1)); in xnn_qs8_dwconv_minmax_ukernel_up8x9__sse2_mul16()
397 const __m128i vabsprod02 = _mm_mul_epu32(vabsacc0123, vmultiplier); in xnn_qs8_dwconv_minmax_ukernel_up8x9__sse2_mul16()
|
| D | up8x9-minmax-ssse3-mul16.c | 212 const __m128i vabsacc0123 = _mm_abs_epi32(vacc0123); in xnn_qs8_dwconv_minmax_ukernel_up8x9__ssse3_mul16() local
215 const __m128i vabsacc13 = _mm_shuffle_epi32(vabsacc0123, _MM_SHUFFLE(3, 3, 1, 1)); in xnn_qs8_dwconv_minmax_ukernel_up8x9__ssse3_mul16()
216 const __m128i vabsprod02 = _mm_mul_epu32(vabsacc0123, vmultiplier); in xnn_qs8_dwconv_minmax_ukernel_up8x9__ssse3_mul16()
391 const __m128i vabsacc0123 = _mm_abs_epi32(vacc0123); in xnn_qs8_dwconv_minmax_ukernel_up8x9__ssse3_mul16() local
394 const __m128i vabsacc13 = _mm_shuffle_epi32(vabsacc0123, _MM_SHUFFLE(3, 3, 1, 1)); in xnn_qs8_dwconv_minmax_ukernel_up8x9__ssse3_mul16()
397 const __m128i vabsprod02 = _mm_mul_epu32(vabsacc0123, vmultiplier); in xnn_qs8_dwconv_minmax_ukernel_up8x9__ssse3_mul16()
|
| D | up16x9-minmax-ssse3-mul16.c | 288 const __m128i vabsacc0123 = _mm_abs_epi32(vacc0123); in xnn_qs8_dwconv_minmax_ukernel_up16x9__ssse3_mul16() local
293 const __m128i vabsacc13 = _mm_shuffle_epi32(vabsacc0123, _MM_SHUFFLE(3, 3, 1, 1)); in xnn_qs8_dwconv_minmax_ukernel_up16x9__ssse3_mul16()
294 const __m128i vabsprod02 = _mm_mul_epu32(vabsacc0123, vmultiplier); in xnn_qs8_dwconv_minmax_ukernel_up16x9__ssse3_mul16()
515 const __m128i vabsacc0123 = _mm_abs_epi32(vacc0123); in xnn_qs8_dwconv_minmax_ukernel_up16x9__ssse3_mul16() local
518 const __m128i vabsacc13 = _mm_shuffle_epi32(vabsacc0123, _MM_SHUFFLE(3, 3, 1, 1)); in xnn_qs8_dwconv_minmax_ukernel_up16x9__ssse3_mul16()
521 const __m128i vabsprod02 = _mm_mul_epu32(vabsacc0123, vmultiplier); in xnn_qs8_dwconv_minmax_ukernel_up16x9__ssse3_mul16()
|