| /external/XNNPACK/src/qs8-gavgpool/gen/ |
| D | 7x-minmax-wasmsimd-c8-acc2.c | 91 const v128_t vabsacc4567 = wasm_i32x4_abs(vacc4567); in xnn_qs8_gavgpool_minmax_ukernel_7x__wasmsimd_c8_acc2() local
94 const v128_t vsgnacc4567 = wasm_i32x4_gt(vabsacc4567, vacc4567); in xnn_qs8_gavgpool_minmax_ukernel_7x__wasmsimd_c8_acc2()
98 const v128_t vabsacc45 = wasm_v32x4_shuffle(vabsacc4567, vzero, 0, 4, 1, 5); in xnn_qs8_gavgpool_minmax_ukernel_7x__wasmsimd_c8_acc2()
99 const v128_t vabsacc67 = wasm_v32x4_shuffle(vabsacc4567, vzero, 2, 6, 3, 7); in xnn_qs8_gavgpool_minmax_ukernel_7x__wasmsimd_c8_acc2()
160 const v128_t vabsacc4567 = wasm_i32x4_abs(vacc4567); in xnn_qs8_gavgpool_minmax_ukernel_7x__wasmsimd_c8_acc2() local
163 const v128_t vsgnacc4567 = wasm_i32x4_gt(vabsacc4567, vacc4567); in xnn_qs8_gavgpool_minmax_ukernel_7x__wasmsimd_c8_acc2()
167 const v128_t vabsacc45 = wasm_v32x4_shuffle(vabsacc4567, vzero, 0, 4, 1, 5); in xnn_qs8_gavgpool_minmax_ukernel_7x__wasmsimd_c8_acc2()
168 const v128_t vabsacc67 = wasm_v32x4_shuffle(vabsacc4567, vzero, 2, 6, 3, 7); in xnn_qs8_gavgpool_minmax_ukernel_7x__wasmsimd_c8_acc2()
|
| D | 7x-minmax-wasmsimd-c16-acc2.c | 106 const v128_t vabsacc4567 = wasm_i32x4_abs(vacc4567); in xnn_qs8_gavgpool_minmax_ukernel_7x__wasmsimd_c16_acc2() local
111 const v128_t vsgnacc4567 = wasm_i32x4_gt(vabsacc4567, vacc4567); in xnn_qs8_gavgpool_minmax_ukernel_7x__wasmsimd_c16_acc2()
117 const v128_t vabsacc45 = wasm_v32x4_shuffle(vabsacc4567, vzero, 0, 4, 1, 5); in xnn_qs8_gavgpool_minmax_ukernel_7x__wasmsimd_c16_acc2()
118 const v128_t vabsacc67 = wasm_v32x4_shuffle(vabsacc4567, vzero, 2, 6, 3, 7); in xnn_qs8_gavgpool_minmax_ukernel_7x__wasmsimd_c16_acc2()
196 const v128_t vabsacc4567 = wasm_i32x4_abs(vacc4567); in xnn_qs8_gavgpool_minmax_ukernel_7x__wasmsimd_c16_acc2() local
199 const v128_t vsgnacc4567 = wasm_i32x4_gt(vabsacc4567, vacc4567); in xnn_qs8_gavgpool_minmax_ukernel_7x__wasmsimd_c16_acc2()
203 const v128_t vabsacc45 = wasm_v32x4_shuffle(vabsacc4567, vzero, 0, 4, 1, 5); in xnn_qs8_gavgpool_minmax_ukernel_7x__wasmsimd_c16_acc2()
204 const v128_t vabsacc67 = wasm_v32x4_shuffle(vabsacc4567, vzero, 2, 6, 3, 7); in xnn_qs8_gavgpool_minmax_ukernel_7x__wasmsimd_c16_acc2()
|
| D | 7x-minmax-sse41-c8-acc2.c | 91 const __m128i vabsacc4567 = _mm_abs_epi32(vacc4567); in xnn_qs8_gavgpool_minmax_ukernel_7x__sse41_c8_acc2() local
94 const __m128i vabsacc57 = _mm_shuffle_epi32(vabsacc4567, _MM_SHUFFLE(3, 3, 1, 1)); in xnn_qs8_gavgpool_minmax_ukernel_7x__sse41_c8_acc2()
98 const __m128i vabsprod46 = _mm_mul_epu32(vabsacc4567, vmultiplier); in xnn_qs8_gavgpool_minmax_ukernel_7x__sse41_c8_acc2()
158 const __m128i vabsacc4567 = _mm_abs_epi32(vacc4567); in xnn_qs8_gavgpool_minmax_ukernel_7x__sse41_c8_acc2() local
161 const __m128i vabsacc57 = _mm_shuffle_epi32(vabsacc4567, _MM_SHUFFLE(3, 3, 1, 1)); in xnn_qs8_gavgpool_minmax_ukernel_7x__sse41_c8_acc2()
165 const __m128i vabsprod46 = _mm_mul_epu32(vabsacc4567, vmultiplier); in xnn_qs8_gavgpool_minmax_ukernel_7x__sse41_c8_acc2()
|
| D | 7p7x-minmax-wasmsimd-c8-acc2.c | 184 const v128_t vabsacc4567 = wasm_i32x4_abs(vacc4567); in xnn_qs8_gavgpool_minmax_ukernel_7p7x__wasmsimd_c8_acc2() local
187 const v128_t vsgnacc4567 = wasm_i32x4_gt(vabsacc4567, vacc4567); in xnn_qs8_gavgpool_minmax_ukernel_7p7x__wasmsimd_c8_acc2()
191 const v128_t vabsacc45 = wasm_v32x4_shuffle(vabsacc4567, vzero, 0, 4, 1, 5); in xnn_qs8_gavgpool_minmax_ukernel_7p7x__wasmsimd_c8_acc2()
192 const v128_t vabsacc67 = wasm_v32x4_shuffle(vabsacc4567, vzero, 2, 6, 3, 7); in xnn_qs8_gavgpool_minmax_ukernel_7p7x__wasmsimd_c8_acc2()
254 const v128_t vabsacc4567 = wasm_i32x4_abs(vacc4567); in xnn_qs8_gavgpool_minmax_ukernel_7p7x__wasmsimd_c8_acc2() local
257 const v128_t vsgnacc4567 = wasm_i32x4_gt(vabsacc4567, vacc4567); in xnn_qs8_gavgpool_minmax_ukernel_7p7x__wasmsimd_c8_acc2()
261 const v128_t vabsacc45 = wasm_v32x4_shuffle(vabsacc4567, vzero, 0, 4, 1, 5); in xnn_qs8_gavgpool_minmax_ukernel_7p7x__wasmsimd_c8_acc2()
262 const v128_t vabsacc67 = wasm_v32x4_shuffle(vabsacc4567, vzero, 2, 6, 3, 7); in xnn_qs8_gavgpool_minmax_ukernel_7p7x__wasmsimd_c8_acc2()
|
| D | 7x-minmax-wasmsimd-c24-acc2.c | 121 const v128_t vabsacc4567 = wasm_i32x4_abs(vacc4567); in xnn_qs8_gavgpool_minmax_ukernel_7x__wasmsimd_c24_acc2() local
128 const v128_t vsgnacc4567 = wasm_i32x4_gt(vabsacc4567, vacc4567); in xnn_qs8_gavgpool_minmax_ukernel_7x__wasmsimd_c24_acc2()
136 const v128_t vabsacc45 = wasm_v32x4_shuffle(vabsacc4567, vzero, 0, 4, 1, 5); in xnn_qs8_gavgpool_minmax_ukernel_7x__wasmsimd_c24_acc2()
137 const v128_t vabsacc67 = wasm_v32x4_shuffle(vabsacc4567, vzero, 2, 6, 3, 7); in xnn_qs8_gavgpool_minmax_ukernel_7x__wasmsimd_c24_acc2()
234 const v128_t vabsacc4567 = wasm_i32x4_abs(vacc4567); in xnn_qs8_gavgpool_minmax_ukernel_7x__wasmsimd_c24_acc2() local
237 const v128_t vsgnacc4567 = wasm_i32x4_gt(vabsacc4567, vacc4567); in xnn_qs8_gavgpool_minmax_ukernel_7x__wasmsimd_c24_acc2()
241 const v128_t vabsacc45 = wasm_v32x4_shuffle(vabsacc4567, vzero, 0, 4, 1, 5); in xnn_qs8_gavgpool_minmax_ukernel_7x__wasmsimd_c24_acc2()
242 const v128_t vabsacc67 = wasm_v32x4_shuffle(vabsacc4567, vzero, 2, 6, 3, 7); in xnn_qs8_gavgpool_minmax_ukernel_7x__wasmsimd_c24_acc2()
|
| D | 7x-minmax-ssse3-c8-acc2.c | 99 const __m128i vabsacc4567 = _mm_abs_epi32(vacc4567); in xnn_qs8_gavgpool_minmax_ukernel_7x__ssse3_c8_acc2() local
102 const __m128i vabsacc57 = _mm_shuffle_epi32(vabsacc4567, _MM_SHUFFLE(3, 3, 1, 1)); in xnn_qs8_gavgpool_minmax_ukernel_7x__ssse3_c8_acc2()
106 const __m128i vabsprod46 = _mm_mul_epu32(vabsacc4567, vmultiplier); in xnn_qs8_gavgpool_minmax_ukernel_7x__ssse3_c8_acc2()
179 const __m128i vabsacc4567 = _mm_abs_epi32(vacc4567); in xnn_qs8_gavgpool_minmax_ukernel_7x__ssse3_c8_acc2() local
182 const __m128i vabsacc57 = _mm_shuffle_epi32(vabsacc4567, _MM_SHUFFLE(3, 3, 1, 1)); in xnn_qs8_gavgpool_minmax_ukernel_7x__ssse3_c8_acc2()
186 const __m128i vabsprod46 = _mm_mul_epu32(vabsacc4567, vmultiplier); in xnn_qs8_gavgpool_minmax_ukernel_7x__ssse3_c8_acc2()
|
| D | 7x-minmax-sse2-c8-acc2.c | 102 const __m128i vabsacc4567 = _mm_sub_epi32(_mm_xor_si128(vacc4567, vsgnacc4567), vsgnacc4567); in xnn_qs8_gavgpool_minmax_ukernel_7x__sse2_c8_acc2() local
105 const __m128i vabsacc57 = _mm_shuffle_epi32(vabsacc4567, _MM_SHUFFLE(3, 3, 1, 1)); in xnn_qs8_gavgpool_minmax_ukernel_7x__sse2_c8_acc2()
109 const __m128i vabsprod46 = _mm_mul_epu32(vabsacc4567, vmultiplier); in xnn_qs8_gavgpool_minmax_ukernel_7x__sse2_c8_acc2()
185 const __m128i vabsacc4567 = _mm_sub_epi32(_mm_xor_si128(vacc4567, vsgnacc4567), vsgnacc4567); in xnn_qs8_gavgpool_minmax_ukernel_7x__sse2_c8_acc2() local
188 const __m128i vabsacc57 = _mm_shuffle_epi32(vabsacc4567, _MM_SHUFFLE(3, 3, 1, 1)); in xnn_qs8_gavgpool_minmax_ukernel_7x__sse2_c8_acc2()
192 const __m128i vabsprod46 = _mm_mul_epu32(vabsacc4567, vmultiplier); in xnn_qs8_gavgpool_minmax_ukernel_7x__sse2_c8_acc2()
|
| D | 7x-minmax-sse41-c16-acc2.c | 106 const __m128i vabsacc4567 = _mm_abs_epi32(vacc4567); in xnn_qs8_gavgpool_minmax_ukernel_7x__sse41_c16_acc2() local
111 const __m128i vabsacc57 = _mm_shuffle_epi32(vabsacc4567, _MM_SHUFFLE(3, 3, 1, 1)); in xnn_qs8_gavgpool_minmax_ukernel_7x__sse41_c16_acc2()
117 const __m128i vabsprod46 = _mm_mul_epu32(vabsacc4567, vmultiplier); in xnn_qs8_gavgpool_minmax_ukernel_7x__sse41_c16_acc2()
191 const __m128i vabsacc4567 = _mm_abs_epi32(vacc4567); in xnn_qs8_gavgpool_minmax_ukernel_7x__sse41_c16_acc2() local
194 const __m128i vabsacc57 = _mm_shuffle_epi32(vabsacc4567, _MM_SHUFFLE(3, 3, 1, 1)); in xnn_qs8_gavgpool_minmax_ukernel_7x__sse41_c16_acc2()
198 const __m128i vabsprod46 = _mm_mul_epu32(vabsacc4567, vmultiplier); in xnn_qs8_gavgpool_minmax_ukernel_7x__sse41_c16_acc2()
|
| D | 7p7x-minmax-wasmsimd-c16-acc2.c | 233 const v128_t vabsacc4567 = wasm_i32x4_abs(vacc4567); in xnn_qs8_gavgpool_minmax_ukernel_7p7x__wasmsimd_c16_acc2() local
238 const v128_t vsgnacc4567 = wasm_i32x4_gt(vabsacc4567, vacc4567); in xnn_qs8_gavgpool_minmax_ukernel_7p7x__wasmsimd_c16_acc2()
244 const v128_t vabsacc45 = wasm_v32x4_shuffle(vabsacc4567, vzero, 0, 4, 1, 5); in xnn_qs8_gavgpool_minmax_ukernel_7p7x__wasmsimd_c16_acc2()
245 const v128_t vabsacc67 = wasm_v32x4_shuffle(vabsacc4567, vzero, 2, 6, 3, 7); in xnn_qs8_gavgpool_minmax_ukernel_7p7x__wasmsimd_c16_acc2()
324 const v128_t vabsacc4567 = wasm_i32x4_abs(vacc4567); in xnn_qs8_gavgpool_minmax_ukernel_7p7x__wasmsimd_c16_acc2() local
327 const v128_t vsgnacc4567 = wasm_i32x4_gt(vabsacc4567, vacc4567); in xnn_qs8_gavgpool_minmax_ukernel_7p7x__wasmsimd_c16_acc2()
331 const v128_t vabsacc45 = wasm_v32x4_shuffle(vabsacc4567, vzero, 0, 4, 1, 5); in xnn_qs8_gavgpool_minmax_ukernel_7p7x__wasmsimd_c16_acc2()
332 const v128_t vabsacc67 = wasm_v32x4_shuffle(vabsacc4567, vzero, 2, 6, 3, 7); in xnn_qs8_gavgpool_minmax_ukernel_7p7x__wasmsimd_c16_acc2()
|
| D | 7p7x-minmax-sse41-c8-acc2.c | 186 const __m128i vabsacc4567 = _mm_abs_epi32(vacc4567); in xnn_qs8_gavgpool_minmax_ukernel_7p7x__sse41_c8_acc2() local
189 const __m128i vabsacc57 = _mm_shuffle_epi32(vabsacc4567, _MM_SHUFFLE(3, 3, 1, 1)); in xnn_qs8_gavgpool_minmax_ukernel_7p7x__sse41_c8_acc2()
193 const __m128i vabsprod46 = _mm_mul_epu32(vabsacc4567, vmultiplier); in xnn_qs8_gavgpool_minmax_ukernel_7p7x__sse41_c8_acc2()
254 const __m128i vabsacc4567 = _mm_abs_epi32(vacc4567); in xnn_qs8_gavgpool_minmax_ukernel_7p7x__sse41_c8_acc2() local
257 const __m128i vabsacc57 = _mm_shuffle_epi32(vabsacc4567, _MM_SHUFFLE(3, 3, 1, 1)); in xnn_qs8_gavgpool_minmax_ukernel_7p7x__sse41_c8_acc2()
261 const __m128i vabsprod46 = _mm_mul_epu32(vabsacc4567, vmultiplier); in xnn_qs8_gavgpool_minmax_ukernel_7p7x__sse41_c8_acc2()
|
| D | 7x-minmax-ssse3-c16-acc2.c | 122 const __m128i vabsacc4567 = _mm_abs_epi32(vacc4567); in xnn_qs8_gavgpool_minmax_ukernel_7x__ssse3_c16_acc2() local
127 const __m128i vabsacc57 = _mm_shuffle_epi32(vabsacc4567, _MM_SHUFFLE(3, 3, 1, 1)); in xnn_qs8_gavgpool_minmax_ukernel_7x__ssse3_c16_acc2()
133 const __m128i vabsprod46 = _mm_mul_epu32(vabsacc4567, vmultiplier); in xnn_qs8_gavgpool_minmax_ukernel_7x__ssse3_c16_acc2()
224 const __m128i vabsacc4567 = _mm_abs_epi32(vacc4567); in xnn_qs8_gavgpool_minmax_ukernel_7x__ssse3_c16_acc2() local
227 const __m128i vabsacc57 = _mm_shuffle_epi32(vabsacc4567, _MM_SHUFFLE(3, 3, 1, 1)); in xnn_qs8_gavgpool_minmax_ukernel_7x__ssse3_c16_acc2()
231 const __m128i vabsprod46 = _mm_mul_epu32(vabsacc4567, vmultiplier); in xnn_qs8_gavgpool_minmax_ukernel_7x__ssse3_c16_acc2()
|
| D | 7x-minmax-sse41-c24-acc2.c | 121 const __m128i vabsacc4567 = _mm_abs_epi32(vacc4567); in xnn_qs8_gavgpool_minmax_ukernel_7x__sse41_c24_acc2() local
128 const __m128i vabsacc57 = _mm_shuffle_epi32(vabsacc4567, _MM_SHUFFLE(3, 3, 1, 1)); in xnn_qs8_gavgpool_minmax_ukernel_7x__sse41_c24_acc2()
136 const __m128i vabsprod46 = _mm_mul_epu32(vabsacc4567, vmultiplier); in xnn_qs8_gavgpool_minmax_ukernel_7x__sse41_c24_acc2()
226 const __m128i vabsacc4567 = _mm_abs_epi32(vacc4567); in xnn_qs8_gavgpool_minmax_ukernel_7x__sse41_c24_acc2() local
229 const __m128i vabsacc57 = _mm_shuffle_epi32(vabsacc4567, _MM_SHUFFLE(3, 3, 1, 1)); in xnn_qs8_gavgpool_minmax_ukernel_7x__sse41_c24_acc2()
233 const __m128i vabsprod46 = _mm_mul_epu32(vabsacc4567, vmultiplier); in xnn_qs8_gavgpool_minmax_ukernel_7x__sse41_c24_acc2()
|
| D | 7x-minmax-sse2-c16-acc2.c | 127 const __m128i vabsacc4567 = _mm_sub_epi32(_mm_xor_si128(vacc4567, vsgnacc4567), vsgnacc4567); in xnn_qs8_gavgpool_minmax_ukernel_7x__sse2_c16_acc2() local
132 const __m128i vabsacc57 = _mm_shuffle_epi32(vabsacc4567, _MM_SHUFFLE(3, 3, 1, 1)); in xnn_qs8_gavgpool_minmax_ukernel_7x__sse2_c16_acc2()
138 const __m128i vabsprod46 = _mm_mul_epu32(vabsacc4567, vmultiplier); in xnn_qs8_gavgpool_minmax_ukernel_7x__sse2_c16_acc2()
232 const __m128i vabsacc4567 = _mm_sub_epi32(_mm_xor_si128(vacc4567, vsgnacc4567), vsgnacc4567); in xnn_qs8_gavgpool_minmax_ukernel_7x__sse2_c16_acc2() local
235 const __m128i vabsacc57 = _mm_shuffle_epi32(vabsacc4567, _MM_SHUFFLE(3, 3, 1, 1)); in xnn_qs8_gavgpool_minmax_ukernel_7x__sse2_c16_acc2()
239 const __m128i vabsprod46 = _mm_mul_epu32(vabsacc4567, vmultiplier); in xnn_qs8_gavgpool_minmax_ukernel_7x__sse2_c16_acc2()
|
| D | 7p7x-minmax-ssse3-c8-acc2.c | 210 const __m128i vabsacc4567 = _mm_abs_epi32(vacc4567); in xnn_qs8_gavgpool_minmax_ukernel_7p7x__ssse3_c8_acc2() local
213 const __m128i vabsacc57 = _mm_shuffle_epi32(vabsacc4567, _MM_SHUFFLE(3, 3, 1, 1)); in xnn_qs8_gavgpool_minmax_ukernel_7p7x__ssse3_c8_acc2()
217 const __m128i vabsprod46 = _mm_mul_epu32(vabsacc4567, vmultiplier); in xnn_qs8_gavgpool_minmax_ukernel_7p7x__ssse3_c8_acc2()
291 const __m128i vabsacc4567 = _mm_abs_epi32(vacc4567); in xnn_qs8_gavgpool_minmax_ukernel_7p7x__ssse3_c8_acc2() local
294 const __m128i vabsacc57 = _mm_shuffle_epi32(vabsacc4567, _MM_SHUFFLE(3, 3, 1, 1)); in xnn_qs8_gavgpool_minmax_ukernel_7p7x__ssse3_c8_acc2()
298 const __m128i vabsprod46 = _mm_mul_epu32(vabsacc4567, vmultiplier); in xnn_qs8_gavgpool_minmax_ukernel_7p7x__ssse3_c8_acc2()
|
| D | 7p7x-minmax-sse2-c8-acc2.c | 213 const __m128i vabsacc4567 = _mm_sub_epi32(_mm_xor_si128(vacc4567, vsgnacc4567), vsgnacc4567); in xnn_qs8_gavgpool_minmax_ukernel_7p7x__sse2_c8_acc2() local
216 const __m128i vabsacc57 = _mm_shuffle_epi32(vabsacc4567, _MM_SHUFFLE(3, 3, 1, 1)); in xnn_qs8_gavgpool_minmax_ukernel_7p7x__sse2_c8_acc2()
220 const __m128i vabsprod46 = _mm_mul_epu32(vabsacc4567, vmultiplier); in xnn_qs8_gavgpool_minmax_ukernel_7p7x__sse2_c8_acc2()
297 const __m128i vabsacc4567 = _mm_sub_epi32(_mm_xor_si128(vacc4567, vsgnacc4567), vsgnacc4567); in xnn_qs8_gavgpool_minmax_ukernel_7p7x__sse2_c8_acc2() local
300 const __m128i vabsacc57 = _mm_shuffle_epi32(vabsacc4567, _MM_SHUFFLE(3, 3, 1, 1)); in xnn_qs8_gavgpool_minmax_ukernel_7p7x__sse2_c8_acc2()
304 const __m128i vabsprod46 = _mm_mul_epu32(vabsacc4567, vmultiplier); in xnn_qs8_gavgpool_minmax_ukernel_7p7x__sse2_c8_acc2()
|
| D | 7p7x-minmax-sse41-c16-acc2.c | 235 const __m128i vabsacc4567 = _mm_abs_epi32(vacc4567); in xnn_qs8_gavgpool_minmax_ukernel_7p7x__sse41_c16_acc2() local
240 const __m128i vabsacc57 = _mm_shuffle_epi32(vabsacc4567, _MM_SHUFFLE(3, 3, 1, 1)); in xnn_qs8_gavgpool_minmax_ukernel_7p7x__sse41_c16_acc2()
246 const __m128i vabsprod46 = _mm_mul_epu32(vabsacc4567, vmultiplier); in xnn_qs8_gavgpool_minmax_ukernel_7p7x__sse41_c16_acc2()
321 const __m128i vabsacc4567 = _mm_abs_epi32(vacc4567); in xnn_qs8_gavgpool_minmax_ukernel_7p7x__sse41_c16_acc2() local
324 const __m128i vabsacc57 = _mm_shuffle_epi32(vabsacc4567, _MM_SHUFFLE(3, 3, 1, 1)); in xnn_qs8_gavgpool_minmax_ukernel_7p7x__sse41_c16_acc2()
328 const __m128i vabsprod46 = _mm_mul_epu32(vabsacc4567, vmultiplier); in xnn_qs8_gavgpool_minmax_ukernel_7p7x__sse41_c16_acc2()
|
| D | 7x-minmax-sse2-c24-acc2.c | 152 const __m128i vabsacc4567 = _mm_sub_epi32(_mm_xor_si128(vacc4567, vsgnacc4567), vsgnacc4567); in xnn_qs8_gavgpool_minmax_ukernel_7x__sse2_c24_acc2() local
159 const __m128i vabsacc57 = _mm_shuffle_epi32(vabsacc4567, _MM_SHUFFLE(3, 3, 1, 1)); in xnn_qs8_gavgpool_minmax_ukernel_7x__sse2_c24_acc2()
167 const __m128i vabsprod46 = _mm_mul_epu32(vabsacc4567, vmultiplier); in xnn_qs8_gavgpool_minmax_ukernel_7x__sse2_c24_acc2()
281 const __m128i vabsacc4567 = _mm_sub_epi32(_mm_xor_si128(vacc4567, vsgnacc4567), vsgnacc4567); in xnn_qs8_gavgpool_minmax_ukernel_7x__sse2_c24_acc2() local
284 const __m128i vabsacc57 = _mm_shuffle_epi32(vabsacc4567, _MM_SHUFFLE(3, 3, 1, 1)); in xnn_qs8_gavgpool_minmax_ukernel_7x__sse2_c24_acc2()
288 const __m128i vabsprod46 = _mm_mul_epu32(vabsacc4567, vmultiplier); in xnn_qs8_gavgpool_minmax_ukernel_7x__sse2_c24_acc2()
|
| D | 7x-minmax-ssse3-c24-acc2.c | 145 const __m128i vabsacc4567 = _mm_abs_epi32(vacc4567); in xnn_qs8_gavgpool_minmax_ukernel_7x__ssse3_c24_acc2() local
152 const __m128i vabsacc57 = _mm_shuffle_epi32(vabsacc4567, _MM_SHUFFLE(3, 3, 1, 1)); in xnn_qs8_gavgpool_minmax_ukernel_7x__ssse3_c24_acc2()
160 const __m128i vabsprod46 = _mm_mul_epu32(vabsacc4567, vmultiplier); in xnn_qs8_gavgpool_minmax_ukernel_7x__ssse3_c24_acc2()
271 const __m128i vabsacc4567 = _mm_abs_epi32(vacc4567); in xnn_qs8_gavgpool_minmax_ukernel_7x__ssse3_c24_acc2() local
274 const __m128i vabsacc57 = _mm_shuffle_epi32(vabsacc4567, _MM_SHUFFLE(3, 3, 1, 1)); in xnn_qs8_gavgpool_minmax_ukernel_7x__ssse3_c24_acc2()
278 const __m128i vabsprod46 = _mm_mul_epu32(vabsacc4567, vmultiplier); in xnn_qs8_gavgpool_minmax_ukernel_7x__ssse3_c24_acc2()
|
| D | 7p7x-minmax-wasmsimd-c24-acc2.c | 356 const v128_t vabsacc4567 = wasm_i32x4_abs(vacc4567); in xnn_qs8_gavgpool_minmax_ukernel_7p7x__wasmsimd_c24_acc2() local
363 const v128_t vsgnacc4567 = wasm_i32x4_gt(vabsacc4567, vacc4567); in xnn_qs8_gavgpool_minmax_ukernel_7p7x__wasmsimd_c24_acc2()
371 const v128_t vabsacc45 = wasm_v32x4_shuffle(vabsacc4567, vzero, 0, 4, 1, 5); in xnn_qs8_gavgpool_minmax_ukernel_7p7x__wasmsimd_c24_acc2()
372 const v128_t vabsacc67 = wasm_v32x4_shuffle(vabsacc4567, vzero, 2, 6, 3, 7); in xnn_qs8_gavgpool_minmax_ukernel_7p7x__wasmsimd_c24_acc2()
470 const v128_t vabsacc4567 = wasm_i32x4_abs(vacc4567); in xnn_qs8_gavgpool_minmax_ukernel_7p7x__wasmsimd_c24_acc2() local
473 const v128_t vsgnacc4567 = wasm_i32x4_gt(vabsacc4567, vacc4567); in xnn_qs8_gavgpool_minmax_ukernel_7p7x__wasmsimd_c24_acc2()
477 const v128_t vabsacc45 = wasm_v32x4_shuffle(vabsacc4567, vzero, 0, 4, 1, 5); in xnn_qs8_gavgpool_minmax_ukernel_7p7x__wasmsimd_c24_acc2()
478 const v128_t vabsacc67 = wasm_v32x4_shuffle(vabsacc4567, vzero, 2, 6, 3, 7); in xnn_qs8_gavgpool_minmax_ukernel_7p7x__wasmsimd_c24_acc2()
|
| D | 7p7x-minmax-ssse3-c16-acc2.c | 283 const __m128i vabsacc4567 = _mm_abs_epi32(vacc4567); in xnn_qs8_gavgpool_minmax_ukernel_7p7x__ssse3_c16_acc2() local
288 const __m128i vabsacc57 = _mm_shuffle_epi32(vabsacc4567, _MM_SHUFFLE(3, 3, 1, 1)); in xnn_qs8_gavgpool_minmax_ukernel_7p7x__ssse3_c16_acc2()
294 const __m128i vabsprod46 = _mm_mul_epu32(vabsacc4567, vmultiplier); in xnn_qs8_gavgpool_minmax_ukernel_7p7x__ssse3_c16_acc2()
386 const __m128i vabsacc4567 = _mm_abs_epi32(vacc4567); in xnn_qs8_gavgpool_minmax_ukernel_7p7x__ssse3_c16_acc2() local
389 const __m128i vabsacc57 = _mm_shuffle_epi32(vabsacc4567, _MM_SHUFFLE(3, 3, 1, 1)); in xnn_qs8_gavgpool_minmax_ukernel_7p7x__ssse3_c16_acc2()
393 const __m128i vabsprod46 = _mm_mul_epu32(vabsacc4567, vmultiplier); in xnn_qs8_gavgpool_minmax_ukernel_7p7x__ssse3_c16_acc2()
|
| D | 7p7x-minmax-sse2-c16-acc2.c | 288 const __m128i vabsacc4567 = _mm_sub_epi32(_mm_xor_si128(vacc4567, vsgnacc4567), vsgnacc4567); in xnn_qs8_gavgpool_minmax_ukernel_7p7x__sse2_c16_acc2() local
293 const __m128i vabsacc57 = _mm_shuffle_epi32(vabsacc4567, _MM_SHUFFLE(3, 3, 1, 1)); in xnn_qs8_gavgpool_minmax_ukernel_7p7x__sse2_c16_acc2()
299 const __m128i vabsprod46 = _mm_mul_epu32(vabsacc4567, vmultiplier); in xnn_qs8_gavgpool_minmax_ukernel_7p7x__sse2_c16_acc2()
394 const __m128i vabsacc4567 = _mm_sub_epi32(_mm_xor_si128(vacc4567, vsgnacc4567), vsgnacc4567); in xnn_qs8_gavgpool_minmax_ukernel_7p7x__sse2_c16_acc2() local
397 const __m128i vabsacc57 = _mm_shuffle_epi32(vabsacc4567, _MM_SHUFFLE(3, 3, 1, 1)); in xnn_qs8_gavgpool_minmax_ukernel_7p7x__sse2_c16_acc2()
401 const __m128i vabsprod46 = _mm_mul_epu32(vabsacc4567, vmultiplier); in xnn_qs8_gavgpool_minmax_ukernel_7p7x__sse2_c16_acc2()
|
| D | 7p7x-minmax-sse41-c24-acc2.c | 360 const __m128i vabsacc4567 = _mm_abs_epi32(vacc4567); in xnn_qs8_gavgpool_minmax_ukernel_7p7x__sse41_c24_acc2() local
367 const __m128i vabsacc57 = _mm_shuffle_epi32(vabsacc4567, _MM_SHUFFLE(3, 3, 1, 1)); in xnn_qs8_gavgpool_minmax_ukernel_7p7x__sse41_c24_acc2()
375 const __m128i vabsprod46 = _mm_mul_epu32(vabsacc4567, vmultiplier); in xnn_qs8_gavgpool_minmax_ukernel_7p7x__sse41_c24_acc2()
466 const __m128i vabsacc4567 = _mm_abs_epi32(vacc4567); in xnn_qs8_gavgpool_minmax_ukernel_7p7x__sse41_c24_acc2() local
469 const __m128i vabsacc57 = _mm_shuffle_epi32(vabsacc4567, _MM_SHUFFLE(3, 3, 1, 1)); in xnn_qs8_gavgpool_minmax_ukernel_7p7x__sse41_c24_acc2()
473 const __m128i vabsprod46 = _mm_mul_epu32(vabsacc4567, vmultiplier); in xnn_qs8_gavgpool_minmax_ukernel_7p7x__sse41_c24_acc2()
|
| /external/XNNPACK/src/qs8-dwconv/gen/ |
| D | up8x9-minmax-sse2-mul16.c | 213 const __m128i vabsacc4567 = _mm_sub_epi32(_mm_xor_si128(vacc4567, vnmask4567), vnmask4567); in xnn_qs8_dwconv_minmax_ukernel_up8x9__sse2_mul16() local
218 const __m128i vabsacc57 = _mm_shuffle_epi32(vabsacc4567, _MM_SHUFFLE(3, 3, 1, 1)); in xnn_qs8_dwconv_minmax_ukernel_up8x9__sse2_mul16()
219 const __m128i vabsprod46 = _mm_mul_epu32(vabsacc4567, vmultiplier); in xnn_qs8_dwconv_minmax_ukernel_up8x9__sse2_mul16()
392 const __m128i vabsacc4567 = _mm_sub_epi32(_mm_xor_si128(vacc4567, vnmask4567), vnmask4567); in xnn_qs8_dwconv_minmax_ukernel_up8x9__sse2_mul16() local
395 const __m128i vabsacc57 = _mm_shuffle_epi32(vabsacc4567, _MM_SHUFFLE(3, 3, 1, 1)); in xnn_qs8_dwconv_minmax_ukernel_up8x9__sse2_mul16()
399 const __m128i vabsprod46 = _mm_mul_epu32(vabsacc4567, vmultiplier); in xnn_qs8_dwconv_minmax_ukernel_up8x9__sse2_mul16()
|
| D | up8x9-minmax-ssse3-mul16.c | 213 const __m128i vabsacc4567 = _mm_abs_epi32(vacc4567); in xnn_qs8_dwconv_minmax_ukernel_up8x9__ssse3_mul16() local
218 const __m128i vabsacc57 = _mm_shuffle_epi32(vabsacc4567, _MM_SHUFFLE(3, 3, 1, 1)); in xnn_qs8_dwconv_minmax_ukernel_up8x9__ssse3_mul16()
219 const __m128i vabsprod46 = _mm_mul_epu32(vabsacc4567, vmultiplier); in xnn_qs8_dwconv_minmax_ukernel_up8x9__ssse3_mul16()
392 const __m128i vabsacc4567 = _mm_abs_epi32(vacc4567); in xnn_qs8_dwconv_minmax_ukernel_up8x9__ssse3_mul16() local
395 const __m128i vabsacc57 = _mm_shuffle_epi32(vabsacc4567, _MM_SHUFFLE(3, 3, 1, 1)); in xnn_qs8_dwconv_minmax_ukernel_up8x9__ssse3_mul16()
399 const __m128i vabsprod46 = _mm_mul_epu32(vabsacc4567, vmultiplier); in xnn_qs8_dwconv_minmax_ukernel_up8x9__ssse3_mul16()
|
| D | up16x9-minmax-ssse3-mul16.c | 289 const __m128i vabsacc4567 = _mm_abs_epi32(vacc4567); in xnn_qs8_dwconv_minmax_ukernel_up16x9__ssse3_mul16() local
296 const __m128i vabsacc57 = _mm_shuffle_epi32(vabsacc4567, _MM_SHUFFLE(3, 3, 1, 1)); in xnn_qs8_dwconv_minmax_ukernel_up16x9__ssse3_mul16()
297 const __m128i vabsprod46 = _mm_mul_epu32(vabsacc4567, vmultiplier); in xnn_qs8_dwconv_minmax_ukernel_up16x9__ssse3_mul16()
516 const __m128i vabsacc4567 = _mm_abs_epi32(vacc4567); in xnn_qs8_dwconv_minmax_ukernel_up16x9__ssse3_mul16() local
519 const __m128i vabsacc57 = _mm_shuffle_epi32(vabsacc4567, _MM_SHUFFLE(3, 3, 1, 1)); in xnn_qs8_dwconv_minmax_ukernel_up16x9__ssse3_mul16()
523 const __m128i vabsprod46 = _mm_mul_epu32(vabsacc4567, vmultiplier); in xnn_qs8_dwconv_minmax_ukernel_up16x9__ssse3_mul16()
|