| /external/XNNPACK/src/qs8-gavgpool/gen/ |
| D | 7p7x-minmax-wasmsimd-c8-acc2.c | 54 const v128_t vxi5x01234567 = wasm_i16x8_load_8x8(i5); in xnn_qs8_gavgpool_minmax_ukernel_7p7x__wasmsimd_c8_acc2() local
63 vacc1x01234567 = wasm_i16x8_add(vacc1x01234567, vxi5x01234567); in xnn_qs8_gavgpool_minmax_ukernel_7p7x__wasmsimd_c8_acc2()
99 const v128_t vxi5x01234567 = wasm_i16x8_load_8x8(i5); in xnn_qs8_gavgpool_minmax_ukernel_7p7x__wasmsimd_c8_acc2() local
108 vacc1x01234567 = wasm_i16x8_add(vacc1x01234567, vxi5x01234567); in xnn_qs8_gavgpool_minmax_ukernel_7p7x__wasmsimd_c8_acc2()
164 const v128_t vxi5x01234567 = wasm_i16x8_load_8x8(i5); in xnn_qs8_gavgpool_minmax_ukernel_7p7x__wasmsimd_c8_acc2() local
173 vacc1x01234567 = wasm_i16x8_add(vacc1x01234567, vxi5x01234567); in xnn_qs8_gavgpool_minmax_ukernel_7p7x__wasmsimd_c8_acc2()
234 const v128_t vxi5x01234567 = wasm_i16x8_load_8x8(i5); in xnn_qs8_gavgpool_minmax_ukernel_7p7x__wasmsimd_c8_acc2() local
243 vacc1x01234567 = wasm_i16x8_add(vacc1x01234567, vxi5x01234567); in xnn_qs8_gavgpool_minmax_ukernel_7p7x__wasmsimd_c8_acc2()
|
| D | 7p7x-minmax-sse41-c8-acc2.c | 54 const __m128i vxi5x01234567 = _mm_cvtepi8_epi16(_mm_loadl_epi64((const __m128i*) i5)); in xnn_qs8_gavgpool_minmax_ukernel_7p7x__sse41_c8_acc2() local
64 vacc1x01234567 = _mm_add_epi16(vacc1x01234567, vxi5x01234567); in xnn_qs8_gavgpool_minmax_ukernel_7p7x__sse41_c8_acc2()
100 const __m128i vxi5x01234567 = _mm_cvtepi8_epi16(_mm_loadl_epi64((const __m128i*) i5)); in xnn_qs8_gavgpool_minmax_ukernel_7p7x__sse41_c8_acc2() local
110 vacc1x01234567 = _mm_add_epi16(vacc1x01234567, vxi5x01234567); in xnn_qs8_gavgpool_minmax_ukernel_7p7x__sse41_c8_acc2()
165 const __m128i vxi5x01234567 = _mm_cvtepi8_epi16(_mm_loadl_epi64((const __m128i*) i5)); in xnn_qs8_gavgpool_minmax_ukernel_7p7x__sse41_c8_acc2() local
175 vacc1x01234567 = _mm_add_epi16(vacc1x01234567, vxi5x01234567); in xnn_qs8_gavgpool_minmax_ukernel_7p7x__sse41_c8_acc2()
233 const __m128i vxi5x01234567 = _mm_cvtepi8_epi16(_mm_loadl_epi64((const __m128i*) i5)); in xnn_qs8_gavgpool_minmax_ukernel_7p7x__sse41_c8_acc2() local
243 vacc1x01234567 = _mm_add_epi16(vacc1x01234567, vxi5x01234567); in xnn_qs8_gavgpool_minmax_ukernel_7p7x__sse41_c8_acc2()
|
| D | 7p7x-minmax-ssse3-c8-acc2.c | 64 …const __m128i vxi5x01234567 = _mm_unpacklo_epi8(vi5x01234567, _mm_cmpgt_epi8(_mm_setzero_si128(), … in xnn_qs8_gavgpool_minmax_ukernel_7p7x__ssse3_c8_acc2() local
71 vacc1x01234567 = _mm_add_epi16(vacc1x01234567, vxi5x01234567); in xnn_qs8_gavgpool_minmax_ukernel_7p7x__ssse3_c8_acc2()
118 …const __m128i vxi5x01234567 = _mm_unpacklo_epi8(vi5x01234567, _mm_cmpgt_epi8(_mm_setzero_si128(), … in xnn_qs8_gavgpool_minmax_ukernel_7p7x__ssse3_c8_acc2() local
125 vacc1x01234567 = _mm_add_epi16(vacc1x01234567, vxi5x01234567); in xnn_qs8_gavgpool_minmax_ukernel_7p7x__ssse3_c8_acc2()
191 …const __m128i vxi5x01234567 = _mm_unpacklo_epi8(vi5x01234567, _mm_cmpgt_epi8(_mm_setzero_si128(), … in xnn_qs8_gavgpool_minmax_ukernel_7p7x__ssse3_c8_acc2() local
198 vacc1x01234567 = _mm_add_epi16(vacc1x01234567, vxi5x01234567); in xnn_qs8_gavgpool_minmax_ukernel_7p7x__ssse3_c8_acc2()
272 …const __m128i vxi5x01234567 = _mm_unpacklo_epi8(vi5x01234567, _mm_cmpgt_epi8(_mm_setzero_si128(), … in xnn_qs8_gavgpool_minmax_ukernel_7p7x__ssse3_c8_acc2() local
279 vacc1x01234567 = _mm_add_epi16(vacc1x01234567, vxi5x01234567); in xnn_qs8_gavgpool_minmax_ukernel_7p7x__ssse3_c8_acc2()
|
| D | 7p7x-minmax-sse2-c8-acc2.c | 64 …const __m128i vxi5x01234567 = _mm_unpacklo_epi8(vi5x01234567, _mm_cmpgt_epi8(_mm_setzero_si128(), … in xnn_qs8_gavgpool_minmax_ukernel_7p7x__sse2_c8_acc2() local
71 vacc1x01234567 = _mm_add_epi16(vacc1x01234567, vxi5x01234567); in xnn_qs8_gavgpool_minmax_ukernel_7p7x__sse2_c8_acc2()
118 …const __m128i vxi5x01234567 = _mm_unpacklo_epi8(vi5x01234567, _mm_cmpgt_epi8(_mm_setzero_si128(), … in xnn_qs8_gavgpool_minmax_ukernel_7p7x__sse2_c8_acc2() local
125 vacc1x01234567 = _mm_add_epi16(vacc1x01234567, vxi5x01234567); in xnn_qs8_gavgpool_minmax_ukernel_7p7x__sse2_c8_acc2()
191 …const __m128i vxi5x01234567 = _mm_unpacklo_epi8(vi5x01234567, _mm_cmpgt_epi8(_mm_setzero_si128(), … in xnn_qs8_gavgpool_minmax_ukernel_7p7x__sse2_c8_acc2() local
198 vacc1x01234567 = _mm_add_epi16(vacc1x01234567, vxi5x01234567); in xnn_qs8_gavgpool_minmax_ukernel_7p7x__sse2_c8_acc2()
275 …const __m128i vxi5x01234567 = _mm_unpacklo_epi8(vi5x01234567, _mm_cmpgt_epi8(_mm_setzero_si128(), … in xnn_qs8_gavgpool_minmax_ukernel_7p7x__sse2_c8_acc2() local
282 vacc1x01234567 = _mm_add_epi16(vacc1x01234567, vxi5x01234567); in xnn_qs8_gavgpool_minmax_ukernel_7p7x__sse2_c8_acc2()
|
| D | 7p7x-minmax-wasmsimd-c24-acc2.c | 64 const v128_t vxi5x01234567 = wasm_i16x8_load_8x8(i5); in xnn_qs8_gavgpool_minmax_ukernel_7p7x__wasmsimd_c24_acc2() local
83 vacc1x01234567 = wasm_i16x8_add(vacc1x01234567, vxi5x01234567); in xnn_qs8_gavgpool_minmax_ukernel_7p7x__wasmsimd_c24_acc2()
122 const v128_t vxi5x01234567 = wasm_i16x8_load_8x8(i5); in xnn_qs8_gavgpool_minmax_ukernel_7p7x__wasmsimd_c24_acc2() local
131 vacc1x01234567 = wasm_i16x8_add(vacc1x01234567, vxi5x01234567); in xnn_qs8_gavgpool_minmax_ukernel_7p7x__wasmsimd_c24_acc2()
180 const v128_t vxi5x01234567 = wasm_i16x8_load_8x8(i5); in xnn_qs8_gavgpool_minmax_ukernel_7p7x__wasmsimd_c24_acc2() local
199 vacc1x01234567 = wasm_i16x8_add(vacc1x01234567, vxi5x01234567); in xnn_qs8_gavgpool_minmax_ukernel_7p7x__wasmsimd_c24_acc2()
238 const v128_t vxi5x01234567 = wasm_i16x8_load_8x8(i5); in xnn_qs8_gavgpool_minmax_ukernel_7p7x__wasmsimd_c24_acc2() local
247 vacc1x01234567 = wasm_i16x8_add(vacc1x01234567, vxi5x01234567); in xnn_qs8_gavgpool_minmax_ukernel_7p7x__wasmsimd_c24_acc2()
316 const v128_t vxi5x01234567 = wasm_i16x8_load_8x8(i5); in xnn_qs8_gavgpool_minmax_ukernel_7p7x__wasmsimd_c24_acc2() local
335 vacc1x01234567 = wasm_i16x8_add(vacc1x01234567, vxi5x01234567); in xnn_qs8_gavgpool_minmax_ukernel_7p7x__wasmsimd_c24_acc2()
[all …]
|
| D | 7p7x-minmax-sse41-c24-acc2.c | 64 const __m128i vxi5x01234567 = _mm_cvtepi8_epi16(_mm_loadl_epi64((const __m128i*) i5)); in xnn_qs8_gavgpool_minmax_ukernel_7p7x__sse41_c24_acc2() local
84 vacc1x01234567 = _mm_add_epi16(vacc1x01234567, vxi5x01234567); in xnn_qs8_gavgpool_minmax_ukernel_7p7x__sse41_c24_acc2()
123 const __m128i vxi5x01234567 = _mm_cvtepi8_epi16(_mm_loadl_epi64((const __m128i*) i5)); in xnn_qs8_gavgpool_minmax_ukernel_7p7x__sse41_c24_acc2() local
133 vacc1x01234567 = _mm_add_epi16(vacc1x01234567, vxi5x01234567); in xnn_qs8_gavgpool_minmax_ukernel_7p7x__sse41_c24_acc2()
182 const __m128i vxi5x01234567 = _mm_cvtepi8_epi16(_mm_loadl_epi64((const __m128i*) i5)); in xnn_qs8_gavgpool_minmax_ukernel_7p7x__sse41_c24_acc2() local
202 vacc1x01234567 = _mm_add_epi16(vacc1x01234567, vxi5x01234567); in xnn_qs8_gavgpool_minmax_ukernel_7p7x__sse41_c24_acc2()
241 const __m128i vxi5x01234567 = _mm_cvtepi8_epi16(_mm_loadl_epi64((const __m128i*) i5)); in xnn_qs8_gavgpool_minmax_ukernel_7p7x__sse41_c24_acc2() local
251 vacc1x01234567 = _mm_add_epi16(vacc1x01234567, vxi5x01234567); in xnn_qs8_gavgpool_minmax_ukernel_7p7x__sse41_c24_acc2()
319 const __m128i vxi5x01234567 = _mm_cvtepi8_epi16(_mm_loadl_epi64((const __m128i*) i5)); in xnn_qs8_gavgpool_minmax_ukernel_7p7x__sse41_c24_acc2() local
339 vacc1x01234567 = _mm_add_epi16(vacc1x01234567, vxi5x01234567); in xnn_qs8_gavgpool_minmax_ukernel_7p7x__sse41_c24_acc2()
[all …]
|
| D | 7p7x-minmax-wasmsimd-c16-acc2.c | 59 const v128_t vxi5x01234567 = wasm_i16x8_load_8x8(i5); in xnn_qs8_gavgpool_minmax_ukernel_7p7x__wasmsimd_c16_acc2() local
73 vacc1x01234567 = wasm_i16x8_add(vacc1x01234567, vxi5x01234567); in xnn_qs8_gavgpool_minmax_ukernel_7p7x__wasmsimd_c16_acc2()
121 const v128_t vxi5x01234567 = wasm_i16x8_load_8x8(i5); in xnn_qs8_gavgpool_minmax_ukernel_7p7x__wasmsimd_c16_acc2() local
135 vacc1x01234567 = wasm_i16x8_add(vacc1x01234567, vxi5x01234567); in xnn_qs8_gavgpool_minmax_ukernel_7p7x__wasmsimd_c16_acc2()
203 const v128_t vxi5x01234567 = wasm_i16x8_load_8x8(i5); in xnn_qs8_gavgpool_minmax_ukernel_7p7x__wasmsimd_c16_acc2() local
217 vacc1x01234567 = wasm_i16x8_add(vacc1x01234567, vxi5x01234567); in xnn_qs8_gavgpool_minmax_ukernel_7p7x__wasmsimd_c16_acc2()
304 const v128_t vxi5x01234567 = wasm_i16x8_load_8x8(i5); in xnn_qs8_gavgpool_minmax_ukernel_7p7x__wasmsimd_c16_acc2() local
313 vacc1x01234567 = wasm_i16x8_add(vacc1x01234567, vxi5x01234567); in xnn_qs8_gavgpool_minmax_ukernel_7p7x__wasmsimd_c16_acc2()
|
| D | 7p7x-minmax-sse41-c16-acc2.c | 59 const __m128i vxi5x01234567 = _mm_cvtepi8_epi16(_mm_loadl_epi64((const __m128i*) i5)); in xnn_qs8_gavgpool_minmax_ukernel_7p7x__sse41_c16_acc2() local
74 vacc1x01234567 = _mm_add_epi16(vacc1x01234567, vxi5x01234567); in xnn_qs8_gavgpool_minmax_ukernel_7p7x__sse41_c16_acc2()
122 const __m128i vxi5x01234567 = _mm_cvtepi8_epi16(_mm_loadl_epi64((const __m128i*) i5)); in xnn_qs8_gavgpool_minmax_ukernel_7p7x__sse41_c16_acc2() local
137 vacc1x01234567 = _mm_add_epi16(vacc1x01234567, vxi5x01234567); in xnn_qs8_gavgpool_minmax_ukernel_7p7x__sse41_c16_acc2()
204 const __m128i vxi5x01234567 = _mm_cvtepi8_epi16(_mm_loadl_epi64((const __m128i*) i5)); in xnn_qs8_gavgpool_minmax_ukernel_7p7x__sse41_c16_acc2() local
219 vacc1x01234567 = _mm_add_epi16(vacc1x01234567, vxi5x01234567); in xnn_qs8_gavgpool_minmax_ukernel_7p7x__sse41_c16_acc2()
300 const __m128i vxi5x01234567 = _mm_cvtepi8_epi16(_mm_loadl_epi64((const __m128i*) i5)); in xnn_qs8_gavgpool_minmax_ukernel_7p7x__sse41_c16_acc2() local
310 vacc1x01234567 = _mm_add_epi16(vacc1x01234567, vxi5x01234567); in xnn_qs8_gavgpool_minmax_ukernel_7p7x__sse41_c16_acc2()
|
| D | 7x-minmax-wasmsimd-c8-acc2.c | 72 const v128_t vxi5x01234567 = wasm_i16x8_load_8x8(i5); in xnn_qs8_gavgpool_minmax_ukernel_7x__wasmsimd_c8_acc2() local
81 vacc1x01234567 = wasm_i16x8_add(vacc1x01234567, vxi5x01234567); in xnn_qs8_gavgpool_minmax_ukernel_7x__wasmsimd_c8_acc2()
141 const v128_t vxi5x01234567 = wasm_i16x8_load_8x8(i5); in xnn_qs8_gavgpool_minmax_ukernel_7x__wasmsimd_c8_acc2() local
150 vacc1x01234567 = wasm_i16x8_add(vacc1x01234567, vxi5x01234567); in xnn_qs8_gavgpool_minmax_ukernel_7x__wasmsimd_c8_acc2()
|
| D | 7x-minmax-sse41-c8-acc2.c | 71 const __m128i vxi5x01234567 = _mm_cvtepi8_epi16(_mm_loadl_epi64((const __m128i*) i5)); in xnn_qs8_gavgpool_minmax_ukernel_7x__sse41_c8_acc2() local
81 vacc1x01234567 = _mm_add_epi16(vacc1x01234567, vxi5x01234567); in xnn_qs8_gavgpool_minmax_ukernel_7x__sse41_c8_acc2()
138 const __m128i vxi5x01234567 = _mm_cvtepi8_epi16(_mm_loadl_epi64((const __m128i*) i5)); in xnn_qs8_gavgpool_minmax_ukernel_7x__sse41_c8_acc2() local
148 vacc1x01234567 = _mm_add_epi16(vacc1x01234567, vxi5x01234567); in xnn_qs8_gavgpool_minmax_ukernel_7x__sse41_c8_acc2()
|
| D | 7p7x-minmax-sse2-c24-acc2.c | 88 …const __m128i vxi5x01234567 = _mm_unpacklo_epi8(vi5x01234567, _mm_cmpgt_epi8(_mm_setzero_si128(), … in xnn_qs8_gavgpool_minmax_ukernel_7p7x__sse2_c24_acc2() local
105 vacc1x01234567 = _mm_add_epi16(vacc1x01234567, vxi5x01234567); in xnn_qs8_gavgpool_minmax_ukernel_7p7x__sse2_c24_acc2()
157 …const __m128i vxi5x01234567 = _mm_unpacklo_epi8(vi5x01234567, _mm_cmpgt_epi8(_mm_setzero_si128(), … in xnn_qs8_gavgpool_minmax_ukernel_7p7x__sse2_c24_acc2() local
164 vacc1x01234567 = _mm_add_epi16(vacc1x01234567, vxi5x01234567); in xnn_qs8_gavgpool_minmax_ukernel_7p7x__sse2_c24_acc2()
238 …const __m128i vxi5x01234567 = _mm_unpacklo_epi8(vi5x01234567, _mm_cmpgt_epi8(_mm_setzero_si128(), … in xnn_qs8_gavgpool_minmax_ukernel_7p7x__sse2_c24_acc2() local
255 vacc1x01234567 = _mm_add_epi16(vacc1x01234567, vxi5x01234567); in xnn_qs8_gavgpool_minmax_ukernel_7p7x__sse2_c24_acc2()
307 …const __m128i vxi5x01234567 = _mm_unpacklo_epi8(vi5x01234567, _mm_cmpgt_epi8(_mm_setzero_si128(), … in xnn_qs8_gavgpool_minmax_ukernel_7p7x__sse2_c24_acc2() local
314 vacc1x01234567 = _mm_add_epi16(vacc1x01234567, vxi5x01234567); in xnn_qs8_gavgpool_minmax_ukernel_7p7x__sse2_c24_acc2()
407 …const __m128i vxi5x01234567 = _mm_unpacklo_epi8(vi5x01234567, _mm_cmpgt_epi8(_mm_setzero_si128(), … in xnn_qs8_gavgpool_minmax_ukernel_7p7x__sse2_c24_acc2() local
424 vacc1x01234567 = _mm_add_epi16(vacc1x01234567, vxi5x01234567); in xnn_qs8_gavgpool_minmax_ukernel_7p7x__sse2_c24_acc2()
[all …]
|
| D | 7p7x-minmax-ssse3-c24-acc2.c | 88 …const __m128i vxi5x01234567 = _mm_unpacklo_epi8(vi5x01234567, _mm_cmpgt_epi8(_mm_setzero_si128(), … in xnn_qs8_gavgpool_minmax_ukernel_7p7x__ssse3_c24_acc2() local
105 vacc1x01234567 = _mm_add_epi16(vacc1x01234567, vxi5x01234567); in xnn_qs8_gavgpool_minmax_ukernel_7p7x__ssse3_c24_acc2()
157 …const __m128i vxi5x01234567 = _mm_unpacklo_epi8(vi5x01234567, _mm_cmpgt_epi8(_mm_setzero_si128(), … in xnn_qs8_gavgpool_minmax_ukernel_7p7x__ssse3_c24_acc2() local
164 vacc1x01234567 = _mm_add_epi16(vacc1x01234567, vxi5x01234567); in xnn_qs8_gavgpool_minmax_ukernel_7p7x__ssse3_c24_acc2()
238 …const __m128i vxi5x01234567 = _mm_unpacklo_epi8(vi5x01234567, _mm_cmpgt_epi8(_mm_setzero_si128(), … in xnn_qs8_gavgpool_minmax_ukernel_7p7x__ssse3_c24_acc2() local
255 vacc1x01234567 = _mm_add_epi16(vacc1x01234567, vxi5x01234567); in xnn_qs8_gavgpool_minmax_ukernel_7p7x__ssse3_c24_acc2()
307 …const __m128i vxi5x01234567 = _mm_unpacklo_epi8(vi5x01234567, _mm_cmpgt_epi8(_mm_setzero_si128(), … in xnn_qs8_gavgpool_minmax_ukernel_7p7x__ssse3_c24_acc2() local
314 vacc1x01234567 = _mm_add_epi16(vacc1x01234567, vxi5x01234567); in xnn_qs8_gavgpool_minmax_ukernel_7p7x__ssse3_c24_acc2()
407 …const __m128i vxi5x01234567 = _mm_unpacklo_epi8(vi5x01234567, _mm_cmpgt_epi8(_mm_setzero_si128(), … in xnn_qs8_gavgpool_minmax_ukernel_7p7x__ssse3_c24_acc2() local
424 vacc1x01234567 = _mm_add_epi16(vacc1x01234567, vxi5x01234567); in xnn_qs8_gavgpool_minmax_ukernel_7p7x__ssse3_c24_acc2()
[all …]
|
| D | 7x-minmax-ssse3-c8-acc2.c | 81 …const __m128i vxi5x01234567 = _mm_unpacklo_epi8(vi5x01234567, _mm_cmpgt_epi8(_mm_setzero_si128(), … in xnn_qs8_gavgpool_minmax_ukernel_7x__ssse3_c8_acc2() local
88 vacc1x01234567 = _mm_add_epi16(vacc1x01234567, vxi5x01234567); in xnn_qs8_gavgpool_minmax_ukernel_7x__ssse3_c8_acc2()
161 …const __m128i vxi5x01234567 = _mm_unpacklo_epi8(vi5x01234567, _mm_cmpgt_epi8(_mm_setzero_si128(), … in xnn_qs8_gavgpool_minmax_ukernel_7x__ssse3_c8_acc2() local
168 vacc1x01234567 = _mm_add_epi16(vacc1x01234567, vxi5x01234567); in xnn_qs8_gavgpool_minmax_ukernel_7x__ssse3_c8_acc2()
|
| D | 7p7x-minmax-ssse3-c16-acc2.c | 76 …const __m128i vxi5x01234567 = _mm_unpacklo_epi8(vi5x01234567, _mm_cmpgt_epi8(_mm_setzero_si128(), … in xnn_qs8_gavgpool_minmax_ukernel_7p7x__ssse3_c16_acc2() local
88 vacc1x01234567 = _mm_add_epi16(vacc1x01234567, vxi5x01234567); in xnn_qs8_gavgpool_minmax_ukernel_7p7x__ssse3_c16_acc2()
155 …const __m128i vxi5x01234567 = _mm_unpacklo_epi8(vi5x01234567, _mm_cmpgt_epi8(_mm_setzero_si128(), … in xnn_qs8_gavgpool_minmax_ukernel_7p7x__ssse3_c16_acc2() local
167 vacc1x01234567 = _mm_add_epi16(vacc1x01234567, vxi5x01234567); in xnn_qs8_gavgpool_minmax_ukernel_7p7x__ssse3_c16_acc2()
253 …const __m128i vxi5x01234567 = _mm_unpacklo_epi8(vi5x01234567, _mm_cmpgt_epi8(_mm_setzero_si128(), … in xnn_qs8_gavgpool_minmax_ukernel_7p7x__ssse3_c16_acc2() local
265 vacc1x01234567 = _mm_add_epi16(vacc1x01234567, vxi5x01234567); in xnn_qs8_gavgpool_minmax_ukernel_7p7x__ssse3_c16_acc2()
367 …const __m128i vxi5x01234567 = _mm_unpacklo_epi8(vi5x01234567, _mm_cmpgt_epi8(_mm_setzero_si128(), … in xnn_qs8_gavgpool_minmax_ukernel_7p7x__ssse3_c16_acc2() local
374 vacc1x01234567 = _mm_add_epi16(vacc1x01234567, vxi5x01234567); in xnn_qs8_gavgpool_minmax_ukernel_7p7x__ssse3_c16_acc2()
|
| D | 7x-minmax-sse2-c8-acc2.c | 81 …const __m128i vxi5x01234567 = _mm_unpacklo_epi8(vi5x01234567, _mm_cmpgt_epi8(_mm_setzero_si128(), … in xnn_qs8_gavgpool_minmax_ukernel_7x__sse2_c8_acc2() local
88 vacc1x01234567 = _mm_add_epi16(vacc1x01234567, vxi5x01234567); in xnn_qs8_gavgpool_minmax_ukernel_7x__sse2_c8_acc2()
164 …const __m128i vxi5x01234567 = _mm_unpacklo_epi8(vi5x01234567, _mm_cmpgt_epi8(_mm_setzero_si128(), … in xnn_qs8_gavgpool_minmax_ukernel_7x__sse2_c8_acc2() local
171 vacc1x01234567 = _mm_add_epi16(vacc1x01234567, vxi5x01234567); in xnn_qs8_gavgpool_minmax_ukernel_7x__sse2_c8_acc2()
|
| D | 7x-minmax-sse41-c16-acc2.c | 76 const __m128i vxi5x01234567 = _mm_cvtepi8_epi16(_mm_loadl_epi64((const __m128i*) i5)); in xnn_qs8_gavgpool_minmax_ukernel_7x__sse41_c16_acc2() local
91 vacc1x01234567 = _mm_add_epi16(vacc1x01234567, vxi5x01234567); in xnn_qs8_gavgpool_minmax_ukernel_7x__sse41_c16_acc2()
171 const __m128i vxi5x01234567 = _mm_cvtepi8_epi16(_mm_loadl_epi64((const __m128i*) i5)); in xnn_qs8_gavgpool_minmax_ukernel_7x__sse41_c16_acc2() local
181 vacc1x01234567 = _mm_add_epi16(vacc1x01234567, vxi5x01234567); in xnn_qs8_gavgpool_minmax_ukernel_7x__sse41_c16_acc2()
|
| D | 7x-minmax-wasmsimd-c16-acc2.c | 77 const v128_t vxi5x01234567 = wasm_i16x8_load_8x8(i5); in xnn_qs8_gavgpool_minmax_ukernel_7x__wasmsimd_c16_acc2() local
91 vacc1x01234567 = wasm_i16x8_add(vacc1x01234567, vxi5x01234567); in xnn_qs8_gavgpool_minmax_ukernel_7x__wasmsimd_c16_acc2()
177 const v128_t vxi5x01234567 = wasm_i16x8_load_8x8(i5); in xnn_qs8_gavgpool_minmax_ukernel_7x__wasmsimd_c16_acc2() local
186 vacc1x01234567 = wasm_i16x8_add(vacc1x01234567, vxi5x01234567); in xnn_qs8_gavgpool_minmax_ukernel_7x__wasmsimd_c16_acc2()
|
| D | 7p7x-minmax-sse2-c16-acc2.c | 76 …const __m128i vxi5x01234567 = _mm_unpacklo_epi8(vi5x01234567, _mm_cmpgt_epi8(_mm_setzero_si128(), … in xnn_qs8_gavgpool_minmax_ukernel_7p7x__sse2_c16_acc2() local
88 vacc1x01234567 = _mm_add_epi16(vacc1x01234567, vxi5x01234567); in xnn_qs8_gavgpool_minmax_ukernel_7p7x__sse2_c16_acc2()
155 …const __m128i vxi5x01234567 = _mm_unpacklo_epi8(vi5x01234567, _mm_cmpgt_epi8(_mm_setzero_si128(), … in xnn_qs8_gavgpool_minmax_ukernel_7p7x__sse2_c16_acc2() local
167 vacc1x01234567 = _mm_add_epi16(vacc1x01234567, vxi5x01234567); in xnn_qs8_gavgpool_minmax_ukernel_7p7x__sse2_c16_acc2()
253 …const __m128i vxi5x01234567 = _mm_unpacklo_epi8(vi5x01234567, _mm_cmpgt_epi8(_mm_setzero_si128(), … in xnn_qs8_gavgpool_minmax_ukernel_7p7x__sse2_c16_acc2() local
265 vacc1x01234567 = _mm_add_epi16(vacc1x01234567, vxi5x01234567); in xnn_qs8_gavgpool_minmax_ukernel_7p7x__sse2_c16_acc2()
372 …const __m128i vxi5x01234567 = _mm_unpacklo_epi8(vi5x01234567, _mm_cmpgt_epi8(_mm_setzero_si128(), … in xnn_qs8_gavgpool_minmax_ukernel_7p7x__sse2_c16_acc2() local
379 vacc1x01234567 = _mm_add_epi16(vacc1x01234567, vxi5x01234567); in xnn_qs8_gavgpool_minmax_ukernel_7p7x__sse2_c16_acc2()
|
| D | 7x-minmax-ssse3-c16-acc2.c | 93 …const __m128i vxi5x01234567 = _mm_unpacklo_epi8(vi5x01234567, _mm_cmpgt_epi8(_mm_setzero_si128(), … in xnn_qs8_gavgpool_minmax_ukernel_7x__ssse3_c16_acc2() local
105 vacc1x01234567 = _mm_add_epi16(vacc1x01234567, vxi5x01234567); in xnn_qs8_gavgpool_minmax_ukernel_7x__ssse3_c16_acc2()
206 …const __m128i vxi5x01234567 = _mm_unpacklo_epi8(vi5x01234567, _mm_cmpgt_epi8(_mm_setzero_si128(), … in xnn_qs8_gavgpool_minmax_ukernel_7x__ssse3_c16_acc2() local
213 vacc1x01234567 = _mm_add_epi16(vacc1x01234567, vxi5x01234567); in xnn_qs8_gavgpool_minmax_ukernel_7x__ssse3_c16_acc2()
|
| D | 7x-minmax-sse41-c24-acc2.c | 81 const __m128i vxi5x01234567 = _mm_cvtepi8_epi16(_mm_loadl_epi64((const __m128i*) i5)); in xnn_qs8_gavgpool_minmax_ukernel_7x__sse41_c24_acc2() local
101 vacc1x01234567 = _mm_add_epi16(vacc1x01234567, vxi5x01234567); in xnn_qs8_gavgpool_minmax_ukernel_7x__sse41_c24_acc2()
206 const __m128i vxi5x01234567 = _mm_cvtepi8_epi16(_mm_loadl_epi64((const __m128i*) i5)); in xnn_qs8_gavgpool_minmax_ukernel_7x__sse41_c24_acc2() local
216 vacc1x01234567 = _mm_add_epi16(vacc1x01234567, vxi5x01234567); in xnn_qs8_gavgpool_minmax_ukernel_7x__sse41_c24_acc2()
|
| D | 7x-minmax-wasmsimd-c24-acc2.c | 82 const v128_t vxi5x01234567 = wasm_i16x8_load_8x8(i5); in xnn_qs8_gavgpool_minmax_ukernel_7x__wasmsimd_c24_acc2() local
101 vacc1x01234567 = wasm_i16x8_add(vacc1x01234567, vxi5x01234567); in xnn_qs8_gavgpool_minmax_ukernel_7x__wasmsimd_c24_acc2()
215 const v128_t vxi5x01234567 = wasm_i16x8_load_8x8(i5); in xnn_qs8_gavgpool_minmax_ukernel_7x__wasmsimd_c24_acc2() local
224 vacc1x01234567 = wasm_i16x8_add(vacc1x01234567, vxi5x01234567); in xnn_qs8_gavgpool_minmax_ukernel_7x__wasmsimd_c24_acc2()
|
| D | 7x-minmax-sse2-c16-acc2.c | 93 …const __m128i vxi5x01234567 = _mm_unpacklo_epi8(vi5x01234567, _mm_cmpgt_epi8(_mm_setzero_si128(), … in xnn_qs8_gavgpool_minmax_ukernel_7x__sse2_c16_acc2() local
105 vacc1x01234567 = _mm_add_epi16(vacc1x01234567, vxi5x01234567); in xnn_qs8_gavgpool_minmax_ukernel_7x__sse2_c16_acc2()
211 …const __m128i vxi5x01234567 = _mm_unpacklo_epi8(vi5x01234567, _mm_cmpgt_epi8(_mm_setzero_si128(), … in xnn_qs8_gavgpool_minmax_ukernel_7x__sse2_c16_acc2() local
218 vacc1x01234567 = _mm_add_epi16(vacc1x01234567, vxi5x01234567); in xnn_qs8_gavgpool_minmax_ukernel_7x__sse2_c16_acc2()
|
| /external/XNNPACK/src/qs8-dwconv/gen/ |
| D | up8x9-minmax-sse41-mul16.c | 153 const __m128i vxi5x01234567 = _mm_cvtepi8_epi16(vi5x01234567); in xnn_qs8_dwconv_minmax_ukernel_up8x9__sse41_mul16() local
159 const __m128i vp5x01234567lo = _mm_mullo_epi16(vxi5x01234567, vxk5x01234567); in xnn_qs8_dwconv_minmax_ukernel_up8x9__sse41_mul16()
160 const __m128i vp5x01234567hi = _mm_mulhi_epi16(vxi5x01234567, vxk5x01234567); in xnn_qs8_dwconv_minmax_ukernel_up8x9__sse41_mul16()
316 const __m128i vxi5x01234567 = _mm_cvtepi8_epi16(vi5x01234567); in xnn_qs8_dwconv_minmax_ukernel_up8x9__sse41_mul16() local
321 const __m128i vp5x01234567lo = _mm_mullo_epi16(vxi5x01234567, vxk5x01234567); in xnn_qs8_dwconv_minmax_ukernel_up8x9__sse41_mul16()
322 const __m128i vp5x01234567hi = _mm_mulhi_epi16(vxi5x01234567, vxk5x01234567); in xnn_qs8_dwconv_minmax_ukernel_up8x9__sse41_mul16()
|
| D | up8x9-minmax-sse2-mul16.c | 156 …const __m128i vxi5x01234567 = _mm_unpacklo_epi8(vi5x01234567, _mm_cmpgt_epi8(_mm_setzero_si128(), … in xnn_qs8_dwconv_minmax_ukernel_up8x9__sse2_mul16() local
159 const __m128i vp5x01234567lo = _mm_mullo_epi16(vxi5x01234567, vxk5x01234567); in xnn_qs8_dwconv_minmax_ukernel_up8x9__sse2_mul16()
160 const __m128i vp5x01234567hi = _mm_mulhi_epi16(vxi5x01234567, vxk5x01234567); in xnn_qs8_dwconv_minmax_ukernel_up8x9__sse2_mul16()
339 …const __m128i vxi5x01234567 = _mm_unpacklo_epi8(vi5x01234567, _mm_cmpgt_epi8(_mm_setzero_si128(), … in xnn_qs8_dwconv_minmax_ukernel_up8x9__sse2_mul16() local
342 const __m128i vp5x01234567lo = _mm_mullo_epi16(vxi5x01234567, vxk5x01234567); in xnn_qs8_dwconv_minmax_ukernel_up8x9__sse2_mul16()
343 const __m128i vp5x01234567hi = _mm_mulhi_epi16(vxi5x01234567, vxk5x01234567); in xnn_qs8_dwconv_minmax_ukernel_up8x9__sse2_mul16()
|
| D | up8x9-minmax-ssse3-mul16.c | 156 …const __m128i vxi5x01234567 = _mm_unpacklo_epi8(vi5x01234567, _mm_cmpgt_epi8(_mm_setzero_si128(), … in xnn_qs8_dwconv_minmax_ukernel_up8x9__ssse3_mul16() local
159 const __m128i vp5x01234567lo = _mm_mullo_epi16(vxi5x01234567, vxk5x01234567); in xnn_qs8_dwconv_minmax_ukernel_up8x9__ssse3_mul16()
160 const __m128i vp5x01234567hi = _mm_mulhi_epi16(vxi5x01234567, vxk5x01234567); in xnn_qs8_dwconv_minmax_ukernel_up8x9__ssse3_mul16()
339 …const __m128i vxi5x01234567 = _mm_unpacklo_epi8(vi5x01234567, _mm_cmpgt_epi8(_mm_setzero_si128(), … in xnn_qs8_dwconv_minmax_ukernel_up8x9__ssse3_mul16() local
342 const __m128i vp5x01234567lo = _mm_mullo_epi16(vxi5x01234567, vxk5x01234567); in xnn_qs8_dwconv_minmax_ukernel_up8x9__ssse3_mul16()
343 const __m128i vp5x01234567hi = _mm_mulhi_epi16(vxi5x01234567, vxk5x01234567); in xnn_qs8_dwconv_minmax_ukernel_up8x9__ssse3_mul16()
|