| /external/XNNPACK/src/qs8-gavgpool/gen/ |
| D | 7p7x-minmax-wasmsimd-c8-acc2.c | 50 const v128_t vxi3x01234567 = wasm_i16x8_load_8x8(i3); in xnn_qs8_gavgpool_minmax_ukernel_7p7x__wasmsimd_c8_acc2() local
60 v128_t vacc1x01234567 = wasm_i16x8_add(vxi2x01234567, vxi3x01234567); in xnn_qs8_gavgpool_minmax_ukernel_7p7x__wasmsimd_c8_acc2()
95 const v128_t vxi3x01234567 = wasm_i16x8_load_8x8(i3); in xnn_qs8_gavgpool_minmax_ukernel_7p7x__wasmsimd_c8_acc2() local
105 v128_t vacc1x01234567 = wasm_i16x8_add(vxi2x01234567, vxi3x01234567); in xnn_qs8_gavgpool_minmax_ukernel_7p7x__wasmsimd_c8_acc2()
160 const v128_t vxi3x01234567 = wasm_i16x8_load_8x8(i3); in xnn_qs8_gavgpool_minmax_ukernel_7p7x__wasmsimd_c8_acc2() local
170 v128_t vacc1x01234567 = wasm_i16x8_add(vxi2x01234567, vxi3x01234567); in xnn_qs8_gavgpool_minmax_ukernel_7p7x__wasmsimd_c8_acc2()
230 const v128_t vxi3x01234567 = wasm_i16x8_load_8x8(i3); in xnn_qs8_gavgpool_minmax_ukernel_7p7x__wasmsimd_c8_acc2() local
240 v128_t vacc1x01234567 = wasm_i16x8_add(vxi2x01234567, vxi3x01234567); in xnn_qs8_gavgpool_minmax_ukernel_7p7x__wasmsimd_c8_acc2()
|
| D | 7p7x-minmax-sse41-c8-acc2.c | 50 const __m128i vxi3x01234567 = _mm_cvtepi8_epi16(_mm_loadl_epi64((const __m128i*) i3)); in xnn_qs8_gavgpool_minmax_ukernel_7p7x__sse41_c8_acc2() local
61 __m128i vacc1x01234567 = _mm_add_epi16(vxi2x01234567, vxi3x01234567); in xnn_qs8_gavgpool_minmax_ukernel_7p7x__sse41_c8_acc2()
96 const __m128i vxi3x01234567 = _mm_cvtepi8_epi16(_mm_loadl_epi64((const __m128i*) i3)); in xnn_qs8_gavgpool_minmax_ukernel_7p7x__sse41_c8_acc2() local
107 __m128i vacc1x01234567 = _mm_add_epi16(vxi2x01234567, vxi3x01234567); in xnn_qs8_gavgpool_minmax_ukernel_7p7x__sse41_c8_acc2()
161 const __m128i vxi3x01234567 = _mm_cvtepi8_epi16(_mm_loadl_epi64((const __m128i*) i3)); in xnn_qs8_gavgpool_minmax_ukernel_7p7x__sse41_c8_acc2() local
172 __m128i vacc1x01234567 = _mm_add_epi16(vxi2x01234567, vxi3x01234567); in xnn_qs8_gavgpool_minmax_ukernel_7p7x__sse41_c8_acc2()
229 const __m128i vxi3x01234567 = _mm_cvtepi8_epi16(_mm_loadl_epi64((const __m128i*) i3)); in xnn_qs8_gavgpool_minmax_ukernel_7p7x__sse41_c8_acc2() local
240 __m128i vacc1x01234567 = _mm_add_epi16(vxi2x01234567, vxi3x01234567); in xnn_qs8_gavgpool_minmax_ukernel_7p7x__sse41_c8_acc2()
|
| D | 7p7x-minmax-sse2-c8-acc2.c | 62 …const __m128i vxi3x01234567 = _mm_unpacklo_epi8(vi3x01234567, _mm_cmpgt_epi8(_mm_setzero_si128(), … in xnn_qs8_gavgpool_minmax_ukernel_7p7x__sse2_c8_acc2() local
68 __m128i vacc1x01234567 = _mm_add_epi16(vxi2x01234567, vxi3x01234567); in xnn_qs8_gavgpool_minmax_ukernel_7p7x__sse2_c8_acc2()
116 …const __m128i vxi3x01234567 = _mm_unpacklo_epi8(vi3x01234567, _mm_cmpgt_epi8(_mm_setzero_si128(), … in xnn_qs8_gavgpool_minmax_ukernel_7p7x__sse2_c8_acc2() local
122 __m128i vacc1x01234567 = _mm_add_epi16(vxi2x01234567, vxi3x01234567); in xnn_qs8_gavgpool_minmax_ukernel_7p7x__sse2_c8_acc2()
189 …const __m128i vxi3x01234567 = _mm_unpacklo_epi8(vi3x01234567, _mm_cmpgt_epi8(_mm_setzero_si128(), … in xnn_qs8_gavgpool_minmax_ukernel_7p7x__sse2_c8_acc2() local
195 __m128i vacc1x01234567 = _mm_add_epi16(vxi2x01234567, vxi3x01234567); in xnn_qs8_gavgpool_minmax_ukernel_7p7x__sse2_c8_acc2()
273 …const __m128i vxi3x01234567 = _mm_unpacklo_epi8(vi3x01234567, _mm_cmpgt_epi8(_mm_setzero_si128(), … in xnn_qs8_gavgpool_minmax_ukernel_7p7x__sse2_c8_acc2() local
279 __m128i vacc1x01234567 = _mm_add_epi16(vxi2x01234567, vxi3x01234567); in xnn_qs8_gavgpool_minmax_ukernel_7p7x__sse2_c8_acc2()
|
| D | 7p7x-minmax-ssse3-c8-acc2.c | 62 …const __m128i vxi3x01234567 = _mm_unpacklo_epi8(vi3x01234567, _mm_cmpgt_epi8(_mm_setzero_si128(), … in xnn_qs8_gavgpool_minmax_ukernel_7p7x__ssse3_c8_acc2() local
68 __m128i vacc1x01234567 = _mm_add_epi16(vxi2x01234567, vxi3x01234567); in xnn_qs8_gavgpool_minmax_ukernel_7p7x__ssse3_c8_acc2()
116 …const __m128i vxi3x01234567 = _mm_unpacklo_epi8(vi3x01234567, _mm_cmpgt_epi8(_mm_setzero_si128(), … in xnn_qs8_gavgpool_minmax_ukernel_7p7x__ssse3_c8_acc2() local
122 __m128i vacc1x01234567 = _mm_add_epi16(vxi2x01234567, vxi3x01234567); in xnn_qs8_gavgpool_minmax_ukernel_7p7x__ssse3_c8_acc2()
189 …const __m128i vxi3x01234567 = _mm_unpacklo_epi8(vi3x01234567, _mm_cmpgt_epi8(_mm_setzero_si128(), … in xnn_qs8_gavgpool_minmax_ukernel_7p7x__ssse3_c8_acc2() local
195 __m128i vacc1x01234567 = _mm_add_epi16(vxi2x01234567, vxi3x01234567); in xnn_qs8_gavgpool_minmax_ukernel_7p7x__ssse3_c8_acc2()
270 …const __m128i vxi3x01234567 = _mm_unpacklo_epi8(vi3x01234567, _mm_cmpgt_epi8(_mm_setzero_si128(), … in xnn_qs8_gavgpool_minmax_ukernel_7p7x__ssse3_c8_acc2() local
276 __m128i vacc1x01234567 = _mm_add_epi16(vxi2x01234567, vxi3x01234567); in xnn_qs8_gavgpool_minmax_ukernel_7p7x__ssse3_c8_acc2()
|
| D | 7p7x-minmax-wasmsimd-c24-acc2.c | 56 const v128_t vxi3x01234567 = wasm_i16x8_load_8x8(i3); in xnn_qs8_gavgpool_minmax_ukernel_7p7x__wasmsimd_c24_acc2() local
76 v128_t vacc1x01234567 = wasm_i16x8_add(vxi2x01234567, vxi3x01234567); in xnn_qs8_gavgpool_minmax_ukernel_7p7x__wasmsimd_c24_acc2()
118 const v128_t vxi3x01234567 = wasm_i16x8_load_8x8(i3); in xnn_qs8_gavgpool_minmax_ukernel_7p7x__wasmsimd_c24_acc2() local
128 v128_t vacc1x01234567 = wasm_i16x8_add(vxi2x01234567, vxi3x01234567); in xnn_qs8_gavgpool_minmax_ukernel_7p7x__wasmsimd_c24_acc2()
172 const v128_t vxi3x01234567 = wasm_i16x8_load_8x8(i3); in xnn_qs8_gavgpool_minmax_ukernel_7p7x__wasmsimd_c24_acc2() local
192 v128_t vacc1x01234567 = wasm_i16x8_add(vxi2x01234567, vxi3x01234567); in xnn_qs8_gavgpool_minmax_ukernel_7p7x__wasmsimd_c24_acc2()
234 const v128_t vxi3x01234567 = wasm_i16x8_load_8x8(i3); in xnn_qs8_gavgpool_minmax_ukernel_7p7x__wasmsimd_c24_acc2() local
244 v128_t vacc1x01234567 = wasm_i16x8_add(vxi2x01234567, vxi3x01234567); in xnn_qs8_gavgpool_minmax_ukernel_7p7x__wasmsimd_c24_acc2()
308 const v128_t vxi3x01234567 = wasm_i16x8_load_8x8(i3); in xnn_qs8_gavgpool_minmax_ukernel_7p7x__wasmsimd_c24_acc2() local
328 v128_t vacc1x01234567 = wasm_i16x8_add(vxi2x01234567, vxi3x01234567); in xnn_qs8_gavgpool_minmax_ukernel_7p7x__wasmsimd_c24_acc2()
[all …]
|
| D | 7p7x-minmax-sse41-c24-acc2.c | 56 const __m128i vxi3x01234567 = _mm_cvtepi8_epi16(_mm_loadl_epi64((const __m128i*) i3)); in xnn_qs8_gavgpool_minmax_ukernel_7p7x__sse41_c24_acc2() local
77 __m128i vacc1x01234567 = _mm_add_epi16(vxi2x01234567, vxi3x01234567); in xnn_qs8_gavgpool_minmax_ukernel_7p7x__sse41_c24_acc2()
119 const __m128i vxi3x01234567 = _mm_cvtepi8_epi16(_mm_loadl_epi64((const __m128i*) i3)); in xnn_qs8_gavgpool_minmax_ukernel_7p7x__sse41_c24_acc2() local
130 __m128i vacc1x01234567 = _mm_add_epi16(vxi2x01234567, vxi3x01234567); in xnn_qs8_gavgpool_minmax_ukernel_7p7x__sse41_c24_acc2()
174 const __m128i vxi3x01234567 = _mm_cvtepi8_epi16(_mm_loadl_epi64((const __m128i*) i3)); in xnn_qs8_gavgpool_minmax_ukernel_7p7x__sse41_c24_acc2() local
195 __m128i vacc1x01234567 = _mm_add_epi16(vxi2x01234567, vxi3x01234567); in xnn_qs8_gavgpool_minmax_ukernel_7p7x__sse41_c24_acc2()
237 const __m128i vxi3x01234567 = _mm_cvtepi8_epi16(_mm_loadl_epi64((const __m128i*) i3)); in xnn_qs8_gavgpool_minmax_ukernel_7p7x__sse41_c24_acc2() local
248 __m128i vacc1x01234567 = _mm_add_epi16(vxi2x01234567, vxi3x01234567); in xnn_qs8_gavgpool_minmax_ukernel_7p7x__sse41_c24_acc2()
311 const __m128i vxi3x01234567 = _mm_cvtepi8_epi16(_mm_loadl_epi64((const __m128i*) i3)); in xnn_qs8_gavgpool_minmax_ukernel_7p7x__sse41_c24_acc2() local
332 __m128i vacc1x01234567 = _mm_add_epi16(vxi2x01234567, vxi3x01234567); in xnn_qs8_gavgpool_minmax_ukernel_7p7x__sse41_c24_acc2()
[all …]
|
| D | 7p7x-minmax-wasmsimd-c16-acc2.c | 53 const v128_t vxi3x01234567 = wasm_i16x8_load_8x8(i3); in xnn_qs8_gavgpool_minmax_ukernel_7p7x__wasmsimd_c16_acc2() local
68 v128_t vacc1x01234567 = wasm_i16x8_add(vxi2x01234567, vxi3x01234567); in xnn_qs8_gavgpool_minmax_ukernel_7p7x__wasmsimd_c16_acc2()
115 const v128_t vxi3x01234567 = wasm_i16x8_load_8x8(i3); in xnn_qs8_gavgpool_minmax_ukernel_7p7x__wasmsimd_c16_acc2() local
130 v128_t vacc1x01234567 = wasm_i16x8_add(vxi2x01234567, vxi3x01234567); in xnn_qs8_gavgpool_minmax_ukernel_7p7x__wasmsimd_c16_acc2()
197 const v128_t vxi3x01234567 = wasm_i16x8_load_8x8(i3); in xnn_qs8_gavgpool_minmax_ukernel_7p7x__wasmsimd_c16_acc2() local
212 v128_t vacc1x01234567 = wasm_i16x8_add(vxi2x01234567, vxi3x01234567); in xnn_qs8_gavgpool_minmax_ukernel_7p7x__wasmsimd_c16_acc2()
300 const v128_t vxi3x01234567 = wasm_i16x8_load_8x8(i3); in xnn_qs8_gavgpool_minmax_ukernel_7p7x__wasmsimd_c16_acc2() local
310 v128_t vacc1x01234567 = wasm_i16x8_add(vxi2x01234567, vxi3x01234567); in xnn_qs8_gavgpool_minmax_ukernel_7p7x__wasmsimd_c16_acc2()
|
| D | 7p7x-minmax-sse41-c16-acc2.c | 53 const __m128i vxi3x01234567 = _mm_cvtepi8_epi16(_mm_loadl_epi64((const __m128i*) i3)); in xnn_qs8_gavgpool_minmax_ukernel_7p7x__sse41_c16_acc2() local
69 __m128i vacc1x01234567 = _mm_add_epi16(vxi2x01234567, vxi3x01234567); in xnn_qs8_gavgpool_minmax_ukernel_7p7x__sse41_c16_acc2()
116 const __m128i vxi3x01234567 = _mm_cvtepi8_epi16(_mm_loadl_epi64((const __m128i*) i3)); in xnn_qs8_gavgpool_minmax_ukernel_7p7x__sse41_c16_acc2() local
132 __m128i vacc1x01234567 = _mm_add_epi16(vxi2x01234567, vxi3x01234567); in xnn_qs8_gavgpool_minmax_ukernel_7p7x__sse41_c16_acc2()
198 const __m128i vxi3x01234567 = _mm_cvtepi8_epi16(_mm_loadl_epi64((const __m128i*) i3)); in xnn_qs8_gavgpool_minmax_ukernel_7p7x__sse41_c16_acc2() local
214 __m128i vacc1x01234567 = _mm_add_epi16(vxi2x01234567, vxi3x01234567); in xnn_qs8_gavgpool_minmax_ukernel_7p7x__sse41_c16_acc2()
296 const __m128i vxi3x01234567 = _mm_cvtepi8_epi16(_mm_loadl_epi64((const __m128i*) i3)); in xnn_qs8_gavgpool_minmax_ukernel_7p7x__sse41_c16_acc2() local
307 __m128i vacc1x01234567 = _mm_add_epi16(vxi2x01234567, vxi3x01234567); in xnn_qs8_gavgpool_minmax_ukernel_7p7x__sse41_c16_acc2()
|
| D | 7x-minmax-wasmsimd-c8-acc2.c | 68 const v128_t vxi3x01234567 = wasm_i16x8_load_8x8(i3); in xnn_qs8_gavgpool_minmax_ukernel_7x__wasmsimd_c8_acc2() local
78 v128_t vacc1x01234567 = wasm_i16x8_add(vxi2x01234567, vxi3x01234567); in xnn_qs8_gavgpool_minmax_ukernel_7x__wasmsimd_c8_acc2()
137 const v128_t vxi3x01234567 = wasm_i16x8_load_8x8(i3); in xnn_qs8_gavgpool_minmax_ukernel_7x__wasmsimd_c8_acc2() local
147 v128_t vacc1x01234567 = wasm_i16x8_add(vxi2x01234567, vxi3x01234567); in xnn_qs8_gavgpool_minmax_ukernel_7x__wasmsimd_c8_acc2()
|
| D | 7x-minmax-sse41-c8-acc2.c | 67 const __m128i vxi3x01234567 = _mm_cvtepi8_epi16(_mm_loadl_epi64((const __m128i*) i3)); in xnn_qs8_gavgpool_minmax_ukernel_7x__sse41_c8_acc2() local
78 __m128i vacc1x01234567 = _mm_add_epi16(vxi2x01234567, vxi3x01234567); in xnn_qs8_gavgpool_minmax_ukernel_7x__sse41_c8_acc2()
134 const __m128i vxi3x01234567 = _mm_cvtepi8_epi16(_mm_loadl_epi64((const __m128i*) i3)); in xnn_qs8_gavgpool_minmax_ukernel_7x__sse41_c8_acc2() local
145 __m128i vacc1x01234567 = _mm_add_epi16(vxi2x01234567, vxi3x01234567); in xnn_qs8_gavgpool_minmax_ukernel_7x__sse41_c8_acc2()
|
| D | 7p7x-minmax-sse2-c24-acc2.c | 82 …const __m128i vxi3x01234567 = _mm_unpacklo_epi8(vi3x01234567, _mm_cmpgt_epi8(_mm_setzero_si128(), … in xnn_qs8_gavgpool_minmax_ukernel_7p7x__sse2_c24_acc2() local
98 __m128i vacc1x01234567 = _mm_add_epi16(vxi2x01234567, vxi3x01234567); in xnn_qs8_gavgpool_minmax_ukernel_7p7x__sse2_c24_acc2()
155 …const __m128i vxi3x01234567 = _mm_unpacklo_epi8(vi3x01234567, _mm_cmpgt_epi8(_mm_setzero_si128(), … in xnn_qs8_gavgpool_minmax_ukernel_7p7x__sse2_c24_acc2() local
161 __m128i vacc1x01234567 = _mm_add_epi16(vxi2x01234567, vxi3x01234567); in xnn_qs8_gavgpool_minmax_ukernel_7p7x__sse2_c24_acc2()
232 …const __m128i vxi3x01234567 = _mm_unpacklo_epi8(vi3x01234567, _mm_cmpgt_epi8(_mm_setzero_si128(), … in xnn_qs8_gavgpool_minmax_ukernel_7p7x__sse2_c24_acc2() local
248 __m128i vacc1x01234567 = _mm_add_epi16(vxi2x01234567, vxi3x01234567); in xnn_qs8_gavgpool_minmax_ukernel_7p7x__sse2_c24_acc2()
305 …const __m128i vxi3x01234567 = _mm_unpacklo_epi8(vi3x01234567, _mm_cmpgt_epi8(_mm_setzero_si128(), … in xnn_qs8_gavgpool_minmax_ukernel_7p7x__sse2_c24_acc2() local
311 __m128i vacc1x01234567 = _mm_add_epi16(vxi2x01234567, vxi3x01234567); in xnn_qs8_gavgpool_minmax_ukernel_7p7x__sse2_c24_acc2()
401 …const __m128i vxi3x01234567 = _mm_unpacklo_epi8(vi3x01234567, _mm_cmpgt_epi8(_mm_setzero_si128(), … in xnn_qs8_gavgpool_minmax_ukernel_7p7x__sse2_c24_acc2() local
417 __m128i vacc1x01234567 = _mm_add_epi16(vxi2x01234567, vxi3x01234567); in xnn_qs8_gavgpool_minmax_ukernel_7p7x__sse2_c24_acc2()
[all …]
|
| D | 7p7x-minmax-ssse3-c24-acc2.c | 82 …const __m128i vxi3x01234567 = _mm_unpacklo_epi8(vi3x01234567, _mm_cmpgt_epi8(_mm_setzero_si128(), … in xnn_qs8_gavgpool_minmax_ukernel_7p7x__ssse3_c24_acc2() local
98 __m128i vacc1x01234567 = _mm_add_epi16(vxi2x01234567, vxi3x01234567); in xnn_qs8_gavgpool_minmax_ukernel_7p7x__ssse3_c24_acc2()
155 …const __m128i vxi3x01234567 = _mm_unpacklo_epi8(vi3x01234567, _mm_cmpgt_epi8(_mm_setzero_si128(), … in xnn_qs8_gavgpool_minmax_ukernel_7p7x__ssse3_c24_acc2() local
161 __m128i vacc1x01234567 = _mm_add_epi16(vxi2x01234567, vxi3x01234567); in xnn_qs8_gavgpool_minmax_ukernel_7p7x__ssse3_c24_acc2()
232 …const __m128i vxi3x01234567 = _mm_unpacklo_epi8(vi3x01234567, _mm_cmpgt_epi8(_mm_setzero_si128(), … in xnn_qs8_gavgpool_minmax_ukernel_7p7x__ssse3_c24_acc2() local
248 __m128i vacc1x01234567 = _mm_add_epi16(vxi2x01234567, vxi3x01234567); in xnn_qs8_gavgpool_minmax_ukernel_7p7x__ssse3_c24_acc2()
305 …const __m128i vxi3x01234567 = _mm_unpacklo_epi8(vi3x01234567, _mm_cmpgt_epi8(_mm_setzero_si128(), … in xnn_qs8_gavgpool_minmax_ukernel_7p7x__ssse3_c24_acc2() local
311 __m128i vacc1x01234567 = _mm_add_epi16(vxi2x01234567, vxi3x01234567); in xnn_qs8_gavgpool_minmax_ukernel_7p7x__ssse3_c24_acc2()
401 …const __m128i vxi3x01234567 = _mm_unpacklo_epi8(vi3x01234567, _mm_cmpgt_epi8(_mm_setzero_si128(), … in xnn_qs8_gavgpool_minmax_ukernel_7p7x__ssse3_c24_acc2() local
417 __m128i vacc1x01234567 = _mm_add_epi16(vxi2x01234567, vxi3x01234567); in xnn_qs8_gavgpool_minmax_ukernel_7p7x__ssse3_c24_acc2()
[all …]
|
| D | 7x-minmax-ssse3-c8-acc2.c | 79 …const __m128i vxi3x01234567 = _mm_unpacklo_epi8(vi3x01234567, _mm_cmpgt_epi8(_mm_setzero_si128(), … in xnn_qs8_gavgpool_minmax_ukernel_7x__ssse3_c8_acc2() local
85 __m128i vacc1x01234567 = _mm_add_epi16(vxi2x01234567, vxi3x01234567); in xnn_qs8_gavgpool_minmax_ukernel_7x__ssse3_c8_acc2()
159 …const __m128i vxi3x01234567 = _mm_unpacklo_epi8(vi3x01234567, _mm_cmpgt_epi8(_mm_setzero_si128(), … in xnn_qs8_gavgpool_minmax_ukernel_7x__ssse3_c8_acc2() local
165 __m128i vacc1x01234567 = _mm_add_epi16(vxi2x01234567, vxi3x01234567); in xnn_qs8_gavgpool_minmax_ukernel_7x__ssse3_c8_acc2()
|
| D | 7p7x-minmax-ssse3-c16-acc2.c | 72 …const __m128i vxi3x01234567 = _mm_unpacklo_epi8(vi3x01234567, _mm_cmpgt_epi8(_mm_setzero_si128(), … in xnn_qs8_gavgpool_minmax_ukernel_7p7x__ssse3_c16_acc2() local
83 __m128i vacc1x01234567 = _mm_add_epi16(vxi2x01234567, vxi3x01234567); in xnn_qs8_gavgpool_minmax_ukernel_7p7x__ssse3_c16_acc2()
151 …const __m128i vxi3x01234567 = _mm_unpacklo_epi8(vi3x01234567, _mm_cmpgt_epi8(_mm_setzero_si128(), … in xnn_qs8_gavgpool_minmax_ukernel_7p7x__ssse3_c16_acc2() local
162 __m128i vacc1x01234567 = _mm_add_epi16(vxi2x01234567, vxi3x01234567); in xnn_qs8_gavgpool_minmax_ukernel_7p7x__ssse3_c16_acc2()
249 …const __m128i vxi3x01234567 = _mm_unpacklo_epi8(vi3x01234567, _mm_cmpgt_epi8(_mm_setzero_si128(), … in xnn_qs8_gavgpool_minmax_ukernel_7p7x__ssse3_c16_acc2() local
260 __m128i vacc1x01234567 = _mm_add_epi16(vxi2x01234567, vxi3x01234567); in xnn_qs8_gavgpool_minmax_ukernel_7p7x__ssse3_c16_acc2()
365 …const __m128i vxi3x01234567 = _mm_unpacklo_epi8(vi3x01234567, _mm_cmpgt_epi8(_mm_setzero_si128(), … in xnn_qs8_gavgpool_minmax_ukernel_7p7x__ssse3_c16_acc2() local
371 __m128i vacc1x01234567 = _mm_add_epi16(vxi2x01234567, vxi3x01234567); in xnn_qs8_gavgpool_minmax_ukernel_7p7x__ssse3_c16_acc2()
|
| D | 7x-minmax-sse2-c8-acc2.c | 79 …const __m128i vxi3x01234567 = _mm_unpacklo_epi8(vi3x01234567, _mm_cmpgt_epi8(_mm_setzero_si128(), … in xnn_qs8_gavgpool_minmax_ukernel_7x__sse2_c8_acc2() local
85 __m128i vacc1x01234567 = _mm_add_epi16(vxi2x01234567, vxi3x01234567); in xnn_qs8_gavgpool_minmax_ukernel_7x__sse2_c8_acc2()
162 …const __m128i vxi3x01234567 = _mm_unpacklo_epi8(vi3x01234567, _mm_cmpgt_epi8(_mm_setzero_si128(), … in xnn_qs8_gavgpool_minmax_ukernel_7x__sse2_c8_acc2() local
168 __m128i vacc1x01234567 = _mm_add_epi16(vxi2x01234567, vxi3x01234567); in xnn_qs8_gavgpool_minmax_ukernel_7x__sse2_c8_acc2()
|
| D | 7x-minmax-sse41-c16-acc2.c | 70 const __m128i vxi3x01234567 = _mm_cvtepi8_epi16(_mm_loadl_epi64((const __m128i*) i3)); in xnn_qs8_gavgpool_minmax_ukernel_7x__sse41_c16_acc2() local
86 __m128i vacc1x01234567 = _mm_add_epi16(vxi2x01234567, vxi3x01234567); in xnn_qs8_gavgpool_minmax_ukernel_7x__sse41_c16_acc2()
167 const __m128i vxi3x01234567 = _mm_cvtepi8_epi16(_mm_loadl_epi64((const __m128i*) i3)); in xnn_qs8_gavgpool_minmax_ukernel_7x__sse41_c16_acc2() local
178 __m128i vacc1x01234567 = _mm_add_epi16(vxi2x01234567, vxi3x01234567); in xnn_qs8_gavgpool_minmax_ukernel_7x__sse41_c16_acc2()
|
| D | 7x-minmax-wasmsimd-c16-acc2.c | 71 const v128_t vxi3x01234567 = wasm_i16x8_load_8x8(i3); in xnn_qs8_gavgpool_minmax_ukernel_7x__wasmsimd_c16_acc2() local
86 v128_t vacc1x01234567 = wasm_i16x8_add(vxi2x01234567, vxi3x01234567); in xnn_qs8_gavgpool_minmax_ukernel_7x__wasmsimd_c16_acc2()
173 const v128_t vxi3x01234567 = wasm_i16x8_load_8x8(i3); in xnn_qs8_gavgpool_minmax_ukernel_7x__wasmsimd_c16_acc2() local
183 v128_t vacc1x01234567 = wasm_i16x8_add(vxi2x01234567, vxi3x01234567); in xnn_qs8_gavgpool_minmax_ukernel_7x__wasmsimd_c16_acc2()
|
| D | 7p7x-minmax-sse2-c16-acc2.c | 72 …const __m128i vxi3x01234567 = _mm_unpacklo_epi8(vi3x01234567, _mm_cmpgt_epi8(_mm_setzero_si128(), … in xnn_qs8_gavgpool_minmax_ukernel_7p7x__sse2_c16_acc2() local
83 __m128i vacc1x01234567 = _mm_add_epi16(vxi2x01234567, vxi3x01234567); in xnn_qs8_gavgpool_minmax_ukernel_7p7x__sse2_c16_acc2()
151 …const __m128i vxi3x01234567 = _mm_unpacklo_epi8(vi3x01234567, _mm_cmpgt_epi8(_mm_setzero_si128(), … in xnn_qs8_gavgpool_minmax_ukernel_7p7x__sse2_c16_acc2() local
162 __m128i vacc1x01234567 = _mm_add_epi16(vxi2x01234567, vxi3x01234567); in xnn_qs8_gavgpool_minmax_ukernel_7p7x__sse2_c16_acc2()
249 …const __m128i vxi3x01234567 = _mm_unpacklo_epi8(vi3x01234567, _mm_cmpgt_epi8(_mm_setzero_si128(), … in xnn_qs8_gavgpool_minmax_ukernel_7p7x__sse2_c16_acc2() local
260 __m128i vacc1x01234567 = _mm_add_epi16(vxi2x01234567, vxi3x01234567); in xnn_qs8_gavgpool_minmax_ukernel_7p7x__sse2_c16_acc2()
370 …const __m128i vxi3x01234567 = _mm_unpacklo_epi8(vi3x01234567, _mm_cmpgt_epi8(_mm_setzero_si128(), … in xnn_qs8_gavgpool_minmax_ukernel_7p7x__sse2_c16_acc2() local
376 __m128i vacc1x01234567 = _mm_add_epi16(vxi2x01234567, vxi3x01234567); in xnn_qs8_gavgpool_minmax_ukernel_7p7x__sse2_c16_acc2()
|
| D | 7x-minmax-sse41-c24-acc2.c | 73 const __m128i vxi3x01234567 = _mm_cvtepi8_epi16(_mm_loadl_epi64((const __m128i*) i3)); in xnn_qs8_gavgpool_minmax_ukernel_7x__sse41_c24_acc2() local
94 __m128i vacc1x01234567 = _mm_add_epi16(vxi2x01234567, vxi3x01234567); in xnn_qs8_gavgpool_minmax_ukernel_7x__sse41_c24_acc2()
202 const __m128i vxi3x01234567 = _mm_cvtepi8_epi16(_mm_loadl_epi64((const __m128i*) i3)); in xnn_qs8_gavgpool_minmax_ukernel_7x__sse41_c24_acc2() local
213 __m128i vacc1x01234567 = _mm_add_epi16(vxi2x01234567, vxi3x01234567); in xnn_qs8_gavgpool_minmax_ukernel_7x__sse41_c24_acc2()
|
| D | 7x-minmax-ssse3-c16-acc2.c | 89 …const __m128i vxi3x01234567 = _mm_unpacklo_epi8(vi3x01234567, _mm_cmpgt_epi8(_mm_setzero_si128(), … in xnn_qs8_gavgpool_minmax_ukernel_7x__ssse3_c16_acc2() local
100 __m128i vacc1x01234567 = _mm_add_epi16(vxi2x01234567, vxi3x01234567); in xnn_qs8_gavgpool_minmax_ukernel_7x__ssse3_c16_acc2()
204 …const __m128i vxi3x01234567 = _mm_unpacklo_epi8(vi3x01234567, _mm_cmpgt_epi8(_mm_setzero_si128(), … in xnn_qs8_gavgpool_minmax_ukernel_7x__ssse3_c16_acc2() local
210 __m128i vacc1x01234567 = _mm_add_epi16(vxi2x01234567, vxi3x01234567); in xnn_qs8_gavgpool_minmax_ukernel_7x__ssse3_c16_acc2()
|
| D | 7x-minmax-wasmsimd-c24-acc2.c | 74 const v128_t vxi3x01234567 = wasm_i16x8_load_8x8(i3); in xnn_qs8_gavgpool_minmax_ukernel_7x__wasmsimd_c24_acc2() local
94 v128_t vacc1x01234567 = wasm_i16x8_add(vxi2x01234567, vxi3x01234567); in xnn_qs8_gavgpool_minmax_ukernel_7x__wasmsimd_c24_acc2()
211 const v128_t vxi3x01234567 = wasm_i16x8_load_8x8(i3); in xnn_qs8_gavgpool_minmax_ukernel_7x__wasmsimd_c24_acc2() local
221 v128_t vacc1x01234567 = wasm_i16x8_add(vxi2x01234567, vxi3x01234567); in xnn_qs8_gavgpool_minmax_ukernel_7x__wasmsimd_c24_acc2()
|
| D | 7x-minmax-sse2-c16-acc2.c | 89 …const __m128i vxi3x01234567 = _mm_unpacklo_epi8(vi3x01234567, _mm_cmpgt_epi8(_mm_setzero_si128(), … in xnn_qs8_gavgpool_minmax_ukernel_7x__sse2_c16_acc2() local
100 __m128i vacc1x01234567 = _mm_add_epi16(vxi2x01234567, vxi3x01234567); in xnn_qs8_gavgpool_minmax_ukernel_7x__sse2_c16_acc2()
209 …const __m128i vxi3x01234567 = _mm_unpacklo_epi8(vi3x01234567, _mm_cmpgt_epi8(_mm_setzero_si128(), … in xnn_qs8_gavgpool_minmax_ukernel_7x__sse2_c16_acc2() local
215 __m128i vacc1x01234567 = _mm_add_epi16(vxi2x01234567, vxi3x01234567); in xnn_qs8_gavgpool_minmax_ukernel_7x__sse2_c16_acc2()
|
| /external/XNNPACK/src/qs8-dwconv/gen/ |
| D | up8x9-minmax-sse41-mul16.c | 127 const __m128i vxi3x01234567 = _mm_cvtepi8_epi16(vi3x01234567); in xnn_qs8_dwconv_minmax_ukernel_up8x9__sse41_mul16() local
133 const __m128i vp3x01234567lo = _mm_mullo_epi16(vxi3x01234567, vxk3x01234567); in xnn_qs8_dwconv_minmax_ukernel_up8x9__sse41_mul16()
134 const __m128i vp3x01234567hi = _mm_mulhi_epi16(vxi3x01234567, vxk3x01234567); in xnn_qs8_dwconv_minmax_ukernel_up8x9__sse41_mul16()
292 const __m128i vxi3x01234567 = _mm_cvtepi8_epi16(vi3x01234567); in xnn_qs8_dwconv_minmax_ukernel_up8x9__sse41_mul16() local
297 const __m128i vp3x01234567lo = _mm_mullo_epi16(vxi3x01234567, vxk3x01234567); in xnn_qs8_dwconv_minmax_ukernel_up8x9__sse41_mul16()
298 const __m128i vp3x01234567hi = _mm_mulhi_epi16(vxi3x01234567, vxk3x01234567); in xnn_qs8_dwconv_minmax_ukernel_up8x9__sse41_mul16()
|
| D | up8x9-minmax-sse2-mul16.c | 130 …const __m128i vxi3x01234567 = _mm_unpacklo_epi8(vi3x01234567, _mm_cmpgt_epi8(_mm_setzero_si128(), … in xnn_qs8_dwconv_minmax_ukernel_up8x9__sse2_mul16() local
133 const __m128i vp3x01234567lo = _mm_mullo_epi16(vxi3x01234567, vxk3x01234567); in xnn_qs8_dwconv_minmax_ukernel_up8x9__sse2_mul16()
134 const __m128i vp3x01234567hi = _mm_mulhi_epi16(vxi3x01234567, vxk3x01234567); in xnn_qs8_dwconv_minmax_ukernel_up8x9__sse2_mul16()
315 …const __m128i vxi3x01234567 = _mm_unpacklo_epi8(vi3x01234567, _mm_cmpgt_epi8(_mm_setzero_si128(), … in xnn_qs8_dwconv_minmax_ukernel_up8x9__sse2_mul16() local
318 const __m128i vp3x01234567lo = _mm_mullo_epi16(vxi3x01234567, vxk3x01234567); in xnn_qs8_dwconv_minmax_ukernel_up8x9__sse2_mul16()
319 const __m128i vp3x01234567hi = _mm_mulhi_epi16(vxi3x01234567, vxk3x01234567); in xnn_qs8_dwconv_minmax_ukernel_up8x9__sse2_mul16()
|
| D | up8x9-minmax-ssse3-mul16.c | 130 …const __m128i vxi3x01234567 = _mm_unpacklo_epi8(vi3x01234567, _mm_cmpgt_epi8(_mm_setzero_si128(), … in xnn_qs8_dwconv_minmax_ukernel_up8x9__ssse3_mul16() local
133 const __m128i vp3x01234567lo = _mm_mullo_epi16(vxi3x01234567, vxk3x01234567); in xnn_qs8_dwconv_minmax_ukernel_up8x9__ssse3_mul16()
134 const __m128i vp3x01234567hi = _mm_mulhi_epi16(vxi3x01234567, vxk3x01234567); in xnn_qs8_dwconv_minmax_ukernel_up8x9__ssse3_mul16()
315 …const __m128i vxi3x01234567 = _mm_unpacklo_epi8(vi3x01234567, _mm_cmpgt_epi8(_mm_setzero_si128(), … in xnn_qs8_dwconv_minmax_ukernel_up8x9__ssse3_mul16() local
318 const __m128i vp3x01234567lo = _mm_mullo_epi16(vxi3x01234567, vxk3x01234567); in xnn_qs8_dwconv_minmax_ukernel_up8x9__ssse3_mul16()
319 const __m128i vp3x01234567hi = _mm_mulhi_epi16(vxi3x01234567, vxk3x01234567); in xnn_qs8_dwconv_minmax_ukernel_up8x9__ssse3_mul16()
|