| /external/XNNPACK/src/qs8-gavgpool/gen/ |
| D | 7p7x-minmax-wasmsimd-c8-acc2.c | 52 const v128_t vxi4x01234567 = wasm_i16x8_load_8x8(i4); in xnn_qs8_gavgpool_minmax_ukernel_7p7x__wasmsimd_c8_acc2() local
62 vacc0x01234567 = wasm_i16x8_add(vacc0x01234567, vxi4x01234567); in xnn_qs8_gavgpool_minmax_ukernel_7p7x__wasmsimd_c8_acc2()
97 const v128_t vxi4x01234567 = wasm_i16x8_load_8x8(i4); in xnn_qs8_gavgpool_minmax_ukernel_7p7x__wasmsimd_c8_acc2() local
107 vacc0x01234567 = wasm_i16x8_add(vacc0x01234567, vxi4x01234567); in xnn_qs8_gavgpool_minmax_ukernel_7p7x__wasmsimd_c8_acc2()
162 const v128_t vxi4x01234567 = wasm_i16x8_load_8x8(i4); in xnn_qs8_gavgpool_minmax_ukernel_7p7x__wasmsimd_c8_acc2() local
172 vacc0x01234567 = wasm_i16x8_add(vacc0x01234567, vxi4x01234567); in xnn_qs8_gavgpool_minmax_ukernel_7p7x__wasmsimd_c8_acc2()
232 const v128_t vxi4x01234567 = wasm_i16x8_load_8x8(i4); in xnn_qs8_gavgpool_minmax_ukernel_7p7x__wasmsimd_c8_acc2() local
242 vacc0x01234567 = wasm_i16x8_add(vacc0x01234567, vxi4x01234567); in xnn_qs8_gavgpool_minmax_ukernel_7p7x__wasmsimd_c8_acc2()
|
| D | 7p7x-minmax-sse41-c8-acc2.c | 52 const __m128i vxi4x01234567 = _mm_cvtepi8_epi16(_mm_loadl_epi64((const __m128i*) i4)); in xnn_qs8_gavgpool_minmax_ukernel_7p7x__sse41_c8_acc2() local
63 vacc0x01234567 = _mm_add_epi16(vacc0x01234567, vxi4x01234567); in xnn_qs8_gavgpool_minmax_ukernel_7p7x__sse41_c8_acc2()
98 const __m128i vxi4x01234567 = _mm_cvtepi8_epi16(_mm_loadl_epi64((const __m128i*) i4)); in xnn_qs8_gavgpool_minmax_ukernel_7p7x__sse41_c8_acc2() local
109 vacc0x01234567 = _mm_add_epi16(vacc0x01234567, vxi4x01234567); in xnn_qs8_gavgpool_minmax_ukernel_7p7x__sse41_c8_acc2()
163 const __m128i vxi4x01234567 = _mm_cvtepi8_epi16(_mm_loadl_epi64((const __m128i*) i4)); in xnn_qs8_gavgpool_minmax_ukernel_7p7x__sse41_c8_acc2() local
174 vacc0x01234567 = _mm_add_epi16(vacc0x01234567, vxi4x01234567); in xnn_qs8_gavgpool_minmax_ukernel_7p7x__sse41_c8_acc2()
231 const __m128i vxi4x01234567 = _mm_cvtepi8_epi16(_mm_loadl_epi64((const __m128i*) i4)); in xnn_qs8_gavgpool_minmax_ukernel_7p7x__sse41_c8_acc2() local
242 vacc0x01234567 = _mm_add_epi16(vacc0x01234567, vxi4x01234567); in xnn_qs8_gavgpool_minmax_ukernel_7p7x__sse41_c8_acc2()
|
| D | 7p7x-minmax-ssse3-c8-acc2.c | 63 …const __m128i vxi4x01234567 = _mm_unpacklo_epi8(vi4x01234567, _mm_cmpgt_epi8(_mm_setzero_si128(), … in xnn_qs8_gavgpool_minmax_ukernel_7p7x__ssse3_c8_acc2() local
70 vacc0x01234567 = _mm_add_epi16(vacc0x01234567, vxi4x01234567); in xnn_qs8_gavgpool_minmax_ukernel_7p7x__ssse3_c8_acc2()
117 …const __m128i vxi4x01234567 = _mm_unpacklo_epi8(vi4x01234567, _mm_cmpgt_epi8(_mm_setzero_si128(), … in xnn_qs8_gavgpool_minmax_ukernel_7p7x__ssse3_c8_acc2() local
124 vacc0x01234567 = _mm_add_epi16(vacc0x01234567, vxi4x01234567); in xnn_qs8_gavgpool_minmax_ukernel_7p7x__ssse3_c8_acc2()
190 …const __m128i vxi4x01234567 = _mm_unpacklo_epi8(vi4x01234567, _mm_cmpgt_epi8(_mm_setzero_si128(), … in xnn_qs8_gavgpool_minmax_ukernel_7p7x__ssse3_c8_acc2() local
197 vacc0x01234567 = _mm_add_epi16(vacc0x01234567, vxi4x01234567); in xnn_qs8_gavgpool_minmax_ukernel_7p7x__ssse3_c8_acc2()
271 …const __m128i vxi4x01234567 = _mm_unpacklo_epi8(vi4x01234567, _mm_cmpgt_epi8(_mm_setzero_si128(), … in xnn_qs8_gavgpool_minmax_ukernel_7p7x__ssse3_c8_acc2() local
278 vacc0x01234567 = _mm_add_epi16(vacc0x01234567, vxi4x01234567); in xnn_qs8_gavgpool_minmax_ukernel_7p7x__ssse3_c8_acc2()
|
| D | 7p7x-minmax-sse2-c8-acc2.c | 63 …const __m128i vxi4x01234567 = _mm_unpacklo_epi8(vi4x01234567, _mm_cmpgt_epi8(_mm_setzero_si128(), … in xnn_qs8_gavgpool_minmax_ukernel_7p7x__sse2_c8_acc2() local
70 vacc0x01234567 = _mm_add_epi16(vacc0x01234567, vxi4x01234567); in xnn_qs8_gavgpool_minmax_ukernel_7p7x__sse2_c8_acc2()
117 …const __m128i vxi4x01234567 = _mm_unpacklo_epi8(vi4x01234567, _mm_cmpgt_epi8(_mm_setzero_si128(), … in xnn_qs8_gavgpool_minmax_ukernel_7p7x__sse2_c8_acc2() local
124 vacc0x01234567 = _mm_add_epi16(vacc0x01234567, vxi4x01234567); in xnn_qs8_gavgpool_minmax_ukernel_7p7x__sse2_c8_acc2()
190 …const __m128i vxi4x01234567 = _mm_unpacklo_epi8(vi4x01234567, _mm_cmpgt_epi8(_mm_setzero_si128(), … in xnn_qs8_gavgpool_minmax_ukernel_7p7x__sse2_c8_acc2() local
197 vacc0x01234567 = _mm_add_epi16(vacc0x01234567, vxi4x01234567); in xnn_qs8_gavgpool_minmax_ukernel_7p7x__sse2_c8_acc2()
274 …const __m128i vxi4x01234567 = _mm_unpacklo_epi8(vi4x01234567, _mm_cmpgt_epi8(_mm_setzero_si128(), … in xnn_qs8_gavgpool_minmax_ukernel_7p7x__sse2_c8_acc2() local
281 vacc0x01234567 = _mm_add_epi16(vacc0x01234567, vxi4x01234567); in xnn_qs8_gavgpool_minmax_ukernel_7p7x__sse2_c8_acc2()
|
| D | 7p7x-minmax-wasmsimd-c24-acc2.c | 60 const v128_t vxi4x01234567 = wasm_i16x8_load_8x8(i4); in xnn_qs8_gavgpool_minmax_ukernel_7p7x__wasmsimd_c24_acc2() local
80 vacc0x01234567 = wasm_i16x8_add(vacc0x01234567, vxi4x01234567); in xnn_qs8_gavgpool_minmax_ukernel_7p7x__wasmsimd_c24_acc2()
120 const v128_t vxi4x01234567 = wasm_i16x8_load_8x8(i4); in xnn_qs8_gavgpool_minmax_ukernel_7p7x__wasmsimd_c24_acc2() local
130 vacc0x01234567 = wasm_i16x8_add(vacc0x01234567, vxi4x01234567); in xnn_qs8_gavgpool_minmax_ukernel_7p7x__wasmsimd_c24_acc2()
176 const v128_t vxi4x01234567 = wasm_i16x8_load_8x8(i4); in xnn_qs8_gavgpool_minmax_ukernel_7p7x__wasmsimd_c24_acc2() local
196 vacc0x01234567 = wasm_i16x8_add(vacc0x01234567, vxi4x01234567); in xnn_qs8_gavgpool_minmax_ukernel_7p7x__wasmsimd_c24_acc2()
236 const v128_t vxi4x01234567 = wasm_i16x8_load_8x8(i4); in xnn_qs8_gavgpool_minmax_ukernel_7p7x__wasmsimd_c24_acc2() local
246 vacc0x01234567 = wasm_i16x8_add(vacc0x01234567, vxi4x01234567); in xnn_qs8_gavgpool_minmax_ukernel_7p7x__wasmsimd_c24_acc2()
312 const v128_t vxi4x01234567 = wasm_i16x8_load_8x8(i4); in xnn_qs8_gavgpool_minmax_ukernel_7p7x__wasmsimd_c24_acc2() local
332 vacc0x01234567 = wasm_i16x8_add(vacc0x01234567, vxi4x01234567); in xnn_qs8_gavgpool_minmax_ukernel_7p7x__wasmsimd_c24_acc2()
[all …]
|
| D | 7p7x-minmax-sse41-c24-acc2.c | 60 const __m128i vxi4x01234567 = _mm_cvtepi8_epi16(_mm_loadl_epi64((const __m128i*) i4)); in xnn_qs8_gavgpool_minmax_ukernel_7p7x__sse41_c24_acc2() local
81 vacc0x01234567 = _mm_add_epi16(vacc0x01234567, vxi4x01234567); in xnn_qs8_gavgpool_minmax_ukernel_7p7x__sse41_c24_acc2()
121 const __m128i vxi4x01234567 = _mm_cvtepi8_epi16(_mm_loadl_epi64((const __m128i*) i4)); in xnn_qs8_gavgpool_minmax_ukernel_7p7x__sse41_c24_acc2() local
132 vacc0x01234567 = _mm_add_epi16(vacc0x01234567, vxi4x01234567); in xnn_qs8_gavgpool_minmax_ukernel_7p7x__sse41_c24_acc2()
178 const __m128i vxi4x01234567 = _mm_cvtepi8_epi16(_mm_loadl_epi64((const __m128i*) i4)); in xnn_qs8_gavgpool_minmax_ukernel_7p7x__sse41_c24_acc2() local
199 vacc0x01234567 = _mm_add_epi16(vacc0x01234567, vxi4x01234567); in xnn_qs8_gavgpool_minmax_ukernel_7p7x__sse41_c24_acc2()
239 const __m128i vxi4x01234567 = _mm_cvtepi8_epi16(_mm_loadl_epi64((const __m128i*) i4)); in xnn_qs8_gavgpool_minmax_ukernel_7p7x__sse41_c24_acc2() local
250 vacc0x01234567 = _mm_add_epi16(vacc0x01234567, vxi4x01234567); in xnn_qs8_gavgpool_minmax_ukernel_7p7x__sse41_c24_acc2()
315 const __m128i vxi4x01234567 = _mm_cvtepi8_epi16(_mm_loadl_epi64((const __m128i*) i4)); in xnn_qs8_gavgpool_minmax_ukernel_7p7x__sse41_c24_acc2() local
336 vacc0x01234567 = _mm_add_epi16(vacc0x01234567, vxi4x01234567); in xnn_qs8_gavgpool_minmax_ukernel_7p7x__sse41_c24_acc2()
[all …]
|
| D | 7p7x-minmax-wasmsimd-c16-acc2.c | 56 const v128_t vxi4x01234567 = wasm_i16x8_load_8x8(i4); in xnn_qs8_gavgpool_minmax_ukernel_7p7x__wasmsimd_c16_acc2() local
71 vacc0x01234567 = wasm_i16x8_add(vacc0x01234567, vxi4x01234567); in xnn_qs8_gavgpool_minmax_ukernel_7p7x__wasmsimd_c16_acc2()
118 const v128_t vxi4x01234567 = wasm_i16x8_load_8x8(i4); in xnn_qs8_gavgpool_minmax_ukernel_7p7x__wasmsimd_c16_acc2() local
133 vacc0x01234567 = wasm_i16x8_add(vacc0x01234567, vxi4x01234567); in xnn_qs8_gavgpool_minmax_ukernel_7p7x__wasmsimd_c16_acc2()
200 const v128_t vxi4x01234567 = wasm_i16x8_load_8x8(i4); in xnn_qs8_gavgpool_minmax_ukernel_7p7x__wasmsimd_c16_acc2() local
215 vacc0x01234567 = wasm_i16x8_add(vacc0x01234567, vxi4x01234567); in xnn_qs8_gavgpool_minmax_ukernel_7p7x__wasmsimd_c16_acc2()
302 const v128_t vxi4x01234567 = wasm_i16x8_load_8x8(i4); in xnn_qs8_gavgpool_minmax_ukernel_7p7x__wasmsimd_c16_acc2() local
312 vacc0x01234567 = wasm_i16x8_add(vacc0x01234567, vxi4x01234567); in xnn_qs8_gavgpool_minmax_ukernel_7p7x__wasmsimd_c16_acc2()
|
| D | 7p7x-minmax-sse41-c16-acc2.c | 56 const __m128i vxi4x01234567 = _mm_cvtepi8_epi16(_mm_loadl_epi64((const __m128i*) i4)); in xnn_qs8_gavgpool_minmax_ukernel_7p7x__sse41_c16_acc2() local
72 vacc0x01234567 = _mm_add_epi16(vacc0x01234567, vxi4x01234567); in xnn_qs8_gavgpool_minmax_ukernel_7p7x__sse41_c16_acc2()
119 const __m128i vxi4x01234567 = _mm_cvtepi8_epi16(_mm_loadl_epi64((const __m128i*) i4)); in xnn_qs8_gavgpool_minmax_ukernel_7p7x__sse41_c16_acc2() local
135 vacc0x01234567 = _mm_add_epi16(vacc0x01234567, vxi4x01234567); in xnn_qs8_gavgpool_minmax_ukernel_7p7x__sse41_c16_acc2()
201 const __m128i vxi4x01234567 = _mm_cvtepi8_epi16(_mm_loadl_epi64((const __m128i*) i4)); in xnn_qs8_gavgpool_minmax_ukernel_7p7x__sse41_c16_acc2() local
217 vacc0x01234567 = _mm_add_epi16(vacc0x01234567, vxi4x01234567); in xnn_qs8_gavgpool_minmax_ukernel_7p7x__sse41_c16_acc2()
298 const __m128i vxi4x01234567 = _mm_cvtepi8_epi16(_mm_loadl_epi64((const __m128i*) i4)); in xnn_qs8_gavgpool_minmax_ukernel_7p7x__sse41_c16_acc2() local
309 vacc0x01234567 = _mm_add_epi16(vacc0x01234567, vxi4x01234567); in xnn_qs8_gavgpool_minmax_ukernel_7p7x__sse41_c16_acc2()
|
| D | 7x-minmax-wasmsimd-c8-acc2.c | 70 const v128_t vxi4x01234567 = wasm_i16x8_load_8x8(i4); in xnn_qs8_gavgpool_minmax_ukernel_7x__wasmsimd_c8_acc2() local
80 vacc0x01234567 = wasm_i16x8_add(vacc0x01234567, vxi4x01234567); in xnn_qs8_gavgpool_minmax_ukernel_7x__wasmsimd_c8_acc2()
139 const v128_t vxi4x01234567 = wasm_i16x8_load_8x8(i4); in xnn_qs8_gavgpool_minmax_ukernel_7x__wasmsimd_c8_acc2() local
149 vacc0x01234567 = wasm_i16x8_add(vacc0x01234567, vxi4x01234567); in xnn_qs8_gavgpool_minmax_ukernel_7x__wasmsimd_c8_acc2()
|
| D | 7x-minmax-sse41-c8-acc2.c | 69 const __m128i vxi4x01234567 = _mm_cvtepi8_epi16(_mm_loadl_epi64((const __m128i*) i4)); in xnn_qs8_gavgpool_minmax_ukernel_7x__sse41_c8_acc2() local
80 vacc0x01234567 = _mm_add_epi16(vacc0x01234567, vxi4x01234567); in xnn_qs8_gavgpool_minmax_ukernel_7x__sse41_c8_acc2()
136 const __m128i vxi4x01234567 = _mm_cvtepi8_epi16(_mm_loadl_epi64((const __m128i*) i4)); in xnn_qs8_gavgpool_minmax_ukernel_7x__sse41_c8_acc2() local
147 vacc0x01234567 = _mm_add_epi16(vacc0x01234567, vxi4x01234567); in xnn_qs8_gavgpool_minmax_ukernel_7x__sse41_c8_acc2()
|
| D | 7p7x-minmax-sse2-c24-acc2.c | 85 …const __m128i vxi4x01234567 = _mm_unpacklo_epi8(vi4x01234567, _mm_cmpgt_epi8(_mm_setzero_si128(), … in xnn_qs8_gavgpool_minmax_ukernel_7p7x__sse2_c24_acc2() local
102 vacc0x01234567 = _mm_add_epi16(vacc0x01234567, vxi4x01234567); in xnn_qs8_gavgpool_minmax_ukernel_7p7x__sse2_c24_acc2()
156 …const __m128i vxi4x01234567 = _mm_unpacklo_epi8(vi4x01234567, _mm_cmpgt_epi8(_mm_setzero_si128(), … in xnn_qs8_gavgpool_minmax_ukernel_7p7x__sse2_c24_acc2() local
163 vacc0x01234567 = _mm_add_epi16(vacc0x01234567, vxi4x01234567); in xnn_qs8_gavgpool_minmax_ukernel_7p7x__sse2_c24_acc2()
235 …const __m128i vxi4x01234567 = _mm_unpacklo_epi8(vi4x01234567, _mm_cmpgt_epi8(_mm_setzero_si128(), … in xnn_qs8_gavgpool_minmax_ukernel_7p7x__sse2_c24_acc2() local
252 vacc0x01234567 = _mm_add_epi16(vacc0x01234567, vxi4x01234567); in xnn_qs8_gavgpool_minmax_ukernel_7p7x__sse2_c24_acc2()
306 …const __m128i vxi4x01234567 = _mm_unpacklo_epi8(vi4x01234567, _mm_cmpgt_epi8(_mm_setzero_si128(), … in xnn_qs8_gavgpool_minmax_ukernel_7p7x__sse2_c24_acc2() local
313 vacc0x01234567 = _mm_add_epi16(vacc0x01234567, vxi4x01234567); in xnn_qs8_gavgpool_minmax_ukernel_7p7x__sse2_c24_acc2()
404 …const __m128i vxi4x01234567 = _mm_unpacklo_epi8(vi4x01234567, _mm_cmpgt_epi8(_mm_setzero_si128(), … in xnn_qs8_gavgpool_minmax_ukernel_7p7x__sse2_c24_acc2() local
421 vacc0x01234567 = _mm_add_epi16(vacc0x01234567, vxi4x01234567); in xnn_qs8_gavgpool_minmax_ukernel_7p7x__sse2_c24_acc2()
[all …]
|
| D | 7p7x-minmax-ssse3-c24-acc2.c | 85 …const __m128i vxi4x01234567 = _mm_unpacklo_epi8(vi4x01234567, _mm_cmpgt_epi8(_mm_setzero_si128(), … in xnn_qs8_gavgpool_minmax_ukernel_7p7x__ssse3_c24_acc2() local
102 vacc0x01234567 = _mm_add_epi16(vacc0x01234567, vxi4x01234567); in xnn_qs8_gavgpool_minmax_ukernel_7p7x__ssse3_c24_acc2()
156 …const __m128i vxi4x01234567 = _mm_unpacklo_epi8(vi4x01234567, _mm_cmpgt_epi8(_mm_setzero_si128(), … in xnn_qs8_gavgpool_minmax_ukernel_7p7x__ssse3_c24_acc2() local
163 vacc0x01234567 = _mm_add_epi16(vacc0x01234567, vxi4x01234567); in xnn_qs8_gavgpool_minmax_ukernel_7p7x__ssse3_c24_acc2()
235 …const __m128i vxi4x01234567 = _mm_unpacklo_epi8(vi4x01234567, _mm_cmpgt_epi8(_mm_setzero_si128(), … in xnn_qs8_gavgpool_minmax_ukernel_7p7x__ssse3_c24_acc2() local
252 vacc0x01234567 = _mm_add_epi16(vacc0x01234567, vxi4x01234567); in xnn_qs8_gavgpool_minmax_ukernel_7p7x__ssse3_c24_acc2()
306 …const __m128i vxi4x01234567 = _mm_unpacklo_epi8(vi4x01234567, _mm_cmpgt_epi8(_mm_setzero_si128(), … in xnn_qs8_gavgpool_minmax_ukernel_7p7x__ssse3_c24_acc2() local
313 vacc0x01234567 = _mm_add_epi16(vacc0x01234567, vxi4x01234567); in xnn_qs8_gavgpool_minmax_ukernel_7p7x__ssse3_c24_acc2()
404 …const __m128i vxi4x01234567 = _mm_unpacklo_epi8(vi4x01234567, _mm_cmpgt_epi8(_mm_setzero_si128(), … in xnn_qs8_gavgpool_minmax_ukernel_7p7x__ssse3_c24_acc2() local
421 vacc0x01234567 = _mm_add_epi16(vacc0x01234567, vxi4x01234567); in xnn_qs8_gavgpool_minmax_ukernel_7p7x__ssse3_c24_acc2()
[all …]
|
| D | 7x-minmax-ssse3-c8-acc2.c | 80 …const __m128i vxi4x01234567 = _mm_unpacklo_epi8(vi4x01234567, _mm_cmpgt_epi8(_mm_setzero_si128(), … in xnn_qs8_gavgpool_minmax_ukernel_7x__ssse3_c8_acc2() local
87 vacc0x01234567 = _mm_add_epi16(vacc0x01234567, vxi4x01234567); in xnn_qs8_gavgpool_minmax_ukernel_7x__ssse3_c8_acc2()
160 …const __m128i vxi4x01234567 = _mm_unpacklo_epi8(vi4x01234567, _mm_cmpgt_epi8(_mm_setzero_si128(), … in xnn_qs8_gavgpool_minmax_ukernel_7x__ssse3_c8_acc2() local
167 vacc0x01234567 = _mm_add_epi16(vacc0x01234567, vxi4x01234567); in xnn_qs8_gavgpool_minmax_ukernel_7x__ssse3_c8_acc2()
|
| D | 7p7x-minmax-ssse3-c16-acc2.c | 74 …const __m128i vxi4x01234567 = _mm_unpacklo_epi8(vi4x01234567, _mm_cmpgt_epi8(_mm_setzero_si128(), … in xnn_qs8_gavgpool_minmax_ukernel_7p7x__ssse3_c16_acc2() local
86 vacc0x01234567 = _mm_add_epi16(vacc0x01234567, vxi4x01234567); in xnn_qs8_gavgpool_minmax_ukernel_7p7x__ssse3_c16_acc2()
153 …const __m128i vxi4x01234567 = _mm_unpacklo_epi8(vi4x01234567, _mm_cmpgt_epi8(_mm_setzero_si128(), … in xnn_qs8_gavgpool_minmax_ukernel_7p7x__ssse3_c16_acc2() local
165 vacc0x01234567 = _mm_add_epi16(vacc0x01234567, vxi4x01234567); in xnn_qs8_gavgpool_minmax_ukernel_7p7x__ssse3_c16_acc2()
251 …const __m128i vxi4x01234567 = _mm_unpacklo_epi8(vi4x01234567, _mm_cmpgt_epi8(_mm_setzero_si128(), … in xnn_qs8_gavgpool_minmax_ukernel_7p7x__ssse3_c16_acc2() local
263 vacc0x01234567 = _mm_add_epi16(vacc0x01234567, vxi4x01234567); in xnn_qs8_gavgpool_minmax_ukernel_7p7x__ssse3_c16_acc2()
366 …const __m128i vxi4x01234567 = _mm_unpacklo_epi8(vi4x01234567, _mm_cmpgt_epi8(_mm_setzero_si128(), … in xnn_qs8_gavgpool_minmax_ukernel_7p7x__ssse3_c16_acc2() local
373 vacc0x01234567 = _mm_add_epi16(vacc0x01234567, vxi4x01234567); in xnn_qs8_gavgpool_minmax_ukernel_7p7x__ssse3_c16_acc2()
|
| D | 7x-minmax-sse2-c8-acc2.c | 80 …const __m128i vxi4x01234567 = _mm_unpacklo_epi8(vi4x01234567, _mm_cmpgt_epi8(_mm_setzero_si128(), … in xnn_qs8_gavgpool_minmax_ukernel_7x__sse2_c8_acc2() local
87 vacc0x01234567 = _mm_add_epi16(vacc0x01234567, vxi4x01234567); in xnn_qs8_gavgpool_minmax_ukernel_7x__sse2_c8_acc2()
163 …const __m128i vxi4x01234567 = _mm_unpacklo_epi8(vi4x01234567, _mm_cmpgt_epi8(_mm_setzero_si128(), … in xnn_qs8_gavgpool_minmax_ukernel_7x__sse2_c8_acc2() local
170 vacc0x01234567 = _mm_add_epi16(vacc0x01234567, vxi4x01234567); in xnn_qs8_gavgpool_minmax_ukernel_7x__sse2_c8_acc2()
|
| D | 7x-minmax-sse41-c16-acc2.c | 73 const __m128i vxi4x01234567 = _mm_cvtepi8_epi16(_mm_loadl_epi64((const __m128i*) i4)); in xnn_qs8_gavgpool_minmax_ukernel_7x__sse41_c16_acc2() local
89 vacc0x01234567 = _mm_add_epi16(vacc0x01234567, vxi4x01234567); in xnn_qs8_gavgpool_minmax_ukernel_7x__sse41_c16_acc2()
169 const __m128i vxi4x01234567 = _mm_cvtepi8_epi16(_mm_loadl_epi64((const __m128i*) i4)); in xnn_qs8_gavgpool_minmax_ukernel_7x__sse41_c16_acc2() local
180 vacc0x01234567 = _mm_add_epi16(vacc0x01234567, vxi4x01234567); in xnn_qs8_gavgpool_minmax_ukernel_7x__sse41_c16_acc2()
|
| D | 7x-minmax-wasmsimd-c16-acc2.c | 74 const v128_t vxi4x01234567 = wasm_i16x8_load_8x8(i4); in xnn_qs8_gavgpool_minmax_ukernel_7x__wasmsimd_c16_acc2() local
89 vacc0x01234567 = wasm_i16x8_add(vacc0x01234567, vxi4x01234567); in xnn_qs8_gavgpool_minmax_ukernel_7x__wasmsimd_c16_acc2()
175 const v128_t vxi4x01234567 = wasm_i16x8_load_8x8(i4); in xnn_qs8_gavgpool_minmax_ukernel_7x__wasmsimd_c16_acc2() local
185 vacc0x01234567 = wasm_i16x8_add(vacc0x01234567, vxi4x01234567); in xnn_qs8_gavgpool_minmax_ukernel_7x__wasmsimd_c16_acc2()
|
| D | 7p7x-minmax-sse2-c16-acc2.c | 74 …const __m128i vxi4x01234567 = _mm_unpacklo_epi8(vi4x01234567, _mm_cmpgt_epi8(_mm_setzero_si128(), … in xnn_qs8_gavgpool_minmax_ukernel_7p7x__sse2_c16_acc2() local
86 vacc0x01234567 = _mm_add_epi16(vacc0x01234567, vxi4x01234567); in xnn_qs8_gavgpool_minmax_ukernel_7p7x__sse2_c16_acc2()
153 …const __m128i vxi4x01234567 = _mm_unpacklo_epi8(vi4x01234567, _mm_cmpgt_epi8(_mm_setzero_si128(), … in xnn_qs8_gavgpool_minmax_ukernel_7p7x__sse2_c16_acc2() local
165 vacc0x01234567 = _mm_add_epi16(vacc0x01234567, vxi4x01234567); in xnn_qs8_gavgpool_minmax_ukernel_7p7x__sse2_c16_acc2()
251 …const __m128i vxi4x01234567 = _mm_unpacklo_epi8(vi4x01234567, _mm_cmpgt_epi8(_mm_setzero_si128(), … in xnn_qs8_gavgpool_minmax_ukernel_7p7x__sse2_c16_acc2() local
263 vacc0x01234567 = _mm_add_epi16(vacc0x01234567, vxi4x01234567); in xnn_qs8_gavgpool_minmax_ukernel_7p7x__sse2_c16_acc2()
371 …const __m128i vxi4x01234567 = _mm_unpacklo_epi8(vi4x01234567, _mm_cmpgt_epi8(_mm_setzero_si128(), … in xnn_qs8_gavgpool_minmax_ukernel_7p7x__sse2_c16_acc2() local
378 vacc0x01234567 = _mm_add_epi16(vacc0x01234567, vxi4x01234567); in xnn_qs8_gavgpool_minmax_ukernel_7p7x__sse2_c16_acc2()
|
| D | 7x-minmax-ssse3-c16-acc2.c | 91 …const __m128i vxi4x01234567 = _mm_unpacklo_epi8(vi4x01234567, _mm_cmpgt_epi8(_mm_setzero_si128(), … in xnn_qs8_gavgpool_minmax_ukernel_7x__ssse3_c16_acc2() local
103 vacc0x01234567 = _mm_add_epi16(vacc0x01234567, vxi4x01234567); in xnn_qs8_gavgpool_minmax_ukernel_7x__ssse3_c16_acc2()
205 …const __m128i vxi4x01234567 = _mm_unpacklo_epi8(vi4x01234567, _mm_cmpgt_epi8(_mm_setzero_si128(), … in xnn_qs8_gavgpool_minmax_ukernel_7x__ssse3_c16_acc2() local
212 vacc0x01234567 = _mm_add_epi16(vacc0x01234567, vxi4x01234567); in xnn_qs8_gavgpool_minmax_ukernel_7x__ssse3_c16_acc2()
|
| D | 7x-minmax-sse41-c24-acc2.c | 77 const __m128i vxi4x01234567 = _mm_cvtepi8_epi16(_mm_loadl_epi64((const __m128i*) i4)); in xnn_qs8_gavgpool_minmax_ukernel_7x__sse41_c24_acc2() local
98 vacc0x01234567 = _mm_add_epi16(vacc0x01234567, vxi4x01234567); in xnn_qs8_gavgpool_minmax_ukernel_7x__sse41_c24_acc2()
204 const __m128i vxi4x01234567 = _mm_cvtepi8_epi16(_mm_loadl_epi64((const __m128i*) i4)); in xnn_qs8_gavgpool_minmax_ukernel_7x__sse41_c24_acc2() local
215 vacc0x01234567 = _mm_add_epi16(vacc0x01234567, vxi4x01234567); in xnn_qs8_gavgpool_minmax_ukernel_7x__sse41_c24_acc2()
|
| D | 7x-minmax-wasmsimd-c24-acc2.c | 78 const v128_t vxi4x01234567 = wasm_i16x8_load_8x8(i4); in xnn_qs8_gavgpool_minmax_ukernel_7x__wasmsimd_c24_acc2() local
98 vacc0x01234567 = wasm_i16x8_add(vacc0x01234567, vxi4x01234567); in xnn_qs8_gavgpool_minmax_ukernel_7x__wasmsimd_c24_acc2()
213 const v128_t vxi4x01234567 = wasm_i16x8_load_8x8(i4); in xnn_qs8_gavgpool_minmax_ukernel_7x__wasmsimd_c24_acc2() local
223 vacc0x01234567 = wasm_i16x8_add(vacc0x01234567, vxi4x01234567); in xnn_qs8_gavgpool_minmax_ukernel_7x__wasmsimd_c24_acc2()
|
| D | 7x-minmax-sse2-c16-acc2.c | 91 …const __m128i vxi4x01234567 = _mm_unpacklo_epi8(vi4x01234567, _mm_cmpgt_epi8(_mm_setzero_si128(), … in xnn_qs8_gavgpool_minmax_ukernel_7x__sse2_c16_acc2() local
103 vacc0x01234567 = _mm_add_epi16(vacc0x01234567, vxi4x01234567); in xnn_qs8_gavgpool_minmax_ukernel_7x__sse2_c16_acc2()
210 …const __m128i vxi4x01234567 = _mm_unpacklo_epi8(vi4x01234567, _mm_cmpgt_epi8(_mm_setzero_si128(), … in xnn_qs8_gavgpool_minmax_ukernel_7x__sse2_c16_acc2() local
217 vacc0x01234567 = _mm_add_epi16(vacc0x01234567, vxi4x01234567); in xnn_qs8_gavgpool_minmax_ukernel_7x__sse2_c16_acc2()
|
| /external/XNNPACK/src/qs8-dwconv/gen/ |
| D | up8x9-minmax-sse41-mul16.c | 140 const __m128i vxi4x01234567 = _mm_cvtepi8_epi16(vi4x01234567); in xnn_qs8_dwconv_minmax_ukernel_up8x9__sse41_mul16() local
146 const __m128i vp4x01234567lo = _mm_mullo_epi16(vxi4x01234567, vxk4x01234567); in xnn_qs8_dwconv_minmax_ukernel_up8x9__sse41_mul16()
147 const __m128i vp4x01234567hi = _mm_mulhi_epi16(vxi4x01234567, vxk4x01234567); in xnn_qs8_dwconv_minmax_ukernel_up8x9__sse41_mul16()
304 const __m128i vxi4x01234567 = _mm_cvtepi8_epi16(vi4x01234567); in xnn_qs8_dwconv_minmax_ukernel_up8x9__sse41_mul16() local
309 const __m128i vp4x01234567lo = _mm_mullo_epi16(vxi4x01234567, vxk4x01234567); in xnn_qs8_dwconv_minmax_ukernel_up8x9__sse41_mul16()
310 const __m128i vp4x01234567hi = _mm_mulhi_epi16(vxi4x01234567, vxk4x01234567); in xnn_qs8_dwconv_minmax_ukernel_up8x9__sse41_mul16()
|
| D | up8x9-minmax-sse2-mul16.c | 143 …const __m128i vxi4x01234567 = _mm_unpacklo_epi8(vi4x01234567, _mm_cmpgt_epi8(_mm_setzero_si128(), … in xnn_qs8_dwconv_minmax_ukernel_up8x9__sse2_mul16() local
146 const __m128i vp4x01234567lo = _mm_mullo_epi16(vxi4x01234567, vxk4x01234567); in xnn_qs8_dwconv_minmax_ukernel_up8x9__sse2_mul16()
147 const __m128i vp4x01234567hi = _mm_mulhi_epi16(vxi4x01234567, vxk4x01234567); in xnn_qs8_dwconv_minmax_ukernel_up8x9__sse2_mul16()
327 …const __m128i vxi4x01234567 = _mm_unpacklo_epi8(vi4x01234567, _mm_cmpgt_epi8(_mm_setzero_si128(), … in xnn_qs8_dwconv_minmax_ukernel_up8x9__sse2_mul16() local
330 const __m128i vp4x01234567lo = _mm_mullo_epi16(vxi4x01234567, vxk4x01234567); in xnn_qs8_dwconv_minmax_ukernel_up8x9__sse2_mul16()
331 const __m128i vp4x01234567hi = _mm_mulhi_epi16(vxi4x01234567, vxk4x01234567); in xnn_qs8_dwconv_minmax_ukernel_up8x9__sse2_mul16()
|
| D | up8x9-minmax-ssse3-mul16.c | 143 …const __m128i vxi4x01234567 = _mm_unpacklo_epi8(vi4x01234567, _mm_cmpgt_epi8(_mm_setzero_si128(), … in xnn_qs8_dwconv_minmax_ukernel_up8x9__ssse3_mul16() local
146 const __m128i vp4x01234567lo = _mm_mullo_epi16(vxi4x01234567, vxk4x01234567); in xnn_qs8_dwconv_minmax_ukernel_up8x9__ssse3_mul16()
147 const __m128i vp4x01234567hi = _mm_mulhi_epi16(vxi4x01234567, vxk4x01234567); in xnn_qs8_dwconv_minmax_ukernel_up8x9__ssse3_mul16()
327 …const __m128i vxi4x01234567 = _mm_unpacklo_epi8(vi4x01234567, _mm_cmpgt_epi8(_mm_setzero_si128(), … in xnn_qs8_dwconv_minmax_ukernel_up8x9__ssse3_mul16() local
330 const __m128i vp4x01234567lo = _mm_mullo_epi16(vxi4x01234567, vxk4x01234567); in xnn_qs8_dwconv_minmax_ukernel_up8x9__ssse3_mul16()
331 const __m128i vp4x01234567hi = _mm_mulhi_epi16(vxi4x01234567, vxk4x01234567); in xnn_qs8_dwconv_minmax_ukernel_up8x9__ssse3_mul16()
|