Home
last modified time | relevance | path

Searched refs:vabsprod02 (Results 1 – 24 of 24) sorted by relevance

/external/XNNPACK/src/qs8-gavgpool/gen/
D7x-minmax-sse41-c8-acc2.c96 const __m128i vabsprod02 = _mm_mul_epu32(vabsacc0123, vmultiplier); in xnn_qs8_gavgpool_minmax_ukernel_7x__sse41_c8_acc2() local
101 const __m128i vabsout02 = _mm_srl_epi64(_mm_add_epi64(vabsprod02, vrounding), vshift); in xnn_qs8_gavgpool_minmax_ukernel_7x__sse41_c8_acc2()
163 const __m128i vabsprod02 = _mm_mul_epu32(vabsacc0123, vmultiplier); in xnn_qs8_gavgpool_minmax_ukernel_7x__sse41_c8_acc2() local
168 const __m128i vabsout02 = _mm_srl_epi64(_mm_add_epi64(vabsprod02, vrounding), vshift); in xnn_qs8_gavgpool_minmax_ukernel_7x__sse41_c8_acc2()
D7x-minmax-ssse3-c8-acc2.c104 const __m128i vabsprod02 = _mm_mul_epu32(vabsacc0123, vmultiplier); in xnn_qs8_gavgpool_minmax_ukernel_7x__ssse3_c8_acc2() local
109 const __m128i vabsout02 = _mm_srl_epi64(_mm_add_epi64(vabsprod02, vrounding), vshift); in xnn_qs8_gavgpool_minmax_ukernel_7x__ssse3_c8_acc2()
184 const __m128i vabsprod02 = _mm_mul_epu32(vabsacc0123, vmultiplier); in xnn_qs8_gavgpool_minmax_ukernel_7x__ssse3_c8_acc2() local
189 const __m128i vabsout02 = _mm_srl_epi64(_mm_add_epi64(vabsprod02, vrounding), vshift); in xnn_qs8_gavgpool_minmax_ukernel_7x__ssse3_c8_acc2()
D7x-minmax-sse2-c8-acc2.c107 const __m128i vabsprod02 = _mm_mul_epu32(vabsacc0123, vmultiplier); in xnn_qs8_gavgpool_minmax_ukernel_7x__sse2_c8_acc2() local
112 const __m128i vabsout02 = _mm_srl_epi64(_mm_add_epi64(vabsprod02, vrounding), vshift); in xnn_qs8_gavgpool_minmax_ukernel_7x__sse2_c8_acc2()
190 const __m128i vabsprod02 = _mm_mul_epu32(vabsacc0123, vmultiplier); in xnn_qs8_gavgpool_minmax_ukernel_7x__sse2_c8_acc2() local
195 const __m128i vabsout02 = _mm_srl_epi64(_mm_add_epi64(vabsprod02, vrounding), vshift); in xnn_qs8_gavgpool_minmax_ukernel_7x__sse2_c8_acc2()
D7x-minmax-sse41-c16-acc2.c115 const __m128i vabsprod02 = _mm_mul_epu32(vabsacc0123, vmultiplier); in xnn_qs8_gavgpool_minmax_ukernel_7x__sse41_c16_acc2() local
124 const __m128i vabsout02 = _mm_srl_epi64(_mm_add_epi64(vabsprod02, vrounding), vshift); in xnn_qs8_gavgpool_minmax_ukernel_7x__sse41_c16_acc2()
196 const __m128i vabsprod02 = _mm_mul_epu32(vabsacc0123, vmultiplier); in xnn_qs8_gavgpool_minmax_ukernel_7x__sse41_c16_acc2() local
201 const __m128i vabsout02 = _mm_srl_epi64(_mm_add_epi64(vabsprod02, vrounding), vshift); in xnn_qs8_gavgpool_minmax_ukernel_7x__sse41_c16_acc2()
D7p7x-minmax-sse41-c8-acc2.c191 const __m128i vabsprod02 = _mm_mul_epu32(vabsacc0123, vmultiplier); in xnn_qs8_gavgpool_minmax_ukernel_7p7x__sse41_c8_acc2() local
196 const __m128i vabsout02 = _mm_srl_epi64(_mm_add_epi64(vabsprod02, vrounding), vshift); in xnn_qs8_gavgpool_minmax_ukernel_7p7x__sse41_c8_acc2()
259 const __m128i vabsprod02 = _mm_mul_epu32(vabsacc0123, vmultiplier); in xnn_qs8_gavgpool_minmax_ukernel_7p7x__sse41_c8_acc2() local
264 const __m128i vabsout02 = _mm_srl_epi64(_mm_add_epi64(vabsprod02, vrounding), vshift); in xnn_qs8_gavgpool_minmax_ukernel_7p7x__sse41_c8_acc2()
D7x-minmax-sse41-c24-acc2.c134 const __m128i vabsprod02 = _mm_mul_epu32(vabsacc0123, vmultiplier); in xnn_qs8_gavgpool_minmax_ukernel_7x__sse41_c24_acc2() local
147 const __m128i vabsout02 = _mm_srl_epi64(_mm_add_epi64(vabsprod02, vrounding), vshift); in xnn_qs8_gavgpool_minmax_ukernel_7x__sse41_c24_acc2()
231 const __m128i vabsprod02 = _mm_mul_epu32(vabsacc0123, vmultiplier); in xnn_qs8_gavgpool_minmax_ukernel_7x__sse41_c24_acc2() local
236 const __m128i vabsout02 = _mm_srl_epi64(_mm_add_epi64(vabsprod02, vrounding), vshift); in xnn_qs8_gavgpool_minmax_ukernel_7x__sse41_c24_acc2()
D7x-minmax-ssse3-c16-acc2.c131 const __m128i vabsprod02 = _mm_mul_epu32(vabsacc0123, vmultiplier); in xnn_qs8_gavgpool_minmax_ukernel_7x__ssse3_c16_acc2() local
140 const __m128i vabsout02 = _mm_srl_epi64(_mm_add_epi64(vabsprod02, vrounding), vshift); in xnn_qs8_gavgpool_minmax_ukernel_7x__ssse3_c16_acc2()
229 const __m128i vabsprod02 = _mm_mul_epu32(vabsacc0123, vmultiplier); in xnn_qs8_gavgpool_minmax_ukernel_7x__ssse3_c16_acc2() local
234 const __m128i vabsout02 = _mm_srl_epi64(_mm_add_epi64(vabsprod02, vrounding), vshift); in xnn_qs8_gavgpool_minmax_ukernel_7x__ssse3_c16_acc2()
D7x-minmax-sse2-c16-acc2.c136 const __m128i vabsprod02 = _mm_mul_epu32(vabsacc0123, vmultiplier); in xnn_qs8_gavgpool_minmax_ukernel_7x__sse2_c16_acc2() local
145 const __m128i vabsout02 = _mm_srl_epi64(_mm_add_epi64(vabsprod02, vrounding), vshift); in xnn_qs8_gavgpool_minmax_ukernel_7x__sse2_c16_acc2()
237 const __m128i vabsprod02 = _mm_mul_epu32(vabsacc0123, vmultiplier); in xnn_qs8_gavgpool_minmax_ukernel_7x__sse2_c16_acc2() local
242 const __m128i vabsout02 = _mm_srl_epi64(_mm_add_epi64(vabsprod02, vrounding), vshift); in xnn_qs8_gavgpool_minmax_ukernel_7x__sse2_c16_acc2()
D7p7x-minmax-sse2-c8-acc2.c218 const __m128i vabsprod02 = _mm_mul_epu32(vabsacc0123, vmultiplier); in xnn_qs8_gavgpool_minmax_ukernel_7p7x__sse2_c8_acc2() local
223 const __m128i vabsout02 = _mm_srl_epi64(_mm_add_epi64(vabsprod02, vrounding), vshift); in xnn_qs8_gavgpool_minmax_ukernel_7p7x__sse2_c8_acc2()
302 const __m128i vabsprod02 = _mm_mul_epu32(vabsacc0123, vmultiplier); in xnn_qs8_gavgpool_minmax_ukernel_7p7x__sse2_c8_acc2() local
307 const __m128i vabsout02 = _mm_srl_epi64(_mm_add_epi64(vabsprod02, vrounding), vshift); in xnn_qs8_gavgpool_minmax_ukernel_7p7x__sse2_c8_acc2()
D7p7x-minmax-ssse3-c8-acc2.c215 const __m128i vabsprod02 = _mm_mul_epu32(vabsacc0123, vmultiplier); in xnn_qs8_gavgpool_minmax_ukernel_7p7x__ssse3_c8_acc2() local
220 const __m128i vabsout02 = _mm_srl_epi64(_mm_add_epi64(vabsprod02, vrounding), vshift); in xnn_qs8_gavgpool_minmax_ukernel_7p7x__ssse3_c8_acc2()
296 const __m128i vabsprod02 = _mm_mul_epu32(vabsacc0123, vmultiplier); in xnn_qs8_gavgpool_minmax_ukernel_7p7x__ssse3_c8_acc2() local
301 const __m128i vabsout02 = _mm_srl_epi64(_mm_add_epi64(vabsprod02, vrounding), vshift); in xnn_qs8_gavgpool_minmax_ukernel_7p7x__ssse3_c8_acc2()
D7x-minmax-ssse3-c24-acc2.c158 const __m128i vabsprod02 = _mm_mul_epu32(vabsacc0123, vmultiplier); in xnn_qs8_gavgpool_minmax_ukernel_7x__ssse3_c24_acc2() local
171 const __m128i vabsout02 = _mm_srl_epi64(_mm_add_epi64(vabsprod02, vrounding), vshift); in xnn_qs8_gavgpool_minmax_ukernel_7x__ssse3_c24_acc2()
276 const __m128i vabsprod02 = _mm_mul_epu32(vabsacc0123, vmultiplier); in xnn_qs8_gavgpool_minmax_ukernel_7x__ssse3_c24_acc2() local
281 const __m128i vabsout02 = _mm_srl_epi64(_mm_add_epi64(vabsprod02, vrounding), vshift); in xnn_qs8_gavgpool_minmax_ukernel_7x__ssse3_c24_acc2()
D7x-minmax-sse2-c24-acc2.c165 const __m128i vabsprod02 = _mm_mul_epu32(vabsacc0123, vmultiplier); in xnn_qs8_gavgpool_minmax_ukernel_7x__sse2_c24_acc2() local
178 const __m128i vabsout02 = _mm_srl_epi64(_mm_add_epi64(vabsprod02, vrounding), vshift); in xnn_qs8_gavgpool_minmax_ukernel_7x__sse2_c24_acc2()
286 const __m128i vabsprod02 = _mm_mul_epu32(vabsacc0123, vmultiplier); in xnn_qs8_gavgpool_minmax_ukernel_7x__sse2_c24_acc2() local
291 const __m128i vabsout02 = _mm_srl_epi64(_mm_add_epi64(vabsprod02, vrounding), vshift); in xnn_qs8_gavgpool_minmax_ukernel_7x__sse2_c24_acc2()
D7p7x-minmax-sse41-c16-acc2.c244 const __m128i vabsprod02 = _mm_mul_epu32(vabsacc0123, vmultiplier); in xnn_qs8_gavgpool_minmax_ukernel_7p7x__sse41_c16_acc2() local
253 const __m128i vabsout02 = _mm_srl_epi64(_mm_add_epi64(vabsprod02, vrounding), vshift); in xnn_qs8_gavgpool_minmax_ukernel_7p7x__sse41_c16_acc2()
326 const __m128i vabsprod02 = _mm_mul_epu32(vabsacc0123, vmultiplier); in xnn_qs8_gavgpool_minmax_ukernel_7p7x__sse41_c16_acc2() local
331 const __m128i vabsout02 = _mm_srl_epi64(_mm_add_epi64(vabsprod02, vrounding), vshift); in xnn_qs8_gavgpool_minmax_ukernel_7p7x__sse41_c16_acc2()
D7p7x-minmax-ssse3-c16-acc2.c292 const __m128i vabsprod02 = _mm_mul_epu32(vabsacc0123, vmultiplier); in xnn_qs8_gavgpool_minmax_ukernel_7p7x__ssse3_c16_acc2() local
301 const __m128i vabsout02 = _mm_srl_epi64(_mm_add_epi64(vabsprod02, vrounding), vshift); in xnn_qs8_gavgpool_minmax_ukernel_7p7x__ssse3_c16_acc2()
391 const __m128i vabsprod02 = _mm_mul_epu32(vabsacc0123, vmultiplier); in xnn_qs8_gavgpool_minmax_ukernel_7p7x__ssse3_c16_acc2() local
396 const __m128i vabsout02 = _mm_srl_epi64(_mm_add_epi64(vabsprod02, vrounding), vshift); in xnn_qs8_gavgpool_minmax_ukernel_7p7x__ssse3_c16_acc2()
D7p7x-minmax-sse2-c16-acc2.c297 const __m128i vabsprod02 = _mm_mul_epu32(vabsacc0123, vmultiplier); in xnn_qs8_gavgpool_minmax_ukernel_7p7x__sse2_c16_acc2() local
306 const __m128i vabsout02 = _mm_srl_epi64(_mm_add_epi64(vabsprod02, vrounding), vshift); in xnn_qs8_gavgpool_minmax_ukernel_7p7x__sse2_c16_acc2()
399 const __m128i vabsprod02 = _mm_mul_epu32(vabsacc0123, vmultiplier); in xnn_qs8_gavgpool_minmax_ukernel_7p7x__sse2_c16_acc2() local
404 const __m128i vabsout02 = _mm_srl_epi64(_mm_add_epi64(vabsprod02, vrounding), vshift); in xnn_qs8_gavgpool_minmax_ukernel_7p7x__sse2_c16_acc2()
D7p7x-minmax-sse41-c24-acc2.c373 const __m128i vabsprod02 = _mm_mul_epu32(vabsacc0123, vmultiplier); in xnn_qs8_gavgpool_minmax_ukernel_7p7x__sse41_c24_acc2() local
386 const __m128i vabsout02 = _mm_srl_epi64(_mm_add_epi64(vabsprod02, vrounding), vshift); in xnn_qs8_gavgpool_minmax_ukernel_7p7x__sse41_c24_acc2()
471 const __m128i vabsprod02 = _mm_mul_epu32(vabsacc0123, vmultiplier); in xnn_qs8_gavgpool_minmax_ukernel_7p7x__sse41_c24_acc2() local
476 const __m128i vabsout02 = _mm_srl_epi64(_mm_add_epi64(vabsprod02, vrounding), vshift); in xnn_qs8_gavgpool_minmax_ukernel_7p7x__sse41_c24_acc2()
D7p7x-minmax-sse2-c24-acc2.c468 const __m128i vabsprod02 = _mm_mul_epu32(vabsacc0123, vmultiplier); in xnn_qs8_gavgpool_minmax_ukernel_7p7x__sse2_c24_acc2() local
481 const __m128i vabsout02 = _mm_srl_epi64(_mm_add_epi64(vabsprod02, vrounding), vshift); in xnn_qs8_gavgpool_minmax_ukernel_7p7x__sse2_c24_acc2()
590 const __m128i vabsprod02 = _mm_mul_epu32(vabsacc0123, vmultiplier); in xnn_qs8_gavgpool_minmax_ukernel_7p7x__sse2_c24_acc2() local
595 const __m128i vabsout02 = _mm_srl_epi64(_mm_add_epi64(vabsprod02, vrounding), vshift); in xnn_qs8_gavgpool_minmax_ukernel_7p7x__sse2_c24_acc2()
D7p7x-minmax-ssse3-c24-acc2.c461 const __m128i vabsprod02 = _mm_mul_epu32(vabsacc0123, vmultiplier); in xnn_qs8_gavgpool_minmax_ukernel_7p7x__ssse3_c24_acc2() local
474 const __m128i vabsout02 = _mm_srl_epi64(_mm_add_epi64(vabsprod02, vrounding), vshift); in xnn_qs8_gavgpool_minmax_ukernel_7p7x__ssse3_c24_acc2()
580 const __m128i vabsprod02 = _mm_mul_epu32(vabsacc0123, vmultiplier); in xnn_qs8_gavgpool_minmax_ukernel_7p7x__ssse3_c24_acc2() local
585 const __m128i vabsout02 = _mm_srl_epi64(_mm_add_epi64(vabsprod02, vrounding), vshift); in xnn_qs8_gavgpool_minmax_ukernel_7p7x__ssse3_c24_acc2()
/external/XNNPACK/src/qs8-dwconv/gen/
Dup8x9-minmax-sse2-mul16.c216 const __m128i vabsprod02 = _mm_mul_epu32(vabsacc0123, vmultiplier); in xnn_qs8_dwconv_minmax_ukernel_up8x9__sse2_mul16() local
227 const __m128i vprod02 = _mm_sub_epi64(_mm_xor_si128(vabsprod02, vnmask02), vnmask02); in xnn_qs8_dwconv_minmax_ukernel_up8x9__sse2_mul16()
397 const __m128i vabsprod02 = _mm_mul_epu32(vabsacc0123, vmultiplier); in xnn_qs8_dwconv_minmax_ukernel_up8x9__sse2_mul16() local
407 const __m128i vprod02 = _mm_sub_epi64(_mm_xor_si128(vabsprod02, vnmask02), vnmask02); in xnn_qs8_dwconv_minmax_ukernel_up8x9__sse2_mul16()
Dup8x9-minmax-ssse3-mul16.c216 const __m128i vabsprod02 = _mm_mul_epu32(vabsacc0123, vmultiplier); in xnn_qs8_dwconv_minmax_ukernel_up8x9__ssse3_mul16() local
227 const __m128i vprod02 = _mm_sub_epi64(_mm_xor_si128(vabsprod02, vnmask02), vnmask02); in xnn_qs8_dwconv_minmax_ukernel_up8x9__ssse3_mul16()
397 const __m128i vabsprod02 = _mm_mul_epu32(vabsacc0123, vmultiplier); in xnn_qs8_dwconv_minmax_ukernel_up8x9__ssse3_mul16() local
407 const __m128i vprod02 = _mm_sub_epi64(_mm_xor_si128(vabsprod02, vnmask02), vnmask02); in xnn_qs8_dwconv_minmax_ukernel_up8x9__ssse3_mul16()
Dup16x9-minmax-ssse3-mul16.c294 const __m128i vabsprod02 = _mm_mul_epu32(vabsacc0123, vmultiplier); in xnn_qs8_dwconv_minmax_ukernel_up16x9__ssse3_mul16() local
315 const __m128i vprod02 = _mm_sub_epi64(_mm_xor_si128(vabsprod02, vnmask02), vnmask02); in xnn_qs8_dwconv_minmax_ukernel_up16x9__ssse3_mul16()
521 const __m128i vabsprod02 = _mm_mul_epu32(vabsacc0123, vmultiplier); in xnn_qs8_dwconv_minmax_ukernel_up16x9__ssse3_mul16() local
531 const __m128i vprod02 = _mm_sub_epi64(_mm_xor_si128(vabsprod02, vnmask02), vnmask02); in xnn_qs8_dwconv_minmax_ukernel_up16x9__ssse3_mul16()
Dup16x9-minmax-sse2-mul16.c294 const __m128i vabsprod02 = _mm_mul_epu32(vabsacc0123, vmultiplier); in xnn_qs8_dwconv_minmax_ukernel_up16x9__sse2_mul16() local
315 const __m128i vprod02 = _mm_sub_epi64(_mm_xor_si128(vabsprod02, vnmask02), vnmask02); in xnn_qs8_dwconv_minmax_ukernel_up16x9__sse2_mul16()
521 const __m128i vabsprod02 = _mm_mul_epu32(vabsacc0123, vmultiplier); in xnn_qs8_dwconv_minmax_ukernel_up16x9__sse2_mul16() local
531 const __m128i vprod02 = _mm_sub_epi64(_mm_xor_si128(vabsprod02, vnmask02), vnmask02); in xnn_qs8_dwconv_minmax_ukernel_up16x9__sse2_mul16()
Dup24x9-minmax-ssse3-mul16.c372 const __m128i vabsprod02 = _mm_mul_epu32(vabsacc0123, vmultiplier); in xnn_qs8_dwconv_minmax_ukernel_up24x9__ssse3_mul16() local
403 const __m128i vprod02 = _mm_sub_epi64(_mm_xor_si128(vabsprod02, vnmask02), vnmask02); in xnn_qs8_dwconv_minmax_ukernel_up24x9__ssse3_mul16()
635 const __m128i vabsprod02 = _mm_mul_epu32(vabsacc0123, vmultiplier); in xnn_qs8_dwconv_minmax_ukernel_up24x9__ssse3_mul16() local
645 const __m128i vprod02 = _mm_sub_epi64(_mm_xor_si128(vabsprod02, vnmask02), vnmask02); in xnn_qs8_dwconv_minmax_ukernel_up24x9__ssse3_mul16()
Dup24x9-minmax-sse2-mul16.c372 const __m128i vabsprod02 = _mm_mul_epu32(vabsacc0123, vmultiplier); in xnn_qs8_dwconv_minmax_ukernel_up24x9__sse2_mul16() local
403 const __m128i vprod02 = _mm_sub_epi64(_mm_xor_si128(vabsprod02, vnmask02), vnmask02); in xnn_qs8_dwconv_minmax_ukernel_up24x9__sse2_mul16()
635 const __m128i vabsprod02 = _mm_mul_epu32(vabsacc0123, vmultiplier); in xnn_qs8_dwconv_minmax_ukernel_up24x9__sse2_mul16() local
645 const __m128i vprod02 = _mm_sub_epi64(_mm_xor_si128(vabsprod02, vnmask02), vnmask02); in xnn_qs8_dwconv_minmax_ukernel_up24x9__sse2_mul16()