/external/XNNPACK/src/qs8-dwconv/gen/ |
D | up8x9-minmax-sse2-mul16.c | 223 const __m128i vnmask13 = _mm_shuffle_epi32(vnmask0123, _MM_SHUFFLE(3, 3, 1, 1)); in xnn_qs8_dwconv_minmax_ukernel_up8x9__sse2_mul16() local 228 const __m128i vprod13 = _mm_sub_epi64(_mm_xor_si128(vabsprod13, vnmask13), vnmask13); in xnn_qs8_dwconv_minmax_ukernel_up8x9__sse2_mul16() 403 const __m128i vnmask13 = _mm_shuffle_epi32(vnmask0123, _MM_SHUFFLE(3, 3, 1, 1)); in xnn_qs8_dwconv_minmax_ukernel_up8x9__sse2_mul16() local 408 const __m128i vprod13 = _mm_sub_epi64(_mm_xor_si128(vabsprod13, vnmask13), vnmask13); in xnn_qs8_dwconv_minmax_ukernel_up8x9__sse2_mul16()
|
D | up8x9-minmax-ssse3-mul16.c | 223 const __m128i vnmask13 = _mm_shuffle_epi32(vnmask0123, _MM_SHUFFLE(3, 3, 1, 1)); in xnn_qs8_dwconv_minmax_ukernel_up8x9__ssse3_mul16() local 228 const __m128i vprod13 = _mm_sub_epi64(_mm_xor_si128(vabsprod13, vnmask13), vnmask13); in xnn_qs8_dwconv_minmax_ukernel_up8x9__ssse3_mul16() 403 const __m128i vnmask13 = _mm_shuffle_epi32(vnmask0123, _MM_SHUFFLE(3, 3, 1, 1)); in xnn_qs8_dwconv_minmax_ukernel_up8x9__ssse3_mul16() local 408 const __m128i vprod13 = _mm_sub_epi64(_mm_xor_si128(vabsprod13, vnmask13), vnmask13); in xnn_qs8_dwconv_minmax_ukernel_up8x9__ssse3_mul16()
|
D | up16x9-minmax-ssse3-mul16.c | 307 const __m128i vnmask13 = _mm_shuffle_epi32(vnmask0123, _MM_SHUFFLE(3, 3, 1, 1)); in xnn_qs8_dwconv_minmax_ukernel_up16x9__ssse3_mul16() local 316 const __m128i vprod13 = _mm_sub_epi64(_mm_xor_si128(vabsprod13, vnmask13), vnmask13); in xnn_qs8_dwconv_minmax_ukernel_up16x9__ssse3_mul16() 527 const __m128i vnmask13 = _mm_shuffle_epi32(vnmask0123, _MM_SHUFFLE(3, 3, 1, 1)); in xnn_qs8_dwconv_minmax_ukernel_up16x9__ssse3_mul16() local 532 const __m128i vprod13 = _mm_sub_epi64(_mm_xor_si128(vabsprod13, vnmask13), vnmask13); in xnn_qs8_dwconv_minmax_ukernel_up16x9__ssse3_mul16()
|
D | up16x9-minmax-sse2-mul16.c | 307 const __m128i vnmask13 = _mm_shuffle_epi32(vnmask0123, _MM_SHUFFLE(3, 3, 1, 1)); in xnn_qs8_dwconv_minmax_ukernel_up16x9__sse2_mul16() local 316 const __m128i vprod13 = _mm_sub_epi64(_mm_xor_si128(vabsprod13, vnmask13), vnmask13); in xnn_qs8_dwconv_minmax_ukernel_up16x9__sse2_mul16() 527 const __m128i vnmask13 = _mm_shuffle_epi32(vnmask0123, _MM_SHUFFLE(3, 3, 1, 1)); in xnn_qs8_dwconv_minmax_ukernel_up16x9__sse2_mul16() local 532 const __m128i vprod13 = _mm_sub_epi64(_mm_xor_si128(vabsprod13, vnmask13), vnmask13); in xnn_qs8_dwconv_minmax_ukernel_up16x9__sse2_mul16()
|
D | up24x9-minmax-ssse3-mul16.c | 391 const __m128i vnmask13 = _mm_shuffle_epi32(vnmask0123, _MM_SHUFFLE(3, 3, 1, 1)); in xnn_qs8_dwconv_minmax_ukernel_up24x9__ssse3_mul16() local 404 const __m128i vprod13 = _mm_sub_epi64(_mm_xor_si128(vabsprod13, vnmask13), vnmask13); in xnn_qs8_dwconv_minmax_ukernel_up24x9__ssse3_mul16() 641 const __m128i vnmask13 = _mm_shuffle_epi32(vnmask0123, _MM_SHUFFLE(3, 3, 1, 1)); in xnn_qs8_dwconv_minmax_ukernel_up24x9__ssse3_mul16() local 646 const __m128i vprod13 = _mm_sub_epi64(_mm_xor_si128(vabsprod13, vnmask13), vnmask13); in xnn_qs8_dwconv_minmax_ukernel_up24x9__ssse3_mul16()
|
D | up24x9-minmax-sse2-mul16.c | 391 const __m128i vnmask13 = _mm_shuffle_epi32(vnmask0123, _MM_SHUFFLE(3, 3, 1, 1)); in xnn_qs8_dwconv_minmax_ukernel_up24x9__sse2_mul16() local 404 const __m128i vprod13 = _mm_sub_epi64(_mm_xor_si128(vabsprod13, vnmask13), vnmask13); in xnn_qs8_dwconv_minmax_ukernel_up24x9__sse2_mul16() 641 const __m128i vnmask13 = _mm_shuffle_epi32(vnmask0123, _MM_SHUFFLE(3, 3, 1, 1)); in xnn_qs8_dwconv_minmax_ukernel_up24x9__sse2_mul16() local 646 const __m128i vprod13 = _mm_sub_epi64(_mm_xor_si128(vabsprod13, vnmask13), vnmask13); in xnn_qs8_dwconv_minmax_ukernel_up24x9__sse2_mul16()
|