| /external/XNNPACK/src/qs8-gavgpool/gen/ |
| D | 7p7x-minmax-sse2-c24-acc2.c | 73 …const __m128i vxi0x01234567 = _mm_unpacklo_epi8(vi0x01234567, _mm_cmpgt_epi8(_mm_setzero_si128(), … in xnn_qs8_gavgpool_minmax_ukernel_7p7x__sse2_c24_acc2()
74 …const __m128i vxi0x89ABCDEF = _mm_unpacklo_epi8(vi0x89ABCDEF, _mm_cmpgt_epi8(_mm_setzero_si128(), … in xnn_qs8_gavgpool_minmax_ukernel_7p7x__sse2_c24_acc2()
75 …const __m128i vxi0xGHIJKLMN = _mm_unpacklo_epi8(vi0xGHIJKLMN, _mm_cmpgt_epi8(_mm_setzero_si128(), … in xnn_qs8_gavgpool_minmax_ukernel_7p7x__sse2_c24_acc2()
76 …const __m128i vxi1x01234567 = _mm_unpacklo_epi8(vi1x01234567, _mm_cmpgt_epi8(_mm_setzero_si128(), … in xnn_qs8_gavgpool_minmax_ukernel_7p7x__sse2_c24_acc2()
77 …const __m128i vxi1x89ABCDEF = _mm_unpacklo_epi8(vi1x89ABCDEF, _mm_cmpgt_epi8(_mm_setzero_si128(), … in xnn_qs8_gavgpool_minmax_ukernel_7p7x__sse2_c24_acc2()
78 …const __m128i vxi1xGHIJKLMN = _mm_unpacklo_epi8(vi1xGHIJKLMN, _mm_cmpgt_epi8(_mm_setzero_si128(), … in xnn_qs8_gavgpool_minmax_ukernel_7p7x__sse2_c24_acc2()
79 …const __m128i vxi2x01234567 = _mm_unpacklo_epi8(vi2x01234567, _mm_cmpgt_epi8(_mm_setzero_si128(), … in xnn_qs8_gavgpool_minmax_ukernel_7p7x__sse2_c24_acc2()
80 …const __m128i vxi2x89ABCDEF = _mm_unpacklo_epi8(vi2x89ABCDEF, _mm_cmpgt_epi8(_mm_setzero_si128(), … in xnn_qs8_gavgpool_minmax_ukernel_7p7x__sse2_c24_acc2()
81 …const __m128i vxi2xGHIJKLMN = _mm_unpacklo_epi8(vi2xGHIJKLMN, _mm_cmpgt_epi8(_mm_setzero_si128(), … in xnn_qs8_gavgpool_minmax_ukernel_7p7x__sse2_c24_acc2()
82 …const __m128i vxi3x01234567 = _mm_unpacklo_epi8(vi3x01234567, _mm_cmpgt_epi8(_mm_setzero_si128(), … in xnn_qs8_gavgpool_minmax_ukernel_7p7x__sse2_c24_acc2()
[all …]
|
| D | 7p7x-minmax-ssse3-c24-acc2.c | 73 …const __m128i vxi0x01234567 = _mm_unpacklo_epi8(vi0x01234567, _mm_cmpgt_epi8(_mm_setzero_si128(), … in xnn_qs8_gavgpool_minmax_ukernel_7p7x__ssse3_c24_acc2()
74 …const __m128i vxi0x89ABCDEF = _mm_unpacklo_epi8(vi0x89ABCDEF, _mm_cmpgt_epi8(_mm_setzero_si128(), … in xnn_qs8_gavgpool_minmax_ukernel_7p7x__ssse3_c24_acc2()
75 …const __m128i vxi0xGHIJKLMN = _mm_unpacklo_epi8(vi0xGHIJKLMN, _mm_cmpgt_epi8(_mm_setzero_si128(), … in xnn_qs8_gavgpool_minmax_ukernel_7p7x__ssse3_c24_acc2()
76 …const __m128i vxi1x01234567 = _mm_unpacklo_epi8(vi1x01234567, _mm_cmpgt_epi8(_mm_setzero_si128(), … in xnn_qs8_gavgpool_minmax_ukernel_7p7x__ssse3_c24_acc2()
77 …const __m128i vxi1x89ABCDEF = _mm_unpacklo_epi8(vi1x89ABCDEF, _mm_cmpgt_epi8(_mm_setzero_si128(), … in xnn_qs8_gavgpool_minmax_ukernel_7p7x__ssse3_c24_acc2()
78 …const __m128i vxi1xGHIJKLMN = _mm_unpacklo_epi8(vi1xGHIJKLMN, _mm_cmpgt_epi8(_mm_setzero_si128(), … in xnn_qs8_gavgpool_minmax_ukernel_7p7x__ssse3_c24_acc2()
79 …const __m128i vxi2x01234567 = _mm_unpacklo_epi8(vi2x01234567, _mm_cmpgt_epi8(_mm_setzero_si128(), … in xnn_qs8_gavgpool_minmax_ukernel_7p7x__ssse3_c24_acc2()
80 …const __m128i vxi2x89ABCDEF = _mm_unpacklo_epi8(vi2x89ABCDEF, _mm_cmpgt_epi8(_mm_setzero_si128(), … in xnn_qs8_gavgpool_minmax_ukernel_7p7x__ssse3_c24_acc2()
81 …const __m128i vxi2xGHIJKLMN = _mm_unpacklo_epi8(vi2xGHIJKLMN, _mm_cmpgt_epi8(_mm_setzero_si128(), … in xnn_qs8_gavgpool_minmax_ukernel_7p7x__ssse3_c24_acc2()
82 …const __m128i vxi3x01234567 = _mm_unpacklo_epi8(vi3x01234567, _mm_cmpgt_epi8(_mm_setzero_si128(), … in xnn_qs8_gavgpool_minmax_ukernel_7p7x__ssse3_c24_acc2()
[all …]
|
| D | 7p7x-minmax-sse2-c16-acc2.c | 66 …const __m128i vxi0x01234567 = _mm_unpacklo_epi8(vi0x01234567, _mm_cmpgt_epi8(_mm_setzero_si128(), … in xnn_qs8_gavgpool_minmax_ukernel_7p7x__sse2_c16_acc2()
67 …const __m128i vxi0x89ABCDEF = _mm_unpacklo_epi8(vi0x89ABCDEF, _mm_cmpgt_epi8(_mm_setzero_si128(), … in xnn_qs8_gavgpool_minmax_ukernel_7p7x__sse2_c16_acc2()
68 …const __m128i vxi1x01234567 = _mm_unpacklo_epi8(vi1x01234567, _mm_cmpgt_epi8(_mm_setzero_si128(), … in xnn_qs8_gavgpool_minmax_ukernel_7p7x__sse2_c16_acc2()
69 …const __m128i vxi1x89ABCDEF = _mm_unpacklo_epi8(vi1x89ABCDEF, _mm_cmpgt_epi8(_mm_setzero_si128(), … in xnn_qs8_gavgpool_minmax_ukernel_7p7x__sse2_c16_acc2()
70 …const __m128i vxi2x01234567 = _mm_unpacklo_epi8(vi2x01234567, _mm_cmpgt_epi8(_mm_setzero_si128(), … in xnn_qs8_gavgpool_minmax_ukernel_7p7x__sse2_c16_acc2()
71 …const __m128i vxi2x89ABCDEF = _mm_unpacklo_epi8(vi2x89ABCDEF, _mm_cmpgt_epi8(_mm_setzero_si128(), … in xnn_qs8_gavgpool_minmax_ukernel_7p7x__sse2_c16_acc2()
72 …const __m128i vxi3x01234567 = _mm_unpacklo_epi8(vi3x01234567, _mm_cmpgt_epi8(_mm_setzero_si128(), … in xnn_qs8_gavgpool_minmax_ukernel_7p7x__sse2_c16_acc2()
73 …const __m128i vxi3x89ABCDEF = _mm_unpacklo_epi8(vi3x89ABCDEF, _mm_cmpgt_epi8(_mm_setzero_si128(), … in xnn_qs8_gavgpool_minmax_ukernel_7p7x__sse2_c16_acc2()
74 …const __m128i vxi4x01234567 = _mm_unpacklo_epi8(vi4x01234567, _mm_cmpgt_epi8(_mm_setzero_si128(), … in xnn_qs8_gavgpool_minmax_ukernel_7p7x__sse2_c16_acc2()
75 …const __m128i vxi4x89ABCDEF = _mm_unpacklo_epi8(vi4x89ABCDEF, _mm_cmpgt_epi8(_mm_setzero_si128(), … in xnn_qs8_gavgpool_minmax_ukernel_7p7x__sse2_c16_acc2()
[all …]
|
| D | 7p7x-minmax-ssse3-c16-acc2.c | 66 …const __m128i vxi0x01234567 = _mm_unpacklo_epi8(vi0x01234567, _mm_cmpgt_epi8(_mm_setzero_si128(), … in xnn_qs8_gavgpool_minmax_ukernel_7p7x__ssse3_c16_acc2()
67 …const __m128i vxi0x89ABCDEF = _mm_unpacklo_epi8(vi0x89ABCDEF, _mm_cmpgt_epi8(_mm_setzero_si128(), … in xnn_qs8_gavgpool_minmax_ukernel_7p7x__ssse3_c16_acc2()
68 …const __m128i vxi1x01234567 = _mm_unpacklo_epi8(vi1x01234567, _mm_cmpgt_epi8(_mm_setzero_si128(), … in xnn_qs8_gavgpool_minmax_ukernel_7p7x__ssse3_c16_acc2()
69 …const __m128i vxi1x89ABCDEF = _mm_unpacklo_epi8(vi1x89ABCDEF, _mm_cmpgt_epi8(_mm_setzero_si128(), … in xnn_qs8_gavgpool_minmax_ukernel_7p7x__ssse3_c16_acc2()
70 …const __m128i vxi2x01234567 = _mm_unpacklo_epi8(vi2x01234567, _mm_cmpgt_epi8(_mm_setzero_si128(), … in xnn_qs8_gavgpool_minmax_ukernel_7p7x__ssse3_c16_acc2()
71 …const __m128i vxi2x89ABCDEF = _mm_unpacklo_epi8(vi2x89ABCDEF, _mm_cmpgt_epi8(_mm_setzero_si128(), … in xnn_qs8_gavgpool_minmax_ukernel_7p7x__ssse3_c16_acc2()
72 …const __m128i vxi3x01234567 = _mm_unpacklo_epi8(vi3x01234567, _mm_cmpgt_epi8(_mm_setzero_si128(), … in xnn_qs8_gavgpool_minmax_ukernel_7p7x__ssse3_c16_acc2()
73 …const __m128i vxi3x89ABCDEF = _mm_unpacklo_epi8(vi3x89ABCDEF, _mm_cmpgt_epi8(_mm_setzero_si128(), … in xnn_qs8_gavgpool_minmax_ukernel_7p7x__ssse3_c16_acc2()
74 …const __m128i vxi4x01234567 = _mm_unpacklo_epi8(vi4x01234567, _mm_cmpgt_epi8(_mm_setzero_si128(), … in xnn_qs8_gavgpool_minmax_ukernel_7p7x__ssse3_c16_acc2()
75 …const __m128i vxi4x89ABCDEF = _mm_unpacklo_epi8(vi4x89ABCDEF, _mm_cmpgt_epi8(_mm_setzero_si128(), … in xnn_qs8_gavgpool_minmax_ukernel_7p7x__ssse3_c16_acc2()
[all …]
|
| D | 7x-minmax-sse2-c24-acc2.c | 90 …const __m128i vxi0x01234567 = _mm_unpacklo_epi8(vi0x01234567, _mm_cmpgt_epi8(_mm_setzero_si128(), … in xnn_qs8_gavgpool_minmax_ukernel_7x__sse2_c24_acc2()
91 …const __m128i vxi0x89ABCDEF = _mm_unpacklo_epi8(vi0x89ABCDEF, _mm_cmpgt_epi8(_mm_setzero_si128(), … in xnn_qs8_gavgpool_minmax_ukernel_7x__sse2_c24_acc2()
92 …const __m128i vxi0xGHIJKLMN = _mm_unpacklo_epi8(vi0xGHIJKLMN, _mm_cmpgt_epi8(_mm_setzero_si128(), … in xnn_qs8_gavgpool_minmax_ukernel_7x__sse2_c24_acc2()
93 …const __m128i vxi1x01234567 = _mm_unpacklo_epi8(vi1x01234567, _mm_cmpgt_epi8(_mm_setzero_si128(), … in xnn_qs8_gavgpool_minmax_ukernel_7x__sse2_c24_acc2()
94 …const __m128i vxi1x89ABCDEF = _mm_unpacklo_epi8(vi1x89ABCDEF, _mm_cmpgt_epi8(_mm_setzero_si128(), … in xnn_qs8_gavgpool_minmax_ukernel_7x__sse2_c24_acc2()
95 …const __m128i vxi1xGHIJKLMN = _mm_unpacklo_epi8(vi1xGHIJKLMN, _mm_cmpgt_epi8(_mm_setzero_si128(), … in xnn_qs8_gavgpool_minmax_ukernel_7x__sse2_c24_acc2()
96 …const __m128i vxi2x01234567 = _mm_unpacklo_epi8(vi2x01234567, _mm_cmpgt_epi8(_mm_setzero_si128(), … in xnn_qs8_gavgpool_minmax_ukernel_7x__sse2_c24_acc2()
97 …const __m128i vxi2x89ABCDEF = _mm_unpacklo_epi8(vi2x89ABCDEF, _mm_cmpgt_epi8(_mm_setzero_si128(), … in xnn_qs8_gavgpool_minmax_ukernel_7x__sse2_c24_acc2()
98 …const __m128i vxi2xGHIJKLMN = _mm_unpacklo_epi8(vi2xGHIJKLMN, _mm_cmpgt_epi8(_mm_setzero_si128(), … in xnn_qs8_gavgpool_minmax_ukernel_7x__sse2_c24_acc2()
99 …const __m128i vxi3x01234567 = _mm_unpacklo_epi8(vi3x01234567, _mm_cmpgt_epi8(_mm_setzero_si128(), … in xnn_qs8_gavgpool_minmax_ukernel_7x__sse2_c24_acc2()
[all …]
|
| D | 7p7x-minmax-sse2-c8-acc2.c | 59 …const __m128i vxi0x01234567 = _mm_unpacklo_epi8(vi0x01234567, _mm_cmpgt_epi8(_mm_setzero_si128(), … in xnn_qs8_gavgpool_minmax_ukernel_7p7x__sse2_c8_acc2()
60 …const __m128i vxi1x01234567 = _mm_unpacklo_epi8(vi1x01234567, _mm_cmpgt_epi8(_mm_setzero_si128(), … in xnn_qs8_gavgpool_minmax_ukernel_7p7x__sse2_c8_acc2()
61 …const __m128i vxi2x01234567 = _mm_unpacklo_epi8(vi2x01234567, _mm_cmpgt_epi8(_mm_setzero_si128(), … in xnn_qs8_gavgpool_minmax_ukernel_7p7x__sse2_c8_acc2()
62 …const __m128i vxi3x01234567 = _mm_unpacklo_epi8(vi3x01234567, _mm_cmpgt_epi8(_mm_setzero_si128(), … in xnn_qs8_gavgpool_minmax_ukernel_7p7x__sse2_c8_acc2()
63 …const __m128i vxi4x01234567 = _mm_unpacklo_epi8(vi4x01234567, _mm_cmpgt_epi8(_mm_setzero_si128(), … in xnn_qs8_gavgpool_minmax_ukernel_7p7x__sse2_c8_acc2()
64 …const __m128i vxi5x01234567 = _mm_unpacklo_epi8(vi5x01234567, _mm_cmpgt_epi8(_mm_setzero_si128(), … in xnn_qs8_gavgpool_minmax_ukernel_7p7x__sse2_c8_acc2()
65 …const __m128i vxi6x01234567 = _mm_unpacklo_epi8(vi6x01234567, _mm_cmpgt_epi8(_mm_setzero_si128(), … in xnn_qs8_gavgpool_minmax_ukernel_7p7x__sse2_c8_acc2()
77 const __m128i vsgnacc0x01234567 = _mm_cmpgt_epi16(_mm_setzero_si128(), vacc0x01234567); in xnn_qs8_gavgpool_minmax_ukernel_7p7x__sse2_c8_acc2()
113 …128i vxi0x01234567 = _mm_unpacklo_epi8(vi0x01234567, _mm_cmpgt_epi8(_mm_setzero_si128(), vi0x01234… in xnn_qs8_gavgpool_minmax_ukernel_7p7x__sse2_c8_acc2()
114 …128i vxi1x01234567 = _mm_unpacklo_epi8(vi1x01234567, _mm_cmpgt_epi8(_mm_setzero_si128(), vi1x01234… in xnn_qs8_gavgpool_minmax_ukernel_7p7x__sse2_c8_acc2()
[all …]
|
| D | 7p7x-minmax-ssse3-c8-acc2.c | 59 …const __m128i vxi0x01234567 = _mm_unpacklo_epi8(vi0x01234567, _mm_cmpgt_epi8(_mm_setzero_si128(), … in xnn_qs8_gavgpool_minmax_ukernel_7p7x__ssse3_c8_acc2()
60 …const __m128i vxi1x01234567 = _mm_unpacklo_epi8(vi1x01234567, _mm_cmpgt_epi8(_mm_setzero_si128(), … in xnn_qs8_gavgpool_minmax_ukernel_7p7x__ssse3_c8_acc2()
61 …const __m128i vxi2x01234567 = _mm_unpacklo_epi8(vi2x01234567, _mm_cmpgt_epi8(_mm_setzero_si128(), … in xnn_qs8_gavgpool_minmax_ukernel_7p7x__ssse3_c8_acc2()
62 …const __m128i vxi3x01234567 = _mm_unpacklo_epi8(vi3x01234567, _mm_cmpgt_epi8(_mm_setzero_si128(), … in xnn_qs8_gavgpool_minmax_ukernel_7p7x__ssse3_c8_acc2()
63 …const __m128i vxi4x01234567 = _mm_unpacklo_epi8(vi4x01234567, _mm_cmpgt_epi8(_mm_setzero_si128(), … in xnn_qs8_gavgpool_minmax_ukernel_7p7x__ssse3_c8_acc2()
64 …const __m128i vxi5x01234567 = _mm_unpacklo_epi8(vi5x01234567, _mm_cmpgt_epi8(_mm_setzero_si128(), … in xnn_qs8_gavgpool_minmax_ukernel_7p7x__ssse3_c8_acc2()
65 …const __m128i vxi6x01234567 = _mm_unpacklo_epi8(vi6x01234567, _mm_cmpgt_epi8(_mm_setzero_si128(), … in xnn_qs8_gavgpool_minmax_ukernel_7p7x__ssse3_c8_acc2()
77 const __m128i vsgnacc0x01234567 = _mm_cmpgt_epi16(_mm_setzero_si128(), vacc0x01234567); in xnn_qs8_gavgpool_minmax_ukernel_7p7x__ssse3_c8_acc2()
113 …128i vxi0x01234567 = _mm_unpacklo_epi8(vi0x01234567, _mm_cmpgt_epi8(_mm_setzero_si128(), vi0x01234… in xnn_qs8_gavgpool_minmax_ukernel_7p7x__ssse3_c8_acc2()
114 …128i vxi1x01234567 = _mm_unpacklo_epi8(vi1x01234567, _mm_cmpgt_epi8(_mm_setzero_si128(), vi1x01234… in xnn_qs8_gavgpool_minmax_ukernel_7p7x__ssse3_c8_acc2()
[all …]
|
| D | 7x-minmax-sse2-c16-acc2.c | 83 …const __m128i vxi0x01234567 = _mm_unpacklo_epi8(vi0x01234567, _mm_cmpgt_epi8(_mm_setzero_si128(), … in xnn_qs8_gavgpool_minmax_ukernel_7x__sse2_c16_acc2()
84 …const __m128i vxi0x89ABCDEF = _mm_unpacklo_epi8(vi0x89ABCDEF, _mm_cmpgt_epi8(_mm_setzero_si128(), … in xnn_qs8_gavgpool_minmax_ukernel_7x__sse2_c16_acc2()
85 …const __m128i vxi1x01234567 = _mm_unpacklo_epi8(vi1x01234567, _mm_cmpgt_epi8(_mm_setzero_si128(), … in xnn_qs8_gavgpool_minmax_ukernel_7x__sse2_c16_acc2()
86 …const __m128i vxi1x89ABCDEF = _mm_unpacklo_epi8(vi1x89ABCDEF, _mm_cmpgt_epi8(_mm_setzero_si128(), … in xnn_qs8_gavgpool_minmax_ukernel_7x__sse2_c16_acc2()
87 …const __m128i vxi2x01234567 = _mm_unpacklo_epi8(vi2x01234567, _mm_cmpgt_epi8(_mm_setzero_si128(), … in xnn_qs8_gavgpool_minmax_ukernel_7x__sse2_c16_acc2()
88 …const __m128i vxi2x89ABCDEF = _mm_unpacklo_epi8(vi2x89ABCDEF, _mm_cmpgt_epi8(_mm_setzero_si128(), … in xnn_qs8_gavgpool_minmax_ukernel_7x__sse2_c16_acc2()
89 …const __m128i vxi3x01234567 = _mm_unpacklo_epi8(vi3x01234567, _mm_cmpgt_epi8(_mm_setzero_si128(), … in xnn_qs8_gavgpool_minmax_ukernel_7x__sse2_c16_acc2()
90 …const __m128i vxi3x89ABCDEF = _mm_unpacklo_epi8(vi3x89ABCDEF, _mm_cmpgt_epi8(_mm_setzero_si128(), … in xnn_qs8_gavgpool_minmax_ukernel_7x__sse2_c16_acc2()
91 …const __m128i vxi4x01234567 = _mm_unpacklo_epi8(vi4x01234567, _mm_cmpgt_epi8(_mm_setzero_si128(), … in xnn_qs8_gavgpool_minmax_ukernel_7x__sse2_c16_acc2()
92 …const __m128i vxi4x89ABCDEF = _mm_unpacklo_epi8(vi4x89ABCDEF, _mm_cmpgt_epi8(_mm_setzero_si128(), … in xnn_qs8_gavgpool_minmax_ukernel_7x__sse2_c16_acc2()
[all …]
|
| D | 7x-minmax-ssse3-c24-acc2.c | 90 …const __m128i vxi0x01234567 = _mm_unpacklo_epi8(vi0x01234567, _mm_cmpgt_epi8(_mm_setzero_si128(), … in xnn_qs8_gavgpool_minmax_ukernel_7x__ssse3_c24_acc2()
91 …const __m128i vxi0x89ABCDEF = _mm_unpacklo_epi8(vi0x89ABCDEF, _mm_cmpgt_epi8(_mm_setzero_si128(), … in xnn_qs8_gavgpool_minmax_ukernel_7x__ssse3_c24_acc2()
92 …const __m128i vxi0xGHIJKLMN = _mm_unpacklo_epi8(vi0xGHIJKLMN, _mm_cmpgt_epi8(_mm_setzero_si128(), … in xnn_qs8_gavgpool_minmax_ukernel_7x__ssse3_c24_acc2()
93 …const __m128i vxi1x01234567 = _mm_unpacklo_epi8(vi1x01234567, _mm_cmpgt_epi8(_mm_setzero_si128(), … in xnn_qs8_gavgpool_minmax_ukernel_7x__ssse3_c24_acc2()
94 …const __m128i vxi1x89ABCDEF = _mm_unpacklo_epi8(vi1x89ABCDEF, _mm_cmpgt_epi8(_mm_setzero_si128(), … in xnn_qs8_gavgpool_minmax_ukernel_7x__ssse3_c24_acc2()
95 …const __m128i vxi1xGHIJKLMN = _mm_unpacklo_epi8(vi1xGHIJKLMN, _mm_cmpgt_epi8(_mm_setzero_si128(), … in xnn_qs8_gavgpool_minmax_ukernel_7x__ssse3_c24_acc2()
96 …const __m128i vxi2x01234567 = _mm_unpacklo_epi8(vi2x01234567, _mm_cmpgt_epi8(_mm_setzero_si128(), … in xnn_qs8_gavgpool_minmax_ukernel_7x__ssse3_c24_acc2()
97 …const __m128i vxi2x89ABCDEF = _mm_unpacklo_epi8(vi2x89ABCDEF, _mm_cmpgt_epi8(_mm_setzero_si128(), … in xnn_qs8_gavgpool_minmax_ukernel_7x__ssse3_c24_acc2()
98 …const __m128i vxi2xGHIJKLMN = _mm_unpacklo_epi8(vi2xGHIJKLMN, _mm_cmpgt_epi8(_mm_setzero_si128(), … in xnn_qs8_gavgpool_minmax_ukernel_7x__ssse3_c24_acc2()
99 …const __m128i vxi3x01234567 = _mm_unpacklo_epi8(vi3x01234567, _mm_cmpgt_epi8(_mm_setzero_si128(), … in xnn_qs8_gavgpool_minmax_ukernel_7x__ssse3_c24_acc2()
[all …]
|
| D | 7x-minmax-sse2-c8-acc2.c | 76 …const __m128i vxi0x01234567 = _mm_unpacklo_epi8(vi0x01234567, _mm_cmpgt_epi8(_mm_setzero_si128(), … in xnn_qs8_gavgpool_minmax_ukernel_7x__sse2_c8_acc2()
77 …const __m128i vxi1x01234567 = _mm_unpacklo_epi8(vi1x01234567, _mm_cmpgt_epi8(_mm_setzero_si128(), … in xnn_qs8_gavgpool_minmax_ukernel_7x__sse2_c8_acc2()
78 …const __m128i vxi2x01234567 = _mm_unpacklo_epi8(vi2x01234567, _mm_cmpgt_epi8(_mm_setzero_si128(), … in xnn_qs8_gavgpool_minmax_ukernel_7x__sse2_c8_acc2()
79 …const __m128i vxi3x01234567 = _mm_unpacklo_epi8(vi3x01234567, _mm_cmpgt_epi8(_mm_setzero_si128(), … in xnn_qs8_gavgpool_minmax_ukernel_7x__sse2_c8_acc2()
80 …const __m128i vxi4x01234567 = _mm_unpacklo_epi8(vi4x01234567, _mm_cmpgt_epi8(_mm_setzero_si128(), … in xnn_qs8_gavgpool_minmax_ukernel_7x__sse2_c8_acc2()
81 …const __m128i vxi5x01234567 = _mm_unpacklo_epi8(vi5x01234567, _mm_cmpgt_epi8(_mm_setzero_si128(), … in xnn_qs8_gavgpool_minmax_ukernel_7x__sse2_c8_acc2()
82 …const __m128i vxi6x01234567 = _mm_unpacklo_epi8(vi6x01234567, _mm_cmpgt_epi8(_mm_setzero_si128(), … in xnn_qs8_gavgpool_minmax_ukernel_7x__sse2_c8_acc2()
94 const __m128i vsgnacc0x01234567 = _mm_cmpgt_epi16(_mm_setzero_si128(), vacc0x01234567); in xnn_qs8_gavgpool_minmax_ukernel_7x__sse2_c8_acc2()
98 const __m128i vsgnacc0123 = _mm_cmpgt_epi32(_mm_setzero_si128(), vacc0123); in xnn_qs8_gavgpool_minmax_ukernel_7x__sse2_c8_acc2()
99 const __m128i vsgnacc4567 = _mm_cmpgt_epi32(_mm_setzero_si128(), vacc4567); in xnn_qs8_gavgpool_minmax_ukernel_7x__sse2_c8_acc2()
[all …]
|
| D | 7x-minmax-ssse3-c16-acc2.c | 83 …const __m128i vxi0x01234567 = _mm_unpacklo_epi8(vi0x01234567, _mm_cmpgt_epi8(_mm_setzero_si128(), … in xnn_qs8_gavgpool_minmax_ukernel_7x__ssse3_c16_acc2()
84 …const __m128i vxi0x89ABCDEF = _mm_unpacklo_epi8(vi0x89ABCDEF, _mm_cmpgt_epi8(_mm_setzero_si128(), … in xnn_qs8_gavgpool_minmax_ukernel_7x__ssse3_c16_acc2()
85 …const __m128i vxi1x01234567 = _mm_unpacklo_epi8(vi1x01234567, _mm_cmpgt_epi8(_mm_setzero_si128(), … in xnn_qs8_gavgpool_minmax_ukernel_7x__ssse3_c16_acc2()
86 …const __m128i vxi1x89ABCDEF = _mm_unpacklo_epi8(vi1x89ABCDEF, _mm_cmpgt_epi8(_mm_setzero_si128(), … in xnn_qs8_gavgpool_minmax_ukernel_7x__ssse3_c16_acc2()
87 …const __m128i vxi2x01234567 = _mm_unpacklo_epi8(vi2x01234567, _mm_cmpgt_epi8(_mm_setzero_si128(), … in xnn_qs8_gavgpool_minmax_ukernel_7x__ssse3_c16_acc2()
88 …const __m128i vxi2x89ABCDEF = _mm_unpacklo_epi8(vi2x89ABCDEF, _mm_cmpgt_epi8(_mm_setzero_si128(), … in xnn_qs8_gavgpool_minmax_ukernel_7x__ssse3_c16_acc2()
89 …const __m128i vxi3x01234567 = _mm_unpacklo_epi8(vi3x01234567, _mm_cmpgt_epi8(_mm_setzero_si128(), … in xnn_qs8_gavgpool_minmax_ukernel_7x__ssse3_c16_acc2()
90 …const __m128i vxi3x89ABCDEF = _mm_unpacklo_epi8(vi3x89ABCDEF, _mm_cmpgt_epi8(_mm_setzero_si128(), … in xnn_qs8_gavgpool_minmax_ukernel_7x__ssse3_c16_acc2()
91 …const __m128i vxi4x01234567 = _mm_unpacklo_epi8(vi4x01234567, _mm_cmpgt_epi8(_mm_setzero_si128(), … in xnn_qs8_gavgpool_minmax_ukernel_7x__ssse3_c16_acc2()
92 …const __m128i vxi4x89ABCDEF = _mm_unpacklo_epi8(vi4x89ABCDEF, _mm_cmpgt_epi8(_mm_setzero_si128(), … in xnn_qs8_gavgpool_minmax_ukernel_7x__ssse3_c16_acc2()
[all …]
|
| /external/libaom/libaom/aom_dsp/x86/ |
| D | aom_subpixel_8t_intrin_sse2.c | 48 __m128i ss_1_1 = _mm_unpacklo_epi8(ss_2, _mm_setzero_si128()); in aom_filter_block1d16_h4_sse2()
49 __m128i ss_2_1 = _mm_unpacklo_epi8(ss_4, _mm_setzero_si128()); in aom_filter_block1d16_h4_sse2()
56 __m128i ss_1_2 = _mm_unpacklo_epi8(ss_1, _mm_setzero_si128()); in aom_filter_block1d16_h4_sse2()
57 __m128i ss_2_2 = _mm_unpacklo_epi8(ss_3, _mm_setzero_si128()); in aom_filter_block1d16_h4_sse2()
72 ss_1_1 = _mm_unpacklo_epi8(ss_2, _mm_setzero_si128()); in aom_filter_block1d16_h4_sse2()
73 ss_2_1 = _mm_unpacklo_epi8(ss_4, _mm_setzero_si128()); in aom_filter_block1d16_h4_sse2()
80 ss_1_2 = _mm_unpacklo_epi8(ss_1, _mm_setzero_si128()); in aom_filter_block1d16_h4_sse2()
81 ss_2_2 = _mm_unpacklo_epi8(ss_3, _mm_setzero_si128()); in aom_filter_block1d16_h4_sse2()
145 __m128i resReg23_lo_1 = _mm_unpacklo_epi8(srcReg23_lo, _mm_setzero_si128()); in aom_filter_block1d16_v4_sse2()
146 __m128i resReg23_lo_2 = _mm_unpackhi_epi8(srcReg23_lo, _mm_setzero_si128()); in aom_filter_block1d16_v4_sse2()
[all …]
|
| D | convolve_sse2.h | 51 ss[0] = _mm_unpacklo_epi8(s[0], _mm_setzero_si128()); in convolve_lo_x()
52 ss[1] = _mm_unpacklo_epi8(s[1], _mm_setzero_si128()); in convolve_lo_x()
53 ss[2] = _mm_unpacklo_epi8(s[2], _mm_setzero_si128()); in convolve_lo_x()
54 ss[3] = _mm_unpacklo_epi8(s[3], _mm_setzero_si128()); in convolve_lo_x()
61 ss[0] = _mm_unpacklo_epi8(s[0], _mm_setzero_si128()); in convolve_lo_y()
62 ss[1] = _mm_unpacklo_epi8(s[2], _mm_setzero_si128()); in convolve_lo_y()
63 ss[2] = _mm_unpacklo_epi8(s[4], _mm_setzero_si128()); in convolve_lo_y()
64 ss[3] = _mm_unpacklo_epi8(s[6], _mm_setzero_si128()); in convolve_lo_y()
71 ss[0] = _mm_unpackhi_epi8(s[0], _mm_setzero_si128()); in convolve_hi_y()
72 ss[1] = _mm_unpackhi_epi8(s[2], _mm_setzero_si128()); in convolve_hi_y()
[all …]
|
| /external/XNNPACK/src/qs8-dwconv/gen/ |
| D | up16x9-minmax-ssse3-mul16.c | 95 …128i vxi0x01234567 = _mm_unpacklo_epi8(vi0x01234567, _mm_cmpgt_epi8(_mm_setzero_si128(), vi0x01234… in xnn_qs8_dwconv_minmax_ukernel_up16x9__ssse3_mul16()
96 …128i vxk0x01234567 = _mm_unpacklo_epi8(vk0x01234567, _mm_cmpgt_epi8(_mm_setzero_si128(), vk0x01234… in xnn_qs8_dwconv_minmax_ukernel_up16x9__ssse3_mul16()
97 …128i vxi0x89ABCDEF = _mm_unpacklo_epi8(vi0x89ABCDEF, _mm_cmpgt_epi8(_mm_setzero_si128(), vi0x89ABC… in xnn_qs8_dwconv_minmax_ukernel_up16x9__ssse3_mul16()
98 …128i vxk0x89ABCDEF = _mm_unpacklo_epi8(vk0x89ABCDEF, _mm_cmpgt_epi8(_mm_setzero_si128(), vk0x89ABC… in xnn_qs8_dwconv_minmax_ukernel_up16x9__ssse3_mul16()
116 …128i vxi1x01234567 = _mm_unpacklo_epi8(vi1x01234567, _mm_cmpgt_epi8(_mm_setzero_si128(), vi1x01234… in xnn_qs8_dwconv_minmax_ukernel_up16x9__ssse3_mul16()
117 …128i vxk1x01234567 = _mm_unpacklo_epi8(vk1x01234567, _mm_cmpgt_epi8(_mm_setzero_si128(), vk1x01234… in xnn_qs8_dwconv_minmax_ukernel_up16x9__ssse3_mul16()
118 …128i vxi1x89ABCDEF = _mm_unpacklo_epi8(vi1x89ABCDEF, _mm_cmpgt_epi8(_mm_setzero_si128(), vi1x89ABC… in xnn_qs8_dwconv_minmax_ukernel_up16x9__ssse3_mul16()
119 …128i vxk1x89ABCDEF = _mm_unpacklo_epi8(vk1x89ABCDEF, _mm_cmpgt_epi8(_mm_setzero_si128(), vk1x89ABC… in xnn_qs8_dwconv_minmax_ukernel_up16x9__ssse3_mul16()
137 …128i vxi2x01234567 = _mm_unpacklo_epi8(vi2x01234567, _mm_cmpgt_epi8(_mm_setzero_si128(), vi2x01234… in xnn_qs8_dwconv_minmax_ukernel_up16x9__ssse3_mul16()
138 …128i vxk2x01234567 = _mm_unpacklo_epi8(vk2x01234567, _mm_cmpgt_epi8(_mm_setzero_si128(), vk2x01234… in xnn_qs8_dwconv_minmax_ukernel_up16x9__ssse3_mul16()
[all …]
|
| D | up16x9-minmax-sse2-mul16.c | 95 …128i vxi0x01234567 = _mm_unpacklo_epi8(vi0x01234567, _mm_cmpgt_epi8(_mm_setzero_si128(), vi0x01234… in xnn_qs8_dwconv_minmax_ukernel_up16x9__sse2_mul16()
96 …128i vxk0x01234567 = _mm_unpacklo_epi8(vk0x01234567, _mm_cmpgt_epi8(_mm_setzero_si128(), vk0x01234… in xnn_qs8_dwconv_minmax_ukernel_up16x9__sse2_mul16()
97 …128i vxi0x89ABCDEF = _mm_unpacklo_epi8(vi0x89ABCDEF, _mm_cmpgt_epi8(_mm_setzero_si128(), vi0x89ABC… in xnn_qs8_dwconv_minmax_ukernel_up16x9__sse2_mul16()
98 …128i vxk0x89ABCDEF = _mm_unpacklo_epi8(vk0x89ABCDEF, _mm_cmpgt_epi8(_mm_setzero_si128(), vk0x89ABC… in xnn_qs8_dwconv_minmax_ukernel_up16x9__sse2_mul16()
116 …128i vxi1x01234567 = _mm_unpacklo_epi8(vi1x01234567, _mm_cmpgt_epi8(_mm_setzero_si128(), vi1x01234… in xnn_qs8_dwconv_minmax_ukernel_up16x9__sse2_mul16()
117 …128i vxk1x01234567 = _mm_unpacklo_epi8(vk1x01234567, _mm_cmpgt_epi8(_mm_setzero_si128(), vk1x01234… in xnn_qs8_dwconv_minmax_ukernel_up16x9__sse2_mul16()
118 …128i vxi1x89ABCDEF = _mm_unpacklo_epi8(vi1x89ABCDEF, _mm_cmpgt_epi8(_mm_setzero_si128(), vi1x89ABC… in xnn_qs8_dwconv_minmax_ukernel_up16x9__sse2_mul16()
119 …128i vxk1x89ABCDEF = _mm_unpacklo_epi8(vk1x89ABCDEF, _mm_cmpgt_epi8(_mm_setzero_si128(), vk1x89ABC… in xnn_qs8_dwconv_minmax_ukernel_up16x9__sse2_mul16()
137 …128i vxi2x01234567 = _mm_unpacklo_epi8(vi2x01234567, _mm_cmpgt_epi8(_mm_setzero_si128(), vi2x01234… in xnn_qs8_dwconv_minmax_ukernel_up16x9__sse2_mul16()
138 …128i vxk2x01234567 = _mm_unpacklo_epi8(vk2x01234567, _mm_cmpgt_epi8(_mm_setzero_si128(), vk2x01234… in xnn_qs8_dwconv_minmax_ukernel_up16x9__sse2_mul16()
[all …]
|
| D | up24x9-minmax-ssse3-mul16.c | 99 …128i vxi0x01234567 = _mm_unpacklo_epi8(vi0x01234567, _mm_cmpgt_epi8(_mm_setzero_si128(), vi0x01234… in xnn_qs8_dwconv_minmax_ukernel_up24x9__ssse3_mul16()
100 …128i vxk0x01234567 = _mm_unpacklo_epi8(vk0x01234567, _mm_cmpgt_epi8(_mm_setzero_si128(), vk0x01234… in xnn_qs8_dwconv_minmax_ukernel_up24x9__ssse3_mul16()
101 …128i vxi0x89ABCDEF = _mm_unpacklo_epi8(vi0x89ABCDEF, _mm_cmpgt_epi8(_mm_setzero_si128(), vi0x89ABC… in xnn_qs8_dwconv_minmax_ukernel_up24x9__ssse3_mul16()
102 …128i vxk0x89ABCDEF = _mm_unpacklo_epi8(vk0x89ABCDEF, _mm_cmpgt_epi8(_mm_setzero_si128(), vk0x89ABC… in xnn_qs8_dwconv_minmax_ukernel_up24x9__ssse3_mul16()
103 …128i vxi0xGHIJKLMN = _mm_unpacklo_epi8(vi0xGHIJKLMN, _mm_cmpgt_epi8(_mm_setzero_si128(), vi0xGHIJK… in xnn_qs8_dwconv_minmax_ukernel_up24x9__ssse3_mul16()
104 …128i vxk0xGHIJKLMN = _mm_unpacklo_epi8(vk0xGHIJKLMN, _mm_cmpgt_epi8(_mm_setzero_si128(), vk0xGHIJK… in xnn_qs8_dwconv_minmax_ukernel_up24x9__ssse3_mul16()
128 …128i vxi1x01234567 = _mm_unpacklo_epi8(vi1x01234567, _mm_cmpgt_epi8(_mm_setzero_si128(), vi1x01234… in xnn_qs8_dwconv_minmax_ukernel_up24x9__ssse3_mul16()
129 …128i vxk1x01234567 = _mm_unpacklo_epi8(vk1x01234567, _mm_cmpgt_epi8(_mm_setzero_si128(), vk1x01234… in xnn_qs8_dwconv_minmax_ukernel_up24x9__ssse3_mul16()
130 …128i vxi1x89ABCDEF = _mm_unpacklo_epi8(vi1x89ABCDEF, _mm_cmpgt_epi8(_mm_setzero_si128(), vi1x89ABC… in xnn_qs8_dwconv_minmax_ukernel_up24x9__ssse3_mul16()
131 …128i vxk1x89ABCDEF = _mm_unpacklo_epi8(vk1x89ABCDEF, _mm_cmpgt_epi8(_mm_setzero_si128(), vk1x89ABC… in xnn_qs8_dwconv_minmax_ukernel_up24x9__ssse3_mul16()
[all …]
|
| D | up24x9-minmax-sse2-mul16.c | 99 …128i vxi0x01234567 = _mm_unpacklo_epi8(vi0x01234567, _mm_cmpgt_epi8(_mm_setzero_si128(), vi0x01234… in xnn_qs8_dwconv_minmax_ukernel_up24x9__sse2_mul16()
100 …128i vxk0x01234567 = _mm_unpacklo_epi8(vk0x01234567, _mm_cmpgt_epi8(_mm_setzero_si128(), vk0x01234… in xnn_qs8_dwconv_minmax_ukernel_up24x9__sse2_mul16()
101 …128i vxi0x89ABCDEF = _mm_unpacklo_epi8(vi0x89ABCDEF, _mm_cmpgt_epi8(_mm_setzero_si128(), vi0x89ABC… in xnn_qs8_dwconv_minmax_ukernel_up24x9__sse2_mul16()
102 …128i vxk0x89ABCDEF = _mm_unpacklo_epi8(vk0x89ABCDEF, _mm_cmpgt_epi8(_mm_setzero_si128(), vk0x89ABC… in xnn_qs8_dwconv_minmax_ukernel_up24x9__sse2_mul16()
103 …128i vxi0xGHIJKLMN = _mm_unpacklo_epi8(vi0xGHIJKLMN, _mm_cmpgt_epi8(_mm_setzero_si128(), vi0xGHIJK… in xnn_qs8_dwconv_minmax_ukernel_up24x9__sse2_mul16()
104 …128i vxk0xGHIJKLMN = _mm_unpacklo_epi8(vk0xGHIJKLMN, _mm_cmpgt_epi8(_mm_setzero_si128(), vk0xGHIJK… in xnn_qs8_dwconv_minmax_ukernel_up24x9__sse2_mul16()
128 …128i vxi1x01234567 = _mm_unpacklo_epi8(vi1x01234567, _mm_cmpgt_epi8(_mm_setzero_si128(), vi1x01234… in xnn_qs8_dwconv_minmax_ukernel_up24x9__sse2_mul16()
129 …128i vxk1x01234567 = _mm_unpacklo_epi8(vk1x01234567, _mm_cmpgt_epi8(_mm_setzero_si128(), vk1x01234… in xnn_qs8_dwconv_minmax_ukernel_up24x9__sse2_mul16()
130 …128i vxi1x89ABCDEF = _mm_unpacklo_epi8(vi1x89ABCDEF, _mm_cmpgt_epi8(_mm_setzero_si128(), vi1x89ABC… in xnn_qs8_dwconv_minmax_ukernel_up24x9__sse2_mul16()
131 …128i vxk1x89ABCDEF = _mm_unpacklo_epi8(vk1x89ABCDEF, _mm_cmpgt_epi8(_mm_setzero_si128(), vk1x89ABC… in xnn_qs8_dwconv_minmax_ukernel_up24x9__sse2_mul16()
[all …]
|
| D | up8x9-minmax-sse2-mul16.c | 91 …128i vxi0x01234567 = _mm_unpacklo_epi8(vi0x01234567, _mm_cmpgt_epi8(_mm_setzero_si128(), vi0x01234… in xnn_qs8_dwconv_minmax_ukernel_up8x9__sse2_mul16()
92 …128i vxk0x01234567 = _mm_unpacklo_epi8(vk0x01234567, _mm_cmpgt_epi8(_mm_setzero_si128(), vk0x01234… in xnn_qs8_dwconv_minmax_ukernel_up8x9__sse2_mul16()
104 …128i vxi1x01234567 = _mm_unpacklo_epi8(vi1x01234567, _mm_cmpgt_epi8(_mm_setzero_si128(), vi1x01234… in xnn_qs8_dwconv_minmax_ukernel_up8x9__sse2_mul16()
105 …128i vxk1x01234567 = _mm_unpacklo_epi8(vk1x01234567, _mm_cmpgt_epi8(_mm_setzero_si128(), vk1x01234… in xnn_qs8_dwconv_minmax_ukernel_up8x9__sse2_mul16()
117 …128i vxi2x01234567 = _mm_unpacklo_epi8(vi2x01234567, _mm_cmpgt_epi8(_mm_setzero_si128(), vi2x01234… in xnn_qs8_dwconv_minmax_ukernel_up8x9__sse2_mul16()
118 …128i vxk2x01234567 = _mm_unpacklo_epi8(vk2x01234567, _mm_cmpgt_epi8(_mm_setzero_si128(), vk2x01234… in xnn_qs8_dwconv_minmax_ukernel_up8x9__sse2_mul16()
130 …128i vxi3x01234567 = _mm_unpacklo_epi8(vi3x01234567, _mm_cmpgt_epi8(_mm_setzero_si128(), vi3x01234… in xnn_qs8_dwconv_minmax_ukernel_up8x9__sse2_mul16()
131 …128i vxk3x01234567 = _mm_unpacklo_epi8(vk3x01234567, _mm_cmpgt_epi8(_mm_setzero_si128(), vk3x01234… in xnn_qs8_dwconv_minmax_ukernel_up8x9__sse2_mul16()
143 …128i vxi4x01234567 = _mm_unpacklo_epi8(vi4x01234567, _mm_cmpgt_epi8(_mm_setzero_si128(), vi4x01234… in xnn_qs8_dwconv_minmax_ukernel_up8x9__sse2_mul16()
144 …128i vxk4x01234567 = _mm_unpacklo_epi8(vk4x01234567, _mm_cmpgt_epi8(_mm_setzero_si128(), vk4x01234… in xnn_qs8_dwconv_minmax_ukernel_up8x9__sse2_mul16()
[all …]
|
| D | up8x9-minmax-ssse3-mul16.c | 91 …128i vxi0x01234567 = _mm_unpacklo_epi8(vi0x01234567, _mm_cmpgt_epi8(_mm_setzero_si128(), vi0x01234… in xnn_qs8_dwconv_minmax_ukernel_up8x9__ssse3_mul16()
92 …128i vxk0x01234567 = _mm_unpacklo_epi8(vk0x01234567, _mm_cmpgt_epi8(_mm_setzero_si128(), vk0x01234… in xnn_qs8_dwconv_minmax_ukernel_up8x9__ssse3_mul16()
104 …128i vxi1x01234567 = _mm_unpacklo_epi8(vi1x01234567, _mm_cmpgt_epi8(_mm_setzero_si128(), vi1x01234… in xnn_qs8_dwconv_minmax_ukernel_up8x9__ssse3_mul16()
105 …128i vxk1x01234567 = _mm_unpacklo_epi8(vk1x01234567, _mm_cmpgt_epi8(_mm_setzero_si128(), vk1x01234… in xnn_qs8_dwconv_minmax_ukernel_up8x9__ssse3_mul16()
117 …128i vxi2x01234567 = _mm_unpacklo_epi8(vi2x01234567, _mm_cmpgt_epi8(_mm_setzero_si128(), vi2x01234… in xnn_qs8_dwconv_minmax_ukernel_up8x9__ssse3_mul16()
118 …128i vxk2x01234567 = _mm_unpacklo_epi8(vk2x01234567, _mm_cmpgt_epi8(_mm_setzero_si128(), vk2x01234… in xnn_qs8_dwconv_minmax_ukernel_up8x9__ssse3_mul16()
130 …128i vxi3x01234567 = _mm_unpacklo_epi8(vi3x01234567, _mm_cmpgt_epi8(_mm_setzero_si128(), vi3x01234… in xnn_qs8_dwconv_minmax_ukernel_up8x9__ssse3_mul16()
131 …128i vxk3x01234567 = _mm_unpacklo_epi8(vk3x01234567, _mm_cmpgt_epi8(_mm_setzero_si128(), vk3x01234… in xnn_qs8_dwconv_minmax_ukernel_up8x9__ssse3_mul16()
143 …128i vxi4x01234567 = _mm_unpacklo_epi8(vi4x01234567, _mm_cmpgt_epi8(_mm_setzero_si128(), vi4x01234… in xnn_qs8_dwconv_minmax_ukernel_up8x9__ssse3_mul16()
144 …128i vxk4x01234567 = _mm_unpacklo_epi8(vk4x01234567, _mm_cmpgt_epi8(_mm_setzero_si128(), vk4x01234… in xnn_qs8_dwconv_minmax_ukernel_up8x9__ssse3_mul16()
[all …]
|
| /external/libvpx/libvpx/vpx_dsp/x86/ |
| D | vpx_subpixel_4t_intrin_sse2.c | 159 src_reg_m10_lo_1 = _mm_unpacklo_epi8(src_reg_m10_lo, _mm_setzero_si128()); in vpx_filter_block1d16_v4_sse2()
160 src_reg_m10_lo_2 = _mm_unpackhi_epi8(src_reg_m10_lo, _mm_setzero_si128()); in vpx_filter_block1d16_v4_sse2()
161 src_reg_m10_hi_1 = _mm_unpacklo_epi8(src_reg_m10_hi, _mm_setzero_si128()); in vpx_filter_block1d16_v4_sse2()
162 src_reg_m10_hi_2 = _mm_unpackhi_epi8(src_reg_m10_hi, _mm_setzero_si128()); in vpx_filter_block1d16_v4_sse2()
168 src_reg_01_lo_1 = _mm_unpacklo_epi8(src_reg_01_lo, _mm_setzero_si128()); in vpx_filter_block1d16_v4_sse2()
169 src_reg_01_lo_2 = _mm_unpackhi_epi8(src_reg_01_lo, _mm_setzero_si128()); in vpx_filter_block1d16_v4_sse2()
170 src_reg_01_hi_1 = _mm_unpacklo_epi8(src_reg_01_hi, _mm_setzero_si128()); in vpx_filter_block1d16_v4_sse2()
171 src_reg_01_hi_2 = _mm_unpackhi_epi8(src_reg_01_hi, _mm_setzero_si128()); in vpx_filter_block1d16_v4_sse2()
191 src_reg_12_lo_1 = _mm_unpacklo_epi8(src_reg_12_lo, _mm_setzero_si128()); in vpx_filter_block1d16_v4_sse2()
192 src_reg_12_lo_2 = _mm_unpackhi_epi8(src_reg_12_lo, _mm_setzero_si128()); in vpx_filter_block1d16_v4_sse2()
[all …]
|
| /external/XNNPACK/src/qs8-vadd/gen/ |
| D | minmax-sse2-mul16-ld64-x32.c | 48 vx01234567 = _mm_unpacklo_epi8(vx01234567, _mm_cmpgt_epi8(_mm_setzero_si128(), vx01234567)); in xnn_qs8_vadd_minmax_ukernel__sse2_mul16_ld64_x32()
49 vy01234567 = _mm_unpacklo_epi8(vy01234567, _mm_cmpgt_epi8(_mm_setzero_si128(), vy01234567)); in xnn_qs8_vadd_minmax_ukernel__sse2_mul16_ld64_x32()
50 vx89ABCDEF = _mm_unpacklo_epi8(vx89ABCDEF, _mm_cmpgt_epi8(_mm_setzero_si128(), vx89ABCDEF)); in xnn_qs8_vadd_minmax_ukernel__sse2_mul16_ld64_x32()
51 vy89ABCDEF = _mm_unpacklo_epi8(vy89ABCDEF, _mm_cmpgt_epi8(_mm_setzero_si128(), vy89ABCDEF)); in xnn_qs8_vadd_minmax_ukernel__sse2_mul16_ld64_x32()
52 vxGHIJKLMN = _mm_unpacklo_epi8(vxGHIJKLMN, _mm_cmpgt_epi8(_mm_setzero_si128(), vxGHIJKLMN)); in xnn_qs8_vadd_minmax_ukernel__sse2_mul16_ld64_x32()
53 vyGHIJKLMN = _mm_unpacklo_epi8(vyGHIJKLMN, _mm_cmpgt_epi8(_mm_setzero_si128(), vyGHIJKLMN)); in xnn_qs8_vadd_minmax_ukernel__sse2_mul16_ld64_x32()
54 vxOPQRSTUV = _mm_unpacklo_epi8(vxOPQRSTUV, _mm_cmpgt_epi8(_mm_setzero_si128(), vxOPQRSTUV)); in xnn_qs8_vadd_minmax_ukernel__sse2_mul16_ld64_x32()
55 vyOPQRSTUV = _mm_unpacklo_epi8(vyOPQRSTUV, _mm_cmpgt_epi8(_mm_setzero_si128(), vyOPQRSTUV)); in xnn_qs8_vadd_minmax_ukernel__sse2_mul16_ld64_x32()
110 …add_epi32(_mm_and_si128(vacc0123, vremainder_mask), _mm_cmpgt_epi32(_mm_setzero_si128(), vacc0123)… in xnn_qs8_vadd_minmax_ukernel__sse2_mul16_ld64_x32()
111 …add_epi32(_mm_and_si128(vacc4567, vremainder_mask), _mm_cmpgt_epi32(_mm_setzero_si128(), vacc4567)… in xnn_qs8_vadd_minmax_ukernel__sse2_mul16_ld64_x32()
[all …]
|
| /external/XNNPACK/src/qs8-vaddc/gen/ |
| D | minmax-sse2-mul16-ld64-x32.c | 43 vx01234567 = _mm_unpacklo_epi8(vx01234567, _mm_cmpgt_epi8(_mm_setzero_si128(), vx01234567)); in xnn_qs8_vaddc_minmax_ukernel__sse2_mul16_ld64_x32()
44 vx89ABCDEF = _mm_unpacklo_epi8(vx89ABCDEF, _mm_cmpgt_epi8(_mm_setzero_si128(), vx89ABCDEF)); in xnn_qs8_vaddc_minmax_ukernel__sse2_mul16_ld64_x32()
45 vxGHIJKLMN = _mm_unpacklo_epi8(vxGHIJKLMN, _mm_cmpgt_epi8(_mm_setzero_si128(), vxGHIJKLMN)); in xnn_qs8_vaddc_minmax_ukernel__sse2_mul16_ld64_x32()
46 vxOPQRSTUV = _mm_unpacklo_epi8(vxOPQRSTUV, _mm_cmpgt_epi8(_mm_setzero_si128(), vxOPQRSTUV)); in xnn_qs8_vaddc_minmax_ukernel__sse2_mul16_ld64_x32()
76 …add_epi32(_mm_and_si128(vacc0123, vremainder_mask), _mm_cmpgt_epi32(_mm_setzero_si128(), vacc0123)… in xnn_qs8_vaddc_minmax_ukernel__sse2_mul16_ld64_x32()
77 …add_epi32(_mm_and_si128(vacc4567, vremainder_mask), _mm_cmpgt_epi32(_mm_setzero_si128(), vacc4567)… in xnn_qs8_vaddc_minmax_ukernel__sse2_mul16_ld64_x32()
78 …add_epi32(_mm_and_si128(vacc89AB, vremainder_mask), _mm_cmpgt_epi32(_mm_setzero_si128(), vacc89AB)… in xnn_qs8_vaddc_minmax_ukernel__sse2_mul16_ld64_x32()
79 …add_epi32(_mm_and_si128(vaccCDEF, vremainder_mask), _mm_cmpgt_epi32(_mm_setzero_si128(), vaccCDEF)… in xnn_qs8_vaddc_minmax_ukernel__sse2_mul16_ld64_x32()
80 …add_epi32(_mm_and_si128(vaccGHIJ, vremainder_mask), _mm_cmpgt_epi32(_mm_setzero_si128(), vaccGHIJ)… in xnn_qs8_vaddc_minmax_ukernel__sse2_mul16_ld64_x32()
81 …add_epi32(_mm_and_si128(vaccKLMN, vremainder_mask), _mm_cmpgt_epi32(_mm_setzero_si128(), vaccKLMN)… in xnn_qs8_vaddc_minmax_ukernel__sse2_mul16_ld64_x32()
[all …]
|
| /external/libaom/libaom/av1/encoder/x86/ |
| D | wedge_utils_sse2.c | 36 __m128i v_acc0_q = _mm_setzero_si128(); in av1_wedge_sse_from_residuals_sse2()
55 const __m128i v_m0_w = _mm_unpacklo_epi8(v_m01_b, _mm_setzero_si128()); in av1_wedge_sse_from_residuals_sse2()
56 const __m128i v_m1_w = _mm_unpackhi_epi8(v_m01_b, _mm_setzero_si128()); in av1_wedge_sse_from_residuals_sse2()
105 __m128i v_acc0_d = _mm_setzero_si128(); in av1_wedge_sign_from_residuals_sse2()
106 __m128i v_acc1_d = _mm_setzero_si128(); in av1_wedge_sign_from_residuals_sse2()
130 const __m128i v_m0_w = _mm_unpacklo_epi8(v_m01_b, _mm_setzero_si128()); in av1_wedge_sign_from_residuals_sse2()
131 const __m128i v_m1_w = _mm_unpackhi_epi8(v_m01_b, _mm_setzero_si128()); in av1_wedge_sign_from_residuals_sse2()
132 const __m128i v_m2_w = _mm_unpacklo_epi8(v_m23_b, _mm_setzero_si128()); in av1_wedge_sign_from_residuals_sse2()
133 const __m128i v_m3_w = _mm_unpackhi_epi8(v_m23_b, _mm_setzero_si128()); in av1_wedge_sign_from_residuals_sse2()
134 const __m128i v_m4_w = _mm_unpacklo_epi8(v_m45_b, _mm_setzero_si128()); in av1_wedge_sign_from_residuals_sse2()
[all …]
|
| /external/XNNPACK/src/qs8-igemm/gen/ |
| D | 4x4c2-minmax-ssse3-ld64.c | 88 const __m128i vxa0 = _mm_unpacklo_epi8(va0, _mm_cmpgt_epi8(_mm_setzero_si128(), va0)); in xnn_qs8_igemm_minmax_ukernel_4x4c2__ssse3_ld64()
91 const __m128i vxa1 = _mm_unpacklo_epi8(va1, _mm_cmpgt_epi8(_mm_setzero_si128(), va1)); in xnn_qs8_igemm_minmax_ukernel_4x4c2__ssse3_ld64()
94 const __m128i vxa2 = _mm_unpacklo_epi8(va2, _mm_cmpgt_epi8(_mm_setzero_si128(), va2)); in xnn_qs8_igemm_minmax_ukernel_4x4c2__ssse3_ld64()
97 const __m128i vxa3 = _mm_unpacklo_epi8(va3, _mm_cmpgt_epi8(_mm_setzero_si128(), va3)); in xnn_qs8_igemm_minmax_ukernel_4x4c2__ssse3_ld64()
101 const __m128i vxb0 = _mm_unpacklo_epi8(vb0, _mm_cmpgt_epi8(_mm_setzero_si128(), vb0)); in xnn_qs8_igemm_minmax_ukernel_4x4c2__ssse3_ld64()
112 const __m128i vxb1 = _mm_unpacklo_epi8(vb1, _mm_cmpgt_epi8(_mm_setzero_si128(), vb1)); in xnn_qs8_igemm_minmax_ukernel_4x4c2__ssse3_ld64()
123 const __m128i vxb2 = _mm_unpacklo_epi8(vb2, _mm_cmpgt_epi8(_mm_setzero_si128(), vb2)); in xnn_qs8_igemm_minmax_ukernel_4x4c2__ssse3_ld64()
134 const __m128i vxb3 = _mm_unpacklo_epi8(vb3, _mm_cmpgt_epi8(_mm_setzero_si128(), vb3)); in xnn_qs8_igemm_minmax_ukernel_4x4c2__ssse3_ld64()
150 const __m128i vxa0 = _mm_unpacklo_epi8(va0, _mm_cmpgt_epi8(_mm_setzero_si128(), va0)); in xnn_qs8_igemm_minmax_ukernel_4x4c2__ssse3_ld64()
153 const __m128i vxa1 = _mm_unpacklo_epi8(va1, _mm_cmpgt_epi8(_mm_setzero_si128(), va1)); in xnn_qs8_igemm_minmax_ukernel_4x4c2__ssse3_ld64()
[all …]
|
| /external/XNNPACK/src/qs8-gemm/gen/ |
| D | 4x4c2-minmax-sse2-ld64.c | 71 const __m128i vxa0 = _mm_unpacklo_epi8(va0, _mm_cmpgt_epi8(_mm_setzero_si128(), va0)); in xnn_qs8_gemm_minmax_ukernel_4x4c2__sse2_ld64()
74 const __m128i vxa1 = _mm_unpacklo_epi8(va1, _mm_cmpgt_epi8(_mm_setzero_si128(), va1)); in xnn_qs8_gemm_minmax_ukernel_4x4c2__sse2_ld64()
77 const __m128i vxa2 = _mm_unpacklo_epi8(va2, _mm_cmpgt_epi8(_mm_setzero_si128(), va2)); in xnn_qs8_gemm_minmax_ukernel_4x4c2__sse2_ld64()
80 const __m128i vxa3 = _mm_unpacklo_epi8(va3, _mm_cmpgt_epi8(_mm_setzero_si128(), va3)); in xnn_qs8_gemm_minmax_ukernel_4x4c2__sse2_ld64()
84 const __m128i vxb0 = _mm_unpacklo_epi8(vb0, _mm_cmpgt_epi8(_mm_setzero_si128(), vb0)); in xnn_qs8_gemm_minmax_ukernel_4x4c2__sse2_ld64()
95 const __m128i vxb1 = _mm_unpacklo_epi8(vb1, _mm_cmpgt_epi8(_mm_setzero_si128(), vb1)); in xnn_qs8_gemm_minmax_ukernel_4x4c2__sse2_ld64()
106 const __m128i vxb2 = _mm_unpacklo_epi8(vb2, _mm_cmpgt_epi8(_mm_setzero_si128(), vb2)); in xnn_qs8_gemm_minmax_ukernel_4x4c2__sse2_ld64()
117 const __m128i vxb3 = _mm_unpacklo_epi8(vb3, _mm_cmpgt_epi8(_mm_setzero_si128(), vb3)); in xnn_qs8_gemm_minmax_ukernel_4x4c2__sse2_ld64()
133 const __m128i vxa0 = _mm_unpacklo_epi8(va0, _mm_cmpgt_epi8(_mm_setzero_si128(), va0)); in xnn_qs8_gemm_minmax_ukernel_4x4c2__sse2_ld64()
136 const __m128i vxa1 = _mm_unpacklo_epi8(va1, _mm_cmpgt_epi8(_mm_setzero_si128(), va1)); in xnn_qs8_gemm_minmax_ukernel_4x4c2__sse2_ld64()
[all …]
|