| /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-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 | 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-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()
113 …const __m128i vxi0x01234567 = _mm_unpacklo_epi8(vi0x01234567, _mm_cmpgt_epi8(_mm_setzero_si128(), … in xnn_qs8_gavgpool_minmax_ukernel_7p7x__sse2_c8_acc2()
114 …const __m128i vxi1x01234567 = _mm_unpacklo_epi8(vi1x01234567, _mm_cmpgt_epi8(_mm_setzero_si128(), … in xnn_qs8_gavgpool_minmax_ukernel_7p7x__sse2_c8_acc2()
115 …const __m128i vxi2x01234567 = _mm_unpacklo_epi8(vi2x01234567, _mm_cmpgt_epi8(_mm_setzero_si128(), … 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()
113 …const __m128i vxi0x01234567 = _mm_unpacklo_epi8(vi0x01234567, _mm_cmpgt_epi8(_mm_setzero_si128(), … in xnn_qs8_gavgpool_minmax_ukernel_7p7x__ssse3_c8_acc2()
114 …const __m128i vxi1x01234567 = _mm_unpacklo_epi8(vi1x01234567, _mm_cmpgt_epi8(_mm_setzero_si128(), … in xnn_qs8_gavgpool_minmax_ukernel_7p7x__ssse3_c8_acc2()
115 …const __m128i vxi2x01234567 = _mm_unpacklo_epi8(vi2x01234567, _mm_cmpgt_epi8(_mm_setzero_si128(), … in xnn_qs8_gavgpool_minmax_ukernel_7p7x__ssse3_c8_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-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 …]
|
| /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()
143 srcReg23_lo = _mm_unpacklo_epi8(srcReg2, srcReg3); in aom_filter_block1d16_v4_sse2()
145 __m128i resReg23_lo_1 = _mm_unpacklo_epi8(srcReg23_lo, _mm_setzero_si128()); in aom_filter_block1d16_v4_sse2()
[all …]
|
| D | avg_intrin_sse2.c | 24 s0 = _mm_unpacklo_epi8(_mm_loadl_epi64((const __m128i *)(s)), u0); in aom_minmax_8x8_sse2()
25 d0 = _mm_unpacklo_epi8(_mm_loadl_epi64((const __m128i *)(d)), u0); in aom_minmax_8x8_sse2()
30 s0 = _mm_unpacklo_epi8(_mm_loadl_epi64((const __m128i *)(s + p)), u0); in aom_minmax_8x8_sse2()
31 d0 = _mm_unpacklo_epi8(_mm_loadl_epi64((const __m128i *)(d + dp)), u0); in aom_minmax_8x8_sse2()
38 s0 = _mm_unpacklo_epi8(_mm_loadl_epi64((const __m128i *)(s + 2 * p)), u0); in aom_minmax_8x8_sse2()
39 d0 = _mm_unpacklo_epi8(_mm_loadl_epi64((const __m128i *)(d + 2 * dp)), u0); in aom_minmax_8x8_sse2()
46 s0 = _mm_unpacklo_epi8(_mm_loadl_epi64((const __m128i *)(s + 3 * p)), u0); in aom_minmax_8x8_sse2()
47 d0 = _mm_unpacklo_epi8(_mm_loadl_epi64((const __m128i *)(d + 3 * dp)), u0); in aom_minmax_8x8_sse2()
54 s0 = _mm_unpacklo_epi8(_mm_loadl_epi64((const __m128i *)(s + 4 * p)), u0); in aom_minmax_8x8_sse2()
55 d0 = _mm_unpacklo_epi8(_mm_loadl_epi64((const __m128i *)(d + 4 * dp)), u0); in aom_minmax_8x8_sse2()
[all …]
|
| /external/libaom/libaom/av1/common/x86/ |
| D | convolve_sse2.c | 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()
104 s[0] = _mm_unpacklo_epi8( in av1_convolve_y_sr_sse2()
107 s[1] = _mm_unpacklo_epi8( in av1_convolve_y_sr_sse2()
[all …]
|
| /external/XNNPACK/src/qs8-dwconv/gen/ |
| D | up16x9-minmax-ssse3-mul16.c | 95 …const __m128i vxi0x01234567 = _mm_unpacklo_epi8(vi0x01234567, _mm_cmpgt_epi8(_mm_setzero_si128(), … in xnn_qs8_dwconv_minmax_ukernel_up16x9__ssse3_mul16()
96 …const __m128i vxk0x01234567 = _mm_unpacklo_epi8(vk0x01234567, _mm_cmpgt_epi8(_mm_setzero_si128(), … in xnn_qs8_dwconv_minmax_ukernel_up16x9__ssse3_mul16()
97 …const __m128i vxi0x89ABCDEF = _mm_unpacklo_epi8(vi0x89ABCDEF, _mm_cmpgt_epi8(_mm_setzero_si128(), … in xnn_qs8_dwconv_minmax_ukernel_up16x9__ssse3_mul16()
98 …const __m128i vxk0x89ABCDEF = _mm_unpacklo_epi8(vk0x89ABCDEF, _mm_cmpgt_epi8(_mm_setzero_si128(), … in xnn_qs8_dwconv_minmax_ukernel_up16x9__ssse3_mul16()
116 …const __m128i vxi1x01234567 = _mm_unpacklo_epi8(vi1x01234567, _mm_cmpgt_epi8(_mm_setzero_si128(), … in xnn_qs8_dwconv_minmax_ukernel_up16x9__ssse3_mul16()
117 …const __m128i vxk1x01234567 = _mm_unpacklo_epi8(vk1x01234567, _mm_cmpgt_epi8(_mm_setzero_si128(), … in xnn_qs8_dwconv_minmax_ukernel_up16x9__ssse3_mul16()
118 …const __m128i vxi1x89ABCDEF = _mm_unpacklo_epi8(vi1x89ABCDEF, _mm_cmpgt_epi8(_mm_setzero_si128(), … in xnn_qs8_dwconv_minmax_ukernel_up16x9__ssse3_mul16()
119 …const __m128i vxk1x89ABCDEF = _mm_unpacklo_epi8(vk1x89ABCDEF, _mm_cmpgt_epi8(_mm_setzero_si128(), … in xnn_qs8_dwconv_minmax_ukernel_up16x9__ssse3_mul16()
137 …const __m128i vxi2x01234567 = _mm_unpacklo_epi8(vi2x01234567, _mm_cmpgt_epi8(_mm_setzero_si128(), … in xnn_qs8_dwconv_minmax_ukernel_up16x9__ssse3_mul16()
138 …const __m128i vxk2x01234567 = _mm_unpacklo_epi8(vk2x01234567, _mm_cmpgt_epi8(_mm_setzero_si128(), … in xnn_qs8_dwconv_minmax_ukernel_up16x9__ssse3_mul16()
[all …]
|
| D | up16x9-minmax-sse2-mul16.c | 95 …const __m128i vxi0x01234567 = _mm_unpacklo_epi8(vi0x01234567, _mm_cmpgt_epi8(_mm_setzero_si128(), … in xnn_qs8_dwconv_minmax_ukernel_up16x9__sse2_mul16()
96 …const __m128i vxk0x01234567 = _mm_unpacklo_epi8(vk0x01234567, _mm_cmpgt_epi8(_mm_setzero_si128(), … in xnn_qs8_dwconv_minmax_ukernel_up16x9__sse2_mul16()
97 …const __m128i vxi0x89ABCDEF = _mm_unpacklo_epi8(vi0x89ABCDEF, _mm_cmpgt_epi8(_mm_setzero_si128(), … in xnn_qs8_dwconv_minmax_ukernel_up16x9__sse2_mul16()
98 …const __m128i vxk0x89ABCDEF = _mm_unpacklo_epi8(vk0x89ABCDEF, _mm_cmpgt_epi8(_mm_setzero_si128(), … in xnn_qs8_dwconv_minmax_ukernel_up16x9__sse2_mul16()
116 …const __m128i vxi1x01234567 = _mm_unpacklo_epi8(vi1x01234567, _mm_cmpgt_epi8(_mm_setzero_si128(), … in xnn_qs8_dwconv_minmax_ukernel_up16x9__sse2_mul16()
117 …const __m128i vxk1x01234567 = _mm_unpacklo_epi8(vk1x01234567, _mm_cmpgt_epi8(_mm_setzero_si128(), … in xnn_qs8_dwconv_minmax_ukernel_up16x9__sse2_mul16()
118 …const __m128i vxi1x89ABCDEF = _mm_unpacklo_epi8(vi1x89ABCDEF, _mm_cmpgt_epi8(_mm_setzero_si128(), … in xnn_qs8_dwconv_minmax_ukernel_up16x9__sse2_mul16()
119 …const __m128i vxk1x89ABCDEF = _mm_unpacklo_epi8(vk1x89ABCDEF, _mm_cmpgt_epi8(_mm_setzero_si128(), … in xnn_qs8_dwconv_minmax_ukernel_up16x9__sse2_mul16()
137 …const __m128i vxi2x01234567 = _mm_unpacklo_epi8(vi2x01234567, _mm_cmpgt_epi8(_mm_setzero_si128(), … in xnn_qs8_dwconv_minmax_ukernel_up16x9__sse2_mul16()
138 …const __m128i vxk2x01234567 = _mm_unpacklo_epi8(vk2x01234567, _mm_cmpgt_epi8(_mm_setzero_si128(), … in xnn_qs8_dwconv_minmax_ukernel_up16x9__sse2_mul16()
[all …]
|
| D | up24x9-minmax-ssse3-mul16.c | 99 …const __m128i vxi0x01234567 = _mm_unpacklo_epi8(vi0x01234567, _mm_cmpgt_epi8(_mm_setzero_si128(), … in xnn_qs8_dwconv_minmax_ukernel_up24x9__ssse3_mul16()
100 …const __m128i vxk0x01234567 = _mm_unpacklo_epi8(vk0x01234567, _mm_cmpgt_epi8(_mm_setzero_si128(), … in xnn_qs8_dwconv_minmax_ukernel_up24x9__ssse3_mul16()
101 …const __m128i vxi0x89ABCDEF = _mm_unpacklo_epi8(vi0x89ABCDEF, _mm_cmpgt_epi8(_mm_setzero_si128(), … in xnn_qs8_dwconv_minmax_ukernel_up24x9__ssse3_mul16()
102 …const __m128i vxk0x89ABCDEF = _mm_unpacklo_epi8(vk0x89ABCDEF, _mm_cmpgt_epi8(_mm_setzero_si128(), … in xnn_qs8_dwconv_minmax_ukernel_up24x9__ssse3_mul16()
103 …const __m128i vxi0xGHIJKLMN = _mm_unpacklo_epi8(vi0xGHIJKLMN, _mm_cmpgt_epi8(_mm_setzero_si128(), … in xnn_qs8_dwconv_minmax_ukernel_up24x9__ssse3_mul16()
104 …const __m128i vxk0xGHIJKLMN = _mm_unpacklo_epi8(vk0xGHIJKLMN, _mm_cmpgt_epi8(_mm_setzero_si128(), … in xnn_qs8_dwconv_minmax_ukernel_up24x9__ssse3_mul16()
128 …const __m128i vxi1x01234567 = _mm_unpacklo_epi8(vi1x01234567, _mm_cmpgt_epi8(_mm_setzero_si128(), … in xnn_qs8_dwconv_minmax_ukernel_up24x9__ssse3_mul16()
129 …const __m128i vxk1x01234567 = _mm_unpacklo_epi8(vk1x01234567, _mm_cmpgt_epi8(_mm_setzero_si128(), … in xnn_qs8_dwconv_minmax_ukernel_up24x9__ssse3_mul16()
130 …const __m128i vxi1x89ABCDEF = _mm_unpacklo_epi8(vi1x89ABCDEF, _mm_cmpgt_epi8(_mm_setzero_si128(), … in xnn_qs8_dwconv_minmax_ukernel_up24x9__ssse3_mul16()
131 …const __m128i vxk1x89ABCDEF = _mm_unpacklo_epi8(vk1x89ABCDEF, _mm_cmpgt_epi8(_mm_setzero_si128(), … in xnn_qs8_dwconv_minmax_ukernel_up24x9__ssse3_mul16()
[all …]
|
| D | up24x9-minmax-sse2-mul16.c | 99 …const __m128i vxi0x01234567 = _mm_unpacklo_epi8(vi0x01234567, _mm_cmpgt_epi8(_mm_setzero_si128(), … in xnn_qs8_dwconv_minmax_ukernel_up24x9__sse2_mul16()
100 …const __m128i vxk0x01234567 = _mm_unpacklo_epi8(vk0x01234567, _mm_cmpgt_epi8(_mm_setzero_si128(), … in xnn_qs8_dwconv_minmax_ukernel_up24x9__sse2_mul16()
101 …const __m128i vxi0x89ABCDEF = _mm_unpacklo_epi8(vi0x89ABCDEF, _mm_cmpgt_epi8(_mm_setzero_si128(), … in xnn_qs8_dwconv_minmax_ukernel_up24x9__sse2_mul16()
102 …const __m128i vxk0x89ABCDEF = _mm_unpacklo_epi8(vk0x89ABCDEF, _mm_cmpgt_epi8(_mm_setzero_si128(), … in xnn_qs8_dwconv_minmax_ukernel_up24x9__sse2_mul16()
103 …const __m128i vxi0xGHIJKLMN = _mm_unpacklo_epi8(vi0xGHIJKLMN, _mm_cmpgt_epi8(_mm_setzero_si128(), … in xnn_qs8_dwconv_minmax_ukernel_up24x9__sse2_mul16()
104 …const __m128i vxk0xGHIJKLMN = _mm_unpacklo_epi8(vk0xGHIJKLMN, _mm_cmpgt_epi8(_mm_setzero_si128(), … in xnn_qs8_dwconv_minmax_ukernel_up24x9__sse2_mul16()
128 …const __m128i vxi1x01234567 = _mm_unpacklo_epi8(vi1x01234567, _mm_cmpgt_epi8(_mm_setzero_si128(), … in xnn_qs8_dwconv_minmax_ukernel_up24x9__sse2_mul16()
129 …const __m128i vxk1x01234567 = _mm_unpacklo_epi8(vk1x01234567, _mm_cmpgt_epi8(_mm_setzero_si128(), … in xnn_qs8_dwconv_minmax_ukernel_up24x9__sse2_mul16()
130 …const __m128i vxi1x89ABCDEF = _mm_unpacklo_epi8(vi1x89ABCDEF, _mm_cmpgt_epi8(_mm_setzero_si128(), … in xnn_qs8_dwconv_minmax_ukernel_up24x9__sse2_mul16()
131 …const __m128i vxk1x89ABCDEF = _mm_unpacklo_epi8(vk1x89ABCDEF, _mm_cmpgt_epi8(_mm_setzero_si128(), … in xnn_qs8_dwconv_minmax_ukernel_up24x9__sse2_mul16()
[all …]
|
| D | up8x9-minmax-sse2-mul16.c | 91 …const __m128i vxi0x01234567 = _mm_unpacklo_epi8(vi0x01234567, _mm_cmpgt_epi8(_mm_setzero_si128(), … in xnn_qs8_dwconv_minmax_ukernel_up8x9__sse2_mul16()
92 …const __m128i vxk0x01234567 = _mm_unpacklo_epi8(vk0x01234567, _mm_cmpgt_epi8(_mm_setzero_si128(), … in xnn_qs8_dwconv_minmax_ukernel_up8x9__sse2_mul16()
104 …const __m128i vxi1x01234567 = _mm_unpacklo_epi8(vi1x01234567, _mm_cmpgt_epi8(_mm_setzero_si128(), … in xnn_qs8_dwconv_minmax_ukernel_up8x9__sse2_mul16()
105 …const __m128i vxk1x01234567 = _mm_unpacklo_epi8(vk1x01234567, _mm_cmpgt_epi8(_mm_setzero_si128(), … in xnn_qs8_dwconv_minmax_ukernel_up8x9__sse2_mul16()
117 …const __m128i vxi2x01234567 = _mm_unpacklo_epi8(vi2x01234567, _mm_cmpgt_epi8(_mm_setzero_si128(), … in xnn_qs8_dwconv_minmax_ukernel_up8x9__sse2_mul16()
118 …const __m128i vxk2x01234567 = _mm_unpacklo_epi8(vk2x01234567, _mm_cmpgt_epi8(_mm_setzero_si128(), … in xnn_qs8_dwconv_minmax_ukernel_up8x9__sse2_mul16()
130 …const __m128i vxi3x01234567 = _mm_unpacklo_epi8(vi3x01234567, _mm_cmpgt_epi8(_mm_setzero_si128(), … in xnn_qs8_dwconv_minmax_ukernel_up8x9__sse2_mul16()
131 …const __m128i vxk3x01234567 = _mm_unpacklo_epi8(vk3x01234567, _mm_cmpgt_epi8(_mm_setzero_si128(), … in xnn_qs8_dwconv_minmax_ukernel_up8x9__sse2_mul16()
143 …const __m128i vxi4x01234567 = _mm_unpacklo_epi8(vi4x01234567, _mm_cmpgt_epi8(_mm_setzero_si128(), … in xnn_qs8_dwconv_minmax_ukernel_up8x9__sse2_mul16()
144 …const __m128i vxk4x01234567 = _mm_unpacklo_epi8(vk4x01234567, _mm_cmpgt_epi8(_mm_setzero_si128(), … in xnn_qs8_dwconv_minmax_ukernel_up8x9__sse2_mul16()
[all …]
|
| D | up8x9-minmax-ssse3-mul16.c | 91 …const __m128i vxi0x01234567 = _mm_unpacklo_epi8(vi0x01234567, _mm_cmpgt_epi8(_mm_setzero_si128(), … in xnn_qs8_dwconv_minmax_ukernel_up8x9__ssse3_mul16()
92 …const __m128i vxk0x01234567 = _mm_unpacklo_epi8(vk0x01234567, _mm_cmpgt_epi8(_mm_setzero_si128(), … in xnn_qs8_dwconv_minmax_ukernel_up8x9__ssse3_mul16()
104 …const __m128i vxi1x01234567 = _mm_unpacklo_epi8(vi1x01234567, _mm_cmpgt_epi8(_mm_setzero_si128(), … in xnn_qs8_dwconv_minmax_ukernel_up8x9__ssse3_mul16()
105 …const __m128i vxk1x01234567 = _mm_unpacklo_epi8(vk1x01234567, _mm_cmpgt_epi8(_mm_setzero_si128(), … in xnn_qs8_dwconv_minmax_ukernel_up8x9__ssse3_mul16()
117 …const __m128i vxi2x01234567 = _mm_unpacklo_epi8(vi2x01234567, _mm_cmpgt_epi8(_mm_setzero_si128(), … in xnn_qs8_dwconv_minmax_ukernel_up8x9__ssse3_mul16()
118 …const __m128i vxk2x01234567 = _mm_unpacklo_epi8(vk2x01234567, _mm_cmpgt_epi8(_mm_setzero_si128(), … in xnn_qs8_dwconv_minmax_ukernel_up8x9__ssse3_mul16()
130 …const __m128i vxi3x01234567 = _mm_unpacklo_epi8(vi3x01234567, _mm_cmpgt_epi8(_mm_setzero_si128(), … in xnn_qs8_dwconv_minmax_ukernel_up8x9__ssse3_mul16()
131 …const __m128i vxk3x01234567 = _mm_unpacklo_epi8(vk3x01234567, _mm_cmpgt_epi8(_mm_setzero_si128(), … in xnn_qs8_dwconv_minmax_ukernel_up8x9__ssse3_mul16()
143 …const __m128i vxi4x01234567 = _mm_unpacklo_epi8(vi4x01234567, _mm_cmpgt_epi8(_mm_setzero_si128(), … in xnn_qs8_dwconv_minmax_ukernel_up8x9__ssse3_mul16()
144 …const __m128i vxk4x01234567 = _mm_unpacklo_epi8(vk4x01234567, _mm_cmpgt_epi8(_mm_setzero_si128(), … in xnn_qs8_dwconv_minmax_ukernel_up8x9__ssse3_mul16()
[all …]
|
| /external/libavc/common/x86/ |
| D | ih264_deblk_luma_ssse3.c | 122 temp1 = _mm_unpacklo_epi8(line1, line2); in ih264_deblk_luma_vert_bs4_ssse3()
123 temp2 = _mm_unpacklo_epi8(line3, line4); in ih264_deblk_luma_vert_bs4_ssse3()
124 temp3 = _mm_unpacklo_epi8(line5, line6); in ih264_deblk_luma_vert_bs4_ssse3()
125 temp4 = _mm_unpacklo_epi8(line7, line8); in ih264_deblk_luma_vert_bs4_ssse3()
146 temp1 = _mm_unpacklo_epi8(line1, line2); in ih264_deblk_luma_vert_bs4_ssse3()
147 temp2 = _mm_unpacklo_epi8(line3, line4); in ih264_deblk_luma_vert_bs4_ssse3()
148 temp3 = _mm_unpacklo_epi8(line5, line6); in ih264_deblk_luma_vert_bs4_ssse3()
149 temp4 = _mm_unpacklo_epi8(line7, line8); in ih264_deblk_luma_vert_bs4_ssse3()
175 temp2 = _mm_unpacklo_epi8(temp1, zero); in ih264_deblk_luma_vert_bs4_ssse3()
188 temp2 = _mm_unpacklo_epi8(temp1, zero); in ih264_deblk_luma_vert_bs4_ssse3()
[all …]
|
| /external/XNNPACK/src/qu8-gavgpool/ |
| D | 7p7x-minmax-sse2-c8.c | 52 const __m128i vxi0 = _mm_unpacklo_epi8(vi0, vzero); in xnn_qu8_gavgpool_minmax_ukernel_7p7x__sse2_c8()
53 const __m128i vxi1 = _mm_unpacklo_epi8(vi1, vzero); in xnn_qu8_gavgpool_minmax_ukernel_7p7x__sse2_c8()
54 const __m128i vxi2 = _mm_unpacklo_epi8(vi2, vzero); in xnn_qu8_gavgpool_minmax_ukernel_7p7x__sse2_c8()
55 const __m128i vxi3 = _mm_unpacklo_epi8(vi3, vzero); in xnn_qu8_gavgpool_minmax_ukernel_7p7x__sse2_c8()
56 const __m128i vxi4 = _mm_unpacklo_epi8(vi4, vzero); in xnn_qu8_gavgpool_minmax_ukernel_7p7x__sse2_c8()
57 const __m128i vxi5 = _mm_unpacklo_epi8(vi5, vzero); in xnn_qu8_gavgpool_minmax_ukernel_7p7x__sse2_c8()
58 const __m128i vxi6 = _mm_unpacklo_epi8(vi6, vzero); in xnn_qu8_gavgpool_minmax_ukernel_7p7x__sse2_c8()
96 const __m128i vxi0 = _mm_unpacklo_epi8(vi0, vzero); in xnn_qu8_gavgpool_minmax_ukernel_7p7x__sse2_c8()
97 const __m128i vxi1 = _mm_unpacklo_epi8(vi1, vzero); in xnn_qu8_gavgpool_minmax_ukernel_7p7x__sse2_c8()
98 const __m128i vxi2 = _mm_unpacklo_epi8(vi2, vzero); in xnn_qu8_gavgpool_minmax_ukernel_7p7x__sse2_c8()
[all …]
|
| /external/XNNPACK/src/qu8-dwconv/ |
| D | up8x9-minmax-sse2.c | 76 const __m128i vxi0 = _mm_unpacklo_epi8(vi0, vzero); in xnn_qu8_dwconv_minmax_ukernel_up8x9__sse2()
78 const __m128i vxk0 = _mm_sub_epi16(_mm_unpacklo_epi8(vk0, vzero), vkernel_zero_point); in xnn_qu8_dwconv_minmax_ukernel_up8x9__sse2()
85 const __m128i vxi1 = _mm_unpacklo_epi8(vi1, vzero); in xnn_qu8_dwconv_minmax_ukernel_up8x9__sse2()
87 const __m128i vxk1 = _mm_sub_epi16(_mm_unpacklo_epi8(vk1, vzero), vkernel_zero_point); in xnn_qu8_dwconv_minmax_ukernel_up8x9__sse2()
94 const __m128i vxi2 = _mm_unpacklo_epi8(vi2, vzero); in xnn_qu8_dwconv_minmax_ukernel_up8x9__sse2()
96 const __m128i vxk2 = _mm_sub_epi16(_mm_unpacklo_epi8(vk2, vzero), vkernel_zero_point); in xnn_qu8_dwconv_minmax_ukernel_up8x9__sse2()
103 const __m128i vxi3 = _mm_unpacklo_epi8(vi3, vzero); in xnn_qu8_dwconv_minmax_ukernel_up8x9__sse2()
105 const __m128i vxk3 = _mm_sub_epi16(_mm_unpacklo_epi8(vk3, vzero), vkernel_zero_point); in xnn_qu8_dwconv_minmax_ukernel_up8x9__sse2()
112 const __m128i vxi4 = _mm_unpacklo_epi8(vi4, vzero); in xnn_qu8_dwconv_minmax_ukernel_up8x9__sse2()
114 const __m128i vxk4 = _mm_sub_epi16(_mm_unpacklo_epi8(vk4, vzero), vkernel_zero_point); in xnn_qu8_dwconv_minmax_ukernel_up8x9__sse2()
[all …]
|
| /external/XNNPACK/src/qu8-avgpool/ |
| D | 9p8x-minmax-sse2-c8.c | 99 const __m128i vxi0 = _mm_unpacklo_epi8(vi0, vzero); in xnn_qu8_avgpool_minmax_ukernel_9p8x__sse2_c8()
100 const __m128i vxi1 = _mm_unpacklo_epi8(vi1, vzero); in xnn_qu8_avgpool_minmax_ukernel_9p8x__sse2_c8()
101 const __m128i vxi2 = _mm_unpacklo_epi8(vi2, vzero); in xnn_qu8_avgpool_minmax_ukernel_9p8x__sse2_c8()
102 const __m128i vxi3 = _mm_unpacklo_epi8(vi3, vzero); in xnn_qu8_avgpool_minmax_ukernel_9p8x__sse2_c8()
103 const __m128i vxi4 = _mm_unpacklo_epi8(vi4, vzero); in xnn_qu8_avgpool_minmax_ukernel_9p8x__sse2_c8()
104 const __m128i vxi5 = _mm_unpacklo_epi8(vi5, vzero); in xnn_qu8_avgpool_minmax_ukernel_9p8x__sse2_c8()
105 const __m128i vxi6 = _mm_unpacklo_epi8(vi6, vzero); in xnn_qu8_avgpool_minmax_ukernel_9p8x__sse2_c8()
106 const __m128i vxi7 = _mm_unpacklo_epi8(vi7, vzero); in xnn_qu8_avgpool_minmax_ukernel_9p8x__sse2_c8()
107 const __m128i vxi8 = _mm_unpacklo_epi8(vi8, vzero); in xnn_qu8_avgpool_minmax_ukernel_9p8x__sse2_c8()
183 const __m128i vxi0 = _mm_unpacklo_epi8(vi0, vzero); in xnn_qu8_avgpool_minmax_ukernel_9p8x__sse2_c8()
[all …]
|
| /external/libhevc/common/x86/ |
| D | ihevc_inter_pred_filters_ssse3_intr.c | 335 … src_temp3_16x8b = _mm_unpacklo_epi8(src_temp1_16x8b, src_temp2_16x8b); /* row = 0 */ in ihevc_inter_pred_luma_horz_ssse3()
341 … src_temp4_16x8b = _mm_unpacklo_epi8(src_temp1_16x8b, src_temp2_16x8b); /* row = 0 */ in ihevc_inter_pred_luma_horz_ssse3()
347 … src_temp5_16x8b = _mm_unpacklo_epi8(src_temp1_16x8b, src_temp2_16x8b); /* row = 0 */ in ihevc_inter_pred_luma_horz_ssse3()
353 … src_temp6_16x8b = _mm_unpacklo_epi8(src_temp1_16x8b, src_temp2_16x8b); /* row = 0 */ in ihevc_inter_pred_luma_horz_ssse3()
368 … src_temp13_16x8b = _mm_unpacklo_epi8(src_temp11_16x8b, src_temp12_16x8b); /* row = 1 */ in ihevc_inter_pred_luma_horz_ssse3()
374 … src_temp14_16x8b = _mm_unpacklo_epi8(src_temp11_16x8b, src_temp12_16x8b); /* row = 1 */ in ihevc_inter_pred_luma_horz_ssse3()
380 … src_temp15_16x8b = _mm_unpacklo_epi8(src_temp11_16x8b, src_temp12_16x8b); /* row = 1 */ in ihevc_inter_pred_luma_horz_ssse3()
386 … src_temp16_16x8b = _mm_unpacklo_epi8(src_temp11_16x8b, src_temp12_16x8b); /* row = 1 */ in ihevc_inter_pred_luma_horz_ssse3()
424 … src_temp3_16x8b = _mm_unpacklo_epi8(src_temp1_16x8b, src_temp2_16x8b); /* row = 0 */ in ihevc_inter_pred_luma_horz_ssse3()
430 … src_temp4_16x8b = _mm_unpacklo_epi8(src_temp1_16x8b, src_temp2_16x8b); /* row = 0 */ in ihevc_inter_pred_luma_horz_ssse3()
[all …]
|
| D | ihevc_chroma_intra_pred_filters_ssse3_intr.c | 219 src_temp_8x16b = _mm_unpacklo_epi8(src_temp_8x16b, zero_8x16b); in ihevc_intra_pred_chroma_planar_ssse3()
336 src_temp5 = _mm_unpacklo_epi8(src_temp3, m_zero); in ihevc_intra_pred_chroma_dc_ssse3()
337 src_temp6 = _mm_unpacklo_epi8(src_temp4, m_zero); in ihevc_intra_pred_chroma_dc_ssse3()
338 src_temp9 = _mm_unpacklo_epi8(src_temp7, m_zero); in ihevc_intra_pred_chroma_dc_ssse3()
339 src_temp10 = _mm_unpacklo_epi8(src_temp8, m_zero); in ihevc_intra_pred_chroma_dc_ssse3()
346 src_temp3 = _mm_unpacklo_epi8(src_temp3, m_zero); in ihevc_intra_pred_chroma_dc_ssse3()
347 src_temp4 = _mm_unpacklo_epi8(src_temp4, m_zero); in ihevc_intra_pred_chroma_dc_ssse3()
348 src_temp7 = _mm_unpacklo_epi8(src_temp7, m_zero); in ihevc_intra_pred_chroma_dc_ssse3()
349 src_temp8 = _mm_unpacklo_epi8(src_temp8, m_zero); in ihevc_intra_pred_chroma_dc_ssse3()
378 src_temp5 = _mm_unpacklo_epi8(src_temp3, m_zero); in ihevc_intra_pred_chroma_dc_ssse3()
[all …]
|
| /external/libhevc/decoder/x86/ |
| D | ihevcd_it_rec_dc_ssse3_intr.c | 106 m_temp_reg_4 = _mm_unpacklo_epi8(m_temp_reg_4, zero_8x16b); in ihevcd_itrans_recon_dc_luma_ssse3()
107 m_temp_reg_5 = _mm_unpacklo_epi8(m_temp_reg_5, zero_8x16b); in ihevcd_itrans_recon_dc_luma_ssse3()
147 m_temp_reg_4 = _mm_unpacklo_epi8(m_temp_reg_0, zero_8x16b); in ihevcd_itrans_recon_dc_luma_ssse3()
148 m_temp_reg_5 = _mm_unpacklo_epi8(m_temp_reg_1, zero_8x16b); in ihevcd_itrans_recon_dc_luma_ssse3()
149 m_temp_reg_6 = _mm_unpacklo_epi8(m_temp_reg_2, zero_8x16b); in ihevcd_itrans_recon_dc_luma_ssse3()
150 m_temp_reg_7 = _mm_unpacklo_epi8(m_temp_reg_3, zero_8x16b); in ihevcd_itrans_recon_dc_luma_ssse3()
257 m_temp_reg_4 = _mm_unpacklo_epi8(m_temp_reg_14, zero_8x16b); in ihevcd_itrans_recon_dc_chroma_ssse3()
258 m_temp_reg_5 = _mm_unpacklo_epi8(m_temp_reg_15, zero_8x16b); in ihevcd_itrans_recon_dc_chroma_ssse3()
276 m_temp_reg_9 = _mm_unpacklo_epi8(m_temp_reg_8, m_temp_reg_0); in ihevcd_itrans_recon_dc_chroma_ssse3()
278 m_temp_reg_10 = _mm_unpacklo_epi8(m_temp_reg_8, m_temp_reg_1); in ihevcd_itrans_recon_dc_chroma_ssse3()
[all …]
|
| D | ihevcd_it_rec_dc_sse42_intr.c | 105 m_temp_reg_4 = _mm_unpacklo_epi8(m_temp_reg_4, zero_8x16b); in ihevcd_itrans_recon_dc_luma_sse42()
106 m_temp_reg_5 = _mm_unpacklo_epi8(m_temp_reg_5, zero_8x16b); in ihevcd_itrans_recon_dc_luma_sse42()
146 m_temp_reg_4 = _mm_unpacklo_epi8(m_temp_reg_0, zero_8x16b); in ihevcd_itrans_recon_dc_luma_sse42()
147 m_temp_reg_5 = _mm_unpacklo_epi8(m_temp_reg_1, zero_8x16b); in ihevcd_itrans_recon_dc_luma_sse42()
148 m_temp_reg_6 = _mm_unpacklo_epi8(m_temp_reg_2, zero_8x16b); in ihevcd_itrans_recon_dc_luma_sse42()
149 m_temp_reg_7 = _mm_unpacklo_epi8(m_temp_reg_3, zero_8x16b); in ihevcd_itrans_recon_dc_luma_sse42()
257 m_temp_reg_4 = _mm_unpacklo_epi8(m_temp_reg_14, zero_8x16b); in ihevcd_itrans_recon_dc_chroma_sse42()
258 m_temp_reg_5 = _mm_unpacklo_epi8(m_temp_reg_15, zero_8x16b); in ihevcd_itrans_recon_dc_chroma_sse42()
276 m_temp_reg_9 = _mm_unpacklo_epi8(m_temp_reg_8, m_temp_reg_0); in ihevcd_itrans_recon_dc_chroma_sse42()
278 m_temp_reg_10 = _mm_unpacklo_epi8(m_temp_reg_8, m_temp_reg_1); in ihevcd_itrans_recon_dc_chroma_sse42()
[all …]
|