Home
last modified time | relevance | path

Searched refs:_mm_loadl_epi64 (Results 1 – 25 of 367) sorted by relevance

12345678910>>...15

/external/libmpeg2/common/x86/
Dimpeg2_inter_pred_sse42_intr.c143 src_r0 = _mm_loadl_epi64((__m128i *)src); in impeg2_copy_mb_sse42()
144 src_r1 = _mm_loadl_epi64((__m128i *)(src + src_wd)); in impeg2_copy_mb_sse42()
145 src_r2 = _mm_loadl_epi64((__m128i *)(src + 2 * src_wd)); in impeg2_copy_mb_sse42()
146 src_r3 = _mm_loadl_epi64((__m128i *)(src + 3 * src_wd)); in impeg2_copy_mb_sse42()
157 src_r0 = _mm_loadl_epi64((__m128i *)src); in impeg2_copy_mb_sse42()
158 src_r1 = _mm_loadl_epi64((__m128i *)(src + src_wd)); in impeg2_copy_mb_sse42()
159 src_r2 = _mm_loadl_epi64((__m128i *)(src + 2 * src_wd)); in impeg2_copy_mb_sse42()
160 src_r3 = _mm_loadl_epi64((__m128i *)(src + 3 * src_wd)); in impeg2_copy_mb_sse42()
173 src_r0 = _mm_loadl_epi64((__m128i *)src); in impeg2_copy_mb_sse42()
174 src_r1 = _mm_loadl_epi64((__m128i *)(src + src_wd)); in impeg2_copy_mb_sse42()
[all …]
/external/XNNPACK/src/qs8-dwconv/gen/
Dup32x9-minmax-avx2-mul32.c89 const __m256i vi0x01234567 = _mm256_cvtepi8_epi32(_mm_loadl_epi64((const __m128i*) i0)); in xnn_qs8_dwconv_minmax_ukernel_up32x9__avx2_mul32()
90 …const __m256i vk0x01234567 = _mm256_cvtepi8_epi32(_mm_loadl_epi64((const __m128i*) ((uintptr_t) w … in xnn_qs8_dwconv_minmax_ukernel_up32x9__avx2_mul32()
91 const __m256i vi0x89ABCDEF = _mm256_cvtepi8_epi32(_mm_loadl_epi64((const __m128i*) (i0 + 8))); in xnn_qs8_dwconv_minmax_ukernel_up32x9__avx2_mul32()
92 …const __m256i vk0x89ABCDEF = _mm256_cvtepi8_epi32(_mm_loadl_epi64((const __m128i*) ((uintptr_t) w … in xnn_qs8_dwconv_minmax_ukernel_up32x9__avx2_mul32()
93 … const __m256i vi0xGHIJKLMN = _mm256_cvtepi8_epi32(_mm_loadl_epi64((const __m128i*) (i0 + 16))); in xnn_qs8_dwconv_minmax_ukernel_up32x9__avx2_mul32()
94 …const __m256i vk0xGHIJKLMN = _mm256_cvtepi8_epi32(_mm_loadl_epi64((const __m128i*) ((uintptr_t) w … in xnn_qs8_dwconv_minmax_ukernel_up32x9__avx2_mul32()
95 … const __m256i vi0xOPQRSTUV = _mm256_cvtepi8_epi32(_mm_loadl_epi64((const __m128i*) (i0 + 24))); in xnn_qs8_dwconv_minmax_ukernel_up32x9__avx2_mul32()
96 …const __m256i vk0xOPQRSTUV = _mm256_cvtepi8_epi32(_mm_loadl_epi64((const __m128i*) ((uintptr_t) w … in xnn_qs8_dwconv_minmax_ukernel_up32x9__avx2_mul32()
104 const __m256i vi1x01234567 = _mm256_cvtepi8_epi32(_mm_loadl_epi64((const __m128i*) i1)); in xnn_qs8_dwconv_minmax_ukernel_up32x9__avx2_mul32()
105 …const __m256i vk1x01234567 = _mm256_cvtepi8_epi32(_mm_loadl_epi64((const __m128i*) ((uintptr_t) w … in xnn_qs8_dwconv_minmax_ukernel_up32x9__avx2_mul32()
[all …]
Dup24x9-minmax-avx2-mul32.c88 const __m256i vi0x01234567 = _mm256_cvtepi8_epi32(_mm_loadl_epi64((const __m128i*) i0)); in xnn_qs8_dwconv_minmax_ukernel_up24x9__avx2_mul32()
89 …const __m256i vk0x01234567 = _mm256_cvtepi8_epi32(_mm_loadl_epi64((const __m128i*) ((uintptr_t) w … in xnn_qs8_dwconv_minmax_ukernel_up24x9__avx2_mul32()
90 const __m256i vi0x89ABCDEF = _mm256_cvtepi8_epi32(_mm_loadl_epi64((const __m128i*) (i0 + 8))); in xnn_qs8_dwconv_minmax_ukernel_up24x9__avx2_mul32()
91 …const __m256i vk0x89ABCDEF = _mm256_cvtepi8_epi32(_mm_loadl_epi64((const __m128i*) ((uintptr_t) w … in xnn_qs8_dwconv_minmax_ukernel_up24x9__avx2_mul32()
92 … const __m256i vi0xGHIJKLMN = _mm256_cvtepi8_epi32(_mm_loadl_epi64((const __m128i*) (i0 + 16))); in xnn_qs8_dwconv_minmax_ukernel_up24x9__avx2_mul32()
93 …const __m256i vk0xGHIJKLMN = _mm256_cvtepi8_epi32(_mm_loadl_epi64((const __m128i*) ((uintptr_t) w … in xnn_qs8_dwconv_minmax_ukernel_up24x9__avx2_mul32()
100 const __m256i vi1x01234567 = _mm256_cvtepi8_epi32(_mm_loadl_epi64((const __m128i*) i1)); in xnn_qs8_dwconv_minmax_ukernel_up24x9__avx2_mul32()
101 …const __m256i vk1x01234567 = _mm256_cvtepi8_epi32(_mm_loadl_epi64((const __m128i*) ((uintptr_t) w … in xnn_qs8_dwconv_minmax_ukernel_up24x9__avx2_mul32()
102 const __m256i vi1x89ABCDEF = _mm256_cvtepi8_epi32(_mm_loadl_epi64((const __m128i*) (i1 + 8))); in xnn_qs8_dwconv_minmax_ukernel_up24x9__avx2_mul32()
103 …const __m256i vk1x89ABCDEF = _mm256_cvtepi8_epi32(_mm_loadl_epi64((const __m128i*) ((uintptr_t) w … in xnn_qs8_dwconv_minmax_ukernel_up24x9__avx2_mul32()
[all …]
Dup16x9-minmax-avx2-mul32.c87 const __m256i vi0x01234567 = _mm256_cvtepi8_epi32(_mm_loadl_epi64((const __m128i*) i0)); in xnn_qs8_dwconv_minmax_ukernel_up16x9__avx2_mul32()
88 …const __m256i vk0x01234567 = _mm256_cvtepi8_epi32(_mm_loadl_epi64((const __m128i*) ((uintptr_t) w … in xnn_qs8_dwconv_minmax_ukernel_up16x9__avx2_mul32()
89 const __m256i vi0x89ABCDEF = _mm256_cvtepi8_epi32(_mm_loadl_epi64((const __m128i*) (i0 + 8))); in xnn_qs8_dwconv_minmax_ukernel_up16x9__avx2_mul32()
90 …const __m256i vk0x89ABCDEF = _mm256_cvtepi8_epi32(_mm_loadl_epi64((const __m128i*) ((uintptr_t) w … in xnn_qs8_dwconv_minmax_ukernel_up16x9__avx2_mul32()
96 const __m256i vi1x01234567 = _mm256_cvtepi8_epi32(_mm_loadl_epi64((const __m128i*) i1)); in xnn_qs8_dwconv_minmax_ukernel_up16x9__avx2_mul32()
97 …const __m256i vk1x01234567 = _mm256_cvtepi8_epi32(_mm_loadl_epi64((const __m128i*) ((uintptr_t) w … in xnn_qs8_dwconv_minmax_ukernel_up16x9__avx2_mul32()
98 const __m256i vi1x89ABCDEF = _mm256_cvtepi8_epi32(_mm_loadl_epi64((const __m128i*) (i1 + 8))); in xnn_qs8_dwconv_minmax_ukernel_up16x9__avx2_mul32()
99 …const __m256i vk1x89ABCDEF = _mm256_cvtepi8_epi32(_mm_loadl_epi64((const __m128i*) ((uintptr_t) w … in xnn_qs8_dwconv_minmax_ukernel_up16x9__avx2_mul32()
105 const __m256i vi2x01234567 = _mm256_cvtepi8_epi32(_mm_loadl_epi64((const __m128i*) i2)); in xnn_qs8_dwconv_minmax_ukernel_up16x9__avx2_mul32()
106 …const __m256i vk2x01234567 = _mm256_cvtepi8_epi32(_mm_loadl_epi64((const __m128i*) ((uintptr_t) w … in xnn_qs8_dwconv_minmax_ukernel_up16x9__avx2_mul32()
[all …]
Dup8x9-minmax-avx2-mul32.c86 const __m256i vi0x01234567 = _mm256_cvtepi8_epi32(_mm_loadl_epi64((const __m128i*) i0)); in xnn_qs8_dwconv_minmax_ukernel_up8x9__avx2_mul32()
87 …const __m256i vk0x01234567 = _mm256_cvtepi8_epi32(_mm_loadl_epi64((const __m128i*) ((uintptr_t) w … in xnn_qs8_dwconv_minmax_ukernel_up8x9__avx2_mul32()
92 const __m256i vi1x01234567 = _mm256_cvtepi8_epi32(_mm_loadl_epi64((const __m128i*) i1)); in xnn_qs8_dwconv_minmax_ukernel_up8x9__avx2_mul32()
93 …const __m256i vk1x01234567 = _mm256_cvtepi8_epi32(_mm_loadl_epi64((const __m128i*) ((uintptr_t) w … in xnn_qs8_dwconv_minmax_ukernel_up8x9__avx2_mul32()
98 const __m256i vi2x01234567 = _mm256_cvtepi8_epi32(_mm_loadl_epi64((const __m128i*) i2)); in xnn_qs8_dwconv_minmax_ukernel_up8x9__avx2_mul32()
99 …const __m256i vk2x01234567 = _mm256_cvtepi8_epi32(_mm_loadl_epi64((const __m128i*) ((uintptr_t) w … in xnn_qs8_dwconv_minmax_ukernel_up8x9__avx2_mul32()
104 const __m256i vi3x01234567 = _mm256_cvtepi8_epi32(_mm_loadl_epi64((const __m128i*) i3)); in xnn_qs8_dwconv_minmax_ukernel_up8x9__avx2_mul32()
105 …const __m256i vk3x01234567 = _mm256_cvtepi8_epi32(_mm_loadl_epi64((const __m128i*) ((uintptr_t) w … in xnn_qs8_dwconv_minmax_ukernel_up8x9__avx2_mul32()
110 const __m256i vi4x01234567 = _mm256_cvtepi8_epi32(_mm_loadl_epi64((const __m128i*) i4)); in xnn_qs8_dwconv_minmax_ukernel_up8x9__avx2_mul32()
111 …const __m256i vk4x01234567 = _mm256_cvtepi8_epi32(_mm_loadl_epi64((const __m128i*) ((uintptr_t) w … in xnn_qs8_dwconv_minmax_ukernel_up8x9__avx2_mul32()
[all …]
Dup24x9-minmax-sse41-mul16.c91 const __m128i vi0x01234567 = _mm_loadl_epi64((const __m128i*) i0); in xnn_qs8_dwconv_minmax_ukernel_up24x9__sse41_mul16()
93 …const __m128i vk0x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 24 * sizeof(int32_t… in xnn_qs8_dwconv_minmax_ukernel_up24x9__sse41_mul16()
95 const __m128i vi0x89ABCDEF = _mm_loadl_epi64((const __m128i*) (i0 + 8)); in xnn_qs8_dwconv_minmax_ukernel_up24x9__sse41_mul16()
97 …const __m128i vk0x89ABCDEF = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 24 * sizeof(int32_t… in xnn_qs8_dwconv_minmax_ukernel_up24x9__sse41_mul16()
99 const __m128i vi0xGHIJKLMN = _mm_loadl_epi64((const __m128i*) (i0 + 16)); in xnn_qs8_dwconv_minmax_ukernel_up24x9__sse41_mul16()
101 …const __m128i vk0xGHIJKLMN = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 24 * sizeof(int32_t… in xnn_qs8_dwconv_minmax_ukernel_up24x9__sse41_mul16()
120 const __m128i vi1x01234567 = _mm_loadl_epi64((const __m128i*) i1); in xnn_qs8_dwconv_minmax_ukernel_up24x9__sse41_mul16()
122 …const __m128i vk1x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 24 * sizeof(int32_t… in xnn_qs8_dwconv_minmax_ukernel_up24x9__sse41_mul16()
124 const __m128i vi1x89ABCDEF = _mm_loadl_epi64((const __m128i*) (i1 + 8)); in xnn_qs8_dwconv_minmax_ukernel_up24x9__sse41_mul16()
126 …const __m128i vk1x89ABCDEF = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 24 * sizeof(int32_t… in xnn_qs8_dwconv_minmax_ukernel_up24x9__sse41_mul16()
[all …]
Dup16x9-minmax-sse41-mul16.c89 const __m128i vi0x01234567 = _mm_loadl_epi64((const __m128i*) i0); in xnn_qs8_dwconv_minmax_ukernel_up16x9__sse41_mul16()
91 …const __m128i vk0x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 16 * sizeof(int32_t… in xnn_qs8_dwconv_minmax_ukernel_up16x9__sse41_mul16()
93 const __m128i vi0x89ABCDEF = _mm_loadl_epi64((const __m128i*) (i0 + 8)); in xnn_qs8_dwconv_minmax_ukernel_up16x9__sse41_mul16()
95 …const __m128i vk0x89ABCDEF = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 16 * sizeof(int32_t… in xnn_qs8_dwconv_minmax_ukernel_up16x9__sse41_mul16()
110 const __m128i vi1x01234567 = _mm_loadl_epi64((const __m128i*) i1); in xnn_qs8_dwconv_minmax_ukernel_up16x9__sse41_mul16()
112 …const __m128i vk1x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 16 * sizeof(int32_t… in xnn_qs8_dwconv_minmax_ukernel_up16x9__sse41_mul16()
114 const __m128i vi1x89ABCDEF = _mm_loadl_epi64((const __m128i*) (i1 + 8)); in xnn_qs8_dwconv_minmax_ukernel_up16x9__sse41_mul16()
116 …const __m128i vk1x89ABCDEF = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 16 * sizeof(int32_t… in xnn_qs8_dwconv_minmax_ukernel_up16x9__sse41_mul16()
131 const __m128i vi2x01234567 = _mm_loadl_epi64((const __m128i*) i2); in xnn_qs8_dwconv_minmax_ukernel_up16x9__sse41_mul16()
133 …const __m128i vk2x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 16 * sizeof(int32_t… in xnn_qs8_dwconv_minmax_ukernel_up16x9__sse41_mul16()
[all …]
Dup8x9-minmax-sse41-mul16.c87 const __m128i vi0x01234567 = _mm_loadl_epi64((const __m128i*) i0); in xnn_qs8_dwconv_minmax_ukernel_up8x9__sse41_mul16()
89 …const __m128i vk0x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 8 * sizeof(int32_t)… in xnn_qs8_dwconv_minmax_ukernel_up8x9__sse41_mul16()
100 const __m128i vi1x01234567 = _mm_loadl_epi64((const __m128i*) i1); in xnn_qs8_dwconv_minmax_ukernel_up8x9__sse41_mul16()
102 …const __m128i vk1x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 8 * sizeof(int32_t)… in xnn_qs8_dwconv_minmax_ukernel_up8x9__sse41_mul16()
113 const __m128i vi2x01234567 = _mm_loadl_epi64((const __m128i*) i2); in xnn_qs8_dwconv_minmax_ukernel_up8x9__sse41_mul16()
115 …const __m128i vk2x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 8 * sizeof(int32_t)… in xnn_qs8_dwconv_minmax_ukernel_up8x9__sse41_mul16()
126 const __m128i vi3x01234567 = _mm_loadl_epi64((const __m128i*) i3); in xnn_qs8_dwconv_minmax_ukernel_up8x9__sse41_mul16()
128 …const __m128i vk3x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 8 * sizeof(int32_t)… in xnn_qs8_dwconv_minmax_ukernel_up8x9__sse41_mul16()
139 const __m128i vi4x01234567 = _mm_loadl_epi64((const __m128i*) i4); in xnn_qs8_dwconv_minmax_ukernel_up8x9__sse41_mul16()
141 …const __m128i vk4x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 8 * sizeof(int32_t)… in xnn_qs8_dwconv_minmax_ukernel_up8x9__sse41_mul16()
[all …]
Dup16x9-minmax-ssse3-mul16.c89 const __m128i vi0x01234567 = _mm_loadl_epi64((const __m128i*) i0); in xnn_qs8_dwconv_minmax_ukernel_up16x9__ssse3_mul16()
90 …const __m128i vk0x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 16 * sizeof(int32_t… in xnn_qs8_dwconv_minmax_ukernel_up16x9__ssse3_mul16()
91 const __m128i vi0x89ABCDEF = _mm_loadl_epi64((const __m128i*) (i0 + 8)); in xnn_qs8_dwconv_minmax_ukernel_up16x9__ssse3_mul16()
92 …const __m128i vk0x89ABCDEF = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 16 * sizeof(int32_t… in xnn_qs8_dwconv_minmax_ukernel_up16x9__ssse3_mul16()
110 const __m128i vi1x01234567 = _mm_loadl_epi64((const __m128i*) i1); in xnn_qs8_dwconv_minmax_ukernel_up16x9__ssse3_mul16()
111 …const __m128i vk1x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 16 * sizeof(int32_t… in xnn_qs8_dwconv_minmax_ukernel_up16x9__ssse3_mul16()
112 const __m128i vi1x89ABCDEF = _mm_loadl_epi64((const __m128i*) (i1 + 8)); in xnn_qs8_dwconv_minmax_ukernel_up16x9__ssse3_mul16()
113 …const __m128i vk1x89ABCDEF = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 16 * sizeof(int32_t… in xnn_qs8_dwconv_minmax_ukernel_up16x9__ssse3_mul16()
131 const __m128i vi2x01234567 = _mm_loadl_epi64((const __m128i*) i2); in xnn_qs8_dwconv_minmax_ukernel_up16x9__ssse3_mul16()
132 …const __m128i vk2x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 16 * sizeof(int32_t… in xnn_qs8_dwconv_minmax_ukernel_up16x9__ssse3_mul16()
[all …]
Dup16x9-minmax-sse2-mul16.c89 const __m128i vi0x01234567 = _mm_loadl_epi64((const __m128i*) i0); in xnn_qs8_dwconv_minmax_ukernel_up16x9__sse2_mul16()
90 …const __m128i vk0x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 16 * sizeof(int32_t… in xnn_qs8_dwconv_minmax_ukernel_up16x9__sse2_mul16()
91 const __m128i vi0x89ABCDEF = _mm_loadl_epi64((const __m128i*) (i0 + 8)); in xnn_qs8_dwconv_minmax_ukernel_up16x9__sse2_mul16()
92 …const __m128i vk0x89ABCDEF = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 16 * sizeof(int32_t… in xnn_qs8_dwconv_minmax_ukernel_up16x9__sse2_mul16()
110 const __m128i vi1x01234567 = _mm_loadl_epi64((const __m128i*) i1); in xnn_qs8_dwconv_minmax_ukernel_up16x9__sse2_mul16()
111 …const __m128i vk1x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 16 * sizeof(int32_t… in xnn_qs8_dwconv_minmax_ukernel_up16x9__sse2_mul16()
112 const __m128i vi1x89ABCDEF = _mm_loadl_epi64((const __m128i*) (i1 + 8)); in xnn_qs8_dwconv_minmax_ukernel_up16x9__sse2_mul16()
113 …const __m128i vk1x89ABCDEF = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 16 * sizeof(int32_t… in xnn_qs8_dwconv_minmax_ukernel_up16x9__sse2_mul16()
131 const __m128i vi2x01234567 = _mm_loadl_epi64((const __m128i*) i2); in xnn_qs8_dwconv_minmax_ukernel_up16x9__sse2_mul16()
132 …const __m128i vk2x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 16 * sizeof(int32_t… in xnn_qs8_dwconv_minmax_ukernel_up16x9__sse2_mul16()
[all …]
Dup24x9-minmax-ssse3-mul16.c91 const __m128i vi0x01234567 = _mm_loadl_epi64((const __m128i*) i0); in xnn_qs8_dwconv_minmax_ukernel_up24x9__ssse3_mul16()
92 …const __m128i vk0x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 24 * sizeof(int32_t… in xnn_qs8_dwconv_minmax_ukernel_up24x9__ssse3_mul16()
93 const __m128i vi0x89ABCDEF = _mm_loadl_epi64((const __m128i*) (i0 + 8)); in xnn_qs8_dwconv_minmax_ukernel_up24x9__ssse3_mul16()
94 …const __m128i vk0x89ABCDEF = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 24 * sizeof(int32_t… in xnn_qs8_dwconv_minmax_ukernel_up24x9__ssse3_mul16()
95 const __m128i vi0xGHIJKLMN = _mm_loadl_epi64((const __m128i*) (i0 + 16)); in xnn_qs8_dwconv_minmax_ukernel_up24x9__ssse3_mul16()
96 …const __m128i vk0xGHIJKLMN = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 24 * sizeof(int32_t… in xnn_qs8_dwconv_minmax_ukernel_up24x9__ssse3_mul16()
120 const __m128i vi1x01234567 = _mm_loadl_epi64((const __m128i*) i1); in xnn_qs8_dwconv_minmax_ukernel_up24x9__ssse3_mul16()
121 …const __m128i vk1x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 24 * sizeof(int32_t… in xnn_qs8_dwconv_minmax_ukernel_up24x9__ssse3_mul16()
122 const __m128i vi1x89ABCDEF = _mm_loadl_epi64((const __m128i*) (i1 + 8)); in xnn_qs8_dwconv_minmax_ukernel_up24x9__ssse3_mul16()
123 …const __m128i vk1x89ABCDEF = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 24 * sizeof(int32_t… in xnn_qs8_dwconv_minmax_ukernel_up24x9__ssse3_mul16()
[all …]
Dup24x9-minmax-sse2-mul16.c91 const __m128i vi0x01234567 = _mm_loadl_epi64((const __m128i*) i0); in xnn_qs8_dwconv_minmax_ukernel_up24x9__sse2_mul16()
92 …const __m128i vk0x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 24 * sizeof(int32_t… in xnn_qs8_dwconv_minmax_ukernel_up24x9__sse2_mul16()
93 const __m128i vi0x89ABCDEF = _mm_loadl_epi64((const __m128i*) (i0 + 8)); in xnn_qs8_dwconv_minmax_ukernel_up24x9__sse2_mul16()
94 …const __m128i vk0x89ABCDEF = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 24 * sizeof(int32_t… in xnn_qs8_dwconv_minmax_ukernel_up24x9__sse2_mul16()
95 const __m128i vi0xGHIJKLMN = _mm_loadl_epi64((const __m128i*) (i0 + 16)); in xnn_qs8_dwconv_minmax_ukernel_up24x9__sse2_mul16()
96 …const __m128i vk0xGHIJKLMN = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 24 * sizeof(int32_t… in xnn_qs8_dwconv_minmax_ukernel_up24x9__sse2_mul16()
120 const __m128i vi1x01234567 = _mm_loadl_epi64((const __m128i*) i1); in xnn_qs8_dwconv_minmax_ukernel_up24x9__sse2_mul16()
121 …const __m128i vk1x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 24 * sizeof(int32_t… in xnn_qs8_dwconv_minmax_ukernel_up24x9__sse2_mul16()
122 const __m128i vi1x89ABCDEF = _mm_loadl_epi64((const __m128i*) (i1 + 8)); in xnn_qs8_dwconv_minmax_ukernel_up24x9__sse2_mul16()
123 …const __m128i vk1x89ABCDEF = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 24 * sizeof(int32_t… in xnn_qs8_dwconv_minmax_ukernel_up24x9__sse2_mul16()
[all …]
/external/XNNPACK/src/qs8-gavgpool/gen/
D7p7x-minmax-sse41-c24-acc2.c44 const __m128i vxi0x01234567 = _mm_cvtepi8_epi16(_mm_loadl_epi64((const __m128i*) i0)); in xnn_qs8_gavgpool_minmax_ukernel_7p7x__sse41_c24_acc2()
45 const __m128i vxi0x89ABCDEF = _mm_cvtepi8_epi16(_mm_loadl_epi64((const __m128i*) (i0 + 8))); in xnn_qs8_gavgpool_minmax_ukernel_7p7x__sse41_c24_acc2()
46 const __m128i vxi0xGHIJKLMN = _mm_cvtepi8_epi16(_mm_loadl_epi64((const __m128i*) (i0 + 16))); in xnn_qs8_gavgpool_minmax_ukernel_7p7x__sse41_c24_acc2()
48 const __m128i vxi1x01234567 = _mm_cvtepi8_epi16(_mm_loadl_epi64((const __m128i*) i1)); in xnn_qs8_gavgpool_minmax_ukernel_7p7x__sse41_c24_acc2()
49 const __m128i vxi1x89ABCDEF = _mm_cvtepi8_epi16(_mm_loadl_epi64((const __m128i*) (i1 + 8))); in xnn_qs8_gavgpool_minmax_ukernel_7p7x__sse41_c24_acc2()
50 const __m128i vxi1xGHIJKLMN = _mm_cvtepi8_epi16(_mm_loadl_epi64((const __m128i*) (i1 + 16))); in xnn_qs8_gavgpool_minmax_ukernel_7p7x__sse41_c24_acc2()
52 const __m128i vxi2x01234567 = _mm_cvtepi8_epi16(_mm_loadl_epi64((const __m128i*) i2)); in xnn_qs8_gavgpool_minmax_ukernel_7p7x__sse41_c24_acc2()
53 const __m128i vxi2x89ABCDEF = _mm_cvtepi8_epi16(_mm_loadl_epi64((const __m128i*) (i2 + 8))); in xnn_qs8_gavgpool_minmax_ukernel_7p7x__sse41_c24_acc2()
54 const __m128i vxi2xGHIJKLMN = _mm_cvtepi8_epi16(_mm_loadl_epi64((const __m128i*) (i2 + 16))); in xnn_qs8_gavgpool_minmax_ukernel_7p7x__sse41_c24_acc2()
56 const __m128i vxi3x01234567 = _mm_cvtepi8_epi16(_mm_loadl_epi64((const __m128i*) i3)); in xnn_qs8_gavgpool_minmax_ukernel_7p7x__sse41_c24_acc2()
[all …]
D7p7x-minmax-sse41-c16-acc2.c44 const __m128i vxi0x01234567 = _mm_cvtepi8_epi16(_mm_loadl_epi64((const __m128i*) i0)); in xnn_qs8_gavgpool_minmax_ukernel_7p7x__sse41_c16_acc2()
45 const __m128i vxi0x89ABCDEF = _mm_cvtepi8_epi16(_mm_loadl_epi64((const __m128i*) (i0 + 8))); in xnn_qs8_gavgpool_minmax_ukernel_7p7x__sse41_c16_acc2()
47 const __m128i vxi1x01234567 = _mm_cvtepi8_epi16(_mm_loadl_epi64((const __m128i*) i1)); in xnn_qs8_gavgpool_minmax_ukernel_7p7x__sse41_c16_acc2()
48 const __m128i vxi1x89ABCDEF = _mm_cvtepi8_epi16(_mm_loadl_epi64((const __m128i*) (i1 + 8))); in xnn_qs8_gavgpool_minmax_ukernel_7p7x__sse41_c16_acc2()
50 const __m128i vxi2x01234567 = _mm_cvtepi8_epi16(_mm_loadl_epi64((const __m128i*) i2)); in xnn_qs8_gavgpool_minmax_ukernel_7p7x__sse41_c16_acc2()
51 const __m128i vxi2x89ABCDEF = _mm_cvtepi8_epi16(_mm_loadl_epi64((const __m128i*) (i2 + 8))); in xnn_qs8_gavgpool_minmax_ukernel_7p7x__sse41_c16_acc2()
53 const __m128i vxi3x01234567 = _mm_cvtepi8_epi16(_mm_loadl_epi64((const __m128i*) i3)); in xnn_qs8_gavgpool_minmax_ukernel_7p7x__sse41_c16_acc2()
54 const __m128i vxi3x89ABCDEF = _mm_cvtepi8_epi16(_mm_loadl_epi64((const __m128i*) (i3 + 8))); in xnn_qs8_gavgpool_minmax_ukernel_7p7x__sse41_c16_acc2()
56 const __m128i vxi4x01234567 = _mm_cvtepi8_epi16(_mm_loadl_epi64((const __m128i*) i4)); in xnn_qs8_gavgpool_minmax_ukernel_7p7x__sse41_c16_acc2()
57 const __m128i vxi4x89ABCDEF = _mm_cvtepi8_epi16(_mm_loadl_epi64((const __m128i*) (i4 + 8))); in xnn_qs8_gavgpool_minmax_ukernel_7p7x__sse41_c16_acc2()
[all …]
D7p7x-minmax-sse2-c24-acc2.c44 const __m128i vi0x01234567 = _mm_loadl_epi64((const __m128i*) i0); in xnn_qs8_gavgpool_minmax_ukernel_7p7x__sse2_c24_acc2()
45 const __m128i vi0x89ABCDEF = _mm_loadl_epi64((const __m128i*) (i0 + 8)); in xnn_qs8_gavgpool_minmax_ukernel_7p7x__sse2_c24_acc2()
46 const __m128i vi0xGHIJKLMN = _mm_loadl_epi64((const __m128i*) (i0 + 16)); in xnn_qs8_gavgpool_minmax_ukernel_7p7x__sse2_c24_acc2()
48 const __m128i vi1x01234567 = _mm_loadl_epi64((const __m128i*) i1); in xnn_qs8_gavgpool_minmax_ukernel_7p7x__sse2_c24_acc2()
49 const __m128i vi1x89ABCDEF = _mm_loadl_epi64((const __m128i*) (i1 + 8)); in xnn_qs8_gavgpool_minmax_ukernel_7p7x__sse2_c24_acc2()
50 const __m128i vi1xGHIJKLMN = _mm_loadl_epi64((const __m128i*) (i1 + 16)); in xnn_qs8_gavgpool_minmax_ukernel_7p7x__sse2_c24_acc2()
52 const __m128i vi2x01234567 = _mm_loadl_epi64((const __m128i*) i2); in xnn_qs8_gavgpool_minmax_ukernel_7p7x__sse2_c24_acc2()
53 const __m128i vi2x89ABCDEF = _mm_loadl_epi64((const __m128i*) (i2 + 8)); in xnn_qs8_gavgpool_minmax_ukernel_7p7x__sse2_c24_acc2()
54 const __m128i vi2xGHIJKLMN = _mm_loadl_epi64((const __m128i*) (i2 + 16)); in xnn_qs8_gavgpool_minmax_ukernel_7p7x__sse2_c24_acc2()
56 const __m128i vi3x01234567 = _mm_loadl_epi64((const __m128i*) i3); in xnn_qs8_gavgpool_minmax_ukernel_7p7x__sse2_c24_acc2()
[all …]
D7p7x-minmax-ssse3-c24-acc2.c44 const __m128i vi0x01234567 = _mm_loadl_epi64((const __m128i*) i0); in xnn_qs8_gavgpool_minmax_ukernel_7p7x__ssse3_c24_acc2()
45 const __m128i vi0x89ABCDEF = _mm_loadl_epi64((const __m128i*) (i0 + 8)); in xnn_qs8_gavgpool_minmax_ukernel_7p7x__ssse3_c24_acc2()
46 const __m128i vi0xGHIJKLMN = _mm_loadl_epi64((const __m128i*) (i0 + 16)); in xnn_qs8_gavgpool_minmax_ukernel_7p7x__ssse3_c24_acc2()
48 const __m128i vi1x01234567 = _mm_loadl_epi64((const __m128i*) i1); in xnn_qs8_gavgpool_minmax_ukernel_7p7x__ssse3_c24_acc2()
49 const __m128i vi1x89ABCDEF = _mm_loadl_epi64((const __m128i*) (i1 + 8)); in xnn_qs8_gavgpool_minmax_ukernel_7p7x__ssse3_c24_acc2()
50 const __m128i vi1xGHIJKLMN = _mm_loadl_epi64((const __m128i*) (i1 + 16)); in xnn_qs8_gavgpool_minmax_ukernel_7p7x__ssse3_c24_acc2()
52 const __m128i vi2x01234567 = _mm_loadl_epi64((const __m128i*) i2); in xnn_qs8_gavgpool_minmax_ukernel_7p7x__ssse3_c24_acc2()
53 const __m128i vi2x89ABCDEF = _mm_loadl_epi64((const __m128i*) (i2 + 8)); in xnn_qs8_gavgpool_minmax_ukernel_7p7x__ssse3_c24_acc2()
54 const __m128i vi2xGHIJKLMN = _mm_loadl_epi64((const __m128i*) (i2 + 16)); in xnn_qs8_gavgpool_minmax_ukernel_7p7x__ssse3_c24_acc2()
56 const __m128i vi3x01234567 = _mm_loadl_epi64((const __m128i*) i3); in xnn_qs8_gavgpool_minmax_ukernel_7p7x__ssse3_c24_acc2()
[all …]
D7p7x-minmax-ssse3-c16-acc2.c44 const __m128i vi0x01234567 = _mm_loadl_epi64((const __m128i*) i0); in xnn_qs8_gavgpool_minmax_ukernel_7p7x__ssse3_c16_acc2()
45 const __m128i vi0x89ABCDEF = _mm_loadl_epi64((const __m128i*) (i0 + 8)); in xnn_qs8_gavgpool_minmax_ukernel_7p7x__ssse3_c16_acc2()
47 const __m128i vi1x01234567 = _mm_loadl_epi64((const __m128i*) i1); in xnn_qs8_gavgpool_minmax_ukernel_7p7x__ssse3_c16_acc2()
48 const __m128i vi1x89ABCDEF = _mm_loadl_epi64((const __m128i*) (i1 + 8)); in xnn_qs8_gavgpool_minmax_ukernel_7p7x__ssse3_c16_acc2()
50 const __m128i vi2x01234567 = _mm_loadl_epi64((const __m128i*) i2); in xnn_qs8_gavgpool_minmax_ukernel_7p7x__ssse3_c16_acc2()
51 const __m128i vi2x89ABCDEF = _mm_loadl_epi64((const __m128i*) (i2 + 8)); in xnn_qs8_gavgpool_minmax_ukernel_7p7x__ssse3_c16_acc2()
53 const __m128i vi3x01234567 = _mm_loadl_epi64((const __m128i*) i3); in xnn_qs8_gavgpool_minmax_ukernel_7p7x__ssse3_c16_acc2()
54 const __m128i vi3x89ABCDEF = _mm_loadl_epi64((const __m128i*) (i3 + 8)); in xnn_qs8_gavgpool_minmax_ukernel_7p7x__ssse3_c16_acc2()
56 const __m128i vi4x01234567 = _mm_loadl_epi64((const __m128i*) i4); in xnn_qs8_gavgpool_minmax_ukernel_7p7x__ssse3_c16_acc2()
57 const __m128i vi4x89ABCDEF = _mm_loadl_epi64((const __m128i*) (i4 + 8)); in xnn_qs8_gavgpool_minmax_ukernel_7p7x__ssse3_c16_acc2()
[all …]
D7p7x-minmax-sse41-c8-acc2.c44 const __m128i vxi0x01234567 = _mm_cvtepi8_epi16(_mm_loadl_epi64((const __m128i*) i0)); in xnn_qs8_gavgpool_minmax_ukernel_7p7x__sse41_c8_acc2()
46 const __m128i vxi1x01234567 = _mm_cvtepi8_epi16(_mm_loadl_epi64((const __m128i*) i1)); in xnn_qs8_gavgpool_minmax_ukernel_7p7x__sse41_c8_acc2()
48 const __m128i vxi2x01234567 = _mm_cvtepi8_epi16(_mm_loadl_epi64((const __m128i*) i2)); in xnn_qs8_gavgpool_minmax_ukernel_7p7x__sse41_c8_acc2()
50 const __m128i vxi3x01234567 = _mm_cvtepi8_epi16(_mm_loadl_epi64((const __m128i*) i3)); in xnn_qs8_gavgpool_minmax_ukernel_7p7x__sse41_c8_acc2()
52 const __m128i vxi4x01234567 = _mm_cvtepi8_epi16(_mm_loadl_epi64((const __m128i*) i4)); in xnn_qs8_gavgpool_minmax_ukernel_7p7x__sse41_c8_acc2()
54 const __m128i vxi5x01234567 = _mm_cvtepi8_epi16(_mm_loadl_epi64((const __m128i*) i5)); in xnn_qs8_gavgpool_minmax_ukernel_7p7x__sse41_c8_acc2()
56 const __m128i vxi6x01234567 = _mm_cvtepi8_epi16(_mm_loadl_epi64((const __m128i*) i6)); in xnn_qs8_gavgpool_minmax_ukernel_7p7x__sse41_c8_acc2()
90 const __m128i vxi0x01234567 = _mm_cvtepi8_epi16(_mm_loadl_epi64((const __m128i*) i0)); in xnn_qs8_gavgpool_minmax_ukernel_7p7x__sse41_c8_acc2()
92 const __m128i vxi1x01234567 = _mm_cvtepi8_epi16(_mm_loadl_epi64((const __m128i*) i1)); in xnn_qs8_gavgpool_minmax_ukernel_7p7x__sse41_c8_acc2()
94 const __m128i vxi2x01234567 = _mm_cvtepi8_epi16(_mm_loadl_epi64((const __m128i*) i2)); in xnn_qs8_gavgpool_minmax_ukernel_7p7x__sse41_c8_acc2()
[all …]
D7p7x-minmax-sse2-c16-acc2.c44 const __m128i vi0x01234567 = _mm_loadl_epi64((const __m128i*) i0); in xnn_qs8_gavgpool_minmax_ukernel_7p7x__sse2_c16_acc2()
45 const __m128i vi0x89ABCDEF = _mm_loadl_epi64((const __m128i*) (i0 + 8)); in xnn_qs8_gavgpool_minmax_ukernel_7p7x__sse2_c16_acc2()
47 const __m128i vi1x01234567 = _mm_loadl_epi64((const __m128i*) i1); in xnn_qs8_gavgpool_minmax_ukernel_7p7x__sse2_c16_acc2()
48 const __m128i vi1x89ABCDEF = _mm_loadl_epi64((const __m128i*) (i1 + 8)); in xnn_qs8_gavgpool_minmax_ukernel_7p7x__sse2_c16_acc2()
50 const __m128i vi2x01234567 = _mm_loadl_epi64((const __m128i*) i2); in xnn_qs8_gavgpool_minmax_ukernel_7p7x__sse2_c16_acc2()
51 const __m128i vi2x89ABCDEF = _mm_loadl_epi64((const __m128i*) (i2 + 8)); in xnn_qs8_gavgpool_minmax_ukernel_7p7x__sse2_c16_acc2()
53 const __m128i vi3x01234567 = _mm_loadl_epi64((const __m128i*) i3); in xnn_qs8_gavgpool_minmax_ukernel_7p7x__sse2_c16_acc2()
54 const __m128i vi3x89ABCDEF = _mm_loadl_epi64((const __m128i*) (i3 + 8)); in xnn_qs8_gavgpool_minmax_ukernel_7p7x__sse2_c16_acc2()
56 const __m128i vi4x01234567 = _mm_loadl_epi64((const __m128i*) i4); in xnn_qs8_gavgpool_minmax_ukernel_7p7x__sse2_c16_acc2()
57 const __m128i vi4x89ABCDEF = _mm_loadl_epi64((const __m128i*) (i4 + 8)); in xnn_qs8_gavgpool_minmax_ukernel_7p7x__sse2_c16_acc2()
[all …]
D7x-minmax-sse41-c24-acc2.c59 const __m128i vshift = _mm_loadl_epi64((const __m128i*) params->sse2.shift); in xnn_qs8_gavgpool_minmax_ukernel_7x__sse41_c24_acc2()
61 const __m128i vxi0x01234567 = _mm_cvtepi8_epi16(_mm_loadl_epi64((const __m128i*) i0)); in xnn_qs8_gavgpool_minmax_ukernel_7x__sse41_c24_acc2()
62 const __m128i vxi0x89ABCDEF = _mm_cvtepi8_epi16(_mm_loadl_epi64((const __m128i*) (i0 + 8))); in xnn_qs8_gavgpool_minmax_ukernel_7x__sse41_c24_acc2()
63 const __m128i vxi0xGHIJKLMN = _mm_cvtepi8_epi16(_mm_loadl_epi64((const __m128i*) (i0 + 16))); in xnn_qs8_gavgpool_minmax_ukernel_7x__sse41_c24_acc2()
65 const __m128i vxi1x01234567 = _mm_cvtepi8_epi16(_mm_loadl_epi64((const __m128i*) i1)); in xnn_qs8_gavgpool_minmax_ukernel_7x__sse41_c24_acc2()
66 const __m128i vxi1x89ABCDEF = _mm_cvtepi8_epi16(_mm_loadl_epi64((const __m128i*) (i1 + 8))); in xnn_qs8_gavgpool_minmax_ukernel_7x__sse41_c24_acc2()
67 const __m128i vxi1xGHIJKLMN = _mm_cvtepi8_epi16(_mm_loadl_epi64((const __m128i*) (i1 + 16))); in xnn_qs8_gavgpool_minmax_ukernel_7x__sse41_c24_acc2()
69 const __m128i vxi2x01234567 = _mm_cvtepi8_epi16(_mm_loadl_epi64((const __m128i*) i2)); in xnn_qs8_gavgpool_minmax_ukernel_7x__sse41_c24_acc2()
70 const __m128i vxi2x89ABCDEF = _mm_cvtepi8_epi16(_mm_loadl_epi64((const __m128i*) (i2 + 8))); in xnn_qs8_gavgpool_minmax_ukernel_7x__sse41_c24_acc2()
71 const __m128i vxi2xGHIJKLMN = _mm_cvtepi8_epi16(_mm_loadl_epi64((const __m128i*) (i2 + 16))); in xnn_qs8_gavgpool_minmax_ukernel_7x__sse41_c24_acc2()
[all …]
D7x-minmax-sse41-c16-acc2.c59 const __m128i vshift = _mm_loadl_epi64((const __m128i*) params->sse2.shift); in xnn_qs8_gavgpool_minmax_ukernel_7x__sse41_c16_acc2()
61 const __m128i vxi0x01234567 = _mm_cvtepi8_epi16(_mm_loadl_epi64((const __m128i*) i0)); in xnn_qs8_gavgpool_minmax_ukernel_7x__sse41_c16_acc2()
62 const __m128i vxi0x89ABCDEF = _mm_cvtepi8_epi16(_mm_loadl_epi64((const __m128i*) (i0 + 8))); in xnn_qs8_gavgpool_minmax_ukernel_7x__sse41_c16_acc2()
64 const __m128i vxi1x01234567 = _mm_cvtepi8_epi16(_mm_loadl_epi64((const __m128i*) i1)); in xnn_qs8_gavgpool_minmax_ukernel_7x__sse41_c16_acc2()
65 const __m128i vxi1x89ABCDEF = _mm_cvtepi8_epi16(_mm_loadl_epi64((const __m128i*) (i1 + 8))); in xnn_qs8_gavgpool_minmax_ukernel_7x__sse41_c16_acc2()
67 const __m128i vxi2x01234567 = _mm_cvtepi8_epi16(_mm_loadl_epi64((const __m128i*) i2)); in xnn_qs8_gavgpool_minmax_ukernel_7x__sse41_c16_acc2()
68 const __m128i vxi2x89ABCDEF = _mm_cvtepi8_epi16(_mm_loadl_epi64((const __m128i*) (i2 + 8))); in xnn_qs8_gavgpool_minmax_ukernel_7x__sse41_c16_acc2()
70 const __m128i vxi3x01234567 = _mm_cvtepi8_epi16(_mm_loadl_epi64((const __m128i*) i3)); in xnn_qs8_gavgpool_minmax_ukernel_7x__sse41_c16_acc2()
71 const __m128i vxi3x89ABCDEF = _mm_cvtepi8_epi16(_mm_loadl_epi64((const __m128i*) (i3 + 8))); in xnn_qs8_gavgpool_minmax_ukernel_7x__sse41_c16_acc2()
73 const __m128i vxi4x01234567 = _mm_cvtepi8_epi16(_mm_loadl_epi64((const __m128i*) i4)); in xnn_qs8_gavgpool_minmax_ukernel_7x__sse41_c16_acc2()
[all …]
D7p7x-minmax-sse2-c8-acc2.c44 const __m128i vi0x01234567 = _mm_loadl_epi64((const __m128i*) i0); in xnn_qs8_gavgpool_minmax_ukernel_7p7x__sse2_c8_acc2()
46 const __m128i vi1x01234567 = _mm_loadl_epi64((const __m128i*) i1); in xnn_qs8_gavgpool_minmax_ukernel_7p7x__sse2_c8_acc2()
48 const __m128i vi2x01234567 = _mm_loadl_epi64((const __m128i*) i2); in xnn_qs8_gavgpool_minmax_ukernel_7p7x__sse2_c8_acc2()
50 const __m128i vi3x01234567 = _mm_loadl_epi64((const __m128i*) i3); in xnn_qs8_gavgpool_minmax_ukernel_7p7x__sse2_c8_acc2()
52 const __m128i vi4x01234567 = _mm_loadl_epi64((const __m128i*) i4); in xnn_qs8_gavgpool_minmax_ukernel_7p7x__sse2_c8_acc2()
54 const __m128i vi5x01234567 = _mm_loadl_epi64((const __m128i*) i5); in xnn_qs8_gavgpool_minmax_ukernel_7p7x__sse2_c8_acc2()
56 const __m128i vi6x01234567 = _mm_loadl_epi64((const __m128i*) i6); in xnn_qs8_gavgpool_minmax_ukernel_7p7x__sse2_c8_acc2()
98 const __m128i vi0x01234567 = _mm_loadl_epi64((const __m128i*) i0); in xnn_qs8_gavgpool_minmax_ukernel_7p7x__sse2_c8_acc2()
100 const __m128i vi1x01234567 = _mm_loadl_epi64((const __m128i*) i1); in xnn_qs8_gavgpool_minmax_ukernel_7p7x__sse2_c8_acc2()
102 const __m128i vi2x01234567 = _mm_loadl_epi64((const __m128i*) i2); in xnn_qs8_gavgpool_minmax_ukernel_7p7x__sse2_c8_acc2()
[all …]
/external/libhevc/common/x86/
Dihevc_inter_pred_filters_sse42_intr.c130 src0_16x8b = _mm_loadl_epi64((__m128i *)(pu1_src)); /* row =0 */ in ihevc_inter_pred_luma_copy_w16out_sse42()
131 src1_16x8b = _mm_loadl_epi64((__m128i *)(pu1_src + 1 * src_strd)); /* row =1 */ in ihevc_inter_pred_luma_copy_w16out_sse42()
132 src2_16x8b = _mm_loadl_epi64((__m128i *)(pu1_src + 2 * src_strd)); /* row =2 */ in ihevc_inter_pred_luma_copy_w16out_sse42()
133 src3_16x8b = _mm_loadl_epi64((__m128i *)(pu1_src + 3 * src_strd)); /* row =3 */ in ihevc_inter_pred_luma_copy_w16out_sse42()
166 src0_16x8b = _mm_loadl_epi64((__m128i *)(pu1_src)); /* row =0 */ in ihevc_inter_pred_luma_copy_w16out_sse42()
167 src1_16x8b = _mm_loadl_epi64((__m128i *)(pu1_src + 1 * src_strd)); /* row =1 */ in ihevc_inter_pred_luma_copy_w16out_sse42()
168 src2_16x8b = _mm_loadl_epi64((__m128i *)(pu1_src + 2 * src_strd)); /* row =2 */ in ihevc_inter_pred_luma_copy_w16out_sse42()
169 src3_16x8b = _mm_loadl_epi64((__m128i *)(pu1_src + 3 * src_strd)); /* row =3 */ in ihevc_inter_pred_luma_copy_w16out_sse42()
288 src0_16x8b = _mm_loadl_epi64((__m128i *)(pu1_src)); /* row =0 */ in ihevc_inter_pred_chroma_copy_sse42()
289 src1_16x8b = _mm_loadl_epi64((__m128i *)(pu1_src + 1 * src_strd)); /* row =1 */ in ihevc_inter_pred_chroma_copy_sse42()
[all …]
/external/flac/src/libFLAC/
Dlpc_intrin_sse41.c70 xmm0 = _mm_loadl_epi64((const __m128i*)(qlp_coeff+0)); // 0 0 q[1] q[0] in FLAC__lpc_compute_residual_from_qlp_coefficients_wide_intrin_sse41()
71 xmm1 = _mm_loadl_epi64((const __m128i*)(qlp_coeff+2)); // 0 0 q[3] q[2] in FLAC__lpc_compute_residual_from_qlp_coefficients_wide_intrin_sse41()
72 xmm2 = _mm_loadl_epi64((const __m128i*)(qlp_coeff+4)); // 0 0 q[5] q[4] in FLAC__lpc_compute_residual_from_qlp_coefficients_wide_intrin_sse41()
73 xmm3 = _mm_loadl_epi64((const __m128i*)(qlp_coeff+6)); // 0 0 q[7] q[6] in FLAC__lpc_compute_residual_from_qlp_coefficients_wide_intrin_sse41()
74 xmm4 = _mm_loadl_epi64((const __m128i*)(qlp_coeff+8)); // 0 0 q[9] q[8] in FLAC__lpc_compute_residual_from_qlp_coefficients_wide_intrin_sse41()
75 xmm5 = _mm_loadl_epi64((const __m128i*)(qlp_coeff+10)); // 0 0 q[11] q[10] in FLAC__lpc_compute_residual_from_qlp_coefficients_wide_intrin_sse41()
88 xmm7 = _mm_loadl_epi64((const __m128i*)(data+i-12)); // 0 0 d[i-11] d[i-12] in FLAC__lpc_compute_residual_from_qlp_coefficients_wide_intrin_sse41()
94 xmm6 = _mm_loadl_epi64((const __m128i*)(data+i-10)); in FLAC__lpc_compute_residual_from_qlp_coefficients_wide_intrin_sse41()
101 xmm6 = _mm_loadl_epi64((const __m128i*)(data+i-8)); in FLAC__lpc_compute_residual_from_qlp_coefficients_wide_intrin_sse41()
108 xmm6 = _mm_loadl_epi64((const __m128i*)(data+i-6)); in FLAC__lpc_compute_residual_from_qlp_coefficients_wide_intrin_sse41()
[all …]
/external/XNNPACK/src/qu8-gavgpool/
D7p7x-minmax-sse2-c8.c44 const __m128i vi0 = _mm_loadl_epi64((const __m128i*) i0); i0 += 8; in xnn_qu8_gavgpool_minmax_ukernel_7p7x__sse2_c8()
45 const __m128i vi1 = _mm_loadl_epi64((const __m128i*) i1); i1 += 8; in xnn_qu8_gavgpool_minmax_ukernel_7p7x__sse2_c8()
46 const __m128i vi2 = _mm_loadl_epi64((const __m128i*) i2); i2 += 8; in xnn_qu8_gavgpool_minmax_ukernel_7p7x__sse2_c8()
47 const __m128i vi3 = _mm_loadl_epi64((const __m128i*) i3); i3 += 8; in xnn_qu8_gavgpool_minmax_ukernel_7p7x__sse2_c8()
48 const __m128i vi4 = _mm_loadl_epi64((const __m128i*) i4); i4 += 8; in xnn_qu8_gavgpool_minmax_ukernel_7p7x__sse2_c8()
49 const __m128i vi5 = _mm_loadl_epi64((const __m128i*) i5); i5 += 8; in xnn_qu8_gavgpool_minmax_ukernel_7p7x__sse2_c8()
50 const __m128i vi6 = _mm_loadl_epi64((const __m128i*) i6); i6 += 8; in xnn_qu8_gavgpool_minmax_ukernel_7p7x__sse2_c8()
86 const __m128i vi0 = _mm_loadl_epi64((const __m128i*) i0); i0 += 8; in xnn_qu8_gavgpool_minmax_ukernel_7p7x__sse2_c8()
87 const __m128i vi1 = _mm_loadl_epi64((const __m128i*) i1); i1 += 8; in xnn_qu8_gavgpool_minmax_ukernel_7p7x__sse2_c8()
88 const __m128i vi2 = _mm_loadl_epi64((const __m128i*) i2); i2 += 8; in xnn_qu8_gavgpool_minmax_ukernel_7p7x__sse2_c8()
[all …]

12345678910>>...15