Home
last modified time | relevance | path

Searched refs:_mm_cvtepi8_epi32 (Results 1 – 25 of 28) sorted by relevance

12

/external/XNNPACK/src/qs8-vadd/gen/
Dminmax-sse41-mul32-ld32-x32.c36 const __m128i vx0123 = _mm_cvtepi8_epi32(_mm_loadu_si32(input_x)); in xnn_qs8_vadd_minmax_ukernel__sse41_mul32_ld32_x32()
37 const __m128i vy0123 = _mm_cvtepi8_epi32(_mm_loadu_si32(input_y)); in xnn_qs8_vadd_minmax_ukernel__sse41_mul32_ld32_x32()
38 const __m128i vx4567 = _mm_cvtepi8_epi32(_mm_loadu_si32(input_x + 4)); in xnn_qs8_vadd_minmax_ukernel__sse41_mul32_ld32_x32()
39 const __m128i vy4567 = _mm_cvtepi8_epi32(_mm_loadu_si32(input_y + 4)); in xnn_qs8_vadd_minmax_ukernel__sse41_mul32_ld32_x32()
40 const __m128i vx89AB = _mm_cvtepi8_epi32(_mm_loadu_si32(input_x + 8)); in xnn_qs8_vadd_minmax_ukernel__sse41_mul32_ld32_x32()
41 const __m128i vy89AB = _mm_cvtepi8_epi32(_mm_loadu_si32(input_y + 8)); in xnn_qs8_vadd_minmax_ukernel__sse41_mul32_ld32_x32()
42 const __m128i vxCDEF = _mm_cvtepi8_epi32(_mm_loadu_si32(input_x + 12)); in xnn_qs8_vadd_minmax_ukernel__sse41_mul32_ld32_x32()
43 const __m128i vyCDEF = _mm_cvtepi8_epi32(_mm_loadu_si32(input_y + 12)); in xnn_qs8_vadd_minmax_ukernel__sse41_mul32_ld32_x32()
44 const __m128i vxGHIJ = _mm_cvtepi8_epi32(_mm_loadu_si32(input_x + 16)); in xnn_qs8_vadd_minmax_ukernel__sse41_mul32_ld32_x32()
45 const __m128i vyGHIJ = _mm_cvtepi8_epi32(_mm_loadu_si32(input_y + 16)); in xnn_qs8_vadd_minmax_ukernel__sse41_mul32_ld32_x32()
[all …]
Dminmax-xop-mul32-ld32-x32.c41 const __m128i vx0123 = _mm_cvtepi8_epi32(_mm_loadu_si32(input_x)); in xnn_qs8_vadd_minmax_ukernel__xop_mul32_ld32_x32()
42 const __m128i vy0123 = _mm_cvtepi8_epi32(_mm_loadu_si32(input_y)); in xnn_qs8_vadd_minmax_ukernel__xop_mul32_ld32_x32()
43 const __m128i vx4567 = _mm_cvtepi8_epi32(_mm_loadu_si32(input_x + 4)); in xnn_qs8_vadd_minmax_ukernel__xop_mul32_ld32_x32()
44 const __m128i vy4567 = _mm_cvtepi8_epi32(_mm_loadu_si32(input_y + 4)); in xnn_qs8_vadd_minmax_ukernel__xop_mul32_ld32_x32()
45 const __m128i vx89AB = _mm_cvtepi8_epi32(_mm_loadu_si32(input_x + 8)); in xnn_qs8_vadd_minmax_ukernel__xop_mul32_ld32_x32()
46 const __m128i vy89AB = _mm_cvtepi8_epi32(_mm_loadu_si32(input_y + 8)); in xnn_qs8_vadd_minmax_ukernel__xop_mul32_ld32_x32()
47 const __m128i vxCDEF = _mm_cvtepi8_epi32(_mm_loadu_si32(input_x + 12)); in xnn_qs8_vadd_minmax_ukernel__xop_mul32_ld32_x32()
48 const __m128i vyCDEF = _mm_cvtepi8_epi32(_mm_loadu_si32(input_y + 12)); in xnn_qs8_vadd_minmax_ukernel__xop_mul32_ld32_x32()
49 const __m128i vxGHIJ = _mm_cvtepi8_epi32(_mm_loadu_si32(input_x + 16)); in xnn_qs8_vadd_minmax_ukernel__xop_mul32_ld32_x32()
50 const __m128i vyGHIJ = _mm_cvtepi8_epi32(_mm_loadu_si32(input_y + 16)); in xnn_qs8_vadd_minmax_ukernel__xop_mul32_ld32_x32()
[all …]
Dminmax-sse41-mul32-ld32-x24.c36 const __m128i vx0123 = _mm_cvtepi8_epi32(_mm_loadu_si32(input_x)); in xnn_qs8_vadd_minmax_ukernel__sse41_mul32_ld32_x24()
37 const __m128i vy0123 = _mm_cvtepi8_epi32(_mm_loadu_si32(input_y)); in xnn_qs8_vadd_minmax_ukernel__sse41_mul32_ld32_x24()
38 const __m128i vx4567 = _mm_cvtepi8_epi32(_mm_loadu_si32(input_x + 4)); in xnn_qs8_vadd_minmax_ukernel__sse41_mul32_ld32_x24()
39 const __m128i vy4567 = _mm_cvtepi8_epi32(_mm_loadu_si32(input_y + 4)); in xnn_qs8_vadd_minmax_ukernel__sse41_mul32_ld32_x24()
40 const __m128i vx89AB = _mm_cvtepi8_epi32(_mm_loadu_si32(input_x + 8)); in xnn_qs8_vadd_minmax_ukernel__sse41_mul32_ld32_x24()
41 const __m128i vy89AB = _mm_cvtepi8_epi32(_mm_loadu_si32(input_y + 8)); in xnn_qs8_vadd_minmax_ukernel__sse41_mul32_ld32_x24()
42 const __m128i vxCDEF = _mm_cvtepi8_epi32(_mm_loadu_si32(input_x + 12)); in xnn_qs8_vadd_minmax_ukernel__sse41_mul32_ld32_x24()
43 const __m128i vyCDEF = _mm_cvtepi8_epi32(_mm_loadu_si32(input_y + 12)); in xnn_qs8_vadd_minmax_ukernel__sse41_mul32_ld32_x24()
44 const __m128i vxGHIJ = _mm_cvtepi8_epi32(_mm_loadu_si32(input_x + 16)); in xnn_qs8_vadd_minmax_ukernel__sse41_mul32_ld32_x24()
45 const __m128i vyGHIJ = _mm_cvtepi8_epi32(_mm_loadu_si32(input_y + 16)); in xnn_qs8_vadd_minmax_ukernel__sse41_mul32_ld32_x24()
[all …]
Dminmax-xop-mul32-ld32-x24.c41 const __m128i vx0123 = _mm_cvtepi8_epi32(_mm_loadu_si32(input_x)); in xnn_qs8_vadd_minmax_ukernel__xop_mul32_ld32_x24()
42 const __m128i vy0123 = _mm_cvtepi8_epi32(_mm_loadu_si32(input_y)); in xnn_qs8_vadd_minmax_ukernel__xop_mul32_ld32_x24()
43 const __m128i vx4567 = _mm_cvtepi8_epi32(_mm_loadu_si32(input_x + 4)); in xnn_qs8_vadd_minmax_ukernel__xop_mul32_ld32_x24()
44 const __m128i vy4567 = _mm_cvtepi8_epi32(_mm_loadu_si32(input_y + 4)); in xnn_qs8_vadd_minmax_ukernel__xop_mul32_ld32_x24()
45 const __m128i vx89AB = _mm_cvtepi8_epi32(_mm_loadu_si32(input_x + 8)); in xnn_qs8_vadd_minmax_ukernel__xop_mul32_ld32_x24()
46 const __m128i vy89AB = _mm_cvtepi8_epi32(_mm_loadu_si32(input_y + 8)); in xnn_qs8_vadd_minmax_ukernel__xop_mul32_ld32_x24()
47 const __m128i vxCDEF = _mm_cvtepi8_epi32(_mm_loadu_si32(input_x + 12)); in xnn_qs8_vadd_minmax_ukernel__xop_mul32_ld32_x24()
48 const __m128i vyCDEF = _mm_cvtepi8_epi32(_mm_loadu_si32(input_y + 12)); in xnn_qs8_vadd_minmax_ukernel__xop_mul32_ld32_x24()
49 const __m128i vxGHIJ = _mm_cvtepi8_epi32(_mm_loadu_si32(input_x + 16)); in xnn_qs8_vadd_minmax_ukernel__xop_mul32_ld32_x24()
50 const __m128i vyGHIJ = _mm_cvtepi8_epi32(_mm_loadu_si32(input_y + 16)); in xnn_qs8_vadd_minmax_ukernel__xop_mul32_ld32_x24()
[all …]
Dminmax-xop-mul32-ld32-x16.c41 const __m128i vx0123 = _mm_cvtepi8_epi32(_mm_loadu_si32(input_x)); in xnn_qs8_vadd_minmax_ukernel__xop_mul32_ld32_x16()
42 const __m128i vy0123 = _mm_cvtepi8_epi32(_mm_loadu_si32(input_y)); in xnn_qs8_vadd_minmax_ukernel__xop_mul32_ld32_x16()
43 const __m128i vx4567 = _mm_cvtepi8_epi32(_mm_loadu_si32(input_x + 4)); in xnn_qs8_vadd_minmax_ukernel__xop_mul32_ld32_x16()
44 const __m128i vy4567 = _mm_cvtepi8_epi32(_mm_loadu_si32(input_y + 4)); in xnn_qs8_vadd_minmax_ukernel__xop_mul32_ld32_x16()
45 const __m128i vx89AB = _mm_cvtepi8_epi32(_mm_loadu_si32(input_x + 8)); in xnn_qs8_vadd_minmax_ukernel__xop_mul32_ld32_x16()
46 const __m128i vy89AB = _mm_cvtepi8_epi32(_mm_loadu_si32(input_y + 8)); in xnn_qs8_vadd_minmax_ukernel__xop_mul32_ld32_x16()
47 const __m128i vxCDEF = _mm_cvtepi8_epi32(_mm_loadu_si32(input_x + 12)); in xnn_qs8_vadd_minmax_ukernel__xop_mul32_ld32_x16()
48 const __m128i vyCDEF = _mm_cvtepi8_epi32(_mm_loadu_si32(input_y + 12)); in xnn_qs8_vadd_minmax_ukernel__xop_mul32_ld32_x16()
88 const __m128i vx0123 = _mm_cvtepi8_epi32(_mm_loadu_si32(input_x)); in xnn_qs8_vadd_minmax_ukernel__xop_mul32_ld32_x16()
89 const __m128i vy0123 = _mm_cvtepi8_epi32(_mm_loadu_si32(input_y)); in xnn_qs8_vadd_minmax_ukernel__xop_mul32_ld32_x16()
[all …]
Dminmax-sse41-mul32-ld32-x16.c36 const __m128i vx0123 = _mm_cvtepi8_epi32(_mm_loadu_si32(input_x)); in xnn_qs8_vadd_minmax_ukernel__sse41_mul32_ld32_x16()
37 const __m128i vy0123 = _mm_cvtepi8_epi32(_mm_loadu_si32(input_y)); in xnn_qs8_vadd_minmax_ukernel__sse41_mul32_ld32_x16()
38 const __m128i vx4567 = _mm_cvtepi8_epi32(_mm_loadu_si32(input_x + 4)); in xnn_qs8_vadd_minmax_ukernel__sse41_mul32_ld32_x16()
39 const __m128i vy4567 = _mm_cvtepi8_epi32(_mm_loadu_si32(input_y + 4)); in xnn_qs8_vadd_minmax_ukernel__sse41_mul32_ld32_x16()
40 const __m128i vx89AB = _mm_cvtepi8_epi32(_mm_loadu_si32(input_x + 8)); in xnn_qs8_vadd_minmax_ukernel__sse41_mul32_ld32_x16()
41 const __m128i vy89AB = _mm_cvtepi8_epi32(_mm_loadu_si32(input_y + 8)); in xnn_qs8_vadd_minmax_ukernel__sse41_mul32_ld32_x16()
42 const __m128i vxCDEF = _mm_cvtepi8_epi32(_mm_loadu_si32(input_x + 12)); in xnn_qs8_vadd_minmax_ukernel__sse41_mul32_ld32_x16()
43 const __m128i vyCDEF = _mm_cvtepi8_epi32(_mm_loadu_si32(input_y + 12)); in xnn_qs8_vadd_minmax_ukernel__sse41_mul32_ld32_x16()
83 const __m128i vx0123 = _mm_cvtepi8_epi32(_mm_loadu_si32(input_x)); in xnn_qs8_vadd_minmax_ukernel__sse41_mul32_ld32_x16()
84 const __m128i vy0123 = _mm_cvtepi8_epi32(_mm_loadu_si32(input_y)); in xnn_qs8_vadd_minmax_ukernel__sse41_mul32_ld32_x16()
[all …]
Dminmax-xop-mul32-ld32-x8.c41 const __m128i vx0123 = _mm_cvtepi8_epi32(_mm_loadu_si32(input_x)); in xnn_qs8_vadd_minmax_ukernel__xop_mul32_ld32_x8()
42 const __m128i vy0123 = _mm_cvtepi8_epi32(_mm_loadu_si32(input_y)); in xnn_qs8_vadd_minmax_ukernel__xop_mul32_ld32_x8()
43 const __m128i vx4567 = _mm_cvtepi8_epi32(_mm_loadu_si32(input_x + 4)); in xnn_qs8_vadd_minmax_ukernel__xop_mul32_ld32_x8()
44 const __m128i vy4567 = _mm_cvtepi8_epi32(_mm_loadu_si32(input_y + 4)); in xnn_qs8_vadd_minmax_ukernel__xop_mul32_ld32_x8()
73 const __m128i vx0123 = _mm_cvtepi8_epi32(_mm_loadu_si32(input_x)); in xnn_qs8_vadd_minmax_ukernel__xop_mul32_ld32_x8()
74 const __m128i vy0123 = _mm_cvtepi8_epi32(_mm_loadu_si32(input_y)); in xnn_qs8_vadd_minmax_ukernel__xop_mul32_ld32_x8()
75 const __m128i vx4567 = _mm_cvtepi8_epi32(_mm_loadu_si32(input_x + 4)); in xnn_qs8_vadd_minmax_ukernel__xop_mul32_ld32_x8()
76 const __m128i vy4567 = _mm_cvtepi8_epi32(_mm_loadu_si32(input_y + 4)); in xnn_qs8_vadd_minmax_ukernel__xop_mul32_ld32_x8()
Dminmax-sse41-mul32-ld32-x8.c36 const __m128i vx0123 = _mm_cvtepi8_epi32(_mm_loadu_si32(input_x)); in xnn_qs8_vadd_minmax_ukernel__sse41_mul32_ld32_x8()
37 const __m128i vy0123 = _mm_cvtepi8_epi32(_mm_loadu_si32(input_y)); in xnn_qs8_vadd_minmax_ukernel__sse41_mul32_ld32_x8()
38 const __m128i vx4567 = _mm_cvtepi8_epi32(_mm_loadu_si32(input_x + 4)); in xnn_qs8_vadd_minmax_ukernel__sse41_mul32_ld32_x8()
39 const __m128i vy4567 = _mm_cvtepi8_epi32(_mm_loadu_si32(input_y + 4)); in xnn_qs8_vadd_minmax_ukernel__sse41_mul32_ld32_x8()
68 const __m128i vx0123 = _mm_cvtepi8_epi32(_mm_loadu_si32(input_x)); in xnn_qs8_vadd_minmax_ukernel__sse41_mul32_ld32_x8()
69 const __m128i vy0123 = _mm_cvtepi8_epi32(_mm_loadu_si32(input_y)); in xnn_qs8_vadd_minmax_ukernel__sse41_mul32_ld32_x8()
70 const __m128i vx4567 = _mm_cvtepi8_epi32(_mm_loadu_si32(input_x + 4)); in xnn_qs8_vadd_minmax_ukernel__sse41_mul32_ld32_x8()
71 const __m128i vy4567 = _mm_cvtepi8_epi32(_mm_loadu_si32(input_y + 4)); in xnn_qs8_vadd_minmax_ukernel__sse41_mul32_ld32_x8()
/external/XNNPACK/src/qs8-vaddc/gen/
Dminmax-sse41-mul32-ld32-x32.c37 const __m128i vx0123 = _mm_cvtepi8_epi32(_mm_loadu_si32(input_x)); in xnn_qs8_vaddc_minmax_ukernel__sse41_mul32_ld32_x32()
38 const __m128i vx4567 = _mm_cvtepi8_epi32(_mm_loadu_si32(input_x + 4)); in xnn_qs8_vaddc_minmax_ukernel__sse41_mul32_ld32_x32()
39 const __m128i vx89AB = _mm_cvtepi8_epi32(_mm_loadu_si32(input_x + 8)); in xnn_qs8_vaddc_minmax_ukernel__sse41_mul32_ld32_x32()
40 const __m128i vxCDEF = _mm_cvtepi8_epi32(_mm_loadu_si32(input_x + 12)); in xnn_qs8_vaddc_minmax_ukernel__sse41_mul32_ld32_x32()
41 const __m128i vxGHIJ = _mm_cvtepi8_epi32(_mm_loadu_si32(input_x + 16)); in xnn_qs8_vaddc_minmax_ukernel__sse41_mul32_ld32_x32()
42 const __m128i vxKLMN = _mm_cvtepi8_epi32(_mm_loadu_si32(input_x + 20)); in xnn_qs8_vaddc_minmax_ukernel__sse41_mul32_ld32_x32()
43 const __m128i vxOPQR = _mm_cvtepi8_epi32(_mm_loadu_si32(input_x + 24)); in xnn_qs8_vaddc_minmax_ukernel__sse41_mul32_ld32_x32()
44 const __m128i vxSTUV = _mm_cvtepi8_epi32(_mm_loadu_si32(input_x + 28)); in xnn_qs8_vaddc_minmax_ukernel__sse41_mul32_ld32_x32()
99 const __m128i vx0123 = _mm_cvtepi8_epi32(_mm_loadu_si32(input_x)); in xnn_qs8_vaddc_minmax_ukernel__sse41_mul32_ld32_x32()
100 const __m128i vx4567 = _mm_cvtepi8_epi32(_mm_loadu_si32(input_x + 4)); in xnn_qs8_vaddc_minmax_ukernel__sse41_mul32_ld32_x32()
Dminmax-xop-mul32-ld32-x32.c42 const __m128i vx0123 = _mm_cvtepi8_epi32(_mm_loadu_si32(input_x)); in xnn_qs8_vaddc_minmax_ukernel__xop_mul32_ld32_x32()
43 const __m128i vx4567 = _mm_cvtepi8_epi32(_mm_loadu_si32(input_x + 4)); in xnn_qs8_vaddc_minmax_ukernel__xop_mul32_ld32_x32()
44 const __m128i vx89AB = _mm_cvtepi8_epi32(_mm_loadu_si32(input_x + 8)); in xnn_qs8_vaddc_minmax_ukernel__xop_mul32_ld32_x32()
45 const __m128i vxCDEF = _mm_cvtepi8_epi32(_mm_loadu_si32(input_x + 12)); in xnn_qs8_vaddc_minmax_ukernel__xop_mul32_ld32_x32()
46 const __m128i vxGHIJ = _mm_cvtepi8_epi32(_mm_loadu_si32(input_x + 16)); in xnn_qs8_vaddc_minmax_ukernel__xop_mul32_ld32_x32()
47 const __m128i vxKLMN = _mm_cvtepi8_epi32(_mm_loadu_si32(input_x + 20)); in xnn_qs8_vaddc_minmax_ukernel__xop_mul32_ld32_x32()
48 const __m128i vxOPQR = _mm_cvtepi8_epi32(_mm_loadu_si32(input_x + 24)); in xnn_qs8_vaddc_minmax_ukernel__xop_mul32_ld32_x32()
49 const __m128i vxSTUV = _mm_cvtepi8_epi32(_mm_loadu_si32(input_x + 28)); in xnn_qs8_vaddc_minmax_ukernel__xop_mul32_ld32_x32()
104 const __m128i vx0123 = _mm_cvtepi8_epi32(_mm_loadu_si32(input_x)); in xnn_qs8_vaddc_minmax_ukernel__xop_mul32_ld32_x32()
105 const __m128i vx4567 = _mm_cvtepi8_epi32(_mm_loadu_si32(input_x + 4)); in xnn_qs8_vaddc_minmax_ukernel__xop_mul32_ld32_x32()
Dminmax-xop-mul32-ld32-x24.c42 const __m128i vx0123 = _mm_cvtepi8_epi32(_mm_loadu_si32(input_x)); in xnn_qs8_vaddc_minmax_ukernel__xop_mul32_ld32_x24()
43 const __m128i vx4567 = _mm_cvtepi8_epi32(_mm_loadu_si32(input_x + 4)); in xnn_qs8_vaddc_minmax_ukernel__xop_mul32_ld32_x24()
44 const __m128i vx89AB = _mm_cvtepi8_epi32(_mm_loadu_si32(input_x + 8)); in xnn_qs8_vaddc_minmax_ukernel__xop_mul32_ld32_x24()
45 const __m128i vxCDEF = _mm_cvtepi8_epi32(_mm_loadu_si32(input_x + 12)); in xnn_qs8_vaddc_minmax_ukernel__xop_mul32_ld32_x24()
46 const __m128i vxGHIJ = _mm_cvtepi8_epi32(_mm_loadu_si32(input_x + 16)); in xnn_qs8_vaddc_minmax_ukernel__xop_mul32_ld32_x24()
47 const __m128i vxKLMN = _mm_cvtepi8_epi32(_mm_loadu_si32(input_x + 20)); in xnn_qs8_vaddc_minmax_ukernel__xop_mul32_ld32_x24()
93 const __m128i vx0123 = _mm_cvtepi8_epi32(_mm_loadu_si32(input_x)); in xnn_qs8_vaddc_minmax_ukernel__xop_mul32_ld32_x24()
94 const __m128i vx4567 = _mm_cvtepi8_epi32(_mm_loadu_si32(input_x + 4)); in xnn_qs8_vaddc_minmax_ukernel__xop_mul32_ld32_x24()
Dminmax-sse41-mul32-ld32-x24.c37 const __m128i vx0123 = _mm_cvtepi8_epi32(_mm_loadu_si32(input_x)); in xnn_qs8_vaddc_minmax_ukernel__sse41_mul32_ld32_x24()
38 const __m128i vx4567 = _mm_cvtepi8_epi32(_mm_loadu_si32(input_x + 4)); in xnn_qs8_vaddc_minmax_ukernel__sse41_mul32_ld32_x24()
39 const __m128i vx89AB = _mm_cvtepi8_epi32(_mm_loadu_si32(input_x + 8)); in xnn_qs8_vaddc_minmax_ukernel__sse41_mul32_ld32_x24()
40 const __m128i vxCDEF = _mm_cvtepi8_epi32(_mm_loadu_si32(input_x + 12)); in xnn_qs8_vaddc_minmax_ukernel__sse41_mul32_ld32_x24()
41 const __m128i vxGHIJ = _mm_cvtepi8_epi32(_mm_loadu_si32(input_x + 16)); in xnn_qs8_vaddc_minmax_ukernel__sse41_mul32_ld32_x24()
42 const __m128i vxKLMN = _mm_cvtepi8_epi32(_mm_loadu_si32(input_x + 20)); in xnn_qs8_vaddc_minmax_ukernel__sse41_mul32_ld32_x24()
88 const __m128i vx0123 = _mm_cvtepi8_epi32(_mm_loadu_si32(input_x)); in xnn_qs8_vaddc_minmax_ukernel__sse41_mul32_ld32_x24()
89 const __m128i vx4567 = _mm_cvtepi8_epi32(_mm_loadu_si32(input_x + 4)); in xnn_qs8_vaddc_minmax_ukernel__sse41_mul32_ld32_x24()
Dminmax-sse41-mul32-ld32-x16.c37 const __m128i vx0123 = _mm_cvtepi8_epi32(_mm_loadu_si32(input_x)); in xnn_qs8_vaddc_minmax_ukernel__sse41_mul32_ld32_x16()
38 const __m128i vx4567 = _mm_cvtepi8_epi32(_mm_loadu_si32(input_x + 4)); in xnn_qs8_vaddc_minmax_ukernel__sse41_mul32_ld32_x16()
39 const __m128i vx89AB = _mm_cvtepi8_epi32(_mm_loadu_si32(input_x + 8)); in xnn_qs8_vaddc_minmax_ukernel__sse41_mul32_ld32_x16()
40 const __m128i vxCDEF = _mm_cvtepi8_epi32(_mm_loadu_si32(input_x + 12)); in xnn_qs8_vaddc_minmax_ukernel__sse41_mul32_ld32_x16()
75 const __m128i vx0123 = _mm_cvtepi8_epi32(_mm_loadu_si32(input_x)); in xnn_qs8_vaddc_minmax_ukernel__sse41_mul32_ld32_x16()
76 const __m128i vx4567 = _mm_cvtepi8_epi32(_mm_loadu_si32(input_x + 4)); in xnn_qs8_vaddc_minmax_ukernel__sse41_mul32_ld32_x16()
Dminmax-xop-mul32-ld32-x16.c42 const __m128i vx0123 = _mm_cvtepi8_epi32(_mm_loadu_si32(input_x)); in xnn_qs8_vaddc_minmax_ukernel__xop_mul32_ld32_x16()
43 const __m128i vx4567 = _mm_cvtepi8_epi32(_mm_loadu_si32(input_x + 4)); in xnn_qs8_vaddc_minmax_ukernel__xop_mul32_ld32_x16()
44 const __m128i vx89AB = _mm_cvtepi8_epi32(_mm_loadu_si32(input_x + 8)); in xnn_qs8_vaddc_minmax_ukernel__xop_mul32_ld32_x16()
45 const __m128i vxCDEF = _mm_cvtepi8_epi32(_mm_loadu_si32(input_x + 12)); in xnn_qs8_vaddc_minmax_ukernel__xop_mul32_ld32_x16()
80 const __m128i vx0123 = _mm_cvtepi8_epi32(_mm_loadu_si32(input_x)); in xnn_qs8_vaddc_minmax_ukernel__xop_mul32_ld32_x16()
81 const __m128i vx4567 = _mm_cvtepi8_epi32(_mm_loadu_si32(input_x + 4)); in xnn_qs8_vaddc_minmax_ukernel__xop_mul32_ld32_x16()
Dminmax-sse41-mul32-ld32-x8.c37 const __m128i vx0123 = _mm_cvtepi8_epi32(_mm_loadu_si32(input_x)); in xnn_qs8_vaddc_minmax_ukernel__sse41_mul32_ld32_x8()
38 const __m128i vx4567 = _mm_cvtepi8_epi32(_mm_loadu_si32(input_x + 4)); in xnn_qs8_vaddc_minmax_ukernel__sse41_mul32_ld32_x8()
64 const __m128i vx0123 = _mm_cvtepi8_epi32(_mm_loadu_si32(input_x)); in xnn_qs8_vaddc_minmax_ukernel__sse41_mul32_ld32_x8()
65 const __m128i vx4567 = _mm_cvtepi8_epi32(_mm_loadu_si32(input_x + 4)); in xnn_qs8_vaddc_minmax_ukernel__sse41_mul32_ld32_x8()
Dminmax-xop-mul32-ld32-x8.c42 const __m128i vx0123 = _mm_cvtepi8_epi32(_mm_loadu_si32(input_x)); in xnn_qs8_vaddc_minmax_ukernel__xop_mul32_ld32_x8()
43 const __m128i vx4567 = _mm_cvtepi8_epi32(_mm_loadu_si32(input_x + 4)); in xnn_qs8_vaddc_minmax_ukernel__xop_mul32_ld32_x8()
69 const __m128i vx0123 = _mm_cvtepi8_epi32(_mm_loadu_si32(input_x)); in xnn_qs8_vaddc_minmax_ukernel__xop_mul32_ld32_x8()
70 const __m128i vx4567 = _mm_cvtepi8_epi32(_mm_loadu_si32(input_x + 4)); in xnn_qs8_vaddc_minmax_ukernel__xop_mul32_ld32_x8()
/external/XNNPACK/src/qs8-vadd/
Dsse-mul32-ld32.c.in44 const __m128i vx${ABC[0:4]} = _mm_cvtepi8_epi32(_mm_loadu_si32(input_x));
45 const __m128i vy${ABC[0:4]} = _mm_cvtepi8_epi32(_mm_loadu_si32(input_y));
47 const __m128i vx${ABC[N:N+4]} = _mm_cvtepi8_epi32(_mm_loadu_si32(input_x + ${N}));
48 const __m128i vy${ABC[N:N+4]} = _mm_cvtepi8_epi32(_mm_loadu_si32(input_y + ${N}));
99 const __m128i vx${ABC[0:4]} = _mm_cvtepi8_epi32(_mm_loadu_si32(input_x));
100 const __m128i vy${ABC[0:4]} = _mm_cvtepi8_epi32(_mm_loadu_si32(input_y));
101 const __m128i vx${ABC[4:8]} = _mm_cvtepi8_epi32(_mm_loadu_si32(input_x + 4));
102 const __m128i vy${ABC[4:8]} = _mm_cvtepi8_epi32(_mm_loadu_si32(input_y + 4));
/external/libopus/celt/x86/
Dx86cpu.h79 (_mm_cvtepi8_epi32(_mm_cvtsi32_si128(*(int *)(x))))
82 (_mm_cvtepi8_epi32(*(__m128i *)(x)))
/external/XNNPACK/src/qs8-vaddc/
Dsse-mul32-ld32.c.in45 const __m128i vx${ABC[0:4]} = _mm_cvtepi8_epi32(_mm_loadu_si32(input_x));
47 const __m128i vx${ABC[N:N+4]} = _mm_cvtepi8_epi32(_mm_loadu_si32(input_x + ${N}));
92 const __m128i vx${ABC[0:4]} = _mm_cvtepi8_epi32(_mm_loadu_si32(input_x));
93 const __m128i vx${ABC[4:8]} = _mm_cvtepi8_epi32(_mm_loadu_si32(input_x + 4));
/external/tensorflow/tensorflow/lite/kernels/internal/optimized/
Dsse_tensor_utils.cc142 const __m128i vec_32x4 = _mm_cvtepi8_epi32( in SseMatrixBatchVectorMultiplyAccumulateImpl()
144 const __m128i row_32x4 = _mm_cvtepi8_epi32( in SseMatrixBatchVectorMultiplyAccumulateImpl()
/external/clang/test/CodeGen/
Dsse41-builtins.c89 return _mm_cvtepi8_epi32(a); in test_mm_cvtepi8_epi32()
/external/llvm-project/clang/test/CodeGen/X86/
Dsse41-builtins.c87 return _mm_cvtepi8_epi32(a); in test_mm_cvtepi8_epi32()
/external/libaom/libaom/av1/common/x86/
Dselfguided_sse4.c188 mask[idx] = _mm_cvtepi8_epi32(_mm_srl_epi64(ones32, shift)); in calc_ab()
329 mask[idx] = _mm_cvtepi8_epi32(_mm_srl_epi64(ones32, shift)); in calc_ab_fast()
/external/llvm-project/clang/lib/Headers/
Dsmmintrin.h1251 _mm_cvtepi8_epi32(__m128i __V) in _mm_cvtepi8_epi32() function
/external/clang/lib/Headers/
Dsmmintrin.h290 _mm_cvtepi8_epi32(__m128i __V) in _mm_cvtepi8_epi32() function

12