Home
last modified time | relevance | path

Searched refs:_mm512_broadcast_i32x4 (Results 1 – 22 of 22) sorted by relevance

/external/XNNPACK/src/qs8-gemm/gen/
D2x16c8-minmax-avx512skx.c52 …const __m512i vmultiplier = _mm512_broadcast_i32x4(_mm_load_si128((const __m128i*) params->sse2.mu… in xnn_qs8_gemm_minmax_ukernel_2x16c8__avx512skx()
53 …const __m512i vrounding = _mm512_broadcast_i32x4(_mm_load_si128((const __m128i*) params->sse2.roun… in xnn_qs8_gemm_minmax_ukernel_2x16c8__avx512skx()
54 …const __m512i vremainder_mask = _mm512_broadcast_i32x4(_mm_load_si128((const __m128i*) params->sse… in xnn_qs8_gemm_minmax_ukernel_2x16c8__avx512skx()
55 …const __m512i vremainder_threshold = _mm512_broadcast_i32x4(_mm_load_si128((const __m128i*) params… in xnn_qs8_gemm_minmax_ukernel_2x16c8__avx512skx()
57 …const __m512i voutput_zero_point = _mm512_broadcast_i32x4(_mm_load_si128((const __m128i*) params->… in xnn_qs8_gemm_minmax_ukernel_2x16c8__avx512skx()
58 …const __m512i voutput_min = _mm512_broadcast_i32x4(_mm_load_si128((const __m128i*) params->sse2.ou… in xnn_qs8_gemm_minmax_ukernel_2x16c8__avx512skx()
59 …const __m512i voutput_max = _mm512_broadcast_i32x4(_mm_load_si128((const __m128i*) params->sse2.ou… in xnn_qs8_gemm_minmax_ukernel_2x16c8__avx512skx()
73 …const __m512i va0 = _mm512_broadcast_i32x4(_mm_cvtepi8_epi16(_mm_loadl_epi64((const __m128i*) a0))… in xnn_qs8_gemm_minmax_ukernel_2x16c8__avx512skx()
75 …const __m512i va1 = _mm512_broadcast_i32x4(_mm_cvtepi8_epi16(_mm_loadl_epi64((const __m128i*) a1))… in xnn_qs8_gemm_minmax_ukernel_2x16c8__avx512skx()
D3x16c8-minmax-avx512skx.c58 …const __m512i vmultiplier = _mm512_broadcast_i32x4(_mm_load_si128((const __m128i*) params->sse2.mu… in xnn_qs8_gemm_minmax_ukernel_3x16c8__avx512skx()
59 …const __m512i vrounding = _mm512_broadcast_i32x4(_mm_load_si128((const __m128i*) params->sse2.roun… in xnn_qs8_gemm_minmax_ukernel_3x16c8__avx512skx()
60 …const __m512i vremainder_mask = _mm512_broadcast_i32x4(_mm_load_si128((const __m128i*) params->sse… in xnn_qs8_gemm_minmax_ukernel_3x16c8__avx512skx()
61 …const __m512i vremainder_threshold = _mm512_broadcast_i32x4(_mm_load_si128((const __m128i*) params… in xnn_qs8_gemm_minmax_ukernel_3x16c8__avx512skx()
63 …const __m512i voutput_zero_point = _mm512_broadcast_i32x4(_mm_load_si128((const __m128i*) params->… in xnn_qs8_gemm_minmax_ukernel_3x16c8__avx512skx()
64 …const __m512i voutput_min = _mm512_broadcast_i32x4(_mm_load_si128((const __m128i*) params->sse2.ou… in xnn_qs8_gemm_minmax_ukernel_3x16c8__avx512skx()
65 …const __m512i voutput_max = _mm512_broadcast_i32x4(_mm_load_si128((const __m128i*) params->sse2.ou… in xnn_qs8_gemm_minmax_ukernel_3x16c8__avx512skx()
83 …const __m512i va0 = _mm512_broadcast_i32x4(_mm_cvtepi8_epi16(_mm_loadl_epi64((const __m128i*) a0))… in xnn_qs8_gemm_minmax_ukernel_3x16c8__avx512skx()
85 …const __m512i va1 = _mm512_broadcast_i32x4(_mm_cvtepi8_epi16(_mm_loadl_epi64((const __m128i*) a1))… in xnn_qs8_gemm_minmax_ukernel_3x16c8__avx512skx()
87 …const __m512i va2 = _mm512_broadcast_i32x4(_mm_cvtepi8_epi16(_mm_loadl_epi64((const __m128i*) a2))… in xnn_qs8_gemm_minmax_ukernel_3x16c8__avx512skx()
D4x16c8-minmax-avx512skx.c64 …const __m512i vmultiplier = _mm512_broadcast_i32x4(_mm_load_si128((const __m128i*) params->sse2.mu… in xnn_qs8_gemm_minmax_ukernel_4x16c8__avx512skx()
65 …const __m512i vrounding = _mm512_broadcast_i32x4(_mm_load_si128((const __m128i*) params->sse2.roun… in xnn_qs8_gemm_minmax_ukernel_4x16c8__avx512skx()
66 …const __m512i vremainder_mask = _mm512_broadcast_i32x4(_mm_load_si128((const __m128i*) params->sse… in xnn_qs8_gemm_minmax_ukernel_4x16c8__avx512skx()
67 …const __m512i vremainder_threshold = _mm512_broadcast_i32x4(_mm_load_si128((const __m128i*) params… in xnn_qs8_gemm_minmax_ukernel_4x16c8__avx512skx()
69 …const __m512i voutput_zero_point = _mm512_broadcast_i32x4(_mm_load_si128((const __m128i*) params->… in xnn_qs8_gemm_minmax_ukernel_4x16c8__avx512skx()
70 …const __m512i voutput_min = _mm512_broadcast_i32x4(_mm_load_si128((const __m128i*) params->sse2.ou… in xnn_qs8_gemm_minmax_ukernel_4x16c8__avx512skx()
71 …const __m512i voutput_max = _mm512_broadcast_i32x4(_mm_load_si128((const __m128i*) params->sse2.ou… in xnn_qs8_gemm_minmax_ukernel_4x16c8__avx512skx()
93 …const __m512i va0 = _mm512_broadcast_i32x4(_mm_cvtepi8_epi16(_mm_loadl_epi64((const __m128i*) a0))… in xnn_qs8_gemm_minmax_ukernel_4x16c8__avx512skx()
95 …const __m512i va1 = _mm512_broadcast_i32x4(_mm_cvtepi8_epi16(_mm_loadl_epi64((const __m128i*) a1))… in xnn_qs8_gemm_minmax_ukernel_4x16c8__avx512skx()
97 …const __m512i va2 = _mm512_broadcast_i32x4(_mm_cvtepi8_epi16(_mm_loadl_epi64((const __m128i*) a2))… in xnn_qs8_gemm_minmax_ukernel_4x16c8__avx512skx()
[all …]
D1x16c8-minmax-avx512skx.c46 …const __m512i vmultiplier = _mm512_broadcast_i32x4(_mm_load_si128((const __m128i*) params->sse2.mu… in xnn_qs8_gemm_minmax_ukernel_1x16c8__avx512skx()
47 …const __m512i vrounding = _mm512_broadcast_i32x4(_mm_load_si128((const __m128i*) params->sse2.roun… in xnn_qs8_gemm_minmax_ukernel_1x16c8__avx512skx()
48 …const __m512i vremainder_mask = _mm512_broadcast_i32x4(_mm_load_si128((const __m128i*) params->sse… in xnn_qs8_gemm_minmax_ukernel_1x16c8__avx512skx()
49 …const __m512i vremainder_threshold = _mm512_broadcast_i32x4(_mm_load_si128((const __m128i*) params… in xnn_qs8_gemm_minmax_ukernel_1x16c8__avx512skx()
63 …const __m512i va0 = _mm512_broadcast_i32x4(_mm_cvtepi8_epi16(_mm_loadl_epi64((const __m128i*) a0))… in xnn_qs8_gemm_minmax_ukernel_1x16c8__avx512skx()
/external/XNNPACK/src/qs8-igemm/gen/
D2x16c8-minmax-avx512skx.c51 …const __m512i vmultiplier = _mm512_broadcast_i32x4(_mm_load_si128((const __m128i*) params->sse2.mu… in xnn_qs8_igemm_minmax_ukernel_2x16c8__avx512skx()
52 …const __m512i vrounding = _mm512_broadcast_i32x4(_mm_load_si128((const __m128i*) params->sse2.roun… in xnn_qs8_igemm_minmax_ukernel_2x16c8__avx512skx()
53 …const __m512i vremainder_mask = _mm512_broadcast_i32x4(_mm_load_si128((const __m128i*) params->sse… in xnn_qs8_igemm_minmax_ukernel_2x16c8__avx512skx()
54 …const __m512i vremainder_threshold = _mm512_broadcast_i32x4(_mm_load_si128((const __m128i*) params… in xnn_qs8_igemm_minmax_ukernel_2x16c8__avx512skx()
56 …const __m512i voutput_zero_point = _mm512_broadcast_i32x4(_mm_load_si128((const __m128i*) params->… in xnn_qs8_igemm_minmax_ukernel_2x16c8__avx512skx()
57 …const __m512i voutput_min = _mm512_broadcast_i32x4(_mm_load_si128((const __m128i*) params->sse2.ou… in xnn_qs8_igemm_minmax_ukernel_2x16c8__avx512skx()
58 …const __m512i voutput_max = _mm512_broadcast_i32x4(_mm_load_si128((const __m128i*) params->sse2.ou… in xnn_qs8_igemm_minmax_ukernel_2x16c8__avx512skx()
84 …const __m512i va0 = _mm512_broadcast_i32x4(_mm_cvtepi8_epi16(_mm_loadl_epi64((const __m128i*) a0))… in xnn_qs8_igemm_minmax_ukernel_2x16c8__avx512skx()
86 …const __m512i va1 = _mm512_broadcast_i32x4(_mm_cvtepi8_epi16(_mm_loadl_epi64((const __m128i*) a1))… in xnn_qs8_igemm_minmax_ukernel_2x16c8__avx512skx()
D3x16c8-minmax-avx512skx.c55 …const __m512i vmultiplier = _mm512_broadcast_i32x4(_mm_load_si128((const __m128i*) params->sse2.mu… in xnn_qs8_igemm_minmax_ukernel_3x16c8__avx512skx()
56 …const __m512i vrounding = _mm512_broadcast_i32x4(_mm_load_si128((const __m128i*) params->sse2.roun… in xnn_qs8_igemm_minmax_ukernel_3x16c8__avx512skx()
57 …const __m512i vremainder_mask = _mm512_broadcast_i32x4(_mm_load_si128((const __m128i*) params->sse… in xnn_qs8_igemm_minmax_ukernel_3x16c8__avx512skx()
58 …const __m512i vremainder_threshold = _mm512_broadcast_i32x4(_mm_load_si128((const __m128i*) params… in xnn_qs8_igemm_minmax_ukernel_3x16c8__avx512skx()
60 …const __m512i voutput_zero_point = _mm512_broadcast_i32x4(_mm_load_si128((const __m128i*) params->… in xnn_qs8_igemm_minmax_ukernel_3x16c8__avx512skx()
61 …const __m512i voutput_min = _mm512_broadcast_i32x4(_mm_load_si128((const __m128i*) params->sse2.ou… in xnn_qs8_igemm_minmax_ukernel_3x16c8__avx512skx()
62 …const __m512i voutput_max = _mm512_broadcast_i32x4(_mm_load_si128((const __m128i*) params->sse2.ou… in xnn_qs8_igemm_minmax_ukernel_3x16c8__avx512skx()
96 …const __m512i va0 = _mm512_broadcast_i32x4(_mm_cvtepi8_epi16(_mm_loadl_epi64((const __m128i*) a0))… in xnn_qs8_igemm_minmax_ukernel_3x16c8__avx512skx()
98 …const __m512i va1 = _mm512_broadcast_i32x4(_mm_cvtepi8_epi16(_mm_loadl_epi64((const __m128i*) a1))… in xnn_qs8_igemm_minmax_ukernel_3x16c8__avx512skx()
100 …const __m512i va2 = _mm512_broadcast_i32x4(_mm_cvtepi8_epi16(_mm_loadl_epi64((const __m128i*) a2))… in xnn_qs8_igemm_minmax_ukernel_3x16c8__avx512skx()
D4x16c8-minmax-avx512skx.c59 …const __m512i vmultiplier = _mm512_broadcast_i32x4(_mm_load_si128((const __m128i*) params->sse2.mu… in xnn_qs8_igemm_minmax_ukernel_4x16c8__avx512skx()
60 …const __m512i vrounding = _mm512_broadcast_i32x4(_mm_load_si128((const __m128i*) params->sse2.roun… in xnn_qs8_igemm_minmax_ukernel_4x16c8__avx512skx()
61 …const __m512i vremainder_mask = _mm512_broadcast_i32x4(_mm_load_si128((const __m128i*) params->sse… in xnn_qs8_igemm_minmax_ukernel_4x16c8__avx512skx()
62 …const __m512i vremainder_threshold = _mm512_broadcast_i32x4(_mm_load_si128((const __m128i*) params… in xnn_qs8_igemm_minmax_ukernel_4x16c8__avx512skx()
64 …const __m512i voutput_zero_point = _mm512_broadcast_i32x4(_mm_load_si128((const __m128i*) params->… in xnn_qs8_igemm_minmax_ukernel_4x16c8__avx512skx()
65 …const __m512i voutput_min = _mm512_broadcast_i32x4(_mm_load_si128((const __m128i*) params->sse2.ou… in xnn_qs8_igemm_minmax_ukernel_4x16c8__avx512skx()
66 …const __m512i voutput_max = _mm512_broadcast_i32x4(_mm_load_si128((const __m128i*) params->sse2.ou… in xnn_qs8_igemm_minmax_ukernel_4x16c8__avx512skx()
108 …const __m512i va0 = _mm512_broadcast_i32x4(_mm_cvtepi8_epi16(_mm_loadl_epi64((const __m128i*) a0))… in xnn_qs8_igemm_minmax_ukernel_4x16c8__avx512skx()
110 …const __m512i va1 = _mm512_broadcast_i32x4(_mm_cvtepi8_epi16(_mm_loadl_epi64((const __m128i*) a1))… in xnn_qs8_igemm_minmax_ukernel_4x16c8__avx512skx()
112 …const __m512i va2 = _mm512_broadcast_i32x4(_mm_cvtepi8_epi16(_mm_loadl_epi64((const __m128i*) a2))… in xnn_qs8_igemm_minmax_ukernel_4x16c8__avx512skx()
[all …]
D1x16c8-minmax-avx512skx.c47 …const __m512i vmultiplier = _mm512_broadcast_i32x4(_mm_load_si128((const __m128i*) params->sse2.mu… in xnn_qs8_igemm_minmax_ukernel_1x16c8__avx512skx()
48 …const __m512i vrounding = _mm512_broadcast_i32x4(_mm_load_si128((const __m128i*) params->sse2.roun… in xnn_qs8_igemm_minmax_ukernel_1x16c8__avx512skx()
49 …const __m512i vremainder_mask = _mm512_broadcast_i32x4(_mm_load_si128((const __m128i*) params->sse… in xnn_qs8_igemm_minmax_ukernel_1x16c8__avx512skx()
50 …const __m512i vremainder_threshold = _mm512_broadcast_i32x4(_mm_load_si128((const __m128i*) params… in xnn_qs8_igemm_minmax_ukernel_1x16c8__avx512skx()
72 …const __m512i va0 = _mm512_broadcast_i32x4(_mm_cvtepi8_epi16(_mm_loadl_epi64((const __m128i*) a0))… in xnn_qs8_igemm_minmax_ukernel_1x16c8__avx512skx()
/external/XNNPACK/src/qs8-gemm/
DMRx16c8-avx512skx.c.in64 …const __m512i vmultiplier = _mm512_broadcast_i32x4(_mm_load_si128((const __m128i*) params->sse2.mu…
65 …const __m512i vrounding = _mm512_broadcast_i32x4(_mm_load_si128((const __m128i*) params->sse2.roun…
66 …const __m512i vremainder_mask = _mm512_broadcast_i32x4(_mm_load_si128((const __m128i*) params->sse…
67 …const __m512i vremainder_threshold = _mm512_broadcast_i32x4(_mm_load_si128((const __m128i*) params…
70 …const __m512i voutput_zero_point = _mm512_broadcast_i32x4(_mm_load_si128((const __m128i*) params->…
71 …const __m512i voutput_min = _mm512_broadcast_i32x4(_mm_load_si128((const __m128i*) params->sse2.ou…
72 …const __m512i voutput_max = _mm512_broadcast_i32x4(_mm_load_si128((const __m128i*) params->sse2.ou…
89 …const __m512i va${M} = _mm512_broadcast_i32x4(_mm_cvtepi8_epi16(_mm_loadl_epi64((const __m128i*) a…
/external/XNNPACK/src/qs8-igemm/
DMRx16c8-avx512skx.c.in61 …const __m512i vmultiplier = _mm512_broadcast_i32x4(_mm_load_si128((const __m128i*) params->sse2.mu…
62 …const __m512i vrounding = _mm512_broadcast_i32x4(_mm_load_si128((const __m128i*) params->sse2.roun…
63 …const __m512i vremainder_mask = _mm512_broadcast_i32x4(_mm_load_si128((const __m128i*) params->sse…
64 …const __m512i vremainder_threshold = _mm512_broadcast_i32x4(_mm_load_si128((const __m128i*) params…
67 …const __m512i voutput_zero_point = _mm512_broadcast_i32x4(_mm_load_si128((const __m128i*) params->…
68 …const __m512i voutput_min = _mm512_broadcast_i32x4(_mm_load_si128((const __m128i*) params->sse2.ou…
69 …const __m512i voutput_max = _mm512_broadcast_i32x4(_mm_load_si128((const __m128i*) params->sse2.ou…
95 …const __m512i va${M} = _mm512_broadcast_i32x4(_mm_cvtepi8_epi16(_mm_loadl_epi64((const __m128i*) a…
/external/XNNPACK/src/qs8-dwconv/gen/
Dup32x9-minmax-avx512skx-mul32.c34 …const __m512i vmultiplier = _mm512_broadcast_i32x4(_mm_load_si128((const __m128i*) params->sse2.mu… in xnn_qs8_dwconv_minmax_ukernel_up32x9__avx512skx_mul32()
35 …const __m512i vrounding = _mm512_broadcast_i32x4(_mm_load_si128((const __m128i*) params->sse2.roun… in xnn_qs8_dwconv_minmax_ukernel_up32x9__avx512skx_mul32()
36 …const __m512i vremainder_mask = _mm512_broadcast_i32x4(_mm_load_si128((const __m128i*) params->sse… in xnn_qs8_dwconv_minmax_ukernel_up32x9__avx512skx_mul32()
37 …const __m512i vremainder_threshold = _mm512_broadcast_i32x4(_mm_load_si128((const __m128i*) params… in xnn_qs8_dwconv_minmax_ukernel_up32x9__avx512skx_mul32()
39 …const __m512i voutput_zero_point = _mm512_broadcast_i32x4(_mm_load_si128((const __m128i*) params->… in xnn_qs8_dwconv_minmax_ukernel_up32x9__avx512skx_mul32()
40 …const __m512i voutput_min = _mm512_broadcast_i32x4(_mm_load_si128((const __m128i*) params->sse2.ou… in xnn_qs8_dwconv_minmax_ukernel_up32x9__avx512skx_mul32()
41 …const __m512i voutput_max = _mm512_broadcast_i32x4(_mm_load_si128((const __m128i*) params->sse2.ou… in xnn_qs8_dwconv_minmax_ukernel_up32x9__avx512skx_mul32()
Dup16x9-minmax-avx512skx-mul32.c34 …const __m512i vmultiplier = _mm512_broadcast_i32x4(_mm_load_si128((const __m128i*) params->sse2.mu… in xnn_qs8_dwconv_minmax_ukernel_up16x9__avx512skx_mul32()
35 …const __m512i vrounding = _mm512_broadcast_i32x4(_mm_load_si128((const __m128i*) params->sse2.roun… in xnn_qs8_dwconv_minmax_ukernel_up16x9__avx512skx_mul32()
36 …const __m512i vremainder_mask = _mm512_broadcast_i32x4(_mm_load_si128((const __m128i*) params->sse… in xnn_qs8_dwconv_minmax_ukernel_up16x9__avx512skx_mul32()
37 …const __m512i vremainder_threshold = _mm512_broadcast_i32x4(_mm_load_si128((const __m128i*) params… in xnn_qs8_dwconv_minmax_ukernel_up16x9__avx512skx_mul32()
/external/XNNPACK/src/qs8-dwconv/
Dunipass-avx512skx-mul32.c.in34 …const __m512i vmultiplier = _mm512_broadcast_i32x4(_mm_load_si128((const __m128i*) params->sse2.mu…
35 …const __m512i vrounding = _mm512_broadcast_i32x4(_mm_load_si128((const __m128i*) params->sse2.roun…
36 …const __m512i vremainder_mask = _mm512_broadcast_i32x4(_mm_load_si128((const __m128i*) params->sse…
37 …const __m512i vremainder_threshold = _mm512_broadcast_i32x4(_mm_load_si128((const __m128i*) params…
40 …const __m512i voutput_zero_point = _mm512_broadcast_i32x4(_mm_load_si128((const __m128i*) params->…
41 …const __m512i voutput_min = _mm512_broadcast_i32x4(_mm_load_si128((const __m128i*) params->sse2.ou…
42 …const __m512i voutput_max = _mm512_broadcast_i32x4(_mm_load_si128((const __m128i*) params->sse2.ou…
/external/XNNPACK/src/f32-vunary/
Davx512f.c.in66 …const __m512i vnonsign_mask = _mm512_broadcast_i32x4(_mm_load_si128((const __m128i*) params->sse.n…
68 …const __m512i vsign_mask = _mm512_broadcast_i32x4(_mm_load_si128((const __m128i*) params->sse.sign…
/external/XNNPACK/src/f32-vunary/gen/
Dvabs-avx512f-x16.c30 …const __m512i vnonsign_mask = _mm512_broadcast_i32x4(_mm_load_si128((const __m128i*) params->sse.n… in xnn_f32_vabs_ukernel__avx512f_x16()
Dvneg-avx512f-x16.c30 …const __m512i vsign_mask = _mm512_broadcast_i32x4(_mm_load_si128((const __m128i*) params->sse.sign… in xnn_f32_vneg_ukernel__avx512f_x16()
Dvabs-avx512f-x32.c30 …const __m512i vnonsign_mask = _mm512_broadcast_i32x4(_mm_load_si128((const __m128i*) params->sse.n… in xnn_f32_vabs_ukernel__avx512f_x32()
Dvneg-avx512f-x32.c30 …const __m512i vsign_mask = _mm512_broadcast_i32x4(_mm_load_si128((const __m128i*) params->sse.sign… in xnn_f32_vneg_ukernel__avx512f_x32()
/external/llvm-project/clang/lib/Headers/
Davx512fintrin.h6852 _mm512_broadcast_i32x4(__m128i __A) in _mm512_broadcast_i32x4() function
6863 (__v16si)_mm512_broadcast_i32x4(__A), in _mm512_mask_broadcast_i32x4()
6871 (__v16si)_mm512_broadcast_i32x4(__A), in _mm512_maskz_broadcast_i32x4()
/external/clang/test/CodeGen/
Davx512f-builtins.c4372 return _mm512_broadcast_i32x4(__A); in test_mm512_broadcast_i32x4()
/external/llvm-project/clang/test/CodeGen/X86/
Davx512f-builtins.c6295 return _mm512_broadcast_i32x4(_mm_loadu_si128(__A)); in test_mm512_broadcast_i32x4()
/external/clang/lib/Headers/
Davx512fintrin.h7314 _mm512_broadcast_i32x4 (__m128i __A) in _mm512_broadcast_i32x4() function