/external/XNNPACK/src/f32-gemm/gen/ |
D | 1x16-avx512f-broadcast.c | 43 __m512 vacc0x0123456789ABCDEF = _mm512_load_ps(w); in xnn_f32_gemm_ukernel_1x16__avx512f_broadcast() local 51 …vacc0x0123456789ABCDEF = _mm512_fmadd_ps(_mm512_set1_ps(*a0), vb0123456789ABCDEF, vacc0x0123456789… in xnn_f32_gemm_ukernel_1x16__avx512f_broadcast() 59 vacc0x0123456789ABCDEF = _mm512_min_ps(vacc0x0123456789ABCDEF, vmax); in xnn_f32_gemm_ukernel_1x16__avx512f_broadcast() 62 vacc0x0123456789ABCDEF = _mm512_max_ps(vacc0x0123456789ABCDEF, vmin); in xnn_f32_gemm_ukernel_1x16__avx512f_broadcast() 65 _mm512_storeu_ps(c0, vacc0x0123456789ABCDEF); in xnn_f32_gemm_ukernel_1x16__avx512f_broadcast() 76 _mm512_mask_storeu_ps(c0, vmask, vacc0x0123456789ABCDEF); in xnn_f32_gemm_ukernel_1x16__avx512f_broadcast()
|
D | 5x16-avx512f-broadcast.c | 67 __m512 vacc0x0123456789ABCDEF = _mm512_load_ps(w); in xnn_f32_gemm_ukernel_5x16__avx512f_broadcast() local 68 __m512 vacc1x0123456789ABCDEF = vacc0x0123456789ABCDEF; in xnn_f32_gemm_ukernel_5x16__avx512f_broadcast() 69 __m512 vacc2x0123456789ABCDEF = vacc0x0123456789ABCDEF; in xnn_f32_gemm_ukernel_5x16__avx512f_broadcast() 70 __m512 vacc3x0123456789ABCDEF = vacc0x0123456789ABCDEF; in xnn_f32_gemm_ukernel_5x16__avx512f_broadcast() 71 __m512 vacc4x0123456789ABCDEF = vacc0x0123456789ABCDEF; in xnn_f32_gemm_ukernel_5x16__avx512f_broadcast() 79 …vacc0x0123456789ABCDEF = _mm512_fmadd_ps(_mm512_set1_ps(*a0), vb0123456789ABCDEF, vacc0x0123456789… in xnn_f32_gemm_ukernel_5x16__avx512f_broadcast() 95 vacc0x0123456789ABCDEF = _mm512_min_ps(vacc0x0123456789ABCDEF, vmax); in xnn_f32_gemm_ukernel_5x16__avx512f_broadcast() 102 vacc0x0123456789ABCDEF = _mm512_max_ps(vacc0x0123456789ABCDEF, vmin); in xnn_f32_gemm_ukernel_5x16__avx512f_broadcast() 117 _mm512_storeu_ps(c0, vacc0x0123456789ABCDEF); in xnn_f32_gemm_ukernel_5x16__avx512f_broadcast() 136 _mm512_mask_storeu_ps(c0, vmask, vacc0x0123456789ABCDEF); in xnn_f32_gemm_ukernel_5x16__avx512f_broadcast()
|
D | 4x16-avx512f-broadcast.c | 61 __m512 vacc0x0123456789ABCDEF = _mm512_load_ps(w); in xnn_f32_gemm_ukernel_4x16__avx512f_broadcast() local 62 __m512 vacc1x0123456789ABCDEF = vacc0x0123456789ABCDEF; in xnn_f32_gemm_ukernel_4x16__avx512f_broadcast() 63 __m512 vacc2x0123456789ABCDEF = vacc0x0123456789ABCDEF; in xnn_f32_gemm_ukernel_4x16__avx512f_broadcast() 64 __m512 vacc3x0123456789ABCDEF = vacc0x0123456789ABCDEF; in xnn_f32_gemm_ukernel_4x16__avx512f_broadcast() 72 …vacc0x0123456789ABCDEF = _mm512_fmadd_ps(_mm512_set1_ps(*a0), vb0123456789ABCDEF, vacc0x0123456789… in xnn_f32_gemm_ukernel_4x16__avx512f_broadcast() 86 vacc0x0123456789ABCDEF = _mm512_min_ps(vacc0x0123456789ABCDEF, vmax); in xnn_f32_gemm_ukernel_4x16__avx512f_broadcast() 92 vacc0x0123456789ABCDEF = _mm512_max_ps(vacc0x0123456789ABCDEF, vmin); in xnn_f32_gemm_ukernel_4x16__avx512f_broadcast() 104 _mm512_storeu_ps(c0, vacc0x0123456789ABCDEF); in xnn_f32_gemm_ukernel_4x16__avx512f_broadcast() 121 _mm512_mask_storeu_ps(c0, vmask, vacc0x0123456789ABCDEF); in xnn_f32_gemm_ukernel_4x16__avx512f_broadcast()
|
D | 7x16-avx512f-broadcast.c | 79 __m512 vacc0x0123456789ABCDEF = _mm512_load_ps(w); in xnn_f32_gemm_ukernel_7x16__avx512f_broadcast() local 80 __m512 vacc1x0123456789ABCDEF = vacc0x0123456789ABCDEF; in xnn_f32_gemm_ukernel_7x16__avx512f_broadcast() 81 __m512 vacc2x0123456789ABCDEF = vacc0x0123456789ABCDEF; in xnn_f32_gemm_ukernel_7x16__avx512f_broadcast() 82 __m512 vacc3x0123456789ABCDEF = vacc0x0123456789ABCDEF; in xnn_f32_gemm_ukernel_7x16__avx512f_broadcast() 83 __m512 vacc4x0123456789ABCDEF = vacc0x0123456789ABCDEF; in xnn_f32_gemm_ukernel_7x16__avx512f_broadcast() 84 __m512 vacc5x0123456789ABCDEF = vacc0x0123456789ABCDEF; in xnn_f32_gemm_ukernel_7x16__avx512f_broadcast() 85 __m512 vacc6x0123456789ABCDEF = vacc0x0123456789ABCDEF; in xnn_f32_gemm_ukernel_7x16__avx512f_broadcast() 93 …vacc0x0123456789ABCDEF = _mm512_fmadd_ps(_mm512_set1_ps(*a0), vb0123456789ABCDEF, vacc0x0123456789… in xnn_f32_gemm_ukernel_7x16__avx512f_broadcast() 113 vacc0x0123456789ABCDEF = _mm512_min_ps(vacc0x0123456789ABCDEF, vmax); in xnn_f32_gemm_ukernel_7x16__avx512f_broadcast() 122 vacc0x0123456789ABCDEF = _mm512_max_ps(vacc0x0123456789ABCDEF, vmin); in xnn_f32_gemm_ukernel_7x16__avx512f_broadcast() [all …]
|
D | 6x16-avx512f-broadcast.c | 73 __m512 vacc0x0123456789ABCDEF = _mm512_load_ps(w); in xnn_f32_gemm_ukernel_6x16__avx512f_broadcast() local 74 __m512 vacc1x0123456789ABCDEF = vacc0x0123456789ABCDEF; in xnn_f32_gemm_ukernel_6x16__avx512f_broadcast() 75 __m512 vacc2x0123456789ABCDEF = vacc0x0123456789ABCDEF; in xnn_f32_gemm_ukernel_6x16__avx512f_broadcast() 76 __m512 vacc3x0123456789ABCDEF = vacc0x0123456789ABCDEF; in xnn_f32_gemm_ukernel_6x16__avx512f_broadcast() 77 __m512 vacc4x0123456789ABCDEF = vacc0x0123456789ABCDEF; in xnn_f32_gemm_ukernel_6x16__avx512f_broadcast() 78 __m512 vacc5x0123456789ABCDEF = vacc0x0123456789ABCDEF; in xnn_f32_gemm_ukernel_6x16__avx512f_broadcast() 86 …vacc0x0123456789ABCDEF = _mm512_fmadd_ps(_mm512_set1_ps(*a0), vb0123456789ABCDEF, vacc0x0123456789… in xnn_f32_gemm_ukernel_6x16__avx512f_broadcast() 104 vacc0x0123456789ABCDEF = _mm512_min_ps(vacc0x0123456789ABCDEF, vmax); in xnn_f32_gemm_ukernel_6x16__avx512f_broadcast() 112 vacc0x0123456789ABCDEF = _mm512_max_ps(vacc0x0123456789ABCDEF, vmin); in xnn_f32_gemm_ukernel_6x16__avx512f_broadcast() 130 _mm512_storeu_ps(c0, vacc0x0123456789ABCDEF); in xnn_f32_gemm_ukernel_6x16__avx512f_broadcast() [all …]
|
D | 8x16-avx512f-broadcast.c | 85 __m512 vacc0x0123456789ABCDEF = _mm512_load_ps(w); in xnn_f32_gemm_ukernel_8x16__avx512f_broadcast() local 86 __m512 vacc1x0123456789ABCDEF = vacc0x0123456789ABCDEF; in xnn_f32_gemm_ukernel_8x16__avx512f_broadcast() 87 __m512 vacc2x0123456789ABCDEF = vacc0x0123456789ABCDEF; in xnn_f32_gemm_ukernel_8x16__avx512f_broadcast() 88 __m512 vacc3x0123456789ABCDEF = vacc0x0123456789ABCDEF; in xnn_f32_gemm_ukernel_8x16__avx512f_broadcast() 89 __m512 vacc4x0123456789ABCDEF = vacc0x0123456789ABCDEF; in xnn_f32_gemm_ukernel_8x16__avx512f_broadcast() 90 __m512 vacc5x0123456789ABCDEF = vacc0x0123456789ABCDEF; in xnn_f32_gemm_ukernel_8x16__avx512f_broadcast() 91 __m512 vacc6x0123456789ABCDEF = vacc0x0123456789ABCDEF; in xnn_f32_gemm_ukernel_8x16__avx512f_broadcast() 92 __m512 vacc7x0123456789ABCDEF = vacc0x0123456789ABCDEF; in xnn_f32_gemm_ukernel_8x16__avx512f_broadcast() 100 …vacc0x0123456789ABCDEF = _mm512_fmadd_ps(_mm512_set1_ps(*a0), vb0123456789ABCDEF, vacc0x0123456789… in xnn_f32_gemm_ukernel_8x16__avx512f_broadcast() 122 vacc0x0123456789ABCDEF = _mm512_min_ps(vacc0x0123456789ABCDEF, vmax); in xnn_f32_gemm_ukernel_8x16__avx512f_broadcast() [all …]
|
/external/XNNPACK/src/f32-gemm/gen-inc/ |
D | 1x16-avx512f-broadcast.c | 45 __m512 vacc0x0123456789ABCDEF = _mm512_load_ps(acc + 0); in xnn_f32_gemminc_ukernel_1x16__avx512f_broadcast() local 53 …vacc0x0123456789ABCDEF = _mm512_fmadd_ps(_mm512_set1_ps(*a0), vb0123456789ABCDEF, vacc0x0123456789… in xnn_f32_gemminc_ukernel_1x16__avx512f_broadcast() 61 vacc0x0123456789ABCDEF = _mm512_min_ps(vacc0x0123456789ABCDEF, vmax); in xnn_f32_gemminc_ukernel_1x16__avx512f_broadcast() 64 vacc0x0123456789ABCDEF = _mm512_max_ps(vacc0x0123456789ABCDEF, vmin); in xnn_f32_gemminc_ukernel_1x16__avx512f_broadcast() 67 _mm512_storeu_ps(c0, vacc0x0123456789ABCDEF); in xnn_f32_gemminc_ukernel_1x16__avx512f_broadcast() 78 _mm512_mask_storeu_ps(c0, vmask, vacc0x0123456789ABCDEF); in xnn_f32_gemminc_ukernel_1x16__avx512f_broadcast()
|
D | 4x16-avx512f-broadcast.c | 63 __m512 vacc0x0123456789ABCDEF = _mm512_load_ps(acc + 0); in xnn_f32_gemminc_ukernel_4x16__avx512f_broadcast() local 74 …vacc0x0123456789ABCDEF = _mm512_fmadd_ps(_mm512_set1_ps(*a0), vb0123456789ABCDEF, vacc0x0123456789… in xnn_f32_gemminc_ukernel_4x16__avx512f_broadcast() 88 vacc0x0123456789ABCDEF = _mm512_min_ps(vacc0x0123456789ABCDEF, vmax); in xnn_f32_gemminc_ukernel_4x16__avx512f_broadcast() 94 vacc0x0123456789ABCDEF = _mm512_max_ps(vacc0x0123456789ABCDEF, vmin); in xnn_f32_gemminc_ukernel_4x16__avx512f_broadcast() 106 _mm512_storeu_ps(c0, vacc0x0123456789ABCDEF); in xnn_f32_gemminc_ukernel_4x16__avx512f_broadcast() 123 _mm512_mask_storeu_ps(c0, vmask, vacc0x0123456789ABCDEF); in xnn_f32_gemminc_ukernel_4x16__avx512f_broadcast()
|
D | 5x16-avx512f-broadcast.c | 69 __m512 vacc0x0123456789ABCDEF = _mm512_load_ps(acc + 0); in xnn_f32_gemminc_ukernel_5x16__avx512f_broadcast() local 81 …vacc0x0123456789ABCDEF = _mm512_fmadd_ps(_mm512_set1_ps(*a0), vb0123456789ABCDEF, vacc0x0123456789… in xnn_f32_gemminc_ukernel_5x16__avx512f_broadcast() 97 vacc0x0123456789ABCDEF = _mm512_min_ps(vacc0x0123456789ABCDEF, vmax); in xnn_f32_gemminc_ukernel_5x16__avx512f_broadcast() 104 vacc0x0123456789ABCDEF = _mm512_max_ps(vacc0x0123456789ABCDEF, vmin); in xnn_f32_gemminc_ukernel_5x16__avx512f_broadcast() 119 _mm512_storeu_ps(c0, vacc0x0123456789ABCDEF); in xnn_f32_gemminc_ukernel_5x16__avx512f_broadcast() 138 _mm512_mask_storeu_ps(c0, vmask, vacc0x0123456789ABCDEF); in xnn_f32_gemminc_ukernel_5x16__avx512f_broadcast()
|
D | 6x16-avx512f-broadcast.c | 75 __m512 vacc0x0123456789ABCDEF = _mm512_load_ps(acc + 0); in xnn_f32_gemminc_ukernel_6x16__avx512f_broadcast() local 88 …vacc0x0123456789ABCDEF = _mm512_fmadd_ps(_mm512_set1_ps(*a0), vb0123456789ABCDEF, vacc0x0123456789… in xnn_f32_gemminc_ukernel_6x16__avx512f_broadcast() 106 vacc0x0123456789ABCDEF = _mm512_min_ps(vacc0x0123456789ABCDEF, vmax); in xnn_f32_gemminc_ukernel_6x16__avx512f_broadcast() 114 vacc0x0123456789ABCDEF = _mm512_max_ps(vacc0x0123456789ABCDEF, vmin); in xnn_f32_gemminc_ukernel_6x16__avx512f_broadcast() 132 _mm512_storeu_ps(c0, vacc0x0123456789ABCDEF); in xnn_f32_gemminc_ukernel_6x16__avx512f_broadcast() 153 _mm512_mask_storeu_ps(c0, vmask, vacc0x0123456789ABCDEF); in xnn_f32_gemminc_ukernel_6x16__avx512f_broadcast()
|
D | 7x16-avx512f-broadcast.c | 81 __m512 vacc0x0123456789ABCDEF = _mm512_load_ps(acc + 0); in xnn_f32_gemminc_ukernel_7x16__avx512f_broadcast() local 95 …vacc0x0123456789ABCDEF = _mm512_fmadd_ps(_mm512_set1_ps(*a0), vb0123456789ABCDEF, vacc0x0123456789… in xnn_f32_gemminc_ukernel_7x16__avx512f_broadcast() 115 vacc0x0123456789ABCDEF = _mm512_min_ps(vacc0x0123456789ABCDEF, vmax); in xnn_f32_gemminc_ukernel_7x16__avx512f_broadcast() 124 vacc0x0123456789ABCDEF = _mm512_max_ps(vacc0x0123456789ABCDEF, vmin); in xnn_f32_gemminc_ukernel_7x16__avx512f_broadcast() 145 _mm512_storeu_ps(c0, vacc0x0123456789ABCDEF); in xnn_f32_gemminc_ukernel_7x16__avx512f_broadcast() 168 _mm512_mask_storeu_ps(c0, vmask, vacc0x0123456789ABCDEF); in xnn_f32_gemminc_ukernel_7x16__avx512f_broadcast()
|
D | 8x16-avx512f-broadcast.c | 87 __m512 vacc0x0123456789ABCDEF = _mm512_load_ps(acc + 0); in xnn_f32_gemminc_ukernel_8x16__avx512f_broadcast() local 102 …vacc0x0123456789ABCDEF = _mm512_fmadd_ps(_mm512_set1_ps(*a0), vb0123456789ABCDEF, vacc0x0123456789… in xnn_f32_gemminc_ukernel_8x16__avx512f_broadcast() 124 vacc0x0123456789ABCDEF = _mm512_min_ps(vacc0x0123456789ABCDEF, vmax); in xnn_f32_gemminc_ukernel_8x16__avx512f_broadcast() 134 vacc0x0123456789ABCDEF = _mm512_max_ps(vacc0x0123456789ABCDEF, vmin); in xnn_f32_gemminc_ukernel_8x16__avx512f_broadcast() 158 _mm512_storeu_ps(c0, vacc0x0123456789ABCDEF); in xnn_f32_gemminc_ukernel_8x16__avx512f_broadcast() 183 _mm512_mask_storeu_ps(c0, vmask, vacc0x0123456789ABCDEF); in xnn_f32_gemminc_ukernel_8x16__avx512f_broadcast()
|
/external/XNNPACK/src/f32-igemm/gen/ |
D | 1x16-avx512f-broadcast.c | 47 __m512 vacc0x0123456789ABCDEF = _mm512_load_ps(w); in xnn_f32_igemm_ukernel_1x16__avx512f_broadcast() local 64 …vacc0x0123456789ABCDEF = _mm512_fmadd_ps(_mm512_set1_ps(*a0), vb0123456789ABCDEF, vacc0x0123456789… in xnn_f32_igemm_ukernel_1x16__avx512f_broadcast() 74 vacc0x0123456789ABCDEF = _mm512_min_ps(vacc0x0123456789ABCDEF, vmax); in xnn_f32_igemm_ukernel_1x16__avx512f_broadcast() 77 vacc0x0123456789ABCDEF = _mm512_max_ps(vacc0x0123456789ABCDEF, vmin); in xnn_f32_igemm_ukernel_1x16__avx512f_broadcast() 80 _mm512_storeu_ps(c0, vacc0x0123456789ABCDEF); in xnn_f32_igemm_ukernel_1x16__avx512f_broadcast() 90 _mm512_mask_storeu_ps(c0, vmask, vacc0x0123456789ABCDEF); in xnn_f32_igemm_ukernel_1x16__avx512f_broadcast()
|
D | 5x16-avx512f-broadcast.c | 63 __m512 vacc0x0123456789ABCDEF = _mm512_load_ps(w); in xnn_f32_igemm_ukernel_5x16__avx512f_broadcast() local 64 __m512 vacc1x0123456789ABCDEF = vacc0x0123456789ABCDEF; in xnn_f32_igemm_ukernel_5x16__avx512f_broadcast() 65 __m512 vacc2x0123456789ABCDEF = vacc0x0123456789ABCDEF; in xnn_f32_igemm_ukernel_5x16__avx512f_broadcast() 66 __m512 vacc3x0123456789ABCDEF = vacc0x0123456789ABCDEF; in xnn_f32_igemm_ukernel_5x16__avx512f_broadcast() 67 __m512 vacc4x0123456789ABCDEF = vacc0x0123456789ABCDEF; in xnn_f32_igemm_ukernel_5x16__avx512f_broadcast() 104 …vacc0x0123456789ABCDEF = _mm512_fmadd_ps(_mm512_set1_ps(*a0), vb0123456789ABCDEF, vacc0x0123456789… in xnn_f32_igemm_ukernel_5x16__avx512f_broadcast() 122 vacc0x0123456789ABCDEF = _mm512_min_ps(vacc0x0123456789ABCDEF, vmax); in xnn_f32_igemm_ukernel_5x16__avx512f_broadcast() 129 vacc0x0123456789ABCDEF = _mm512_max_ps(vacc0x0123456789ABCDEF, vmin); in xnn_f32_igemm_ukernel_5x16__avx512f_broadcast() 144 _mm512_storeu_ps(c0, vacc0x0123456789ABCDEF); in xnn_f32_igemm_ukernel_5x16__avx512f_broadcast() 158 _mm512_mask_storeu_ps(c0, vmask, vacc0x0123456789ABCDEF); in xnn_f32_igemm_ukernel_5x16__avx512f_broadcast()
|
D | 7x16-avx512f-broadcast.c | 71 __m512 vacc0x0123456789ABCDEF = _mm512_load_ps(w); in xnn_f32_igemm_ukernel_7x16__avx512f_broadcast() local 72 __m512 vacc1x0123456789ABCDEF = vacc0x0123456789ABCDEF; in xnn_f32_igemm_ukernel_7x16__avx512f_broadcast() 73 __m512 vacc2x0123456789ABCDEF = vacc0x0123456789ABCDEF; in xnn_f32_igemm_ukernel_7x16__avx512f_broadcast() 74 __m512 vacc3x0123456789ABCDEF = vacc0x0123456789ABCDEF; in xnn_f32_igemm_ukernel_7x16__avx512f_broadcast() 75 __m512 vacc4x0123456789ABCDEF = vacc0x0123456789ABCDEF; in xnn_f32_igemm_ukernel_7x16__avx512f_broadcast() 76 __m512 vacc5x0123456789ABCDEF = vacc0x0123456789ABCDEF; in xnn_f32_igemm_ukernel_7x16__avx512f_broadcast() 77 __m512 vacc6x0123456789ABCDEF = vacc0x0123456789ABCDEF; in xnn_f32_igemm_ukernel_7x16__avx512f_broadcast() 124 …vacc0x0123456789ABCDEF = _mm512_fmadd_ps(_mm512_set1_ps(*a0), vb0123456789ABCDEF, vacc0x0123456789… in xnn_f32_igemm_ukernel_7x16__avx512f_broadcast() 146 vacc0x0123456789ABCDEF = _mm512_min_ps(vacc0x0123456789ABCDEF, vmax); in xnn_f32_igemm_ukernel_7x16__avx512f_broadcast() 155 vacc0x0123456789ABCDEF = _mm512_max_ps(vacc0x0123456789ABCDEF, vmin); in xnn_f32_igemm_ukernel_7x16__avx512f_broadcast() [all …]
|
D | 4x16-avx512f-broadcast.c | 59 __m512 vacc0x0123456789ABCDEF = _mm512_load_ps(w); in xnn_f32_igemm_ukernel_4x16__avx512f_broadcast() local 60 __m512 vacc1x0123456789ABCDEF = vacc0x0123456789ABCDEF; in xnn_f32_igemm_ukernel_4x16__avx512f_broadcast() 61 __m512 vacc2x0123456789ABCDEF = vacc0x0123456789ABCDEF; in xnn_f32_igemm_ukernel_4x16__avx512f_broadcast() 62 __m512 vacc3x0123456789ABCDEF = vacc0x0123456789ABCDEF; in xnn_f32_igemm_ukernel_4x16__avx512f_broadcast() 94 …vacc0x0123456789ABCDEF = _mm512_fmadd_ps(_mm512_set1_ps(*a0), vb0123456789ABCDEF, vacc0x0123456789… in xnn_f32_igemm_ukernel_4x16__avx512f_broadcast() 110 vacc0x0123456789ABCDEF = _mm512_min_ps(vacc0x0123456789ABCDEF, vmax); in xnn_f32_igemm_ukernel_4x16__avx512f_broadcast() 116 vacc0x0123456789ABCDEF = _mm512_max_ps(vacc0x0123456789ABCDEF, vmin); in xnn_f32_igemm_ukernel_4x16__avx512f_broadcast() 128 _mm512_storeu_ps(c0, vacc0x0123456789ABCDEF); in xnn_f32_igemm_ukernel_4x16__avx512f_broadcast() 141 _mm512_mask_storeu_ps(c0, vmask, vacc0x0123456789ABCDEF); in xnn_f32_igemm_ukernel_4x16__avx512f_broadcast()
|
D | 6x16-avx512f-broadcast.c | 67 __m512 vacc0x0123456789ABCDEF = _mm512_load_ps(w); in xnn_f32_igemm_ukernel_6x16__avx512f_broadcast() local 68 __m512 vacc1x0123456789ABCDEF = vacc0x0123456789ABCDEF; in xnn_f32_igemm_ukernel_6x16__avx512f_broadcast() 69 __m512 vacc2x0123456789ABCDEF = vacc0x0123456789ABCDEF; in xnn_f32_igemm_ukernel_6x16__avx512f_broadcast() 70 __m512 vacc3x0123456789ABCDEF = vacc0x0123456789ABCDEF; in xnn_f32_igemm_ukernel_6x16__avx512f_broadcast() 71 __m512 vacc4x0123456789ABCDEF = vacc0x0123456789ABCDEF; in xnn_f32_igemm_ukernel_6x16__avx512f_broadcast() 72 __m512 vacc5x0123456789ABCDEF = vacc0x0123456789ABCDEF; in xnn_f32_igemm_ukernel_6x16__avx512f_broadcast() 114 …vacc0x0123456789ABCDEF = _mm512_fmadd_ps(_mm512_set1_ps(*a0), vb0123456789ABCDEF, vacc0x0123456789… in xnn_f32_igemm_ukernel_6x16__avx512f_broadcast() 134 vacc0x0123456789ABCDEF = _mm512_min_ps(vacc0x0123456789ABCDEF, vmax); in xnn_f32_igemm_ukernel_6x16__avx512f_broadcast() 142 vacc0x0123456789ABCDEF = _mm512_max_ps(vacc0x0123456789ABCDEF, vmin); in xnn_f32_igemm_ukernel_6x16__avx512f_broadcast() 160 _mm512_storeu_ps(c0, vacc0x0123456789ABCDEF); in xnn_f32_igemm_ukernel_6x16__avx512f_broadcast() [all …]
|
D | 8x16-avx512f-broadcast.c | 75 __m512 vacc0x0123456789ABCDEF = _mm512_load_ps(w); in xnn_f32_igemm_ukernel_8x16__avx512f_broadcast() local 76 __m512 vacc1x0123456789ABCDEF = vacc0x0123456789ABCDEF; in xnn_f32_igemm_ukernel_8x16__avx512f_broadcast() 77 __m512 vacc2x0123456789ABCDEF = vacc0x0123456789ABCDEF; in xnn_f32_igemm_ukernel_8x16__avx512f_broadcast() 78 __m512 vacc3x0123456789ABCDEF = vacc0x0123456789ABCDEF; in xnn_f32_igemm_ukernel_8x16__avx512f_broadcast() 79 __m512 vacc4x0123456789ABCDEF = vacc0x0123456789ABCDEF; in xnn_f32_igemm_ukernel_8x16__avx512f_broadcast() 80 __m512 vacc5x0123456789ABCDEF = vacc0x0123456789ABCDEF; in xnn_f32_igemm_ukernel_8x16__avx512f_broadcast() 81 __m512 vacc6x0123456789ABCDEF = vacc0x0123456789ABCDEF; in xnn_f32_igemm_ukernel_8x16__avx512f_broadcast() 82 __m512 vacc7x0123456789ABCDEF = vacc0x0123456789ABCDEF; in xnn_f32_igemm_ukernel_8x16__avx512f_broadcast() 134 …vacc0x0123456789ABCDEF = _mm512_fmadd_ps(_mm512_set1_ps(*a0), vb0123456789ABCDEF, vacc0x0123456789… in xnn_f32_igemm_ukernel_8x16__avx512f_broadcast() 158 vacc0x0123456789ABCDEF = _mm512_min_ps(vacc0x0123456789ABCDEF, vmax); in xnn_f32_igemm_ukernel_8x16__avx512f_broadcast() [all …]
|