/external/XNNPACK/src/f32-gemm/gen-inc/ |
D | 5x16inc-minmax-avx512f-broadcast.c | 73 __m512 vacc4x0123456789ABCDEF = _mm512_load_ps(acc + 64); in xnn_f32_gemminc_minmax_ukernel_5x16__avx512f_broadcast() local 85 …vacc4x0123456789ABCDEF = _mm512_fmadd_ps(_mm512_set1_ps(*a4), vb0123456789ABCDEF, vacc4x0123456789… in xnn_f32_gemminc_minmax_ukernel_5x16__avx512f_broadcast() 101 vacc4x0123456789ABCDEF = _mm512_min_ps(vacc4x0123456789ABCDEF, vmax); in xnn_f32_gemminc_minmax_ukernel_5x16__avx512f_broadcast() 108 vacc4x0123456789ABCDEF = _mm512_max_ps(vacc4x0123456789ABCDEF, vmin); in xnn_f32_gemminc_minmax_ukernel_5x16__avx512f_broadcast() 111 _mm512_storeu_ps(c4, vacc4x0123456789ABCDEF); in xnn_f32_gemminc_minmax_ukernel_5x16__avx512f_broadcast() 134 _mm512_mask_storeu_ps(c4, vmask, vacc4x0123456789ABCDEF); in xnn_f32_gemminc_minmax_ukernel_5x16__avx512f_broadcast()
|
D | 6x16inc-minmax-avx512f-broadcast.c | 79 __m512 vacc4x0123456789ABCDEF = _mm512_load_ps(acc + 64); in xnn_f32_gemminc_minmax_ukernel_6x16__avx512f_broadcast() local 92 …vacc4x0123456789ABCDEF = _mm512_fmadd_ps(_mm512_set1_ps(*a4), vb0123456789ABCDEF, vacc4x0123456789… in xnn_f32_gemminc_minmax_ukernel_6x16__avx512f_broadcast() 110 vacc4x0123456789ABCDEF = _mm512_min_ps(vacc4x0123456789ABCDEF, vmax); in xnn_f32_gemminc_minmax_ukernel_6x16__avx512f_broadcast() 118 vacc4x0123456789ABCDEF = _mm512_max_ps(vacc4x0123456789ABCDEF, vmin); in xnn_f32_gemminc_minmax_ukernel_6x16__avx512f_broadcast() 124 _mm512_storeu_ps(c4, vacc4x0123456789ABCDEF); in xnn_f32_gemminc_minmax_ukernel_6x16__avx512f_broadcast() 149 _mm512_mask_storeu_ps(c4, vmask, vacc4x0123456789ABCDEF); in xnn_f32_gemminc_minmax_ukernel_6x16__avx512f_broadcast()
|
D | 7x16inc-minmax-avx512f-broadcast.c | 85 __m512 vacc4x0123456789ABCDEF = _mm512_load_ps(acc + 64); in xnn_f32_gemminc_minmax_ukernel_7x16__avx512f_broadcast() local 99 …vacc4x0123456789ABCDEF = _mm512_fmadd_ps(_mm512_set1_ps(*a4), vb0123456789ABCDEF, vacc4x0123456789… in xnn_f32_gemminc_minmax_ukernel_7x16__avx512f_broadcast() 119 vacc4x0123456789ABCDEF = _mm512_min_ps(vacc4x0123456789ABCDEF, vmax); in xnn_f32_gemminc_minmax_ukernel_7x16__avx512f_broadcast() 128 vacc4x0123456789ABCDEF = _mm512_max_ps(vacc4x0123456789ABCDEF, vmin); in xnn_f32_gemminc_minmax_ukernel_7x16__avx512f_broadcast() 137 _mm512_storeu_ps(c4, vacc4x0123456789ABCDEF); in xnn_f32_gemminc_minmax_ukernel_7x16__avx512f_broadcast() 164 _mm512_mask_storeu_ps(c4, vmask, vacc4x0123456789ABCDEF); in xnn_f32_gemminc_minmax_ukernel_7x16__avx512f_broadcast()
|
D | 8x16inc-minmax-avx512f-broadcast.c | 91 __m512 vacc4x0123456789ABCDEF = _mm512_load_ps(acc + 64); in xnn_f32_gemminc_minmax_ukernel_8x16__avx512f_broadcast() local 106 …vacc4x0123456789ABCDEF = _mm512_fmadd_ps(_mm512_set1_ps(*a4), vb0123456789ABCDEF, vacc4x0123456789… in xnn_f32_gemminc_minmax_ukernel_8x16__avx512f_broadcast() 128 vacc4x0123456789ABCDEF = _mm512_min_ps(vacc4x0123456789ABCDEF, vmax); in xnn_f32_gemminc_minmax_ukernel_8x16__avx512f_broadcast() 138 vacc4x0123456789ABCDEF = _mm512_max_ps(vacc4x0123456789ABCDEF, vmin); in xnn_f32_gemminc_minmax_ukernel_8x16__avx512f_broadcast() 150 _mm512_storeu_ps(c4, vacc4x0123456789ABCDEF); in xnn_f32_gemminc_minmax_ukernel_8x16__avx512f_broadcast() 179 _mm512_mask_storeu_ps(c4, vmask, vacc4x0123456789ABCDEF); in xnn_f32_gemminc_minmax_ukernel_8x16__avx512f_broadcast()
|
/external/XNNPACK/src/f32-gemm/gen/ |
D | 5x16-minmax-avx512f-broadcast.c | 71 __m512 vacc4x0123456789ABCDEF = vacc0x0123456789ABCDEF; in xnn_f32_gemm_minmax_ukernel_5x16__avx512f_broadcast() local 83 …vacc4x0123456789ABCDEF = _mm512_fmadd_ps(_mm512_set1_ps(*a4), vb0123456789ABCDEF, vacc4x0123456789… in xnn_f32_gemm_minmax_ukernel_5x16__avx512f_broadcast() 99 vacc4x0123456789ABCDEF = _mm512_min_ps(vacc4x0123456789ABCDEF, vmax); in xnn_f32_gemm_minmax_ukernel_5x16__avx512f_broadcast() 106 vacc4x0123456789ABCDEF = _mm512_max_ps(vacc4x0123456789ABCDEF, vmin); in xnn_f32_gemm_minmax_ukernel_5x16__avx512f_broadcast() 109 _mm512_storeu_ps(c4, vacc4x0123456789ABCDEF); in xnn_f32_gemm_minmax_ukernel_5x16__avx512f_broadcast() 132 _mm512_mask_storeu_ps(c4, vmask, vacc4x0123456789ABCDEF); in xnn_f32_gemm_minmax_ukernel_5x16__avx512f_broadcast()
|
D | 6x16-minmax-avx512f-broadcast.c | 77 __m512 vacc4x0123456789ABCDEF = vacc0x0123456789ABCDEF; in xnn_f32_gemm_minmax_ukernel_6x16__avx512f_broadcast() local 90 …vacc4x0123456789ABCDEF = _mm512_fmadd_ps(_mm512_set1_ps(*a4), vb0123456789ABCDEF, vacc4x0123456789… in xnn_f32_gemm_minmax_ukernel_6x16__avx512f_broadcast() 108 vacc4x0123456789ABCDEF = _mm512_min_ps(vacc4x0123456789ABCDEF, vmax); in xnn_f32_gemm_minmax_ukernel_6x16__avx512f_broadcast() 116 vacc4x0123456789ABCDEF = _mm512_max_ps(vacc4x0123456789ABCDEF, vmin); in xnn_f32_gemm_minmax_ukernel_6x16__avx512f_broadcast() 122 _mm512_storeu_ps(c4, vacc4x0123456789ABCDEF); in xnn_f32_gemm_minmax_ukernel_6x16__avx512f_broadcast() 147 _mm512_mask_storeu_ps(c4, vmask, vacc4x0123456789ABCDEF); in xnn_f32_gemm_minmax_ukernel_6x16__avx512f_broadcast()
|
D | 7x16-minmax-avx512f-broadcast.c | 83 __m512 vacc4x0123456789ABCDEF = vacc0x0123456789ABCDEF; in xnn_f32_gemm_minmax_ukernel_7x16__avx512f_broadcast() local 97 …vacc4x0123456789ABCDEF = _mm512_fmadd_ps(_mm512_set1_ps(*a4), vb0123456789ABCDEF, vacc4x0123456789… in xnn_f32_gemm_minmax_ukernel_7x16__avx512f_broadcast() 117 vacc4x0123456789ABCDEF = _mm512_min_ps(vacc4x0123456789ABCDEF, vmax); in xnn_f32_gemm_minmax_ukernel_7x16__avx512f_broadcast() 126 vacc4x0123456789ABCDEF = _mm512_max_ps(vacc4x0123456789ABCDEF, vmin); in xnn_f32_gemm_minmax_ukernel_7x16__avx512f_broadcast() 135 _mm512_storeu_ps(c4, vacc4x0123456789ABCDEF); in xnn_f32_gemm_minmax_ukernel_7x16__avx512f_broadcast() 162 _mm512_mask_storeu_ps(c4, vmask, vacc4x0123456789ABCDEF); in xnn_f32_gemm_minmax_ukernel_7x16__avx512f_broadcast()
|
D | 8x16-minmax-avx512f-broadcast.c | 89 __m512 vacc4x0123456789ABCDEF = vacc0x0123456789ABCDEF; in xnn_f32_gemm_minmax_ukernel_8x16__avx512f_broadcast() local 104 …vacc4x0123456789ABCDEF = _mm512_fmadd_ps(_mm512_set1_ps(*a4), vb0123456789ABCDEF, vacc4x0123456789… in xnn_f32_gemm_minmax_ukernel_8x16__avx512f_broadcast() 126 vacc4x0123456789ABCDEF = _mm512_min_ps(vacc4x0123456789ABCDEF, vmax); in xnn_f32_gemm_minmax_ukernel_8x16__avx512f_broadcast() 136 vacc4x0123456789ABCDEF = _mm512_max_ps(vacc4x0123456789ABCDEF, vmin); in xnn_f32_gemm_minmax_ukernel_8x16__avx512f_broadcast() 148 _mm512_storeu_ps(c4, vacc4x0123456789ABCDEF); in xnn_f32_gemm_minmax_ukernel_8x16__avx512f_broadcast() 177 _mm512_mask_storeu_ps(c4, vmask, vacc4x0123456789ABCDEF); in xnn_f32_gemm_minmax_ukernel_8x16__avx512f_broadcast()
|
/external/XNNPACK/src/f32-igemm/gen/ |
D | 5x16-minmax-avx512f-broadcast.c | 67 __m512 vacc4x0123456789ABCDEF = vacc0x0123456789ABCDEF; in xnn_f32_igemm_minmax_ukernel_5x16__avx512f_broadcast() local 108 …vacc4x0123456789ABCDEF = _mm512_fmadd_ps(_mm512_set1_ps(*a4), vb0123456789ABCDEF, vacc4x0123456789… in xnn_f32_igemm_minmax_ukernel_5x16__avx512f_broadcast() 126 vacc4x0123456789ABCDEF = _mm512_min_ps(vacc4x0123456789ABCDEF, vmax); in xnn_f32_igemm_minmax_ukernel_5x16__avx512f_broadcast() 133 vacc4x0123456789ABCDEF = _mm512_max_ps(vacc4x0123456789ABCDEF, vmin); in xnn_f32_igemm_minmax_ukernel_5x16__avx512f_broadcast() 136 _mm512_storeu_ps(c4, vacc4x0123456789ABCDEF); in xnn_f32_igemm_minmax_ukernel_5x16__avx512f_broadcast() 154 _mm512_mask_storeu_ps(c4, vmask, vacc4x0123456789ABCDEF); in xnn_f32_igemm_minmax_ukernel_5x16__avx512f_broadcast()
|
D | 6x16-minmax-avx512f-broadcast.c | 71 __m512 vacc4x0123456789ABCDEF = vacc0x0123456789ABCDEF; in xnn_f32_igemm_minmax_ukernel_6x16__avx512f_broadcast() local 118 …vacc4x0123456789ABCDEF = _mm512_fmadd_ps(_mm512_set1_ps(*a4), vb0123456789ABCDEF, vacc4x0123456789… in xnn_f32_igemm_minmax_ukernel_6x16__avx512f_broadcast() 138 vacc4x0123456789ABCDEF = _mm512_min_ps(vacc4x0123456789ABCDEF, vmax); in xnn_f32_igemm_minmax_ukernel_6x16__avx512f_broadcast() 146 vacc4x0123456789ABCDEF = _mm512_max_ps(vacc4x0123456789ABCDEF, vmin); in xnn_f32_igemm_minmax_ukernel_6x16__avx512f_broadcast() 152 _mm512_storeu_ps(c4, vacc4x0123456789ABCDEF); in xnn_f32_igemm_minmax_ukernel_6x16__avx512f_broadcast() 171 _mm512_mask_storeu_ps(c4, vmask, vacc4x0123456789ABCDEF); in xnn_f32_igemm_minmax_ukernel_6x16__avx512f_broadcast()
|
D | 7x16-minmax-avx512f-broadcast.c | 75 __m512 vacc4x0123456789ABCDEF = vacc0x0123456789ABCDEF; in xnn_f32_igemm_minmax_ukernel_7x16__avx512f_broadcast() local 128 …vacc4x0123456789ABCDEF = _mm512_fmadd_ps(_mm512_set1_ps(*a4), vb0123456789ABCDEF, vacc4x0123456789… in xnn_f32_igemm_minmax_ukernel_7x16__avx512f_broadcast() 150 vacc4x0123456789ABCDEF = _mm512_min_ps(vacc4x0123456789ABCDEF, vmax); in xnn_f32_igemm_minmax_ukernel_7x16__avx512f_broadcast() 159 vacc4x0123456789ABCDEF = _mm512_max_ps(vacc4x0123456789ABCDEF, vmin); in xnn_f32_igemm_minmax_ukernel_7x16__avx512f_broadcast() 168 _mm512_storeu_ps(c4, vacc4x0123456789ABCDEF); in xnn_f32_igemm_minmax_ukernel_7x16__avx512f_broadcast() 188 _mm512_mask_storeu_ps(c4, vmask, vacc4x0123456789ABCDEF); in xnn_f32_igemm_minmax_ukernel_7x16__avx512f_broadcast()
|
D | 8x16-minmax-avx512f-broadcast.c | 79 __m512 vacc4x0123456789ABCDEF = vacc0x0123456789ABCDEF; in xnn_f32_igemm_minmax_ukernel_8x16__avx512f_broadcast() local 138 …vacc4x0123456789ABCDEF = _mm512_fmadd_ps(_mm512_set1_ps(*a4), vb0123456789ABCDEF, vacc4x0123456789… in xnn_f32_igemm_minmax_ukernel_8x16__avx512f_broadcast() 162 vacc4x0123456789ABCDEF = _mm512_min_ps(vacc4x0123456789ABCDEF, vmax); in xnn_f32_igemm_minmax_ukernel_8x16__avx512f_broadcast() 172 vacc4x0123456789ABCDEF = _mm512_max_ps(vacc4x0123456789ABCDEF, vmin); in xnn_f32_igemm_minmax_ukernel_8x16__avx512f_broadcast() 184 _mm512_storeu_ps(c4, vacc4x0123456789ABCDEF); in xnn_f32_igemm_minmax_ukernel_8x16__avx512f_broadcast() 205 _mm512_mask_storeu_ps(c4, vmask, vacc4x0123456789ABCDEF); in xnn_f32_igemm_minmax_ukernel_8x16__avx512f_broadcast()
|