/external/XNNPACK/src/f32-dwconv/gen/ |
D | up32x4-avx512f.c | 78 const __m512 vi3x0123456789ABCDEF = _mm512_loadu_ps(i3); in xnn_f32_dwconv_ukernel_up32x4__avx512f() local 84 …vacc0123456789ABCDEFp0 = _mm512_fmadd_ps(vi3x0123456789ABCDEF, vk3x0123456789ABCDEF, vacc012345678… in xnn_f32_dwconv_ukernel_up32x4__avx512f() 120 const __m512 vi3x0123456789ABCDEF = _mm512_loadu_ps(i3); in xnn_f32_dwconv_ukernel_up32x4__avx512f() local 124 …vacc0123456789ABCDEFp0 = _mm512_fmadd_ps(vi3x0123456789ABCDEF, vk3x0123456789ABCDEF, vacc012345678… in xnn_f32_dwconv_ukernel_up32x4__avx512f() 155 const __m512 vi3x0123456789ABCDEF = _mm512_maskz_loadu_ps(vmask, i3); in xnn_f32_dwconv_ukernel_up32x4__avx512f() local 157 …vacc0123456789ABCDEFp0 = _mm512_fmadd_ps(vi3x0123456789ABCDEF, vk3x0123456789ABCDEF, vacc012345678… in xnn_f32_dwconv_ukernel_up32x4__avx512f()
|
D | up32x4-avx512f-acc2.c | 78 const __m512 vi3x0123456789ABCDEF = _mm512_loadu_ps(i3); in xnn_f32_dwconv_ukernel_up32x4__avx512f_acc2() local 84 …vacc0123456789ABCDEFp1 = _mm512_fmadd_ps(vi3x0123456789ABCDEF, vk3x0123456789ABCDEF, vacc012345678… in xnn_f32_dwconv_ukernel_up32x4__avx512f_acc2() 123 const __m512 vi3x0123456789ABCDEF = _mm512_loadu_ps(i3); in xnn_f32_dwconv_ukernel_up32x4__avx512f_acc2() local 127 …vacc0123456789ABCDEFp1 = _mm512_fmadd_ps(vi3x0123456789ABCDEF, vk3x0123456789ABCDEF, vacc012345678… in xnn_f32_dwconv_ukernel_up32x4__avx512f_acc2() 160 const __m512 vi3x0123456789ABCDEF = _mm512_maskz_loadu_ps(vmask, i3); in xnn_f32_dwconv_ukernel_up32x4__avx512f_acc2() local 162 …vacc0123456789ABCDEFp1 = _mm512_fmadd_ps(vi3x0123456789ABCDEF, vk3x0123456789ABCDEF, vacc012345678… in xnn_f32_dwconv_ukernel_up32x4__avx512f_acc2()
|
D | up16x4-avx512f.c | 68 const __m512 vi3x0123456789ABCDEF = _mm512_loadu_ps(i3); in xnn_f32_dwconv_ukernel_up16x4__avx512f() local 72 …vacc0123456789ABCDEFp0 = _mm512_fmadd_ps(vi3x0123456789ABCDEF, vk3x0123456789ABCDEF, vacc012345678… in xnn_f32_dwconv_ukernel_up16x4__avx512f() 103 const __m512 vi3x0123456789ABCDEF = _mm512_maskz_loadu_ps(vmask, i3); in xnn_f32_dwconv_ukernel_up16x4__avx512f() local 105 …vacc0123456789ABCDEFp0 = _mm512_fmadd_ps(vi3x0123456789ABCDEF, vk3x0123456789ABCDEF, vacc012345678… in xnn_f32_dwconv_ukernel_up16x4__avx512f()
|
D | up16x4-avx512f-acc2.c | 68 const __m512 vi3x0123456789ABCDEF = _mm512_loadu_ps(i3); in xnn_f32_dwconv_ukernel_up16x4__avx512f_acc2() local 72 …vacc0123456789ABCDEFp1 = _mm512_fmadd_ps(vi3x0123456789ABCDEF, vk3x0123456789ABCDEF, vacc012345678… in xnn_f32_dwconv_ukernel_up16x4__avx512f_acc2() 105 const __m512 vi3x0123456789ABCDEF = _mm512_maskz_loadu_ps(vmask, i3); in xnn_f32_dwconv_ukernel_up16x4__avx512f_acc2() local 107 …vacc0123456789ABCDEFp1 = _mm512_fmadd_ps(vi3x0123456789ABCDEF, vk3x0123456789ABCDEF, vacc012345678… in xnn_f32_dwconv_ukernel_up16x4__avx512f_acc2()
|
D | up32x9-avx512f.c | 88 const __m512 vi3x0123456789ABCDEF = _mm512_loadu_ps(i3); in xnn_f32_dwconv_ukernel_up32x9__avx512f() local 94 …vacc0123456789ABCDEFp0 = _mm512_fmadd_ps(vi3x0123456789ABCDEF, vk3x0123456789ABCDEF, vacc012345678… in xnn_f32_dwconv_ukernel_up32x9__avx512f() 175 const __m512 vi3x0123456789ABCDEF = _mm512_loadu_ps(i3); in xnn_f32_dwconv_ukernel_up32x9__avx512f() local 179 …vacc0123456789ABCDEFp0 = _mm512_fmadd_ps(vi3x0123456789ABCDEF, vk3x0123456789ABCDEF, vacc012345678… in xnn_f32_dwconv_ukernel_up32x9__avx512f() 240 const __m512 vi3x0123456789ABCDEF = _mm512_maskz_loadu_ps(vmask, i3); in xnn_f32_dwconv_ukernel_up32x9__avx512f() local 242 …vacc0123456789ABCDEFp0 = _mm512_fmadd_ps(vi3x0123456789ABCDEF, vk3x0123456789ABCDEF, vacc012345678… in xnn_f32_dwconv_ukernel_up32x9__avx512f()
|
D | up32x9-avx512f-acc2.c | 88 const __m512 vi3x0123456789ABCDEF = _mm512_loadu_ps(i3); in xnn_f32_dwconv_ukernel_up32x9__avx512f_acc2() local 94 …vacc0123456789ABCDEFp1 = _mm512_fmadd_ps(vi3x0123456789ABCDEF, vk3x0123456789ABCDEF, vacc012345678… in xnn_f32_dwconv_ukernel_up32x9__avx512f_acc2() 178 const __m512 vi3x0123456789ABCDEF = _mm512_loadu_ps(i3); in xnn_f32_dwconv_ukernel_up32x9__avx512f_acc2() local 182 …vacc0123456789ABCDEFp1 = _mm512_fmadd_ps(vi3x0123456789ABCDEF, vk3x0123456789ABCDEF, vacc012345678… in xnn_f32_dwconv_ukernel_up32x9__avx512f_acc2() 245 const __m512 vi3x0123456789ABCDEF = _mm512_maskz_loadu_ps(vmask, i3); in xnn_f32_dwconv_ukernel_up32x9__avx512f_acc2() local 247 …vacc0123456789ABCDEFp1 = _mm512_fmadd_ps(vi3x0123456789ABCDEF, vk3x0123456789ABCDEF, vacc012345678… in xnn_f32_dwconv_ukernel_up32x9__avx512f_acc2()
|
D | up16x9-avx512f.c | 78 const __m512 vi3x0123456789ABCDEF = _mm512_loadu_ps(i3); in xnn_f32_dwconv_ukernel_up16x9__avx512f() local 82 …vacc0123456789ABCDEFp0 = _mm512_fmadd_ps(vi3x0123456789ABCDEF, vk3x0123456789ABCDEF, vacc012345678… in xnn_f32_dwconv_ukernel_up16x9__avx512f() 143 const __m512 vi3x0123456789ABCDEF = _mm512_maskz_loadu_ps(vmask, i3); in xnn_f32_dwconv_ukernel_up16x9__avx512f() local 145 …vacc0123456789ABCDEFp0 = _mm512_fmadd_ps(vi3x0123456789ABCDEF, vk3x0123456789ABCDEF, vacc012345678… in xnn_f32_dwconv_ukernel_up16x9__avx512f()
|
D | up16x9-avx512f-acc2.c | 78 const __m512 vi3x0123456789ABCDEF = _mm512_loadu_ps(i3); in xnn_f32_dwconv_ukernel_up16x9__avx512f_acc2() local 82 …vacc0123456789ABCDEFp1 = _mm512_fmadd_ps(vi3x0123456789ABCDEF, vk3x0123456789ABCDEF, vacc012345678… in xnn_f32_dwconv_ukernel_up16x9__avx512f_acc2() 145 const __m512 vi3x0123456789ABCDEF = _mm512_maskz_loadu_ps(vmask, i3); in xnn_f32_dwconv_ukernel_up16x9__avx512f_acc2() local 147 …vacc0123456789ABCDEFp1 = _mm512_fmadd_ps(vi3x0123456789ABCDEF, vk3x0123456789ABCDEF, vacc012345678… in xnn_f32_dwconv_ukernel_up16x9__avx512f_acc2()
|
D | up16x25-avx512f.c | 110 const __m512 vi3x0123456789ABCDEF = _mm512_loadu_ps(i3); in xnn_f32_dwconv_ukernel_up16x25__avx512f() local 114 …vacc0123456789ABCDEFp0 = _mm512_fmadd_ps(vi3x0123456789ABCDEF, vk3x0123456789ABCDEF, vacc012345678… in xnn_f32_dwconv_ukernel_up16x25__avx512f() 271 const __m512 vi3x0123456789ABCDEF = _mm512_maskz_loadu_ps(vmask, i3); in xnn_f32_dwconv_ukernel_up16x25__avx512f() local 273 …vacc0123456789ABCDEFp0 = _mm512_fmadd_ps(vi3x0123456789ABCDEF, vk3x0123456789ABCDEF, vacc012345678… in xnn_f32_dwconv_ukernel_up16x25__avx512f()
|
D | up16x25-avx512f-acc2.c | 110 const __m512 vi3x0123456789ABCDEF = _mm512_loadu_ps(i3); in xnn_f32_dwconv_ukernel_up16x25__avx512f_acc2() local 114 …vacc0123456789ABCDEFp1 = _mm512_fmadd_ps(vi3x0123456789ABCDEF, vk3x0123456789ABCDEF, vacc012345678… in xnn_f32_dwconv_ukernel_up16x25__avx512f_acc2() 273 const __m512 vi3x0123456789ABCDEF = _mm512_maskz_loadu_ps(vmask, i3); in xnn_f32_dwconv_ukernel_up16x25__avx512f_acc2() local 275 …vacc0123456789ABCDEFp1 = _mm512_fmadd_ps(vi3x0123456789ABCDEF, vk3x0123456789ABCDEF, vacc012345678… in xnn_f32_dwconv_ukernel_up16x25__avx512f_acc2()
|
D | up32x25-avx512f-acc2.c | 120 const __m512 vi3x0123456789ABCDEF = _mm512_loadu_ps(i3); in xnn_f32_dwconv_ukernel_up32x25__avx512f_acc2() local 126 …vacc0123456789ABCDEFp1 = _mm512_fmadd_ps(vi3x0123456789ABCDEF, vk3x0123456789ABCDEF, vacc012345678… in xnn_f32_dwconv_ukernel_up32x25__avx512f_acc2() 354 const __m512 vi3x0123456789ABCDEF = _mm512_loadu_ps(i3); in xnn_f32_dwconv_ukernel_up32x25__avx512f_acc2() local 358 …vacc0123456789ABCDEFp1 = _mm512_fmadd_ps(vi3x0123456789ABCDEF, vk3x0123456789ABCDEF, vacc012345678… in xnn_f32_dwconv_ukernel_up32x25__avx512f_acc2() 517 const __m512 vi3x0123456789ABCDEF = _mm512_maskz_loadu_ps(vmask, i3); in xnn_f32_dwconv_ukernel_up32x25__avx512f_acc2() local 519 …vacc0123456789ABCDEFp1 = _mm512_fmadd_ps(vi3x0123456789ABCDEF, vk3x0123456789ABCDEF, vacc012345678… in xnn_f32_dwconv_ukernel_up32x25__avx512f_acc2()
|
D | up32x25-avx512f.c | 120 const __m512 vi3x0123456789ABCDEF = _mm512_loadu_ps(i3); in xnn_f32_dwconv_ukernel_up32x25__avx512f() local 126 …vacc0123456789ABCDEFp0 = _mm512_fmadd_ps(vi3x0123456789ABCDEF, vk3x0123456789ABCDEF, vacc012345678… in xnn_f32_dwconv_ukernel_up32x25__avx512f() 351 const __m512 vi3x0123456789ABCDEF = _mm512_loadu_ps(i3); in xnn_f32_dwconv_ukernel_up32x25__avx512f() local 355 …vacc0123456789ABCDEFp0 = _mm512_fmadd_ps(vi3x0123456789ABCDEF, vk3x0123456789ABCDEF, vacc012345678… in xnn_f32_dwconv_ukernel_up32x25__avx512f() 512 const __m512 vi3x0123456789ABCDEF = _mm512_maskz_loadu_ps(vmask, i3); in xnn_f32_dwconv_ukernel_up32x25__avx512f() local 514 …vacc0123456789ABCDEFp0 = _mm512_fmadd_ps(vi3x0123456789ABCDEF, vk3x0123456789ABCDEF, vacc012345678… in xnn_f32_dwconv_ukernel_up32x25__avx512f()
|