/external/XNNPACK/src/f32-dwconv/gen/ |
D | up32x4-avx512f.c | 51 const __m512 vi0x0123456789ABCDEF = _mm512_loadu_ps(i0); in xnn_f32_dwconv_ukernel_up32x4__avx512f() local 57 …vacc0123456789ABCDEFp0 = _mm512_fmadd_ps(vi0x0123456789ABCDEF, vk0x0123456789ABCDEF, vacc012345678… in xnn_f32_dwconv_ukernel_up32x4__avx512f() 102 const __m512 vi0x0123456789ABCDEF = _mm512_loadu_ps(i0); in xnn_f32_dwconv_ukernel_up32x4__avx512f() local 106 …vacc0123456789ABCDEFp0 = _mm512_fmadd_ps(vi0x0123456789ABCDEF, vk0x0123456789ABCDEF, vacc012345678… in xnn_f32_dwconv_ukernel_up32x4__avx512f() 143 const __m512 vi0x0123456789ABCDEF = _mm512_maskz_loadu_ps(vmask, i0); in xnn_f32_dwconv_ukernel_up32x4__avx512f() local 145 …vacc0123456789ABCDEFp0 = _mm512_fmadd_ps(vi0x0123456789ABCDEF, vk0x0123456789ABCDEF, vacc012345678… in xnn_f32_dwconv_ukernel_up32x4__avx512f()
|
D | up32x4-avx512f-acc2.c | 51 const __m512 vi0x0123456789ABCDEF = _mm512_loadu_ps(i0); in xnn_f32_dwconv_ukernel_up32x4__avx512f_acc2() local 57 …vacc0123456789ABCDEFp0 = _mm512_fmadd_ps(vi0x0123456789ABCDEF, vk0x0123456789ABCDEF, vacc012345678… in xnn_f32_dwconv_ukernel_up32x4__avx512f_acc2() 105 const __m512 vi0x0123456789ABCDEF = _mm512_loadu_ps(i0); in xnn_f32_dwconv_ukernel_up32x4__avx512f_acc2() local 109 …vacc0123456789ABCDEFp0 = _mm512_fmadd_ps(vi0x0123456789ABCDEF, vk0x0123456789ABCDEF, vacc012345678… in xnn_f32_dwconv_ukernel_up32x4__avx512f_acc2() 148 const __m512 vi0x0123456789ABCDEF = _mm512_maskz_loadu_ps(vmask, i0); in xnn_f32_dwconv_ukernel_up32x4__avx512f_acc2() local 150 …vacc0123456789ABCDEFp0 = _mm512_fmadd_ps(vi0x0123456789ABCDEF, vk0x0123456789ABCDEF, vacc012345678… in xnn_f32_dwconv_ukernel_up32x4__avx512f_acc2()
|
D | up16x4-avx512f.c | 50 const __m512 vi0x0123456789ABCDEF = _mm512_loadu_ps(i0); in xnn_f32_dwconv_ukernel_up16x4__avx512f() local 54 …vacc0123456789ABCDEFp0 = _mm512_fmadd_ps(vi0x0123456789ABCDEF, vk0x0123456789ABCDEF, vacc012345678… in xnn_f32_dwconv_ukernel_up16x4__avx512f() 91 const __m512 vi0x0123456789ABCDEF = _mm512_maskz_loadu_ps(vmask, i0); in xnn_f32_dwconv_ukernel_up16x4__avx512f() local 93 …vacc0123456789ABCDEFp0 = _mm512_fmadd_ps(vi0x0123456789ABCDEF, vk0x0123456789ABCDEF, vacc012345678… in xnn_f32_dwconv_ukernel_up16x4__avx512f()
|
D | up16x4-avx512f-acc2.c | 50 const __m512 vi0x0123456789ABCDEF = _mm512_loadu_ps(i0); in xnn_f32_dwconv_ukernel_up16x4__avx512f_acc2() local 54 …vacc0123456789ABCDEFp0 = _mm512_fmadd_ps(vi0x0123456789ABCDEF, vk0x0123456789ABCDEF, vacc012345678… in xnn_f32_dwconv_ukernel_up16x4__avx512f_acc2() 93 const __m512 vi0x0123456789ABCDEF = _mm512_maskz_loadu_ps(vmask, i0); in xnn_f32_dwconv_ukernel_up16x4__avx512f_acc2() local 95 …vacc0123456789ABCDEFp0 = _mm512_fmadd_ps(vi0x0123456789ABCDEF, vk0x0123456789ABCDEF, vacc012345678… in xnn_f32_dwconv_ukernel_up16x4__avx512f_acc2()
|
D | up32x9-avx512f.c | 61 const __m512 vi0x0123456789ABCDEF = _mm512_loadu_ps(i0); in xnn_f32_dwconv_ukernel_up32x9__avx512f() local 67 …vacc0123456789ABCDEFp0 = _mm512_fmadd_ps(vi0x0123456789ABCDEF, vk0x0123456789ABCDEF, vacc012345678… in xnn_f32_dwconv_ukernel_up32x9__avx512f() 157 const __m512 vi0x0123456789ABCDEF = _mm512_loadu_ps(i0); in xnn_f32_dwconv_ukernel_up32x9__avx512f() local 161 …vacc0123456789ABCDEFp0 = _mm512_fmadd_ps(vi0x0123456789ABCDEF, vk0x0123456789ABCDEF, vacc012345678… in xnn_f32_dwconv_ukernel_up32x9__avx512f() 228 const __m512 vi0x0123456789ABCDEF = _mm512_maskz_loadu_ps(vmask, i0); in xnn_f32_dwconv_ukernel_up32x9__avx512f() local 230 …vacc0123456789ABCDEFp0 = _mm512_fmadd_ps(vi0x0123456789ABCDEF, vk0x0123456789ABCDEF, vacc012345678… in xnn_f32_dwconv_ukernel_up32x9__avx512f()
|
D | up32x9-avx512f-acc2.c | 61 const __m512 vi0x0123456789ABCDEF = _mm512_loadu_ps(i0); in xnn_f32_dwconv_ukernel_up32x9__avx512f_acc2() local 67 …vacc0123456789ABCDEFp0 = _mm512_fmadd_ps(vi0x0123456789ABCDEF, vk0x0123456789ABCDEF, vacc012345678… in xnn_f32_dwconv_ukernel_up32x9__avx512f_acc2() 160 const __m512 vi0x0123456789ABCDEF = _mm512_loadu_ps(i0); in xnn_f32_dwconv_ukernel_up32x9__avx512f_acc2() local 164 …vacc0123456789ABCDEFp0 = _mm512_fmadd_ps(vi0x0123456789ABCDEF, vk0x0123456789ABCDEF, vacc012345678… in xnn_f32_dwconv_ukernel_up32x9__avx512f_acc2() 233 const __m512 vi0x0123456789ABCDEF = _mm512_maskz_loadu_ps(vmask, i0); in xnn_f32_dwconv_ukernel_up32x9__avx512f_acc2() local 235 …vacc0123456789ABCDEFp0 = _mm512_fmadd_ps(vi0x0123456789ABCDEF, vk0x0123456789ABCDEF, vacc012345678… in xnn_f32_dwconv_ukernel_up32x9__avx512f_acc2()
|
D | up16x9-avx512f.c | 60 const __m512 vi0x0123456789ABCDEF = _mm512_loadu_ps(i0); in xnn_f32_dwconv_ukernel_up16x9__avx512f() local 64 …vacc0123456789ABCDEFp0 = _mm512_fmadd_ps(vi0x0123456789ABCDEF, vk0x0123456789ABCDEF, vacc012345678… in xnn_f32_dwconv_ukernel_up16x9__avx512f() 131 const __m512 vi0x0123456789ABCDEF = _mm512_maskz_loadu_ps(vmask, i0); in xnn_f32_dwconv_ukernel_up16x9__avx512f() local 133 …vacc0123456789ABCDEFp0 = _mm512_fmadd_ps(vi0x0123456789ABCDEF, vk0x0123456789ABCDEF, vacc012345678… in xnn_f32_dwconv_ukernel_up16x9__avx512f()
|
D | up16x9-avx512f-acc2.c | 60 const __m512 vi0x0123456789ABCDEF = _mm512_loadu_ps(i0); in xnn_f32_dwconv_ukernel_up16x9__avx512f_acc2() local 64 …vacc0123456789ABCDEFp0 = _mm512_fmadd_ps(vi0x0123456789ABCDEF, vk0x0123456789ABCDEF, vacc012345678… in xnn_f32_dwconv_ukernel_up16x9__avx512f_acc2() 133 const __m512 vi0x0123456789ABCDEF = _mm512_maskz_loadu_ps(vmask, i0); in xnn_f32_dwconv_ukernel_up16x9__avx512f_acc2() local 135 …vacc0123456789ABCDEFp0 = _mm512_fmadd_ps(vi0x0123456789ABCDEF, vk0x0123456789ABCDEF, vacc012345678… in xnn_f32_dwconv_ukernel_up16x9__avx512f_acc2()
|
D | up16x25-avx512f.c | 92 const __m512 vi0x0123456789ABCDEF = _mm512_loadu_ps(i0); in xnn_f32_dwconv_ukernel_up16x25__avx512f() local 96 …vacc0123456789ABCDEFp0 = _mm512_fmadd_ps(vi0x0123456789ABCDEF, vk0x0123456789ABCDEF, vacc012345678… in xnn_f32_dwconv_ukernel_up16x25__avx512f() 259 const __m512 vi0x0123456789ABCDEF = _mm512_maskz_loadu_ps(vmask, i0); in xnn_f32_dwconv_ukernel_up16x25__avx512f() local 261 …vacc0123456789ABCDEFp0 = _mm512_fmadd_ps(vi0x0123456789ABCDEF, vk0x0123456789ABCDEF, vacc012345678… in xnn_f32_dwconv_ukernel_up16x25__avx512f()
|
D | up16x25-avx512f-acc2.c | 92 const __m512 vi0x0123456789ABCDEF = _mm512_loadu_ps(i0); in xnn_f32_dwconv_ukernel_up16x25__avx512f_acc2() local 96 …vacc0123456789ABCDEFp0 = _mm512_fmadd_ps(vi0x0123456789ABCDEF, vk0x0123456789ABCDEF, vacc012345678… in xnn_f32_dwconv_ukernel_up16x25__avx512f_acc2() 261 const __m512 vi0x0123456789ABCDEF = _mm512_maskz_loadu_ps(vmask, i0); in xnn_f32_dwconv_ukernel_up16x25__avx512f_acc2() local 263 …vacc0123456789ABCDEFp0 = _mm512_fmadd_ps(vi0x0123456789ABCDEF, vk0x0123456789ABCDEF, vacc012345678… in xnn_f32_dwconv_ukernel_up16x25__avx512f_acc2()
|
D | up32x25-avx512f-acc2.c | 93 const __m512 vi0x0123456789ABCDEF = _mm512_loadu_ps(i0); in xnn_f32_dwconv_ukernel_up32x25__avx512f_acc2() local 99 …vacc0123456789ABCDEFp0 = _mm512_fmadd_ps(vi0x0123456789ABCDEF, vk0x0123456789ABCDEF, vacc012345678… in xnn_f32_dwconv_ukernel_up32x25__avx512f_acc2() 336 const __m512 vi0x0123456789ABCDEF = _mm512_loadu_ps(i0); in xnn_f32_dwconv_ukernel_up32x25__avx512f_acc2() local 340 …vacc0123456789ABCDEFp0 = _mm512_fmadd_ps(vi0x0123456789ABCDEF, vk0x0123456789ABCDEF, vacc012345678… in xnn_f32_dwconv_ukernel_up32x25__avx512f_acc2() 505 const __m512 vi0x0123456789ABCDEF = _mm512_maskz_loadu_ps(vmask, i0); in xnn_f32_dwconv_ukernel_up32x25__avx512f_acc2() local 507 …vacc0123456789ABCDEFp0 = _mm512_fmadd_ps(vi0x0123456789ABCDEF, vk0x0123456789ABCDEF, vacc012345678… in xnn_f32_dwconv_ukernel_up32x25__avx512f_acc2()
|
D | up32x25-avx512f.c | 93 const __m512 vi0x0123456789ABCDEF = _mm512_loadu_ps(i0); in xnn_f32_dwconv_ukernel_up32x25__avx512f() local 99 …vacc0123456789ABCDEFp0 = _mm512_fmadd_ps(vi0x0123456789ABCDEF, vk0x0123456789ABCDEF, vacc012345678… in xnn_f32_dwconv_ukernel_up32x25__avx512f() 333 const __m512 vi0x0123456789ABCDEF = _mm512_loadu_ps(i0); in xnn_f32_dwconv_ukernel_up32x25__avx512f() local 337 …vacc0123456789ABCDEFp0 = _mm512_fmadd_ps(vi0x0123456789ABCDEF, vk0x0123456789ABCDEF, vacc012345678… in xnn_f32_dwconv_ukernel_up32x25__avx512f() 500 const __m512 vi0x0123456789ABCDEF = _mm512_maskz_loadu_ps(vmask, i0); in xnn_f32_dwconv_ukernel_up32x25__avx512f() local 502 …vacc0123456789ABCDEFp0 = _mm512_fmadd_ps(vi0x0123456789ABCDEF, vk0x0123456789ABCDEF, vacc012345678… in xnn_f32_dwconv_ukernel_up32x25__avx512f()
|