/external/XNNPACK/src/f32-dwconv/gen/ |
D | up32x9-avx512f.c | 137 const __m512 vk8x0123456789ABCDEF = _mm512_load_ps(w + 288); in xnn_f32_dwconv_ukernel_up32x9__avx512f() local 139 …vacc0123456789ABCDEFp0 = _mm512_fmadd_ps(vi8x0123456789ABCDEF, vk8x0123456789ABCDEF, vacc012345678… in xnn_f32_dwconv_ukernel_up32x9__avx512f() 208 const __m512 vk8x0123456789ABCDEF = _mm512_load_ps(w + 288); in xnn_f32_dwconv_ukernel_up32x9__avx512f() local 209 …vacc0123456789ABCDEFp0 = _mm512_fmadd_ps(vi8x0123456789ABCDEF, vk8x0123456789ABCDEF, vacc012345678… in xnn_f32_dwconv_ukernel_up32x9__avx512f() 261 const __m512 vk8x0123456789ABCDEF = _mm512_maskz_loadu_ps(vmask, w + 288); in xnn_f32_dwconv_ukernel_up32x9__avx512f() local 262 …vacc0123456789ABCDEFp0 = _mm512_fmadd_ps(vi8x0123456789ABCDEF, vk8x0123456789ABCDEF, vacc012345678… in xnn_f32_dwconv_ukernel_up32x9__avx512f()
|
D | up32x9-avx512f-acc2.c | 137 const __m512 vk8x0123456789ABCDEF = _mm512_load_ps(w + 288); in xnn_f32_dwconv_ukernel_up32x9__avx512f_acc2() local 139 …vacc0123456789ABCDEFp0 = _mm512_fmadd_ps(vi8x0123456789ABCDEF, vk8x0123456789ABCDEF, vacc012345678… in xnn_f32_dwconv_ukernel_up32x9__avx512f_acc2() 211 const __m512 vk8x0123456789ABCDEF = _mm512_load_ps(w + 288); in xnn_f32_dwconv_ukernel_up32x9__avx512f_acc2() local 212 …vacc0123456789ABCDEFp0 = _mm512_fmadd_ps(vi8x0123456789ABCDEF, vk8x0123456789ABCDEF, vacc012345678… in xnn_f32_dwconv_ukernel_up32x9__avx512f_acc2() 266 const __m512 vk8x0123456789ABCDEF = _mm512_maskz_loadu_ps(vmask, w + 288); in xnn_f32_dwconv_ukernel_up32x9__avx512f_acc2() local 267 …vacc0123456789ABCDEFp0 = _mm512_fmadd_ps(vi8x0123456789ABCDEF, vk8x0123456789ABCDEF, vacc012345678… in xnn_f32_dwconv_ukernel_up32x9__avx512f_acc2()
|
D | up16x9-avx512f.c | 111 const __m512 vk8x0123456789ABCDEF = _mm512_load_ps(w + 144); in xnn_f32_dwconv_ukernel_up16x9__avx512f() local 112 …vacc0123456789ABCDEFp0 = _mm512_fmadd_ps(vi8x0123456789ABCDEF, vk8x0123456789ABCDEF, vacc012345678… in xnn_f32_dwconv_ukernel_up16x9__avx512f() 164 const __m512 vk8x0123456789ABCDEF = _mm512_maskz_loadu_ps(vmask, w + 144); in xnn_f32_dwconv_ukernel_up16x9__avx512f() local 165 …vacc0123456789ABCDEFp0 = _mm512_fmadd_ps(vi8x0123456789ABCDEF, vk8x0123456789ABCDEF, vacc012345678… in xnn_f32_dwconv_ukernel_up16x9__avx512f()
|
D | up16x9-avx512f-acc2.c | 111 const __m512 vk8x0123456789ABCDEF = _mm512_load_ps(w + 144); in xnn_f32_dwconv_ukernel_up16x9__avx512f_acc2() local 112 …vacc0123456789ABCDEFp0 = _mm512_fmadd_ps(vi8x0123456789ABCDEF, vk8x0123456789ABCDEF, vacc012345678… in xnn_f32_dwconv_ukernel_up16x9__avx512f_acc2() 166 const __m512 vk8x0123456789ABCDEF = _mm512_maskz_loadu_ps(vmask, w + 144); in xnn_f32_dwconv_ukernel_up16x9__avx512f_acc2() local 167 …vacc0123456789ABCDEFp0 = _mm512_fmadd_ps(vi8x0123456789ABCDEF, vk8x0123456789ABCDEF, vacc012345678… in xnn_f32_dwconv_ukernel_up16x9__avx512f_acc2()
|
D | up16x25-avx512f.c | 143 const __m512 vk8x0123456789ABCDEF = _mm512_load_ps(w + 144); in xnn_f32_dwconv_ukernel_up16x25__avx512f() local 144 …vacc0123456789ABCDEFp0 = _mm512_fmadd_ps(vi8x0123456789ABCDEF, vk8x0123456789ABCDEF, vacc012345678… in xnn_f32_dwconv_ukernel_up16x25__avx512f() 292 const __m512 vk8x0123456789ABCDEF = _mm512_maskz_loadu_ps(vmask, w + 144); in xnn_f32_dwconv_ukernel_up16x25__avx512f() local 293 …vacc0123456789ABCDEFp0 = _mm512_fmadd_ps(vi8x0123456789ABCDEF, vk8x0123456789ABCDEF, vacc012345678… in xnn_f32_dwconv_ukernel_up16x25__avx512f()
|
D | up16x25-avx512f-acc2.c | 143 const __m512 vk8x0123456789ABCDEF = _mm512_load_ps(w + 144); in xnn_f32_dwconv_ukernel_up16x25__avx512f_acc2() local 144 …vacc0123456789ABCDEFp0 = _mm512_fmadd_ps(vi8x0123456789ABCDEF, vk8x0123456789ABCDEF, vacc012345678… in xnn_f32_dwconv_ukernel_up16x25__avx512f_acc2() 294 const __m512 vk8x0123456789ABCDEF = _mm512_maskz_loadu_ps(vmask, w + 144); in xnn_f32_dwconv_ukernel_up16x25__avx512f_acc2() local 295 …vacc0123456789ABCDEFp0 = _mm512_fmadd_ps(vi8x0123456789ABCDEF, vk8x0123456789ABCDEF, vacc012345678… in xnn_f32_dwconv_ukernel_up16x25__avx512f_acc2()
|
D | up32x25-avx512f-acc2.c | 169 const __m512 vk8x0123456789ABCDEF = _mm512_load_ps(w + 288); in xnn_f32_dwconv_ukernel_up32x25__avx512f_acc2() local 171 …vacc0123456789ABCDEFp0 = _mm512_fmadd_ps(vi8x0123456789ABCDEF, vk8x0123456789ABCDEF, vacc012345678… in xnn_f32_dwconv_ukernel_up32x25__avx512f_acc2() 387 const __m512 vk8x0123456789ABCDEF = _mm512_load_ps(w + 288); in xnn_f32_dwconv_ukernel_up32x25__avx512f_acc2() local 388 …vacc0123456789ABCDEFp0 = _mm512_fmadd_ps(vi8x0123456789ABCDEF, vk8x0123456789ABCDEF, vacc012345678… in xnn_f32_dwconv_ukernel_up32x25__avx512f_acc2() 538 const __m512 vk8x0123456789ABCDEF = _mm512_maskz_loadu_ps(vmask, w + 288); in xnn_f32_dwconv_ukernel_up32x25__avx512f_acc2() local 539 …vacc0123456789ABCDEFp0 = _mm512_fmadd_ps(vi8x0123456789ABCDEF, vk8x0123456789ABCDEF, vacc012345678… in xnn_f32_dwconv_ukernel_up32x25__avx512f_acc2()
|
D | up32x25-avx512f.c | 169 const __m512 vk8x0123456789ABCDEF = _mm512_load_ps(w + 288); in xnn_f32_dwconv_ukernel_up32x25__avx512f() local 171 …vacc0123456789ABCDEFp0 = _mm512_fmadd_ps(vi8x0123456789ABCDEF, vk8x0123456789ABCDEF, vacc012345678… in xnn_f32_dwconv_ukernel_up32x25__avx512f() 384 const __m512 vk8x0123456789ABCDEF = _mm512_load_ps(w + 288); in xnn_f32_dwconv_ukernel_up32x25__avx512f() local 385 …vacc0123456789ABCDEFp0 = _mm512_fmadd_ps(vi8x0123456789ABCDEF, vk8x0123456789ABCDEF, vacc012345678… in xnn_f32_dwconv_ukernel_up32x25__avx512f() 533 const __m512 vk8x0123456789ABCDEF = _mm512_maskz_loadu_ps(vmask, w + 288); in xnn_f32_dwconv_ukernel_up32x25__avx512f() local 534 …vacc0123456789ABCDEFp0 = _mm512_fmadd_ps(vi8x0123456789ABCDEF, vk8x0123456789ABCDEF, vacc012345678… in xnn_f32_dwconv_ukernel_up32x25__avx512f()
|