/external/XNNPACK/src/f32-dwconv/gen/ |
D | up32x9-minmax-avx512f-acc2.c | 139 const __m512 vk5x0123456789ABCDEF = _mm512_load_ps(w + 192); in xnn_f32_dwconv_minmax_ukernel_up32x9__avx512f_acc2() local 141 …vacc0123456789ABCDEFp1 = _mm512_fmadd_ps(vi5x0123456789ABCDEF, vk5x0123456789ABCDEF, vacc012345678… in xnn_f32_dwconv_minmax_ukernel_up32x9__avx512f_acc2() 222 const __m512 vk5x0123456789ABCDEF = _mm512_load_ps(w + 192); in xnn_f32_dwconv_minmax_ukernel_up32x9__avx512f_acc2() local 223 …vacc0123456789ABCDEFp1 = _mm512_fmadd_ps(vi5x0123456789ABCDEF, vk5x0123456789ABCDEF, vacc012345678… in xnn_f32_dwconv_minmax_ukernel_up32x9__avx512f_acc2() 283 const __m512 vk5x0123456789ABCDEF = _mm512_maskz_loadu_ps(vmask, w + 192); in xnn_f32_dwconv_minmax_ukernel_up32x9__avx512f_acc2() local 284 …vacc0123456789ABCDEFp1 = _mm512_fmadd_ps(vi5x0123456789ABCDEF, vk5x0123456789ABCDEF, vacc012345678… in xnn_f32_dwconv_minmax_ukernel_up32x9__avx512f_acc2()
|
D | up32x9-minmax-avx512f.c | 139 const __m512 vk5x0123456789ABCDEF = _mm512_load_ps(w + 192); in xnn_f32_dwconv_minmax_ukernel_up32x9__avx512f() local 141 …vacc0123456789ABCDEFp0 = _mm512_fmadd_ps(vi5x0123456789ABCDEF, vk5x0123456789ABCDEF, vacc012345678… in xnn_f32_dwconv_minmax_ukernel_up32x9__avx512f() 219 const __m512 vk5x0123456789ABCDEF = _mm512_load_ps(w + 192); in xnn_f32_dwconv_minmax_ukernel_up32x9__avx512f() local 220 …vacc0123456789ABCDEFp0 = _mm512_fmadd_ps(vi5x0123456789ABCDEF, vk5x0123456789ABCDEF, vacc012345678… in xnn_f32_dwconv_minmax_ukernel_up32x9__avx512f() 278 const __m512 vk5x0123456789ABCDEF = _mm512_maskz_loadu_ps(vmask, w + 192); in xnn_f32_dwconv_minmax_ukernel_up32x9__avx512f() local 279 …vacc0123456789ABCDEFp0 = _mm512_fmadd_ps(vi5x0123456789ABCDEF, vk5x0123456789ABCDEF, vacc012345678… in xnn_f32_dwconv_minmax_ukernel_up32x9__avx512f()
|
D | up16x9-minmax-avx512f-acc2.c | 122 const __m512 vk5x0123456789ABCDEF = _mm512_load_ps(w + 96); in xnn_f32_dwconv_minmax_ukernel_up16x9__avx512f_acc2() local 123 …vacc0123456789ABCDEFp1 = _mm512_fmadd_ps(vi5x0123456789ABCDEF, vk5x0123456789ABCDEF, vacc012345678… in xnn_f32_dwconv_minmax_ukernel_up16x9__avx512f_acc2() 183 const __m512 vk5x0123456789ABCDEF = _mm512_maskz_loadu_ps(vmask, w + 96); in xnn_f32_dwconv_minmax_ukernel_up16x9__avx512f_acc2() local 184 …vacc0123456789ABCDEFp1 = _mm512_fmadd_ps(vi5x0123456789ABCDEF, vk5x0123456789ABCDEF, vacc012345678… in xnn_f32_dwconv_minmax_ukernel_up16x9__avx512f_acc2()
|
D | up16x9-minmax-avx512f.c | 122 const __m512 vk5x0123456789ABCDEF = _mm512_load_ps(w + 96); in xnn_f32_dwconv_minmax_ukernel_up16x9__avx512f() local 123 …vacc0123456789ABCDEFp0 = _mm512_fmadd_ps(vi5x0123456789ABCDEF, vk5x0123456789ABCDEF, vacc012345678… in xnn_f32_dwconv_minmax_ukernel_up16x9__avx512f() 181 const __m512 vk5x0123456789ABCDEF = _mm512_maskz_loadu_ps(vmask, w + 96); in xnn_f32_dwconv_minmax_ukernel_up16x9__avx512f() local 182 …vacc0123456789ABCDEFp0 = _mm512_fmadd_ps(vi5x0123456789ABCDEF, vk5x0123456789ABCDEF, vacc012345678… in xnn_f32_dwconv_minmax_ukernel_up16x9__avx512f()
|
D | up16x25-minmax-avx512f.c | 202 const __m512 vk5x0123456789ABCDEF = _mm512_load_ps(w + 96); in xnn_f32_dwconv_minmax_ukernel_up16x25__avx512f() local 203 …vacc0123456789ABCDEFp0 = _mm512_fmadd_ps(vi5x0123456789ABCDEF, vk5x0123456789ABCDEF, vacc012345678… in xnn_f32_dwconv_minmax_ukernel_up16x25__avx512f() 357 const __m512 vk5x0123456789ABCDEF = _mm512_maskz_loadu_ps(vmask, w + 96); in xnn_f32_dwconv_minmax_ukernel_up16x25__avx512f() local 358 …vacc0123456789ABCDEFp0 = _mm512_fmadd_ps(vi5x0123456789ABCDEF, vk5x0123456789ABCDEF, vacc012345678… in xnn_f32_dwconv_minmax_ukernel_up16x25__avx512f()
|
D | up16x25-minmax-avx512f-acc2.c | 202 const __m512 vk5x0123456789ABCDEF = _mm512_load_ps(w + 96); in xnn_f32_dwconv_minmax_ukernel_up16x25__avx512f_acc2() local 203 …vacc0123456789ABCDEFp1 = _mm512_fmadd_ps(vi5x0123456789ABCDEF, vk5x0123456789ABCDEF, vacc012345678… in xnn_f32_dwconv_minmax_ukernel_up16x25__avx512f_acc2() 359 const __m512 vk5x0123456789ABCDEF = _mm512_maskz_loadu_ps(vmask, w + 96); in xnn_f32_dwconv_minmax_ukernel_up16x25__avx512f_acc2() local 360 …vacc0123456789ABCDEFp1 = _mm512_fmadd_ps(vi5x0123456789ABCDEF, vk5x0123456789ABCDEF, vacc012345678… in xnn_f32_dwconv_minmax_ukernel_up16x25__avx512f_acc2()
|
D | up32x25-minmax-avx512f-acc2.c | 219 const __m512 vk5x0123456789ABCDEF = _mm512_load_ps(w + 192); in xnn_f32_dwconv_minmax_ukernel_up32x25__avx512f_acc2() local 221 …vacc0123456789ABCDEFp1 = _mm512_fmadd_ps(vi5x0123456789ABCDEF, vk5x0123456789ABCDEF, vacc012345678… in xnn_f32_dwconv_minmax_ukernel_up32x25__avx512f_acc2() 446 const __m512 vk5x0123456789ABCDEF = _mm512_load_ps(w + 192); in xnn_f32_dwconv_minmax_ukernel_up32x25__avx512f_acc2() local 447 …vacc0123456789ABCDEFp1 = _mm512_fmadd_ps(vi5x0123456789ABCDEF, vk5x0123456789ABCDEF, vacc012345678… in xnn_f32_dwconv_minmax_ukernel_up32x25__avx512f_acc2() 603 const __m512 vk5x0123456789ABCDEF = _mm512_maskz_loadu_ps(vmask, w + 192); in xnn_f32_dwconv_minmax_ukernel_up32x25__avx512f_acc2() local 604 …vacc0123456789ABCDEFp1 = _mm512_fmadd_ps(vi5x0123456789ABCDEF, vk5x0123456789ABCDEF, vacc012345678… in xnn_f32_dwconv_minmax_ukernel_up32x25__avx512f_acc2()
|
D | up32x25-minmax-avx512f.c | 219 const __m512 vk5x0123456789ABCDEF = _mm512_load_ps(w + 192); in xnn_f32_dwconv_minmax_ukernel_up32x25__avx512f() local 221 …vacc0123456789ABCDEFp0 = _mm512_fmadd_ps(vi5x0123456789ABCDEF, vk5x0123456789ABCDEF, vacc012345678… in xnn_f32_dwconv_minmax_ukernel_up32x25__avx512f() 443 const __m512 vk5x0123456789ABCDEF = _mm512_load_ps(w + 192); in xnn_f32_dwconv_minmax_ukernel_up32x25__avx512f() local 444 …vacc0123456789ABCDEFp0 = _mm512_fmadd_ps(vi5x0123456789ABCDEF, vk5x0123456789ABCDEF, vacc012345678… in xnn_f32_dwconv_minmax_ukernel_up32x25__avx512f() 598 const __m512 vk5x0123456789ABCDEF = _mm512_maskz_loadu_ps(vmask, w + 192); in xnn_f32_dwconv_minmax_ukernel_up32x25__avx512f() local 599 …vacc0123456789ABCDEFp0 = _mm512_fmadd_ps(vi5x0123456789ABCDEF, vk5x0123456789ABCDEF, vacc012345678… in xnn_f32_dwconv_minmax_ukernel_up32x25__avx512f()
|
/external/XNNPACK/src/qs8-dwconv/gen/ |
D | up16x9-minmax-avx512skx-mul32.c | 128 …const __m512i vk5x0123456789ABCDEF = _mm512_cvtepi8_epi32(_mm_load_si128((const __m128i*) ((uintpt… in xnn_qs8_dwconv_minmax_ukernel_up16x9__avx512skx_mul32() local 131 …12_add_epi32(vacc0123456789ABCDEF, _mm512_mullo_epi32(vi5x0123456789ABCDEF, vk5x0123456789ABCDEF)); in xnn_qs8_dwconv_minmax_ukernel_up16x9__avx512skx_mul32() 215 …const __m512i vk5x0123456789ABCDEF = _mm512_cvtepi8_epi32(_mm_loadu_si128((const __m128i*) ((uintp… in xnn_qs8_dwconv_minmax_ukernel_up16x9__avx512skx_mul32() local 217 …12_add_epi32(vacc0123456789ABCDEF, _mm512_mullo_epi32(vi5x0123456789ABCDEF, vk5x0123456789ABCDEF)); in xnn_qs8_dwconv_minmax_ukernel_up16x9__avx512skx_mul32()
|
D | up32x9-minmax-avx512skx-mul32.c | 145 …const __m512i vk5x0123456789ABCDEF = _mm512_cvtepi8_epi32(_mm_load_si128((const __m128i*) ((uintpt… in xnn_qs8_dwconv_minmax_ukernel_up32x9__avx512skx_mul32() local 150 …12_add_epi32(vacc0123456789ABCDEF, _mm512_mullo_epi32(vi5x0123456789ABCDEF, vk5x0123456789ABCDEF)); in xnn_qs8_dwconv_minmax_ukernel_up32x9__avx512skx_mul32() 267 …const __m512i vk5x0123456789ABCDEF = _mm512_cvtepi8_epi32(_mm_loadu_si128((const __m128i*) (k + 16… in xnn_qs8_dwconv_minmax_ukernel_up32x9__avx512skx_mul32() local 270 …12_add_epi32(vacc0123456789ABCDEF, _mm512_mullo_epi32(vi5x0123456789ABCDEF, vk5x0123456789ABCDEF)); in xnn_qs8_dwconv_minmax_ukernel_up32x9__avx512skx_mul32()
|
D | up16x9-minmax-avx2-mul16.c | 133 …const __m256i vk5x0123456789ABCDEF = _mm256_cvtepi8_epi16(_mm_loadu_si128((const __m128i*) ((uintp… in xnn_qs8_dwconv_minmax_ukernel_up16x9__avx2_mul16() local 136 … __m256i vprod5x0123456789ABCDEF = _mm256_mullo_epi16(vi5x0123456789ABCDEF, vk5x0123456789ABCDEF); in xnn_qs8_dwconv_minmax_ukernel_up16x9__avx2_mul16() 261 …const __m256i vk5x0123456789ABCDEF = _mm256_cvtepi8_epi16(_mm_loadu_si128((const __m128i*) ((uintp… in xnn_qs8_dwconv_minmax_ukernel_up16x9__avx2_mul16() local 263 …t __m256i vprod5x0123456789ABCDEF = _mm256_mullo_epi16(vi5x0123456789ABCDEF, vk5x0123456789ABCDEF); in xnn_qs8_dwconv_minmax_ukernel_up16x9__avx2_mul16()
|
D | up32x9-minmax-avx2-mul16.c | 165 …const __m256i vk5x0123456789ABCDEF = _mm256_cvtepi8_epi16(_mm_loadu_si128((const __m128i*) ((uintp… in xnn_qs8_dwconv_minmax_ukernel_up32x9__avx2_mul16() local 170 … __m256i vprod5x0123456789ABCDEF = _mm256_mullo_epi16(vi5x0123456789ABCDEF, vk5x0123456789ABCDEF); in xnn_qs8_dwconv_minmax_ukernel_up32x9__avx2_mul16() 347 …const __m256i vk5x0123456789ABCDEF = _mm256_cvtepi8_epi16(_mm_loadu_si128((const __m128i*) (k + 16… in xnn_qs8_dwconv_minmax_ukernel_up32x9__avx2_mul16() local 350 …t __m256i vprod5x0123456789ABCDEF = _mm256_mullo_epi16(vi5x0123456789ABCDEF, vk5x0123456789ABCDEF); in xnn_qs8_dwconv_minmax_ukernel_up32x9__avx2_mul16()
|