/external/XNNPACK/src/f32-dwconv/gen/ |
D | up32x9-minmax-avx512f.c | 144 const __m512 vi6x0123456789ABCDEF = _mm512_loadu_ps(i6); in xnn_f32_dwconv_minmax_ukernel_up32x9__avx512f() local 150 …vacc0123456789ABCDEFp0 = _mm512_fmadd_ps(vi6x0123456789ABCDEF, vk6x0123456789ABCDEF, vacc012345678… in xnn_f32_dwconv_minmax_ukernel_up32x9__avx512f() 222 const __m512 vi6x0123456789ABCDEF = _mm512_loadu_ps(i6); in xnn_f32_dwconv_minmax_ukernel_up32x9__avx512f() local 226 …vacc0123456789ABCDEFp0 = _mm512_fmadd_ps(vi6x0123456789ABCDEF, vk6x0123456789ABCDEF, vacc012345678… in xnn_f32_dwconv_minmax_ukernel_up32x9__avx512f() 281 const __m512 vi6x0123456789ABCDEF = _mm512_maskz_loadu_ps(vmask, i6); in xnn_f32_dwconv_minmax_ukernel_up32x9__avx512f() local 283 …vacc0123456789ABCDEFp0 = _mm512_fmadd_ps(vi6x0123456789ABCDEF, vk6x0123456789ABCDEF, vacc012345678… in xnn_f32_dwconv_minmax_ukernel_up32x9__avx512f()
|
D | up32x9-minmax-avx512f-acc2.c | 144 const __m512 vi6x0123456789ABCDEF = _mm512_loadu_ps(i6); in xnn_f32_dwconv_minmax_ukernel_up32x9__avx512f_acc2() local 150 …vacc0123456789ABCDEFp0 = _mm512_fmadd_ps(vi6x0123456789ABCDEF, vk6x0123456789ABCDEF, vacc012345678… in xnn_f32_dwconv_minmax_ukernel_up32x9__avx512f_acc2() 225 const __m512 vi6x0123456789ABCDEF = _mm512_loadu_ps(i6); in xnn_f32_dwconv_minmax_ukernel_up32x9__avx512f_acc2() local 229 …vacc0123456789ABCDEFp0 = _mm512_fmadd_ps(vi6x0123456789ABCDEF, vk6x0123456789ABCDEF, vacc012345678… in xnn_f32_dwconv_minmax_ukernel_up32x9__avx512f_acc2() 286 const __m512 vi6x0123456789ABCDEF = _mm512_maskz_loadu_ps(vmask, i6); in xnn_f32_dwconv_minmax_ukernel_up32x9__avx512f_acc2() local 288 …vacc0123456789ABCDEFp0 = _mm512_fmadd_ps(vi6x0123456789ABCDEF, vk6x0123456789ABCDEF, vacc012345678… in xnn_f32_dwconv_minmax_ukernel_up32x9__avx512f_acc2()
|
D | up16x9-minmax-avx512f-acc2.c | 125 const __m512 vi6x0123456789ABCDEF = _mm512_loadu_ps(i6); in xnn_f32_dwconv_minmax_ukernel_up16x9__avx512f_acc2() local 129 …vacc0123456789ABCDEFp0 = _mm512_fmadd_ps(vi6x0123456789ABCDEF, vk6x0123456789ABCDEF, vacc012345678… in xnn_f32_dwconv_minmax_ukernel_up16x9__avx512f_acc2() 186 const __m512 vi6x0123456789ABCDEF = _mm512_maskz_loadu_ps(vmask, i6); in xnn_f32_dwconv_minmax_ukernel_up16x9__avx512f_acc2() local 188 …vacc0123456789ABCDEFp0 = _mm512_fmadd_ps(vi6x0123456789ABCDEF, vk6x0123456789ABCDEF, vacc012345678… in xnn_f32_dwconv_minmax_ukernel_up16x9__avx512f_acc2()
|
D | up16x9-minmax-avx512f.c | 125 const __m512 vi6x0123456789ABCDEF = _mm512_loadu_ps(i6); in xnn_f32_dwconv_minmax_ukernel_up16x9__avx512f() local 129 …vacc0123456789ABCDEFp0 = _mm512_fmadd_ps(vi6x0123456789ABCDEF, vk6x0123456789ABCDEF, vacc012345678… in xnn_f32_dwconv_minmax_ukernel_up16x9__avx512f() 184 const __m512 vi6x0123456789ABCDEF = _mm512_maskz_loadu_ps(vmask, i6); in xnn_f32_dwconv_minmax_ukernel_up16x9__avx512f() local 186 …vacc0123456789ABCDEFp0 = _mm512_fmadd_ps(vi6x0123456789ABCDEF, vk6x0123456789ABCDEF, vacc012345678… in xnn_f32_dwconv_minmax_ukernel_up16x9__avx512f()
|
D | up16x25-minmax-avx512f-acc2.c | 205 const __m512 vi6x0123456789ABCDEF = _mm512_loadu_ps(i6); in xnn_f32_dwconv_minmax_ukernel_up16x25__avx512f_acc2() local 209 …vacc0123456789ABCDEFp0 = _mm512_fmadd_ps(vi6x0123456789ABCDEF, vk6x0123456789ABCDEF, vacc012345678… in xnn_f32_dwconv_minmax_ukernel_up16x25__avx512f_acc2() 362 const __m512 vi6x0123456789ABCDEF = _mm512_maskz_loadu_ps(vmask, i6); in xnn_f32_dwconv_minmax_ukernel_up16x25__avx512f_acc2() local 364 …vacc0123456789ABCDEFp0 = _mm512_fmadd_ps(vi6x0123456789ABCDEF, vk6x0123456789ABCDEF, vacc012345678… in xnn_f32_dwconv_minmax_ukernel_up16x25__avx512f_acc2()
|
D | up16x25-minmax-avx512f.c | 205 const __m512 vi6x0123456789ABCDEF = _mm512_loadu_ps(i6); in xnn_f32_dwconv_minmax_ukernel_up16x25__avx512f() local 209 …vacc0123456789ABCDEFp0 = _mm512_fmadd_ps(vi6x0123456789ABCDEF, vk6x0123456789ABCDEF, vacc012345678… in xnn_f32_dwconv_minmax_ukernel_up16x25__avx512f() 360 const __m512 vi6x0123456789ABCDEF = _mm512_maskz_loadu_ps(vmask, i6); in xnn_f32_dwconv_minmax_ukernel_up16x25__avx512f() local 362 …vacc0123456789ABCDEFp0 = _mm512_fmadd_ps(vi6x0123456789ABCDEF, vk6x0123456789ABCDEF, vacc012345678… in xnn_f32_dwconv_minmax_ukernel_up16x25__avx512f()
|
D | up32x25-minmax-avx512f.c | 224 const __m512 vi6x0123456789ABCDEF = _mm512_loadu_ps(i6); in xnn_f32_dwconv_minmax_ukernel_up32x25__avx512f() local 230 …vacc0123456789ABCDEFp0 = _mm512_fmadd_ps(vi6x0123456789ABCDEF, vk6x0123456789ABCDEF, vacc012345678… in xnn_f32_dwconv_minmax_ukernel_up32x25__avx512f() 446 const __m512 vi6x0123456789ABCDEF = _mm512_loadu_ps(i6); in xnn_f32_dwconv_minmax_ukernel_up32x25__avx512f() local 450 …vacc0123456789ABCDEFp0 = _mm512_fmadd_ps(vi6x0123456789ABCDEF, vk6x0123456789ABCDEF, vacc012345678… in xnn_f32_dwconv_minmax_ukernel_up32x25__avx512f() 601 const __m512 vi6x0123456789ABCDEF = _mm512_maskz_loadu_ps(vmask, i6); in xnn_f32_dwconv_minmax_ukernel_up32x25__avx512f() local 603 …vacc0123456789ABCDEFp0 = _mm512_fmadd_ps(vi6x0123456789ABCDEF, vk6x0123456789ABCDEF, vacc012345678… in xnn_f32_dwconv_minmax_ukernel_up32x25__avx512f()
|
D | up32x25-minmax-avx512f-acc2.c | 224 const __m512 vi6x0123456789ABCDEF = _mm512_loadu_ps(i6); in xnn_f32_dwconv_minmax_ukernel_up32x25__avx512f_acc2() local 230 …vacc0123456789ABCDEFp0 = _mm512_fmadd_ps(vi6x0123456789ABCDEF, vk6x0123456789ABCDEF, vacc012345678… in xnn_f32_dwconv_minmax_ukernel_up32x25__avx512f_acc2() 449 const __m512 vi6x0123456789ABCDEF = _mm512_loadu_ps(i6); in xnn_f32_dwconv_minmax_ukernel_up32x25__avx512f_acc2() local 453 …vacc0123456789ABCDEFp0 = _mm512_fmadd_ps(vi6x0123456789ABCDEF, vk6x0123456789ABCDEF, vacc012345678… in xnn_f32_dwconv_minmax_ukernel_up32x25__avx512f_acc2() 606 const __m512 vi6x0123456789ABCDEF = _mm512_maskz_loadu_ps(vmask, i6); in xnn_f32_dwconv_minmax_ukernel_up32x25__avx512f_acc2() local 608 …vacc0123456789ABCDEFp0 = _mm512_fmadd_ps(vi6x0123456789ABCDEF, vk6x0123456789ABCDEF, vacc012345678… in xnn_f32_dwconv_minmax_ukernel_up32x25__avx512f_acc2()
|
/external/XNNPACK/src/qs8-dwconv/gen/ |
D | up16x9-minmax-avx512skx-mul32.c | 133 … const __m512i vi6x0123456789ABCDEF = _mm512_cvtepi8_epi32(_mm_loadu_si128((const __m128i*) i6)); in xnn_qs8_dwconv_minmax_ukernel_up16x9__avx512skx_mul32() local 137 …BCDEF = _mm512_add_epi32(vacc0123456789ABCDEF, _mm512_mullo_epi32(vi6x0123456789ABCDEF, vk6x012345… in xnn_qs8_dwconv_minmax_ukernel_up16x9__avx512skx_mul32() 219 … const __m512i vi6x0123456789ABCDEF = _mm512_cvtepi8_epi32(_mm_loadu_si128((const __m128i*) i6)); in xnn_qs8_dwconv_minmax_ukernel_up16x9__avx512skx_mul32() local 222 …BCDEF = _mm512_add_epi32(vacc0123456789ABCDEF, _mm512_mullo_epi32(vi6x0123456789ABCDEF, vk6x012345… in xnn_qs8_dwconv_minmax_ukernel_up16x9__avx512skx_mul32()
|
D | up32x9-minmax-avx512skx-mul32.c | 153 … const __m512i vi6x0123456789ABCDEF = _mm512_cvtepi8_epi32(_mm_loadu_si128((const __m128i*) i6)); in xnn_qs8_dwconv_minmax_ukernel_up32x9__avx512skx_mul32() local 159 …BCDEF = _mm512_add_epi32(vacc0123456789ABCDEF, _mm512_mullo_epi32(vi6x0123456789ABCDEF, vk6x012345… in xnn_qs8_dwconv_minmax_ukernel_up32x9__avx512skx_mul32() 272 … const __m512i vi6x0123456789ABCDEF = _mm512_cvtepi8_epi32(_mm_loadu_si128((const __m128i*) i6)); in xnn_qs8_dwconv_minmax_ukernel_up32x9__avx512skx_mul32() local 276 …BCDEF = _mm512_add_epi32(vacc0123456789ABCDEF, _mm512_mullo_epi32(vi6x0123456789ABCDEF, vk6x012345… in xnn_qs8_dwconv_minmax_ukernel_up32x9__avx512skx_mul32()
|
D | up16x9-minmax-avx2-mul16.c | 141 … const __m256i vi6x0123456789ABCDEF = _mm256_cvtepi8_epi16(_mm_loadu_si128((const __m128i*) i6)); in xnn_qs8_dwconv_minmax_ukernel_up16x9__avx2_mul16() local 145 …const __m256i vprod6x0123456789ABCDEF = _mm256_mullo_epi16(vi6x0123456789ABCDEF, vk6x0123456789AB… in xnn_qs8_dwconv_minmax_ukernel_up16x9__avx2_mul16() 268 … const __m256i vi6x0123456789ABCDEF = _mm256_cvtepi8_epi16(_mm_loadu_si128((const __m128i*) i6)); in xnn_qs8_dwconv_minmax_ukernel_up16x9__avx2_mul16() local 271 …const __m256i vprod6x0123456789ABCDEF = _mm256_mullo_epi16(vi6x0123456789ABCDEF, vk6x0123456789ABC… in xnn_qs8_dwconv_minmax_ukernel_up16x9__avx2_mul16()
|
D | up32x9-minmax-avx2-mul16.c | 179 … const __m256i vi6x0123456789ABCDEF = _mm256_cvtepi8_epi16(_mm_loadu_si128((const __m128i*) i6)); in xnn_qs8_dwconv_minmax_ukernel_up32x9__avx2_mul16() local 185 …const __m256i vprod6x0123456789ABCDEF = _mm256_mullo_epi16(vi6x0123456789ABCDEF, vk6x0123456789AB… in xnn_qs8_dwconv_minmax_ukernel_up32x9__avx2_mul16() 355 … const __m256i vi6x0123456789ABCDEF = _mm256_cvtepi8_epi16(_mm_loadu_si128((const __m128i*) i6)); in xnn_qs8_dwconv_minmax_ukernel_up32x9__avx2_mul16() local 359 …const __m256i vprod6x0123456789ABCDEF = _mm256_mullo_epi16(vi6x0123456789ABCDEF, vk6x0123456789ABC… in xnn_qs8_dwconv_minmax_ukernel_up32x9__avx2_mul16()
|