/external/XNNPACK/src/f32-dwconv/gen/ |
D | up32x9-minmax-avx512f.c | 126 const __m512 vi4x0123456789ABCDEF = _mm512_loadu_ps(i4); in xnn_f32_dwconv_minmax_ukernel_up32x9__avx512f() local 132 …vacc0123456789ABCDEFp0 = _mm512_fmadd_ps(vi4x0123456789ABCDEF, vk4x0123456789ABCDEF, vacc012345678… in xnn_f32_dwconv_minmax_ukernel_up32x9__avx512f() 210 const __m512 vi4x0123456789ABCDEF = _mm512_loadu_ps(i4); in xnn_f32_dwconv_minmax_ukernel_up32x9__avx512f() local 214 …vacc0123456789ABCDEFp0 = _mm512_fmadd_ps(vi4x0123456789ABCDEF, vk4x0123456789ABCDEF, vacc012345678… in xnn_f32_dwconv_minmax_ukernel_up32x9__avx512f() 273 const __m512 vi4x0123456789ABCDEF = _mm512_maskz_loadu_ps(vmask, i4); in xnn_f32_dwconv_minmax_ukernel_up32x9__avx512f() local 275 …vacc0123456789ABCDEFp0 = _mm512_fmadd_ps(vi4x0123456789ABCDEF, vk4x0123456789ABCDEF, vacc012345678… in xnn_f32_dwconv_minmax_ukernel_up32x9__avx512f()
|
D | up32x9-minmax-avx512f-acc2.c | 126 const __m512 vi4x0123456789ABCDEF = _mm512_loadu_ps(i4); in xnn_f32_dwconv_minmax_ukernel_up32x9__avx512f_acc2() local 132 …vacc0123456789ABCDEFp0 = _mm512_fmadd_ps(vi4x0123456789ABCDEF, vk4x0123456789ABCDEF, vacc012345678… in xnn_f32_dwconv_minmax_ukernel_up32x9__avx512f_acc2() 213 const __m512 vi4x0123456789ABCDEF = _mm512_loadu_ps(i4); in xnn_f32_dwconv_minmax_ukernel_up32x9__avx512f_acc2() local 217 …vacc0123456789ABCDEFp0 = _mm512_fmadd_ps(vi4x0123456789ABCDEF, vk4x0123456789ABCDEF, vacc012345678… in xnn_f32_dwconv_minmax_ukernel_up32x9__avx512f_acc2() 278 const __m512 vi4x0123456789ABCDEF = _mm512_maskz_loadu_ps(vmask, i4); in xnn_f32_dwconv_minmax_ukernel_up32x9__avx512f_acc2() local 280 …vacc0123456789ABCDEFp0 = _mm512_fmadd_ps(vi4x0123456789ABCDEF, vk4x0123456789ABCDEF, vacc012345678… in xnn_f32_dwconv_minmax_ukernel_up32x9__avx512f_acc2()
|
D | up16x9-minmax-avx512f-acc2.c | 113 const __m512 vi4x0123456789ABCDEF = _mm512_loadu_ps(i4); in xnn_f32_dwconv_minmax_ukernel_up16x9__avx512f_acc2() local 117 …vacc0123456789ABCDEFp0 = _mm512_fmadd_ps(vi4x0123456789ABCDEF, vk4x0123456789ABCDEF, vacc012345678… in xnn_f32_dwconv_minmax_ukernel_up16x9__avx512f_acc2() 178 const __m512 vi4x0123456789ABCDEF = _mm512_maskz_loadu_ps(vmask, i4); in xnn_f32_dwconv_minmax_ukernel_up16x9__avx512f_acc2() local 180 …vacc0123456789ABCDEFp0 = _mm512_fmadd_ps(vi4x0123456789ABCDEF, vk4x0123456789ABCDEF, vacc012345678… in xnn_f32_dwconv_minmax_ukernel_up16x9__avx512f_acc2()
|
D | up16x9-minmax-avx512f.c | 113 const __m512 vi4x0123456789ABCDEF = _mm512_loadu_ps(i4); in xnn_f32_dwconv_minmax_ukernel_up16x9__avx512f() local 117 …vacc0123456789ABCDEFp0 = _mm512_fmadd_ps(vi4x0123456789ABCDEF, vk4x0123456789ABCDEF, vacc012345678… in xnn_f32_dwconv_minmax_ukernel_up16x9__avx512f() 176 const __m512 vi4x0123456789ABCDEF = _mm512_maskz_loadu_ps(vmask, i4); in xnn_f32_dwconv_minmax_ukernel_up16x9__avx512f() local 178 …vacc0123456789ABCDEFp0 = _mm512_fmadd_ps(vi4x0123456789ABCDEF, vk4x0123456789ABCDEF, vacc012345678… in xnn_f32_dwconv_minmax_ukernel_up16x9__avx512f()
|
D | up16x25-minmax-avx512f-acc2.c | 193 const __m512 vi4x0123456789ABCDEF = _mm512_loadu_ps(i4); in xnn_f32_dwconv_minmax_ukernel_up16x25__avx512f_acc2() local 197 …vacc0123456789ABCDEFp0 = _mm512_fmadd_ps(vi4x0123456789ABCDEF, vk4x0123456789ABCDEF, vacc012345678… in xnn_f32_dwconv_minmax_ukernel_up16x25__avx512f_acc2() 354 const __m512 vi4x0123456789ABCDEF = _mm512_maskz_loadu_ps(vmask, i4); in xnn_f32_dwconv_minmax_ukernel_up16x25__avx512f_acc2() local 356 …vacc0123456789ABCDEFp0 = _mm512_fmadd_ps(vi4x0123456789ABCDEF, vk4x0123456789ABCDEF, vacc012345678… in xnn_f32_dwconv_minmax_ukernel_up16x25__avx512f_acc2()
|
D | up16x25-minmax-avx512f.c | 193 const __m512 vi4x0123456789ABCDEF = _mm512_loadu_ps(i4); in xnn_f32_dwconv_minmax_ukernel_up16x25__avx512f() local 197 …vacc0123456789ABCDEFp0 = _mm512_fmadd_ps(vi4x0123456789ABCDEF, vk4x0123456789ABCDEF, vacc012345678… in xnn_f32_dwconv_minmax_ukernel_up16x25__avx512f() 352 const __m512 vi4x0123456789ABCDEF = _mm512_maskz_loadu_ps(vmask, i4); in xnn_f32_dwconv_minmax_ukernel_up16x25__avx512f() local 354 …vacc0123456789ABCDEFp0 = _mm512_fmadd_ps(vi4x0123456789ABCDEF, vk4x0123456789ABCDEF, vacc012345678… in xnn_f32_dwconv_minmax_ukernel_up16x25__avx512f()
|
D | up32x25-minmax-avx512f.c | 206 const __m512 vi4x0123456789ABCDEF = _mm512_loadu_ps(i4); in xnn_f32_dwconv_minmax_ukernel_up32x25__avx512f() local 212 …vacc0123456789ABCDEFp0 = _mm512_fmadd_ps(vi4x0123456789ABCDEF, vk4x0123456789ABCDEF, vacc012345678… in xnn_f32_dwconv_minmax_ukernel_up32x25__avx512f() 434 const __m512 vi4x0123456789ABCDEF = _mm512_loadu_ps(i4); in xnn_f32_dwconv_minmax_ukernel_up32x25__avx512f() local 438 …vacc0123456789ABCDEFp0 = _mm512_fmadd_ps(vi4x0123456789ABCDEF, vk4x0123456789ABCDEF, vacc012345678… in xnn_f32_dwconv_minmax_ukernel_up32x25__avx512f() 593 const __m512 vi4x0123456789ABCDEF = _mm512_maskz_loadu_ps(vmask, i4); in xnn_f32_dwconv_minmax_ukernel_up32x25__avx512f() local 595 …vacc0123456789ABCDEFp0 = _mm512_fmadd_ps(vi4x0123456789ABCDEF, vk4x0123456789ABCDEF, vacc012345678… in xnn_f32_dwconv_minmax_ukernel_up32x25__avx512f()
|
D | up32x25-minmax-avx512f-acc2.c | 206 const __m512 vi4x0123456789ABCDEF = _mm512_loadu_ps(i4); in xnn_f32_dwconv_minmax_ukernel_up32x25__avx512f_acc2() local 212 …vacc0123456789ABCDEFp0 = _mm512_fmadd_ps(vi4x0123456789ABCDEF, vk4x0123456789ABCDEF, vacc012345678… in xnn_f32_dwconv_minmax_ukernel_up32x25__avx512f_acc2() 437 const __m512 vi4x0123456789ABCDEF = _mm512_loadu_ps(i4); in xnn_f32_dwconv_minmax_ukernel_up32x25__avx512f_acc2() local 441 …vacc0123456789ABCDEFp0 = _mm512_fmadd_ps(vi4x0123456789ABCDEF, vk4x0123456789ABCDEF, vacc012345678… in xnn_f32_dwconv_minmax_ukernel_up32x25__avx512f_acc2() 598 const __m512 vi4x0123456789ABCDEF = _mm512_maskz_loadu_ps(vmask, i4); in xnn_f32_dwconv_minmax_ukernel_up32x25__avx512f_acc2() local 600 …vacc0123456789ABCDEFp0 = _mm512_fmadd_ps(vi4x0123456789ABCDEF, vk4x0123456789ABCDEF, vacc012345678… in xnn_f32_dwconv_minmax_ukernel_up32x25__avx512f_acc2()
|
/external/XNNPACK/src/qs8-dwconv/gen/ |
D | up16x9-minmax-avx512skx-mul32.c | 121 … const __m512i vi4x0123456789ABCDEF = _mm512_cvtepi8_epi32(_mm_loadu_si128((const __m128i*) i4)); in xnn_qs8_dwconv_minmax_ukernel_up16x9__avx512skx_mul32() local 125 …BCDEF = _mm512_add_epi32(vacc0123456789ABCDEF, _mm512_mullo_epi32(vi4x0123456789ABCDEF, vk4x012345… in xnn_qs8_dwconv_minmax_ukernel_up16x9__avx512skx_mul32() 209 … const __m512i vi4x0123456789ABCDEF = _mm512_cvtepi8_epi32(_mm_loadu_si128((const __m128i*) i4)); in xnn_qs8_dwconv_minmax_ukernel_up16x9__avx512skx_mul32() local 212 …BCDEF = _mm512_add_epi32(vacc0123456789ABCDEF, _mm512_mullo_epi32(vi4x0123456789ABCDEF, vk4x012345… in xnn_qs8_dwconv_minmax_ukernel_up16x9__avx512skx_mul32()
|
D | up32x9-minmax-avx512skx-mul32.c | 135 … const __m512i vi4x0123456789ABCDEF = _mm512_cvtepi8_epi32(_mm_loadu_si128((const __m128i*) i4)); in xnn_qs8_dwconv_minmax_ukernel_up32x9__avx512skx_mul32() local 141 …BCDEF = _mm512_add_epi32(vacc0123456789ABCDEF, _mm512_mullo_epi32(vi4x0123456789ABCDEF, vk4x012345… in xnn_qs8_dwconv_minmax_ukernel_up32x9__avx512skx_mul32() 260 … const __m512i vi4x0123456789ABCDEF = _mm512_cvtepi8_epi32(_mm_loadu_si128((const __m128i*) i4)); in xnn_qs8_dwconv_minmax_ukernel_up32x9__avx512skx_mul32() local 264 …BCDEF = _mm512_add_epi32(vacc0123456789ABCDEF, _mm512_mullo_epi32(vi4x0123456789ABCDEF, vk4x012345… in xnn_qs8_dwconv_minmax_ukernel_up32x9__avx512skx_mul32()
|
D | up16x9-minmax-avx2-mul16.c | 123 … const __m256i vi4x0123456789ABCDEF = _mm256_cvtepi8_epi16(_mm_loadu_si128((const __m128i*) i4)); in xnn_qs8_dwconv_minmax_ukernel_up16x9__avx2_mul16() local 127 …const __m256i vprod4x0123456789ABCDEF = _mm256_mullo_epi16(vi4x0123456789ABCDEF, vk4x0123456789AB… in xnn_qs8_dwconv_minmax_ukernel_up16x9__avx2_mul16() 252 … const __m256i vi4x0123456789ABCDEF = _mm256_cvtepi8_epi16(_mm_loadu_si128((const __m128i*) i4)); in xnn_qs8_dwconv_minmax_ukernel_up16x9__avx2_mul16() local 255 …const __m256i vprod4x0123456789ABCDEF = _mm256_mullo_epi16(vi4x0123456789ABCDEF, vk4x0123456789ABC… in xnn_qs8_dwconv_minmax_ukernel_up16x9__avx2_mul16()
|
D | up32x9-minmax-avx2-mul16.c | 149 … const __m256i vi4x0123456789ABCDEF = _mm256_cvtepi8_epi16(_mm_loadu_si128((const __m128i*) i4)); in xnn_qs8_dwconv_minmax_ukernel_up32x9__avx2_mul16() local 155 …const __m256i vprod4x0123456789ABCDEF = _mm256_mullo_epi16(vi4x0123456789ABCDEF, vk4x0123456789AB… in xnn_qs8_dwconv_minmax_ukernel_up32x9__avx2_mul16() 337 … const __m256i vi4x0123456789ABCDEF = _mm256_cvtepi8_epi16(_mm_loadu_si128((const __m128i*) i4)); in xnn_qs8_dwconv_minmax_ukernel_up32x9__avx2_mul16() local 341 …const __m256i vprod4x0123456789ABCDEF = _mm256_mullo_epi16(vi4x0123456789ABCDEF, vk4x0123456789ABC… in xnn_qs8_dwconv_minmax_ukernel_up32x9__avx2_mul16()
|