/aosp_15_r20/external/XNNPACK/src/f32-dwconv/gen/ |
H A D | up32x9-minmax-avx512f.c | 162 const __m512 vi8x0123456789ABCDEF = _mm512_loadu_ps(i8); in xnn_f32_dwconv_minmax_ukernel_up32x9__avx512f() local 234 const __m512 vi8x0123456789ABCDEF = _mm512_loadu_ps(i8); in xnn_f32_dwconv_minmax_ukernel_up32x9__avx512f() local 289 const __m512 vi8x0123456789ABCDEF = _mm512_maskz_loadu_ps(vmask, i8); in xnn_f32_dwconv_minmax_ukernel_up32x9__avx512f() local
|
H A D | up32x9-minmax-avx512f-acc2.c | 162 const __m512 vi8x0123456789ABCDEF = _mm512_loadu_ps(i8); in xnn_f32_dwconv_minmax_ukernel_up32x9__avx512f_acc2() local 237 const __m512 vi8x0123456789ABCDEF = _mm512_loadu_ps(i8); in xnn_f32_dwconv_minmax_ukernel_up32x9__avx512f_acc2() local 294 const __m512 vi8x0123456789ABCDEF = _mm512_maskz_loadu_ps(vmask, i8); in xnn_f32_dwconv_minmax_ukernel_up32x9__avx512f_acc2() local
|
H A D | up16x9-minmax-avx512f.c | 137 const __m512 vi8x0123456789ABCDEF = _mm512_loadu_ps(i8); in xnn_f32_dwconv_minmax_ukernel_up16x9__avx512f() local 192 const __m512 vi8x0123456789ABCDEF = _mm512_maskz_loadu_ps(vmask, i8); in xnn_f32_dwconv_minmax_ukernel_up16x9__avx512f() local
|
H A D | up16x9-minmax-avx512f-acc2.c | 137 const __m512 vi8x0123456789ABCDEF = _mm512_loadu_ps(i8); in xnn_f32_dwconv_minmax_ukernel_up16x9__avx512f_acc2() local 194 const __m512 vi8x0123456789ABCDEF = _mm512_maskz_loadu_ps(vmask, i8); in xnn_f32_dwconv_minmax_ukernel_up16x9__avx512f_acc2() local
|
H A D | up32x25-minmax-avx512f-acc2.c | 242 const __m512 vi8x0123456789ABCDEF = _mm512_loadu_ps(i8); in xnn_f32_dwconv_minmax_ukernel_up32x25__avx512f_acc2() local 461 const __m512 vi8x0123456789ABCDEF = _mm512_loadu_ps(i8); in xnn_f32_dwconv_minmax_ukernel_up32x25__avx512f_acc2() local 614 const __m512 vi8x0123456789ABCDEF = _mm512_maskz_loadu_ps(vmask, i8); in xnn_f32_dwconv_minmax_ukernel_up32x25__avx512f_acc2() local
|
H A D | up32x25-minmax-avx512f.c | 242 const __m512 vi8x0123456789ABCDEF = _mm512_loadu_ps(i8); in xnn_f32_dwconv_minmax_ukernel_up32x25__avx512f() local 458 const __m512 vi8x0123456789ABCDEF = _mm512_loadu_ps(i8); in xnn_f32_dwconv_minmax_ukernel_up32x25__avx512f() local 609 const __m512 vi8x0123456789ABCDEF = _mm512_maskz_loadu_ps(vmask, i8); in xnn_f32_dwconv_minmax_ukernel_up32x25__avx512f() local
|
H A D | up16x25-minmax-avx512f.c | 217 const __m512 vi8x0123456789ABCDEF = _mm512_loadu_ps(i8); in xnn_f32_dwconv_minmax_ukernel_up16x25__avx512f() local 368 const __m512 vi8x0123456789ABCDEF = _mm512_maskz_loadu_ps(vmask, i8); in xnn_f32_dwconv_minmax_ukernel_up16x25__avx512f() local
|
H A D | up16x25-minmax-avx512f-acc2.c | 217 const __m512 vi8x0123456789ABCDEF = _mm512_loadu_ps(i8); in xnn_f32_dwconv_minmax_ukernel_up16x25__avx512f_acc2() local 370 const __m512 vi8x0123456789ABCDEF = _mm512_maskz_loadu_ps(vmask, i8); in xnn_f32_dwconv_minmax_ukernel_up16x25__avx512f_acc2() local
|
/aosp_15_r20/external/XNNPACK/src/qs8-dwconv/gen/ |
H A D | up16x9-minmax-fp32-avx512skx-mul32.c | 140 … const __m512i vi8x0123456789ABCDEF = _mm512_cvtepi8_epi32(_mm_loadu_si128((const __m128i*) i8)); in xnn_qs8_dwconv_minmax_fp32_ukernel_up16x9__avx512skx_mul32() local 214 … const __m512i vi8x0123456789ABCDEF = _mm512_cvtepi8_epi32(_mm_loadu_si128((const __m128i*) i8)); in xnn_qs8_dwconv_minmax_fp32_ukernel_up16x9__avx512skx_mul32() local
|
H A D | up16x9-minmax-fp32-avx2-mul16-add16-vpunpck.c | 159 … const __m256i vi8x0123456789ABCDEF = _mm256_cvtepi8_epi16(_mm_loadu_si128((const __m128i*) i8)); in xnn_qs8_dwconv_minmax_fp32_ukernel_up16x9__avx2_mul16_add16_vpunpck() local 280 … const __m256i vi8x0123456789ABCDEF = _mm256_cvtepi8_epi16(_mm_loadu_si128((const __m128i*) i8)); in xnn_qs8_dwconv_minmax_fp32_ukernel_up16x9__avx2_mul16_add16_vpunpck() local
|
H A D | up32x9-minmax-fp32-avx512skx-mul32.c | 166 … const __m512i vi8x0123456789ABCDEF = _mm512_cvtepi8_epi32(_mm_loadu_si128((const __m128i*) i8)); in xnn_qs8_dwconv_minmax_fp32_ukernel_up32x9__avx512skx_mul32() local 263 … const __m512i vi8x0123456789ABCDEF = _mm512_cvtepi8_epi32(_mm_loadu_si128((const __m128i*) i8)); in xnn_qs8_dwconv_minmax_fp32_ukernel_up32x9__avx512skx_mul32() local
|
H A D | up16x9-minmax-fp32-avx2-mul16-vpunpck.c | 171 … const __m256i vi8x0123456789ABCDEF = _mm256_cvtepi8_epi16(_mm_loadu_si128((const __m128i*) i8)); in xnn_qs8_dwconv_minmax_fp32_ukernel_up16x9__avx2_mul16_vpunpck() local 292 … const __m256i vi8x0123456789ABCDEF = _mm256_cvtepi8_epi16(_mm_loadu_si128((const __m128i*) i8)); in xnn_qs8_dwconv_minmax_fp32_ukernel_up16x9__avx2_mul16_vpunpck() local
|
H A D | up16x9-minmax-fp32-avx2-mul16-vpmovsx.c | 160 … const __m256i vi8x0123456789ABCDEF = _mm256_cvtepi8_epi16(_mm_loadu_si128((const __m128i*) i8)); in xnn_qs8_dwconv_minmax_fp32_ukernel_up16x9__avx2_mul16_vpmovsx() local 266 … const __m256i vi8x0123456789ABCDEF = _mm256_cvtepi8_epi16(_mm_loadu_si128((const __m128i*) i8)); in xnn_qs8_dwconv_minmax_fp32_ukernel_up16x9__avx2_mul16_vpmovsx() local
|
H A D | up32x9-minmax-fp32-avx2-mul16-add16-vpunpck.c | 199 … const __m256i vi8x0123456789ABCDEF = _mm256_cvtepi8_epi16(_mm_loadu_si128((const __m128i*) i8)); in xnn_qs8_dwconv_minmax_fp32_ukernel_up32x9__avx2_mul16_add16_vpunpck() local 349 … const __m256i vi8x0123456789ABCDEF = _mm256_cvtepi8_epi16(_mm_loadu_si128((const __m128i*) i8)); in xnn_qs8_dwconv_minmax_fp32_ukernel_up32x9__avx2_mul16_add16_vpunpck() local
|
H A D | up16x25-minmax-fp32-avx512skx-mul32.c | 220 … const __m512i vi8x0123456789ABCDEF = _mm512_cvtepi8_epi32(_mm_loadu_si128((const __m128i*) i8)); in xnn_qs8_dwconv_minmax_fp32_ukernel_up16x25__avx512skx_mul32() local 390 … const __m512i vi8x0123456789ABCDEF = _mm512_cvtepi8_epi32(_mm_loadu_si128((const __m128i*) i8)); in xnn_qs8_dwconv_minmax_fp32_ukernel_up16x25__avx512skx_mul32() local
|
H A D | up32x9-minmax-fp32-avx2-mul16-vpmovsx.c | 210 … const __m256i vi8x0123456789ABCDEF = _mm256_cvtepi8_epi16(_mm_loadu_si128((const __m128i*) i8)); in xnn_qs8_dwconv_minmax_fp32_ukernel_up32x9__avx2_mul16_vpmovsx() local 343 … const __m256i vi8x0123456789ABCDEF = _mm256_cvtepi8_epi16(_mm_loadu_si128((const __m128i*) i8)); in xnn_qs8_dwconv_minmax_fp32_ukernel_up32x9__avx2_mul16_vpmovsx() local
|
/aosp_15_r20/external/XNNPACK/src/qu8-dwconv/gen/ |
H A D | up16x9-minmax-fp32-avx512skx-mul32.c | 141 … const __m512i vi8x0123456789ABCDEF = _mm512_cvtepu8_epi32(_mm_loadu_si128((const __m128i*) i8)); in xnn_qu8_dwconv_minmax_fp32_ukernel_up16x9__avx512skx_mul32() local 215 … const __m512i vi8x0123456789ABCDEF = _mm512_cvtepu8_epi32(_mm_loadu_si128((const __m128i*) i8)); in xnn_qu8_dwconv_minmax_fp32_ukernel_up16x9__avx512skx_mul32() local
|
H A D | up32x9-minmax-fp32-avx512skx-mul32.c | 167 … const __m512i vi8x0123456789ABCDEF = _mm512_cvtepu8_epi32(_mm_loadu_si128((const __m128i*) i8)); in xnn_qu8_dwconv_minmax_fp32_ukernel_up32x9__avx512skx_mul32() local 264 … const __m512i vi8x0123456789ABCDEF = _mm512_cvtepu8_epi32(_mm_loadu_si128((const __m128i*) i8)); in xnn_qu8_dwconv_minmax_fp32_ukernel_up32x9__avx512skx_mul32() local
|
H A D | up16x25-minmax-fp32-avx512skx-mul32.c | 221 … const __m512i vi8x0123456789ABCDEF = _mm512_cvtepu8_epi32(_mm_loadu_si128((const __m128i*) i8)); in xnn_qu8_dwconv_minmax_fp32_ukernel_up16x25__avx512skx_mul32() local 391 … const __m512i vi8x0123456789ABCDEF = _mm512_cvtepu8_epi32(_mm_loadu_si128((const __m128i*) i8)); in xnn_qu8_dwconv_minmax_fp32_ukernel_up16x25__avx512skx_mul32() local
|
/aosp_15_r20/external/XNNPACK/src/qc8-dwconv/gen/ |
H A D | up16x9-minmax-fp32-avx512skx-mul32.c | 139 … const __m512i vi8x0123456789ABCDEF = _mm512_cvtepi8_epi32(_mm_loadu_si128((const __m128i*) i8)); in xnn_qc8_dwconv_minmax_fp32_ukernel_up16x9__avx512skx_mul32() local 215 … const __m512i vi8x0123456789ABCDEF = _mm512_cvtepi8_epi32(_mm_loadu_si128((const __m128i*) i8)); in xnn_qc8_dwconv_minmax_fp32_ukernel_up16x9__avx512skx_mul32() local
|
H A D | up32x9-minmax-fp32-avx512skx-mul32.c | 165 … const __m512i vi8x0123456789ABCDEF = _mm512_cvtepi8_epi32(_mm_loadu_si128((const __m128i*) i8)); in xnn_qc8_dwconv_minmax_fp32_ukernel_up32x9__avx512skx_mul32() local 265 … const __m512i vi8x0123456789ABCDEF = _mm512_cvtepi8_epi32(_mm_loadu_si128((const __m128i*) i8)); in xnn_qc8_dwconv_minmax_fp32_ukernel_up32x9__avx512skx_mul32() local
|
H A D | up16x9-minmax-fp32-avx2-mul16-add16-vpunpck.c | 159 … const __m256i vi8x0123456789ABCDEF = _mm256_cvtepi8_epi16(_mm_loadu_si128((const __m128i*) i8)); in xnn_qc8_dwconv_minmax_fp32_ukernel_up16x9__avx2_mul16_add16_vpunpck() local 282 … const __m256i vi8x0123456789ABCDEF = _mm256_cvtepi8_epi16(_mm_loadu_si128((const __m128i*) i8)); in xnn_qc8_dwconv_minmax_fp32_ukernel_up16x9__avx2_mul16_add16_vpunpck() local
|
H A D | up16x9-minmax-fp32-avx2-mul16-vpmovsx.c | 160 … const __m256i vi8x0123456789ABCDEF = _mm256_cvtepi8_epi16(_mm_loadu_si128((const __m128i*) i8)); in xnn_qc8_dwconv_minmax_fp32_ukernel_up16x9__avx2_mul16_vpmovsx() local 268 … const __m256i vi8x0123456789ABCDEF = _mm256_cvtepi8_epi16(_mm_loadu_si128((const __m128i*) i8)); in xnn_qc8_dwconv_minmax_fp32_ukernel_up16x9__avx2_mul16_vpmovsx() local
|
H A D | up16x9-minmax-fp32-avx2-mul16-vpunpck.c | 171 … const __m256i vi8x0123456789ABCDEF = _mm256_cvtepi8_epi16(_mm_loadu_si128((const __m128i*) i8)); in xnn_qc8_dwconv_minmax_fp32_ukernel_up16x9__avx2_mul16_vpunpck() local 294 … const __m256i vi8x0123456789ABCDEF = _mm256_cvtepi8_epi16(_mm_loadu_si128((const __m128i*) i8)); in xnn_qc8_dwconv_minmax_fp32_ukernel_up16x9__avx2_mul16_vpunpck() local
|
H A D | up32x9-minmax-fp32-avx2-mul16-add16-vpunpck.c | 199 … const __m256i vi8x0123456789ABCDEF = _mm256_cvtepi8_epi16(_mm_loadu_si128((const __m128i*) i8)); in xnn_qc8_dwconv_minmax_fp32_ukernel_up32x9__avx2_mul16_add16_vpunpck() local 353 … const __m256i vi8x0123456789ABCDEF = _mm256_cvtepi8_epi16(_mm_loadu_si128((const __m128i*) i8)); in xnn_qc8_dwconv_minmax_fp32_ukernel_up32x9__avx2_mul16_add16_vpunpck() local
|