/external/XNNPACK/src/f32-dwconv/gen/ |
D | up16x25-minmax-avx512f-acc2.c | 229 const __m512 vi10x0123456789ABCDEF = _mm512_loadu_ps(i10); in xnn_f32_dwconv_minmax_ukernel_up16x25__avx512f_acc2() local 378 const __m512 vi10x0123456789ABCDEF = _mm512_maskz_loadu_ps(vmask, i10); in xnn_f32_dwconv_minmax_ukernel_up16x25__avx512f_acc2() local
|
D | up16x25-minmax-avx512f.c | 229 const __m512 vi10x0123456789ABCDEF = _mm512_loadu_ps(i10); in xnn_f32_dwconv_minmax_ukernel_up16x25__avx512f() local 376 const __m512 vi10x0123456789ABCDEF = _mm512_maskz_loadu_ps(vmask, i10); in xnn_f32_dwconv_minmax_ukernel_up16x25__avx512f() local
|
D | up32x25-minmax-avx512f-acc2.c | 260 const __m512 vi10x0123456789ABCDEF = _mm512_loadu_ps(i10); in xnn_f32_dwconv_minmax_ukernel_up32x25__avx512f_acc2() local 473 const __m512 vi10x0123456789ABCDEF = _mm512_loadu_ps(i10); in xnn_f32_dwconv_minmax_ukernel_up32x25__avx512f_acc2() local 622 const __m512 vi10x0123456789ABCDEF = _mm512_maskz_loadu_ps(vmask, i10); in xnn_f32_dwconv_minmax_ukernel_up32x25__avx512f_acc2() local
|
D | up32x25-minmax-avx512f.c | 260 const __m512 vi10x0123456789ABCDEF = _mm512_loadu_ps(i10); in xnn_f32_dwconv_minmax_ukernel_up32x25__avx512f() local 470 const __m512 vi10x0123456789ABCDEF = _mm512_loadu_ps(i10); in xnn_f32_dwconv_minmax_ukernel_up32x25__avx512f() local 617 const __m512 vi10x0123456789ABCDEF = _mm512_maskz_loadu_ps(vmask, i10); in xnn_f32_dwconv_minmax_ukernel_up32x25__avx512f() local
|
/external/XNNPACK/src/qs8-dwconv/gen/ |
D | up16x25-minmax-fp32-avx512skx-mul32.c | 232 … const __m512i vi10x0123456789ABCDEF = _mm512_cvtepi8_epi32(_mm_loadu_si128((const __m128i*) i10)); in xnn_qs8_dwconv_minmax_fp32_ukernel_up16x25__avx512skx_mul32() local 400 … const __m512i vi10x0123456789ABCDEF = _mm512_cvtepi8_epi32(_mm_loadu_si128((const __m128i*) i10)); in xnn_qs8_dwconv_minmax_fp32_ukernel_up16x25__avx512skx_mul32() local
|
D | up16x25-minmax-fp32-avx2-mul16-add16-vpunpck.c | 256 … const __m256i vi10x0123456789ABCDEF = _mm256_cvtepi8_epi16(_mm_loadu_si128((const __m128i*) i10)); in xnn_qs8_dwconv_minmax_fp32_ukernel_up16x25__avx2_mul16_add16_vpunpck() local 514 … const __m256i vi10x0123456789ABCDEF = _mm256_cvtepi8_epi16(_mm_loadu_si128((const __m128i*) i10)); in xnn_qs8_dwconv_minmax_fp32_ukernel_up16x25__avx2_mul16_add16_vpunpck() local
|
D | up32x25-minmax-fp32-avx512skx-mul32.c | 264 … const __m512i vi10x0123456789ABCDEF = _mm512_cvtepi8_epi32(_mm_loadu_si128((const __m128i*) i10)); in xnn_qs8_dwconv_minmax_fp32_ukernel_up32x25__avx512skx_mul32() local 499 … const __m512i vi10x0123456789ABCDEF = _mm512_cvtepi8_epi32(_mm_loadu_si128((const __m128i*) i10)); in xnn_qs8_dwconv_minmax_fp32_ukernel_up32x25__avx512skx_mul32() local
|
D | up16x25-minmax-fp32-avx2-mul16-vpunpck.c | 271 … const __m256i vi10x0123456789ABCDEF = _mm256_cvtepi8_epi16(_mm_loadu_si128((const __m128i*) i10)); in xnn_qs8_dwconv_minmax_fp32_ukernel_up16x25__avx2_mul16_vpunpck() local 550 … const __m256i vi10x0123456789ABCDEF = _mm256_cvtepi8_epi16(_mm_loadu_si128((const __m128i*) i10)); in xnn_qs8_dwconv_minmax_fp32_ukernel_up16x25__avx2_mul16_vpunpck() local
|
D | up16x25-minmax-fp32-avx2-mul16-vpmovsx.c | 258 … const __m256i vi10x0123456789ABCDEF = _mm256_cvtepi8_epi16(_mm_loadu_si128((const __m128i*) i10)); in xnn_qs8_dwconv_minmax_fp32_ukernel_up16x25__avx2_mul16_vpmovsx() local 506 … const __m256i vi10x0123456789ABCDEF = _mm256_cvtepi8_epi16(_mm_loadu_si128((const __m128i*) i10)); in xnn_qs8_dwconv_minmax_fp32_ukernel_up16x25__avx2_mul16_vpmovsx() local
|
D | up32x25-minmax-fp32-avx2-mul16-add16-vpunpck.c | 305 … const __m256i vi10x0123456789ABCDEF = _mm256_cvtepi8_epi16(_mm_loadu_si128((const __m128i*) i10)); in xnn_qs8_dwconv_minmax_fp32_ukernel_up32x25__avx2_mul16_add16_vpunpck() local 657 … const __m256i vi10x0123456789ABCDEF = _mm256_cvtepi8_epi16(_mm_loadu_si128((const __m128i*) i10)); in xnn_qs8_dwconv_minmax_fp32_ukernel_up32x25__avx2_mul16_add16_vpunpck() local
|
D | up16x25-minmax-rndnu-neon-mul8-ld128.c | 274 const int8x16_t vi10x0123456789ABCDEF = vld1q_s8(i10); i10 += 16; in xnn_qs8_dwconv_minmax_rndnu_ukernel_up16x25__neon_mul8_ld128() local
|
D | up16x25-minmax-rndnu-neon-mla8-ld128.c | 254 const int8x16_t vi10x0123456789ABCDEF = vld1q_s8(i10); i10 += 16; in xnn_qs8_dwconv_minmax_rndnu_ukernel_up16x25__neon_mla8_ld128() local
|
D | up32x25-minmax-fp32-avx2-mul16-vpunpck.c | 335 … const __m256i vi10x0123456789ABCDEF = _mm256_cvtepi8_epi16(_mm_loadu_si128((const __m128i*) i10)); in xnn_qs8_dwconv_minmax_fp32_ukernel_up32x25__avx2_mul16_vpunpck() local 729 … const __m256i vi10x0123456789ABCDEF = _mm256_cvtepi8_epi16(_mm_loadu_si128((const __m128i*) i10)); in xnn_qs8_dwconv_minmax_fp32_ukernel_up32x25__avx2_mul16_vpunpck() local
|
D | up32x25-minmax-fp32-avx2-mul16-vpmovsx.c | 320 … const __m256i vi10x0123456789ABCDEF = _mm256_cvtepi8_epi16(_mm_loadu_si128((const __m128i*) i10)); in xnn_qs8_dwconv_minmax_fp32_ukernel_up32x25__avx2_mul16_vpmovsx() local 681 … const __m256i vi10x0123456789ABCDEF = _mm256_cvtepi8_epi16(_mm_loadu_si128((const __m128i*) i10)); in xnn_qs8_dwconv_minmax_fp32_ukernel_up32x25__avx2_mul16_vpmovsx() local
|
/external/XNNPACK/src/qu8-dwconv/gen/ |
D | up16x25-minmax-fp32-avx512skx-mul32.c | 233 … const __m512i vi10x0123456789ABCDEF = _mm512_cvtepu8_epi32(_mm_loadu_si128((const __m128i*) i10)); in xnn_qu8_dwconv_minmax_fp32_ukernel_up16x25__avx512skx_mul32() local 401 … const __m512i vi10x0123456789ABCDEF = _mm512_cvtepu8_epi32(_mm_loadu_si128((const __m128i*) i10)); in xnn_qu8_dwconv_minmax_fp32_ukernel_up16x25__avx512skx_mul32() local
|
D | up32x25-minmax-fp32-avx512skx-mul32.c | 265 … const __m512i vi10x0123456789ABCDEF = _mm512_cvtepu8_epi32(_mm_loadu_si128((const __m128i*) i10)); in xnn_qu8_dwconv_minmax_fp32_ukernel_up32x25__avx512skx_mul32() local 500 … const __m512i vi10x0123456789ABCDEF = _mm512_cvtepu8_epi32(_mm_loadu_si128((const __m128i*) i10)); in xnn_qu8_dwconv_minmax_fp32_ukernel_up32x25__avx512skx_mul32() local
|
/external/XNNPACK/src/qc8-dwconv/gen/ |
D | up16x25-minmax-fp32-avx512skx-mul32.c | 231 … const __m512i vi10x0123456789ABCDEF = _mm512_cvtepi8_epi32(_mm_loadu_si128((const __m128i*) i10)); in xnn_qc8_dwconv_minmax_fp32_ukernel_up16x25__avx512skx_mul32() local 401 … const __m512i vi10x0123456789ABCDEF = _mm512_cvtepi8_epi32(_mm_loadu_si128((const __m128i*) i10)); in xnn_qc8_dwconv_minmax_fp32_ukernel_up16x25__avx512skx_mul32() local
|
D | up32x25-minmax-fp32-avx512skx-mul32.c | 263 … const __m512i vi10x0123456789ABCDEF = _mm512_cvtepi8_epi32(_mm_loadu_si128((const __m128i*) i10)); in xnn_qc8_dwconv_minmax_fp32_ukernel_up32x25__avx512skx_mul32() local 501 … const __m512i vi10x0123456789ABCDEF = _mm512_cvtepi8_epi32(_mm_loadu_si128((const __m128i*) i10)); in xnn_qc8_dwconv_minmax_fp32_ukernel_up32x25__avx512skx_mul32() local
|
D | up16x25-minmax-fp32-avx2-mul16-add16-vpunpck.c | 256 … const __m256i vi10x0123456789ABCDEF = _mm256_cvtepi8_epi16(_mm_loadu_si128((const __m128i*) i10)); in xnn_qc8_dwconv_minmax_fp32_ukernel_up16x25__avx2_mul16_add16_vpunpck() local 516 … const __m256i vi10x0123456789ABCDEF = _mm256_cvtepi8_epi16(_mm_loadu_si128((const __m128i*) i10)); in xnn_qc8_dwconv_minmax_fp32_ukernel_up16x25__avx2_mul16_add16_vpunpck() local
|
D | up16x25-minmax-fp32-avx2-mul16-vpmovsx.c | 258 … const __m256i vi10x0123456789ABCDEF = _mm256_cvtepi8_epi16(_mm_loadu_si128((const __m128i*) i10)); in xnn_qc8_dwconv_minmax_fp32_ukernel_up16x25__avx2_mul16_vpmovsx() local 508 … const __m256i vi10x0123456789ABCDEF = _mm256_cvtepi8_epi16(_mm_loadu_si128((const __m128i*) i10)); in xnn_qc8_dwconv_minmax_fp32_ukernel_up16x25__avx2_mul16_vpmovsx() local
|
D | up16x25-minmax-fp32-avx2-mul16-vpunpck.c | 271 … const __m256i vi10x0123456789ABCDEF = _mm256_cvtepi8_epi16(_mm_loadu_si128((const __m128i*) i10)); in xnn_qc8_dwconv_minmax_fp32_ukernel_up16x25__avx2_mul16_vpunpck() local 552 … const __m256i vi10x0123456789ABCDEF = _mm256_cvtepi8_epi16(_mm_loadu_si128((const __m128i*) i10)); in xnn_qc8_dwconv_minmax_fp32_ukernel_up16x25__avx2_mul16_vpunpck() local
|
D | up32x25-minmax-fp32-avx2-mul16-add16-vpunpck.c | 305 … const __m256i vi10x0123456789ABCDEF = _mm256_cvtepi8_epi16(_mm_loadu_si128((const __m128i*) i10)); in xnn_qc8_dwconv_minmax_fp32_ukernel_up32x25__avx2_mul16_add16_vpunpck() local 661 … const __m256i vi10x0123456789ABCDEF = _mm256_cvtepi8_epi16(_mm_loadu_si128((const __m128i*) i10)); in xnn_qc8_dwconv_minmax_fp32_ukernel_up32x25__avx2_mul16_add16_vpunpck() local
|
D | up32x25-minmax-fp32-avx2-mul16-vpmovsx.c | 320 … const __m256i vi10x0123456789ABCDEF = _mm256_cvtepi8_epi16(_mm_loadu_si128((const __m128i*) i10)); in xnn_qc8_dwconv_minmax_fp32_ukernel_up32x25__avx2_mul16_vpmovsx() local 685 … const __m256i vi10x0123456789ABCDEF = _mm256_cvtepi8_epi16(_mm_loadu_si128((const __m128i*) i10)); in xnn_qc8_dwconv_minmax_fp32_ukernel_up32x25__avx2_mul16_vpmovsx() local
|
D | up16x25-minmax-fp32-neon-mul8-ld128.c | 272 const int8x16_t vi10x0123456789ABCDEF = vld1q_s8(i10); i10 += 16; in xnn_qc8_dwconv_minmax_fp32_ukernel_up16x25__neon_mul8_ld128() local
|
D | up16x25-minmax-fp32-neonv8-mul8-ld128.c | 272 const int8x16_t vi10x0123456789ABCDEF = vld1q_s8(i10); i10 += 16; in xnn_qc8_dwconv_minmax_fp32_ukernel_up16x25__neonv8_mul8_ld128() local
|