/aosp_15_r20/external/XNNPACK/src/f32-dwconv/gen/ |
H A D | up32x25-minmax-avx512f-acc2.c | 273 const __m512 vk11x0123456789ABCDEF = _mm512_load_ps(w + 384); in xnn_f32_dwconv_minmax_ukernel_up32x25__avx512f_acc2() local 482 const __m512 vk11x0123456789ABCDEF = _mm512_load_ps(w + 384); in xnn_f32_dwconv_minmax_ukernel_up32x25__avx512f_acc2() local 627 const __m512 vk11x0123456789ABCDEF = _mm512_maskz_loadu_ps(vmask, w + 384); in xnn_f32_dwconv_minmax_ukernel_up32x25__avx512f_acc2() local
|
H A D | up32x25-minmax-avx512f.c | 273 const __m512 vk11x0123456789ABCDEF = _mm512_load_ps(w + 384); in xnn_f32_dwconv_minmax_ukernel_up32x25__avx512f() local 479 const __m512 vk11x0123456789ABCDEF = _mm512_load_ps(w + 384); in xnn_f32_dwconv_minmax_ukernel_up32x25__avx512f() local 622 const __m512 vk11x0123456789ABCDEF = _mm512_maskz_loadu_ps(vmask, w + 384); in xnn_f32_dwconv_minmax_ukernel_up32x25__avx512f() local
|
H A D | up16x25-minmax-avx512f.c | 238 const __m512 vk11x0123456789ABCDEF = _mm512_load_ps(w + 192); in xnn_f32_dwconv_minmax_ukernel_up16x25__avx512f() local 381 const __m512 vk11x0123456789ABCDEF = _mm512_maskz_loadu_ps(vmask, w + 192); in xnn_f32_dwconv_minmax_ukernel_up16x25__avx512f() local
|
H A D | up16x25-minmax-avx512f-acc2.c | 238 const __m512 vk11x0123456789ABCDEF = _mm512_load_ps(w + 192); in xnn_f32_dwconv_minmax_ukernel_up16x25__avx512f_acc2() local 383 const __m512 vk11x0123456789ABCDEF = _mm512_maskz_loadu_ps(vmask, w + 192); in xnn_f32_dwconv_minmax_ukernel_up16x25__avx512f_acc2() local
|
/aosp_15_r20/external/XNNPACK/src/qs8-dwconv/gen/ |
H A D | up16x25-minmax-fp32-avx512skx-mul32.c | 239 …const __m512i vk11x0123456789ABCDEF = _mm512_cvtepi8_epi32(_mm_load_si128((const __m128i*) ((uintp… in xnn_qs8_dwconv_minmax_fp32_ukernel_up16x25__avx512skx_mul32() local 406 …const __m512i vk11x0123456789ABCDEF = _mm512_cvtepi8_epi32(_mm_loadu_si128((const __m128i*) ((uint… in xnn_qs8_dwconv_minmax_fp32_ukernel_up16x25__avx512skx_mul32() local
|
H A D | up32x25-minmax-fp32-avx512skx-mul32.c | 274 …const __m512i vk11x0123456789ABCDEF = _mm512_cvtepi8_epi32(_mm_load_si128((const __m128i*) ((uintp… in xnn_qs8_dwconv_minmax_fp32_ukernel_up32x25__avx512skx_mul32() local 506 …const __m512i vk11x0123456789ABCDEF = _mm512_cvtepi8_epi32(_mm_loadu_si128((const __m128i*) (k + 3… in xnn_qs8_dwconv_minmax_fp32_ukernel_up32x25__avx512skx_mul32() local
|
H A D | up16x25-minmax-fp32-avx2-mul16-add16-vpunpck.c | 264 …const __m256i vk11x0123456789ABCDEF = _mm256_cvtepi8_epi16(_mm_loadu_si128((const __m128i*) ((uint… in xnn_qs8_dwconv_minmax_fp32_ukernel_up16x25__avx2_mul16_add16_vpunpck() local 524 …const __m256i vk11x0123456789ABCDEF = _mm256_cvtepi8_epi16(_mm_loadu_si128((const __m128i*) ((uint… in xnn_qs8_dwconv_minmax_fp32_ukernel_up16x25__avx2_mul16_add16_vpunpck() local
|
H A D | up16x25-minmax-fp32-avx2-mul16-vpmovsx.c | 268 …const __m256i vk11x0123456789ABCDEF = _mm256_cvtepi8_epi16(_mm_loadu_si128((const __m128i*) ((uint… in xnn_qs8_dwconv_minmax_fp32_ukernel_up16x25__avx2_mul16_vpmovsx() local 515 …const __m256i vk11x0123456789ABCDEF = _mm256_cvtepi8_epi16(_mm_loadu_si128((const __m128i*) ((uint… in xnn_qs8_dwconv_minmax_fp32_ukernel_up16x25__avx2_mul16_vpmovsx() local
|
H A D | up16x25-minmax-fp32-avx2-mul16-vpunpck.c | 282 …const __m256i vk11x0123456789ABCDEF = _mm256_cvtepi8_epi16(_mm_loadu_si128((const __m128i*) ((uint… in xnn_qs8_dwconv_minmax_fp32_ukernel_up16x25__avx2_mul16_vpunpck() local 560 …const __m256i vk11x0123456789ABCDEF = _mm256_cvtepi8_epi16(_mm_loadu_si128((const __m128i*) ((uint… in xnn_qs8_dwconv_minmax_fp32_ukernel_up16x25__avx2_mul16_vpunpck() local
|
H A D | up32x25-minmax-fp32-avx2-mul16-add16-vpunpck.c | 316 …const __m256i vk11x0123456789ABCDEF = _mm256_cvtepi8_epi16(_mm_loadu_si128((const __m128i*) ((uint… in xnn_qs8_dwconv_minmax_fp32_ukernel_up32x25__avx2_mul16_add16_vpunpck() local 668 …const __m256i vk11x0123456789ABCDEF = _mm256_cvtepi8_epi16(_mm_loadu_si128((const __m128i*) (k + 3… in xnn_qs8_dwconv_minmax_fp32_ukernel_up32x25__avx2_mul16_add16_vpunpck() local
|
H A D | up32x25-minmax-fp32-avx2-mul16-vpmovsx.c | 336 …const __m256i vk11x0123456789ABCDEF = _mm256_cvtepi8_epi16(_mm_loadu_si128((const __m128i*) ((uint… in xnn_qs8_dwconv_minmax_fp32_ukernel_up32x25__avx2_mul16_vpmovsx() local 691 …const __m256i vk11x0123456789ABCDEF = _mm256_cvtepi8_epi16(_mm_loadu_si128((const __m128i*) (k + 3… in xnn_qs8_dwconv_minmax_fp32_ukernel_up32x25__avx2_mul16_vpmovsx() local
|
H A D | up32x25-minmax-fp32-avx2-mul16-vpunpck.c | 352 …const __m256i vk11x0123456789ABCDEF = _mm256_cvtepi8_epi16(_mm_loadu_si128((const __m128i*) ((uint… in xnn_qs8_dwconv_minmax_fp32_ukernel_up32x25__avx2_mul16_vpunpck() local 740 …const __m256i vk11x0123456789ABCDEF = _mm256_cvtepi8_epi16(_mm_loadu_si128((const __m128i*) (k + 3… in xnn_qs8_dwconv_minmax_fp32_ukernel_up32x25__avx2_mul16_vpunpck() local
|
H A D | up16x25-minmax-rndnu-neon-mla8-ld128.c | 261 … const int8x16_t vk11x0123456789ABCDEF = vld1q_s8(w); w = (const void*) ((const int8_t*) w + 16); in xnn_qs8_dwconv_minmax_rndnu_ukernel_up16x25__neon_mla8_ld128() local
|
H A D | up16x25-minmax-rndnu-neon-mul8-ld128.c | 285 … const int8x16_t vk11x0123456789ABCDEF = vld1q_s8(w); w = (const void*) ((const int8_t*) w + 16); in xnn_qs8_dwconv_minmax_rndnu_ukernel_up16x25__neon_mul8_ld128() local
|
/aosp_15_r20/external/XNNPACK/src/qu8-dwconv/gen/ |
H A D | up16x25-minmax-fp32-avx512skx-mul32.c | 240 …const __m512i vk11x0123456789ABCDEF = _mm512_sub_epi32(_mm512_cvtepu8_epi32(_mm_load_si128((const … in xnn_qu8_dwconv_minmax_fp32_ukernel_up16x25__avx512skx_mul32() local 407 …const __m512i vk11x0123456789ABCDEF = _mm512_sub_epi32(_mm512_cvtepu8_epi32(_mm_loadu_si128((const… in xnn_qu8_dwconv_minmax_fp32_ukernel_up16x25__avx512skx_mul32() local
|
H A D | up32x25-minmax-fp32-avx512skx-mul32.c | 275 …const __m512i vk11x0123456789ABCDEF = _mm512_sub_epi32(_mm512_cvtepu8_epi32(_mm_load_si128((const … in xnn_qu8_dwconv_minmax_fp32_ukernel_up32x25__avx512skx_mul32() local 507 …const __m512i vk11x0123456789ABCDEF = _mm512_sub_epi32(_mm512_cvtepu8_epi32(_mm_loadu_si128((const… in xnn_qu8_dwconv_minmax_fp32_ukernel_up32x25__avx512skx_mul32() local
|
/aosp_15_r20/external/XNNPACK/src/qc8-dwconv/gen/ |
H A D | up16x25-minmax-fp32-avx512skx-mul32.c | 238 …const __m512i vk11x0123456789ABCDEF = _mm512_cvtepi8_epi32(_mm_load_si128((const __m128i*) ((uintp… in xnn_qc8_dwconv_minmax_fp32_ukernel_up16x25__avx512skx_mul32() local 407 …const __m512i vk11x0123456789ABCDEF = _mm512_cvtepi8_epi32(_mm_loadu_si128((const __m128i*) ((uint… in xnn_qc8_dwconv_minmax_fp32_ukernel_up16x25__avx512skx_mul32() local
|
H A D | up32x25-minmax-fp32-avx512skx-mul32.c | 273 …const __m512i vk11x0123456789ABCDEF = _mm512_cvtepi8_epi32(_mm_load_si128((const __m128i*) ((uintp… in xnn_qc8_dwconv_minmax_fp32_ukernel_up32x25__avx512skx_mul32() local 508 …const __m512i vk11x0123456789ABCDEF = _mm512_cvtepi8_epi32(_mm_loadu_si128((const __m128i*) (k + 3… in xnn_qc8_dwconv_minmax_fp32_ukernel_up32x25__avx512skx_mul32() local
|
H A D | up16x25-minmax-fp32-avx2-mul16-add16-vpunpck.c | 264 …const __m256i vk11x0123456789ABCDEF = _mm256_cvtepi8_epi16(_mm_loadu_si128((const __m128i*) ((uint… in xnn_qc8_dwconv_minmax_fp32_ukernel_up16x25__avx2_mul16_add16_vpunpck() local 526 …const __m256i vk11x0123456789ABCDEF = _mm256_cvtepi8_epi16(_mm_loadu_si128((const __m128i*) ((uint… in xnn_qc8_dwconv_minmax_fp32_ukernel_up16x25__avx2_mul16_add16_vpunpck() local
|
H A D | up16x25-minmax-fp32-avx2-mul16-vpmovsx.c | 268 …const __m256i vk11x0123456789ABCDEF = _mm256_cvtepi8_epi16(_mm_loadu_si128((const __m128i*) ((uint… in xnn_qc8_dwconv_minmax_fp32_ukernel_up16x25__avx2_mul16_vpmovsx() local 517 …const __m256i vk11x0123456789ABCDEF = _mm256_cvtepi8_epi16(_mm_loadu_si128((const __m128i*) ((uint… in xnn_qc8_dwconv_minmax_fp32_ukernel_up16x25__avx2_mul16_vpmovsx() local
|
H A D | up16x25-minmax-fp32-avx2-mul16-vpunpck.c | 282 …const __m256i vk11x0123456789ABCDEF = _mm256_cvtepi8_epi16(_mm_loadu_si128((const __m128i*) ((uint… in xnn_qc8_dwconv_minmax_fp32_ukernel_up16x25__avx2_mul16_vpunpck() local 562 …const __m256i vk11x0123456789ABCDEF = _mm256_cvtepi8_epi16(_mm_loadu_si128((const __m128i*) ((uint… in xnn_qc8_dwconv_minmax_fp32_ukernel_up16x25__avx2_mul16_vpunpck() local
|
H A D | up32x25-minmax-fp32-avx2-mul16-add16-vpunpck.c | 316 …const __m256i vk11x0123456789ABCDEF = _mm256_cvtepi8_epi16(_mm_loadu_si128((const __m128i*) ((uint… in xnn_qc8_dwconv_minmax_fp32_ukernel_up32x25__avx2_mul16_add16_vpunpck() local 672 …const __m256i vk11x0123456789ABCDEF = _mm256_cvtepi8_epi16(_mm_loadu_si128((const __m128i*) (k + 3… in xnn_qc8_dwconv_minmax_fp32_ukernel_up32x25__avx2_mul16_add16_vpunpck() local
|
H A D | up32x25-minmax-fp32-avx2-mul16-vpmovsx.c | 336 …const __m256i vk11x0123456789ABCDEF = _mm256_cvtepi8_epi16(_mm_loadu_si128((const __m128i*) ((uint… in xnn_qc8_dwconv_minmax_fp32_ukernel_up32x25__avx2_mul16_vpmovsx() local 695 …const __m256i vk11x0123456789ABCDEF = _mm256_cvtepi8_epi16(_mm_loadu_si128((const __m128i*) (k + 3… in xnn_qc8_dwconv_minmax_fp32_ukernel_up32x25__avx2_mul16_vpmovsx() local
|
H A D | up32x25-minmax-fp32-avx2-mul16-vpunpck.c | 352 …const __m256i vk11x0123456789ABCDEF = _mm256_cvtepi8_epi16(_mm_loadu_si128((const __m128i*) ((uint… in xnn_qc8_dwconv_minmax_fp32_ukernel_up32x25__avx2_mul16_vpunpck() local 744 …const __m256i vk11x0123456789ABCDEF = _mm256_cvtepi8_epi16(_mm_loadu_si128((const __m128i*) (k + 3… in xnn_qc8_dwconv_minmax_fp32_ukernel_up32x25__avx2_mul16_vpunpck() local
|
H A D | up16x25-minmax-fp32-neon-mul8-ld128.c | 283 … const int8x16_t vk11x0123456789ABCDEF = vld1q_s8(w); w = (const void*) ((const int8_t*) w + 16); in xnn_qc8_dwconv_minmax_fp32_ukernel_up16x25__neon_mul8_ld128() local
|