| /aosp_15_r20/external/XNNPACK/src/f32-dwconv/gen/ |
| H A D | up16x25-minmax-fma3.c | 371 const __m256 vk22x01234567 = _mm256_load_ps(w + 368); in xnn_f32_dwconv_minmax_ukernel_up16x25__fma3() local
373 vacc01234567p0 = _mm256_fmadd_ps(vi22x01234567, vk22x01234567, vacc01234567p0); in xnn_f32_dwconv_minmax_ukernel_up16x25__fma3()
544 const __m256 vk22x01234567 = _mm256_load_ps(w + 368); in xnn_f32_dwconv_minmax_ukernel_up16x25__fma3() local
545 vacc01234567p0 = _mm256_fmadd_ps(vi22x01234567, vk22x01234567, vacc01234567p0); in xnn_f32_dwconv_minmax_ukernel_up16x25__fma3()
664 const __m256 vk22x01234567 = _mm256_load_ps(w + 368); in xnn_f32_dwconv_minmax_ukernel_up16x25__fma3() local
665 vacc01234567p0 = _mm256_fmadd_ps(vi22x01234567, vk22x01234567, vacc01234567p0); in xnn_f32_dwconv_minmax_ukernel_up16x25__fma3()
|
| H A D | up16x25-minmax-avx.c | 371 const __m256 vk22x01234567 = _mm256_load_ps(w + 368); in xnn_f32_dwconv_minmax_ukernel_up16x25__avx() local
373 vacc01234567p0 = _mm256_add_ps(vacc01234567p0, _mm256_mul_ps(vi22x01234567, vk22x01234567)); in xnn_f32_dwconv_minmax_ukernel_up16x25__avx()
544 const __m256 vk22x01234567 = _mm256_load_ps(w + 368); in xnn_f32_dwconv_minmax_ukernel_up16x25__avx() local
545 vacc01234567p0 = _mm256_add_ps(vacc01234567p0, _mm256_mul_ps(vi22x01234567, vk22x01234567)); in xnn_f32_dwconv_minmax_ukernel_up16x25__avx()
664 const __m256 vk22x01234567 = _mm256_load_ps(w + 368); in xnn_f32_dwconv_minmax_ukernel_up16x25__avx() local
665 vacc01234567p0 = _mm256_add_ps(vacc01234567p0, _mm256_mul_ps(vi22x01234567, vk22x01234567)); in xnn_f32_dwconv_minmax_ukernel_up16x25__avx()
|
| H A D | up16x25-minmax-fma3-acc2.c | 371 const __m256 vk22x01234567 = _mm256_load_ps(w + 368); in xnn_f32_dwconv_minmax_ukernel_up16x25__fma3_acc2() local
373 vacc01234567p0 = _mm256_fmadd_ps(vi22x01234567, vk22x01234567, vacc01234567p0); in xnn_f32_dwconv_minmax_ukernel_up16x25__fma3_acc2()
547 const __m256 vk22x01234567 = _mm256_load_ps(w + 368); in xnn_f32_dwconv_minmax_ukernel_up16x25__fma3_acc2() local
548 vacc01234567p0 = _mm256_fmadd_ps(vi22x01234567, vk22x01234567, vacc01234567p0); in xnn_f32_dwconv_minmax_ukernel_up16x25__fma3_acc2()
669 const __m256 vk22x01234567 = _mm256_load_ps(w + 368); in xnn_f32_dwconv_minmax_ukernel_up16x25__fma3_acc2() local
670 vacc01234567p0 = _mm256_fmadd_ps(vi22x01234567, vk22x01234567, vacc01234567p0); in xnn_f32_dwconv_minmax_ukernel_up16x25__fma3_acc2()
|
| H A D | up8x25-minmax-fma3-acc2.c | 303 const __m256 vk22x01234567 = _mm256_load_ps(w + 184); in xnn_f32_dwconv_minmax_ukernel_up8x25__fma3_acc2() local
304 vacc01234567p0 = _mm256_fmadd_ps(vi22x01234567, vk22x01234567, vacc01234567p0); in xnn_f32_dwconv_minmax_ukernel_up8x25__fma3_acc2()
425 const __m256 vk22x01234567 = _mm256_load_ps(w + 184); in xnn_f32_dwconv_minmax_ukernel_up8x25__fma3_acc2() local
426 vacc01234567p0 = _mm256_fmadd_ps(vi22x01234567, vk22x01234567, vacc01234567p0); in xnn_f32_dwconv_minmax_ukernel_up8x25__fma3_acc2()
|
| H A D | up8x25-minmax-fma3.c | 303 const __m256 vk22x01234567 = _mm256_load_ps(w + 184); in xnn_f32_dwconv_minmax_ukernel_up8x25__fma3() local
304 vacc01234567p0 = _mm256_fmadd_ps(vi22x01234567, vk22x01234567, vacc01234567p0); in xnn_f32_dwconv_minmax_ukernel_up8x25__fma3()
423 const __m256 vk22x01234567 = _mm256_load_ps(w + 184); in xnn_f32_dwconv_minmax_ukernel_up8x25__fma3() local
424 vacc01234567p0 = _mm256_fmadd_ps(vi22x01234567, vk22x01234567, vacc01234567p0); in xnn_f32_dwconv_minmax_ukernel_up8x25__fma3()
|
| H A D | up8x25-minmax-avx.c | 303 const __m256 vk22x01234567 = _mm256_load_ps(w + 184); in xnn_f32_dwconv_minmax_ukernel_up8x25__avx() local
304 vacc01234567p0 = _mm256_add_ps(vacc01234567p0, _mm256_mul_ps(vi22x01234567, vk22x01234567)); in xnn_f32_dwconv_minmax_ukernel_up8x25__avx()
423 const __m256 vk22x01234567 = _mm256_load_ps(w + 184); in xnn_f32_dwconv_minmax_ukernel_up8x25__avx() local
424 vacc01234567p0 = _mm256_add_ps(vacc01234567p0, _mm256_mul_ps(vi22x01234567, vk22x01234567)); in xnn_f32_dwconv_minmax_ukernel_up8x25__avx()
|
| H A D | up16x25-minmax-avx-acc2.c | 371 const __m256 vk22x01234567 = _mm256_load_ps(w + 368); in xnn_f32_dwconv_minmax_ukernel_up16x25__avx_acc2() local
373 vacc01234567p0 = _mm256_add_ps(vacc01234567p0, _mm256_mul_ps(vi22x01234567, vk22x01234567)); in xnn_f32_dwconv_minmax_ukernel_up16x25__avx_acc2()
547 const __m256 vk22x01234567 = _mm256_load_ps(w + 368); in xnn_f32_dwconv_minmax_ukernel_up16x25__avx_acc2() local
548 vacc01234567p0 = _mm256_add_ps(vacc01234567p0, _mm256_mul_ps(vi22x01234567, vk22x01234567)); in xnn_f32_dwconv_minmax_ukernel_up16x25__avx_acc2()
669 const __m256 vk22x01234567 = _mm256_load_ps(w + 368); in xnn_f32_dwconv_minmax_ukernel_up16x25__avx_acc2() local
670 vacc01234567p0 = _mm256_add_ps(vacc01234567p0, _mm256_mul_ps(vi22x01234567, vk22x01234567)); in xnn_f32_dwconv_minmax_ukernel_up16x25__avx_acc2()
|
| H A D | up8x25-minmax-avx-acc2.c | 303 const __m256 vk22x01234567 = _mm256_load_ps(w + 184); in xnn_f32_dwconv_minmax_ukernel_up8x25__avx_acc2() local
304 vacc01234567p0 = _mm256_add_ps(vacc01234567p0, _mm256_mul_ps(vi22x01234567, vk22x01234567)); in xnn_f32_dwconv_minmax_ukernel_up8x25__avx_acc2()
425 const __m256 vk22x01234567 = _mm256_load_ps(w + 184); in xnn_f32_dwconv_minmax_ukernel_up8x25__avx_acc2() local
426 vacc01234567p0 = _mm256_add_ps(vacc01234567p0, _mm256_mul_ps(vi22x01234567, vk22x01234567)); in xnn_f32_dwconv_minmax_ukernel_up8x25__avx_acc2()
|
| /aosp_15_r20/external/XNNPACK/src/qs8-dwconv/gen/ |
| H A D | up8x25-minmax-fp32-neonv8-mul16.c | 305 … const int16x8_t vk22x01234567 = vmovl_s8(vld1_s8(w)); w = (const void*) ((const int8_t*) w + 8); in xnn_qs8_dwconv_minmax_fp32_ukernel_up8x25__neonv8_mul16() local
307 vacc0123 = vmlal_s16(vacc0123, vget_low_s16(vi22x01234567), vget_low_s16(vk22x01234567)); in xnn_qs8_dwconv_minmax_fp32_ukernel_up8x25__neonv8_mul16()
308 vacc4567 = vmlal_s16(vacc4567, vget_high_s16(vi22x01234567), vget_high_s16(vk22x01234567)); in xnn_qs8_dwconv_minmax_fp32_ukernel_up8x25__neonv8_mul16()
467 const int16x8_t vk22x01234567 = vmovl_s8(vld1_s8((const void*) ((const int8_t*) w + 176))); in xnn_qs8_dwconv_minmax_fp32_ukernel_up8x25__neonv8_mul16() local
469 vacc0123 = vmlal_s16(vacc0123, vget_low_s16(vi22x01234567), vget_low_s16(vk22x01234567)); in xnn_qs8_dwconv_minmax_fp32_ukernel_up8x25__neonv8_mul16()
470 vacc4567 = vmlal_s16(vacc4567, vget_high_s16(vi22x01234567), vget_high_s16(vk22x01234567)); in xnn_qs8_dwconv_minmax_fp32_ukernel_up8x25__neonv8_mul16()
|
| H A D | up8x25-minmax-fp32-neon-mul16.c | 305 … const int16x8_t vk22x01234567 = vmovl_s8(vld1_s8(w)); w = (const void*) ((const int8_t*) w + 8); in xnn_qs8_dwconv_minmax_fp32_ukernel_up8x25__neon_mul16() local
307 vacc0123 = vmlal_s16(vacc0123, vget_low_s16(vi22x01234567), vget_low_s16(vk22x01234567)); in xnn_qs8_dwconv_minmax_fp32_ukernel_up8x25__neon_mul16()
308 vacc4567 = vmlal_s16(vacc4567, vget_high_s16(vi22x01234567), vget_high_s16(vk22x01234567)); in xnn_qs8_dwconv_minmax_fp32_ukernel_up8x25__neon_mul16()
468 const int16x8_t vk22x01234567 = vmovl_s8(vld1_s8((const void*) ((const int8_t*) w + 176))); in xnn_qs8_dwconv_minmax_fp32_ukernel_up8x25__neon_mul16() local
470 vacc0123 = vmlal_s16(vacc0123, vget_low_s16(vi22x01234567), vget_low_s16(vk22x01234567)); in xnn_qs8_dwconv_minmax_fp32_ukernel_up8x25__neon_mul16()
471 vacc4567 = vmlal_s16(vacc4567, vget_high_s16(vi22x01234567), vget_high_s16(vk22x01234567)); in xnn_qs8_dwconv_minmax_fp32_ukernel_up8x25__neon_mul16()
|
| H A D | up8x25-minmax-rndnu-neon-mul16.c | 306 … const int16x8_t vk22x01234567 = vmovl_s8(vld1_s8(w)); w = (const void*) ((const int8_t*) w + 8); in xnn_qs8_dwconv_minmax_rndnu_ukernel_up8x25__neon_mul16() local
308 vacc0123 = vmlal_s16(vacc0123, vget_low_s16(vi22x01234567), vget_low_s16(vk22x01234567)); in xnn_qs8_dwconv_minmax_rndnu_ukernel_up8x25__neon_mul16()
309 vacc4567 = vmlal_s16(vacc4567, vget_high_s16(vi22x01234567), vget_high_s16(vk22x01234567)); in xnn_qs8_dwconv_minmax_rndnu_ukernel_up8x25__neon_mul16()
468 const int16x8_t vk22x01234567 = vmovl_s8(vld1_s8((const void*) ((const int8_t*) w + 176))); in xnn_qs8_dwconv_minmax_rndnu_ukernel_up8x25__neon_mul16() local
470 vacc0123 = vmlal_s16(vacc0123, vget_low_s16(vi22x01234567), vget_low_s16(vk22x01234567)); in xnn_qs8_dwconv_minmax_rndnu_ukernel_up8x25__neon_mul16()
471 vacc4567 = vmlal_s16(vacc4567, vget_high_s16(vi22x01234567), vget_high_s16(vk22x01234567)); in xnn_qs8_dwconv_minmax_rndnu_ukernel_up8x25__neon_mul16()
|
| H A D | up8x25-minmax-fp32-sse2-mul16-add16.c | 422 …const __m128i vk22x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 8 * sizeof(int32_t… in xnn_qs8_dwconv_minmax_fp32_ukernel_up8x25__sse2_mul16_add16() local
426 … const __m128i vxk22x01234567 = _mm_srai_epi16(_mm_unpacklo_epi8(vk22x01234567, vk22x01234567), 8); in xnn_qs8_dwconv_minmax_fp32_ukernel_up8x25__sse2_mul16_add16()
723 …const __m128i vk22x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 8 * sizeof(int32_t… in xnn_qs8_dwconv_minmax_fp32_ukernel_up8x25__sse2_mul16_add16() local
726 … const __m128i vxk22x01234567 = _mm_srai_epi16(_mm_unpacklo_epi8(vk22x01234567, vk22x01234567), 8); in xnn_qs8_dwconv_minmax_fp32_ukernel_up8x25__sse2_mul16_add16()
|
| /aosp_15_r20/external/XNNPACK/src/qc8-dwconv/gen/ |
| H A D | up8x25-minmax-fp32-neonv8-mul16.c | 304 … const int16x8_t vk22x01234567 = vmovl_s8(vld1_s8(w)); w = (const void*) ((const int8_t*) w + 8); in xnn_qc8_dwconv_minmax_fp32_ukernel_up8x25__neonv8_mul16() local
306 vacc0123 = vmlal_s16(vacc0123, vget_low_s16(vi22x01234567), vget_low_s16(vk22x01234567)); in xnn_qc8_dwconv_minmax_fp32_ukernel_up8x25__neonv8_mul16()
307 vacc4567 = vmlal_s16(vacc4567, vget_high_s16(vi22x01234567), vget_high_s16(vk22x01234567)); in xnn_qc8_dwconv_minmax_fp32_ukernel_up8x25__neonv8_mul16()
469 const int16x8_t vk22x01234567 = vmovl_s8(vld1_s8((const void*) ((const int8_t*) w + 176))); in xnn_qc8_dwconv_minmax_fp32_ukernel_up8x25__neonv8_mul16() local
471 vacc0123 = vmlal_s16(vacc0123, vget_low_s16(vi22x01234567), vget_low_s16(vk22x01234567)); in xnn_qc8_dwconv_minmax_fp32_ukernel_up8x25__neonv8_mul16()
472 vacc4567 = vmlal_s16(vacc4567, vget_high_s16(vi22x01234567), vget_high_s16(vk22x01234567)); in xnn_qc8_dwconv_minmax_fp32_ukernel_up8x25__neonv8_mul16()
|
| H A D | up8x25-minmax-fp32-neon-mul16.c | 304 … const int16x8_t vk22x01234567 = vmovl_s8(vld1_s8(w)); w = (const void*) ((const int8_t*) w + 8); in xnn_qc8_dwconv_minmax_fp32_ukernel_up8x25__neon_mul16() local
306 vacc0123 = vmlal_s16(vacc0123, vget_low_s16(vi22x01234567), vget_low_s16(vk22x01234567)); in xnn_qc8_dwconv_minmax_fp32_ukernel_up8x25__neon_mul16()
307 vacc4567 = vmlal_s16(vacc4567, vget_high_s16(vi22x01234567), vget_high_s16(vk22x01234567)); in xnn_qc8_dwconv_minmax_fp32_ukernel_up8x25__neon_mul16()
470 const int16x8_t vk22x01234567 = vmovl_s8(vld1_s8((const void*) ((const int8_t*) w + 176))); in xnn_qc8_dwconv_minmax_fp32_ukernel_up8x25__neon_mul16() local
472 vacc0123 = vmlal_s16(vacc0123, vget_low_s16(vi22x01234567), vget_low_s16(vk22x01234567)); in xnn_qc8_dwconv_minmax_fp32_ukernel_up8x25__neon_mul16()
473 vacc4567 = vmlal_s16(vacc4567, vget_high_s16(vi22x01234567), vget_high_s16(vk22x01234567)); in xnn_qc8_dwconv_minmax_fp32_ukernel_up8x25__neon_mul16()
|
| H A D | up8x25-minmax-fp32-sse2-mul16-add16.c | 422 …const __m128i vk22x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 8 * sizeof(int32_t… in xnn_qc8_dwconv_minmax_fp32_ukernel_up8x25__sse2_mul16_add16() local
426 … const __m128i vxk22x01234567 = _mm_srai_epi16(_mm_unpacklo_epi8(vk22x01234567, vk22x01234567), 8); in xnn_qc8_dwconv_minmax_fp32_ukernel_up8x25__sse2_mul16_add16()
725 …const __m128i vk22x01234567 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 8 * sizeof(int32_t… in xnn_qc8_dwconv_minmax_fp32_ukernel_up8x25__sse2_mul16_add16() local
728 … const __m128i vxk22x01234567 = _mm_srai_epi16(_mm_unpacklo_epi8(vk22x01234567, vk22x01234567), 8); in xnn_qc8_dwconv_minmax_fp32_ukernel_up8x25__sse2_mul16_add16()
|
| /aosp_15_r20/external/XNNPACK/src/qu8-dwconv/gen/ |
| H A D | up8x25-minmax-fp32-neonv8-mul16.c | 306 …const int16x8_t vk22x01234567 = vreinterpretq_s16_u16(vsubl_u8(vld1_u8(w), vkernel_zero_point)); w… in xnn_qu8_dwconv_minmax_fp32_ukernel_up8x25__neonv8_mul16() local
308 vacc0123 = vmlal_s16(vacc0123, vget_low_s16(vi22x01234567), vget_low_s16(vk22x01234567)); in xnn_qu8_dwconv_minmax_fp32_ukernel_up8x25__neonv8_mul16()
309 vacc4567 = vmlal_s16(vacc4567, vget_high_s16(vi22x01234567), vget_high_s16(vk22x01234567)); in xnn_qu8_dwconv_minmax_fp32_ukernel_up8x25__neonv8_mul16()
468 …const int16x8_t vk22x01234567 = vreinterpretq_s16_u16(vsubl_u8(vld1_u8((const void*) ((const uint8… in xnn_qu8_dwconv_minmax_fp32_ukernel_up8x25__neonv8_mul16() local
470 vacc0123 = vmlal_s16(vacc0123, vget_low_s16(vi22x01234567), vget_low_s16(vk22x01234567)); in xnn_qu8_dwconv_minmax_fp32_ukernel_up8x25__neonv8_mul16()
471 vacc4567 = vmlal_s16(vacc4567, vget_high_s16(vi22x01234567), vget_high_s16(vk22x01234567)); in xnn_qu8_dwconv_minmax_fp32_ukernel_up8x25__neonv8_mul16()
|
| H A D | up8x25-minmax-fp32-neon-mul16.c | 306 …const int16x8_t vk22x01234567 = vreinterpretq_s16_u16(vsubl_u8(vld1_u8(w), vkernel_zero_point)); w… in xnn_qu8_dwconv_minmax_fp32_ukernel_up8x25__neon_mul16() local
308 vacc0123 = vmlal_s16(vacc0123, vget_low_s16(vi22x01234567), vget_low_s16(vk22x01234567)); in xnn_qu8_dwconv_minmax_fp32_ukernel_up8x25__neon_mul16()
309 vacc4567 = vmlal_s16(vacc4567, vget_high_s16(vi22x01234567), vget_high_s16(vk22x01234567)); in xnn_qu8_dwconv_minmax_fp32_ukernel_up8x25__neon_mul16()
469 …const int16x8_t vk22x01234567 = vreinterpretq_s16_u16(vsubl_u8(vld1_u8((const void*) ((const uint8… in xnn_qu8_dwconv_minmax_fp32_ukernel_up8x25__neon_mul16() local
471 vacc0123 = vmlal_s16(vacc0123, vget_low_s16(vi22x01234567), vget_low_s16(vk22x01234567)); in xnn_qu8_dwconv_minmax_fp32_ukernel_up8x25__neon_mul16()
472 vacc4567 = vmlal_s16(vacc4567, vget_high_s16(vi22x01234567), vget_high_s16(vk22x01234567)); in xnn_qu8_dwconv_minmax_fp32_ukernel_up8x25__neon_mul16()
|
| H A D | up8x25-minmax-rndnu-neon-mul16.c | 307 …const int16x8_t vk22x01234567 = vreinterpretq_s16_u16(vsubl_u8(vld1_u8(w), vkernel_zero_point)); w… in xnn_qu8_dwconv_minmax_rndnu_ukernel_up8x25__neon_mul16() local
309 vacc0123 = vmlal_s16(vacc0123, vget_low_s16(vi22x01234567), vget_low_s16(vk22x01234567)); in xnn_qu8_dwconv_minmax_rndnu_ukernel_up8x25__neon_mul16()
310 vacc4567 = vmlal_s16(vacc4567, vget_high_s16(vi22x01234567), vget_high_s16(vk22x01234567)); in xnn_qu8_dwconv_minmax_rndnu_ukernel_up8x25__neon_mul16()
469 …const int16x8_t vk22x01234567 = vreinterpretq_s16_u16(vsubl_u8(vld1_u8((const void*) ((const uint8… in xnn_qu8_dwconv_minmax_rndnu_ukernel_up8x25__neon_mul16() local
471 vacc0123 = vmlal_s16(vacc0123, vget_low_s16(vi22x01234567), vget_low_s16(vk22x01234567)); in xnn_qu8_dwconv_minmax_rndnu_ukernel_up8x25__neon_mul16()
472 vacc4567 = vmlal_s16(vacc4567, vget_high_s16(vi22x01234567), vget_high_s16(vk22x01234567)); in xnn_qu8_dwconv_minmax_rndnu_ukernel_up8x25__neon_mul16()
|
| /aosp_15_r20/external/XNNPACK/src/f16-dwconv/gen/ |
| H A D | up16x25-minmax-neonfp16arith-acc2.c | 327 const float16x8_t vk22x01234567 = vld1q_f16(w); w += 8; in xnn_f16_dwconv_minmax_ukernel_up16x25__neonfp16arith_acc2() local
329 vacc01234567p0 = vfmaq_f16(vacc01234567p0, vi22x01234567, vk22x01234567); in xnn_f16_dwconv_minmax_ukernel_up16x25__neonfp16arith_acc2()
451 const float16x8_t vk22x01234567 = vld1q_f16(w + 360); in xnn_f16_dwconv_minmax_ukernel_up16x25__neonfp16arith_acc2() local
452 vacc01234567p0 = vfmaq_f16(vacc01234567p0, vi22x01234567, vk22x01234567); in xnn_f16_dwconv_minmax_ukernel_up16x25__neonfp16arith_acc2()
563 const float16x8_t vk22x01234567 = vld1q_f16(w + 368); in xnn_f16_dwconv_minmax_ukernel_up16x25__neonfp16arith_acc2() local
564 vacc01234567p0 = vfmaq_f16(vacc01234567p0, vi22x01234567, vk22x01234567); in xnn_f16_dwconv_minmax_ukernel_up16x25__neonfp16arith_acc2()
|
| H A D | up16x25-minmax-neonfp16arith.c | 327 const float16x8_t vk22x01234567 = vld1q_f16(w); w += 8; in xnn_f16_dwconv_minmax_ukernel_up16x25__neonfp16arith() local
329 vacc01234567p0 = vfmaq_f16(vacc01234567p0, vi22x01234567, vk22x01234567); in xnn_f16_dwconv_minmax_ukernel_up16x25__neonfp16arith()
448 const float16x8_t vk22x01234567 = vld1q_f16(w + 360); in xnn_f16_dwconv_minmax_ukernel_up16x25__neonfp16arith() local
449 vacc01234567p0 = vfmaq_f16(vacc01234567p0, vi22x01234567, vk22x01234567); in xnn_f16_dwconv_minmax_ukernel_up16x25__neonfp16arith()
558 const float16x8_t vk22x01234567 = vld1q_f16(w + 368); in xnn_f16_dwconv_minmax_ukernel_up16x25__neonfp16arith() local
559 vacc01234567p0 = vfmaq_f16(vacc01234567p0, vi22x01234567, vk22x01234567); in xnn_f16_dwconv_minmax_ukernel_up16x25__neonfp16arith()
|
| H A D | up8x25-minmax-neonfp16arith-acc2.c | 259 const float16x8_t vk22x01234567 = vld1q_f16(w); w += 8; in xnn_f16_dwconv_minmax_ukernel_up8x25__neonfp16arith_acc2() local
260 vacc01234567p0 = vfmaq_f16(vacc01234567p0, vi22x01234567, vk22x01234567); in xnn_f16_dwconv_minmax_ukernel_up8x25__neonfp16arith_acc2()
371 const float16x8_t vk22x01234567 = vld1q_f16(w); w += 8; in xnn_f16_dwconv_minmax_ukernel_up8x25__neonfp16arith_acc2() local
372 vacc01234567p0 = vfmaq_f16(vacc01234567p0, vi22x01234567, vk22x01234567); in xnn_f16_dwconv_minmax_ukernel_up8x25__neonfp16arith_acc2()
|
| H A D | up8x25-minmax-neonfp16arith.c | 259 const float16x8_t vk22x01234567 = vld1q_f16(w); w += 8; in xnn_f16_dwconv_minmax_ukernel_up8x25__neonfp16arith() local
260 vacc01234567p0 = vfmaq_f16(vacc01234567p0, vi22x01234567, vk22x01234567); in xnn_f16_dwconv_minmax_ukernel_up8x25__neonfp16arith()
369 const float16x8_t vk22x01234567 = vld1q_f16(w); w += 8; in xnn_f16_dwconv_minmax_ukernel_up8x25__neonfp16arith() local
370 vacc01234567p0 = vfmaq_f16(vacc01234567p0, vi22x01234567, vk22x01234567); in xnn_f16_dwconv_minmax_ukernel_up8x25__neonfp16arith()
|
| H A D | up16x25-minmax-fma3.c | 374 const __m256 vk22x01234567 = _mm256_cvtph_ps(_mm_loadu_si128((const __m128i*) (w + 368))); in xnn_f16_dwconv_minmax_ukernel_up16x25__fma3() local
376 …vacc01234567p0 = _mm256_cvtph_ps(_mm256_cvtps_ph(_mm256_fmadd_ps(vi22x01234567, vk22x01234567, vac… in xnn_f16_dwconv_minmax_ukernel_up16x25__fma3()
547 const __m256 vk22x01234567 = _mm256_cvtph_ps(_mm_load_si128((const __m128i*) (w + 368))); in xnn_f16_dwconv_minmax_ukernel_up16x25__fma3() local
548 …vacc01234567p0 = _mm256_cvtph_ps(_mm256_cvtps_ph(_mm256_fmadd_ps(vi22x01234567, vk22x01234567, vac… in xnn_f16_dwconv_minmax_ukernel_up16x25__fma3()
689 const __m256 vk22x01234567 = _mm256_cvtph_ps(_mm_load_si128((const __m128i*) (w + 368))); in xnn_f16_dwconv_minmax_ukernel_up16x25__fma3() local
690 …vacc01234567p0 = _mm256_cvtph_ps(_mm256_cvtps_ph(_mm256_fmadd_ps(vi22x01234567, vk22x01234567, vac… in xnn_f16_dwconv_minmax_ukernel_up16x25__fma3()
|
| H A D | up8x25-minmax-fma3-acc2.c | 306 const __m256 vk22x01234567 = _mm256_cvtph_ps(_mm_loadu_si128((const __m128i*) (w + 184))); in xnn_f16_dwconv_minmax_ukernel_up8x25__fma3_acc2() local
307 …vacc01234567p0 = _mm256_cvtph_ps(_mm256_cvtps_ph(_mm256_fmadd_ps(vi22x01234567, vk22x01234567, vac… in xnn_f16_dwconv_minmax_ukernel_up8x25__fma3_acc2()
450 const __m256 vk22x01234567 = _mm256_cvtph_ps(_mm_load_si128((const __m128i*) (w + 184))); in xnn_f16_dwconv_minmax_ukernel_up8x25__fma3_acc2() local
451 …vacc01234567p0 = _mm256_cvtph_ps(_mm256_cvtps_ph(_mm256_fmadd_ps(vi22x01234567, vk22x01234567, vac… in xnn_f16_dwconv_minmax_ukernel_up8x25__fma3_acc2()
|
| H A D | up8x25-minmax-fma3.c | 306 const __m256 vk22x01234567 = _mm256_cvtph_ps(_mm_loadu_si128((const __m128i*) (w + 184))); in xnn_f16_dwconv_minmax_ukernel_up8x25__fma3() local
307 …vacc01234567p0 = _mm256_cvtph_ps(_mm256_cvtps_ph(_mm256_fmadd_ps(vi22x01234567, vk22x01234567, vac… in xnn_f16_dwconv_minmax_ukernel_up8x25__fma3()
448 const __m256 vk22x01234567 = _mm256_cvtph_ps(_mm_load_si128((const __m128i*) (w + 184))); in xnn_f16_dwconv_minmax_ukernel_up8x25__fma3() local
449 …vacc01234567p0 = _mm256_cvtph_ps(_mm256_cvtps_ph(_mm256_fmadd_ps(vi22x01234567, vk22x01234567, vac… in xnn_f16_dwconv_minmax_ukernel_up8x25__fma3()
|