/external/XNNPACK/src/qu8-gavgpool/gen/ |
D | 7x-minmax-fp32-sse41-c24.c | 132 __m128 vfpaccGHIJ = _mm_cvtepi32_ps(vaccGHIJ); in xnn_qu8_gavgpool_minmax_fp32_ukernel_7x__sse41_c24() local 139 vfpaccGHIJ = _mm_mul_ps(vfpaccGHIJ, vscale); in xnn_qu8_gavgpool_minmax_fp32_ukernel_7x__sse41_c24() 146 vfpaccGHIJ = _mm_min_ps(vfpaccGHIJ, voutput_max_less_zero_point); in xnn_qu8_gavgpool_minmax_fp32_ukernel_7x__sse41_c24() 153 vaccGHIJ = _mm_cvtps_epi32(vfpaccGHIJ); in xnn_qu8_gavgpool_minmax_fp32_ukernel_7x__sse41_c24()
|
D | 7x-minmax-fp32-sse2-c24.c | 159 __m128 vfpaccGHIJ = _mm_cvtepi32_ps(vaccGHIJ); in xnn_qu8_gavgpool_minmax_fp32_ukernel_7x__sse2_c24() local 166 vfpaccGHIJ = _mm_mul_ps(vfpaccGHIJ, vscale); in xnn_qu8_gavgpool_minmax_fp32_ukernel_7x__sse2_c24() 173 vfpaccGHIJ = _mm_min_ps(vfpaccGHIJ, voutput_max_less_zero_point); in xnn_qu8_gavgpool_minmax_fp32_ukernel_7x__sse2_c24() 180 vaccGHIJ = _mm_cvtps_epi32(vfpaccGHIJ); in xnn_qu8_gavgpool_minmax_fp32_ukernel_7x__sse2_c24()
|
D | 7x-minmax-fp32-neonv8-c24.c | 116 float32x4_t vfpaccGHIJ = vcvtq_f32_s32(vaccGHIJ); in xnn_qu8_gavgpool_minmax_fp32_ukernel_7x__neonv8_c24() local 123 vfpaccGHIJ = vmulq_f32(vfpaccGHIJ, vscale); in xnn_qu8_gavgpool_minmax_fp32_ukernel_7x__neonv8_c24() 130 vaccGHIJ = vcvtnq_s32_f32(vfpaccGHIJ); in xnn_qu8_gavgpool_minmax_fp32_ukernel_7x__neonv8_c24()
|
D | 7x-minmax-fp32-neon-c24.c | 116 float32x4_t vfpaccGHIJ = vcvtq_f32_s32(vaccGHIJ); in xnn_qu8_gavgpool_minmax_fp32_ukernel_7x__neon_c24() local 123 vfpaccGHIJ = vmulq_f32(vfpaccGHIJ, vscale); in xnn_qu8_gavgpool_minmax_fp32_ukernel_7x__neon_c24() 130 vaccGHIJ = vreinterpretq_s32_f32(vaddq_f32(vfpaccGHIJ, vmagic_bias)); in xnn_qu8_gavgpool_minmax_fp32_ukernel_7x__neon_c24()
|
D | 7x-minmax-fp32-neon-c32.c | 131 float32x4_t vfpaccGHIJ = vcvtq_f32_s32(vaccGHIJ); in xnn_qu8_gavgpool_minmax_fp32_ukernel_7x__neon_c32() local 140 vfpaccGHIJ = vmulq_f32(vfpaccGHIJ, vscale); in xnn_qu8_gavgpool_minmax_fp32_ukernel_7x__neon_c32() 149 vaccGHIJ = vreinterpretq_s32_f32(vaddq_f32(vfpaccGHIJ, vmagic_bias)); in xnn_qu8_gavgpool_minmax_fp32_ukernel_7x__neon_c32()
|
D | 7x-minmax-fp32-neonv8-c32.c | 131 float32x4_t vfpaccGHIJ = vcvtq_f32_s32(vaccGHIJ); in xnn_qu8_gavgpool_minmax_fp32_ukernel_7x__neonv8_c32() local 140 vfpaccGHIJ = vmulq_f32(vfpaccGHIJ, vscale); in xnn_qu8_gavgpool_minmax_fp32_ukernel_7x__neonv8_c32() 149 vaccGHIJ = vcvtnq_s32_f32(vfpaccGHIJ); in xnn_qu8_gavgpool_minmax_fp32_ukernel_7x__neonv8_c32()
|
D | 7p7x-minmax-fp32-sse41-c24.c | 385 __m128 vfpaccGHIJ = _mm_cvtepi32_ps(vaccGHIJ); in xnn_qu8_gavgpool_minmax_fp32_ukernel_7p7x__sse41_c24() local 392 vfpaccGHIJ = _mm_mul_ps(vfpaccGHIJ, vscale); in xnn_qu8_gavgpool_minmax_fp32_ukernel_7p7x__sse41_c24() 399 vfpaccGHIJ = _mm_min_ps(vfpaccGHIJ, voutput_max_less_zero_point); in xnn_qu8_gavgpool_minmax_fp32_ukernel_7p7x__sse41_c24() 406 vaccGHIJ = _mm_cvtps_epi32(vfpaccGHIJ); in xnn_qu8_gavgpool_minmax_fp32_ukernel_7p7x__sse41_c24()
|
D | 7p7x-minmax-fp32-sse2-c24.c | 492 __m128 vfpaccGHIJ = _mm_cvtepi32_ps(vaccGHIJ); in xnn_qu8_gavgpool_minmax_fp32_ukernel_7p7x__sse2_c24() local 499 vfpaccGHIJ = _mm_mul_ps(vfpaccGHIJ, vscale); in xnn_qu8_gavgpool_minmax_fp32_ukernel_7p7x__sse2_c24() 506 vfpaccGHIJ = _mm_min_ps(vfpaccGHIJ, voutput_max_less_zero_point); in xnn_qu8_gavgpool_minmax_fp32_ukernel_7p7x__sse2_c24() 513 vaccGHIJ = _mm_cvtps_epi32(vfpaccGHIJ); in xnn_qu8_gavgpool_minmax_fp32_ukernel_7p7x__sse2_c24()
|
D | 7p7x-minmax-fp32-neon-c24.c | 322 float32x4_t vfpaccGHIJ = vcvtq_f32_s32(vaccGHIJ); in xnn_qu8_gavgpool_minmax_fp32_ukernel_7p7x__neon_c24() local 329 vfpaccGHIJ = vmulq_f32(vfpaccGHIJ, vscale); in xnn_qu8_gavgpool_minmax_fp32_ukernel_7p7x__neon_c24() 336 vaccGHIJ = vreinterpretq_s32_f32(vaddq_f32(vfpaccGHIJ, vmagic_bias)); in xnn_qu8_gavgpool_minmax_fp32_ukernel_7p7x__neon_c24()
|
D | 7p7x-minmax-fp32-neonv8-c24.c | 322 float32x4_t vfpaccGHIJ = vcvtq_f32_s32(vaccGHIJ); in xnn_qu8_gavgpool_minmax_fp32_ukernel_7p7x__neonv8_c24() local 329 vfpaccGHIJ = vmulq_f32(vfpaccGHIJ, vscale); in xnn_qu8_gavgpool_minmax_fp32_ukernel_7p7x__neonv8_c24() 336 vaccGHIJ = vcvtnq_s32_f32(vfpaccGHIJ); in xnn_qu8_gavgpool_minmax_fp32_ukernel_7p7x__neonv8_c24()
|
/external/XNNPACK/src/qs8-gavgpool/gen/ |
D | 7x-minmax-fp32-sse41-c24.c | 131 __m128 vfpaccGHIJ = _mm_cvtepi32_ps(vaccGHIJ); in xnn_qs8_gavgpool_minmax_fp32_ukernel_7x__sse41_c24() local 138 vfpaccGHIJ = _mm_mul_ps(vfpaccGHIJ, vscale); in xnn_qs8_gavgpool_minmax_fp32_ukernel_7x__sse41_c24() 145 vfpaccGHIJ = _mm_min_ps(vfpaccGHIJ, voutput_max_less_zero_point); in xnn_qs8_gavgpool_minmax_fp32_ukernel_7x__sse41_c24() 152 vaccGHIJ = _mm_cvtps_epi32(vfpaccGHIJ); in xnn_qs8_gavgpool_minmax_fp32_ukernel_7x__sse41_c24()
|
D | 7x-minmax-fp32-sse2-c24.c | 161 __m128 vfpaccGHIJ = _mm_cvtepi32_ps(vaccGHIJ); in xnn_qs8_gavgpool_minmax_fp32_ukernel_7x__sse2_c24() local 168 vfpaccGHIJ = _mm_mul_ps(vfpaccGHIJ, vscale); in xnn_qs8_gavgpool_minmax_fp32_ukernel_7x__sse2_c24() 175 vfpaccGHIJ = _mm_min_ps(vfpaccGHIJ, voutput_max_less_zero_point); in xnn_qs8_gavgpool_minmax_fp32_ukernel_7x__sse2_c24() 182 vaccGHIJ = _mm_cvtps_epi32(vfpaccGHIJ); in xnn_qs8_gavgpool_minmax_fp32_ukernel_7x__sse2_c24()
|
D | 7x-minmax-fp32-neonv8-c24.c | 116 float32x4_t vfpaccGHIJ = vcvtq_f32_s32(vaccGHIJ); in xnn_qs8_gavgpool_minmax_fp32_ukernel_7x__neonv8_c24() local 123 vfpaccGHIJ = vmulq_f32(vfpaccGHIJ, vscale); in xnn_qs8_gavgpool_minmax_fp32_ukernel_7x__neonv8_c24() 130 vaccGHIJ = vcvtnq_s32_f32(vfpaccGHIJ); in xnn_qs8_gavgpool_minmax_fp32_ukernel_7x__neonv8_c24()
|
D | 7x-minmax-fp32-neon-c24.c | 116 float32x4_t vfpaccGHIJ = vcvtq_f32_s32(vaccGHIJ); in xnn_qs8_gavgpool_minmax_fp32_ukernel_7x__neon_c24() local 123 vfpaccGHIJ = vmulq_f32(vfpaccGHIJ, vscale); in xnn_qs8_gavgpool_minmax_fp32_ukernel_7x__neon_c24() 130 vaccGHIJ = vreinterpretq_s32_f32(vaddq_f32(vfpaccGHIJ, vmagic_bias)); in xnn_qs8_gavgpool_minmax_fp32_ukernel_7x__neon_c24()
|
D | 7x-minmax-fp32-neonv8-c32.c | 131 float32x4_t vfpaccGHIJ = vcvtq_f32_s32(vaccGHIJ); in xnn_qs8_gavgpool_minmax_fp32_ukernel_7x__neonv8_c32() local 140 vfpaccGHIJ = vmulq_f32(vfpaccGHIJ, vscale); in xnn_qs8_gavgpool_minmax_fp32_ukernel_7x__neonv8_c32() 149 vaccGHIJ = vcvtnq_s32_f32(vfpaccGHIJ); in xnn_qs8_gavgpool_minmax_fp32_ukernel_7x__neonv8_c32()
|
D | 7x-minmax-fp32-neon-c32.c | 131 float32x4_t vfpaccGHIJ = vcvtq_f32_s32(vaccGHIJ); in xnn_qs8_gavgpool_minmax_fp32_ukernel_7x__neon_c32() local 140 vfpaccGHIJ = vmulq_f32(vfpaccGHIJ, vscale); in xnn_qs8_gavgpool_minmax_fp32_ukernel_7x__neon_c32() 149 vaccGHIJ = vreinterpretq_s32_f32(vaddq_f32(vfpaccGHIJ, vmagic_bias)); in xnn_qs8_gavgpool_minmax_fp32_ukernel_7x__neon_c32()
|
D | 7p7x-minmax-fp32-sse41-c24.c | 382 __m128 vfpaccGHIJ = _mm_cvtepi32_ps(vaccGHIJ); in xnn_qs8_gavgpool_minmax_fp32_ukernel_7p7x__sse41_c24() local 389 vfpaccGHIJ = _mm_mul_ps(vfpaccGHIJ, vscale); in xnn_qs8_gavgpool_minmax_fp32_ukernel_7p7x__sse41_c24() 396 vfpaccGHIJ = _mm_min_ps(vfpaccGHIJ, voutput_max_less_zero_point); in xnn_qs8_gavgpool_minmax_fp32_ukernel_7p7x__sse41_c24() 403 vaccGHIJ = _mm_cvtps_epi32(vfpaccGHIJ); in xnn_qs8_gavgpool_minmax_fp32_ukernel_7p7x__sse41_c24()
|
D | 7p7x-minmax-fp32-sse2-c24.c | 502 __m128 vfpaccGHIJ = _mm_cvtepi32_ps(vaccGHIJ); in xnn_qs8_gavgpool_minmax_fp32_ukernel_7p7x__sse2_c24() local 509 vfpaccGHIJ = _mm_mul_ps(vfpaccGHIJ, vscale); in xnn_qs8_gavgpool_minmax_fp32_ukernel_7p7x__sse2_c24() 516 vfpaccGHIJ = _mm_min_ps(vfpaccGHIJ, voutput_max_less_zero_point); in xnn_qs8_gavgpool_minmax_fp32_ukernel_7p7x__sse2_c24() 523 vaccGHIJ = _mm_cvtps_epi32(vfpaccGHIJ); in xnn_qs8_gavgpool_minmax_fp32_ukernel_7p7x__sse2_c24()
|
D | 7p7x-minmax-fp32-neon-c24.c | 322 float32x4_t vfpaccGHIJ = vcvtq_f32_s32(vaccGHIJ); in xnn_qs8_gavgpool_minmax_fp32_ukernel_7p7x__neon_c24() local 329 vfpaccGHIJ = vmulq_f32(vfpaccGHIJ, vscale); in xnn_qs8_gavgpool_minmax_fp32_ukernel_7p7x__neon_c24() 336 vaccGHIJ = vreinterpretq_s32_f32(vaddq_f32(vfpaccGHIJ, vmagic_bias)); in xnn_qs8_gavgpool_minmax_fp32_ukernel_7p7x__neon_c24()
|
D | 7p7x-minmax-fp32-neonv8-c24.c | 322 float32x4_t vfpaccGHIJ = vcvtq_f32_s32(vaccGHIJ); in xnn_qs8_gavgpool_minmax_fp32_ukernel_7p7x__neonv8_c24() local 329 vfpaccGHIJ = vmulq_f32(vfpaccGHIJ, vscale); in xnn_qs8_gavgpool_minmax_fp32_ukernel_7p7x__neonv8_c24() 336 vaccGHIJ = vcvtnq_s32_f32(vfpaccGHIJ); in xnn_qs8_gavgpool_minmax_fp32_ukernel_7p7x__neonv8_c24()
|
/external/XNNPACK/src/qc8-dwconv/gen/ |
D | up24x9-minmax-fp32-neonv8-mul16.c | 225 float32x4_t vfpaccGHIJ = vcvtq_f32_s32(vaccGHIJ); in xnn_qc8_dwconv_minmax_fp32_ukernel_up24x9__neonv8_mul16() local 239 vfpaccGHIJ = vmulq_f32(vfpaccGHIJ, vscaleGHIJ); in xnn_qc8_dwconv_minmax_fp32_ukernel_up24x9__neonv8_mul16() 246 vaccGHIJ = vcvtnq_s32_f32(vfpaccGHIJ); in xnn_qc8_dwconv_minmax_fp32_ukernel_up24x9__neonv8_mul16()
|
D | up24x9-minmax-fp32-neon-mul16.c | 225 float32x4_t vfpaccGHIJ = vcvtq_f32_s32(vaccGHIJ); in xnn_qc8_dwconv_minmax_fp32_ukernel_up24x9__neon_mul16() local 239 vfpaccGHIJ = vmulq_f32(vfpaccGHIJ, vscaleGHIJ); in xnn_qc8_dwconv_minmax_fp32_ukernel_up24x9__neon_mul16() 246 vaccGHIJ = vreinterpretq_s32_f32(vaddq_f32(vfpaccGHIJ, vmagic_bias)); in xnn_qc8_dwconv_minmax_fp32_ukernel_up24x9__neon_mul16()
|
/external/XNNPACK/src/qs8-dwconv/gen/ |
D | up24x9-minmax-fp32-neon-mul16.c | 226 float32x4_t vfpaccGHIJ = vcvtq_f32_s32(vaccGHIJ); in xnn_qs8_dwconv_minmax_fp32_ukernel_up24x9__neon_mul16() local 233 vfpaccGHIJ = vmulq_f32(vfpaccGHIJ, vscale); in xnn_qs8_dwconv_minmax_fp32_ukernel_up24x9__neon_mul16() 240 vaccGHIJ = vreinterpretq_s32_f32(vaddq_f32(vfpaccGHIJ, vmagic_bias)); in xnn_qs8_dwconv_minmax_fp32_ukernel_up24x9__neon_mul16()
|
D | up24x9-minmax-fp32-neonv8-mul16.c | 226 float32x4_t vfpaccGHIJ = vcvtq_f32_s32(vaccGHIJ); in xnn_qs8_dwconv_minmax_fp32_ukernel_up24x9__neonv8_mul16() local 233 vfpaccGHIJ = vmulq_f32(vfpaccGHIJ, vscale); in xnn_qs8_dwconv_minmax_fp32_ukernel_up24x9__neonv8_mul16() 240 vaccGHIJ = vcvtnq_s32_f32(vfpaccGHIJ); in xnn_qs8_dwconv_minmax_fp32_ukernel_up24x9__neonv8_mul16()
|
/external/XNNPACK/src/qu8-dwconv/gen/ |
D | up24x9-minmax-fp32-neon-mul16.c | 227 float32x4_t vfpaccGHIJ = vcvtq_f32_s32(vaccGHIJ); in xnn_qu8_dwconv_minmax_fp32_ukernel_up24x9__neon_mul16() local 234 vfpaccGHIJ = vmulq_f32(vfpaccGHIJ, vscale); in xnn_qu8_dwconv_minmax_fp32_ukernel_up24x9__neon_mul16() 241 vaccGHIJ = vreinterpretq_s32_f32(vaddq_f32(vfpaccGHIJ, vmagic_bias)); in xnn_qu8_dwconv_minmax_fp32_ukernel_up24x9__neon_mul16()
|