Searched refs:vy_zero_point (Results 1 – 8 of 8) sorted by relevance
/external/XNNPACK/src/qu8-vadd/ |
D | minmax-neon.c | 24 const int16x8_t vy_zero_point = vld1q_dup_s16(¶ms->neon.y_zero_point); in xnn_qu8_vadd_minmax_ukernel__neon() local 87 …const int16x8_t vacc0 = vqaddq_s16(vqmovn_high_s32(vqmovn_s32(vacc0_lo), vacc0_hi), vy_zero_point); in xnn_qu8_vadd_minmax_ukernel__neon() 88 …const int16x8_t vacc1 = vqaddq_s16(vqmovn_high_s32(vqmovn_s32(vacc1_lo), vacc1_hi), vy_zero_point); in xnn_qu8_vadd_minmax_ukernel__neon() 89 …const int16x8_t vacc2 = vqaddq_s16(vqmovn_high_s32(vqmovn_s32(vacc2_lo), vacc2_hi), vy_zero_point); in xnn_qu8_vadd_minmax_ukernel__neon() 90 …const int16x8_t vacc3 = vqaddq_s16(vqmovn_high_s32(vqmovn_s32(vacc3_lo), vacc3_hi), vy_zero_point); in xnn_qu8_vadd_minmax_ukernel__neon() 140 …16x8_t vacc0 = vqaddq_s16(vcombine_s16(vqmovn_s32(vacc0_lo), vqmovn_s32(vacc0_hi)), vy_zero_point); in xnn_qu8_vadd_minmax_ukernel__neon() 141 …16x8_t vacc1 = vqaddq_s16(vcombine_s16(vqmovn_s32(vacc1_lo), vqmovn_s32(vacc1_hi)), vy_zero_point); in xnn_qu8_vadd_minmax_ukernel__neon() 182 const int16x8_t vacc = vqaddq_s16(vqmovn_high_s32(vqmovn_s32(vacc_lo), vacc_hi), vy_zero_point); in xnn_qu8_vadd_minmax_ukernel__neon() 184 …int16x8_t vacc = vqaddq_s16(vcombine_s16(vqmovn_s32(vacc_lo), vqmovn_s32(vacc_hi)), vy_zero_point); in xnn_qu8_vadd_minmax_ukernel__neon() 225 const int16x8_t vacc = vqaddq_s16(vqmovn_high_s32(vqmovn_s32(vacc_lo), vacc_hi), vy_zero_point); in xnn_qu8_vadd_minmax_ukernel__neon() [all …]
|
D | minmax-sse2.c | 68 const __m128i vy_zero_point = _mm_load_si128((const __m128i*) params->sse2.y_zero_point); in xnn_qu8_vadd_minmax_ukernel__sse2() local 69 const __m128i vacc = _mm_adds_epi16(_mm_packs_epi32(vacc_lo, vacc_hi), vy_zero_point); in xnn_qu8_vadd_minmax_ukernel__sse2() 110 const __m128i vy_zero_point = _mm_load_si128((const __m128i*) params->sse2.y_zero_point); in xnn_qu8_vadd_minmax_ukernel__sse2() local 111 const __m128i vacc = _mm_adds_epi16(_mm_packs_epi32(vacc_lo, vacc_hi), vy_zero_point); in xnn_qu8_vadd_minmax_ukernel__sse2()
|
D | minmax-scalar.c | 28 const int32_t vy_zero_point = params->scalar.y_zero_point; in xnn_qu8_vadd_minmax_ukernel__scalar() local 48 vy += vy_zero_point; in xnn_qu8_vadd_minmax_ukernel__scalar()
|
/external/XNNPACK/src/qs8-vadd/gen/ |
D | minmax-neon-ld64-x32.c | 25 const int8x8_t vy_zero_point = vld1_dup_s8(¶ms->neon.y_zero_point); in xnn_qs8_vadd_minmax_ukernel__neon_ld64_x32() local 45 const int16x8_t vey01234567 = vsubl_s8(vy01234567, vy_zero_point); in xnn_qs8_vadd_minmax_ukernel__neon_ld64_x32() 47 const int16x8_t vey89ABCDEF = vsubl_s8(vy89ABCDEF, vy_zero_point); in xnn_qs8_vadd_minmax_ukernel__neon_ld64_x32() 49 const int16x8_t veyGHIJKLMN = vsubl_s8(vyGHIJKLMN, vy_zero_point); in xnn_qs8_vadd_minmax_ukernel__neon_ld64_x32() 51 const int16x8_t veyOPQRSTUV = vsubl_s8(vyOPQRSTUV, vy_zero_point); in xnn_qs8_vadd_minmax_ukernel__neon_ld64_x32() 112 const int16x8_t vey01234567 = vsubl_s8(vy01234567, vy_zero_point); in xnn_qs8_vadd_minmax_ukernel__neon_ld64_x32()
|
D | minmax-neon-ld64-x24.c | 25 const int8x8_t vy_zero_point = vld1_dup_s8(¶ms->neon.y_zero_point); in xnn_qs8_vadd_minmax_ukernel__neon_ld64_x24() local 43 const int16x8_t vey01234567 = vsubl_s8(vy01234567, vy_zero_point); in xnn_qs8_vadd_minmax_ukernel__neon_ld64_x24() 45 const int16x8_t vey89ABCDEF = vsubl_s8(vy89ABCDEF, vy_zero_point); in xnn_qs8_vadd_minmax_ukernel__neon_ld64_x24() 47 const int16x8_t veyGHIJKLMN = vsubl_s8(vyGHIJKLMN, vy_zero_point); in xnn_qs8_vadd_minmax_ukernel__neon_ld64_x24() 99 const int16x8_t vey01234567 = vsubl_s8(vy01234567, vy_zero_point); in xnn_qs8_vadd_minmax_ukernel__neon_ld64_x24()
|
D | minmax-neon-ld64-x16.c | 25 const int8x8_t vy_zero_point = vld1_dup_s8(¶ms->neon.y_zero_point); in xnn_qs8_vadd_minmax_ukernel__neon_ld64_x16() local 41 const int16x8_t vey01234567 = vsubl_s8(vy01234567, vy_zero_point); in xnn_qs8_vadd_minmax_ukernel__neon_ld64_x16() 43 const int16x8_t vey89ABCDEF = vsubl_s8(vy89ABCDEF, vy_zero_point); in xnn_qs8_vadd_minmax_ukernel__neon_ld64_x16() 82 const int16x8_t vey01234567 = vsubl_s8(vy01234567, vy_zero_point); in xnn_qs8_vadd_minmax_ukernel__neon_ld64_x16()
|
D | minmax-neon-ld64-x8.c | 25 const int8x8_t vy_zero_point = vld1_dup_s8(¶ms->neon.y_zero_point); in xnn_qs8_vadd_minmax_ukernel__neon_ld64_x8() local 39 const int16x8_t vey01234567 = vsubl_s8(vy01234567, vy_zero_point); in xnn_qs8_vadd_minmax_ukernel__neon_ld64_x8() 69 const int16x8_t vey01234567 = vsubl_s8(vy01234567, vy_zero_point); in xnn_qs8_vadd_minmax_ukernel__neon_ld64_x8()
|
/external/XNNPACK/src/qs8-vadd/ |
D | neon-ld64.c.in | 24 const int8x8_t vy_zero_point = vld1_dup_s8(¶ms->neon.y_zero_point); variable 40 const int16x8_t vey${ABC[N:N+8]} = vsubl_s8(vy${ABC[N:N+8]}, vy_zero_point); 93 const int16x8_t vey${ABC[0:8]} = vsubl_s8(vy${ABC[0:8]}, vy_zero_point);
|