| /external/XNNPACK/src/qu8-gemm/gen/ |
| D | 1x8c4-minmax-rndnu-neondot.c | 49 uint32x4_t vpacc0x0123 = vld1q_u32(w); w = (const void*) ((const uint32_t*) w + 4); in xnn_qu8_gemm_minmax_rndnu_ukernel_1x8c4__neondot() local
68 vpacc0x0123 = vdotq_lane_u32(vpacc0x0123, vb0123x0123, va0x01234567, 0); in xnn_qu8_gemm_minmax_rndnu_ukernel_1x8c4__neondot()
70 vpacc0x0123 = vdotq_lane_u32(vpacc0x0123, vb4567x0123, va0x01234567, 1); in xnn_qu8_gemm_minmax_rndnu_ukernel_1x8c4__neondot()
86 vpacc0x0123 = vdotq_lane_u32(vpacc0x0123, vb0123x0123, va0x01234567, 0); in xnn_qu8_gemm_minmax_rndnu_ukernel_1x8c4__neondot()
93 int32x4_t vacc0x0123 = vreinterpretq_s32_u32(vsubq_u32(vpacc0x0123, vnacc0x0123)); in xnn_qu8_gemm_minmax_rndnu_ukernel_1x8c4__neondot()
|
| D | 2x8c4-minmax-rndnu-neondot.c | 55 uint32x4_t vpacc0x0123 = vld1q_u32(w); w = (const void*) ((const uint32_t*) w + 4); in xnn_qu8_gemm_minmax_rndnu_ukernel_2x8c4__neondot() local
57 uint32x4_t vpacc1x0123 = vpacc0x0123; in xnn_qu8_gemm_minmax_rndnu_ukernel_2x8c4__neondot()
78 vpacc0x0123 = vdotq_lane_u32(vpacc0x0123, vb0123x0123, va0x01234567, 0); in xnn_qu8_gemm_minmax_rndnu_ukernel_2x8c4__neondot()
80 vpacc0x0123 = vdotq_lane_u32(vpacc0x0123, vb4567x0123, va0x01234567, 1); in xnn_qu8_gemm_minmax_rndnu_ukernel_2x8c4__neondot()
102 vpacc0x0123 = vdotq_lane_u32(vpacc0x0123, vb0123x0123, va0x01234567, 0); in xnn_qu8_gemm_minmax_rndnu_ukernel_2x8c4__neondot()
112 int32x4_t vacc0x0123 = vreinterpretq_s32_u32(vsubq_u32(vpacc0x0123, vnacc0x0123)); in xnn_qu8_gemm_minmax_rndnu_ukernel_2x8c4__neondot()
|
| D | 1x16c4-minmax-rndnu-neondot.c | 49 uint32x4_t vpacc0x0123 = vld1q_u32(w); w = (const void*) ((const uint32_t*) w + 4); in xnn_qu8_gemm_minmax_rndnu_ukernel_1x16c4__neondot() local
74 vpacc0x0123 = vdotq_lane_u32(vpacc0x0123, vb0123x0123, va0x01234567, 0); in xnn_qu8_gemm_minmax_rndnu_ukernel_1x16c4__neondot()
78 vpacc0x0123 = vdotq_lane_u32(vpacc0x0123, vb4567x0123, va0x01234567, 1); in xnn_qu8_gemm_minmax_rndnu_ukernel_1x16c4__neondot()
98 vpacc0x0123 = vdotq_lane_u32(vpacc0x0123, vb0123x0123, va0x01234567, 0); in xnn_qu8_gemm_minmax_rndnu_ukernel_1x16c4__neondot()
107 int32x4_t vacc0x0123 = vreinterpretq_s32_u32(vsubq_u32(vpacc0x0123, vnacc0x0123)); in xnn_qu8_gemm_minmax_rndnu_ukernel_1x16c4__neondot()
|
| D | 1x16c4-minmax-fp32-neondot.c | 50 uint32x4_t vpacc0x0123 = vld1q_u32(w); w = (const void*) ((const uint32_t*) w + 4); in xnn_qu8_gemm_minmax_fp32_ukernel_1x16c4__neondot() local
75 vpacc0x0123 = vdotq_lane_u32(vpacc0x0123, vb0123x0123, va0x01234567, 0); in xnn_qu8_gemm_minmax_fp32_ukernel_1x16c4__neondot()
79 vpacc0x0123 = vdotq_lane_u32(vpacc0x0123, vb4567x0123, va0x01234567, 1); in xnn_qu8_gemm_minmax_fp32_ukernel_1x16c4__neondot()
99 vpacc0x0123 = vdotq_lane_u32(vpacc0x0123, vb0123x0123, va0x01234567, 0); in xnn_qu8_gemm_minmax_fp32_ukernel_1x16c4__neondot()
108 int32x4_t vacc0x0123 = vreinterpretq_s32_u32(vsubq_u32(vpacc0x0123, vnacc0x0123)); in xnn_qu8_gemm_minmax_fp32_ukernel_1x16c4__neondot()
|
| D | 3x8c4-minmax-rndnu-neondot.c | 61 uint32x4_t vpacc0x0123 = vld1q_u32(w); w = (const void*) ((const uint32_t*) w + 4); in xnn_qu8_gemm_minmax_rndnu_ukernel_3x8c4__neondot() local
63 uint32x4_t vpacc1x0123 = vpacc0x0123; in xnn_qu8_gemm_minmax_rndnu_ukernel_3x8c4__neondot()
65 uint32x4_t vpacc2x0123 = vpacc0x0123; in xnn_qu8_gemm_minmax_rndnu_ukernel_3x8c4__neondot()
88 vpacc0x0123 = vdotq_lane_u32(vpacc0x0123, vb0123x0123, va0x01234567, 0); in xnn_qu8_gemm_minmax_rndnu_ukernel_3x8c4__neondot()
90 vpacc0x0123 = vdotq_lane_u32(vpacc0x0123, vb4567x0123, va0x01234567, 1); in xnn_qu8_gemm_minmax_rndnu_ukernel_3x8c4__neondot()
118 vpacc0x0123 = vdotq_lane_u32(vpacc0x0123, vb0123x0123, va0x01234567, 0); in xnn_qu8_gemm_minmax_rndnu_ukernel_3x8c4__neondot()
131 int32x4_t vacc0x0123 = vreinterpretq_s32_u32(vsubq_u32(vpacc0x0123, vnacc0x0123)); in xnn_qu8_gemm_minmax_rndnu_ukernel_3x8c4__neondot()
|
| D | 4x8c4-minmax-rndnu-neondot.c | 67 uint32x4_t vpacc0x0123 = vld1q_u32(w); w = (const void*) ((const uint32_t*) w + 4); in xnn_qu8_gemm_minmax_rndnu_ukernel_4x8c4__neondot() local
69 uint32x4_t vpacc1x0123 = vpacc0x0123; in xnn_qu8_gemm_minmax_rndnu_ukernel_4x8c4__neondot()
71 uint32x4_t vpacc2x0123 = vpacc0x0123; in xnn_qu8_gemm_minmax_rndnu_ukernel_4x8c4__neondot()
73 uint32x4_t vpacc3x0123 = vpacc0x0123; in xnn_qu8_gemm_minmax_rndnu_ukernel_4x8c4__neondot()
98 vpacc0x0123 = vdotq_lane_u32(vpacc0x0123, vb0123x0123, va0x01234567, 0); in xnn_qu8_gemm_minmax_rndnu_ukernel_4x8c4__neondot()
100 vpacc0x0123 = vdotq_lane_u32(vpacc0x0123, vb4567x0123, va0x01234567, 1); in xnn_qu8_gemm_minmax_rndnu_ukernel_4x8c4__neondot()
134 vpacc0x0123 = vdotq_lane_u32(vpacc0x0123, vb0123x0123, va0x01234567, 0); in xnn_qu8_gemm_minmax_rndnu_ukernel_4x8c4__neondot()
150 int32x4_t vacc0x0123 = vreinterpretq_s32_u32(vsubq_u32(vpacc0x0123, vnacc0x0123)); in xnn_qu8_gemm_minmax_rndnu_ukernel_4x8c4__neondot()
|
| D | 5x8c4-minmax-rndnu-neondot.c | 73 uint32x4_t vpacc0x0123 = vld1q_u32(w); w = (const void*) ((const uint32_t*) w + 4); in xnn_qu8_gemm_minmax_rndnu_ukernel_5x8c4__neondot() local
75 uint32x4_t vpacc1x0123 = vpacc0x0123; in xnn_qu8_gemm_minmax_rndnu_ukernel_5x8c4__neondot()
77 uint32x4_t vpacc2x0123 = vpacc0x0123; in xnn_qu8_gemm_minmax_rndnu_ukernel_5x8c4__neondot()
79 uint32x4_t vpacc3x0123 = vpacc0x0123; in xnn_qu8_gemm_minmax_rndnu_ukernel_5x8c4__neondot()
81 uint32x4_t vpacc4x0123 = vpacc0x0123; in xnn_qu8_gemm_minmax_rndnu_ukernel_5x8c4__neondot()
108 vpacc0x0123 = vdotq_lane_u32(vpacc0x0123, vb0123x0123, va0x01234567, 0); in xnn_qu8_gemm_minmax_rndnu_ukernel_5x8c4__neondot()
110 vpacc0x0123 = vdotq_lane_u32(vpacc0x0123, vb4567x0123, va0x01234567, 1); in xnn_qu8_gemm_minmax_rndnu_ukernel_5x8c4__neondot()
150 vpacc0x0123 = vdotq_lane_u32(vpacc0x0123, vb0123x0123, va0x01234567, 0); in xnn_qu8_gemm_minmax_rndnu_ukernel_5x8c4__neondot()
169 int32x4_t vacc0x0123 = vreinterpretq_s32_u32(vsubq_u32(vpacc0x0123, vnacc0x0123)); in xnn_qu8_gemm_minmax_rndnu_ukernel_5x8c4__neondot()
|
| D | 2x16c4-minmax-fp32-neondot.c | 56 uint32x4_t vpacc0x0123 = vld1q_u32(w); w = (const void*) ((const uint32_t*) w + 4); in xnn_qu8_gemm_minmax_fp32_ukernel_2x16c4__neondot() local
60 uint32x4_t vpacc1x0123 = vpacc0x0123; in xnn_qu8_gemm_minmax_fp32_ukernel_2x16c4__neondot()
87 vpacc0x0123 = vdotq_lane_u32(vpacc0x0123, vb0123x0123, va0x01234567, 0); in xnn_qu8_gemm_minmax_fp32_ukernel_2x16c4__neondot()
91 vpacc0x0123 = vdotq_lane_u32(vpacc0x0123, vb4567x0123, va0x01234567, 1); in xnn_qu8_gemm_minmax_fp32_ukernel_2x16c4__neondot()
121 vpacc0x0123 = vdotq_lane_u32(vpacc0x0123, vb0123x0123, va0x01234567, 0); in xnn_qu8_gemm_minmax_fp32_ukernel_2x16c4__neondot()
135 int32x4_t vacc0x0123 = vreinterpretq_s32_u32(vsubq_u32(vpacc0x0123, vnacc0x0123)); in xnn_qu8_gemm_minmax_fp32_ukernel_2x16c4__neondot()
|
| D | 2x16c4-minmax-rndnu-neondot.c | 55 uint32x4_t vpacc0x0123 = vld1q_u32(w); w = (const void*) ((const uint32_t*) w + 4); in xnn_qu8_gemm_minmax_rndnu_ukernel_2x16c4__neondot() local
59 uint32x4_t vpacc1x0123 = vpacc0x0123; in xnn_qu8_gemm_minmax_rndnu_ukernel_2x16c4__neondot()
86 vpacc0x0123 = vdotq_lane_u32(vpacc0x0123, vb0123x0123, va0x01234567, 0); in xnn_qu8_gemm_minmax_rndnu_ukernel_2x16c4__neondot()
90 vpacc0x0123 = vdotq_lane_u32(vpacc0x0123, vb4567x0123, va0x01234567, 1); in xnn_qu8_gemm_minmax_rndnu_ukernel_2x16c4__neondot()
120 vpacc0x0123 = vdotq_lane_u32(vpacc0x0123, vb0123x0123, va0x01234567, 0); in xnn_qu8_gemm_minmax_rndnu_ukernel_2x16c4__neondot()
134 int32x4_t vacc0x0123 = vreinterpretq_s32_u32(vsubq_u32(vpacc0x0123, vnacc0x0123)); in xnn_qu8_gemm_minmax_rndnu_ukernel_2x16c4__neondot()
|
| D | 6x8c4-minmax-rndnu-neondot.c | 79 uint32x4_t vpacc0x0123 = vld1q_u32(w); w = (const void*) ((const uint32_t*) w + 4); in xnn_qu8_gemm_minmax_rndnu_ukernel_6x8c4__neondot() local
81 uint32x4_t vpacc1x0123 = vpacc0x0123; in xnn_qu8_gemm_minmax_rndnu_ukernel_6x8c4__neondot()
83 uint32x4_t vpacc2x0123 = vpacc0x0123; in xnn_qu8_gemm_minmax_rndnu_ukernel_6x8c4__neondot()
85 uint32x4_t vpacc3x0123 = vpacc0x0123; in xnn_qu8_gemm_minmax_rndnu_ukernel_6x8c4__neondot()
87 uint32x4_t vpacc4x0123 = vpacc0x0123; in xnn_qu8_gemm_minmax_rndnu_ukernel_6x8c4__neondot()
89 uint32x4_t vpacc5x0123 = vpacc0x0123; in xnn_qu8_gemm_minmax_rndnu_ukernel_6x8c4__neondot()
118 vpacc0x0123 = vdotq_lane_u32(vpacc0x0123, vb0123x0123, va0x01234567, 0); in xnn_qu8_gemm_minmax_rndnu_ukernel_6x8c4__neondot()
120 vpacc0x0123 = vdotq_lane_u32(vpacc0x0123, vb4567x0123, va0x01234567, 1); in xnn_qu8_gemm_minmax_rndnu_ukernel_6x8c4__neondot()
166 vpacc0x0123 = vdotq_lane_u32(vpacc0x0123, vb0123x0123, va0x01234567, 0); in xnn_qu8_gemm_minmax_rndnu_ukernel_6x8c4__neondot()
188 int32x4_t vacc0x0123 = vreinterpretq_s32_u32(vsubq_u32(vpacc0x0123, vnacc0x0123)); in xnn_qu8_gemm_minmax_rndnu_ukernel_6x8c4__neondot()
|
| D | 8x8c4-minmax-rndnu-neondot.c | 91 uint32x4_t vpacc0x0123 = vld1q_u32(w); w = (const void*) ((const uint32_t*) w + 4); in xnn_qu8_gemm_minmax_rndnu_ukernel_8x8c4__neondot() local
93 uint32x4_t vpacc1x0123 = vpacc0x0123; in xnn_qu8_gemm_minmax_rndnu_ukernel_8x8c4__neondot()
95 uint32x4_t vpacc2x0123 = vpacc0x0123; in xnn_qu8_gemm_minmax_rndnu_ukernel_8x8c4__neondot()
97 uint32x4_t vpacc3x0123 = vpacc0x0123; in xnn_qu8_gemm_minmax_rndnu_ukernel_8x8c4__neondot()
99 uint32x4_t vpacc4x0123 = vpacc0x0123; in xnn_qu8_gemm_minmax_rndnu_ukernel_8x8c4__neondot()
101 uint32x4_t vpacc5x0123 = vpacc0x0123; in xnn_qu8_gemm_minmax_rndnu_ukernel_8x8c4__neondot()
103 uint32x4_t vpacc6x0123 = vpacc0x0123; in xnn_qu8_gemm_minmax_rndnu_ukernel_8x8c4__neondot()
105 uint32x4_t vpacc7x0123 = vpacc0x0123; in xnn_qu8_gemm_minmax_rndnu_ukernel_8x8c4__neondot()
138 vpacc0x0123 = vdotq_lane_u32(vpacc0x0123, vb0123x0123, va0x01234567, 0); in xnn_qu8_gemm_minmax_rndnu_ukernel_8x8c4__neondot()
140 vpacc0x0123 = vdotq_lane_u32(vpacc0x0123, vb4567x0123, va0x01234567, 1); in xnn_qu8_gemm_minmax_rndnu_ukernel_8x8c4__neondot()
[all …]
|
| D | 3x16c4-minmax-rndnu-neondot.c | 61 uint32x4_t vpacc0x0123 = vld1q_u32(w); w = (const void*) ((const uint32_t*) w + 4); in xnn_qu8_gemm_minmax_rndnu_ukernel_3x16c4__neondot() local
65 uint32x4_t vpacc1x0123 = vpacc0x0123; in xnn_qu8_gemm_minmax_rndnu_ukernel_3x16c4__neondot()
69 uint32x4_t vpacc2x0123 = vpacc0x0123; in xnn_qu8_gemm_minmax_rndnu_ukernel_3x16c4__neondot()
98 vpacc0x0123 = vdotq_lane_u32(vpacc0x0123, vb0123x0123, va0x01234567, 0); in xnn_qu8_gemm_minmax_rndnu_ukernel_3x16c4__neondot()
102 vpacc0x0123 = vdotq_lane_u32(vpacc0x0123, vb4567x0123, va0x01234567, 1); in xnn_qu8_gemm_minmax_rndnu_ukernel_3x16c4__neondot()
142 vpacc0x0123 = vdotq_lane_u32(vpacc0x0123, vb0123x0123, va0x01234567, 0); in xnn_qu8_gemm_minmax_rndnu_ukernel_3x16c4__neondot()
161 int32x4_t vacc0x0123 = vreinterpretq_s32_u32(vsubq_u32(vpacc0x0123, vnacc0x0123)); in xnn_qu8_gemm_minmax_rndnu_ukernel_3x16c4__neondot()
|
| /external/XNNPACK/src/qu8-igemm/gen/ |
| D | 1x8c4-minmax-rndnu-neondot.c | 51 uint32x4_t vpacc0x0123 = vld1q_u32(w); w = (const void*) ((const uint32_t*) w + 4); in xnn_qu8_igemm_minmax_rndnu_ukernel_1x8c4__neondot() local
78 vpacc0x0123 = vdotq_lane_u32(vpacc0x0123, vb0123x0123, va0x01234567, 0); in xnn_qu8_igemm_minmax_rndnu_ukernel_1x8c4__neondot()
80 vpacc0x0123 = vdotq_lane_u32(vpacc0x0123, vb4567x0123, va0x01234567, 1); in xnn_qu8_igemm_minmax_rndnu_ukernel_1x8c4__neondot()
96 vpacc0x0123 = vdotq_lane_u32(vpacc0x0123, vb0123x0123, va0x01234567, 0); in xnn_qu8_igemm_minmax_rndnu_ukernel_1x8c4__neondot()
105 int32x4_t vacc0x0123 = vreinterpretq_s32_u32(vsubq_u32(vpacc0x0123, vnacc0x0123)); in xnn_qu8_igemm_minmax_rndnu_ukernel_1x8c4__neondot()
|
| D | 2x8c4-minmax-rndnu-neondot.c | 55 uint32x4_t vpacc0x0123 = vld1q_u32(w); w = (const void*) ((const uint32_t*) w + 4); in xnn_qu8_igemm_minmax_rndnu_ukernel_2x8c4__neondot() local
57 uint32x4_t vpacc1x0123 = vpacc0x0123; in xnn_qu8_igemm_minmax_rndnu_ukernel_2x8c4__neondot()
90 vpacc0x0123 = vdotq_lane_u32(vpacc0x0123, vb0123x0123, va0x01234567, 0); in xnn_qu8_igemm_minmax_rndnu_ukernel_2x8c4__neondot()
92 vpacc0x0123 = vdotq_lane_u32(vpacc0x0123, vb4567x0123, va0x01234567, 1); in xnn_qu8_igemm_minmax_rndnu_ukernel_2x8c4__neondot()
114 vpacc0x0123 = vdotq_lane_u32(vpacc0x0123, vb0123x0123, va0x01234567, 0); in xnn_qu8_igemm_minmax_rndnu_ukernel_2x8c4__neondot()
126 int32x4_t vacc0x0123 = vreinterpretq_s32_u32(vsubq_u32(vpacc0x0123, vnacc0x0123)); in xnn_qu8_igemm_minmax_rndnu_ukernel_2x8c4__neondot()
|
| D | 1x16c4-minmax-fp32-neondot.c | 52 uint32x4_t vpacc0x0123 = vld1q_u32(w); w = (const void*) ((const uint32_t*) w + 4); in xnn_qu8_igemm_minmax_fp32_ukernel_1x16c4__neondot() local
85 vpacc0x0123 = vdotq_lane_u32(vpacc0x0123, vb0123x0123, va0x01234567, 0); in xnn_qu8_igemm_minmax_fp32_ukernel_1x16c4__neondot()
89 vpacc0x0123 = vdotq_lane_u32(vpacc0x0123, vb4567x0123, va0x01234567, 1); in xnn_qu8_igemm_minmax_fp32_ukernel_1x16c4__neondot()
109 vpacc0x0123 = vdotq_lane_u32(vpacc0x0123, vb0123x0123, va0x01234567, 0); in xnn_qu8_igemm_minmax_fp32_ukernel_1x16c4__neondot()
120 int32x4_t vacc0x0123 = vreinterpretq_s32_u32(vsubq_u32(vpacc0x0123, vnacc0x0123)); in xnn_qu8_igemm_minmax_fp32_ukernel_1x16c4__neondot()
|
| D | 3x8c4-minmax-rndnu-neondot.c | 59 uint32x4_t vpacc0x0123 = vld1q_u32(w); w = (const void*) ((const uint32_t*) w + 4); in xnn_qu8_igemm_minmax_rndnu_ukernel_3x8c4__neondot() local
61 uint32x4_t vpacc1x0123 = vpacc0x0123; in xnn_qu8_igemm_minmax_rndnu_ukernel_3x8c4__neondot()
63 uint32x4_t vpacc2x0123 = vpacc0x0123; in xnn_qu8_igemm_minmax_rndnu_ukernel_3x8c4__neondot()
102 vpacc0x0123 = vdotq_lane_u32(vpacc0x0123, vb0123x0123, va0x01234567, 0); in xnn_qu8_igemm_minmax_rndnu_ukernel_3x8c4__neondot()
104 vpacc0x0123 = vdotq_lane_u32(vpacc0x0123, vb4567x0123, va0x01234567, 1); in xnn_qu8_igemm_minmax_rndnu_ukernel_3x8c4__neondot()
132 vpacc0x0123 = vdotq_lane_u32(vpacc0x0123, vb0123x0123, va0x01234567, 0); in xnn_qu8_igemm_minmax_rndnu_ukernel_3x8c4__neondot()
147 int32x4_t vacc0x0123 = vreinterpretq_s32_u32(vsubq_u32(vpacc0x0123, vnacc0x0123)); in xnn_qu8_igemm_minmax_rndnu_ukernel_3x8c4__neondot()
|
| D | 1x16c4-minmax-rndnu-neondot.c | 51 uint32x4_t vpacc0x0123 = vld1q_u32(w); w = (const void*) ((const uint32_t*) w + 4); in xnn_qu8_igemm_minmax_rndnu_ukernel_1x16c4__neondot() local
84 vpacc0x0123 = vdotq_lane_u32(vpacc0x0123, vb0123x0123, va0x01234567, 0); in xnn_qu8_igemm_minmax_rndnu_ukernel_1x16c4__neondot()
88 vpacc0x0123 = vdotq_lane_u32(vpacc0x0123, vb4567x0123, va0x01234567, 1); in xnn_qu8_igemm_minmax_rndnu_ukernel_1x16c4__neondot()
108 vpacc0x0123 = vdotq_lane_u32(vpacc0x0123, vb0123x0123, va0x01234567, 0); in xnn_qu8_igemm_minmax_rndnu_ukernel_1x16c4__neondot()
119 int32x4_t vacc0x0123 = vreinterpretq_s32_u32(vsubq_u32(vpacc0x0123, vnacc0x0123)); in xnn_qu8_igemm_minmax_rndnu_ukernel_1x16c4__neondot()
|
| D | 4x8c4-minmax-rndnu-neondot.c | 63 uint32x4_t vpacc0x0123 = vld1q_u32(w); w = (const void*) ((const uint32_t*) w + 4); in xnn_qu8_igemm_minmax_rndnu_ukernel_4x8c4__neondot() local
65 uint32x4_t vpacc1x0123 = vpacc0x0123; in xnn_qu8_igemm_minmax_rndnu_ukernel_4x8c4__neondot()
67 uint32x4_t vpacc2x0123 = vpacc0x0123; in xnn_qu8_igemm_minmax_rndnu_ukernel_4x8c4__neondot()
69 uint32x4_t vpacc3x0123 = vpacc0x0123; in xnn_qu8_igemm_minmax_rndnu_ukernel_4x8c4__neondot()
114 vpacc0x0123 = vdotq_lane_u32(vpacc0x0123, vb0123x0123, va0x01234567, 0); in xnn_qu8_igemm_minmax_rndnu_ukernel_4x8c4__neondot()
116 vpacc0x0123 = vdotq_lane_u32(vpacc0x0123, vb4567x0123, va0x01234567, 1); in xnn_qu8_igemm_minmax_rndnu_ukernel_4x8c4__neondot()
150 vpacc0x0123 = vdotq_lane_u32(vpacc0x0123, vb0123x0123, va0x01234567, 0); in xnn_qu8_igemm_minmax_rndnu_ukernel_4x8c4__neondot()
168 int32x4_t vacc0x0123 = vreinterpretq_s32_u32(vsubq_u32(vpacc0x0123, vnacc0x0123)); in xnn_qu8_igemm_minmax_rndnu_ukernel_4x8c4__neondot()
|
| D | 5x8c4-minmax-rndnu-neondot.c | 67 uint32x4_t vpacc0x0123 = vld1q_u32(w); w = (const void*) ((const uint32_t*) w + 4); in xnn_qu8_igemm_minmax_rndnu_ukernel_5x8c4__neondot() local
69 uint32x4_t vpacc1x0123 = vpacc0x0123; in xnn_qu8_igemm_minmax_rndnu_ukernel_5x8c4__neondot()
71 uint32x4_t vpacc2x0123 = vpacc0x0123; in xnn_qu8_igemm_minmax_rndnu_ukernel_5x8c4__neondot()
73 uint32x4_t vpacc3x0123 = vpacc0x0123; in xnn_qu8_igemm_minmax_rndnu_ukernel_5x8c4__neondot()
75 uint32x4_t vpacc4x0123 = vpacc0x0123; in xnn_qu8_igemm_minmax_rndnu_ukernel_5x8c4__neondot()
126 vpacc0x0123 = vdotq_lane_u32(vpacc0x0123, vb0123x0123, va0x01234567, 0); in xnn_qu8_igemm_minmax_rndnu_ukernel_5x8c4__neondot()
128 vpacc0x0123 = vdotq_lane_u32(vpacc0x0123, vb4567x0123, va0x01234567, 1); in xnn_qu8_igemm_minmax_rndnu_ukernel_5x8c4__neondot()
168 vpacc0x0123 = vdotq_lane_u32(vpacc0x0123, vb0123x0123, va0x01234567, 0); in xnn_qu8_igemm_minmax_rndnu_ukernel_5x8c4__neondot()
189 int32x4_t vacc0x0123 = vreinterpretq_s32_u32(vsubq_u32(vpacc0x0123, vnacc0x0123)); in xnn_qu8_igemm_minmax_rndnu_ukernel_5x8c4__neondot()
|
| D | 2x16c4-minmax-fp32-neondot.c | 56 uint32x4_t vpacc0x0123 = vld1q_u32(w); w = (const void*) ((const uint32_t*) w + 4); in xnn_qu8_igemm_minmax_fp32_ukernel_2x16c4__neondot() local
60 uint32x4_t vpacc1x0123 = vpacc0x0123; in xnn_qu8_igemm_minmax_fp32_ukernel_2x16c4__neondot()
99 vpacc0x0123 = vdotq_lane_u32(vpacc0x0123, vb0123x0123, va0x01234567, 0); in xnn_qu8_igemm_minmax_fp32_ukernel_2x16c4__neondot()
103 vpacc0x0123 = vdotq_lane_u32(vpacc0x0123, vb4567x0123, va0x01234567, 1); in xnn_qu8_igemm_minmax_fp32_ukernel_2x16c4__neondot()
133 vpacc0x0123 = vdotq_lane_u32(vpacc0x0123, vb0123x0123, va0x01234567, 0); in xnn_qu8_igemm_minmax_fp32_ukernel_2x16c4__neondot()
149 int32x4_t vacc0x0123 = vreinterpretq_s32_u32(vsubq_u32(vpacc0x0123, vnacc0x0123)); in xnn_qu8_igemm_minmax_fp32_ukernel_2x16c4__neondot()
|
| D | 6x8c4-minmax-rndnu-neondot.c | 71 uint32x4_t vpacc0x0123 = vld1q_u32(w); w = (const void*) ((const uint32_t*) w + 4); in xnn_qu8_igemm_minmax_rndnu_ukernel_6x8c4__neondot() local
73 uint32x4_t vpacc1x0123 = vpacc0x0123; in xnn_qu8_igemm_minmax_rndnu_ukernel_6x8c4__neondot()
75 uint32x4_t vpacc2x0123 = vpacc0x0123; in xnn_qu8_igemm_minmax_rndnu_ukernel_6x8c4__neondot()
77 uint32x4_t vpacc3x0123 = vpacc0x0123; in xnn_qu8_igemm_minmax_rndnu_ukernel_6x8c4__neondot()
79 uint32x4_t vpacc4x0123 = vpacc0x0123; in xnn_qu8_igemm_minmax_rndnu_ukernel_6x8c4__neondot()
81 uint32x4_t vpacc5x0123 = vpacc0x0123; in xnn_qu8_igemm_minmax_rndnu_ukernel_6x8c4__neondot()
138 vpacc0x0123 = vdotq_lane_u32(vpacc0x0123, vb0123x0123, va0x01234567, 0); in xnn_qu8_igemm_minmax_rndnu_ukernel_6x8c4__neondot()
140 vpacc0x0123 = vdotq_lane_u32(vpacc0x0123, vb4567x0123, va0x01234567, 1); in xnn_qu8_igemm_minmax_rndnu_ukernel_6x8c4__neondot()
186 vpacc0x0123 = vdotq_lane_u32(vpacc0x0123, vb0123x0123, va0x01234567, 0); in xnn_qu8_igemm_minmax_rndnu_ukernel_6x8c4__neondot()
210 int32x4_t vacc0x0123 = vreinterpretq_s32_u32(vsubq_u32(vpacc0x0123, vnacc0x0123)); in xnn_qu8_igemm_minmax_rndnu_ukernel_6x8c4__neondot()
|
| D | 2x16c4-minmax-rndnu-neondot.c | 55 uint32x4_t vpacc0x0123 = vld1q_u32(w); w = (const void*) ((const uint32_t*) w + 4); in xnn_qu8_igemm_minmax_rndnu_ukernel_2x16c4__neondot() local
59 uint32x4_t vpacc1x0123 = vpacc0x0123; in xnn_qu8_igemm_minmax_rndnu_ukernel_2x16c4__neondot()
98 vpacc0x0123 = vdotq_lane_u32(vpacc0x0123, vb0123x0123, va0x01234567, 0); in xnn_qu8_igemm_minmax_rndnu_ukernel_2x16c4__neondot()
102 vpacc0x0123 = vdotq_lane_u32(vpacc0x0123, vb4567x0123, va0x01234567, 1); in xnn_qu8_igemm_minmax_rndnu_ukernel_2x16c4__neondot()
132 vpacc0x0123 = vdotq_lane_u32(vpacc0x0123, vb0123x0123, va0x01234567, 0); in xnn_qu8_igemm_minmax_rndnu_ukernel_2x16c4__neondot()
148 int32x4_t vacc0x0123 = vreinterpretq_s32_u32(vsubq_u32(vpacc0x0123, vnacc0x0123)); in xnn_qu8_igemm_minmax_rndnu_ukernel_2x16c4__neondot()
|
| D | 8x8c4-minmax-rndnu-neondot.c | 79 uint32x4_t vpacc0x0123 = vld1q_u32(w); w = (const void*) ((const uint32_t*) w + 4); in xnn_qu8_igemm_minmax_rndnu_ukernel_8x8c4__neondot() local
81 uint32x4_t vpacc1x0123 = vpacc0x0123; in xnn_qu8_igemm_minmax_rndnu_ukernel_8x8c4__neondot()
83 uint32x4_t vpacc2x0123 = vpacc0x0123; in xnn_qu8_igemm_minmax_rndnu_ukernel_8x8c4__neondot()
85 uint32x4_t vpacc3x0123 = vpacc0x0123; in xnn_qu8_igemm_minmax_rndnu_ukernel_8x8c4__neondot()
87 uint32x4_t vpacc4x0123 = vpacc0x0123; in xnn_qu8_igemm_minmax_rndnu_ukernel_8x8c4__neondot()
89 uint32x4_t vpacc5x0123 = vpacc0x0123; in xnn_qu8_igemm_minmax_rndnu_ukernel_8x8c4__neondot()
91 uint32x4_t vpacc6x0123 = vpacc0x0123; in xnn_qu8_igemm_minmax_rndnu_ukernel_8x8c4__neondot()
93 uint32x4_t vpacc7x0123 = vpacc0x0123; in xnn_qu8_igemm_minmax_rndnu_ukernel_8x8c4__neondot()
162 vpacc0x0123 = vdotq_lane_u32(vpacc0x0123, vb0123x0123, va0x01234567, 0); in xnn_qu8_igemm_minmax_rndnu_ukernel_8x8c4__neondot()
164 vpacc0x0123 = vdotq_lane_u32(vpacc0x0123, vb4567x0123, va0x01234567, 1); in xnn_qu8_igemm_minmax_rndnu_ukernel_8x8c4__neondot()
[all …]
|
| D | 3x16c4-minmax-rndnu-neondot.c | 59 uint32x4_t vpacc0x0123 = vld1q_u32(w); w = (const void*) ((const uint32_t*) w + 4); in xnn_qu8_igemm_minmax_rndnu_ukernel_3x16c4__neondot() local
63 uint32x4_t vpacc1x0123 = vpacc0x0123; in xnn_qu8_igemm_minmax_rndnu_ukernel_3x16c4__neondot()
67 uint32x4_t vpacc2x0123 = vpacc0x0123; in xnn_qu8_igemm_minmax_rndnu_ukernel_3x16c4__neondot()
112 vpacc0x0123 = vdotq_lane_u32(vpacc0x0123, vb0123x0123, va0x01234567, 0); in xnn_qu8_igemm_minmax_rndnu_ukernel_3x16c4__neondot()
116 vpacc0x0123 = vdotq_lane_u32(vpacc0x0123, vb4567x0123, va0x01234567, 1); in xnn_qu8_igemm_minmax_rndnu_ukernel_3x16c4__neondot()
156 vpacc0x0123 = vdotq_lane_u32(vpacc0x0123, vb0123x0123, va0x01234567, 0); in xnn_qu8_igemm_minmax_rndnu_ukernel_3x16c4__neondot()
177 int32x4_t vacc0x0123 = vreinterpretq_s32_u32(vsubq_u32(vpacc0x0123, vnacc0x0123)); in xnn_qu8_igemm_minmax_rndnu_ukernel_3x16c4__neondot()
|
| D | 1x32c4-minmax-rndnu-neondot.c | 51 uint32x4_t vpacc0x0123 = vld1q_u32(w); w = (const void*) ((const uint32_t*) w + 4); in xnn_qu8_igemm_minmax_rndnu_ukernel_1x32c4__neondot() local
96 vpacc0x0123 = vdotq_lane_u32(vpacc0x0123, vb0123x0123, va0x01234567, 0); in xnn_qu8_igemm_minmax_rndnu_ukernel_1x32c4__neondot()
104 vpacc0x0123 = vdotq_lane_u32(vpacc0x0123, vb4567x0123, va0x01234567, 1); in xnn_qu8_igemm_minmax_rndnu_ukernel_1x32c4__neondot()
132 vpacc0x0123 = vdotq_lane_u32(vpacc0x0123, vb0123x0123, va0x01234567, 0); in xnn_qu8_igemm_minmax_rndnu_ukernel_1x32c4__neondot()
147 int32x4_t vacc0x0123 = vreinterpretq_s32_u32(vsubq_u32(vpacc0x0123, vnacc0x0123)); in xnn_qu8_igemm_minmax_rndnu_ukernel_1x32c4__neondot()
|