| /external/XNNPACK/src/f32-spmm/gen/ |
| D | 32x2-minmax-neonfma.c | 50 float32x4_t vacc0123n1 = vld1q_dup_f32(w); w += 1; in xnn_f32_spmm_minmax_ukernel_32x2__neonfma() local
51 float32x4_t vacc4567n1 = vacc0123n1; in xnn_f32_spmm_minmax_ukernel_32x2__neonfma()
52 float32x4_t vacc89ABn1 = vacc0123n1; in xnn_f32_spmm_minmax_ukernel_32x2__neonfma()
53 float32x4_t vaccCDEFn1 = vacc0123n1; in xnn_f32_spmm_minmax_ukernel_32x2__neonfma()
54 float32x4_t vaccGHIJn1 = vacc0123n1; in xnn_f32_spmm_minmax_ukernel_32x2__neonfma()
55 float32x4_t vaccKLMNn1 = vacc0123n1; in xnn_f32_spmm_minmax_ukernel_32x2__neonfma()
56 float32x4_t vaccOPQRn1 = vacc0123n1; in xnn_f32_spmm_minmax_ukernel_32x2__neonfma()
57 float32x4_t vaccSTUVn1 = vacc0123n1; in xnn_f32_spmm_minmax_ukernel_32x2__neonfma()
82 vacc0123n1 = vfmaq_lane_f32(vacc0123n1, vi0123, vw, 1); in xnn_f32_spmm_minmax_ukernel_32x2__neonfma()
100 float32x4_t vout0123n1 = vminq_f32(vacc0123n1, vmax); in xnn_f32_spmm_minmax_ukernel_32x2__neonfma()
[all …]
|
| D | 16x2-minmax-neonfma.c | 46 float32x4_t vacc0123n1 = vld1q_dup_f32(w); w += 1; in xnn_f32_spmm_minmax_ukernel_16x2__neonfma() local
47 float32x4_t vacc4567n1 = vacc0123n1; in xnn_f32_spmm_minmax_ukernel_16x2__neonfma()
48 float32x4_t vacc89ABn1 = vacc0123n1; in xnn_f32_spmm_minmax_ukernel_16x2__neonfma()
49 float32x4_t vaccCDEFn1 = vacc0123n1; in xnn_f32_spmm_minmax_ukernel_16x2__neonfma()
65 vacc0123n1 = vfmaq_lane_f32(vacc0123n1, vi0123, vw, 1); in xnn_f32_spmm_minmax_ukernel_16x2__neonfma()
75 float32x4_t vout0123n1 = vminq_f32(vacc0123n1, vmax); in xnn_f32_spmm_minmax_ukernel_16x2__neonfma()
160 float32x4_t vacc0123n1 = vld1q_dup_f32(w); w += 1; in xnn_f32_spmm_minmax_ukernel_16x2__neonfma() local
161 float32x4_t vacc4567n1 = vacc0123n1; in xnn_f32_spmm_minmax_ukernel_16x2__neonfma()
172 vacc0123n1 = vfmaq_lane_f32(vacc0123n1, vi0123, vw, 1); in xnn_f32_spmm_minmax_ukernel_16x2__neonfma()
178 float32x4_t vout0123n1 = vminq_f32(vacc0123n1, vmax); in xnn_f32_spmm_minmax_ukernel_16x2__neonfma()
[all …]
|
| D | 12x2-minmax-neonfma.c | 45 float32x4_t vacc0123n1 = vld1q_dup_f32(w); w += 1; in xnn_f32_spmm_minmax_ukernel_12x2__neonfma() local
46 float32x4_t vacc4567n1 = vacc0123n1; in xnn_f32_spmm_minmax_ukernel_12x2__neonfma()
47 float32x4_t vacc89ABn1 = vacc0123n1; in xnn_f32_spmm_minmax_ukernel_12x2__neonfma()
61 vacc0123n1 = vfmaq_lane_f32(vacc0123n1, vi0123, vw, 1); in xnn_f32_spmm_minmax_ukernel_12x2__neonfma()
69 float32x4_t vout0123n1 = vminq_f32(vacc0123n1, vmax); in xnn_f32_spmm_minmax_ukernel_12x2__neonfma()
143 float32x4_t vacc0123n1 = vld1q_dup_f32(w); w += 1; in xnn_f32_spmm_minmax_ukernel_12x2__neonfma() local
144 float32x4_t vacc4567n1 = vacc0123n1; in xnn_f32_spmm_minmax_ukernel_12x2__neonfma()
155 vacc0123n1 = vfmaq_lane_f32(vacc0123n1, vi0123, vw, 1); in xnn_f32_spmm_minmax_ukernel_12x2__neonfma()
161 float32x4_t vout0123n1 = vminq_f32(vacc0123n1, vmax); in xnn_f32_spmm_minmax_ukernel_12x2__neonfma()
219 float32x4_t vacc0123n1 = vld1q_dup_f32(w); w += 1; in xnn_f32_spmm_minmax_ukernel_12x2__neonfma() local
[all …]
|
| D | 32x4-minmax-neonfma.c | 50 float32x4_t vacc0123n1 = vld1q_dup_f32(w); w += 1; in xnn_f32_spmm_minmax_ukernel_32x4__neonfma() local
51 float32x4_t vacc4567n1 = vacc0123n1; in xnn_f32_spmm_minmax_ukernel_32x4__neonfma()
52 float32x4_t vacc89ABn1 = vacc0123n1; in xnn_f32_spmm_minmax_ukernel_32x4__neonfma()
53 float32x4_t vaccCDEFn1 = vacc0123n1; in xnn_f32_spmm_minmax_ukernel_32x4__neonfma()
54 float32x4_t vaccGHIJn1 = vacc0123n1; in xnn_f32_spmm_minmax_ukernel_32x4__neonfma()
55 float32x4_t vaccKLMNn1 = vacc0123n1; in xnn_f32_spmm_minmax_ukernel_32x4__neonfma()
56 float32x4_t vaccOPQRn1 = vacc0123n1; in xnn_f32_spmm_minmax_ukernel_32x4__neonfma()
57 float32x4_t vaccSTUVn1 = vacc0123n1; in xnn_f32_spmm_minmax_ukernel_32x4__neonfma()
98 vacc0123n1 = vfmaq_laneq_f32(vacc0123n1, vi0123, vw, 1); in xnn_f32_spmm_minmax_ukernel_32x4__neonfma()
132 float32x4_t vout0123n1 = vminq_f32(vacc0123n1, vmax); in xnn_f32_spmm_minmax_ukernel_32x4__neonfma()
[all …]
|
| D | 8x2-minmax-neonfma.c | 44 float32x4_t vacc0123n1 = vld1q_dup_f32(w); w += 1; in xnn_f32_spmm_minmax_ukernel_8x2__neonfma() local
45 float32x4_t vacc4567n1 = vacc0123n1; in xnn_f32_spmm_minmax_ukernel_8x2__neonfma()
57 vacc0123n1 = vfmaq_lane_f32(vacc0123n1, vi0123, vw, 1); in xnn_f32_spmm_minmax_ukernel_8x2__neonfma()
63 float32x4_t vout0123n1 = vminq_f32(vacc0123n1, vmax); in xnn_f32_spmm_minmax_ukernel_8x2__neonfma()
125 float32x4_t vacc0123n1 = vld1q_dup_f32(w); w += 1; in xnn_f32_spmm_minmax_ukernel_8x2__neonfma() local
134 vacc0123n1 = vfmaq_lane_f32(vacc0123n1, vi0123, vw, 1); in xnn_f32_spmm_minmax_ukernel_8x2__neonfma()
138 float32x4_t vout0123n1 = vminq_f32(vacc0123n1, vmax); in xnn_f32_spmm_minmax_ukernel_8x2__neonfma()
|
| D | 12x4-minmax-neonfma.c | 45 float32x4_t vacc0123n1 = vld1q_dup_f32(w); w += 1; in xnn_f32_spmm_minmax_ukernel_12x4__neonfma() local
46 float32x4_t vacc4567n1 = vacc0123n1; in xnn_f32_spmm_minmax_ukernel_12x4__neonfma()
47 float32x4_t vacc89ABn1 = vacc0123n1; in xnn_f32_spmm_minmax_ukernel_12x4__neonfma()
67 vacc0123n1 = vfmaq_laneq_f32(vacc0123n1, vi0123, vw, 1); in xnn_f32_spmm_minmax_ukernel_12x4__neonfma()
81 float32x4_t vout0123n1 = vminq_f32(vacc0123n1, vmax); in xnn_f32_spmm_minmax_ukernel_12x4__neonfma()
175 float32x4_t vacc0123n1 = vld1q_dup_f32(w); w += 1; in xnn_f32_spmm_minmax_ukernel_12x4__neonfma() local
176 float32x4_t vacc4567n1 = vacc0123n1; in xnn_f32_spmm_minmax_ukernel_12x4__neonfma()
191 vacc0123n1 = vfmaq_laneq_f32(vacc0123n1, vi0123, vw, 1); in xnn_f32_spmm_minmax_ukernel_12x4__neonfma()
201 float32x4_t vout0123n1 = vminq_f32(vacc0123n1, vmax); in xnn_f32_spmm_minmax_ukernel_12x4__neonfma()
273 float32x4_t vacc0123n1 = vld1q_dup_f32(w); w += 1; in xnn_f32_spmm_minmax_ukernel_12x4__neonfma() local
[all …]
|
| D | 16x4-minmax-neonfma.c | 46 float32x4_t vacc0123n1 = vld1q_dup_f32(w); w += 1; in xnn_f32_spmm_minmax_ukernel_16x4__neonfma() local
47 float32x4_t vacc4567n1 = vacc0123n1; in xnn_f32_spmm_minmax_ukernel_16x4__neonfma()
48 float32x4_t vacc89ABn1 = vacc0123n1; in xnn_f32_spmm_minmax_ukernel_16x4__neonfma()
49 float32x4_t vaccCDEFn1 = vacc0123n1; in xnn_f32_spmm_minmax_ukernel_16x4__neonfma()
73 vacc0123n1 = vfmaq_laneq_f32(vacc0123n1, vi0123, vw, 1); in xnn_f32_spmm_minmax_ukernel_16x4__neonfma()
91 float32x4_t vout0123n1 = vminq_f32(vacc0123n1, vmax); in xnn_f32_spmm_minmax_ukernel_16x4__neonfma()
202 float32x4_t vacc0123n1 = vld1q_dup_f32(w); w += 1; in xnn_f32_spmm_minmax_ukernel_16x4__neonfma() local
203 float32x4_t vacc4567n1 = vacc0123n1; in xnn_f32_spmm_minmax_ukernel_16x4__neonfma()
218 vacc0123n1 = vfmaq_laneq_f32(vacc0123n1, vi0123, vw, 1); in xnn_f32_spmm_minmax_ukernel_16x4__neonfma()
228 float32x4_t vout0123n1 = vminq_f32(vacc0123n1, vmax); in xnn_f32_spmm_minmax_ukernel_16x4__neonfma()
[all …]
|
| D | 8x4-minmax-neonfma.c | 44 float32x4_t vacc0123n1 = vld1q_dup_f32(w); w += 1; in xnn_f32_spmm_minmax_ukernel_8x4__neonfma() local
45 float32x4_t vacc4567n1 = vacc0123n1; in xnn_f32_spmm_minmax_ukernel_8x4__neonfma()
61 vacc0123n1 = vfmaq_laneq_f32(vacc0123n1, vi0123, vw, 1); in xnn_f32_spmm_minmax_ukernel_8x4__neonfma()
71 float32x4_t vout0123n1 = vminq_f32(vacc0123n1, vmax); in xnn_f32_spmm_minmax_ukernel_8x4__neonfma()
147 float32x4_t vacc0123n1 = vld1q_dup_f32(w); w += 1; in xnn_f32_spmm_minmax_ukernel_8x4__neonfma() local
158 vacc0123n1 = vfmaq_laneq_f32(vacc0123n1, vi0123, vw, 1); in xnn_f32_spmm_minmax_ukernel_8x4__neonfma()
164 float32x4_t vout0123n1 = vminq_f32(vacc0123n1, vmax); in xnn_f32_spmm_minmax_ukernel_8x4__neonfma()
|
| D | 4x2-minmax-neonfma.c | 43 float32x4_t vacc0123n1 = vld1q_dup_f32(w); w += 1; in xnn_f32_spmm_minmax_ukernel_4x2__neonfma() local
53 vacc0123n1 = vfmaq_lane_f32(vacc0123n1, vi0123, vw, 1); in xnn_f32_spmm_minmax_ukernel_4x2__neonfma()
57 float32x4_t vout0123n1 = vminq_f32(vacc0123n1, vmax); in xnn_f32_spmm_minmax_ukernel_4x2__neonfma()
|
| D | 4x4-minmax-neonfma.c | 43 float32x4_t vacc0123n1 = vld1q_dup_f32(w); w += 1; in xnn_f32_spmm_minmax_ukernel_4x4__neonfma() local
55 vacc0123n1 = vfmaq_laneq_f32(vacc0123n1, vi0123, vw, 1); in xnn_f32_spmm_minmax_ukernel_4x4__neonfma()
61 float32x4_t vout0123n1 = vminq_f32(vacc0123n1, vmax); in xnn_f32_spmm_minmax_ukernel_4x4__neonfma()
|