| /external/XNNPACK/src/f32-igemm/gen/ |
| D | 4x4-minmax-neon-lane-ld64.c | 101 const float32x4_t vb0123c1 = vld1q_f32(w); w += 4; in xnn_f32_igemm_minmax_ukernel_4x4__neon_lane_ld64() local
103 vacc0x0123 = vmlaq_lane_f32(vacc0x0123, vb0123c1, va0, 1); in xnn_f32_igemm_minmax_ukernel_4x4__neon_lane_ld64()
104 vacc1x0123 = vmlaq_lane_f32(vacc1x0123, vb0123c1, va1, 1); in xnn_f32_igemm_minmax_ukernel_4x4__neon_lane_ld64()
105 vacc2x0123 = vmlaq_lane_f32(vacc2x0123, vb0123c1, va2, 1); in xnn_f32_igemm_minmax_ukernel_4x4__neon_lane_ld64()
106 vacc3x0123 = vmlaq_lane_f32(vacc3x0123, vb0123c1, va3, 1); in xnn_f32_igemm_minmax_ukernel_4x4__neon_lane_ld64()
|
| D | 4x4-minmax-neonfma-lane-ld64.c | 101 const float32x4_t vb0123c1 = vld1q_f32(w); w += 4; in xnn_f32_igemm_minmax_ukernel_4x4__neonfma_lane_ld64() local
103 vacc0x0123 = vfmaq_lane_f32(vacc0x0123, vb0123c1, va0, 1); in xnn_f32_igemm_minmax_ukernel_4x4__neonfma_lane_ld64()
104 vacc1x0123 = vfmaq_lane_f32(vacc1x0123, vb0123c1, va1, 1); in xnn_f32_igemm_minmax_ukernel_4x4__neonfma_lane_ld64()
105 vacc2x0123 = vfmaq_lane_f32(vacc2x0123, vb0123c1, va2, 1); in xnn_f32_igemm_minmax_ukernel_4x4__neonfma_lane_ld64()
106 vacc3x0123 = vfmaq_lane_f32(vacc3x0123, vb0123c1, va3, 1); in xnn_f32_igemm_minmax_ukernel_4x4__neonfma_lane_ld64()
|
| D | 6x8-minmax-neonfma-lane-ld64.c | 138 const float32x4_t vb0123c1 = vld1q_f32(w); w += 4; in xnn_f32_igemm_minmax_ukernel_6x8__neonfma_lane_ld64() local
141 vacc0x0123 = vfmaq_lane_f32(vacc0x0123, vb0123c1, va0, 1); in xnn_f32_igemm_minmax_ukernel_6x8__neonfma_lane_ld64()
142 vacc1x0123 = vfmaq_lane_f32(vacc1x0123, vb0123c1, va1, 1); in xnn_f32_igemm_minmax_ukernel_6x8__neonfma_lane_ld64()
143 vacc2x0123 = vfmaq_lane_f32(vacc2x0123, vb0123c1, va2, 1); in xnn_f32_igemm_minmax_ukernel_6x8__neonfma_lane_ld64()
144 vacc3x0123 = vfmaq_lane_f32(vacc3x0123, vb0123c1, va3, 1); in xnn_f32_igemm_minmax_ukernel_6x8__neonfma_lane_ld64()
145 vacc4x0123 = vfmaq_lane_f32(vacc4x0123, vb0123c1, va4, 1); in xnn_f32_igemm_minmax_ukernel_6x8__neonfma_lane_ld64()
146 vacc5x0123 = vfmaq_lane_f32(vacc5x0123, vb0123c1, va5, 1); in xnn_f32_igemm_minmax_ukernel_6x8__neonfma_lane_ld64()
|
| D | 6x8-minmax-neon-lane-ld64.c | 138 const float32x4_t vb0123c1 = vld1q_f32(w); w += 4; in xnn_f32_igemm_minmax_ukernel_6x8__neon_lane_ld64() local
141 vacc0x0123 = vmlaq_lane_f32(vacc0x0123, vb0123c1, va0, 1); in xnn_f32_igemm_minmax_ukernel_6x8__neon_lane_ld64()
142 vacc1x0123 = vmlaq_lane_f32(vacc1x0123, vb0123c1, va1, 1); in xnn_f32_igemm_minmax_ukernel_6x8__neon_lane_ld64()
143 vacc2x0123 = vmlaq_lane_f32(vacc2x0123, vb0123c1, va2, 1); in xnn_f32_igemm_minmax_ukernel_6x8__neon_lane_ld64()
144 vacc3x0123 = vmlaq_lane_f32(vacc3x0123, vb0123c1, va3, 1); in xnn_f32_igemm_minmax_ukernel_6x8__neon_lane_ld64()
145 vacc4x0123 = vmlaq_lane_f32(vacc4x0123, vb0123c1, va4, 1); in xnn_f32_igemm_minmax_ukernel_6x8__neon_lane_ld64()
146 vacc5x0123 = vmlaq_lane_f32(vacc5x0123, vb0123c1, va5, 1); in xnn_f32_igemm_minmax_ukernel_6x8__neon_lane_ld64()
|
| D | 6x8-minmax-neonfma-dup-ld64.c | 144 const float32x4_t vb0123c1 = vld1q_f32(w); w += 4; in xnn_f32_igemm_minmax_ukernel_6x8__neonfma_dup_ld64() local
153 vacc0x0123 = vfmaq_f32(vacc0x0123, va0c1, vb0123c1); in xnn_f32_igemm_minmax_ukernel_6x8__neonfma_dup_ld64()
154 vacc1x0123 = vfmaq_f32(vacc1x0123, va1c1, vb0123c1); in xnn_f32_igemm_minmax_ukernel_6x8__neonfma_dup_ld64()
155 vacc2x0123 = vfmaq_f32(vacc2x0123, va2c1, vb0123c1); in xnn_f32_igemm_minmax_ukernel_6x8__neonfma_dup_ld64()
156 vacc3x0123 = vfmaq_f32(vacc3x0123, va3c1, vb0123c1); in xnn_f32_igemm_minmax_ukernel_6x8__neonfma_dup_ld64()
157 vacc4x0123 = vfmaq_f32(vacc4x0123, va4c1, vb0123c1); in xnn_f32_igemm_minmax_ukernel_6x8__neonfma_dup_ld64()
158 vacc5x0123 = vfmaq_f32(vacc5x0123, va5c1, vb0123c1); in xnn_f32_igemm_minmax_ukernel_6x8__neonfma_dup_ld64()
|
| D | 6x8-minmax-neon-dup-ld64.c | 144 const float32x4_t vb0123c1 = vld1q_f32(w); w += 4; in xnn_f32_igemm_minmax_ukernel_6x8__neon_dup_ld64() local
153 vacc0x0123 = vmlaq_f32(vacc0x0123, va0c1, vb0123c1); in xnn_f32_igemm_minmax_ukernel_6x8__neon_dup_ld64()
154 vacc1x0123 = vmlaq_f32(vacc1x0123, va1c1, vb0123c1); in xnn_f32_igemm_minmax_ukernel_6x8__neon_dup_ld64()
155 vacc2x0123 = vmlaq_f32(vacc2x0123, va2c1, vb0123c1); in xnn_f32_igemm_minmax_ukernel_6x8__neon_dup_ld64()
156 vacc3x0123 = vmlaq_f32(vacc3x0123, va3c1, vb0123c1); in xnn_f32_igemm_minmax_ukernel_6x8__neon_dup_ld64()
157 vacc4x0123 = vmlaq_f32(vacc4x0123, va4c1, vb0123c1); in xnn_f32_igemm_minmax_ukernel_6x8__neon_dup_ld64()
158 vacc5x0123 = vmlaq_f32(vacc5x0123, va5c1, vb0123c1); in xnn_f32_igemm_minmax_ukernel_6x8__neon_dup_ld64()
|
| D | 4x8-minmax-neon-dup-ld64.c | 114 const float32x4_t vb0123c1 = vld1q_f32(w); w += 4; in xnn_f32_igemm_minmax_ukernel_4x8__neon_dup_ld64() local
121 vacc0x0123 = vmlaq_f32(vacc0x0123, va0c1, vb0123c1); in xnn_f32_igemm_minmax_ukernel_4x8__neon_dup_ld64()
122 vacc1x0123 = vmlaq_f32(vacc1x0123, va1c1, vb0123c1); in xnn_f32_igemm_minmax_ukernel_4x8__neon_dup_ld64()
123 vacc2x0123 = vmlaq_f32(vacc2x0123, va2c1, vb0123c1); in xnn_f32_igemm_minmax_ukernel_4x8__neon_dup_ld64()
124 vacc3x0123 = vmlaq_f32(vacc3x0123, va3c1, vb0123c1); in xnn_f32_igemm_minmax_ukernel_4x8__neon_dup_ld64()
|
| D | 4x8-minmax-neonfma-dup-ld64.c | 114 const float32x4_t vb0123c1 = vld1q_f32(w); w += 4; in xnn_f32_igemm_minmax_ukernel_4x8__neonfma_dup_ld64() local
121 vacc0x0123 = vfmaq_f32(vacc0x0123, va0c1, vb0123c1); in xnn_f32_igemm_minmax_ukernel_4x8__neonfma_dup_ld64()
122 vacc1x0123 = vfmaq_f32(vacc1x0123, va1c1, vb0123c1); in xnn_f32_igemm_minmax_ukernel_4x8__neonfma_dup_ld64()
123 vacc2x0123 = vfmaq_f32(vacc2x0123, va2c1, vb0123c1); in xnn_f32_igemm_minmax_ukernel_4x8__neonfma_dup_ld64()
124 vacc3x0123 = vfmaq_f32(vacc3x0123, va3c1, vb0123c1); in xnn_f32_igemm_minmax_ukernel_4x8__neonfma_dup_ld64()
|
| /external/XNNPACK/src/f32-gemm/gen/ |
| D | 6x8-minmax-neonfma-lane-ld64.c | 110 const float32x4_t vb0123c1 = vld1q_f32(w); w += 4; in xnn_f32_gemm_minmax_ukernel_6x8__neonfma_lane_ld64() local
113 vacc0x0123 = vfmaq_lane_f32(vacc0x0123, vb0123c1, va0, 1); in xnn_f32_gemm_minmax_ukernel_6x8__neonfma_lane_ld64()
114 vacc1x0123 = vfmaq_lane_f32(vacc1x0123, vb0123c1, va1, 1); in xnn_f32_gemm_minmax_ukernel_6x8__neonfma_lane_ld64()
115 vacc2x0123 = vfmaq_lane_f32(vacc2x0123, vb0123c1, va2, 1); in xnn_f32_gemm_minmax_ukernel_6x8__neonfma_lane_ld64()
116 vacc3x0123 = vfmaq_lane_f32(vacc3x0123, vb0123c1, va3, 1); in xnn_f32_gemm_minmax_ukernel_6x8__neonfma_lane_ld64()
117 vacc4x0123 = vfmaq_lane_f32(vacc4x0123, vb0123c1, va4, 1); in xnn_f32_gemm_minmax_ukernel_6x8__neonfma_lane_ld64()
118 vacc5x0123 = vfmaq_lane_f32(vacc5x0123, vb0123c1, va5, 1); in xnn_f32_gemm_minmax_ukernel_6x8__neonfma_lane_ld64()
|
| D | 6x8-minmax-neon-lane-ld64.c | 110 const float32x4_t vb0123c1 = vld1q_f32(w); w += 4; in xnn_f32_gemm_minmax_ukernel_6x8__neon_lane_ld64() local
113 vacc0x0123 = vmlaq_lane_f32(vacc0x0123, vb0123c1, va0, 1); in xnn_f32_gemm_minmax_ukernel_6x8__neon_lane_ld64()
114 vacc1x0123 = vmlaq_lane_f32(vacc1x0123, vb0123c1, va1, 1); in xnn_f32_gemm_minmax_ukernel_6x8__neon_lane_ld64()
115 vacc2x0123 = vmlaq_lane_f32(vacc2x0123, vb0123c1, va2, 1); in xnn_f32_gemm_minmax_ukernel_6x8__neon_lane_ld64()
116 vacc3x0123 = vmlaq_lane_f32(vacc3x0123, vb0123c1, va3, 1); in xnn_f32_gemm_minmax_ukernel_6x8__neon_lane_ld64()
117 vacc4x0123 = vmlaq_lane_f32(vacc4x0123, vb0123c1, va4, 1); in xnn_f32_gemm_minmax_ukernel_6x8__neon_lane_ld64()
118 vacc5x0123 = vmlaq_lane_f32(vacc5x0123, vb0123c1, va5, 1); in xnn_f32_gemm_minmax_ukernel_6x8__neon_lane_ld64()
|
| D | 5x8-minmax-neonfma-lane-ld64.c | 99 const float32x4_t vb0123c1 = vld1q_f32(w); w += 4; in xnn_f32_gemm_minmax_ukernel_5x8__neonfma_lane_ld64() local
102 vacc0x0123 = vfmaq_lane_f32(vacc0x0123, vb0123c1, va0, 1); in xnn_f32_gemm_minmax_ukernel_5x8__neonfma_lane_ld64()
103 vacc1x0123 = vfmaq_lane_f32(vacc1x0123, vb0123c1, va1, 1); in xnn_f32_gemm_minmax_ukernel_5x8__neonfma_lane_ld64()
104 vacc2x0123 = vfmaq_lane_f32(vacc2x0123, vb0123c1, va2, 1); in xnn_f32_gemm_minmax_ukernel_5x8__neonfma_lane_ld64()
105 vacc3x0123 = vfmaq_lane_f32(vacc3x0123, vb0123c1, va3, 1); in xnn_f32_gemm_minmax_ukernel_5x8__neonfma_lane_ld64()
106 vacc4x0123 = vfmaq_lane_f32(vacc4x0123, vb0123c1, va4, 1); in xnn_f32_gemm_minmax_ukernel_5x8__neonfma_lane_ld64()
|
| D | 5x8-minmax-neon-lane-ld64.c | 99 const float32x4_t vb0123c1 = vld1q_f32(w); w += 4; in xnn_f32_gemm_minmax_ukernel_5x8__neon_lane_ld64() local
102 vacc0x0123 = vmlaq_lane_f32(vacc0x0123, vb0123c1, va0, 1); in xnn_f32_gemm_minmax_ukernel_5x8__neon_lane_ld64()
103 vacc1x0123 = vmlaq_lane_f32(vacc1x0123, vb0123c1, va1, 1); in xnn_f32_gemm_minmax_ukernel_5x8__neon_lane_ld64()
104 vacc2x0123 = vmlaq_lane_f32(vacc2x0123, vb0123c1, va2, 1); in xnn_f32_gemm_minmax_ukernel_5x8__neon_lane_ld64()
105 vacc3x0123 = vmlaq_lane_f32(vacc3x0123, vb0123c1, va3, 1); in xnn_f32_gemm_minmax_ukernel_5x8__neon_lane_ld64()
106 vacc4x0123 = vmlaq_lane_f32(vacc4x0123, vb0123c1, va4, 1); in xnn_f32_gemm_minmax_ukernel_5x8__neon_lane_ld64()
|
| D | 6x8-minmax-neon-dup-ld64.c | 116 const float32x4_t vb0123c1 = vld1q_f32(w); w += 4; in xnn_f32_gemm_minmax_ukernel_6x8__neon_dup_ld64() local
125 vacc0x0123 = vmlaq_f32(vacc0x0123, va0c1, vb0123c1); in xnn_f32_gemm_minmax_ukernel_6x8__neon_dup_ld64()
126 vacc1x0123 = vmlaq_f32(vacc1x0123, va1c1, vb0123c1); in xnn_f32_gemm_minmax_ukernel_6x8__neon_dup_ld64()
127 vacc2x0123 = vmlaq_f32(vacc2x0123, va2c1, vb0123c1); in xnn_f32_gemm_minmax_ukernel_6x8__neon_dup_ld64()
128 vacc3x0123 = vmlaq_f32(vacc3x0123, va3c1, vb0123c1); in xnn_f32_gemm_minmax_ukernel_6x8__neon_dup_ld64()
129 vacc4x0123 = vmlaq_f32(vacc4x0123, va4c1, vb0123c1); in xnn_f32_gemm_minmax_ukernel_6x8__neon_dup_ld64()
130 vacc5x0123 = vmlaq_f32(vacc5x0123, va5c1, vb0123c1); in xnn_f32_gemm_minmax_ukernel_6x8__neon_dup_ld64()
|
| D | 6x8-minmax-neonfma-dup-ld64.c | 116 const float32x4_t vb0123c1 = vld1q_f32(w); w += 4; in xnn_f32_gemm_minmax_ukernel_6x8__neonfma_dup_ld64() local
125 vacc0x0123 = vfmaq_f32(vacc0x0123, va0c1, vb0123c1); in xnn_f32_gemm_minmax_ukernel_6x8__neonfma_dup_ld64()
126 vacc1x0123 = vfmaq_f32(vacc1x0123, va1c1, vb0123c1); in xnn_f32_gemm_minmax_ukernel_6x8__neonfma_dup_ld64()
127 vacc2x0123 = vfmaq_f32(vacc2x0123, va2c1, vb0123c1); in xnn_f32_gemm_minmax_ukernel_6x8__neonfma_dup_ld64()
128 vacc3x0123 = vfmaq_f32(vacc3x0123, va3c1, vb0123c1); in xnn_f32_gemm_minmax_ukernel_6x8__neonfma_dup_ld64()
129 vacc4x0123 = vfmaq_f32(vacc4x0123, va4c1, vb0123c1); in xnn_f32_gemm_minmax_ukernel_6x8__neonfma_dup_ld64()
130 vacc5x0123 = vfmaq_f32(vacc5x0123, va5c1, vb0123c1); in xnn_f32_gemm_minmax_ukernel_6x8__neonfma_dup_ld64()
|
| D | 4x8-minmax-neon-lane-ld64.c | 88 const float32x4_t vb0123c1 = vld1q_f32(w); w += 4; in xnn_f32_gemm_minmax_ukernel_4x8__neon_lane_ld64() local
91 vacc0x0123 = vmlaq_lane_f32(vacc0x0123, vb0123c1, va0, 1); in xnn_f32_gemm_minmax_ukernel_4x8__neon_lane_ld64()
92 vacc1x0123 = vmlaq_lane_f32(vacc1x0123, vb0123c1, va1, 1); in xnn_f32_gemm_minmax_ukernel_4x8__neon_lane_ld64()
93 vacc2x0123 = vmlaq_lane_f32(vacc2x0123, vb0123c1, va2, 1); in xnn_f32_gemm_minmax_ukernel_4x8__neon_lane_ld64()
94 vacc3x0123 = vmlaq_lane_f32(vacc3x0123, vb0123c1, va3, 1); in xnn_f32_gemm_minmax_ukernel_4x8__neon_lane_ld64()
|
| D | 4x8-minmax-neonfma-lane-ld64.c | 88 const float32x4_t vb0123c1 = vld1q_f32(w); w += 4; in xnn_f32_gemm_minmax_ukernel_4x8__neonfma_lane_ld64() local
91 vacc0x0123 = vfmaq_lane_f32(vacc0x0123, vb0123c1, va0, 1); in xnn_f32_gemm_minmax_ukernel_4x8__neonfma_lane_ld64()
92 vacc1x0123 = vfmaq_lane_f32(vacc1x0123, vb0123c1, va1, 1); in xnn_f32_gemm_minmax_ukernel_4x8__neonfma_lane_ld64()
93 vacc2x0123 = vfmaq_lane_f32(vacc2x0123, vb0123c1, va2, 1); in xnn_f32_gemm_minmax_ukernel_4x8__neonfma_lane_ld64()
94 vacc3x0123 = vfmaq_lane_f32(vacc3x0123, vb0123c1, va3, 1); in xnn_f32_gemm_minmax_ukernel_4x8__neonfma_lane_ld64()
|
| D | 4x8-minmax-neonfma-dup-ld64.c | 92 const float32x4_t vb0123c1 = vld1q_f32(w); w += 4; in xnn_f32_gemm_minmax_ukernel_4x8__neonfma_dup_ld64() local
99 vacc0x0123 = vfmaq_f32(vacc0x0123, va0c1, vb0123c1); in xnn_f32_gemm_minmax_ukernel_4x8__neonfma_dup_ld64()
100 vacc1x0123 = vfmaq_f32(vacc1x0123, va1c1, vb0123c1); in xnn_f32_gemm_minmax_ukernel_4x8__neonfma_dup_ld64()
101 vacc2x0123 = vfmaq_f32(vacc2x0123, va2c1, vb0123c1); in xnn_f32_gemm_minmax_ukernel_4x8__neonfma_dup_ld64()
102 vacc3x0123 = vfmaq_f32(vacc3x0123, va3c1, vb0123c1); in xnn_f32_gemm_minmax_ukernel_4x8__neonfma_dup_ld64()
|
| /external/XNNPACK/src/f32-gemm/gen-inc/ |
| D | 6x8inc-minmax-neon-lane-ld64.c | 112 const float32x4_t vb0123c1 = vld1q_f32(w); w += 4; in xnn_f32_gemminc_minmax_ukernel_6x8__neon_lane_ld64() local
115 vacc0x0123 = vmlaq_lane_f32(vacc0x0123, vb0123c1, va0, 1); in xnn_f32_gemminc_minmax_ukernel_6x8__neon_lane_ld64()
116 vacc1x0123 = vmlaq_lane_f32(vacc1x0123, vb0123c1, va1, 1); in xnn_f32_gemminc_minmax_ukernel_6x8__neon_lane_ld64()
117 vacc2x0123 = vmlaq_lane_f32(vacc2x0123, vb0123c1, va2, 1); in xnn_f32_gemminc_minmax_ukernel_6x8__neon_lane_ld64()
118 vacc3x0123 = vmlaq_lane_f32(vacc3x0123, vb0123c1, va3, 1); in xnn_f32_gemminc_minmax_ukernel_6x8__neon_lane_ld64()
119 vacc4x0123 = vmlaq_lane_f32(vacc4x0123, vb0123c1, va4, 1); in xnn_f32_gemminc_minmax_ukernel_6x8__neon_lane_ld64()
120 vacc5x0123 = vmlaq_lane_f32(vacc5x0123, vb0123c1, va5, 1); in xnn_f32_gemminc_minmax_ukernel_6x8__neon_lane_ld64()
|
| D | 6x8inc-minmax-neonfma-lane-ld64.c | 112 const float32x4_t vb0123c1 = vld1q_f32(w); w += 4; in xnn_f32_gemminc_minmax_ukernel_6x8__neonfma_lane_ld64() local
115 vacc0x0123 = vfmaq_lane_f32(vacc0x0123, vb0123c1, va0, 1); in xnn_f32_gemminc_minmax_ukernel_6x8__neonfma_lane_ld64()
116 vacc1x0123 = vfmaq_lane_f32(vacc1x0123, vb0123c1, va1, 1); in xnn_f32_gemminc_minmax_ukernel_6x8__neonfma_lane_ld64()
117 vacc2x0123 = vfmaq_lane_f32(vacc2x0123, vb0123c1, va2, 1); in xnn_f32_gemminc_minmax_ukernel_6x8__neonfma_lane_ld64()
118 vacc3x0123 = vfmaq_lane_f32(vacc3x0123, vb0123c1, va3, 1); in xnn_f32_gemminc_minmax_ukernel_6x8__neonfma_lane_ld64()
119 vacc4x0123 = vfmaq_lane_f32(vacc4x0123, vb0123c1, va4, 1); in xnn_f32_gemminc_minmax_ukernel_6x8__neonfma_lane_ld64()
120 vacc5x0123 = vfmaq_lane_f32(vacc5x0123, vb0123c1, va5, 1); in xnn_f32_gemminc_minmax_ukernel_6x8__neonfma_lane_ld64()
|
| D | 5x8inc-minmax-neon-lane-ld64.c | 101 const float32x4_t vb0123c1 = vld1q_f32(w); w += 4; in xnn_f32_gemminc_minmax_ukernel_5x8__neon_lane_ld64() local
104 vacc0x0123 = vmlaq_lane_f32(vacc0x0123, vb0123c1, va0, 1); in xnn_f32_gemminc_minmax_ukernel_5x8__neon_lane_ld64()
105 vacc1x0123 = vmlaq_lane_f32(vacc1x0123, vb0123c1, va1, 1); in xnn_f32_gemminc_minmax_ukernel_5x8__neon_lane_ld64()
106 vacc2x0123 = vmlaq_lane_f32(vacc2x0123, vb0123c1, va2, 1); in xnn_f32_gemminc_minmax_ukernel_5x8__neon_lane_ld64()
107 vacc3x0123 = vmlaq_lane_f32(vacc3x0123, vb0123c1, va3, 1); in xnn_f32_gemminc_minmax_ukernel_5x8__neon_lane_ld64()
108 vacc4x0123 = vmlaq_lane_f32(vacc4x0123, vb0123c1, va4, 1); in xnn_f32_gemminc_minmax_ukernel_5x8__neon_lane_ld64()
|
| D | 5x8inc-minmax-neonfma-lane-ld64.c | 101 const float32x4_t vb0123c1 = vld1q_f32(w); w += 4; in xnn_f32_gemminc_minmax_ukernel_5x8__neonfma_lane_ld64() local
104 vacc0x0123 = vfmaq_lane_f32(vacc0x0123, vb0123c1, va0, 1); in xnn_f32_gemminc_minmax_ukernel_5x8__neonfma_lane_ld64()
105 vacc1x0123 = vfmaq_lane_f32(vacc1x0123, vb0123c1, va1, 1); in xnn_f32_gemminc_minmax_ukernel_5x8__neonfma_lane_ld64()
106 vacc2x0123 = vfmaq_lane_f32(vacc2x0123, vb0123c1, va2, 1); in xnn_f32_gemminc_minmax_ukernel_5x8__neonfma_lane_ld64()
107 vacc3x0123 = vfmaq_lane_f32(vacc3x0123, vb0123c1, va3, 1); in xnn_f32_gemminc_minmax_ukernel_5x8__neonfma_lane_ld64()
108 vacc4x0123 = vfmaq_lane_f32(vacc4x0123, vb0123c1, va4, 1); in xnn_f32_gemminc_minmax_ukernel_5x8__neonfma_lane_ld64()
|
| D | 6x8inc-minmax-neon-dup-ld64.c | 118 const float32x4_t vb0123c1 = vld1q_f32(w); w += 4; in xnn_f32_gemminc_minmax_ukernel_6x8__neon_dup_ld64() local
127 vacc0x0123 = vmlaq_f32(vacc0x0123, va0c1, vb0123c1); in xnn_f32_gemminc_minmax_ukernel_6x8__neon_dup_ld64()
128 vacc1x0123 = vmlaq_f32(vacc1x0123, va1c1, vb0123c1); in xnn_f32_gemminc_minmax_ukernel_6x8__neon_dup_ld64()
129 vacc2x0123 = vmlaq_f32(vacc2x0123, va2c1, vb0123c1); in xnn_f32_gemminc_minmax_ukernel_6x8__neon_dup_ld64()
130 vacc3x0123 = vmlaq_f32(vacc3x0123, va3c1, vb0123c1); in xnn_f32_gemminc_minmax_ukernel_6x8__neon_dup_ld64()
131 vacc4x0123 = vmlaq_f32(vacc4x0123, va4c1, vb0123c1); in xnn_f32_gemminc_minmax_ukernel_6x8__neon_dup_ld64()
132 vacc5x0123 = vmlaq_f32(vacc5x0123, va5c1, vb0123c1); in xnn_f32_gemminc_minmax_ukernel_6x8__neon_dup_ld64()
|
| D | 6x8inc-minmax-neonfma-dup-ld64.c | 118 const float32x4_t vb0123c1 = vld1q_f32(w); w += 4; in xnn_f32_gemminc_minmax_ukernel_6x8__neonfma_dup_ld64() local
127 vacc0x0123 = vfmaq_f32(vacc0x0123, va0c1, vb0123c1); in xnn_f32_gemminc_minmax_ukernel_6x8__neonfma_dup_ld64()
128 vacc1x0123 = vfmaq_f32(vacc1x0123, va1c1, vb0123c1); in xnn_f32_gemminc_minmax_ukernel_6x8__neonfma_dup_ld64()
129 vacc2x0123 = vfmaq_f32(vacc2x0123, va2c1, vb0123c1); in xnn_f32_gemminc_minmax_ukernel_6x8__neonfma_dup_ld64()
130 vacc3x0123 = vfmaq_f32(vacc3x0123, va3c1, vb0123c1); in xnn_f32_gemminc_minmax_ukernel_6x8__neonfma_dup_ld64()
131 vacc4x0123 = vfmaq_f32(vacc4x0123, va4c1, vb0123c1); in xnn_f32_gemminc_minmax_ukernel_6x8__neonfma_dup_ld64()
132 vacc5x0123 = vfmaq_f32(vacc5x0123, va5c1, vb0123c1); in xnn_f32_gemminc_minmax_ukernel_6x8__neonfma_dup_ld64()
|
| D | 4x8inc-minmax-neonfma-lane-ld64.c | 90 const float32x4_t vb0123c1 = vld1q_f32(w); w += 4; in xnn_f32_gemminc_minmax_ukernel_4x8__neonfma_lane_ld64() local
93 vacc0x0123 = vfmaq_lane_f32(vacc0x0123, vb0123c1, va0, 1); in xnn_f32_gemminc_minmax_ukernel_4x8__neonfma_lane_ld64()
94 vacc1x0123 = vfmaq_lane_f32(vacc1x0123, vb0123c1, va1, 1); in xnn_f32_gemminc_minmax_ukernel_4x8__neonfma_lane_ld64()
95 vacc2x0123 = vfmaq_lane_f32(vacc2x0123, vb0123c1, va2, 1); in xnn_f32_gemminc_minmax_ukernel_4x8__neonfma_lane_ld64()
96 vacc3x0123 = vfmaq_lane_f32(vacc3x0123, vb0123c1, va3, 1); in xnn_f32_gemminc_minmax_ukernel_4x8__neonfma_lane_ld64()
|
| D | 4x8inc-minmax-neon-dup-ld64.c | 94 const float32x4_t vb0123c1 = vld1q_f32(w); w += 4; in xnn_f32_gemminc_minmax_ukernel_4x8__neon_dup_ld64() local
101 vacc0x0123 = vmlaq_f32(vacc0x0123, va0c1, vb0123c1); in xnn_f32_gemminc_minmax_ukernel_4x8__neon_dup_ld64()
102 vacc1x0123 = vmlaq_f32(vacc1x0123, va1c1, vb0123c1); in xnn_f32_gemminc_minmax_ukernel_4x8__neon_dup_ld64()
103 vacc2x0123 = vmlaq_f32(vacc2x0123, va2c1, vb0123c1); in xnn_f32_gemminc_minmax_ukernel_4x8__neon_dup_ld64()
104 vacc3x0123 = vmlaq_f32(vacc3x0123, va3c1, vb0123c1); in xnn_f32_gemminc_minmax_ukernel_4x8__neon_dup_ld64()
|