| /external/XNNPACK/src/f32-igemm/gen/ |
| D | 4x4-minmax-neon-lane-ld64.c | 95 const float32x4_t vb0123c0 = vld1q_f32(w); w += 4; in xnn_f32_igemm_minmax_ukernel_4x4__neon_lane_ld64() local
97 vacc0x0123 = vmlaq_lane_f32(vacc0x0123, vb0123c0, va0, 0); in xnn_f32_igemm_minmax_ukernel_4x4__neon_lane_ld64()
98 vacc1x0123 = vmlaq_lane_f32(vacc1x0123, vb0123c0, va1, 0); in xnn_f32_igemm_minmax_ukernel_4x4__neon_lane_ld64()
99 vacc2x0123 = vmlaq_lane_f32(vacc2x0123, vb0123c0, va2, 0); in xnn_f32_igemm_minmax_ukernel_4x4__neon_lane_ld64()
100 vacc3x0123 = vmlaq_lane_f32(vacc3x0123, vb0123c0, va3, 0); in xnn_f32_igemm_minmax_ukernel_4x4__neon_lane_ld64()
|
| D | 4x4-minmax-neonfma-lane-ld64.c | 95 const float32x4_t vb0123c0 = vld1q_f32(w); w += 4; in xnn_f32_igemm_minmax_ukernel_4x4__neonfma_lane_ld64() local
97 vacc0x0123 = vfmaq_lane_f32(vacc0x0123, vb0123c0, va0, 0); in xnn_f32_igemm_minmax_ukernel_4x4__neonfma_lane_ld64()
98 vacc1x0123 = vfmaq_lane_f32(vacc1x0123, vb0123c0, va1, 0); in xnn_f32_igemm_minmax_ukernel_4x4__neonfma_lane_ld64()
99 vacc2x0123 = vfmaq_lane_f32(vacc2x0123, vb0123c0, va2, 0); in xnn_f32_igemm_minmax_ukernel_4x4__neonfma_lane_ld64()
100 vacc3x0123 = vfmaq_lane_f32(vacc3x0123, vb0123c0, va3, 0); in xnn_f32_igemm_minmax_ukernel_4x4__neonfma_lane_ld64()
|
| D | 6x8-minmax-neonfma-lane-ld64.c | 123 const float32x4_t vb0123c0 = vld1q_f32(w); w += 4; in xnn_f32_igemm_minmax_ukernel_6x8__neonfma_lane_ld64() local
126 vacc0x0123 = vfmaq_lane_f32(vacc0x0123, vb0123c0, va0, 0); in xnn_f32_igemm_minmax_ukernel_6x8__neonfma_lane_ld64()
127 vacc1x0123 = vfmaq_lane_f32(vacc1x0123, vb0123c0, va1, 0); in xnn_f32_igemm_minmax_ukernel_6x8__neonfma_lane_ld64()
128 vacc2x0123 = vfmaq_lane_f32(vacc2x0123, vb0123c0, va2, 0); in xnn_f32_igemm_minmax_ukernel_6x8__neonfma_lane_ld64()
129 vacc3x0123 = vfmaq_lane_f32(vacc3x0123, vb0123c0, va3, 0); in xnn_f32_igemm_minmax_ukernel_6x8__neonfma_lane_ld64()
130 vacc4x0123 = vfmaq_lane_f32(vacc4x0123, vb0123c0, va4, 0); in xnn_f32_igemm_minmax_ukernel_6x8__neonfma_lane_ld64()
131 vacc5x0123 = vfmaq_lane_f32(vacc5x0123, vb0123c0, va5, 0); in xnn_f32_igemm_minmax_ukernel_6x8__neonfma_lane_ld64()
|
| D | 6x8-minmax-neon-lane-ld64.c | 123 const float32x4_t vb0123c0 = vld1q_f32(w); w += 4; in xnn_f32_igemm_minmax_ukernel_6x8__neon_lane_ld64() local
126 vacc0x0123 = vmlaq_lane_f32(vacc0x0123, vb0123c0, va0, 0); in xnn_f32_igemm_minmax_ukernel_6x8__neon_lane_ld64()
127 vacc1x0123 = vmlaq_lane_f32(vacc1x0123, vb0123c0, va1, 0); in xnn_f32_igemm_minmax_ukernel_6x8__neon_lane_ld64()
128 vacc2x0123 = vmlaq_lane_f32(vacc2x0123, vb0123c0, va2, 0); in xnn_f32_igemm_minmax_ukernel_6x8__neon_lane_ld64()
129 vacc3x0123 = vmlaq_lane_f32(vacc3x0123, vb0123c0, va3, 0); in xnn_f32_igemm_minmax_ukernel_6x8__neon_lane_ld64()
130 vacc4x0123 = vmlaq_lane_f32(vacc4x0123, vb0123c0, va4, 0); in xnn_f32_igemm_minmax_ukernel_6x8__neon_lane_ld64()
131 vacc5x0123 = vmlaq_lane_f32(vacc5x0123, vb0123c0, va5, 0); in xnn_f32_igemm_minmax_ukernel_6x8__neon_lane_ld64()
|
| D | 6x8-minmax-neonfma-dup-ld64.c | 123 const float32x4_t vb0123c0 = vld1q_f32(w); w += 4; in xnn_f32_igemm_minmax_ukernel_6x8__neonfma_dup_ld64() local
132 vacc0x0123 = vfmaq_f32(vacc0x0123, va0c0, vb0123c0); in xnn_f32_igemm_minmax_ukernel_6x8__neonfma_dup_ld64()
133 vacc1x0123 = vfmaq_f32(vacc1x0123, va1c0, vb0123c0); in xnn_f32_igemm_minmax_ukernel_6x8__neonfma_dup_ld64()
134 vacc2x0123 = vfmaq_f32(vacc2x0123, va2c0, vb0123c0); in xnn_f32_igemm_minmax_ukernel_6x8__neonfma_dup_ld64()
135 vacc3x0123 = vfmaq_f32(vacc3x0123, va3c0, vb0123c0); in xnn_f32_igemm_minmax_ukernel_6x8__neonfma_dup_ld64()
136 vacc4x0123 = vfmaq_f32(vacc4x0123, va4c0, vb0123c0); in xnn_f32_igemm_minmax_ukernel_6x8__neonfma_dup_ld64()
137 vacc5x0123 = vfmaq_f32(vacc5x0123, va5c0, vb0123c0); in xnn_f32_igemm_minmax_ukernel_6x8__neonfma_dup_ld64()
|
| D | 6x8-minmax-neon-dup-ld64.c | 123 const float32x4_t vb0123c0 = vld1q_f32(w); w += 4; in xnn_f32_igemm_minmax_ukernel_6x8__neon_dup_ld64() local
132 vacc0x0123 = vmlaq_f32(vacc0x0123, va0c0, vb0123c0); in xnn_f32_igemm_minmax_ukernel_6x8__neon_dup_ld64()
133 vacc1x0123 = vmlaq_f32(vacc1x0123, va1c0, vb0123c0); in xnn_f32_igemm_minmax_ukernel_6x8__neon_dup_ld64()
134 vacc2x0123 = vmlaq_f32(vacc2x0123, va2c0, vb0123c0); in xnn_f32_igemm_minmax_ukernel_6x8__neon_dup_ld64()
135 vacc3x0123 = vmlaq_f32(vacc3x0123, va3c0, vb0123c0); in xnn_f32_igemm_minmax_ukernel_6x8__neon_dup_ld64()
136 vacc4x0123 = vmlaq_f32(vacc4x0123, va4c0, vb0123c0); in xnn_f32_igemm_minmax_ukernel_6x8__neon_dup_ld64()
137 vacc5x0123 = vmlaq_f32(vacc5x0123, va5c0, vb0123c0); in xnn_f32_igemm_minmax_ukernel_6x8__neon_dup_ld64()
|
| D | 4x8-minmax-neon-dup-ld64.c | 99 const float32x4_t vb0123c0 = vld1q_f32(w); w += 4; in xnn_f32_igemm_minmax_ukernel_4x8__neon_dup_ld64() local
106 vacc0x0123 = vmlaq_f32(vacc0x0123, va0c0, vb0123c0); in xnn_f32_igemm_minmax_ukernel_4x8__neon_dup_ld64()
107 vacc1x0123 = vmlaq_f32(vacc1x0123, va1c0, vb0123c0); in xnn_f32_igemm_minmax_ukernel_4x8__neon_dup_ld64()
108 vacc2x0123 = vmlaq_f32(vacc2x0123, va2c0, vb0123c0); in xnn_f32_igemm_minmax_ukernel_4x8__neon_dup_ld64()
109 vacc3x0123 = vmlaq_f32(vacc3x0123, va3c0, vb0123c0); in xnn_f32_igemm_minmax_ukernel_4x8__neon_dup_ld64()
|
| D | 4x8-minmax-neonfma-dup-ld64.c | 99 const float32x4_t vb0123c0 = vld1q_f32(w); w += 4; in xnn_f32_igemm_minmax_ukernel_4x8__neonfma_dup_ld64() local
106 vacc0x0123 = vfmaq_f32(vacc0x0123, va0c0, vb0123c0); in xnn_f32_igemm_minmax_ukernel_4x8__neonfma_dup_ld64()
107 vacc1x0123 = vfmaq_f32(vacc1x0123, va1c0, vb0123c0); in xnn_f32_igemm_minmax_ukernel_4x8__neonfma_dup_ld64()
108 vacc2x0123 = vfmaq_f32(vacc2x0123, va2c0, vb0123c0); in xnn_f32_igemm_minmax_ukernel_4x8__neonfma_dup_ld64()
109 vacc3x0123 = vfmaq_f32(vacc3x0123, va3c0, vb0123c0); in xnn_f32_igemm_minmax_ukernel_4x8__neonfma_dup_ld64()
|
| /external/XNNPACK/src/f32-gemm/gen/ |
| D | 6x8-minmax-neonfma-lane-ld64.c | 95 const float32x4_t vb0123c0 = vld1q_f32(w); w += 4; in xnn_f32_gemm_minmax_ukernel_6x8__neonfma_lane_ld64() local
98 vacc0x0123 = vfmaq_lane_f32(vacc0x0123, vb0123c0, va0, 0); in xnn_f32_gemm_minmax_ukernel_6x8__neonfma_lane_ld64()
99 vacc1x0123 = vfmaq_lane_f32(vacc1x0123, vb0123c0, va1, 0); in xnn_f32_gemm_minmax_ukernel_6x8__neonfma_lane_ld64()
100 vacc2x0123 = vfmaq_lane_f32(vacc2x0123, vb0123c0, va2, 0); in xnn_f32_gemm_minmax_ukernel_6x8__neonfma_lane_ld64()
101 vacc3x0123 = vfmaq_lane_f32(vacc3x0123, vb0123c0, va3, 0); in xnn_f32_gemm_minmax_ukernel_6x8__neonfma_lane_ld64()
102 vacc4x0123 = vfmaq_lane_f32(vacc4x0123, vb0123c0, va4, 0); in xnn_f32_gemm_minmax_ukernel_6x8__neonfma_lane_ld64()
103 vacc5x0123 = vfmaq_lane_f32(vacc5x0123, vb0123c0, va5, 0); in xnn_f32_gemm_minmax_ukernel_6x8__neonfma_lane_ld64()
|
| D | 6x8-minmax-neon-lane-ld64.c | 95 const float32x4_t vb0123c0 = vld1q_f32(w); w += 4; in xnn_f32_gemm_minmax_ukernel_6x8__neon_lane_ld64() local
98 vacc0x0123 = vmlaq_lane_f32(vacc0x0123, vb0123c0, va0, 0); in xnn_f32_gemm_minmax_ukernel_6x8__neon_lane_ld64()
99 vacc1x0123 = vmlaq_lane_f32(vacc1x0123, vb0123c0, va1, 0); in xnn_f32_gemm_minmax_ukernel_6x8__neon_lane_ld64()
100 vacc2x0123 = vmlaq_lane_f32(vacc2x0123, vb0123c0, va2, 0); in xnn_f32_gemm_minmax_ukernel_6x8__neon_lane_ld64()
101 vacc3x0123 = vmlaq_lane_f32(vacc3x0123, vb0123c0, va3, 0); in xnn_f32_gemm_minmax_ukernel_6x8__neon_lane_ld64()
102 vacc4x0123 = vmlaq_lane_f32(vacc4x0123, vb0123c0, va4, 0); in xnn_f32_gemm_minmax_ukernel_6x8__neon_lane_ld64()
103 vacc5x0123 = vmlaq_lane_f32(vacc5x0123, vb0123c0, va5, 0); in xnn_f32_gemm_minmax_ukernel_6x8__neon_lane_ld64()
|
| D | 5x8-minmax-neonfma-lane-ld64.c | 86 const float32x4_t vb0123c0 = vld1q_f32(w); w += 4; in xnn_f32_gemm_minmax_ukernel_5x8__neonfma_lane_ld64() local
89 vacc0x0123 = vfmaq_lane_f32(vacc0x0123, vb0123c0, va0, 0); in xnn_f32_gemm_minmax_ukernel_5x8__neonfma_lane_ld64()
90 vacc1x0123 = vfmaq_lane_f32(vacc1x0123, vb0123c0, va1, 0); in xnn_f32_gemm_minmax_ukernel_5x8__neonfma_lane_ld64()
91 vacc2x0123 = vfmaq_lane_f32(vacc2x0123, vb0123c0, va2, 0); in xnn_f32_gemm_minmax_ukernel_5x8__neonfma_lane_ld64()
92 vacc3x0123 = vfmaq_lane_f32(vacc3x0123, vb0123c0, va3, 0); in xnn_f32_gemm_minmax_ukernel_5x8__neonfma_lane_ld64()
93 vacc4x0123 = vfmaq_lane_f32(vacc4x0123, vb0123c0, va4, 0); in xnn_f32_gemm_minmax_ukernel_5x8__neonfma_lane_ld64()
|
| D | 5x8-minmax-neon-lane-ld64.c | 86 const float32x4_t vb0123c0 = vld1q_f32(w); w += 4; in xnn_f32_gemm_minmax_ukernel_5x8__neon_lane_ld64() local
89 vacc0x0123 = vmlaq_lane_f32(vacc0x0123, vb0123c0, va0, 0); in xnn_f32_gemm_minmax_ukernel_5x8__neon_lane_ld64()
90 vacc1x0123 = vmlaq_lane_f32(vacc1x0123, vb0123c0, va1, 0); in xnn_f32_gemm_minmax_ukernel_5x8__neon_lane_ld64()
91 vacc2x0123 = vmlaq_lane_f32(vacc2x0123, vb0123c0, va2, 0); in xnn_f32_gemm_minmax_ukernel_5x8__neon_lane_ld64()
92 vacc3x0123 = vmlaq_lane_f32(vacc3x0123, vb0123c0, va3, 0); in xnn_f32_gemm_minmax_ukernel_5x8__neon_lane_ld64()
93 vacc4x0123 = vmlaq_lane_f32(vacc4x0123, vb0123c0, va4, 0); in xnn_f32_gemm_minmax_ukernel_5x8__neon_lane_ld64()
|
| D | 6x8-minmax-neon-dup-ld64.c | 95 const float32x4_t vb0123c0 = vld1q_f32(w); w += 4; in xnn_f32_gemm_minmax_ukernel_6x8__neon_dup_ld64() local
104 vacc0x0123 = vmlaq_f32(vacc0x0123, va0c0, vb0123c0); in xnn_f32_gemm_minmax_ukernel_6x8__neon_dup_ld64()
105 vacc1x0123 = vmlaq_f32(vacc1x0123, va1c0, vb0123c0); in xnn_f32_gemm_minmax_ukernel_6x8__neon_dup_ld64()
106 vacc2x0123 = vmlaq_f32(vacc2x0123, va2c0, vb0123c0); in xnn_f32_gemm_minmax_ukernel_6x8__neon_dup_ld64()
107 vacc3x0123 = vmlaq_f32(vacc3x0123, va3c0, vb0123c0); in xnn_f32_gemm_minmax_ukernel_6x8__neon_dup_ld64()
108 vacc4x0123 = vmlaq_f32(vacc4x0123, va4c0, vb0123c0); in xnn_f32_gemm_minmax_ukernel_6x8__neon_dup_ld64()
109 vacc5x0123 = vmlaq_f32(vacc5x0123, va5c0, vb0123c0); in xnn_f32_gemm_minmax_ukernel_6x8__neon_dup_ld64()
|
| D | 6x8-minmax-neonfma-dup-ld64.c | 95 const float32x4_t vb0123c0 = vld1q_f32(w); w += 4; in xnn_f32_gemm_minmax_ukernel_6x8__neonfma_dup_ld64() local
104 vacc0x0123 = vfmaq_f32(vacc0x0123, va0c0, vb0123c0); in xnn_f32_gemm_minmax_ukernel_6x8__neonfma_dup_ld64()
105 vacc1x0123 = vfmaq_f32(vacc1x0123, va1c0, vb0123c0); in xnn_f32_gemm_minmax_ukernel_6x8__neonfma_dup_ld64()
106 vacc2x0123 = vfmaq_f32(vacc2x0123, va2c0, vb0123c0); in xnn_f32_gemm_minmax_ukernel_6x8__neonfma_dup_ld64()
107 vacc3x0123 = vfmaq_f32(vacc3x0123, va3c0, vb0123c0); in xnn_f32_gemm_minmax_ukernel_6x8__neonfma_dup_ld64()
108 vacc4x0123 = vfmaq_f32(vacc4x0123, va4c0, vb0123c0); in xnn_f32_gemm_minmax_ukernel_6x8__neonfma_dup_ld64()
109 vacc5x0123 = vfmaq_f32(vacc5x0123, va5c0, vb0123c0); in xnn_f32_gemm_minmax_ukernel_6x8__neonfma_dup_ld64()
|
| D | 4x8-minmax-neon-lane-ld64.c | 77 const float32x4_t vb0123c0 = vld1q_f32(w); w += 4; in xnn_f32_gemm_minmax_ukernel_4x8__neon_lane_ld64() local
80 vacc0x0123 = vmlaq_lane_f32(vacc0x0123, vb0123c0, va0, 0); in xnn_f32_gemm_minmax_ukernel_4x8__neon_lane_ld64()
81 vacc1x0123 = vmlaq_lane_f32(vacc1x0123, vb0123c0, va1, 0); in xnn_f32_gemm_minmax_ukernel_4x8__neon_lane_ld64()
82 vacc2x0123 = vmlaq_lane_f32(vacc2x0123, vb0123c0, va2, 0); in xnn_f32_gemm_minmax_ukernel_4x8__neon_lane_ld64()
83 vacc3x0123 = vmlaq_lane_f32(vacc3x0123, vb0123c0, va3, 0); in xnn_f32_gemm_minmax_ukernel_4x8__neon_lane_ld64()
|
| D | 4x8-minmax-neonfma-lane-ld64.c | 77 const float32x4_t vb0123c0 = vld1q_f32(w); w += 4; in xnn_f32_gemm_minmax_ukernel_4x8__neonfma_lane_ld64() local
80 vacc0x0123 = vfmaq_lane_f32(vacc0x0123, vb0123c0, va0, 0); in xnn_f32_gemm_minmax_ukernel_4x8__neonfma_lane_ld64()
81 vacc1x0123 = vfmaq_lane_f32(vacc1x0123, vb0123c0, va1, 0); in xnn_f32_gemm_minmax_ukernel_4x8__neonfma_lane_ld64()
82 vacc2x0123 = vfmaq_lane_f32(vacc2x0123, vb0123c0, va2, 0); in xnn_f32_gemm_minmax_ukernel_4x8__neonfma_lane_ld64()
83 vacc3x0123 = vfmaq_lane_f32(vacc3x0123, vb0123c0, va3, 0); in xnn_f32_gemm_minmax_ukernel_4x8__neonfma_lane_ld64()
|
| D | 4x8-minmax-neonfma-dup-ld64.c | 77 const float32x4_t vb0123c0 = vld1q_f32(w); w += 4; in xnn_f32_gemm_minmax_ukernel_4x8__neonfma_dup_ld64() local
84 vacc0x0123 = vfmaq_f32(vacc0x0123, va0c0, vb0123c0); in xnn_f32_gemm_minmax_ukernel_4x8__neonfma_dup_ld64()
85 vacc1x0123 = vfmaq_f32(vacc1x0123, va1c0, vb0123c0); in xnn_f32_gemm_minmax_ukernel_4x8__neonfma_dup_ld64()
86 vacc2x0123 = vfmaq_f32(vacc2x0123, va2c0, vb0123c0); in xnn_f32_gemm_minmax_ukernel_4x8__neonfma_dup_ld64()
87 vacc3x0123 = vfmaq_f32(vacc3x0123, va3c0, vb0123c0); in xnn_f32_gemm_minmax_ukernel_4x8__neonfma_dup_ld64()
|
| /external/XNNPACK/src/f32-gemm/gen-inc/ |
| D | 6x8inc-minmax-neon-lane-ld64.c | 97 const float32x4_t vb0123c0 = vld1q_f32(w); w += 4; in xnn_f32_gemminc_minmax_ukernel_6x8__neon_lane_ld64() local
100 vacc0x0123 = vmlaq_lane_f32(vacc0x0123, vb0123c0, va0, 0); in xnn_f32_gemminc_minmax_ukernel_6x8__neon_lane_ld64()
101 vacc1x0123 = vmlaq_lane_f32(vacc1x0123, vb0123c0, va1, 0); in xnn_f32_gemminc_minmax_ukernel_6x8__neon_lane_ld64()
102 vacc2x0123 = vmlaq_lane_f32(vacc2x0123, vb0123c0, va2, 0); in xnn_f32_gemminc_minmax_ukernel_6x8__neon_lane_ld64()
103 vacc3x0123 = vmlaq_lane_f32(vacc3x0123, vb0123c0, va3, 0); in xnn_f32_gemminc_minmax_ukernel_6x8__neon_lane_ld64()
104 vacc4x0123 = vmlaq_lane_f32(vacc4x0123, vb0123c0, va4, 0); in xnn_f32_gemminc_minmax_ukernel_6x8__neon_lane_ld64()
105 vacc5x0123 = vmlaq_lane_f32(vacc5x0123, vb0123c0, va5, 0); in xnn_f32_gemminc_minmax_ukernel_6x8__neon_lane_ld64()
|
| D | 6x8inc-minmax-neonfma-lane-ld64.c | 97 const float32x4_t vb0123c0 = vld1q_f32(w); w += 4; in xnn_f32_gemminc_minmax_ukernel_6x8__neonfma_lane_ld64() local
100 vacc0x0123 = vfmaq_lane_f32(vacc0x0123, vb0123c0, va0, 0); in xnn_f32_gemminc_minmax_ukernel_6x8__neonfma_lane_ld64()
101 vacc1x0123 = vfmaq_lane_f32(vacc1x0123, vb0123c0, va1, 0); in xnn_f32_gemminc_minmax_ukernel_6x8__neonfma_lane_ld64()
102 vacc2x0123 = vfmaq_lane_f32(vacc2x0123, vb0123c0, va2, 0); in xnn_f32_gemminc_minmax_ukernel_6x8__neonfma_lane_ld64()
103 vacc3x0123 = vfmaq_lane_f32(vacc3x0123, vb0123c0, va3, 0); in xnn_f32_gemminc_minmax_ukernel_6x8__neonfma_lane_ld64()
104 vacc4x0123 = vfmaq_lane_f32(vacc4x0123, vb0123c0, va4, 0); in xnn_f32_gemminc_minmax_ukernel_6x8__neonfma_lane_ld64()
105 vacc5x0123 = vfmaq_lane_f32(vacc5x0123, vb0123c0, va5, 0); in xnn_f32_gemminc_minmax_ukernel_6x8__neonfma_lane_ld64()
|
| D | 5x8inc-minmax-neon-lane-ld64.c | 88 const float32x4_t vb0123c0 = vld1q_f32(w); w += 4; in xnn_f32_gemminc_minmax_ukernel_5x8__neon_lane_ld64() local
91 vacc0x0123 = vmlaq_lane_f32(vacc0x0123, vb0123c0, va0, 0); in xnn_f32_gemminc_minmax_ukernel_5x8__neon_lane_ld64()
92 vacc1x0123 = vmlaq_lane_f32(vacc1x0123, vb0123c0, va1, 0); in xnn_f32_gemminc_minmax_ukernel_5x8__neon_lane_ld64()
93 vacc2x0123 = vmlaq_lane_f32(vacc2x0123, vb0123c0, va2, 0); in xnn_f32_gemminc_minmax_ukernel_5x8__neon_lane_ld64()
94 vacc3x0123 = vmlaq_lane_f32(vacc3x0123, vb0123c0, va3, 0); in xnn_f32_gemminc_minmax_ukernel_5x8__neon_lane_ld64()
95 vacc4x0123 = vmlaq_lane_f32(vacc4x0123, vb0123c0, va4, 0); in xnn_f32_gemminc_minmax_ukernel_5x8__neon_lane_ld64()
|
| D | 5x8inc-minmax-neonfma-lane-ld64.c | 88 const float32x4_t vb0123c0 = vld1q_f32(w); w += 4; in xnn_f32_gemminc_minmax_ukernel_5x8__neonfma_lane_ld64() local
91 vacc0x0123 = vfmaq_lane_f32(vacc0x0123, vb0123c0, va0, 0); in xnn_f32_gemminc_minmax_ukernel_5x8__neonfma_lane_ld64()
92 vacc1x0123 = vfmaq_lane_f32(vacc1x0123, vb0123c0, va1, 0); in xnn_f32_gemminc_minmax_ukernel_5x8__neonfma_lane_ld64()
93 vacc2x0123 = vfmaq_lane_f32(vacc2x0123, vb0123c0, va2, 0); in xnn_f32_gemminc_minmax_ukernel_5x8__neonfma_lane_ld64()
94 vacc3x0123 = vfmaq_lane_f32(vacc3x0123, vb0123c0, va3, 0); in xnn_f32_gemminc_minmax_ukernel_5x8__neonfma_lane_ld64()
95 vacc4x0123 = vfmaq_lane_f32(vacc4x0123, vb0123c0, va4, 0); in xnn_f32_gemminc_minmax_ukernel_5x8__neonfma_lane_ld64()
|
| D | 6x8inc-minmax-neon-dup-ld64.c | 97 const float32x4_t vb0123c0 = vld1q_f32(w); w += 4; in xnn_f32_gemminc_minmax_ukernel_6x8__neon_dup_ld64() local
106 vacc0x0123 = vmlaq_f32(vacc0x0123, va0c0, vb0123c0); in xnn_f32_gemminc_minmax_ukernel_6x8__neon_dup_ld64()
107 vacc1x0123 = vmlaq_f32(vacc1x0123, va1c0, vb0123c0); in xnn_f32_gemminc_minmax_ukernel_6x8__neon_dup_ld64()
108 vacc2x0123 = vmlaq_f32(vacc2x0123, va2c0, vb0123c0); in xnn_f32_gemminc_minmax_ukernel_6x8__neon_dup_ld64()
109 vacc3x0123 = vmlaq_f32(vacc3x0123, va3c0, vb0123c0); in xnn_f32_gemminc_minmax_ukernel_6x8__neon_dup_ld64()
110 vacc4x0123 = vmlaq_f32(vacc4x0123, va4c0, vb0123c0); in xnn_f32_gemminc_minmax_ukernel_6x8__neon_dup_ld64()
111 vacc5x0123 = vmlaq_f32(vacc5x0123, va5c0, vb0123c0); in xnn_f32_gemminc_minmax_ukernel_6x8__neon_dup_ld64()
|
| D | 6x8inc-minmax-neonfma-dup-ld64.c | 97 const float32x4_t vb0123c0 = vld1q_f32(w); w += 4; in xnn_f32_gemminc_minmax_ukernel_6x8__neonfma_dup_ld64() local
106 vacc0x0123 = vfmaq_f32(vacc0x0123, va0c0, vb0123c0); in xnn_f32_gemminc_minmax_ukernel_6x8__neonfma_dup_ld64()
107 vacc1x0123 = vfmaq_f32(vacc1x0123, va1c0, vb0123c0); in xnn_f32_gemminc_minmax_ukernel_6x8__neonfma_dup_ld64()
108 vacc2x0123 = vfmaq_f32(vacc2x0123, va2c0, vb0123c0); in xnn_f32_gemminc_minmax_ukernel_6x8__neonfma_dup_ld64()
109 vacc3x0123 = vfmaq_f32(vacc3x0123, va3c0, vb0123c0); in xnn_f32_gemminc_minmax_ukernel_6x8__neonfma_dup_ld64()
110 vacc4x0123 = vfmaq_f32(vacc4x0123, va4c0, vb0123c0); in xnn_f32_gemminc_minmax_ukernel_6x8__neonfma_dup_ld64()
111 vacc5x0123 = vfmaq_f32(vacc5x0123, va5c0, vb0123c0); in xnn_f32_gemminc_minmax_ukernel_6x8__neonfma_dup_ld64()
|
| D | 4x8inc-minmax-neonfma-lane-ld64.c | 79 const float32x4_t vb0123c0 = vld1q_f32(w); w += 4; in xnn_f32_gemminc_minmax_ukernel_4x8__neonfma_lane_ld64() local
82 vacc0x0123 = vfmaq_lane_f32(vacc0x0123, vb0123c0, va0, 0); in xnn_f32_gemminc_minmax_ukernel_4x8__neonfma_lane_ld64()
83 vacc1x0123 = vfmaq_lane_f32(vacc1x0123, vb0123c0, va1, 0); in xnn_f32_gemminc_minmax_ukernel_4x8__neonfma_lane_ld64()
84 vacc2x0123 = vfmaq_lane_f32(vacc2x0123, vb0123c0, va2, 0); in xnn_f32_gemminc_minmax_ukernel_4x8__neonfma_lane_ld64()
85 vacc3x0123 = vfmaq_lane_f32(vacc3x0123, vb0123c0, va3, 0); in xnn_f32_gemminc_minmax_ukernel_4x8__neonfma_lane_ld64()
|
| D | 4x8inc-minmax-neon-dup-ld64.c | 79 const float32x4_t vb0123c0 = vld1q_f32(w); w += 4; in xnn_f32_gemminc_minmax_ukernel_4x8__neon_dup_ld64() local
86 vacc0x0123 = vmlaq_f32(vacc0x0123, va0c0, vb0123c0); in xnn_f32_gemminc_minmax_ukernel_4x8__neon_dup_ld64()
87 vacc1x0123 = vmlaq_f32(vacc1x0123, va1c0, vb0123c0); in xnn_f32_gemminc_minmax_ukernel_4x8__neon_dup_ld64()
88 vacc2x0123 = vmlaq_f32(vacc2x0123, va2c0, vb0123c0); in xnn_f32_gemminc_minmax_ukernel_4x8__neon_dup_ld64()
89 vacc3x0123 = vmlaq_f32(vacc3x0123, va3c0, vb0123c0); in xnn_f32_gemminc_minmax_ukernel_4x8__neon_dup_ld64()
|