| /external/XNNPACK/src/f32-igemm/gen/ |
| D | 6x8-minmax-neonfma-lane-ld64.c | 124 const float32x4_t vb4567c0 = vld1q_f32(w); w += 4; in xnn_f32_igemm_minmax_ukernel_6x8__neonfma_lane_ld64() local
132 vacc0x4567 = vfmaq_lane_f32(vacc0x4567, vb4567c0, va0, 0); in xnn_f32_igemm_minmax_ukernel_6x8__neonfma_lane_ld64()
133 vacc1x4567 = vfmaq_lane_f32(vacc1x4567, vb4567c0, va1, 0); in xnn_f32_igemm_minmax_ukernel_6x8__neonfma_lane_ld64()
134 vacc2x4567 = vfmaq_lane_f32(vacc2x4567, vb4567c0, va2, 0); in xnn_f32_igemm_minmax_ukernel_6x8__neonfma_lane_ld64()
135 vacc3x4567 = vfmaq_lane_f32(vacc3x4567, vb4567c0, va3, 0); in xnn_f32_igemm_minmax_ukernel_6x8__neonfma_lane_ld64()
136 vacc4x4567 = vfmaq_lane_f32(vacc4x4567, vb4567c0, va4, 0); in xnn_f32_igemm_minmax_ukernel_6x8__neonfma_lane_ld64()
137 vacc5x4567 = vfmaq_lane_f32(vacc5x4567, vb4567c0, va5, 0); in xnn_f32_igemm_minmax_ukernel_6x8__neonfma_lane_ld64()
|
| D | 6x8-minmax-neon-lane-ld64.c | 124 const float32x4_t vb4567c0 = vld1q_f32(w); w += 4; in xnn_f32_igemm_minmax_ukernel_6x8__neon_lane_ld64() local
132 vacc0x4567 = vmlaq_lane_f32(vacc0x4567, vb4567c0, va0, 0); in xnn_f32_igemm_minmax_ukernel_6x8__neon_lane_ld64()
133 vacc1x4567 = vmlaq_lane_f32(vacc1x4567, vb4567c0, va1, 0); in xnn_f32_igemm_minmax_ukernel_6x8__neon_lane_ld64()
134 vacc2x4567 = vmlaq_lane_f32(vacc2x4567, vb4567c0, va2, 0); in xnn_f32_igemm_minmax_ukernel_6x8__neon_lane_ld64()
135 vacc3x4567 = vmlaq_lane_f32(vacc3x4567, vb4567c0, va3, 0); in xnn_f32_igemm_minmax_ukernel_6x8__neon_lane_ld64()
136 vacc4x4567 = vmlaq_lane_f32(vacc4x4567, vb4567c0, va4, 0); in xnn_f32_igemm_minmax_ukernel_6x8__neon_lane_ld64()
137 vacc5x4567 = vmlaq_lane_f32(vacc5x4567, vb4567c0, va5, 0); in xnn_f32_igemm_minmax_ukernel_6x8__neon_lane_ld64()
|
| D | 6x8-minmax-neonfma-dup-ld64.c | 124 const float32x4_t vb4567c0 = vld1q_f32(w); w += 4; in xnn_f32_igemm_minmax_ukernel_6x8__neonfma_dup_ld64() local
138 vacc0x4567 = vfmaq_f32(vacc0x4567, va0c0, vb4567c0); in xnn_f32_igemm_minmax_ukernel_6x8__neonfma_dup_ld64()
139 vacc1x4567 = vfmaq_f32(vacc1x4567, va1c0, vb4567c0); in xnn_f32_igemm_minmax_ukernel_6x8__neonfma_dup_ld64()
140 vacc2x4567 = vfmaq_f32(vacc2x4567, va2c0, vb4567c0); in xnn_f32_igemm_minmax_ukernel_6x8__neonfma_dup_ld64()
141 vacc3x4567 = vfmaq_f32(vacc3x4567, va3c0, vb4567c0); in xnn_f32_igemm_minmax_ukernel_6x8__neonfma_dup_ld64()
142 vacc4x4567 = vfmaq_f32(vacc4x4567, va4c0, vb4567c0); in xnn_f32_igemm_minmax_ukernel_6x8__neonfma_dup_ld64()
143 vacc5x4567 = vfmaq_f32(vacc5x4567, va5c0, vb4567c0); in xnn_f32_igemm_minmax_ukernel_6x8__neonfma_dup_ld64()
|
| D | 6x8-minmax-neon-dup-ld64.c | 124 const float32x4_t vb4567c0 = vld1q_f32(w); w += 4; in xnn_f32_igemm_minmax_ukernel_6x8__neon_dup_ld64() local
138 vacc0x4567 = vmlaq_f32(vacc0x4567, va0c0, vb4567c0); in xnn_f32_igemm_minmax_ukernel_6x8__neon_dup_ld64()
139 vacc1x4567 = vmlaq_f32(vacc1x4567, va1c0, vb4567c0); in xnn_f32_igemm_minmax_ukernel_6x8__neon_dup_ld64()
140 vacc2x4567 = vmlaq_f32(vacc2x4567, va2c0, vb4567c0); in xnn_f32_igemm_minmax_ukernel_6x8__neon_dup_ld64()
141 vacc3x4567 = vmlaq_f32(vacc3x4567, va3c0, vb4567c0); in xnn_f32_igemm_minmax_ukernel_6x8__neon_dup_ld64()
142 vacc4x4567 = vmlaq_f32(vacc4x4567, va4c0, vb4567c0); in xnn_f32_igemm_minmax_ukernel_6x8__neon_dup_ld64()
143 vacc5x4567 = vmlaq_f32(vacc5x4567, va5c0, vb4567c0); in xnn_f32_igemm_minmax_ukernel_6x8__neon_dup_ld64()
|
| D | 4x8-minmax-neon-dup-ld64.c | 100 const float32x4_t vb4567c0 = vld1q_f32(w); w += 4; in xnn_f32_igemm_minmax_ukernel_4x8__neon_dup_ld64() local
110 vacc0x4567 = vmlaq_f32(vacc0x4567, va0c0, vb4567c0); in xnn_f32_igemm_minmax_ukernel_4x8__neon_dup_ld64()
111 vacc1x4567 = vmlaq_f32(vacc1x4567, va1c0, vb4567c0); in xnn_f32_igemm_minmax_ukernel_4x8__neon_dup_ld64()
112 vacc2x4567 = vmlaq_f32(vacc2x4567, va2c0, vb4567c0); in xnn_f32_igemm_minmax_ukernel_4x8__neon_dup_ld64()
113 vacc3x4567 = vmlaq_f32(vacc3x4567, va3c0, vb4567c0); in xnn_f32_igemm_minmax_ukernel_4x8__neon_dup_ld64()
|
| D | 4x8-minmax-neonfma-dup-ld64.c | 100 const float32x4_t vb4567c0 = vld1q_f32(w); w += 4; in xnn_f32_igemm_minmax_ukernel_4x8__neonfma_dup_ld64() local
110 vacc0x4567 = vfmaq_f32(vacc0x4567, va0c0, vb4567c0); in xnn_f32_igemm_minmax_ukernel_4x8__neonfma_dup_ld64()
111 vacc1x4567 = vfmaq_f32(vacc1x4567, va1c0, vb4567c0); in xnn_f32_igemm_minmax_ukernel_4x8__neonfma_dup_ld64()
112 vacc2x4567 = vfmaq_f32(vacc2x4567, va2c0, vb4567c0); in xnn_f32_igemm_minmax_ukernel_4x8__neonfma_dup_ld64()
113 vacc3x4567 = vfmaq_f32(vacc3x4567, va3c0, vb4567c0); in xnn_f32_igemm_minmax_ukernel_4x8__neonfma_dup_ld64()
|
| D | 4x8-minmax-neon-lane-ld64.c | 100 const float32x4_t vb4567c0 = vld1q_f32(w); w += 4; in xnn_f32_igemm_minmax_ukernel_4x8__neon_lane_ld64() local
106 vacc0x4567 = vmlaq_lane_f32(vacc0x4567, vb4567c0, va0, 0); in xnn_f32_igemm_minmax_ukernel_4x8__neon_lane_ld64()
107 vacc1x4567 = vmlaq_lane_f32(vacc1x4567, vb4567c0, va1, 0); in xnn_f32_igemm_minmax_ukernel_4x8__neon_lane_ld64()
108 vacc2x4567 = vmlaq_lane_f32(vacc2x4567, vb4567c0, va2, 0); in xnn_f32_igemm_minmax_ukernel_4x8__neon_lane_ld64()
109 vacc3x4567 = vmlaq_lane_f32(vacc3x4567, vb4567c0, va3, 0); in xnn_f32_igemm_minmax_ukernel_4x8__neon_lane_ld64()
|
| D | 4x8-minmax-neonfma-lane-ld64.c | 100 const float32x4_t vb4567c0 = vld1q_f32(w); w += 4; in xnn_f32_igemm_minmax_ukernel_4x8__neonfma_lane_ld64() local
106 vacc0x4567 = vfmaq_lane_f32(vacc0x4567, vb4567c0, va0, 0); in xnn_f32_igemm_minmax_ukernel_4x8__neonfma_lane_ld64()
107 vacc1x4567 = vfmaq_lane_f32(vacc1x4567, vb4567c0, va1, 0); in xnn_f32_igemm_minmax_ukernel_4x8__neonfma_lane_ld64()
108 vacc2x4567 = vfmaq_lane_f32(vacc2x4567, vb4567c0, va2, 0); in xnn_f32_igemm_minmax_ukernel_4x8__neonfma_lane_ld64()
109 vacc3x4567 = vfmaq_lane_f32(vacc3x4567, vb4567c0, va3, 0); in xnn_f32_igemm_minmax_ukernel_4x8__neonfma_lane_ld64()
|
| /external/XNNPACK/src/f32-gemm/gen/ |
| D | 6x8-minmax-neonfma-lane-ld64.c | 96 const float32x4_t vb4567c0 = vld1q_f32(w); w += 4; in xnn_f32_gemm_minmax_ukernel_6x8__neonfma_lane_ld64() local
104 vacc0x4567 = vfmaq_lane_f32(vacc0x4567, vb4567c0, va0, 0); in xnn_f32_gemm_minmax_ukernel_6x8__neonfma_lane_ld64()
105 vacc1x4567 = vfmaq_lane_f32(vacc1x4567, vb4567c0, va1, 0); in xnn_f32_gemm_minmax_ukernel_6x8__neonfma_lane_ld64()
106 vacc2x4567 = vfmaq_lane_f32(vacc2x4567, vb4567c0, va2, 0); in xnn_f32_gemm_minmax_ukernel_6x8__neonfma_lane_ld64()
107 vacc3x4567 = vfmaq_lane_f32(vacc3x4567, vb4567c0, va3, 0); in xnn_f32_gemm_minmax_ukernel_6x8__neonfma_lane_ld64()
108 vacc4x4567 = vfmaq_lane_f32(vacc4x4567, vb4567c0, va4, 0); in xnn_f32_gemm_minmax_ukernel_6x8__neonfma_lane_ld64()
109 vacc5x4567 = vfmaq_lane_f32(vacc5x4567, vb4567c0, va5, 0); in xnn_f32_gemm_minmax_ukernel_6x8__neonfma_lane_ld64()
|
| D | 6x8-minmax-neon-lane-ld64.c | 96 const float32x4_t vb4567c0 = vld1q_f32(w); w += 4; in xnn_f32_gemm_minmax_ukernel_6x8__neon_lane_ld64() local
104 vacc0x4567 = vmlaq_lane_f32(vacc0x4567, vb4567c0, va0, 0); in xnn_f32_gemm_minmax_ukernel_6x8__neon_lane_ld64()
105 vacc1x4567 = vmlaq_lane_f32(vacc1x4567, vb4567c0, va1, 0); in xnn_f32_gemm_minmax_ukernel_6x8__neon_lane_ld64()
106 vacc2x4567 = vmlaq_lane_f32(vacc2x4567, vb4567c0, va2, 0); in xnn_f32_gemm_minmax_ukernel_6x8__neon_lane_ld64()
107 vacc3x4567 = vmlaq_lane_f32(vacc3x4567, vb4567c0, va3, 0); in xnn_f32_gemm_minmax_ukernel_6x8__neon_lane_ld64()
108 vacc4x4567 = vmlaq_lane_f32(vacc4x4567, vb4567c0, va4, 0); in xnn_f32_gemm_minmax_ukernel_6x8__neon_lane_ld64()
109 vacc5x4567 = vmlaq_lane_f32(vacc5x4567, vb4567c0, va5, 0); in xnn_f32_gemm_minmax_ukernel_6x8__neon_lane_ld64()
|
| D | 5x8-minmax-neonfma-lane-ld64.c | 87 const float32x4_t vb4567c0 = vld1q_f32(w); w += 4; in xnn_f32_gemm_minmax_ukernel_5x8__neonfma_lane_ld64() local
94 vacc0x4567 = vfmaq_lane_f32(vacc0x4567, vb4567c0, va0, 0); in xnn_f32_gemm_minmax_ukernel_5x8__neonfma_lane_ld64()
95 vacc1x4567 = vfmaq_lane_f32(vacc1x4567, vb4567c0, va1, 0); in xnn_f32_gemm_minmax_ukernel_5x8__neonfma_lane_ld64()
96 vacc2x4567 = vfmaq_lane_f32(vacc2x4567, vb4567c0, va2, 0); in xnn_f32_gemm_minmax_ukernel_5x8__neonfma_lane_ld64()
97 vacc3x4567 = vfmaq_lane_f32(vacc3x4567, vb4567c0, va3, 0); in xnn_f32_gemm_minmax_ukernel_5x8__neonfma_lane_ld64()
98 vacc4x4567 = vfmaq_lane_f32(vacc4x4567, vb4567c0, va4, 0); in xnn_f32_gemm_minmax_ukernel_5x8__neonfma_lane_ld64()
|
| D | 5x8-minmax-neon-lane-ld64.c | 87 const float32x4_t vb4567c0 = vld1q_f32(w); w += 4; in xnn_f32_gemm_minmax_ukernel_5x8__neon_lane_ld64() local
94 vacc0x4567 = vmlaq_lane_f32(vacc0x4567, vb4567c0, va0, 0); in xnn_f32_gemm_minmax_ukernel_5x8__neon_lane_ld64()
95 vacc1x4567 = vmlaq_lane_f32(vacc1x4567, vb4567c0, va1, 0); in xnn_f32_gemm_minmax_ukernel_5x8__neon_lane_ld64()
96 vacc2x4567 = vmlaq_lane_f32(vacc2x4567, vb4567c0, va2, 0); in xnn_f32_gemm_minmax_ukernel_5x8__neon_lane_ld64()
97 vacc3x4567 = vmlaq_lane_f32(vacc3x4567, vb4567c0, va3, 0); in xnn_f32_gemm_minmax_ukernel_5x8__neon_lane_ld64()
98 vacc4x4567 = vmlaq_lane_f32(vacc4x4567, vb4567c0, va4, 0); in xnn_f32_gemm_minmax_ukernel_5x8__neon_lane_ld64()
|
| D | 6x8-minmax-neon-dup-ld64.c | 96 const float32x4_t vb4567c0 = vld1q_f32(w); w += 4; in xnn_f32_gemm_minmax_ukernel_6x8__neon_dup_ld64() local
110 vacc0x4567 = vmlaq_f32(vacc0x4567, va0c0, vb4567c0); in xnn_f32_gemm_minmax_ukernel_6x8__neon_dup_ld64()
111 vacc1x4567 = vmlaq_f32(vacc1x4567, va1c0, vb4567c0); in xnn_f32_gemm_minmax_ukernel_6x8__neon_dup_ld64()
112 vacc2x4567 = vmlaq_f32(vacc2x4567, va2c0, vb4567c0); in xnn_f32_gemm_minmax_ukernel_6x8__neon_dup_ld64()
113 vacc3x4567 = vmlaq_f32(vacc3x4567, va3c0, vb4567c0); in xnn_f32_gemm_minmax_ukernel_6x8__neon_dup_ld64()
114 vacc4x4567 = vmlaq_f32(vacc4x4567, va4c0, vb4567c0); in xnn_f32_gemm_minmax_ukernel_6x8__neon_dup_ld64()
115 vacc5x4567 = vmlaq_f32(vacc5x4567, va5c0, vb4567c0); in xnn_f32_gemm_minmax_ukernel_6x8__neon_dup_ld64()
|
| D | 6x8-minmax-neonfma-dup-ld64.c | 96 const float32x4_t vb4567c0 = vld1q_f32(w); w += 4; in xnn_f32_gemm_minmax_ukernel_6x8__neonfma_dup_ld64() local
110 vacc0x4567 = vfmaq_f32(vacc0x4567, va0c0, vb4567c0); in xnn_f32_gemm_minmax_ukernel_6x8__neonfma_dup_ld64()
111 vacc1x4567 = vfmaq_f32(vacc1x4567, va1c0, vb4567c0); in xnn_f32_gemm_minmax_ukernel_6x8__neonfma_dup_ld64()
112 vacc2x4567 = vfmaq_f32(vacc2x4567, va2c0, vb4567c0); in xnn_f32_gemm_minmax_ukernel_6x8__neonfma_dup_ld64()
113 vacc3x4567 = vfmaq_f32(vacc3x4567, va3c0, vb4567c0); in xnn_f32_gemm_minmax_ukernel_6x8__neonfma_dup_ld64()
114 vacc4x4567 = vfmaq_f32(vacc4x4567, va4c0, vb4567c0); in xnn_f32_gemm_minmax_ukernel_6x8__neonfma_dup_ld64()
115 vacc5x4567 = vfmaq_f32(vacc5x4567, va5c0, vb4567c0); in xnn_f32_gemm_minmax_ukernel_6x8__neonfma_dup_ld64()
|
| D | 4x8-minmax-neon-lane-ld64.c | 78 const float32x4_t vb4567c0 = vld1q_f32(w); w += 4; in xnn_f32_gemm_minmax_ukernel_4x8__neon_lane_ld64() local
84 vacc0x4567 = vmlaq_lane_f32(vacc0x4567, vb4567c0, va0, 0); in xnn_f32_gemm_minmax_ukernel_4x8__neon_lane_ld64()
85 vacc1x4567 = vmlaq_lane_f32(vacc1x4567, vb4567c0, va1, 0); in xnn_f32_gemm_minmax_ukernel_4x8__neon_lane_ld64()
86 vacc2x4567 = vmlaq_lane_f32(vacc2x4567, vb4567c0, va2, 0); in xnn_f32_gemm_minmax_ukernel_4x8__neon_lane_ld64()
87 vacc3x4567 = vmlaq_lane_f32(vacc3x4567, vb4567c0, va3, 0); in xnn_f32_gemm_minmax_ukernel_4x8__neon_lane_ld64()
|
| D | 4x8-minmax-neonfma-lane-ld64.c | 78 const float32x4_t vb4567c0 = vld1q_f32(w); w += 4; in xnn_f32_gemm_minmax_ukernel_4x8__neonfma_lane_ld64() local
84 vacc0x4567 = vfmaq_lane_f32(vacc0x4567, vb4567c0, va0, 0); in xnn_f32_gemm_minmax_ukernel_4x8__neonfma_lane_ld64()
85 vacc1x4567 = vfmaq_lane_f32(vacc1x4567, vb4567c0, va1, 0); in xnn_f32_gemm_minmax_ukernel_4x8__neonfma_lane_ld64()
86 vacc2x4567 = vfmaq_lane_f32(vacc2x4567, vb4567c0, va2, 0); in xnn_f32_gemm_minmax_ukernel_4x8__neonfma_lane_ld64()
87 vacc3x4567 = vfmaq_lane_f32(vacc3x4567, vb4567c0, va3, 0); in xnn_f32_gemm_minmax_ukernel_4x8__neonfma_lane_ld64()
|
| D | 4x8-minmax-neonfma-dup-ld64.c | 78 const float32x4_t vb4567c0 = vld1q_f32(w); w += 4; in xnn_f32_gemm_minmax_ukernel_4x8__neonfma_dup_ld64() local
88 vacc0x4567 = vfmaq_f32(vacc0x4567, va0c0, vb4567c0); in xnn_f32_gemm_minmax_ukernel_4x8__neonfma_dup_ld64()
89 vacc1x4567 = vfmaq_f32(vacc1x4567, va1c0, vb4567c0); in xnn_f32_gemm_minmax_ukernel_4x8__neonfma_dup_ld64()
90 vacc2x4567 = vfmaq_f32(vacc2x4567, va2c0, vb4567c0); in xnn_f32_gemm_minmax_ukernel_4x8__neonfma_dup_ld64()
91 vacc3x4567 = vfmaq_f32(vacc3x4567, va3c0, vb4567c0); in xnn_f32_gemm_minmax_ukernel_4x8__neonfma_dup_ld64()
|
| /external/XNNPACK/src/f32-gemm/gen-inc/ |
| D | 6x8inc-minmax-neon-lane-ld64.c | 98 const float32x4_t vb4567c0 = vld1q_f32(w); w += 4; in xnn_f32_gemminc_minmax_ukernel_6x8__neon_lane_ld64() local
106 vacc0x4567 = vmlaq_lane_f32(vacc0x4567, vb4567c0, va0, 0); in xnn_f32_gemminc_minmax_ukernel_6x8__neon_lane_ld64()
107 vacc1x4567 = vmlaq_lane_f32(vacc1x4567, vb4567c0, va1, 0); in xnn_f32_gemminc_minmax_ukernel_6x8__neon_lane_ld64()
108 vacc2x4567 = vmlaq_lane_f32(vacc2x4567, vb4567c0, va2, 0); in xnn_f32_gemminc_minmax_ukernel_6x8__neon_lane_ld64()
109 vacc3x4567 = vmlaq_lane_f32(vacc3x4567, vb4567c0, va3, 0); in xnn_f32_gemminc_minmax_ukernel_6x8__neon_lane_ld64()
110 vacc4x4567 = vmlaq_lane_f32(vacc4x4567, vb4567c0, va4, 0); in xnn_f32_gemminc_minmax_ukernel_6x8__neon_lane_ld64()
111 vacc5x4567 = vmlaq_lane_f32(vacc5x4567, vb4567c0, va5, 0); in xnn_f32_gemminc_minmax_ukernel_6x8__neon_lane_ld64()
|
| D | 6x8inc-minmax-neonfma-lane-ld64.c | 98 const float32x4_t vb4567c0 = vld1q_f32(w); w += 4; in xnn_f32_gemminc_minmax_ukernel_6x8__neonfma_lane_ld64() local
106 vacc0x4567 = vfmaq_lane_f32(vacc0x4567, vb4567c0, va0, 0); in xnn_f32_gemminc_minmax_ukernel_6x8__neonfma_lane_ld64()
107 vacc1x4567 = vfmaq_lane_f32(vacc1x4567, vb4567c0, va1, 0); in xnn_f32_gemminc_minmax_ukernel_6x8__neonfma_lane_ld64()
108 vacc2x4567 = vfmaq_lane_f32(vacc2x4567, vb4567c0, va2, 0); in xnn_f32_gemminc_minmax_ukernel_6x8__neonfma_lane_ld64()
109 vacc3x4567 = vfmaq_lane_f32(vacc3x4567, vb4567c0, va3, 0); in xnn_f32_gemminc_minmax_ukernel_6x8__neonfma_lane_ld64()
110 vacc4x4567 = vfmaq_lane_f32(vacc4x4567, vb4567c0, va4, 0); in xnn_f32_gemminc_minmax_ukernel_6x8__neonfma_lane_ld64()
111 vacc5x4567 = vfmaq_lane_f32(vacc5x4567, vb4567c0, va5, 0); in xnn_f32_gemminc_minmax_ukernel_6x8__neonfma_lane_ld64()
|
| D | 5x8inc-minmax-neon-lane-ld64.c | 89 const float32x4_t vb4567c0 = vld1q_f32(w); w += 4; in xnn_f32_gemminc_minmax_ukernel_5x8__neon_lane_ld64() local
96 vacc0x4567 = vmlaq_lane_f32(vacc0x4567, vb4567c0, va0, 0); in xnn_f32_gemminc_minmax_ukernel_5x8__neon_lane_ld64()
97 vacc1x4567 = vmlaq_lane_f32(vacc1x4567, vb4567c0, va1, 0); in xnn_f32_gemminc_minmax_ukernel_5x8__neon_lane_ld64()
98 vacc2x4567 = vmlaq_lane_f32(vacc2x4567, vb4567c0, va2, 0); in xnn_f32_gemminc_minmax_ukernel_5x8__neon_lane_ld64()
99 vacc3x4567 = vmlaq_lane_f32(vacc3x4567, vb4567c0, va3, 0); in xnn_f32_gemminc_minmax_ukernel_5x8__neon_lane_ld64()
100 vacc4x4567 = vmlaq_lane_f32(vacc4x4567, vb4567c0, va4, 0); in xnn_f32_gemminc_minmax_ukernel_5x8__neon_lane_ld64()
|
| D | 5x8inc-minmax-neonfma-lane-ld64.c | 89 const float32x4_t vb4567c0 = vld1q_f32(w); w += 4; in xnn_f32_gemminc_minmax_ukernel_5x8__neonfma_lane_ld64() local
96 vacc0x4567 = vfmaq_lane_f32(vacc0x4567, vb4567c0, va0, 0); in xnn_f32_gemminc_minmax_ukernel_5x8__neonfma_lane_ld64()
97 vacc1x4567 = vfmaq_lane_f32(vacc1x4567, vb4567c0, va1, 0); in xnn_f32_gemminc_minmax_ukernel_5x8__neonfma_lane_ld64()
98 vacc2x4567 = vfmaq_lane_f32(vacc2x4567, vb4567c0, va2, 0); in xnn_f32_gemminc_minmax_ukernel_5x8__neonfma_lane_ld64()
99 vacc3x4567 = vfmaq_lane_f32(vacc3x4567, vb4567c0, va3, 0); in xnn_f32_gemminc_minmax_ukernel_5x8__neonfma_lane_ld64()
100 vacc4x4567 = vfmaq_lane_f32(vacc4x4567, vb4567c0, va4, 0); in xnn_f32_gemminc_minmax_ukernel_5x8__neonfma_lane_ld64()
|
| D | 6x8inc-minmax-neon-dup-ld64.c | 98 const float32x4_t vb4567c0 = vld1q_f32(w); w += 4; in xnn_f32_gemminc_minmax_ukernel_6x8__neon_dup_ld64() local
112 vacc0x4567 = vmlaq_f32(vacc0x4567, va0c0, vb4567c0); in xnn_f32_gemminc_minmax_ukernel_6x8__neon_dup_ld64()
113 vacc1x4567 = vmlaq_f32(vacc1x4567, va1c0, vb4567c0); in xnn_f32_gemminc_minmax_ukernel_6x8__neon_dup_ld64()
114 vacc2x4567 = vmlaq_f32(vacc2x4567, va2c0, vb4567c0); in xnn_f32_gemminc_minmax_ukernel_6x8__neon_dup_ld64()
115 vacc3x4567 = vmlaq_f32(vacc3x4567, va3c0, vb4567c0); in xnn_f32_gemminc_minmax_ukernel_6x8__neon_dup_ld64()
116 vacc4x4567 = vmlaq_f32(vacc4x4567, va4c0, vb4567c0); in xnn_f32_gemminc_minmax_ukernel_6x8__neon_dup_ld64()
117 vacc5x4567 = vmlaq_f32(vacc5x4567, va5c0, vb4567c0); in xnn_f32_gemminc_minmax_ukernel_6x8__neon_dup_ld64()
|
| D | 6x8inc-minmax-neonfma-dup-ld64.c | 98 const float32x4_t vb4567c0 = vld1q_f32(w); w += 4; in xnn_f32_gemminc_minmax_ukernel_6x8__neonfma_dup_ld64() local
112 vacc0x4567 = vfmaq_f32(vacc0x4567, va0c0, vb4567c0); in xnn_f32_gemminc_minmax_ukernel_6x8__neonfma_dup_ld64()
113 vacc1x4567 = vfmaq_f32(vacc1x4567, va1c0, vb4567c0); in xnn_f32_gemminc_minmax_ukernel_6x8__neonfma_dup_ld64()
114 vacc2x4567 = vfmaq_f32(vacc2x4567, va2c0, vb4567c0); in xnn_f32_gemminc_minmax_ukernel_6x8__neonfma_dup_ld64()
115 vacc3x4567 = vfmaq_f32(vacc3x4567, va3c0, vb4567c0); in xnn_f32_gemminc_minmax_ukernel_6x8__neonfma_dup_ld64()
116 vacc4x4567 = vfmaq_f32(vacc4x4567, va4c0, vb4567c0); in xnn_f32_gemminc_minmax_ukernel_6x8__neonfma_dup_ld64()
117 vacc5x4567 = vfmaq_f32(vacc5x4567, va5c0, vb4567c0); in xnn_f32_gemminc_minmax_ukernel_6x8__neonfma_dup_ld64()
|
| D | 4x8inc-minmax-neonfma-lane-ld64.c | 80 const float32x4_t vb4567c0 = vld1q_f32(w); w += 4; in xnn_f32_gemminc_minmax_ukernel_4x8__neonfma_lane_ld64() local
86 vacc0x4567 = vfmaq_lane_f32(vacc0x4567, vb4567c0, va0, 0); in xnn_f32_gemminc_minmax_ukernel_4x8__neonfma_lane_ld64()
87 vacc1x4567 = vfmaq_lane_f32(vacc1x4567, vb4567c0, va1, 0); in xnn_f32_gemminc_minmax_ukernel_4x8__neonfma_lane_ld64()
88 vacc2x4567 = vfmaq_lane_f32(vacc2x4567, vb4567c0, va2, 0); in xnn_f32_gemminc_minmax_ukernel_4x8__neonfma_lane_ld64()
89 vacc3x4567 = vfmaq_lane_f32(vacc3x4567, vb4567c0, va3, 0); in xnn_f32_gemminc_minmax_ukernel_4x8__neonfma_lane_ld64()
|
| D | 4x8inc-minmax-neon-dup-ld64.c | 80 const float32x4_t vb4567c0 = vld1q_f32(w); w += 4; in xnn_f32_gemminc_minmax_ukernel_4x8__neon_dup_ld64() local
90 vacc0x4567 = vmlaq_f32(vacc0x4567, va0c0, vb4567c0); in xnn_f32_gemminc_minmax_ukernel_4x8__neon_dup_ld64()
91 vacc1x4567 = vmlaq_f32(vacc1x4567, va1c0, vb4567c0); in xnn_f32_gemminc_minmax_ukernel_4x8__neon_dup_ld64()
92 vacc2x4567 = vmlaq_f32(vacc2x4567, va2c0, vb4567c0); in xnn_f32_gemminc_minmax_ukernel_4x8__neon_dup_ld64()
93 vacc3x4567 = vmlaq_f32(vacc3x4567, va3c0, vb4567c0); in xnn_f32_gemminc_minmax_ukernel_4x8__neon_dup_ld64()
|