/external/XNNPACK/src/f32-maxpool/ |
D | 9p8x-minmax-neon-c4.c | 97 const float32x4_t vmax018 = vmaxq_f32(vmaxq_f32(vi0, vi1), vi8); in xnn_f32_maxpool_minmax_ukernel_9p8x__neon_c4() 98 const float32x4_t vmax23 = vmaxq_f32(vi2, vi3); in xnn_f32_maxpool_minmax_ukernel_9p8x__neon_c4() 99 const float32x4_t vmax45 = vmaxq_f32(vi4, vi5); in xnn_f32_maxpool_minmax_ukernel_9p8x__neon_c4() 100 const float32x4_t vmax67 = vmaxq_f32(vi6, vi7); in xnn_f32_maxpool_minmax_ukernel_9p8x__neon_c4() 102 const float32x4_t vmax2345 = vmaxq_f32(vmax23, vmax45); in xnn_f32_maxpool_minmax_ukernel_9p8x__neon_c4() 103 const float32x4_t vmax01678 = vmaxq_f32(vmax018, vmax67); in xnn_f32_maxpool_minmax_ukernel_9p8x__neon_c4() 104 const float32x4_t vmax = vmaxq_f32(vmax2345, vmax01678); in xnn_f32_maxpool_minmax_ukernel_9p8x__neon_c4() 105 const float32x4_t vout = vmaxq_f32(vminq_f32(vmax, voutput_max), voutput_min); in xnn_f32_maxpool_minmax_ukernel_9p8x__neon_c4() 130 const float32x4_t vmax018 = vmaxq_f32(vmaxq_f32(vi0, vi1), vi8); in xnn_f32_maxpool_minmax_ukernel_9p8x__neon_c4() 131 const float32x4_t vmax23 = vmaxq_f32(vi2, vi3); in xnn_f32_maxpool_minmax_ukernel_9p8x__neon_c4() [all …]
|
/external/XNNPACK/src/f32-spmm/gen/ |
D | 32x4-minmax-neonfma.c | 157 vout0123n0 = vmaxq_f32(vout0123n0, vmin); in xnn_f32_spmm_minmax_ukernel_32x4__neonfma() 158 vout4567n0 = vmaxq_f32(vout4567n0, vmin); in xnn_f32_spmm_minmax_ukernel_32x4__neonfma() 159 vout89ABn0 = vmaxq_f32(vout89ABn0, vmin); in xnn_f32_spmm_minmax_ukernel_32x4__neonfma() 160 voutCDEFn0 = vmaxq_f32(voutCDEFn0, vmin); in xnn_f32_spmm_minmax_ukernel_32x4__neonfma() 161 voutGHIJn0 = vmaxq_f32(voutGHIJn0, vmin); in xnn_f32_spmm_minmax_ukernel_32x4__neonfma() 162 voutKLMNn0 = vmaxq_f32(voutKLMNn0, vmin); in xnn_f32_spmm_minmax_ukernel_32x4__neonfma() 163 voutOPQRn0 = vmaxq_f32(voutOPQRn0, vmin); in xnn_f32_spmm_minmax_ukernel_32x4__neonfma() 164 voutSTUVn0 = vmaxq_f32(voutSTUVn0, vmin); in xnn_f32_spmm_minmax_ukernel_32x4__neonfma() 165 vout0123n1 = vmaxq_f32(vout0123n1, vmin); in xnn_f32_spmm_minmax_ukernel_32x4__neonfma() 166 vout4567n1 = vmaxq_f32(vout4567n1, vmin); in xnn_f32_spmm_minmax_ukernel_32x4__neonfma() [all …]
|
D | 32x2-minmax-neonfma.c | 109 vout0123n0 = vmaxq_f32(vout0123n0, vmin); in xnn_f32_spmm_minmax_ukernel_32x2__neonfma() 110 vout4567n0 = vmaxq_f32(vout4567n0, vmin); in xnn_f32_spmm_minmax_ukernel_32x2__neonfma() 111 vout89ABn0 = vmaxq_f32(vout89ABn0, vmin); in xnn_f32_spmm_minmax_ukernel_32x2__neonfma() 112 voutCDEFn0 = vmaxq_f32(voutCDEFn0, vmin); in xnn_f32_spmm_minmax_ukernel_32x2__neonfma() 113 voutGHIJn0 = vmaxq_f32(voutGHIJn0, vmin); in xnn_f32_spmm_minmax_ukernel_32x2__neonfma() 114 voutKLMNn0 = vmaxq_f32(voutKLMNn0, vmin); in xnn_f32_spmm_minmax_ukernel_32x2__neonfma() 115 voutOPQRn0 = vmaxq_f32(voutOPQRn0, vmin); in xnn_f32_spmm_minmax_ukernel_32x2__neonfma() 116 voutSTUVn0 = vmaxq_f32(voutSTUVn0, vmin); in xnn_f32_spmm_minmax_ukernel_32x2__neonfma() 117 vout0123n1 = vmaxq_f32(vout0123n1, vmin); in xnn_f32_spmm_minmax_ukernel_32x2__neonfma() 118 vout4567n1 = vmaxq_f32(vout4567n1, vmin); in xnn_f32_spmm_minmax_ukernel_32x2__neonfma() [all …]
|
D | 16x4-minmax-neonfma.c | 104 vout0123n0 = vmaxq_f32(vout0123n0, vmin); in xnn_f32_spmm_minmax_ukernel_16x4__neonfma() 105 vout4567n0 = vmaxq_f32(vout4567n0, vmin); in xnn_f32_spmm_minmax_ukernel_16x4__neonfma() 106 vout89ABn0 = vmaxq_f32(vout89ABn0, vmin); in xnn_f32_spmm_minmax_ukernel_16x4__neonfma() 107 voutCDEFn0 = vmaxq_f32(voutCDEFn0, vmin); in xnn_f32_spmm_minmax_ukernel_16x4__neonfma() 108 vout0123n1 = vmaxq_f32(vout0123n1, vmin); in xnn_f32_spmm_minmax_ukernel_16x4__neonfma() 109 vout4567n1 = vmaxq_f32(vout4567n1, vmin); in xnn_f32_spmm_minmax_ukernel_16x4__neonfma() 110 vout89ABn1 = vmaxq_f32(vout89ABn1, vmin); in xnn_f32_spmm_minmax_ukernel_16x4__neonfma() 111 voutCDEFn1 = vmaxq_f32(voutCDEFn1, vmin); in xnn_f32_spmm_minmax_ukernel_16x4__neonfma() 112 vout0123n2 = vmaxq_f32(vout0123n2, vmin); in xnn_f32_spmm_minmax_ukernel_16x4__neonfma() 113 vout4567n2 = vmaxq_f32(vout4567n2, vmin); in xnn_f32_spmm_minmax_ukernel_16x4__neonfma() [all …]
|
D | 12x4-minmax-neonfma.c | 91 vout0123n0 = vmaxq_f32(vout0123n0, vmin); in xnn_f32_spmm_minmax_ukernel_12x4__neonfma() 92 vout4567n0 = vmaxq_f32(vout4567n0, vmin); in xnn_f32_spmm_minmax_ukernel_12x4__neonfma() 93 vout89ABn0 = vmaxq_f32(vout89ABn0, vmin); in xnn_f32_spmm_minmax_ukernel_12x4__neonfma() 94 vout0123n1 = vmaxq_f32(vout0123n1, vmin); in xnn_f32_spmm_minmax_ukernel_12x4__neonfma() 95 vout4567n1 = vmaxq_f32(vout4567n1, vmin); in xnn_f32_spmm_minmax_ukernel_12x4__neonfma() 96 vout89ABn1 = vmaxq_f32(vout89ABn1, vmin); in xnn_f32_spmm_minmax_ukernel_12x4__neonfma() 97 vout0123n2 = vmaxq_f32(vout0123n2, vmin); in xnn_f32_spmm_minmax_ukernel_12x4__neonfma() 98 vout4567n2 = vmaxq_f32(vout4567n2, vmin); in xnn_f32_spmm_minmax_ukernel_12x4__neonfma() 99 vout89ABn2 = vmaxq_f32(vout89ABn2, vmin); in xnn_f32_spmm_minmax_ukernel_12x4__neonfma() 100 vout0123n3 = vmaxq_f32(vout0123n3, vmin); in xnn_f32_spmm_minmax_ukernel_12x4__neonfma() [all …]
|
D | 32x1-minmax-neon.c | 84 vout0123 = vmaxq_f32(vout0123, vmin); in xnn_f32_spmm_minmax_ukernel_32x1__neon() 85 vout4567 = vmaxq_f32(vout4567, vmin); in xnn_f32_spmm_minmax_ukernel_32x1__neon() 86 vout89AB = vmaxq_f32(vout89AB, vmin); in xnn_f32_spmm_minmax_ukernel_32x1__neon() 87 voutCDEF = vmaxq_f32(voutCDEF, vmin); in xnn_f32_spmm_minmax_ukernel_32x1__neon() 88 voutGHIJ = vmaxq_f32(voutGHIJ, vmin); in xnn_f32_spmm_minmax_ukernel_32x1__neon() 89 voutKLMN = vmaxq_f32(voutKLMN, vmin); in xnn_f32_spmm_minmax_ukernel_32x1__neon() 90 voutOPQR = vmaxq_f32(voutOPQR, vmin); in xnn_f32_spmm_minmax_ukernel_32x1__neon() 91 voutSTUV = vmaxq_f32(voutSTUV, vmin); in xnn_f32_spmm_minmax_ukernel_32x1__neon() 138 vout0123 = vmaxq_f32(vout0123, vmin); in xnn_f32_spmm_minmax_ukernel_32x1__neon() 139 vout4567 = vmaxq_f32(vout4567, vmin); in xnn_f32_spmm_minmax_ukernel_32x1__neon() [all …]
|
D | 32x1-minmax-neonfma.c | 84 vout0123 = vmaxq_f32(vout0123, vmin); in xnn_f32_spmm_minmax_ukernel_32x1__neonfma() 85 vout4567 = vmaxq_f32(vout4567, vmin); in xnn_f32_spmm_minmax_ukernel_32x1__neonfma() 86 vout89AB = vmaxq_f32(vout89AB, vmin); in xnn_f32_spmm_minmax_ukernel_32x1__neonfma() 87 voutCDEF = vmaxq_f32(voutCDEF, vmin); in xnn_f32_spmm_minmax_ukernel_32x1__neonfma() 88 voutGHIJ = vmaxq_f32(voutGHIJ, vmin); in xnn_f32_spmm_minmax_ukernel_32x1__neonfma() 89 voutKLMN = vmaxq_f32(voutKLMN, vmin); in xnn_f32_spmm_minmax_ukernel_32x1__neonfma() 90 voutOPQR = vmaxq_f32(voutOPQR, vmin); in xnn_f32_spmm_minmax_ukernel_32x1__neonfma() 91 voutSTUV = vmaxq_f32(voutSTUV, vmin); in xnn_f32_spmm_minmax_ukernel_32x1__neonfma() 138 vout0123 = vmaxq_f32(vout0123, vmin); in xnn_f32_spmm_minmax_ukernel_32x1__neonfma() 139 vout4567 = vmaxq_f32(vout4567, vmin); in xnn_f32_spmm_minmax_ukernel_32x1__neonfma() [all …]
|
D | 16x2-minmax-neonfma.c | 80 vout0123n0 = vmaxq_f32(vout0123n0, vmin); in xnn_f32_spmm_minmax_ukernel_16x2__neonfma() 81 vout4567n0 = vmaxq_f32(vout4567n0, vmin); in xnn_f32_spmm_minmax_ukernel_16x2__neonfma() 82 vout89ABn0 = vmaxq_f32(vout89ABn0, vmin); in xnn_f32_spmm_minmax_ukernel_16x2__neonfma() 83 voutCDEFn0 = vmaxq_f32(voutCDEFn0, vmin); in xnn_f32_spmm_minmax_ukernel_16x2__neonfma() 84 vout0123n1 = vmaxq_f32(vout0123n1, vmin); in xnn_f32_spmm_minmax_ukernel_16x2__neonfma() 85 vout4567n1 = vmaxq_f32(vout4567n1, vmin); in xnn_f32_spmm_minmax_ukernel_16x2__neonfma() 86 vout89ABn1 = vmaxq_f32(vout89ABn1, vmin); in xnn_f32_spmm_minmax_ukernel_16x2__neonfma() 87 voutCDEFn1 = vmaxq_f32(voutCDEFn1, vmin); in xnn_f32_spmm_minmax_ukernel_16x2__neonfma() 132 vout0123 = vmaxq_f32(vout0123, vmin); in xnn_f32_spmm_minmax_ukernel_16x2__neonfma() 133 vout4567 = vmaxq_f32(vout4567, vmin); in xnn_f32_spmm_minmax_ukernel_16x2__neonfma() [all …]
|
D | 32x1-minmax-neonfma-pipelined.c | 95 vout0123 = vmaxq_f32(vout0123, vmin); in xnn_f32_spmm_minmax_ukernel_32x1__neonfma_pipelined() 96 vout4567 = vmaxq_f32(vout4567, vmin); in xnn_f32_spmm_minmax_ukernel_32x1__neonfma_pipelined() 97 vout89AB = vmaxq_f32(vout89AB, vmin); in xnn_f32_spmm_minmax_ukernel_32x1__neonfma_pipelined() 98 voutCDEF = vmaxq_f32(voutCDEF, vmin); in xnn_f32_spmm_minmax_ukernel_32x1__neonfma_pipelined() 99 voutGHIJ = vmaxq_f32(voutGHIJ, vmin); in xnn_f32_spmm_minmax_ukernel_32x1__neonfma_pipelined() 100 voutKLMN = vmaxq_f32(voutKLMN, vmin); in xnn_f32_spmm_minmax_ukernel_32x1__neonfma_pipelined() 101 voutOPQR = vmaxq_f32(voutOPQR, vmin); in xnn_f32_spmm_minmax_ukernel_32x1__neonfma_pipelined() 102 voutSTUV = vmaxq_f32(voutSTUV, vmin); in xnn_f32_spmm_minmax_ukernel_32x1__neonfma_pipelined() 152 vout0123 = vmaxq_f32(vout0123, vmin); in xnn_f32_spmm_minmax_ukernel_32x1__neonfma_pipelined() 153 vout4567 = vmaxq_f32(vout4567, vmin); in xnn_f32_spmm_minmax_ukernel_32x1__neonfma_pipelined() [all …]
|
D | 32x1-minmax-neon-pipelined.c | 95 vout0123 = vmaxq_f32(vout0123, vmin); in xnn_f32_spmm_minmax_ukernel_32x1__neon_pipelined() 96 vout4567 = vmaxq_f32(vout4567, vmin); in xnn_f32_spmm_minmax_ukernel_32x1__neon_pipelined() 97 vout89AB = vmaxq_f32(vout89AB, vmin); in xnn_f32_spmm_minmax_ukernel_32x1__neon_pipelined() 98 voutCDEF = vmaxq_f32(voutCDEF, vmin); in xnn_f32_spmm_minmax_ukernel_32x1__neon_pipelined() 99 voutGHIJ = vmaxq_f32(voutGHIJ, vmin); in xnn_f32_spmm_minmax_ukernel_32x1__neon_pipelined() 100 voutKLMN = vmaxq_f32(voutKLMN, vmin); in xnn_f32_spmm_minmax_ukernel_32x1__neon_pipelined() 101 voutOPQR = vmaxq_f32(voutOPQR, vmin); in xnn_f32_spmm_minmax_ukernel_32x1__neon_pipelined() 102 voutSTUV = vmaxq_f32(voutSTUV, vmin); in xnn_f32_spmm_minmax_ukernel_32x1__neon_pipelined() 152 vout0123 = vmaxq_f32(vout0123, vmin); in xnn_f32_spmm_minmax_ukernel_32x1__neon_pipelined() 153 vout4567 = vmaxq_f32(vout4567, vmin); in xnn_f32_spmm_minmax_ukernel_32x1__neon_pipelined() [all …]
|
D | 12x2-minmax-neonfma.c | 73 vout0123n0 = vmaxq_f32(vout0123n0, vmin); in xnn_f32_spmm_minmax_ukernel_12x2__neonfma() 74 vout4567n0 = vmaxq_f32(vout4567n0, vmin); in xnn_f32_spmm_minmax_ukernel_12x2__neonfma() 75 vout89ABn0 = vmaxq_f32(vout89ABn0, vmin); in xnn_f32_spmm_minmax_ukernel_12x2__neonfma() 76 vout0123n1 = vmaxq_f32(vout0123n1, vmin); in xnn_f32_spmm_minmax_ukernel_12x2__neonfma() 77 vout4567n1 = vmaxq_f32(vout4567n1, vmin); in xnn_f32_spmm_minmax_ukernel_12x2__neonfma() 78 vout89ABn1 = vmaxq_f32(vout89ABn1, vmin); in xnn_f32_spmm_minmax_ukernel_12x2__neonfma() 117 vout0123 = vmaxq_f32(vout0123, vmin); in xnn_f32_spmm_minmax_ukernel_12x2__neonfma() 118 vout4567 = vmaxq_f32(vout4567, vmin); in xnn_f32_spmm_minmax_ukernel_12x2__neonfma() 119 vout89AB = vmaxq_f32(vout89AB, vmin); in xnn_f32_spmm_minmax_ukernel_12x2__neonfma() 164 vout0123n0 = vmaxq_f32(vout0123n0, vmin); in xnn_f32_spmm_minmax_ukernel_12x2__neonfma() [all …]
|
D | 8x4-minmax-neonfma.c | 78 vout0123n0 = vmaxq_f32(vout0123n0, vmin); in xnn_f32_spmm_minmax_ukernel_8x4__neonfma() 79 vout4567n0 = vmaxq_f32(vout4567n0, vmin); in xnn_f32_spmm_minmax_ukernel_8x4__neonfma() 80 vout0123n1 = vmaxq_f32(vout0123n1, vmin); in xnn_f32_spmm_minmax_ukernel_8x4__neonfma() 81 vout4567n1 = vmaxq_f32(vout4567n1, vmin); in xnn_f32_spmm_minmax_ukernel_8x4__neonfma() 82 vout0123n2 = vmaxq_f32(vout0123n2, vmin); in xnn_f32_spmm_minmax_ukernel_8x4__neonfma() 83 vout4567n2 = vmaxq_f32(vout4567n2, vmin); in xnn_f32_spmm_minmax_ukernel_8x4__neonfma() 84 vout0123n3 = vmaxq_f32(vout0123n3, vmin); in xnn_f32_spmm_minmax_ukernel_8x4__neonfma() 85 vout4567n3 = vmaxq_f32(vout4567n3, vmin); in xnn_f32_spmm_minmax_ukernel_8x4__neonfma() 124 vout0123 = vmaxq_f32(vout0123, vmin); in xnn_f32_spmm_minmax_ukernel_8x4__neonfma() 125 vout4567 = vmaxq_f32(vout4567, vmin); in xnn_f32_spmm_minmax_ukernel_8x4__neonfma() [all …]
|
D | 32x1-minmax-neon-x2.c | 155 vout0123 = vmaxq_f32(vout0123, vmin); in xnn_f32_spmm_minmax_ukernel_32x1__neon_x2() 156 vout4567 = vmaxq_f32(vout4567, vmin); in xnn_f32_spmm_minmax_ukernel_32x1__neon_x2() 157 vout89AB = vmaxq_f32(vout89AB, vmin); in xnn_f32_spmm_minmax_ukernel_32x1__neon_x2() 158 voutCDEF = vmaxq_f32(voutCDEF, vmin); in xnn_f32_spmm_minmax_ukernel_32x1__neon_x2() 159 voutGHIJ = vmaxq_f32(voutGHIJ, vmin); in xnn_f32_spmm_minmax_ukernel_32x1__neon_x2() 160 voutKLMN = vmaxq_f32(voutKLMN, vmin); in xnn_f32_spmm_minmax_ukernel_32x1__neon_x2() 161 voutOPQR = vmaxq_f32(voutOPQR, vmin); in xnn_f32_spmm_minmax_ukernel_32x1__neon_x2() 162 voutSTUV = vmaxq_f32(voutSTUV, vmin); in xnn_f32_spmm_minmax_ukernel_32x1__neon_x2() 209 vout0123 = vmaxq_f32(vout0123, vmin); in xnn_f32_spmm_minmax_ukernel_32x1__neon_x2() 210 vout4567 = vmaxq_f32(vout4567, vmin); in xnn_f32_spmm_minmax_ukernel_32x1__neon_x2() [all …]
|
/external/XNNPACK/src/f32-rmax/ |
D | neon.c | 32 vmax0 = vmaxq_f32(vmax0, vx0); in xnn_f32_rmax_ukernel__neon() 33 vmax1 = vmaxq_f32(vmax1, vx1); in xnn_f32_rmax_ukernel__neon() 34 vmax2 = vmaxq_f32(vmax2, vx2); in xnn_f32_rmax_ukernel__neon() 35 vmax3 = vmaxq_f32(vmax3, vx3); in xnn_f32_rmax_ukernel__neon() 37 float32x4_t vmax = vmaxq_f32(vmaxq_f32(vmax0, vmax1), vmaxq_f32(vmax2, vmax3)); in xnn_f32_rmax_ukernel__neon() 40 vmax = vmaxq_f32(vmax, vx); in xnn_f32_rmax_ukernel__neon()
|
/external/XNNPACK/src/f32-ppmm/gen/ |
D | 8x8-minmax-neon.c | 130 vacc0x0123 = vmaxq_f32(vacc0x0123, vmin); in xnn_f32_ppmm_minmax_ukernel_8x8__neon() 131 vacc1x0123 = vmaxq_f32(vacc1x0123, vmin); in xnn_f32_ppmm_minmax_ukernel_8x8__neon() 132 vacc2x0123 = vmaxq_f32(vacc2x0123, vmin); in xnn_f32_ppmm_minmax_ukernel_8x8__neon() 133 vacc3x0123 = vmaxq_f32(vacc3x0123, vmin); in xnn_f32_ppmm_minmax_ukernel_8x8__neon() 134 vacc4x0123 = vmaxq_f32(vacc4x0123, vmin); in xnn_f32_ppmm_minmax_ukernel_8x8__neon() 135 vacc5x0123 = vmaxq_f32(vacc5x0123, vmin); in xnn_f32_ppmm_minmax_ukernel_8x8__neon() 136 vacc6x0123 = vmaxq_f32(vacc6x0123, vmin); in xnn_f32_ppmm_minmax_ukernel_8x8__neon() 137 vacc7x0123 = vmaxq_f32(vacc7x0123, vmin); in xnn_f32_ppmm_minmax_ukernel_8x8__neon() 138 vacc0x4567 = vmaxq_f32(vacc0x4567, vmin); in xnn_f32_ppmm_minmax_ukernel_8x8__neon() 139 vacc1x4567 = vmaxq_f32(vacc1x4567, vmin); in xnn_f32_ppmm_minmax_ukernel_8x8__neon() [all …]
|
D | 8x8-minmax-neonfma.c | 158 vacc0x0123 = vmaxq_f32(vacc0x0123, vmin); in xnn_f32_ppmm_minmax_ukernel_8x8__neonfma() 159 vacc1x0123 = vmaxq_f32(vacc1x0123, vmin); in xnn_f32_ppmm_minmax_ukernel_8x8__neonfma() 160 vacc2x0123 = vmaxq_f32(vacc2x0123, vmin); in xnn_f32_ppmm_minmax_ukernel_8x8__neonfma() 161 vacc3x0123 = vmaxq_f32(vacc3x0123, vmin); in xnn_f32_ppmm_minmax_ukernel_8x8__neonfma() 162 vacc4x0123 = vmaxq_f32(vacc4x0123, vmin); in xnn_f32_ppmm_minmax_ukernel_8x8__neonfma() 163 vacc5x0123 = vmaxq_f32(vacc5x0123, vmin); in xnn_f32_ppmm_minmax_ukernel_8x8__neonfma() 164 vacc6x0123 = vmaxq_f32(vacc6x0123, vmin); in xnn_f32_ppmm_minmax_ukernel_8x8__neonfma() 165 vacc7x0123 = vmaxq_f32(vacc7x0123, vmin); in xnn_f32_ppmm_minmax_ukernel_8x8__neonfma() 166 vacc0x4567 = vmaxq_f32(vacc0x4567, vmin); in xnn_f32_ppmm_minmax_ukernel_8x8__neonfma() 167 vacc1x4567 = vmaxq_f32(vacc1x4567, vmin); in xnn_f32_ppmm_minmax_ukernel_8x8__neonfma() [all …]
|
D | 4x8-minmax-neon.c | 89 vacc0x0123 = vmaxq_f32(vacc0x0123, vmin); in xnn_f32_ppmm_minmax_ukernel_4x8__neon() 90 vacc1x0123 = vmaxq_f32(vacc1x0123, vmin); in xnn_f32_ppmm_minmax_ukernel_4x8__neon() 91 vacc2x0123 = vmaxq_f32(vacc2x0123, vmin); in xnn_f32_ppmm_minmax_ukernel_4x8__neon() 92 vacc3x0123 = vmaxq_f32(vacc3x0123, vmin); in xnn_f32_ppmm_minmax_ukernel_4x8__neon() 93 vacc0x4567 = vmaxq_f32(vacc0x4567, vmin); in xnn_f32_ppmm_minmax_ukernel_4x8__neon() 94 vacc1x4567 = vmaxq_f32(vacc1x4567, vmin); in xnn_f32_ppmm_minmax_ukernel_4x8__neon() 95 vacc2x4567 = vmaxq_f32(vacc2x4567, vmin); in xnn_f32_ppmm_minmax_ukernel_4x8__neon() 96 vacc3x4567 = vmaxq_f32(vacc3x4567, vmin); in xnn_f32_ppmm_minmax_ukernel_4x8__neon()
|
/external/XNNPACK/src/f32-vmulcaddc/gen/ |
D | c8-minmax-neonfma-2x.c | 67 vacc0x0123 = vmaxq_f32(vacc0x0123, vmin); in xnn_f32_vmulcaddc_minmax_ukernel_c8__neonfma_2x() 68 vacc0x4567 = vmaxq_f32(vacc0x4567, vmin); in xnn_f32_vmulcaddc_minmax_ukernel_c8__neonfma_2x() 69 vacc1x0123 = vmaxq_f32(vacc1x0123, vmin); in xnn_f32_vmulcaddc_minmax_ukernel_c8__neonfma_2x() 70 vacc1x4567 = vmaxq_f32(vacc1x4567, vmin); in xnn_f32_vmulcaddc_minmax_ukernel_c8__neonfma_2x() 94 vacc0x0123 = vmaxq_f32(vacc0x0123, vmin); in xnn_f32_vmulcaddc_minmax_ukernel_c8__neonfma_2x() 95 vacc1x0123 = vmaxq_f32(vacc1x0123, vmin); in xnn_f32_vmulcaddc_minmax_ukernel_c8__neonfma_2x() 115 vacc0x0123 = vmaxq_f32(vacc0x0123, vmin); in xnn_f32_vmulcaddc_minmax_ukernel_c8__neonfma_2x() 116 vacc1x0123 = vmaxq_f32(vacc1x0123, vmin); in xnn_f32_vmulcaddc_minmax_ukernel_c8__neonfma_2x()
|
D | c8-minmax-neon-2x.c | 71 vacc0x0123 = vmaxq_f32(vacc0x0123, vmin); in xnn_f32_vmulcaddc_minmax_ukernel_c8__neon_2x() 72 vacc0x4567 = vmaxq_f32(vacc0x4567, vmin); in xnn_f32_vmulcaddc_minmax_ukernel_c8__neon_2x() 73 vacc1x0123 = vmaxq_f32(vacc1x0123, vmin); in xnn_f32_vmulcaddc_minmax_ukernel_c8__neon_2x() 74 vacc1x4567 = vmaxq_f32(vacc1x4567, vmin); in xnn_f32_vmulcaddc_minmax_ukernel_c8__neon_2x() 100 vacc0x0123 = vmaxq_f32(vacc0x0123, vmin); in xnn_f32_vmulcaddc_minmax_ukernel_c8__neon_2x() 101 vacc1x0123 = vmaxq_f32(vacc1x0123, vmin); in xnn_f32_vmulcaddc_minmax_ukernel_c8__neon_2x() 123 vacc0x0123 = vmaxq_f32(vacc0x0123, vmin); in xnn_f32_vmulcaddc_minmax_ukernel_c8__neon_2x() 124 vacc1x0123 = vmaxq_f32(vacc1x0123, vmin); in xnn_f32_vmulcaddc_minmax_ukernel_c8__neon_2x()
|
/external/XNNPACK/src/f32-igemm/gen/ |
D | 6x8-minmax-neon-lane-ld64.c | 197 vacc0x0123 = vmaxq_f32(vacc0x0123, vmin); in xnn_f32_igemm_minmax_ukernel_6x8__neon_lane_ld64() 198 vacc1x0123 = vmaxq_f32(vacc1x0123, vmin); in xnn_f32_igemm_minmax_ukernel_6x8__neon_lane_ld64() 199 vacc2x0123 = vmaxq_f32(vacc2x0123, vmin); in xnn_f32_igemm_minmax_ukernel_6x8__neon_lane_ld64() 200 vacc3x0123 = vmaxq_f32(vacc3x0123, vmin); in xnn_f32_igemm_minmax_ukernel_6x8__neon_lane_ld64() 201 vacc4x0123 = vmaxq_f32(vacc4x0123, vmin); in xnn_f32_igemm_minmax_ukernel_6x8__neon_lane_ld64() 202 vacc5x0123 = vmaxq_f32(vacc5x0123, vmin); in xnn_f32_igemm_minmax_ukernel_6x8__neon_lane_ld64() 203 vacc0x4567 = vmaxq_f32(vacc0x4567, vmin); in xnn_f32_igemm_minmax_ukernel_6x8__neon_lane_ld64() 204 vacc1x4567 = vmaxq_f32(vacc1x4567, vmin); in xnn_f32_igemm_minmax_ukernel_6x8__neon_lane_ld64() 205 vacc2x4567 = vmaxq_f32(vacc2x4567, vmin); in xnn_f32_igemm_minmax_ukernel_6x8__neon_lane_ld64() 206 vacc3x4567 = vmaxq_f32(vacc3x4567, vmin); in xnn_f32_igemm_minmax_ukernel_6x8__neon_lane_ld64() [all …]
|
D | 6x8-minmax-neonfma-lane-ld64.c | 197 vacc0x0123 = vmaxq_f32(vacc0x0123, vmin); in xnn_f32_igemm_minmax_ukernel_6x8__neonfma_lane_ld64() 198 vacc1x0123 = vmaxq_f32(vacc1x0123, vmin); in xnn_f32_igemm_minmax_ukernel_6x8__neonfma_lane_ld64() 199 vacc2x0123 = vmaxq_f32(vacc2x0123, vmin); in xnn_f32_igemm_minmax_ukernel_6x8__neonfma_lane_ld64() 200 vacc3x0123 = vmaxq_f32(vacc3x0123, vmin); in xnn_f32_igemm_minmax_ukernel_6x8__neonfma_lane_ld64() 201 vacc4x0123 = vmaxq_f32(vacc4x0123, vmin); in xnn_f32_igemm_minmax_ukernel_6x8__neonfma_lane_ld64() 202 vacc5x0123 = vmaxq_f32(vacc5x0123, vmin); in xnn_f32_igemm_minmax_ukernel_6x8__neonfma_lane_ld64() 203 vacc0x4567 = vmaxq_f32(vacc0x4567, vmin); in xnn_f32_igemm_minmax_ukernel_6x8__neonfma_lane_ld64() 204 vacc1x4567 = vmaxq_f32(vacc1x4567, vmin); in xnn_f32_igemm_minmax_ukernel_6x8__neonfma_lane_ld64() 205 vacc2x4567 = vmaxq_f32(vacc2x4567, vmin); in xnn_f32_igemm_minmax_ukernel_6x8__neonfma_lane_ld64() 206 vacc3x4567 = vmaxq_f32(vacc3x4567, vmin); in xnn_f32_igemm_minmax_ukernel_6x8__neonfma_lane_ld64() [all …]
|
/external/XNNPACK/src/f32-gemm/gen/ |
D | 6x8-minmax-neon-lane-ld64.c | 165 vacc0x0123 = vmaxq_f32(vacc0x0123, vmin); in xnn_f32_gemm_minmax_ukernel_6x8__neon_lane_ld64() 166 vacc1x0123 = vmaxq_f32(vacc1x0123, vmin); in xnn_f32_gemm_minmax_ukernel_6x8__neon_lane_ld64() 167 vacc2x0123 = vmaxq_f32(vacc2x0123, vmin); in xnn_f32_gemm_minmax_ukernel_6x8__neon_lane_ld64() 168 vacc3x0123 = vmaxq_f32(vacc3x0123, vmin); in xnn_f32_gemm_minmax_ukernel_6x8__neon_lane_ld64() 169 vacc4x0123 = vmaxq_f32(vacc4x0123, vmin); in xnn_f32_gemm_minmax_ukernel_6x8__neon_lane_ld64() 170 vacc5x0123 = vmaxq_f32(vacc5x0123, vmin); in xnn_f32_gemm_minmax_ukernel_6x8__neon_lane_ld64() 171 vacc0x4567 = vmaxq_f32(vacc0x4567, vmin); in xnn_f32_gemm_minmax_ukernel_6x8__neon_lane_ld64() 172 vacc1x4567 = vmaxq_f32(vacc1x4567, vmin); in xnn_f32_gemm_minmax_ukernel_6x8__neon_lane_ld64() 173 vacc2x4567 = vmaxq_f32(vacc2x4567, vmin); in xnn_f32_gemm_minmax_ukernel_6x8__neon_lane_ld64() 174 vacc3x4567 = vmaxq_f32(vacc3x4567, vmin); in xnn_f32_gemm_minmax_ukernel_6x8__neon_lane_ld64() [all …]
|
D | 6x8-minmax-neonfma-lane-ld64.c | 165 vacc0x0123 = vmaxq_f32(vacc0x0123, vmin); in xnn_f32_gemm_minmax_ukernel_6x8__neonfma_lane_ld64() 166 vacc1x0123 = vmaxq_f32(vacc1x0123, vmin); in xnn_f32_gemm_minmax_ukernel_6x8__neonfma_lane_ld64() 167 vacc2x0123 = vmaxq_f32(vacc2x0123, vmin); in xnn_f32_gemm_minmax_ukernel_6x8__neonfma_lane_ld64() 168 vacc3x0123 = vmaxq_f32(vacc3x0123, vmin); in xnn_f32_gemm_minmax_ukernel_6x8__neonfma_lane_ld64() 169 vacc4x0123 = vmaxq_f32(vacc4x0123, vmin); in xnn_f32_gemm_minmax_ukernel_6x8__neonfma_lane_ld64() 170 vacc5x0123 = vmaxq_f32(vacc5x0123, vmin); in xnn_f32_gemm_minmax_ukernel_6x8__neonfma_lane_ld64() 171 vacc0x4567 = vmaxq_f32(vacc0x4567, vmin); in xnn_f32_gemm_minmax_ukernel_6x8__neonfma_lane_ld64() 172 vacc1x4567 = vmaxq_f32(vacc1x4567, vmin); in xnn_f32_gemm_minmax_ukernel_6x8__neonfma_lane_ld64() 173 vacc2x4567 = vmaxq_f32(vacc2x4567, vmin); in xnn_f32_gemm_minmax_ukernel_6x8__neonfma_lane_ld64() 174 vacc3x4567 = vmaxq_f32(vacc3x4567, vmin); in xnn_f32_gemm_minmax_ukernel_6x8__neonfma_lane_ld64() [all …]
|
/external/XNNPACK/src/f32-gemm/gen-inc/ |
D | 6x8inc-minmax-neon-lane-ld64.c | 167 vacc0x0123 = vmaxq_f32(vacc0x0123, vmin); in xnn_f32_gemminc_minmax_ukernel_6x8__neon_lane_ld64() 168 vacc1x0123 = vmaxq_f32(vacc1x0123, vmin); in xnn_f32_gemminc_minmax_ukernel_6x8__neon_lane_ld64() 169 vacc2x0123 = vmaxq_f32(vacc2x0123, vmin); in xnn_f32_gemminc_minmax_ukernel_6x8__neon_lane_ld64() 170 vacc3x0123 = vmaxq_f32(vacc3x0123, vmin); in xnn_f32_gemminc_minmax_ukernel_6x8__neon_lane_ld64() 171 vacc4x0123 = vmaxq_f32(vacc4x0123, vmin); in xnn_f32_gemminc_minmax_ukernel_6x8__neon_lane_ld64() 172 vacc5x0123 = vmaxq_f32(vacc5x0123, vmin); in xnn_f32_gemminc_minmax_ukernel_6x8__neon_lane_ld64() 173 vacc0x4567 = vmaxq_f32(vacc0x4567, vmin); in xnn_f32_gemminc_minmax_ukernel_6x8__neon_lane_ld64() 174 vacc1x4567 = vmaxq_f32(vacc1x4567, vmin); in xnn_f32_gemminc_minmax_ukernel_6x8__neon_lane_ld64() 175 vacc2x4567 = vmaxq_f32(vacc2x4567, vmin); in xnn_f32_gemminc_minmax_ukernel_6x8__neon_lane_ld64() 176 vacc3x4567 = vmaxq_f32(vacc3x4567, vmin); in xnn_f32_gemminc_minmax_ukernel_6x8__neon_lane_ld64() [all …]
|
D | 6x8inc-minmax-neonfma-lane-ld64.c | 167 vacc0x0123 = vmaxq_f32(vacc0x0123, vmin); in xnn_f32_gemminc_minmax_ukernel_6x8__neonfma_lane_ld64() 168 vacc1x0123 = vmaxq_f32(vacc1x0123, vmin); in xnn_f32_gemminc_minmax_ukernel_6x8__neonfma_lane_ld64() 169 vacc2x0123 = vmaxq_f32(vacc2x0123, vmin); in xnn_f32_gemminc_minmax_ukernel_6x8__neonfma_lane_ld64() 170 vacc3x0123 = vmaxq_f32(vacc3x0123, vmin); in xnn_f32_gemminc_minmax_ukernel_6x8__neonfma_lane_ld64() 171 vacc4x0123 = vmaxq_f32(vacc4x0123, vmin); in xnn_f32_gemminc_minmax_ukernel_6x8__neonfma_lane_ld64() 172 vacc5x0123 = vmaxq_f32(vacc5x0123, vmin); in xnn_f32_gemminc_minmax_ukernel_6x8__neonfma_lane_ld64() 173 vacc0x4567 = vmaxq_f32(vacc0x4567, vmin); in xnn_f32_gemminc_minmax_ukernel_6x8__neonfma_lane_ld64() 174 vacc1x4567 = vmaxq_f32(vacc1x4567, vmin); in xnn_f32_gemminc_minmax_ukernel_6x8__neonfma_lane_ld64() 175 vacc2x4567 = vmaxq_f32(vacc2x4567, vmin); in xnn_f32_gemminc_minmax_ukernel_6x8__neonfma_lane_ld64() 176 vacc3x4567 = vmaxq_f32(vacc3x4567, vmin); in xnn_f32_gemminc_minmax_ukernel_6x8__neonfma_lane_ld64() [all …]
|