| /external/XNNPACK/src/qs8-gemm/gen/ |
| D | 4x16-minmax-neon-mlal-lane.c | 92 vacc0x0123 = vmlal_lane_s16(vacc0x0123, vget_low_s16(vxb01234567c0), vget_low_s16(vxa0), 0); in xnn_qs8_gemm_minmax_ukernel_4x16__neon_mlal_lane()
93 vacc0x4567 = vmlal_lane_s16(vacc0x4567, vget_high_s16(vxb01234567c0), vget_low_s16(vxa0), 0); in xnn_qs8_gemm_minmax_ukernel_4x16__neon_mlal_lane()
94 vacc1x0123 = vmlal_lane_s16(vacc1x0123, vget_low_s16(vxb01234567c0), vget_low_s16(vxa1), 0); in xnn_qs8_gemm_minmax_ukernel_4x16__neon_mlal_lane()
95 vacc1x4567 = vmlal_lane_s16(vacc1x4567, vget_high_s16(vxb01234567c0), vget_low_s16(vxa1), 0); in xnn_qs8_gemm_minmax_ukernel_4x16__neon_mlal_lane()
96 vacc2x0123 = vmlal_lane_s16(vacc2x0123, vget_low_s16(vxb01234567c0), vget_low_s16(vxa2), 0); in xnn_qs8_gemm_minmax_ukernel_4x16__neon_mlal_lane()
97 vacc2x4567 = vmlal_lane_s16(vacc2x4567, vget_high_s16(vxb01234567c0), vget_low_s16(vxa2), 0); in xnn_qs8_gemm_minmax_ukernel_4x16__neon_mlal_lane()
98 vacc3x0123 = vmlal_lane_s16(vacc3x0123, vget_low_s16(vxb01234567c0), vget_low_s16(vxa3), 0); in xnn_qs8_gemm_minmax_ukernel_4x16__neon_mlal_lane()
99 vacc3x4567 = vmlal_lane_s16(vacc3x4567, vget_high_s16(vxb01234567c0), vget_low_s16(vxa3), 0); in xnn_qs8_gemm_minmax_ukernel_4x16__neon_mlal_lane()
103 vacc0x89AB = vmlal_lane_s16(vacc0x89AB, vget_low_s16(vxb89ABCDEFc0), vget_low_s16(vxa0), 0); in xnn_qs8_gemm_minmax_ukernel_4x16__neon_mlal_lane()
104 vacc0xCDEF = vmlal_lane_s16(vacc0xCDEF, vget_high_s16(vxb89ABCDEFc0), vget_low_s16(vxa0), 0); in xnn_qs8_gemm_minmax_ukernel_4x16__neon_mlal_lane()
[all …]
|
| D | 3x16-minmax-neon-mlal-lane.c | 80 vacc0x0123 = vmlal_lane_s16(vacc0x0123, vget_low_s16(vxb01234567c0), vget_low_s16(vxa0), 0); in xnn_qs8_gemm_minmax_ukernel_3x16__neon_mlal_lane()
81 vacc0x4567 = vmlal_lane_s16(vacc0x4567, vget_high_s16(vxb01234567c0), vget_low_s16(vxa0), 0); in xnn_qs8_gemm_minmax_ukernel_3x16__neon_mlal_lane()
82 vacc1x0123 = vmlal_lane_s16(vacc1x0123, vget_low_s16(vxb01234567c0), vget_low_s16(vxa1), 0); in xnn_qs8_gemm_minmax_ukernel_3x16__neon_mlal_lane()
83 vacc1x4567 = vmlal_lane_s16(vacc1x4567, vget_high_s16(vxb01234567c0), vget_low_s16(vxa1), 0); in xnn_qs8_gemm_minmax_ukernel_3x16__neon_mlal_lane()
84 vacc2x0123 = vmlal_lane_s16(vacc2x0123, vget_low_s16(vxb01234567c0), vget_low_s16(vxa2), 0); in xnn_qs8_gemm_minmax_ukernel_3x16__neon_mlal_lane()
85 vacc2x4567 = vmlal_lane_s16(vacc2x4567, vget_high_s16(vxb01234567c0), vget_low_s16(vxa2), 0); in xnn_qs8_gemm_minmax_ukernel_3x16__neon_mlal_lane()
89 vacc0x89AB = vmlal_lane_s16(vacc0x89AB, vget_low_s16(vxb89ABCDEFc0), vget_low_s16(vxa0), 0); in xnn_qs8_gemm_minmax_ukernel_3x16__neon_mlal_lane()
90 vacc0xCDEF = vmlal_lane_s16(vacc0xCDEF, vget_high_s16(vxb89ABCDEFc0), vget_low_s16(vxa0), 0); in xnn_qs8_gemm_minmax_ukernel_3x16__neon_mlal_lane()
91 vacc1x89AB = vmlal_lane_s16(vacc1x89AB, vget_low_s16(vxb89ABCDEFc0), vget_low_s16(vxa1), 0); in xnn_qs8_gemm_minmax_ukernel_3x16__neon_mlal_lane()
92 vacc1xCDEF = vmlal_lane_s16(vacc1xCDEF, vget_high_s16(vxb89ABCDEFc0), vget_low_s16(vxa1), 0); in xnn_qs8_gemm_minmax_ukernel_3x16__neon_mlal_lane()
[all …]
|
| D | 4x8-minmax-neon-mlal-lane.c | 84 vacc0x0123 = vmlal_lane_s16(vacc0x0123, vget_low_s16(vxb01234567c0), vget_low_s16(vxa0), 0); in xnn_qs8_gemm_minmax_ukernel_4x8__neon_mlal_lane()
85 vacc0x4567 = vmlal_lane_s16(vacc0x4567, vget_high_s16(vxb01234567c0), vget_low_s16(vxa0), 0); in xnn_qs8_gemm_minmax_ukernel_4x8__neon_mlal_lane()
86 vacc1x0123 = vmlal_lane_s16(vacc1x0123, vget_low_s16(vxb01234567c0), vget_low_s16(vxa1), 0); in xnn_qs8_gemm_minmax_ukernel_4x8__neon_mlal_lane()
87 vacc1x4567 = vmlal_lane_s16(vacc1x4567, vget_high_s16(vxb01234567c0), vget_low_s16(vxa1), 0); in xnn_qs8_gemm_minmax_ukernel_4x8__neon_mlal_lane()
88 vacc2x0123 = vmlal_lane_s16(vacc2x0123, vget_low_s16(vxb01234567c0), vget_low_s16(vxa2), 0); in xnn_qs8_gemm_minmax_ukernel_4x8__neon_mlal_lane()
89 vacc2x4567 = vmlal_lane_s16(vacc2x4567, vget_high_s16(vxb01234567c0), vget_low_s16(vxa2), 0); in xnn_qs8_gemm_minmax_ukernel_4x8__neon_mlal_lane()
90 vacc3x0123 = vmlal_lane_s16(vacc3x0123, vget_low_s16(vxb01234567c0), vget_low_s16(vxa3), 0); in xnn_qs8_gemm_minmax_ukernel_4x8__neon_mlal_lane()
91 vacc3x4567 = vmlal_lane_s16(vacc3x4567, vget_high_s16(vxb01234567c0), vget_low_s16(vxa3), 0); in xnn_qs8_gemm_minmax_ukernel_4x8__neon_mlal_lane()
95 vacc0x0123 = vmlal_lane_s16(vacc0x0123, vget_low_s16(vxb01234567c1), vget_low_s16(vxa0), 1); in xnn_qs8_gemm_minmax_ukernel_4x8__neon_mlal_lane()
96 vacc0x4567 = vmlal_lane_s16(vacc0x4567, vget_high_s16(vxb01234567c1), vget_low_s16(vxa0), 1); in xnn_qs8_gemm_minmax_ukernel_4x8__neon_mlal_lane()
[all …]
|
| D | 2x16-minmax-neon-mlal-lane.c | 68 vacc0x0123 = vmlal_lane_s16(vacc0x0123, vget_low_s16(vxb01234567c0), vget_low_s16(vxa0), 0); in xnn_qs8_gemm_minmax_ukernel_2x16__neon_mlal_lane()
69 vacc0x4567 = vmlal_lane_s16(vacc0x4567, vget_high_s16(vxb01234567c0), vget_low_s16(vxa0), 0); in xnn_qs8_gemm_minmax_ukernel_2x16__neon_mlal_lane()
70 vacc1x0123 = vmlal_lane_s16(vacc1x0123, vget_low_s16(vxb01234567c0), vget_low_s16(vxa1), 0); in xnn_qs8_gemm_minmax_ukernel_2x16__neon_mlal_lane()
71 vacc1x4567 = vmlal_lane_s16(vacc1x4567, vget_high_s16(vxb01234567c0), vget_low_s16(vxa1), 0); in xnn_qs8_gemm_minmax_ukernel_2x16__neon_mlal_lane()
75 vacc0x89AB = vmlal_lane_s16(vacc0x89AB, vget_low_s16(vxb89ABCDEFc0), vget_low_s16(vxa0), 0); in xnn_qs8_gemm_minmax_ukernel_2x16__neon_mlal_lane()
76 vacc0xCDEF = vmlal_lane_s16(vacc0xCDEF, vget_high_s16(vxb89ABCDEFc0), vget_low_s16(vxa0), 0); in xnn_qs8_gemm_minmax_ukernel_2x16__neon_mlal_lane()
77 vacc1x89AB = vmlal_lane_s16(vacc1x89AB, vget_low_s16(vxb89ABCDEFc0), vget_low_s16(vxa1), 0); in xnn_qs8_gemm_minmax_ukernel_2x16__neon_mlal_lane()
78 vacc1xCDEF = vmlal_lane_s16(vacc1xCDEF, vget_high_s16(vxb89ABCDEFc0), vget_low_s16(vxa1), 0); in xnn_qs8_gemm_minmax_ukernel_2x16__neon_mlal_lane()
82 vacc0x0123 = vmlal_lane_s16(vacc0x0123, vget_low_s16(vxb01234567c1), vget_low_s16(vxa0), 1); in xnn_qs8_gemm_minmax_ukernel_2x16__neon_mlal_lane()
83 vacc0x4567 = vmlal_lane_s16(vacc0x4567, vget_high_s16(vxb01234567c1), vget_low_s16(vxa0), 1); in xnn_qs8_gemm_minmax_ukernel_2x16__neon_mlal_lane()
[all …]
|
| D | 3x8-minmax-neon-mlal-lane.c | 74 vacc0x0123 = vmlal_lane_s16(vacc0x0123, vget_low_s16(vxb01234567c0), vget_low_s16(vxa0), 0); in xnn_qs8_gemm_minmax_ukernel_3x8__neon_mlal_lane()
75 vacc0x4567 = vmlal_lane_s16(vacc0x4567, vget_high_s16(vxb01234567c0), vget_low_s16(vxa0), 0); in xnn_qs8_gemm_minmax_ukernel_3x8__neon_mlal_lane()
76 vacc1x0123 = vmlal_lane_s16(vacc1x0123, vget_low_s16(vxb01234567c0), vget_low_s16(vxa1), 0); in xnn_qs8_gemm_minmax_ukernel_3x8__neon_mlal_lane()
77 vacc1x4567 = vmlal_lane_s16(vacc1x4567, vget_high_s16(vxb01234567c0), vget_low_s16(vxa1), 0); in xnn_qs8_gemm_minmax_ukernel_3x8__neon_mlal_lane()
78 vacc2x0123 = vmlal_lane_s16(vacc2x0123, vget_low_s16(vxb01234567c0), vget_low_s16(vxa2), 0); in xnn_qs8_gemm_minmax_ukernel_3x8__neon_mlal_lane()
79 vacc2x4567 = vmlal_lane_s16(vacc2x4567, vget_high_s16(vxb01234567c0), vget_low_s16(vxa2), 0); in xnn_qs8_gemm_minmax_ukernel_3x8__neon_mlal_lane()
83 vacc0x0123 = vmlal_lane_s16(vacc0x0123, vget_low_s16(vxb01234567c1), vget_low_s16(vxa0), 1); in xnn_qs8_gemm_minmax_ukernel_3x8__neon_mlal_lane()
84 vacc0x4567 = vmlal_lane_s16(vacc0x4567, vget_high_s16(vxb01234567c1), vget_low_s16(vxa0), 1); in xnn_qs8_gemm_minmax_ukernel_3x8__neon_mlal_lane()
85 vacc1x0123 = vmlal_lane_s16(vacc1x0123, vget_low_s16(vxb01234567c1), vget_low_s16(vxa1), 1); in xnn_qs8_gemm_minmax_ukernel_3x8__neon_mlal_lane()
86 vacc1x4567 = vmlal_lane_s16(vacc1x4567, vget_high_s16(vxb01234567c1), vget_low_s16(vxa1), 1); in xnn_qs8_gemm_minmax_ukernel_3x8__neon_mlal_lane()
[all …]
|
| D | 2x8-minmax-neon-mlal-lane.c | 64 vacc0x0123 = vmlal_lane_s16(vacc0x0123, vget_low_s16(vxb01234567c0), vget_low_s16(vxa0), 0); in xnn_qs8_gemm_minmax_ukernel_2x8__neon_mlal_lane()
65 vacc0x4567 = vmlal_lane_s16(vacc0x4567, vget_high_s16(vxb01234567c0), vget_low_s16(vxa0), 0); in xnn_qs8_gemm_minmax_ukernel_2x8__neon_mlal_lane()
66 vacc1x0123 = vmlal_lane_s16(vacc1x0123, vget_low_s16(vxb01234567c0), vget_low_s16(vxa1), 0); in xnn_qs8_gemm_minmax_ukernel_2x8__neon_mlal_lane()
67 vacc1x4567 = vmlal_lane_s16(vacc1x4567, vget_high_s16(vxb01234567c0), vget_low_s16(vxa1), 0); in xnn_qs8_gemm_minmax_ukernel_2x8__neon_mlal_lane()
71 vacc0x0123 = vmlal_lane_s16(vacc0x0123, vget_low_s16(vxb01234567c1), vget_low_s16(vxa0), 1); in xnn_qs8_gemm_minmax_ukernel_2x8__neon_mlal_lane()
72 vacc0x4567 = vmlal_lane_s16(vacc0x4567, vget_high_s16(vxb01234567c1), vget_low_s16(vxa0), 1); in xnn_qs8_gemm_minmax_ukernel_2x8__neon_mlal_lane()
73 vacc1x0123 = vmlal_lane_s16(vacc1x0123, vget_low_s16(vxb01234567c1), vget_low_s16(vxa1), 1); in xnn_qs8_gemm_minmax_ukernel_2x8__neon_mlal_lane()
74 vacc1x4567 = vmlal_lane_s16(vacc1x4567, vget_high_s16(vxb01234567c1), vget_low_s16(vxa1), 1); in xnn_qs8_gemm_minmax_ukernel_2x8__neon_mlal_lane()
78 vacc0x0123 = vmlal_lane_s16(vacc0x0123, vget_low_s16(vxb01234567c2), vget_low_s16(vxa0), 2); in xnn_qs8_gemm_minmax_ukernel_2x8__neon_mlal_lane()
79 vacc0x4567 = vmlal_lane_s16(vacc0x4567, vget_high_s16(vxb01234567c2), vget_low_s16(vxa0), 2); in xnn_qs8_gemm_minmax_ukernel_2x8__neon_mlal_lane()
[all …]
|
| D | 1x16-minmax-neon-mlal-lane.c | 56 vacc0x0123 = vmlal_lane_s16(vacc0x0123, vget_low_s16(vxb01234567c0), vget_low_s16(vxa0), 0); in xnn_qs8_gemm_minmax_ukernel_1x16__neon_mlal_lane()
57 vacc0x4567 = vmlal_lane_s16(vacc0x4567, vget_high_s16(vxb01234567c0), vget_low_s16(vxa0), 0); in xnn_qs8_gemm_minmax_ukernel_1x16__neon_mlal_lane()
61 vacc0x89AB = vmlal_lane_s16(vacc0x89AB, vget_low_s16(vxb89ABCDEFc0), vget_low_s16(vxa0), 0); in xnn_qs8_gemm_minmax_ukernel_1x16__neon_mlal_lane()
62 vacc0xCDEF = vmlal_lane_s16(vacc0xCDEF, vget_high_s16(vxb89ABCDEFc0), vget_low_s16(vxa0), 0); in xnn_qs8_gemm_minmax_ukernel_1x16__neon_mlal_lane()
66 vacc0x0123 = vmlal_lane_s16(vacc0x0123, vget_low_s16(vxb01234567c1), vget_low_s16(vxa0), 1); in xnn_qs8_gemm_minmax_ukernel_1x16__neon_mlal_lane()
67 vacc0x4567 = vmlal_lane_s16(vacc0x4567, vget_high_s16(vxb01234567c1), vget_low_s16(vxa0), 1); in xnn_qs8_gemm_minmax_ukernel_1x16__neon_mlal_lane()
71 vacc0x89AB = vmlal_lane_s16(vacc0x89AB, vget_low_s16(vxb89ABCDEFc1), vget_low_s16(vxa0), 1); in xnn_qs8_gemm_minmax_ukernel_1x16__neon_mlal_lane()
72 vacc0xCDEF = vmlal_lane_s16(vacc0xCDEF, vget_high_s16(vxb89ABCDEFc1), vget_low_s16(vxa0), 1); in xnn_qs8_gemm_minmax_ukernel_1x16__neon_mlal_lane()
76 vacc0x0123 = vmlal_lane_s16(vacc0x0123, vget_low_s16(vxb01234567c2), vget_low_s16(vxa0), 2); in xnn_qs8_gemm_minmax_ukernel_1x16__neon_mlal_lane()
77 vacc0x4567 = vmlal_lane_s16(vacc0x4567, vget_high_s16(vxb01234567c2), vget_low_s16(vxa0), 2); in xnn_qs8_gemm_minmax_ukernel_1x16__neon_mlal_lane()
[all …]
|
| D | 4x16-minmax-neon-mull-addw-dup.c | 88 vacc0x0123 = vaddw_s16(vacc0x0123, vget_low_s16(vprod0x01234567c0)); in xnn_qs8_gemm_minmax_ukernel_4x16__neon_mull_addw_dup()
91 vacc1x0123 = vaddw_s16(vacc1x0123, vget_low_s16(vprod1x01234567c0)); in xnn_qs8_gemm_minmax_ukernel_4x16__neon_mull_addw_dup()
94 vacc2x0123 = vaddw_s16(vacc2x0123, vget_low_s16(vprod2x01234567c0)); in xnn_qs8_gemm_minmax_ukernel_4x16__neon_mull_addw_dup()
97 vacc3x0123 = vaddw_s16(vacc3x0123, vget_low_s16(vprod3x01234567c0)); in xnn_qs8_gemm_minmax_ukernel_4x16__neon_mull_addw_dup()
102 vacc0x89AB = vaddw_s16(vacc0x89AB, vget_low_s16(vprod0x89ABCDEFc0)); in xnn_qs8_gemm_minmax_ukernel_4x16__neon_mull_addw_dup()
105 vacc1x89AB = vaddw_s16(vacc1x89AB, vget_low_s16(vprod1x89ABCDEFc0)); in xnn_qs8_gemm_minmax_ukernel_4x16__neon_mull_addw_dup()
108 vacc2x89AB = vaddw_s16(vacc2x89AB, vget_low_s16(vprod2x89ABCDEFc0)); in xnn_qs8_gemm_minmax_ukernel_4x16__neon_mull_addw_dup()
111 vacc3x89AB = vaddw_s16(vacc3x89AB, vget_low_s16(vprod3x89ABCDEFc0)); in xnn_qs8_gemm_minmax_ukernel_4x16__neon_mull_addw_dup()
116 vacc0x0123 = vaddw_s16(vacc0x0123, vget_low_s16(vprod0x01234567c1)); in xnn_qs8_gemm_minmax_ukernel_4x16__neon_mull_addw_dup()
119 vacc1x0123 = vaddw_s16(vacc1x0123, vget_low_s16(vprod1x01234567c1)); in xnn_qs8_gemm_minmax_ukernel_4x16__neon_mull_addw_dup()
[all …]
|
| /external/XNNPACK/src/qu8-gemm/ |
| D | 8x8-minmax-neon.c | 121 vacc0x0123 = vmlal_lane_s16(vacc0x0123, vget_low_s16(vxb01234567c0), vget_low_s16(vxa0), 0); in xnn_qu8_gemm_minmax_ukernel_8x8__neon()
122 vacc0x4567 = vmlal_lane_s16(vacc0x4567, vget_high_s16(vxb01234567c0), vget_low_s16(vxa0), 0); in xnn_qu8_gemm_minmax_ukernel_8x8__neon()
123 vacc1x0123 = vmlal_lane_s16(vacc1x0123, vget_low_s16(vxb01234567c0), vget_low_s16(vxa1), 0); in xnn_qu8_gemm_minmax_ukernel_8x8__neon()
124 vacc1x4567 = vmlal_lane_s16(vacc1x4567, vget_high_s16(vxb01234567c0), vget_low_s16(vxa1), 0); in xnn_qu8_gemm_minmax_ukernel_8x8__neon()
125 vacc2x0123 = vmlal_lane_s16(vacc2x0123, vget_low_s16(vxb01234567c0), vget_low_s16(vxa2), 0); in xnn_qu8_gemm_minmax_ukernel_8x8__neon()
126 vacc2x4567 = vmlal_lane_s16(vacc2x4567, vget_high_s16(vxb01234567c0), vget_low_s16(vxa2), 0); in xnn_qu8_gemm_minmax_ukernel_8x8__neon()
127 vacc3x0123 = vmlal_lane_s16(vacc3x0123, vget_low_s16(vxb01234567c0), vget_low_s16(vxa3), 0); in xnn_qu8_gemm_minmax_ukernel_8x8__neon()
128 vacc3x4567 = vmlal_lane_s16(vacc3x4567, vget_high_s16(vxb01234567c0), vget_low_s16(vxa3), 0); in xnn_qu8_gemm_minmax_ukernel_8x8__neon()
129 vacc4x0123 = vmlal_lane_s16(vacc4x0123, vget_low_s16(vxb01234567c0), vget_low_s16(vxa4), 0); in xnn_qu8_gemm_minmax_ukernel_8x8__neon()
130 vacc4x4567 = vmlal_lane_s16(vacc4x4567, vget_high_s16(vxb01234567c0), vget_low_s16(vxa4), 0); in xnn_qu8_gemm_minmax_ukernel_8x8__neon()
[all …]
|
| D | 4x8-minmax-neon.c | 81 vacc0x0123 = vmlal_lane_s16(vacc0x0123, vget_low_s16(vxb01234567c0), vget_low_s16(vxa0), 0); in xnn_qu8_gemm_minmax_ukernel_4x8__neon()
82 vacc0x4567 = vmlal_lane_s16(vacc0x4567, vget_high_s16(vxb01234567c0), vget_low_s16(vxa0), 0); in xnn_qu8_gemm_minmax_ukernel_4x8__neon()
83 vacc1x0123 = vmlal_lane_s16(vacc1x0123, vget_low_s16(vxb01234567c0), vget_low_s16(vxa1), 0); in xnn_qu8_gemm_minmax_ukernel_4x8__neon()
84 vacc1x4567 = vmlal_lane_s16(vacc1x4567, vget_high_s16(vxb01234567c0), vget_low_s16(vxa1), 0); in xnn_qu8_gemm_minmax_ukernel_4x8__neon()
85 vacc2x0123 = vmlal_lane_s16(vacc2x0123, vget_low_s16(vxb01234567c0), vget_low_s16(vxa2), 0); in xnn_qu8_gemm_minmax_ukernel_4x8__neon()
86 vacc2x4567 = vmlal_lane_s16(vacc2x4567, vget_high_s16(vxb01234567c0), vget_low_s16(vxa2), 0); in xnn_qu8_gemm_minmax_ukernel_4x8__neon()
87 vacc3x0123 = vmlal_lane_s16(vacc3x0123, vget_low_s16(vxb01234567c0), vget_low_s16(vxa3), 0); in xnn_qu8_gemm_minmax_ukernel_4x8__neon()
88 vacc3x4567 = vmlal_lane_s16(vacc3x4567, vget_high_s16(vxb01234567c0), vget_low_s16(vxa3), 0); in xnn_qu8_gemm_minmax_ukernel_4x8__neon()
93 vacc0x0123 = vmlal_lane_s16(vacc0x0123, vget_low_s16(vxb01234567c1), vget_low_s16(vxa0), 1); in xnn_qu8_gemm_minmax_ukernel_4x8__neon()
94 vacc0x4567 = vmlal_lane_s16(vacc0x4567, vget_high_s16(vxb01234567c1), vget_low_s16(vxa0), 1); in xnn_qu8_gemm_minmax_ukernel_4x8__neon()
[all …]
|
| /external/XNNPACK/src/qu8-igemm/ |
| D | 8x8-minmax-neon.c | 147 vacc0x0123 = vmlal_lane_s16(vacc0x0123, vget_low_s16(vxb01234567), vget_low_s16(vxa0), 0); in xnn_qu8_igemm_minmax_ukernel_8x8__neon()
148 … vacc0x4567 = vmlal_lane_s16(vacc0x4567, vget_high_s16(vxb01234567), vget_low_s16(vxa0), 0); in xnn_qu8_igemm_minmax_ukernel_8x8__neon()
149 vacc1x0123 = vmlal_lane_s16(vacc1x0123, vget_low_s16(vxb01234567), vget_low_s16(vxa1), 0); in xnn_qu8_igemm_minmax_ukernel_8x8__neon()
150 … vacc1x4567 = vmlal_lane_s16(vacc1x4567, vget_high_s16(vxb01234567), vget_low_s16(vxa1), 0); in xnn_qu8_igemm_minmax_ukernel_8x8__neon()
151 vacc2x0123 = vmlal_lane_s16(vacc2x0123, vget_low_s16(vxb01234567), vget_low_s16(vxa2), 0); in xnn_qu8_igemm_minmax_ukernel_8x8__neon()
152 … vacc2x4567 = vmlal_lane_s16(vacc2x4567, vget_high_s16(vxb01234567), vget_low_s16(vxa2), 0); in xnn_qu8_igemm_minmax_ukernel_8x8__neon()
153 vacc3x0123 = vmlal_lane_s16(vacc3x0123, vget_low_s16(vxb01234567), vget_low_s16(vxa3), 0); in xnn_qu8_igemm_minmax_ukernel_8x8__neon()
154 … vacc3x4567 = vmlal_lane_s16(vacc3x4567, vget_high_s16(vxb01234567), vget_low_s16(vxa3), 0); in xnn_qu8_igemm_minmax_ukernel_8x8__neon()
155 vacc4x0123 = vmlal_lane_s16(vacc4x0123, vget_low_s16(vxb01234567), vget_low_s16(vxa4), 0); in xnn_qu8_igemm_minmax_ukernel_8x8__neon()
156 … vacc4x4567 = vmlal_lane_s16(vacc4x4567, vget_high_s16(vxb01234567), vget_low_s16(vxa4), 0); in xnn_qu8_igemm_minmax_ukernel_8x8__neon()
[all …]
|
| D | 4x8-minmax-neon.c | 99 vacc0x0123 = vmlal_lane_s16(vacc0x0123, vget_low_s16(vxb01234567), vget_low_s16(vxa0), 0); in xnn_qu8_igemm_minmax_ukernel_4x8__neon()
100 … vacc0x4567 = vmlal_lane_s16(vacc0x4567, vget_high_s16(vxb01234567), vget_low_s16(vxa0), 0); in xnn_qu8_igemm_minmax_ukernel_4x8__neon()
101 vacc1x0123 = vmlal_lane_s16(vacc1x0123, vget_low_s16(vxb01234567), vget_low_s16(vxa1), 0); in xnn_qu8_igemm_minmax_ukernel_4x8__neon()
102 … vacc1x4567 = vmlal_lane_s16(vacc1x4567, vget_high_s16(vxb01234567), vget_low_s16(vxa1), 0); in xnn_qu8_igemm_minmax_ukernel_4x8__neon()
103 vacc2x0123 = vmlal_lane_s16(vacc2x0123, vget_low_s16(vxb01234567), vget_low_s16(vxa2), 0); in xnn_qu8_igemm_minmax_ukernel_4x8__neon()
104 … vacc2x4567 = vmlal_lane_s16(vacc2x4567, vget_high_s16(vxb01234567), vget_low_s16(vxa2), 0); in xnn_qu8_igemm_minmax_ukernel_4x8__neon()
105 vacc3x0123 = vmlal_lane_s16(vacc3x0123, vget_low_s16(vxb01234567), vget_low_s16(vxa3), 0); in xnn_qu8_igemm_minmax_ukernel_4x8__neon()
106 … vacc3x4567 = vmlal_lane_s16(vacc3x4567, vget_high_s16(vxb01234567), vget_low_s16(vxa3), 0); in xnn_qu8_igemm_minmax_ukernel_4x8__neon()
113 vacc0x0123 = vmlal_lane_s16(vacc0x0123, vget_low_s16(vxb01234567), vget_low_s16(vxa0), 1); in xnn_qu8_igemm_minmax_ukernel_4x8__neon()
114 … vacc0x4567 = vmlal_lane_s16(vacc0x4567, vget_high_s16(vxb01234567), vget_low_s16(vxa0), 1); in xnn_qu8_igemm_minmax_ukernel_4x8__neon()
[all …]
|
| /external/XNNPACK/src/qs8-igemm/gen/ |
| D | 4x16-minmax-neon-mlal-lane.c | 109 vacc0x0123 = vmlal_lane_s16(vacc0x0123, vget_low_s16(vxb01234567c0), vget_low_s16(vxa0), 0); in xnn_qs8_igemm_minmax_ukernel_4x16__neon_mlal_lane()
110 … vacc0x4567 = vmlal_lane_s16(vacc0x4567, vget_high_s16(vxb01234567c0), vget_low_s16(vxa0), 0); in xnn_qs8_igemm_minmax_ukernel_4x16__neon_mlal_lane()
111 vacc1x0123 = vmlal_lane_s16(vacc1x0123, vget_low_s16(vxb01234567c0), vget_low_s16(vxa1), 0); in xnn_qs8_igemm_minmax_ukernel_4x16__neon_mlal_lane()
112 … vacc1x4567 = vmlal_lane_s16(vacc1x4567, vget_high_s16(vxb01234567c0), vget_low_s16(vxa1), 0); in xnn_qs8_igemm_minmax_ukernel_4x16__neon_mlal_lane()
113 vacc2x0123 = vmlal_lane_s16(vacc2x0123, vget_low_s16(vxb01234567c0), vget_low_s16(vxa2), 0); in xnn_qs8_igemm_minmax_ukernel_4x16__neon_mlal_lane()
114 … vacc2x4567 = vmlal_lane_s16(vacc2x4567, vget_high_s16(vxb01234567c0), vget_low_s16(vxa2), 0); in xnn_qs8_igemm_minmax_ukernel_4x16__neon_mlal_lane()
115 vacc3x0123 = vmlal_lane_s16(vacc3x0123, vget_low_s16(vxb01234567c0), vget_low_s16(vxa3), 0); in xnn_qs8_igemm_minmax_ukernel_4x16__neon_mlal_lane()
116 … vacc3x4567 = vmlal_lane_s16(vacc3x4567, vget_high_s16(vxb01234567c0), vget_low_s16(vxa3), 0); in xnn_qs8_igemm_minmax_ukernel_4x16__neon_mlal_lane()
120 vacc0x89AB = vmlal_lane_s16(vacc0x89AB, vget_low_s16(vxb89ABCDEFc0), vget_low_s16(vxa0), 0); in xnn_qs8_igemm_minmax_ukernel_4x16__neon_mlal_lane()
121 … vacc0xCDEF = vmlal_lane_s16(vacc0xCDEF, vget_high_s16(vxb89ABCDEFc0), vget_low_s16(vxa0), 0); in xnn_qs8_igemm_minmax_ukernel_4x16__neon_mlal_lane()
[all …]
|
| D | 3x16-minmax-neon-mlal-lane.c | 95 vacc0x0123 = vmlal_lane_s16(vacc0x0123, vget_low_s16(vxb01234567c0), vget_low_s16(vxa0), 0); in xnn_qs8_igemm_minmax_ukernel_3x16__neon_mlal_lane()
96 … vacc0x4567 = vmlal_lane_s16(vacc0x4567, vget_high_s16(vxb01234567c0), vget_low_s16(vxa0), 0); in xnn_qs8_igemm_minmax_ukernel_3x16__neon_mlal_lane()
97 vacc1x0123 = vmlal_lane_s16(vacc1x0123, vget_low_s16(vxb01234567c0), vget_low_s16(vxa1), 0); in xnn_qs8_igemm_minmax_ukernel_3x16__neon_mlal_lane()
98 … vacc1x4567 = vmlal_lane_s16(vacc1x4567, vget_high_s16(vxb01234567c0), vget_low_s16(vxa1), 0); in xnn_qs8_igemm_minmax_ukernel_3x16__neon_mlal_lane()
99 vacc2x0123 = vmlal_lane_s16(vacc2x0123, vget_low_s16(vxb01234567c0), vget_low_s16(vxa2), 0); in xnn_qs8_igemm_minmax_ukernel_3x16__neon_mlal_lane()
100 … vacc2x4567 = vmlal_lane_s16(vacc2x4567, vget_high_s16(vxb01234567c0), vget_low_s16(vxa2), 0); in xnn_qs8_igemm_minmax_ukernel_3x16__neon_mlal_lane()
104 vacc0x89AB = vmlal_lane_s16(vacc0x89AB, vget_low_s16(vxb89ABCDEFc0), vget_low_s16(vxa0), 0); in xnn_qs8_igemm_minmax_ukernel_3x16__neon_mlal_lane()
105 … vacc0xCDEF = vmlal_lane_s16(vacc0xCDEF, vget_high_s16(vxb89ABCDEFc0), vget_low_s16(vxa0), 0); in xnn_qs8_igemm_minmax_ukernel_3x16__neon_mlal_lane()
106 vacc1x89AB = vmlal_lane_s16(vacc1x89AB, vget_low_s16(vxb89ABCDEFc0), vget_low_s16(vxa1), 0); in xnn_qs8_igemm_minmax_ukernel_3x16__neon_mlal_lane()
107 … vacc1xCDEF = vmlal_lane_s16(vacc1xCDEF, vget_high_s16(vxb89ABCDEFc0), vget_low_s16(vxa1), 0); in xnn_qs8_igemm_minmax_ukernel_3x16__neon_mlal_lane()
[all …]
|
| D | 4x8-minmax-neon-mlal-lane.c | 101 vacc0x0123 = vmlal_lane_s16(vacc0x0123, vget_low_s16(vxb01234567c0), vget_low_s16(vxa0), 0); in xnn_qs8_igemm_minmax_ukernel_4x8__neon_mlal_lane()
102 … vacc0x4567 = vmlal_lane_s16(vacc0x4567, vget_high_s16(vxb01234567c0), vget_low_s16(vxa0), 0); in xnn_qs8_igemm_minmax_ukernel_4x8__neon_mlal_lane()
103 vacc1x0123 = vmlal_lane_s16(vacc1x0123, vget_low_s16(vxb01234567c0), vget_low_s16(vxa1), 0); in xnn_qs8_igemm_minmax_ukernel_4x8__neon_mlal_lane()
104 … vacc1x4567 = vmlal_lane_s16(vacc1x4567, vget_high_s16(vxb01234567c0), vget_low_s16(vxa1), 0); in xnn_qs8_igemm_minmax_ukernel_4x8__neon_mlal_lane()
105 vacc2x0123 = vmlal_lane_s16(vacc2x0123, vget_low_s16(vxb01234567c0), vget_low_s16(vxa2), 0); in xnn_qs8_igemm_minmax_ukernel_4x8__neon_mlal_lane()
106 … vacc2x4567 = vmlal_lane_s16(vacc2x4567, vget_high_s16(vxb01234567c0), vget_low_s16(vxa2), 0); in xnn_qs8_igemm_minmax_ukernel_4x8__neon_mlal_lane()
107 vacc3x0123 = vmlal_lane_s16(vacc3x0123, vget_low_s16(vxb01234567c0), vget_low_s16(vxa3), 0); in xnn_qs8_igemm_minmax_ukernel_4x8__neon_mlal_lane()
108 … vacc3x4567 = vmlal_lane_s16(vacc3x4567, vget_high_s16(vxb01234567c0), vget_low_s16(vxa3), 0); in xnn_qs8_igemm_minmax_ukernel_4x8__neon_mlal_lane()
112 vacc0x0123 = vmlal_lane_s16(vacc0x0123, vget_low_s16(vxb01234567c1), vget_low_s16(vxa0), 1); in xnn_qs8_igemm_minmax_ukernel_4x8__neon_mlal_lane()
113 … vacc0x4567 = vmlal_lane_s16(vacc0x4567, vget_high_s16(vxb01234567c1), vget_low_s16(vxa0), 1); in xnn_qs8_igemm_minmax_ukernel_4x8__neon_mlal_lane()
[all …]
|
| D | 2x16-minmax-neon-mlal-lane.c | 81 vacc0x0123 = vmlal_lane_s16(vacc0x0123, vget_low_s16(vxb01234567c0), vget_low_s16(vxa0), 0); in xnn_qs8_igemm_minmax_ukernel_2x16__neon_mlal_lane()
82 … vacc0x4567 = vmlal_lane_s16(vacc0x4567, vget_high_s16(vxb01234567c0), vget_low_s16(vxa0), 0); in xnn_qs8_igemm_minmax_ukernel_2x16__neon_mlal_lane()
83 vacc1x0123 = vmlal_lane_s16(vacc1x0123, vget_low_s16(vxb01234567c0), vget_low_s16(vxa1), 0); in xnn_qs8_igemm_minmax_ukernel_2x16__neon_mlal_lane()
84 … vacc1x4567 = vmlal_lane_s16(vacc1x4567, vget_high_s16(vxb01234567c0), vget_low_s16(vxa1), 0); in xnn_qs8_igemm_minmax_ukernel_2x16__neon_mlal_lane()
88 vacc0x89AB = vmlal_lane_s16(vacc0x89AB, vget_low_s16(vxb89ABCDEFc0), vget_low_s16(vxa0), 0); in xnn_qs8_igemm_minmax_ukernel_2x16__neon_mlal_lane()
89 … vacc0xCDEF = vmlal_lane_s16(vacc0xCDEF, vget_high_s16(vxb89ABCDEFc0), vget_low_s16(vxa0), 0); in xnn_qs8_igemm_minmax_ukernel_2x16__neon_mlal_lane()
90 vacc1x89AB = vmlal_lane_s16(vacc1x89AB, vget_low_s16(vxb89ABCDEFc0), vget_low_s16(vxa1), 0); in xnn_qs8_igemm_minmax_ukernel_2x16__neon_mlal_lane()
91 … vacc1xCDEF = vmlal_lane_s16(vacc1xCDEF, vget_high_s16(vxb89ABCDEFc0), vget_low_s16(vxa1), 0); in xnn_qs8_igemm_minmax_ukernel_2x16__neon_mlal_lane()
95 vacc0x0123 = vmlal_lane_s16(vacc0x0123, vget_low_s16(vxb01234567c1), vget_low_s16(vxa0), 1); in xnn_qs8_igemm_minmax_ukernel_2x16__neon_mlal_lane()
96 … vacc0x4567 = vmlal_lane_s16(vacc0x4567, vget_high_s16(vxb01234567c1), vget_low_s16(vxa0), 1); in xnn_qs8_igemm_minmax_ukernel_2x16__neon_mlal_lane()
[all …]
|
| D | 3x8-minmax-neon-mlal-lane.c | 89 vacc0x0123 = vmlal_lane_s16(vacc0x0123, vget_low_s16(vxb01234567c0), vget_low_s16(vxa0), 0); in xnn_qs8_igemm_minmax_ukernel_3x8__neon_mlal_lane()
90 … vacc0x4567 = vmlal_lane_s16(vacc0x4567, vget_high_s16(vxb01234567c0), vget_low_s16(vxa0), 0); in xnn_qs8_igemm_minmax_ukernel_3x8__neon_mlal_lane()
91 vacc1x0123 = vmlal_lane_s16(vacc1x0123, vget_low_s16(vxb01234567c0), vget_low_s16(vxa1), 0); in xnn_qs8_igemm_minmax_ukernel_3x8__neon_mlal_lane()
92 … vacc1x4567 = vmlal_lane_s16(vacc1x4567, vget_high_s16(vxb01234567c0), vget_low_s16(vxa1), 0); in xnn_qs8_igemm_minmax_ukernel_3x8__neon_mlal_lane()
93 vacc2x0123 = vmlal_lane_s16(vacc2x0123, vget_low_s16(vxb01234567c0), vget_low_s16(vxa2), 0); in xnn_qs8_igemm_minmax_ukernel_3x8__neon_mlal_lane()
94 … vacc2x4567 = vmlal_lane_s16(vacc2x4567, vget_high_s16(vxb01234567c0), vget_low_s16(vxa2), 0); in xnn_qs8_igemm_minmax_ukernel_3x8__neon_mlal_lane()
98 vacc0x0123 = vmlal_lane_s16(vacc0x0123, vget_low_s16(vxb01234567c1), vget_low_s16(vxa0), 1); in xnn_qs8_igemm_minmax_ukernel_3x8__neon_mlal_lane()
99 … vacc0x4567 = vmlal_lane_s16(vacc0x4567, vget_high_s16(vxb01234567c1), vget_low_s16(vxa0), 1); in xnn_qs8_igemm_minmax_ukernel_3x8__neon_mlal_lane()
100 vacc1x0123 = vmlal_lane_s16(vacc1x0123, vget_low_s16(vxb01234567c1), vget_low_s16(vxa1), 1); in xnn_qs8_igemm_minmax_ukernel_3x8__neon_mlal_lane()
101 … vacc1x4567 = vmlal_lane_s16(vacc1x4567, vget_high_s16(vxb01234567c1), vget_low_s16(vxa1), 1); in xnn_qs8_igemm_minmax_ukernel_3x8__neon_mlal_lane()
[all …]
|
| D | 2x8-minmax-neon-mlal-lane.c | 77 vacc0x0123 = vmlal_lane_s16(vacc0x0123, vget_low_s16(vxb01234567c0), vget_low_s16(vxa0), 0); in xnn_qs8_igemm_minmax_ukernel_2x8__neon_mlal_lane()
78 … vacc0x4567 = vmlal_lane_s16(vacc0x4567, vget_high_s16(vxb01234567c0), vget_low_s16(vxa0), 0); in xnn_qs8_igemm_minmax_ukernel_2x8__neon_mlal_lane()
79 vacc1x0123 = vmlal_lane_s16(vacc1x0123, vget_low_s16(vxb01234567c0), vget_low_s16(vxa1), 0); in xnn_qs8_igemm_minmax_ukernel_2x8__neon_mlal_lane()
80 … vacc1x4567 = vmlal_lane_s16(vacc1x4567, vget_high_s16(vxb01234567c0), vget_low_s16(vxa1), 0); in xnn_qs8_igemm_minmax_ukernel_2x8__neon_mlal_lane()
84 vacc0x0123 = vmlal_lane_s16(vacc0x0123, vget_low_s16(vxb01234567c1), vget_low_s16(vxa0), 1); in xnn_qs8_igemm_minmax_ukernel_2x8__neon_mlal_lane()
85 … vacc0x4567 = vmlal_lane_s16(vacc0x4567, vget_high_s16(vxb01234567c1), vget_low_s16(vxa0), 1); in xnn_qs8_igemm_minmax_ukernel_2x8__neon_mlal_lane()
86 vacc1x0123 = vmlal_lane_s16(vacc1x0123, vget_low_s16(vxb01234567c1), vget_low_s16(vxa1), 1); in xnn_qs8_igemm_minmax_ukernel_2x8__neon_mlal_lane()
87 … vacc1x4567 = vmlal_lane_s16(vacc1x4567, vget_high_s16(vxb01234567c1), vget_low_s16(vxa1), 1); in xnn_qs8_igemm_minmax_ukernel_2x8__neon_mlal_lane()
91 vacc0x0123 = vmlal_lane_s16(vacc0x0123, vget_low_s16(vxb01234567c2), vget_low_s16(vxa0), 2); in xnn_qs8_igemm_minmax_ukernel_2x8__neon_mlal_lane()
92 … vacc0x4567 = vmlal_lane_s16(vacc0x4567, vget_high_s16(vxb01234567c2), vget_low_s16(vxa0), 2); in xnn_qs8_igemm_minmax_ukernel_2x8__neon_mlal_lane()
[all …]
|
| D | 1x16-minmax-neon-mlal-lane.c | 67 vacc0x0123 = vmlal_lane_s16(vacc0x0123, vget_low_s16(vxb01234567c0), vget_low_s16(vxa0), 0); in xnn_qs8_igemm_minmax_ukernel_1x16__neon_mlal_lane()
68 … vacc0x4567 = vmlal_lane_s16(vacc0x4567, vget_high_s16(vxb01234567c0), vget_low_s16(vxa0), 0); in xnn_qs8_igemm_minmax_ukernel_1x16__neon_mlal_lane()
72 vacc0x89AB = vmlal_lane_s16(vacc0x89AB, vget_low_s16(vxb89ABCDEFc0), vget_low_s16(vxa0), 0); in xnn_qs8_igemm_minmax_ukernel_1x16__neon_mlal_lane()
73 … vacc0xCDEF = vmlal_lane_s16(vacc0xCDEF, vget_high_s16(vxb89ABCDEFc0), vget_low_s16(vxa0), 0); in xnn_qs8_igemm_minmax_ukernel_1x16__neon_mlal_lane()
77 vacc0x0123 = vmlal_lane_s16(vacc0x0123, vget_low_s16(vxb01234567c1), vget_low_s16(vxa0), 1); in xnn_qs8_igemm_minmax_ukernel_1x16__neon_mlal_lane()
78 … vacc0x4567 = vmlal_lane_s16(vacc0x4567, vget_high_s16(vxb01234567c1), vget_low_s16(vxa0), 1); in xnn_qs8_igemm_minmax_ukernel_1x16__neon_mlal_lane()
82 vacc0x89AB = vmlal_lane_s16(vacc0x89AB, vget_low_s16(vxb89ABCDEFc1), vget_low_s16(vxa0), 1); in xnn_qs8_igemm_minmax_ukernel_1x16__neon_mlal_lane()
83 … vacc0xCDEF = vmlal_lane_s16(vacc0xCDEF, vget_high_s16(vxb89ABCDEFc1), vget_low_s16(vxa0), 1); in xnn_qs8_igemm_minmax_ukernel_1x16__neon_mlal_lane()
87 vacc0x0123 = vmlal_lane_s16(vacc0x0123, vget_low_s16(vxb01234567c2), vget_low_s16(vxa0), 2); in xnn_qs8_igemm_minmax_ukernel_1x16__neon_mlal_lane()
88 … vacc0x4567 = vmlal_lane_s16(vacc0x4567, vget_high_s16(vxb01234567c2), vget_low_s16(vxa0), 2); in xnn_qs8_igemm_minmax_ukernel_1x16__neon_mlal_lane()
[all …]
|
| D | 4x16-minmax-neon-mull-addw-dup.c | 105 vacc0x0123 = vaddw_s16(vacc0x0123, vget_low_s16(vprod0x01234567c0)); in xnn_qs8_igemm_minmax_ukernel_4x16__neon_mull_addw_dup()
108 vacc1x0123 = vaddw_s16(vacc1x0123, vget_low_s16(vprod1x01234567c0)); in xnn_qs8_igemm_minmax_ukernel_4x16__neon_mull_addw_dup()
111 vacc2x0123 = vaddw_s16(vacc2x0123, vget_low_s16(vprod2x01234567c0)); in xnn_qs8_igemm_minmax_ukernel_4x16__neon_mull_addw_dup()
114 vacc3x0123 = vaddw_s16(vacc3x0123, vget_low_s16(vprod3x01234567c0)); in xnn_qs8_igemm_minmax_ukernel_4x16__neon_mull_addw_dup()
119 vacc0x89AB = vaddw_s16(vacc0x89AB, vget_low_s16(vprod0x89ABCDEFc0)); in xnn_qs8_igemm_minmax_ukernel_4x16__neon_mull_addw_dup()
122 vacc1x89AB = vaddw_s16(vacc1x89AB, vget_low_s16(vprod1x89ABCDEFc0)); in xnn_qs8_igemm_minmax_ukernel_4x16__neon_mull_addw_dup()
125 vacc2x89AB = vaddw_s16(vacc2x89AB, vget_low_s16(vprod2x89ABCDEFc0)); in xnn_qs8_igemm_minmax_ukernel_4x16__neon_mull_addw_dup()
128 vacc3x89AB = vaddw_s16(vacc3x89AB, vget_low_s16(vprod3x89ABCDEFc0)); in xnn_qs8_igemm_minmax_ukernel_4x16__neon_mull_addw_dup()
133 vacc0x0123 = vaddw_s16(vacc0x0123, vget_low_s16(vprod0x01234567c1)); in xnn_qs8_igemm_minmax_ukernel_4x16__neon_mull_addw_dup()
136 vacc1x0123 = vaddw_s16(vacc1x0123, vget_low_s16(vprod1x01234567c1)); in xnn_qs8_igemm_minmax_ukernel_4x16__neon_mull_addw_dup()
[all …]
|
| D | 1x8-minmax-neon-mlal-lane.c | 65 vacc0x0123 = vmlal_lane_s16(vacc0x0123, vget_low_s16(vxb01234567c0), vget_low_s16(vxa0), 0); in xnn_qs8_igemm_minmax_ukernel_1x8__neon_mlal_lane()
66 … vacc0x4567 = vmlal_lane_s16(vacc0x4567, vget_high_s16(vxb01234567c0), vget_low_s16(vxa0), 0); in xnn_qs8_igemm_minmax_ukernel_1x8__neon_mlal_lane()
70 vacc0x0123 = vmlal_lane_s16(vacc0x0123, vget_low_s16(vxb01234567c1), vget_low_s16(vxa0), 1); in xnn_qs8_igemm_minmax_ukernel_1x8__neon_mlal_lane()
71 … vacc0x4567 = vmlal_lane_s16(vacc0x4567, vget_high_s16(vxb01234567c1), vget_low_s16(vxa0), 1); in xnn_qs8_igemm_minmax_ukernel_1x8__neon_mlal_lane()
75 vacc0x0123 = vmlal_lane_s16(vacc0x0123, vget_low_s16(vxb01234567c2), vget_low_s16(vxa0), 2); in xnn_qs8_igemm_minmax_ukernel_1x8__neon_mlal_lane()
76 … vacc0x4567 = vmlal_lane_s16(vacc0x4567, vget_high_s16(vxb01234567c2), vget_low_s16(vxa0), 2); in xnn_qs8_igemm_minmax_ukernel_1x8__neon_mlal_lane()
80 vacc0x0123 = vmlal_lane_s16(vacc0x0123, vget_low_s16(vxb01234567c3), vget_low_s16(vxa0), 3); in xnn_qs8_igemm_minmax_ukernel_1x8__neon_mlal_lane()
81 … vacc0x4567 = vmlal_lane_s16(vacc0x4567, vget_high_s16(vxb01234567c3), vget_low_s16(vxa0), 3); in xnn_qs8_igemm_minmax_ukernel_1x8__neon_mlal_lane()
86 … vacc0x0123 = vmlal_lane_s16(vacc0x0123, vget_low_s16(vxb01234567c4), vget_high_s16(vxa0), 0); in xnn_qs8_igemm_minmax_ukernel_1x8__neon_mlal_lane()
91 … vacc0x0123 = vmlal_lane_s16(vacc0x0123, vget_low_s16(vxb01234567c5), vget_high_s16(vxa0), 1); in xnn_qs8_igemm_minmax_ukernel_1x8__neon_mlal_lane()
[all …]
|
| /external/XNNPACK/src/qs8-dwconv/gen/ |
| D | up32x9-minmax-neon-mul16.c | 108 vacc0123 = vmlal_s16(vacc0123, vget_low_s16(vi0x01234567), vget_low_s16(vk0x01234567)); in xnn_qs8_dwconv_minmax_ukernel_up32x9__neon_mul16()
110 vacc89AB = vmlal_s16(vacc89AB, vget_low_s16(vi0x89ABCDEF), vget_low_s16(vk0x89ABCDEF)); in xnn_qs8_dwconv_minmax_ukernel_up32x9__neon_mul16()
112 vaccGHIJ = vmlal_s16(vaccGHIJ, vget_low_s16(vi0xGHIJKLMN), vget_low_s16(vk0xGHIJKLMN)); in xnn_qs8_dwconv_minmax_ukernel_up32x9__neon_mul16()
114 vaccOPQR = vmlal_s16(vaccOPQR, vget_low_s16(vi0xOPQRSTUV), vget_low_s16(vk0xOPQRSTUV)); in xnn_qs8_dwconv_minmax_ukernel_up32x9__neon_mul16()
126 vacc0123 = vmlal_s16(vacc0123, vget_low_s16(vi1x01234567), vget_low_s16(vk1x01234567)); in xnn_qs8_dwconv_minmax_ukernel_up32x9__neon_mul16()
128 vacc89AB = vmlal_s16(vacc89AB, vget_low_s16(vi1x89ABCDEF), vget_low_s16(vk1x89ABCDEF)); in xnn_qs8_dwconv_minmax_ukernel_up32x9__neon_mul16()
130 vaccGHIJ = vmlal_s16(vaccGHIJ, vget_low_s16(vi1xGHIJKLMN), vget_low_s16(vk1xGHIJKLMN)); in xnn_qs8_dwconv_minmax_ukernel_up32x9__neon_mul16()
132 vaccOPQR = vmlal_s16(vaccOPQR, vget_low_s16(vi1xOPQRSTUV), vget_low_s16(vk1xOPQRSTUV)); in xnn_qs8_dwconv_minmax_ukernel_up32x9__neon_mul16()
144 vacc0123 = vmlal_s16(vacc0123, vget_low_s16(vi2x01234567), vget_low_s16(vk2x01234567)); in xnn_qs8_dwconv_minmax_ukernel_up32x9__neon_mul16()
146 vacc89AB = vmlal_s16(vacc89AB, vget_low_s16(vi2x89ABCDEF), vget_low_s16(vk2x89ABCDEF)); in xnn_qs8_dwconv_minmax_ukernel_up32x9__neon_mul16()
[all …]
|
| D | up24x9-minmax-neon-mul16.c | 104 vacc0123 = vmlal_s16(vacc0123, vget_low_s16(vi0x01234567), vget_low_s16(vk0x01234567)); in xnn_qs8_dwconv_minmax_ukernel_up24x9__neon_mul16()
106 vacc89AB = vmlal_s16(vacc89AB, vget_low_s16(vi0x89ABCDEF), vget_low_s16(vk0x89ABCDEF)); in xnn_qs8_dwconv_minmax_ukernel_up24x9__neon_mul16()
108 vaccGHIJ = vmlal_s16(vaccGHIJ, vget_low_s16(vi0xGHIJKLMN), vget_low_s16(vk0xGHIJKLMN)); in xnn_qs8_dwconv_minmax_ukernel_up24x9__neon_mul16()
118 vacc0123 = vmlal_s16(vacc0123, vget_low_s16(vi1x01234567), vget_low_s16(vk1x01234567)); in xnn_qs8_dwconv_minmax_ukernel_up24x9__neon_mul16()
120 vacc89AB = vmlal_s16(vacc89AB, vget_low_s16(vi1x89ABCDEF), vget_low_s16(vk1x89ABCDEF)); in xnn_qs8_dwconv_minmax_ukernel_up24x9__neon_mul16()
122 vaccGHIJ = vmlal_s16(vaccGHIJ, vget_low_s16(vi1xGHIJKLMN), vget_low_s16(vk1xGHIJKLMN)); in xnn_qs8_dwconv_minmax_ukernel_up24x9__neon_mul16()
132 vacc0123 = vmlal_s16(vacc0123, vget_low_s16(vi2x01234567), vget_low_s16(vk2x01234567)); in xnn_qs8_dwconv_minmax_ukernel_up24x9__neon_mul16()
134 vacc89AB = vmlal_s16(vacc89AB, vget_low_s16(vi2x89ABCDEF), vget_low_s16(vk2x89ABCDEF)); in xnn_qs8_dwconv_minmax_ukernel_up24x9__neon_mul16()
136 vaccGHIJ = vmlal_s16(vaccGHIJ, vget_low_s16(vi2xGHIJKLMN), vget_low_s16(vk2xGHIJKLMN)); in xnn_qs8_dwconv_minmax_ukernel_up24x9__neon_mul16()
146 vacc0123 = vmlal_s16(vacc0123, vget_low_s16(vi3x01234567), vget_low_s16(vk3x01234567)); in xnn_qs8_dwconv_minmax_ukernel_up24x9__neon_mul16()
[all …]
|
| D | up16x9-minmax-neon-mul16.c | 100 vacc0123 = vmlal_s16(vacc0123, vget_low_s16(vi0x01234567), vget_low_s16(vk0x01234567)); in xnn_qs8_dwconv_minmax_ukernel_up16x9__neon_mul16()
102 vacc89AB = vmlal_s16(vacc89AB, vget_low_s16(vi0x89ABCDEF), vget_low_s16(vk0x89ABCDEF)); in xnn_qs8_dwconv_minmax_ukernel_up16x9__neon_mul16()
110 vacc0123 = vmlal_s16(vacc0123, vget_low_s16(vi1x01234567), vget_low_s16(vk1x01234567)); in xnn_qs8_dwconv_minmax_ukernel_up16x9__neon_mul16()
112 vacc89AB = vmlal_s16(vacc89AB, vget_low_s16(vi1x89ABCDEF), vget_low_s16(vk1x89ABCDEF)); in xnn_qs8_dwconv_minmax_ukernel_up16x9__neon_mul16()
120 vacc0123 = vmlal_s16(vacc0123, vget_low_s16(vi2x01234567), vget_low_s16(vk2x01234567)); in xnn_qs8_dwconv_minmax_ukernel_up16x9__neon_mul16()
122 vacc89AB = vmlal_s16(vacc89AB, vget_low_s16(vi2x89ABCDEF), vget_low_s16(vk2x89ABCDEF)); in xnn_qs8_dwconv_minmax_ukernel_up16x9__neon_mul16()
130 vacc0123 = vmlal_s16(vacc0123, vget_low_s16(vi3x01234567), vget_low_s16(vk3x01234567)); in xnn_qs8_dwconv_minmax_ukernel_up16x9__neon_mul16()
132 vacc89AB = vmlal_s16(vacc89AB, vget_low_s16(vi3x89ABCDEF), vget_low_s16(vk3x89ABCDEF)); in xnn_qs8_dwconv_minmax_ukernel_up16x9__neon_mul16()
140 vacc0123 = vmlal_s16(vacc0123, vget_low_s16(vi4x01234567), vget_low_s16(vk4x01234567)); in xnn_qs8_dwconv_minmax_ukernel_up16x9__neon_mul16()
142 vacc89AB = vmlal_s16(vacc89AB, vget_low_s16(vi4x89ABCDEF), vget_low_s16(vk4x89ABCDEF)); in xnn_qs8_dwconv_minmax_ukernel_up16x9__neon_mul16()
[all …]
|
| /external/libhevc/common/arm/ |
| D | ihevc_resi_trans_neon_32x32.c | 137 vget_high_s16(diff_16[2][0]), vget_low_s16(diff_16[2][0])); in ihevc_resi_trans_32x32_neon()
141 vget_high_s16(diff_16[3][0]), vget_low_s16(diff_16[3][0])); in ihevc_resi_trans_32x32_neon()
162 vget_high_s16(diff_16[2][1]), vget_low_s16(diff_16[2][1])); in ihevc_resi_trans_32x32_neon()
166 vget_high_s16(diff_16[3][1]), vget_low_s16(diff_16[3][1])); in ihevc_resi_trans_32x32_neon()
239 e0_1 = vcombine_s16(vget_high_s16(e0_1), vget_low_s16(e0_1)); in ihevc_resi_trans_32x32_neon()
244 e1_1 = vcombine_s16(vget_high_s16(e1_1), vget_low_s16(e1_1)); in ihevc_resi_trans_32x32_neon()
253 vcombine_s16(vget_low_s16(ee0), vget_low_s16(ee1)); in ihevc_resi_trans_32x32_neon()
265 vtrn_s32(vreinterpret_s32_s16(vget_low_s16(eee)), in ihevc_resi_trans_32x32_neon()
274 vtrn_s16(vget_low_s16(eeee), vget_high_s16(eeee)); in ihevc_resi_trans_32x32_neon()
286 vmull_s16(vget_low_s16(g_ai2_ihevc_trans_32_01_8), eeee_00); in ihevc_resi_trans_32x32_neon()
[all …]
|