Home
last modified time | relevance | path

Searched refs:vprod0x1 (Results 1 – 25 of 64) sorted by relevance

123

/external/XNNPACK/src/qs8-gemm/gen/
D1x8c8-minmax-rndnu-neon-mlal.c73 int16x8_t vprod0x1 = vmull_s8(vb1x0, va0x0); in xnn_qs8_gemm_minmax_rndnu_ukernel_1x8c8__neon_mlal() local
74 vprod0x1 = vmlal_s8(vprod0x1, vb1x1, va0x1); in xnn_qs8_gemm_minmax_rndnu_ukernel_1x8c8__neon_mlal()
75 vacc0x1 = vpadalq_s16(vacc0x1, vprod0x1); in xnn_qs8_gemm_minmax_rndnu_ukernel_1x8c8__neon_mlal()
112 const int16x8_t vprod0x1 = vmull_s8(vb1, va0); in xnn_qs8_gemm_minmax_rndnu_ukernel_1x8c8__neon_mlal() local
113 vacc0x1 = vpadalq_s16(vacc0x1, vprod0x1); in xnn_qs8_gemm_minmax_rndnu_ukernel_1x8c8__neon_mlal()
D1x8c8-minmax-fp32-neon-mlal.c73 int16x8_t vprod0x1 = vmull_s8(vb1x0, va0x0); in xnn_qs8_gemm_minmax_fp32_ukernel_1x8c8__neon_mlal() local
74 vprod0x1 = vmlal_s8(vprod0x1, vb1x1, va0x1); in xnn_qs8_gemm_minmax_fp32_ukernel_1x8c8__neon_mlal()
75 vacc0x1 = vpadalq_s16(vacc0x1, vprod0x1); in xnn_qs8_gemm_minmax_fp32_ukernel_1x8c8__neon_mlal()
112 const int16x8_t vprod0x1 = vmull_s8(vb1, va0); in xnn_qs8_gemm_minmax_fp32_ukernel_1x8c8__neon_mlal() local
113 vacc0x1 = vpadalq_s16(vacc0x1, vprod0x1); in xnn_qs8_gemm_minmax_fp32_ukernel_1x8c8__neon_mlal()
D1x8c8-minmax-fp32-neonv8-mlal.c74 int16x8_t vprod0x1 = vmull_s8(vb1x0, va0x0); in xnn_qs8_gemm_minmax_fp32_ukernel_1x8c8__neonv8_mlal() local
75 vprod0x1 = vmlal_s8(vprod0x1, vb1x1, va0x1); in xnn_qs8_gemm_minmax_fp32_ukernel_1x8c8__neonv8_mlal()
76 vacc0x1 = vpadalq_s16(vacc0x1, vprod0x1); in xnn_qs8_gemm_minmax_fp32_ukernel_1x8c8__neonv8_mlal()
113 const int16x8_t vprod0x1 = vmull_s8(vb1, va0); in xnn_qs8_gemm_minmax_fp32_ukernel_1x8c8__neonv8_mlal() local
114 vacc0x1 = vpadalq_s16(vacc0x1, vprod0x1); in xnn_qs8_gemm_minmax_fp32_ukernel_1x8c8__neonv8_mlal()
D1x8c16-minmax-rndnu-neon-mlal.c70 int16x8_t vprod0x1 = vmull_s8(vget_low_s8(vb1), vget_low_s8(va0)); in xnn_qs8_gemm_minmax_rndnu_ukernel_1x8c16__neon_mlal() local
71 vprod0x1 = vmlal_s8(vprod0x1, vget_high_s8(vb1), vget_high_s8(va0)); in xnn_qs8_gemm_minmax_rndnu_ukernel_1x8c16__neon_mlal()
72 vacc0x1 = vpadalq_s16(vacc0x1, vprod0x1); in xnn_qs8_gemm_minmax_rndnu_ukernel_1x8c16__neon_mlal()
D2x8c8-minmax-fp32-neonv8-mlal.c93 int16x8_t vprod0x1 = vmull_s8(vb1x0, va0x0); in xnn_qs8_gemm_minmax_fp32_ukernel_2x8c8__neonv8_mlal() local
95 vprod0x1 = vmlal_s8(vprod0x1, vb1x1, va0x1); in xnn_qs8_gemm_minmax_fp32_ukernel_2x8c8__neonv8_mlal()
97 vacc0x1 = vpadalq_s16(vacc0x1, vprod0x1); in xnn_qs8_gemm_minmax_fp32_ukernel_2x8c8__neonv8_mlal()
156 const int16x8_t vprod0x1 = vmull_s8(vb1, va0); in xnn_qs8_gemm_minmax_fp32_ukernel_2x8c8__neonv8_mlal() local
158 vacc0x1 = vpadalq_s16(vacc0x1, vprod0x1); in xnn_qs8_gemm_minmax_fp32_ukernel_2x8c8__neonv8_mlal()
D2x8c8-minmax-fp32-neon-mlal.c92 int16x8_t vprod0x1 = vmull_s8(vb1x0, va0x0); in xnn_qs8_gemm_minmax_fp32_ukernel_2x8c8__neon_mlal() local
94 vprod0x1 = vmlal_s8(vprod0x1, vb1x1, va0x1); in xnn_qs8_gemm_minmax_fp32_ukernel_2x8c8__neon_mlal()
96 vacc0x1 = vpadalq_s16(vacc0x1, vprod0x1); in xnn_qs8_gemm_minmax_fp32_ukernel_2x8c8__neon_mlal()
155 const int16x8_t vprod0x1 = vmull_s8(vb1, va0); in xnn_qs8_gemm_minmax_fp32_ukernel_2x8c8__neon_mlal() local
157 vacc0x1 = vpadalq_s16(vacc0x1, vprod0x1); in xnn_qs8_gemm_minmax_fp32_ukernel_2x8c8__neon_mlal()
D2x8c8-minmax-rndnu-neon-mlal.c92 int16x8_t vprod0x1 = vmull_s8(vb1x0, va0x0); in xnn_qs8_gemm_minmax_rndnu_ukernel_2x8c8__neon_mlal() local
94 vprod0x1 = vmlal_s8(vprod0x1, vb1x1, va0x1); in xnn_qs8_gemm_minmax_rndnu_ukernel_2x8c8__neon_mlal()
96 vacc0x1 = vpadalq_s16(vacc0x1, vprod0x1); in xnn_qs8_gemm_minmax_rndnu_ukernel_2x8c8__neon_mlal()
155 const int16x8_t vprod0x1 = vmull_s8(vb1, va0); in xnn_qs8_gemm_minmax_rndnu_ukernel_2x8c8__neon_mlal() local
157 vacc0x1 = vpadalq_s16(vacc0x1, vprod0x1); in xnn_qs8_gemm_minmax_rndnu_ukernel_2x8c8__neon_mlal()
D1x16c8-minmax-rndnu-neon-mlal.c89 int16x8_t vprod0x1 = vmull_s8(vb1x0, va0x0); in xnn_qs8_gemm_minmax_rndnu_ukernel_1x16c8__neon_mlal() local
90 vprod0x1 = vmlal_s8(vprod0x1, vb1x1, va0x1); in xnn_qs8_gemm_minmax_rndnu_ukernel_1x16c8__neon_mlal()
91 vacc0x1 = vpadalq_s16(vacc0x1, vprod0x1); in xnn_qs8_gemm_minmax_rndnu_ukernel_1x16c8__neon_mlal()
160 const int16x8_t vprod0x1 = vmull_s8(vb1, va0); in xnn_qs8_gemm_minmax_rndnu_ukernel_1x16c8__neon_mlal() local
161 vacc0x1 = vpadalq_s16(vacc0x1, vprod0x1); in xnn_qs8_gemm_minmax_rndnu_ukernel_1x16c8__neon_mlal()
D2x8c16-minmax-rndnu-neon-mlal.c88 int16x8_t vprod0x1 = vmull_s8(vget_low_s8(vb1), vget_low_s8(va0)); in xnn_qs8_gemm_minmax_rndnu_ukernel_2x8c16__neon_mlal() local
90 vprod0x1 = vmlal_s8(vprod0x1, vget_high_s8(vb1), vget_high_s8(va0)); in xnn_qs8_gemm_minmax_rndnu_ukernel_2x8c16__neon_mlal()
92 vacc0x1 = vpadalq_s16(vacc0x1, vprod0x1); in xnn_qs8_gemm_minmax_rndnu_ukernel_2x8c16__neon_mlal()
/external/XNNPACK/src/qc8-gemm/gen/
D1x8c8-minmax-fp32-neon-mlal.c73 int16x8_t vprod0x1 = vmull_s8(vb1x0, va0x0); in xnn_qc8_gemm_minmax_fp32_ukernel_1x8c8__neon_mlal() local
74 vprod0x1 = vmlal_s8(vprod0x1, vb1x1, va0x1); in xnn_qc8_gemm_minmax_fp32_ukernel_1x8c8__neon_mlal()
75 vacc0x1 = vpadalq_s16(vacc0x1, vprod0x1); in xnn_qc8_gemm_minmax_fp32_ukernel_1x8c8__neon_mlal()
112 const int16x8_t vprod0x1 = vmull_s8(vb1, va0); in xnn_qc8_gemm_minmax_fp32_ukernel_1x8c8__neon_mlal() local
113 vacc0x1 = vpadalq_s16(vacc0x1, vprod0x1); in xnn_qc8_gemm_minmax_fp32_ukernel_1x8c8__neon_mlal()
D1x8c8-minmax-fp32-neonv8-mlal.c74 int16x8_t vprod0x1 = vmull_s8(vb1x0, va0x0); in xnn_qc8_gemm_minmax_fp32_ukernel_1x8c8__neonv8_mlal() local
75 vprod0x1 = vmlal_s8(vprod0x1, vb1x1, va0x1); in xnn_qc8_gemm_minmax_fp32_ukernel_1x8c8__neonv8_mlal()
76 vacc0x1 = vpadalq_s16(vacc0x1, vprod0x1); in xnn_qc8_gemm_minmax_fp32_ukernel_1x8c8__neonv8_mlal()
113 const int16x8_t vprod0x1 = vmull_s8(vb1, va0); in xnn_qc8_gemm_minmax_fp32_ukernel_1x8c8__neonv8_mlal() local
114 vacc0x1 = vpadalq_s16(vacc0x1, vprod0x1); in xnn_qc8_gemm_minmax_fp32_ukernel_1x8c8__neonv8_mlal()
D2x8c8-minmax-fp32-neon-mlal.c92 int16x8_t vprod0x1 = vmull_s8(vb1x0, va0x0); in xnn_qc8_gemm_minmax_fp32_ukernel_2x8c8__neon_mlal() local
94 vprod0x1 = vmlal_s8(vprod0x1, vb1x1, va0x1); in xnn_qc8_gemm_minmax_fp32_ukernel_2x8c8__neon_mlal()
96 vacc0x1 = vpadalq_s16(vacc0x1, vprod0x1); in xnn_qc8_gemm_minmax_fp32_ukernel_2x8c8__neon_mlal()
155 const int16x8_t vprod0x1 = vmull_s8(vb1, va0); in xnn_qc8_gemm_minmax_fp32_ukernel_2x8c8__neon_mlal() local
157 vacc0x1 = vpadalq_s16(vacc0x1, vprod0x1); in xnn_qc8_gemm_minmax_fp32_ukernel_2x8c8__neon_mlal()
D2x8c8-minmax-fp32-neonv8-mlal.c93 int16x8_t vprod0x1 = vmull_s8(vb1x0, va0x0); in xnn_qc8_gemm_minmax_fp32_ukernel_2x8c8__neonv8_mlal() local
95 vprod0x1 = vmlal_s8(vprod0x1, vb1x1, va0x1); in xnn_qc8_gemm_minmax_fp32_ukernel_2x8c8__neonv8_mlal()
97 vacc0x1 = vpadalq_s16(vacc0x1, vprod0x1); in xnn_qc8_gemm_minmax_fp32_ukernel_2x8c8__neonv8_mlal()
156 const int16x8_t vprod0x1 = vmull_s8(vb1, va0); in xnn_qc8_gemm_minmax_fp32_ukernel_2x8c8__neonv8_mlal() local
158 vacc0x1 = vpadalq_s16(vacc0x1, vprod0x1); in xnn_qc8_gemm_minmax_fp32_ukernel_2x8c8__neonv8_mlal()
/external/XNNPACK/src/qs8-igemm/gen/
D1x8c8-minmax-fp32-neonv8-mlal.c85 int16x8_t vprod0x1 = vmull_s8(vb1x0, va0x0); in xnn_qs8_igemm_minmax_fp32_ukernel_1x8c8__neonv8_mlal() local
86 vprod0x1 = vmlal_s8(vprod0x1, vb1x1, va0x1); in xnn_qs8_igemm_minmax_fp32_ukernel_1x8c8__neonv8_mlal()
87 vacc0x1 = vpadalq_s16(vacc0x1, vprod0x1); in xnn_qs8_igemm_minmax_fp32_ukernel_1x8c8__neonv8_mlal()
124 const int16x8_t vprod0x1 = vmull_s8(vb1, va0); in xnn_qs8_igemm_minmax_fp32_ukernel_1x8c8__neonv8_mlal() local
125 vacc0x1 = vpadalq_s16(vacc0x1, vprod0x1); in xnn_qs8_igemm_minmax_fp32_ukernel_1x8c8__neonv8_mlal()
D1x8c8-minmax-rndnu-neon-mlal.c84 int16x8_t vprod0x1 = vmull_s8(vb1x0, va0x0); in xnn_qs8_igemm_minmax_rndnu_ukernel_1x8c8__neon_mlal() local
85 vprod0x1 = vmlal_s8(vprod0x1, vb1x1, va0x1); in xnn_qs8_igemm_minmax_rndnu_ukernel_1x8c8__neon_mlal()
86 vacc0x1 = vpadalq_s16(vacc0x1, vprod0x1); in xnn_qs8_igemm_minmax_rndnu_ukernel_1x8c8__neon_mlal()
123 const int16x8_t vprod0x1 = vmull_s8(vb1, va0); in xnn_qs8_igemm_minmax_rndnu_ukernel_1x8c8__neon_mlal() local
124 vacc0x1 = vpadalq_s16(vacc0x1, vprod0x1); in xnn_qs8_igemm_minmax_rndnu_ukernel_1x8c8__neon_mlal()
D1x8c8-minmax-fp32-neon-mlal.c84 int16x8_t vprod0x1 = vmull_s8(vb1x0, va0x0); in xnn_qs8_igemm_minmax_fp32_ukernel_1x8c8__neon_mlal() local
85 vprod0x1 = vmlal_s8(vprod0x1, vb1x1, va0x1); in xnn_qs8_igemm_minmax_fp32_ukernel_1x8c8__neon_mlal()
86 vacc0x1 = vpadalq_s16(vacc0x1, vprod0x1); in xnn_qs8_igemm_minmax_fp32_ukernel_1x8c8__neon_mlal()
123 const int16x8_t vprod0x1 = vmull_s8(vb1, va0); in xnn_qs8_igemm_minmax_fp32_ukernel_1x8c8__neon_mlal() local
124 vacc0x1 = vpadalq_s16(vacc0x1, vprod0x1); in xnn_qs8_igemm_minmax_fp32_ukernel_1x8c8__neon_mlal()
D1x8c16-minmax-rndnu-neon-mlal.c81 int16x8_t vprod0x1 = vmull_s8(vget_low_s8(vb1), vget_low_s8(va0)); in xnn_qs8_igemm_minmax_rndnu_ukernel_1x8c16__neon_mlal() local
82 vprod0x1 = vmlal_s8(vprod0x1, vget_high_s8(vb1), vget_high_s8(va0)); in xnn_qs8_igemm_minmax_rndnu_ukernel_1x8c16__neon_mlal()
83 vacc0x1 = vpadalq_s16(vacc0x1, vprod0x1); in xnn_qs8_igemm_minmax_rndnu_ukernel_1x8c16__neon_mlal()
D2x8c8-minmax-rndnu-neon-mlal.c105 int16x8_t vprod0x1 = vmull_s8(vb1x0, va0x0); in xnn_qs8_igemm_minmax_rndnu_ukernel_2x8c8__neon_mlal() local
107 vprod0x1 = vmlal_s8(vprod0x1, vb1x1, va0x1); in xnn_qs8_igemm_minmax_rndnu_ukernel_2x8c8__neon_mlal()
109 vacc0x1 = vpadalq_s16(vacc0x1, vprod0x1); in xnn_qs8_igemm_minmax_rndnu_ukernel_2x8c8__neon_mlal()
168 const int16x8_t vprod0x1 = vmull_s8(vb1, va0); in xnn_qs8_igemm_minmax_rndnu_ukernel_2x8c8__neon_mlal() local
170 vacc0x1 = vpadalq_s16(vacc0x1, vprod0x1); in xnn_qs8_igemm_minmax_rndnu_ukernel_2x8c8__neon_mlal()
D2x8c8-minmax-fp32-neon-mlal.c105 int16x8_t vprod0x1 = vmull_s8(vb1x0, va0x0); in xnn_qs8_igemm_minmax_fp32_ukernel_2x8c8__neon_mlal() local
107 vprod0x1 = vmlal_s8(vprod0x1, vb1x1, va0x1); in xnn_qs8_igemm_minmax_fp32_ukernel_2x8c8__neon_mlal()
109 vacc0x1 = vpadalq_s16(vacc0x1, vprod0x1); in xnn_qs8_igemm_minmax_fp32_ukernel_2x8c8__neon_mlal()
168 const int16x8_t vprod0x1 = vmull_s8(vb1, va0); in xnn_qs8_igemm_minmax_fp32_ukernel_2x8c8__neon_mlal() local
170 vacc0x1 = vpadalq_s16(vacc0x1, vprod0x1); in xnn_qs8_igemm_minmax_fp32_ukernel_2x8c8__neon_mlal()
D2x8c8-minmax-fp32-neonv8-mlal.c106 int16x8_t vprod0x1 = vmull_s8(vb1x0, va0x0); in xnn_qs8_igemm_minmax_fp32_ukernel_2x8c8__neonv8_mlal() local
108 vprod0x1 = vmlal_s8(vprod0x1, vb1x1, va0x1); in xnn_qs8_igemm_minmax_fp32_ukernel_2x8c8__neonv8_mlal()
110 vacc0x1 = vpadalq_s16(vacc0x1, vprod0x1); in xnn_qs8_igemm_minmax_fp32_ukernel_2x8c8__neonv8_mlal()
169 const int16x8_t vprod0x1 = vmull_s8(vb1, va0); in xnn_qs8_igemm_minmax_fp32_ukernel_2x8c8__neonv8_mlal() local
171 vacc0x1 = vpadalq_s16(vacc0x1, vprod0x1); in xnn_qs8_igemm_minmax_fp32_ukernel_2x8c8__neonv8_mlal()
D1x16c8-minmax-rndnu-neon-mlal.c100 int16x8_t vprod0x1 = vmull_s8(vb1x0, va0x0); in xnn_qs8_igemm_minmax_rndnu_ukernel_1x16c8__neon_mlal() local
101 vprod0x1 = vmlal_s8(vprod0x1, vb1x1, va0x1); in xnn_qs8_igemm_minmax_rndnu_ukernel_1x16c8__neon_mlal()
102 vacc0x1 = vpadalq_s16(vacc0x1, vprod0x1); in xnn_qs8_igemm_minmax_rndnu_ukernel_1x16c8__neon_mlal()
171 const int16x8_t vprod0x1 = vmull_s8(vb1, va0); in xnn_qs8_igemm_minmax_rndnu_ukernel_1x16c8__neon_mlal() local
172 vacc0x1 = vpadalq_s16(vacc0x1, vprod0x1); in xnn_qs8_igemm_minmax_rndnu_ukernel_1x16c8__neon_mlal()
/external/XNNPACK/src/qc8-igemm/gen/
D1x8c8-minmax-fp32-neon-mlal.c84 int16x8_t vprod0x1 = vmull_s8(vb1x0, va0x0); in xnn_qc8_igemm_minmax_fp32_ukernel_1x8c8__neon_mlal() local
85 vprod0x1 = vmlal_s8(vprod0x1, vb1x1, va0x1); in xnn_qc8_igemm_minmax_fp32_ukernel_1x8c8__neon_mlal()
86 vacc0x1 = vpadalq_s16(vacc0x1, vprod0x1); in xnn_qc8_igemm_minmax_fp32_ukernel_1x8c8__neon_mlal()
123 const int16x8_t vprod0x1 = vmull_s8(vb1, va0); in xnn_qc8_igemm_minmax_fp32_ukernel_1x8c8__neon_mlal() local
124 vacc0x1 = vpadalq_s16(vacc0x1, vprod0x1); in xnn_qc8_igemm_minmax_fp32_ukernel_1x8c8__neon_mlal()
D1x8c8-minmax-fp32-neonv8-mlal.c85 int16x8_t vprod0x1 = vmull_s8(vb1x0, va0x0); in xnn_qc8_igemm_minmax_fp32_ukernel_1x8c8__neonv8_mlal() local
86 vprod0x1 = vmlal_s8(vprod0x1, vb1x1, va0x1); in xnn_qc8_igemm_minmax_fp32_ukernel_1x8c8__neonv8_mlal()
87 vacc0x1 = vpadalq_s16(vacc0x1, vprod0x1); in xnn_qc8_igemm_minmax_fp32_ukernel_1x8c8__neonv8_mlal()
124 const int16x8_t vprod0x1 = vmull_s8(vb1, va0); in xnn_qc8_igemm_minmax_fp32_ukernel_1x8c8__neonv8_mlal() local
125 vacc0x1 = vpadalq_s16(vacc0x1, vprod0x1); in xnn_qc8_igemm_minmax_fp32_ukernel_1x8c8__neonv8_mlal()
D2x8c8-minmax-fp32-neonv8-mlal.c106 int16x8_t vprod0x1 = vmull_s8(vb1x0, va0x0); in xnn_qc8_igemm_minmax_fp32_ukernel_2x8c8__neonv8_mlal() local
108 vprod0x1 = vmlal_s8(vprod0x1, vb1x1, va0x1); in xnn_qc8_igemm_minmax_fp32_ukernel_2x8c8__neonv8_mlal()
110 vacc0x1 = vpadalq_s16(vacc0x1, vprod0x1); in xnn_qc8_igemm_minmax_fp32_ukernel_2x8c8__neonv8_mlal()
169 const int16x8_t vprod0x1 = vmull_s8(vb1, va0); in xnn_qc8_igemm_minmax_fp32_ukernel_2x8c8__neonv8_mlal() local
171 vacc0x1 = vpadalq_s16(vacc0x1, vprod0x1); in xnn_qc8_igemm_minmax_fp32_ukernel_2x8c8__neonv8_mlal()
D2x8c8-minmax-fp32-neon-mlal.c105 int16x8_t vprod0x1 = vmull_s8(vb1x0, va0x0); in xnn_qc8_igemm_minmax_fp32_ukernel_2x8c8__neon_mlal() local
107 vprod0x1 = vmlal_s8(vprod0x1, vb1x1, va0x1); in xnn_qc8_igemm_minmax_fp32_ukernel_2x8c8__neon_mlal()
109 vacc0x1 = vpadalq_s16(vacc0x1, vprod0x1); in xnn_qc8_igemm_minmax_fp32_ukernel_2x8c8__neon_mlal()
168 const int16x8_t vprod0x1 = vmull_s8(vb1, va0); in xnn_qc8_igemm_minmax_fp32_ukernel_2x8c8__neon_mlal() local
170 vacc0x1 = vpadalq_s16(vacc0x1, vprod0x1); in xnn_qc8_igemm_minmax_fp32_ukernel_2x8c8__neon_mlal()

123