Home
last modified time | relevance | path

Searched refs:vprod0x0 (Results 1 – 25 of 64) sorted by relevance

123

/external/XNNPACK/src/qs8-gemm/gen/
D1x8c8-minmax-rndnu-neon-mlal.c69 int16x8_t vprod0x0 = vmull_s8(vb0x0, va0x0); in xnn_qs8_gemm_minmax_rndnu_ukernel_1x8c8__neon_mlal() local
70 vprod0x0 = vmlal_s8(vprod0x0, vb0x1, va0x1); in xnn_qs8_gemm_minmax_rndnu_ukernel_1x8c8__neon_mlal()
71 vacc0x0 = vpadalq_s16(vacc0x0, vprod0x0); in xnn_qs8_gemm_minmax_rndnu_ukernel_1x8c8__neon_mlal()
109 const int16x8_t vprod0x0 = vmull_s8(vb0, va0); in xnn_qs8_gemm_minmax_rndnu_ukernel_1x8c8__neon_mlal() local
110 vacc0x0 = vpadalq_s16(vacc0x0, vprod0x0); in xnn_qs8_gemm_minmax_rndnu_ukernel_1x8c8__neon_mlal()
D1x8c8-minmax-fp32-neon-mlal.c69 int16x8_t vprod0x0 = vmull_s8(vb0x0, va0x0); in xnn_qs8_gemm_minmax_fp32_ukernel_1x8c8__neon_mlal() local
70 vprod0x0 = vmlal_s8(vprod0x0, vb0x1, va0x1); in xnn_qs8_gemm_minmax_fp32_ukernel_1x8c8__neon_mlal()
71 vacc0x0 = vpadalq_s16(vacc0x0, vprod0x0); in xnn_qs8_gemm_minmax_fp32_ukernel_1x8c8__neon_mlal()
109 const int16x8_t vprod0x0 = vmull_s8(vb0, va0); in xnn_qs8_gemm_minmax_fp32_ukernel_1x8c8__neon_mlal() local
110 vacc0x0 = vpadalq_s16(vacc0x0, vprod0x0); in xnn_qs8_gemm_minmax_fp32_ukernel_1x8c8__neon_mlal()
D1x8c8-minmax-fp32-neonv8-mlal.c70 int16x8_t vprod0x0 = vmull_s8(vb0x0, va0x0); in xnn_qs8_gemm_minmax_fp32_ukernel_1x8c8__neonv8_mlal() local
71 vprod0x0 = vmlal_s8(vprod0x0, vb0x1, va0x1); in xnn_qs8_gemm_minmax_fp32_ukernel_1x8c8__neonv8_mlal()
72 vacc0x0 = vpadalq_s16(vacc0x0, vprod0x0); in xnn_qs8_gemm_minmax_fp32_ukernel_1x8c8__neonv8_mlal()
110 const int16x8_t vprod0x0 = vmull_s8(vb0, va0); in xnn_qs8_gemm_minmax_fp32_ukernel_1x8c8__neonv8_mlal() local
111 vacc0x0 = vpadalq_s16(vacc0x0, vprod0x0); in xnn_qs8_gemm_minmax_fp32_ukernel_1x8c8__neonv8_mlal()
D1x8c16-minmax-rndnu-neon-mlal.c67 int16x8_t vprod0x0 = vmull_s8(vget_low_s8(vb0), vget_low_s8(va0)); in xnn_qs8_gemm_minmax_rndnu_ukernel_1x8c16__neon_mlal() local
68 vprod0x0 = vmlal_s8(vprod0x0, vget_high_s8(vb0), vget_high_s8(va0)); in xnn_qs8_gemm_minmax_rndnu_ukernel_1x8c16__neon_mlal()
69 vacc0x0 = vpadalq_s16(vacc0x0, vprod0x0); in xnn_qs8_gemm_minmax_rndnu_ukernel_1x8c16__neon_mlal()
D2x8c8-minmax-fp32-neonv8-mlal.c86 int16x8_t vprod0x0 = vmull_s8(vb0x0, va0x0); in xnn_qs8_gemm_minmax_fp32_ukernel_2x8c8__neonv8_mlal() local
88 vprod0x0 = vmlal_s8(vprod0x0, vb0x1, va0x1); in xnn_qs8_gemm_minmax_fp32_ukernel_2x8c8__neonv8_mlal()
90 vacc0x0 = vpadalq_s16(vacc0x0, vprod0x0); in xnn_qs8_gemm_minmax_fp32_ukernel_2x8c8__neonv8_mlal()
151 const int16x8_t vprod0x0 = vmull_s8(vb0, va0); in xnn_qs8_gemm_minmax_fp32_ukernel_2x8c8__neonv8_mlal() local
153 vacc0x0 = vpadalq_s16(vacc0x0, vprod0x0); in xnn_qs8_gemm_minmax_fp32_ukernel_2x8c8__neonv8_mlal()
D2x8c8-minmax-fp32-neon-mlal.c85 int16x8_t vprod0x0 = vmull_s8(vb0x0, va0x0); in xnn_qs8_gemm_minmax_fp32_ukernel_2x8c8__neon_mlal() local
87 vprod0x0 = vmlal_s8(vprod0x0, vb0x1, va0x1); in xnn_qs8_gemm_minmax_fp32_ukernel_2x8c8__neon_mlal()
89 vacc0x0 = vpadalq_s16(vacc0x0, vprod0x0); in xnn_qs8_gemm_minmax_fp32_ukernel_2x8c8__neon_mlal()
150 const int16x8_t vprod0x0 = vmull_s8(vb0, va0); in xnn_qs8_gemm_minmax_fp32_ukernel_2x8c8__neon_mlal() local
152 vacc0x0 = vpadalq_s16(vacc0x0, vprod0x0); in xnn_qs8_gemm_minmax_fp32_ukernel_2x8c8__neon_mlal()
D2x8c8-minmax-rndnu-neon-mlal.c85 int16x8_t vprod0x0 = vmull_s8(vb0x0, va0x0); in xnn_qs8_gemm_minmax_rndnu_ukernel_2x8c8__neon_mlal() local
87 vprod0x0 = vmlal_s8(vprod0x0, vb0x1, va0x1); in xnn_qs8_gemm_minmax_rndnu_ukernel_2x8c8__neon_mlal()
89 vacc0x0 = vpadalq_s16(vacc0x0, vprod0x0); in xnn_qs8_gemm_minmax_rndnu_ukernel_2x8c8__neon_mlal()
150 const int16x8_t vprod0x0 = vmull_s8(vb0, va0); in xnn_qs8_gemm_minmax_rndnu_ukernel_2x8c8__neon_mlal() local
152 vacc0x0 = vpadalq_s16(vacc0x0, vprod0x0); in xnn_qs8_gemm_minmax_rndnu_ukernel_2x8c8__neon_mlal()
D1x16c8-minmax-rndnu-neon-mlal.c85 int16x8_t vprod0x0 = vmull_s8(vb0x0, va0x0); in xnn_qs8_gemm_minmax_rndnu_ukernel_1x16c8__neon_mlal() local
86 vprod0x0 = vmlal_s8(vprod0x0, vb0x1, va0x1); in xnn_qs8_gemm_minmax_rndnu_ukernel_1x16c8__neon_mlal()
87 vacc0x0 = vpadalq_s16(vacc0x0, vprod0x0); in xnn_qs8_gemm_minmax_rndnu_ukernel_1x16c8__neon_mlal()
157 const int16x8_t vprod0x0 = vmull_s8(vb0, va0); in xnn_qs8_gemm_minmax_rndnu_ukernel_1x16c8__neon_mlal() local
158 vacc0x0 = vpadalq_s16(vacc0x0, vprod0x0); in xnn_qs8_gemm_minmax_rndnu_ukernel_1x16c8__neon_mlal()
D2x8c16-minmax-rndnu-neon-mlal.c82 int16x8_t vprod0x0 = vmull_s8(vget_low_s8(vb0), vget_low_s8(va0)); in xnn_qs8_gemm_minmax_rndnu_ukernel_2x8c16__neon_mlal() local
84 vprod0x0 = vmlal_s8(vprod0x0, vget_high_s8(vb0), vget_high_s8(va0)); in xnn_qs8_gemm_minmax_rndnu_ukernel_2x8c16__neon_mlal()
86 vacc0x0 = vpadalq_s16(vacc0x0, vprod0x0); in xnn_qs8_gemm_minmax_rndnu_ukernel_2x8c16__neon_mlal()
/external/XNNPACK/src/qc8-gemm/gen/
D1x8c8-minmax-fp32-neon-mlal.c69 int16x8_t vprod0x0 = vmull_s8(vb0x0, va0x0); in xnn_qc8_gemm_minmax_fp32_ukernel_1x8c8__neon_mlal() local
70 vprod0x0 = vmlal_s8(vprod0x0, vb0x1, va0x1); in xnn_qc8_gemm_minmax_fp32_ukernel_1x8c8__neon_mlal()
71 vacc0x0 = vpadalq_s16(vacc0x0, vprod0x0); in xnn_qc8_gemm_minmax_fp32_ukernel_1x8c8__neon_mlal()
109 const int16x8_t vprod0x0 = vmull_s8(vb0, va0); in xnn_qc8_gemm_minmax_fp32_ukernel_1x8c8__neon_mlal() local
110 vacc0x0 = vpadalq_s16(vacc0x0, vprod0x0); in xnn_qc8_gemm_minmax_fp32_ukernel_1x8c8__neon_mlal()
D1x8c8-minmax-fp32-neonv8-mlal.c70 int16x8_t vprod0x0 = vmull_s8(vb0x0, va0x0); in xnn_qc8_gemm_minmax_fp32_ukernel_1x8c8__neonv8_mlal() local
71 vprod0x0 = vmlal_s8(vprod0x0, vb0x1, va0x1); in xnn_qc8_gemm_minmax_fp32_ukernel_1x8c8__neonv8_mlal()
72 vacc0x0 = vpadalq_s16(vacc0x0, vprod0x0); in xnn_qc8_gemm_minmax_fp32_ukernel_1x8c8__neonv8_mlal()
110 const int16x8_t vprod0x0 = vmull_s8(vb0, va0); in xnn_qc8_gemm_minmax_fp32_ukernel_1x8c8__neonv8_mlal() local
111 vacc0x0 = vpadalq_s16(vacc0x0, vprod0x0); in xnn_qc8_gemm_minmax_fp32_ukernel_1x8c8__neonv8_mlal()
D2x8c8-minmax-fp32-neon-mlal.c85 int16x8_t vprod0x0 = vmull_s8(vb0x0, va0x0); in xnn_qc8_gemm_minmax_fp32_ukernel_2x8c8__neon_mlal() local
87 vprod0x0 = vmlal_s8(vprod0x0, vb0x1, va0x1); in xnn_qc8_gemm_minmax_fp32_ukernel_2x8c8__neon_mlal()
89 vacc0x0 = vpadalq_s16(vacc0x0, vprod0x0); in xnn_qc8_gemm_minmax_fp32_ukernel_2x8c8__neon_mlal()
150 const int16x8_t vprod0x0 = vmull_s8(vb0, va0); in xnn_qc8_gemm_minmax_fp32_ukernel_2x8c8__neon_mlal() local
152 vacc0x0 = vpadalq_s16(vacc0x0, vprod0x0); in xnn_qc8_gemm_minmax_fp32_ukernel_2x8c8__neon_mlal()
D2x8c8-minmax-fp32-neonv8-mlal.c86 int16x8_t vprod0x0 = vmull_s8(vb0x0, va0x0); in xnn_qc8_gemm_minmax_fp32_ukernel_2x8c8__neonv8_mlal() local
88 vprod0x0 = vmlal_s8(vprod0x0, vb0x1, va0x1); in xnn_qc8_gemm_minmax_fp32_ukernel_2x8c8__neonv8_mlal()
90 vacc0x0 = vpadalq_s16(vacc0x0, vprod0x0); in xnn_qc8_gemm_minmax_fp32_ukernel_2x8c8__neonv8_mlal()
151 const int16x8_t vprod0x0 = vmull_s8(vb0, va0); in xnn_qc8_gemm_minmax_fp32_ukernel_2x8c8__neonv8_mlal() local
153 vacc0x0 = vpadalq_s16(vacc0x0, vprod0x0); in xnn_qc8_gemm_minmax_fp32_ukernel_2x8c8__neonv8_mlal()
/external/XNNPACK/src/qs8-igemm/gen/
D1x8c8-minmax-fp32-neonv8-mlal.c81 int16x8_t vprod0x0 = vmull_s8(vb0x0, va0x0); in xnn_qs8_igemm_minmax_fp32_ukernel_1x8c8__neonv8_mlal() local
82 vprod0x0 = vmlal_s8(vprod0x0, vb0x1, va0x1); in xnn_qs8_igemm_minmax_fp32_ukernel_1x8c8__neonv8_mlal()
83 vacc0x0 = vpadalq_s16(vacc0x0, vprod0x0); in xnn_qs8_igemm_minmax_fp32_ukernel_1x8c8__neonv8_mlal()
121 const int16x8_t vprod0x0 = vmull_s8(vb0, va0); in xnn_qs8_igemm_minmax_fp32_ukernel_1x8c8__neonv8_mlal() local
122 vacc0x0 = vpadalq_s16(vacc0x0, vprod0x0); in xnn_qs8_igemm_minmax_fp32_ukernel_1x8c8__neonv8_mlal()
D1x8c8-minmax-rndnu-neon-mlal.c80 int16x8_t vprod0x0 = vmull_s8(vb0x0, va0x0); in xnn_qs8_igemm_minmax_rndnu_ukernel_1x8c8__neon_mlal() local
81 vprod0x0 = vmlal_s8(vprod0x0, vb0x1, va0x1); in xnn_qs8_igemm_minmax_rndnu_ukernel_1x8c8__neon_mlal()
82 vacc0x0 = vpadalq_s16(vacc0x0, vprod0x0); in xnn_qs8_igemm_minmax_rndnu_ukernel_1x8c8__neon_mlal()
120 const int16x8_t vprod0x0 = vmull_s8(vb0, va0); in xnn_qs8_igemm_minmax_rndnu_ukernel_1x8c8__neon_mlal() local
121 vacc0x0 = vpadalq_s16(vacc0x0, vprod0x0); in xnn_qs8_igemm_minmax_rndnu_ukernel_1x8c8__neon_mlal()
D1x8c8-minmax-fp32-neon-mlal.c80 int16x8_t vprod0x0 = vmull_s8(vb0x0, va0x0); in xnn_qs8_igemm_minmax_fp32_ukernel_1x8c8__neon_mlal() local
81 vprod0x0 = vmlal_s8(vprod0x0, vb0x1, va0x1); in xnn_qs8_igemm_minmax_fp32_ukernel_1x8c8__neon_mlal()
82 vacc0x0 = vpadalq_s16(vacc0x0, vprod0x0); in xnn_qs8_igemm_minmax_fp32_ukernel_1x8c8__neon_mlal()
120 const int16x8_t vprod0x0 = vmull_s8(vb0, va0); in xnn_qs8_igemm_minmax_fp32_ukernel_1x8c8__neon_mlal() local
121 vacc0x0 = vpadalq_s16(vacc0x0, vprod0x0); in xnn_qs8_igemm_minmax_fp32_ukernel_1x8c8__neon_mlal()
D1x8c16-minmax-rndnu-neon-mlal.c78 int16x8_t vprod0x0 = vmull_s8(vget_low_s8(vb0), vget_low_s8(va0)); in xnn_qs8_igemm_minmax_rndnu_ukernel_1x8c16__neon_mlal() local
79 vprod0x0 = vmlal_s8(vprod0x0, vget_high_s8(vb0), vget_high_s8(va0)); in xnn_qs8_igemm_minmax_rndnu_ukernel_1x8c16__neon_mlal()
80 vacc0x0 = vpadalq_s16(vacc0x0, vprod0x0); in xnn_qs8_igemm_minmax_rndnu_ukernel_1x8c16__neon_mlal()
D2x8c8-minmax-rndnu-neon-mlal.c98 int16x8_t vprod0x0 = vmull_s8(vb0x0, va0x0); in xnn_qs8_igemm_minmax_rndnu_ukernel_2x8c8__neon_mlal() local
100 vprod0x0 = vmlal_s8(vprod0x0, vb0x1, va0x1); in xnn_qs8_igemm_minmax_rndnu_ukernel_2x8c8__neon_mlal()
102 vacc0x0 = vpadalq_s16(vacc0x0, vprod0x0); in xnn_qs8_igemm_minmax_rndnu_ukernel_2x8c8__neon_mlal()
163 const int16x8_t vprod0x0 = vmull_s8(vb0, va0); in xnn_qs8_igemm_minmax_rndnu_ukernel_2x8c8__neon_mlal() local
165 vacc0x0 = vpadalq_s16(vacc0x0, vprod0x0); in xnn_qs8_igemm_minmax_rndnu_ukernel_2x8c8__neon_mlal()
D2x8c8-minmax-fp32-neon-mlal.c98 int16x8_t vprod0x0 = vmull_s8(vb0x0, va0x0); in xnn_qs8_igemm_minmax_fp32_ukernel_2x8c8__neon_mlal() local
100 vprod0x0 = vmlal_s8(vprod0x0, vb0x1, va0x1); in xnn_qs8_igemm_minmax_fp32_ukernel_2x8c8__neon_mlal()
102 vacc0x0 = vpadalq_s16(vacc0x0, vprod0x0); in xnn_qs8_igemm_minmax_fp32_ukernel_2x8c8__neon_mlal()
163 const int16x8_t vprod0x0 = vmull_s8(vb0, va0); in xnn_qs8_igemm_minmax_fp32_ukernel_2x8c8__neon_mlal() local
165 vacc0x0 = vpadalq_s16(vacc0x0, vprod0x0); in xnn_qs8_igemm_minmax_fp32_ukernel_2x8c8__neon_mlal()
D2x8c8-minmax-fp32-neonv8-mlal.c99 int16x8_t vprod0x0 = vmull_s8(vb0x0, va0x0); in xnn_qs8_igemm_minmax_fp32_ukernel_2x8c8__neonv8_mlal() local
101 vprod0x0 = vmlal_s8(vprod0x0, vb0x1, va0x1); in xnn_qs8_igemm_minmax_fp32_ukernel_2x8c8__neonv8_mlal()
103 vacc0x0 = vpadalq_s16(vacc0x0, vprod0x0); in xnn_qs8_igemm_minmax_fp32_ukernel_2x8c8__neonv8_mlal()
164 const int16x8_t vprod0x0 = vmull_s8(vb0, va0); in xnn_qs8_igemm_minmax_fp32_ukernel_2x8c8__neonv8_mlal() local
166 vacc0x0 = vpadalq_s16(vacc0x0, vprod0x0); in xnn_qs8_igemm_minmax_fp32_ukernel_2x8c8__neonv8_mlal()
D1x16c8-minmax-rndnu-neon-mlal.c96 int16x8_t vprod0x0 = vmull_s8(vb0x0, va0x0); in xnn_qs8_igemm_minmax_rndnu_ukernel_1x16c8__neon_mlal() local
97 vprod0x0 = vmlal_s8(vprod0x0, vb0x1, va0x1); in xnn_qs8_igemm_minmax_rndnu_ukernel_1x16c8__neon_mlal()
98 vacc0x0 = vpadalq_s16(vacc0x0, vprod0x0); in xnn_qs8_igemm_minmax_rndnu_ukernel_1x16c8__neon_mlal()
168 const int16x8_t vprod0x0 = vmull_s8(vb0, va0); in xnn_qs8_igemm_minmax_rndnu_ukernel_1x16c8__neon_mlal() local
169 vacc0x0 = vpadalq_s16(vacc0x0, vprod0x0); in xnn_qs8_igemm_minmax_rndnu_ukernel_1x16c8__neon_mlal()
/external/XNNPACK/src/qc8-igemm/gen/
D1x8c8-minmax-fp32-neon-mlal.c80 int16x8_t vprod0x0 = vmull_s8(vb0x0, va0x0); in xnn_qc8_igemm_minmax_fp32_ukernel_1x8c8__neon_mlal() local
81 vprod0x0 = vmlal_s8(vprod0x0, vb0x1, va0x1); in xnn_qc8_igemm_minmax_fp32_ukernel_1x8c8__neon_mlal()
82 vacc0x0 = vpadalq_s16(vacc0x0, vprod0x0); in xnn_qc8_igemm_minmax_fp32_ukernel_1x8c8__neon_mlal()
120 const int16x8_t vprod0x0 = vmull_s8(vb0, va0); in xnn_qc8_igemm_minmax_fp32_ukernel_1x8c8__neon_mlal() local
121 vacc0x0 = vpadalq_s16(vacc0x0, vprod0x0); in xnn_qc8_igemm_minmax_fp32_ukernel_1x8c8__neon_mlal()
D1x8c8-minmax-fp32-neonv8-mlal.c81 int16x8_t vprod0x0 = vmull_s8(vb0x0, va0x0); in xnn_qc8_igemm_minmax_fp32_ukernel_1x8c8__neonv8_mlal() local
82 vprod0x0 = vmlal_s8(vprod0x0, vb0x1, va0x1); in xnn_qc8_igemm_minmax_fp32_ukernel_1x8c8__neonv8_mlal()
83 vacc0x0 = vpadalq_s16(vacc0x0, vprod0x0); in xnn_qc8_igemm_minmax_fp32_ukernel_1x8c8__neonv8_mlal()
121 const int16x8_t vprod0x0 = vmull_s8(vb0, va0); in xnn_qc8_igemm_minmax_fp32_ukernel_1x8c8__neonv8_mlal() local
122 vacc0x0 = vpadalq_s16(vacc0x0, vprod0x0); in xnn_qc8_igemm_minmax_fp32_ukernel_1x8c8__neonv8_mlal()
D2x8c8-minmax-fp32-neonv8-mlal.c99 int16x8_t vprod0x0 = vmull_s8(vb0x0, va0x0); in xnn_qc8_igemm_minmax_fp32_ukernel_2x8c8__neonv8_mlal() local
101 vprod0x0 = vmlal_s8(vprod0x0, vb0x1, va0x1); in xnn_qc8_igemm_minmax_fp32_ukernel_2x8c8__neonv8_mlal()
103 vacc0x0 = vpadalq_s16(vacc0x0, vprod0x0); in xnn_qc8_igemm_minmax_fp32_ukernel_2x8c8__neonv8_mlal()
164 const int16x8_t vprod0x0 = vmull_s8(vb0, va0); in xnn_qc8_igemm_minmax_fp32_ukernel_2x8c8__neonv8_mlal() local
166 vacc0x0 = vpadalq_s16(vacc0x0, vprod0x0); in xnn_qc8_igemm_minmax_fp32_ukernel_2x8c8__neonv8_mlal()
D2x8c8-minmax-fp32-neon-mlal.c98 int16x8_t vprod0x0 = vmull_s8(vb0x0, va0x0); in xnn_qc8_igemm_minmax_fp32_ukernel_2x8c8__neon_mlal() local
100 vprod0x0 = vmlal_s8(vprod0x0, vb0x1, va0x1); in xnn_qc8_igemm_minmax_fp32_ukernel_2x8c8__neon_mlal()
102 vacc0x0 = vpadalq_s16(vacc0x0, vprod0x0); in xnn_qc8_igemm_minmax_fp32_ukernel_2x8c8__neon_mlal()
163 const int16x8_t vprod0x0 = vmull_s8(vb0, va0); in xnn_qc8_igemm_minmax_fp32_ukernel_2x8c8__neon_mlal() local
165 vacc0x0 = vpadalq_s16(vacc0x0, vprod0x0); in xnn_qc8_igemm_minmax_fp32_ukernel_2x8c8__neon_mlal()

123