Home
last modified time | relevance | path

Searched refs:vprod0x4 (Results 1 – 25 of 64) sorted by relevance

123

/external/XNNPACK/src/qs8-gemm/gen/
D1x8c8-minmax-rndnu-neon-mlal.c85 int16x8_t vprod0x4 = vmull_s8(vb4x0, va0x0); in xnn_qs8_gemm_minmax_rndnu_ukernel_1x8c8__neon_mlal() local
86 vprod0x4 = vmlal_s8(vprod0x4, vb4x1, va0x1); in xnn_qs8_gemm_minmax_rndnu_ukernel_1x8c8__neon_mlal()
87 vacc0x4 = vpadalq_s16(vacc0x4, vprod0x4); in xnn_qs8_gemm_minmax_rndnu_ukernel_1x8c8__neon_mlal()
121 const int16x8_t vprod0x4 = vmull_s8(vb4, va0); in xnn_qs8_gemm_minmax_rndnu_ukernel_1x8c8__neon_mlal() local
122 vacc0x4 = vpadalq_s16(vacc0x4, vprod0x4); in xnn_qs8_gemm_minmax_rndnu_ukernel_1x8c8__neon_mlal()
D1x8c8-minmax-fp32-neon-mlal.c85 int16x8_t vprod0x4 = vmull_s8(vb4x0, va0x0); in xnn_qs8_gemm_minmax_fp32_ukernel_1x8c8__neon_mlal() local
86 vprod0x4 = vmlal_s8(vprod0x4, vb4x1, va0x1); in xnn_qs8_gemm_minmax_fp32_ukernel_1x8c8__neon_mlal()
87 vacc0x4 = vpadalq_s16(vacc0x4, vprod0x4); in xnn_qs8_gemm_minmax_fp32_ukernel_1x8c8__neon_mlal()
121 const int16x8_t vprod0x4 = vmull_s8(vb4, va0); in xnn_qs8_gemm_minmax_fp32_ukernel_1x8c8__neon_mlal() local
122 vacc0x4 = vpadalq_s16(vacc0x4, vprod0x4); in xnn_qs8_gemm_minmax_fp32_ukernel_1x8c8__neon_mlal()
D1x8c8-minmax-fp32-neonv8-mlal.c86 int16x8_t vprod0x4 = vmull_s8(vb4x0, va0x0); in xnn_qs8_gemm_minmax_fp32_ukernel_1x8c8__neonv8_mlal() local
87 vprod0x4 = vmlal_s8(vprod0x4, vb4x1, va0x1); in xnn_qs8_gemm_minmax_fp32_ukernel_1x8c8__neonv8_mlal()
88 vacc0x4 = vpadalq_s16(vacc0x4, vprod0x4); in xnn_qs8_gemm_minmax_fp32_ukernel_1x8c8__neonv8_mlal()
122 const int16x8_t vprod0x4 = vmull_s8(vb4, va0); in xnn_qs8_gemm_minmax_fp32_ukernel_1x8c8__neonv8_mlal() local
123 vacc0x4 = vpadalq_s16(vacc0x4, vprod0x4); in xnn_qs8_gemm_minmax_fp32_ukernel_1x8c8__neonv8_mlal()
D1x8c16-minmax-rndnu-neon-mlal.c79 int16x8_t vprod0x4 = vmull_s8(vget_low_s8(vb4), vget_low_s8(va0)); in xnn_qs8_gemm_minmax_rndnu_ukernel_1x8c16__neon_mlal() local
80 vprod0x4 = vmlal_s8(vprod0x4, vget_high_s8(vb4), vget_high_s8(va0)); in xnn_qs8_gemm_minmax_rndnu_ukernel_1x8c16__neon_mlal()
81 vacc0x4 = vpadalq_s16(vacc0x4, vprod0x4); in xnn_qs8_gemm_minmax_rndnu_ukernel_1x8c16__neon_mlal()
D2x8c8-minmax-fp32-neonv8-mlal.c114 int16x8_t vprod0x4 = vmull_s8(vb4x0, va0x0); in xnn_qs8_gemm_minmax_fp32_ukernel_2x8c8__neonv8_mlal() local
116 vprod0x4 = vmlal_s8(vprod0x4, vb4x1, va0x1); in xnn_qs8_gemm_minmax_fp32_ukernel_2x8c8__neonv8_mlal()
118 vacc0x4 = vpadalq_s16(vacc0x4, vprod0x4); in xnn_qs8_gemm_minmax_fp32_ukernel_2x8c8__neonv8_mlal()
171 const int16x8_t vprod0x4 = vmull_s8(vb4, va0); in xnn_qs8_gemm_minmax_fp32_ukernel_2x8c8__neonv8_mlal() local
173 vacc0x4 = vpadalq_s16(vacc0x4, vprod0x4); in xnn_qs8_gemm_minmax_fp32_ukernel_2x8c8__neonv8_mlal()
D2x8c8-minmax-fp32-neon-mlal.c113 int16x8_t vprod0x4 = vmull_s8(vb4x0, va0x0); in xnn_qs8_gemm_minmax_fp32_ukernel_2x8c8__neon_mlal() local
115 vprod0x4 = vmlal_s8(vprod0x4, vb4x1, va0x1); in xnn_qs8_gemm_minmax_fp32_ukernel_2x8c8__neon_mlal()
117 vacc0x4 = vpadalq_s16(vacc0x4, vprod0x4); in xnn_qs8_gemm_minmax_fp32_ukernel_2x8c8__neon_mlal()
170 const int16x8_t vprod0x4 = vmull_s8(vb4, va0); in xnn_qs8_gemm_minmax_fp32_ukernel_2x8c8__neon_mlal() local
172 vacc0x4 = vpadalq_s16(vacc0x4, vprod0x4); in xnn_qs8_gemm_minmax_fp32_ukernel_2x8c8__neon_mlal()
D2x8c8-minmax-rndnu-neon-mlal.c113 int16x8_t vprod0x4 = vmull_s8(vb4x0, va0x0); in xnn_qs8_gemm_minmax_rndnu_ukernel_2x8c8__neon_mlal() local
115 vprod0x4 = vmlal_s8(vprod0x4, vb4x1, va0x1); in xnn_qs8_gemm_minmax_rndnu_ukernel_2x8c8__neon_mlal()
117 vacc0x4 = vpadalq_s16(vacc0x4, vprod0x4); in xnn_qs8_gemm_minmax_rndnu_ukernel_2x8c8__neon_mlal()
170 const int16x8_t vprod0x4 = vmull_s8(vb4, va0); in xnn_qs8_gemm_minmax_rndnu_ukernel_2x8c8__neon_mlal() local
172 vacc0x4 = vpadalq_s16(vacc0x4, vprod0x4); in xnn_qs8_gemm_minmax_rndnu_ukernel_2x8c8__neon_mlal()
D1x16c8-minmax-rndnu-neon-mlal.c101 int16x8_t vprod0x4 = vmull_s8(vb4x0, va0x0); in xnn_qs8_gemm_minmax_rndnu_ukernel_1x16c8__neon_mlal() local
102 vprod0x4 = vmlal_s8(vprod0x4, vb4x1, va0x1); in xnn_qs8_gemm_minmax_rndnu_ukernel_1x16c8__neon_mlal()
103 vacc0x4 = vpadalq_s16(vacc0x4, vprod0x4); in xnn_qs8_gemm_minmax_rndnu_ukernel_1x16c8__neon_mlal()
169 const int16x8_t vprod0x4 = vmull_s8(vb4, va0); in xnn_qs8_gemm_minmax_rndnu_ukernel_1x16c8__neon_mlal() local
170 vacc0x4 = vpadalq_s16(vacc0x4, vprod0x4); in xnn_qs8_gemm_minmax_rndnu_ukernel_1x16c8__neon_mlal()
D2x8c16-minmax-rndnu-neon-mlal.c106 int16x8_t vprod0x4 = vmull_s8(vget_low_s8(vb4), vget_low_s8(va0)); in xnn_qs8_gemm_minmax_rndnu_ukernel_2x8c16__neon_mlal() local
108 vprod0x4 = vmlal_s8(vprod0x4, vget_high_s8(vb4), vget_high_s8(va0)); in xnn_qs8_gemm_minmax_rndnu_ukernel_2x8c16__neon_mlal()
110 vacc0x4 = vpadalq_s16(vacc0x4, vprod0x4); in xnn_qs8_gemm_minmax_rndnu_ukernel_2x8c16__neon_mlal()
/external/XNNPACK/src/qc8-gemm/gen/
D1x8c8-minmax-fp32-neon-mlal.c85 int16x8_t vprod0x4 = vmull_s8(vb4x0, va0x0); in xnn_qc8_gemm_minmax_fp32_ukernel_1x8c8__neon_mlal() local
86 vprod0x4 = vmlal_s8(vprod0x4, vb4x1, va0x1); in xnn_qc8_gemm_minmax_fp32_ukernel_1x8c8__neon_mlal()
87 vacc0x4 = vpadalq_s16(vacc0x4, vprod0x4); in xnn_qc8_gemm_minmax_fp32_ukernel_1x8c8__neon_mlal()
121 const int16x8_t vprod0x4 = vmull_s8(vb4, va0); in xnn_qc8_gemm_minmax_fp32_ukernel_1x8c8__neon_mlal() local
122 vacc0x4 = vpadalq_s16(vacc0x4, vprod0x4); in xnn_qc8_gemm_minmax_fp32_ukernel_1x8c8__neon_mlal()
D1x8c8-minmax-fp32-neonv8-mlal.c86 int16x8_t vprod0x4 = vmull_s8(vb4x0, va0x0); in xnn_qc8_gemm_minmax_fp32_ukernel_1x8c8__neonv8_mlal() local
87 vprod0x4 = vmlal_s8(vprod0x4, vb4x1, va0x1); in xnn_qc8_gemm_minmax_fp32_ukernel_1x8c8__neonv8_mlal()
88 vacc0x4 = vpadalq_s16(vacc0x4, vprod0x4); in xnn_qc8_gemm_minmax_fp32_ukernel_1x8c8__neonv8_mlal()
122 const int16x8_t vprod0x4 = vmull_s8(vb4, va0); in xnn_qc8_gemm_minmax_fp32_ukernel_1x8c8__neonv8_mlal() local
123 vacc0x4 = vpadalq_s16(vacc0x4, vprod0x4); in xnn_qc8_gemm_minmax_fp32_ukernel_1x8c8__neonv8_mlal()
D2x8c8-minmax-fp32-neon-mlal.c113 int16x8_t vprod0x4 = vmull_s8(vb4x0, va0x0); in xnn_qc8_gemm_minmax_fp32_ukernel_2x8c8__neon_mlal() local
115 vprod0x4 = vmlal_s8(vprod0x4, vb4x1, va0x1); in xnn_qc8_gemm_minmax_fp32_ukernel_2x8c8__neon_mlal()
117 vacc0x4 = vpadalq_s16(vacc0x4, vprod0x4); in xnn_qc8_gemm_minmax_fp32_ukernel_2x8c8__neon_mlal()
170 const int16x8_t vprod0x4 = vmull_s8(vb4, va0); in xnn_qc8_gemm_minmax_fp32_ukernel_2x8c8__neon_mlal() local
172 vacc0x4 = vpadalq_s16(vacc0x4, vprod0x4); in xnn_qc8_gemm_minmax_fp32_ukernel_2x8c8__neon_mlal()
D2x8c8-minmax-fp32-neonv8-mlal.c114 int16x8_t vprod0x4 = vmull_s8(vb4x0, va0x0); in xnn_qc8_gemm_minmax_fp32_ukernel_2x8c8__neonv8_mlal() local
116 vprod0x4 = vmlal_s8(vprod0x4, vb4x1, va0x1); in xnn_qc8_gemm_minmax_fp32_ukernel_2x8c8__neonv8_mlal()
118 vacc0x4 = vpadalq_s16(vacc0x4, vprod0x4); in xnn_qc8_gemm_minmax_fp32_ukernel_2x8c8__neonv8_mlal()
171 const int16x8_t vprod0x4 = vmull_s8(vb4, va0); in xnn_qc8_gemm_minmax_fp32_ukernel_2x8c8__neonv8_mlal() local
173 vacc0x4 = vpadalq_s16(vacc0x4, vprod0x4); in xnn_qc8_gemm_minmax_fp32_ukernel_2x8c8__neonv8_mlal()
/external/XNNPACK/src/qs8-igemm/gen/
D1x8c8-minmax-fp32-neonv8-mlal.c97 int16x8_t vprod0x4 = vmull_s8(vb4x0, va0x0); in xnn_qs8_igemm_minmax_fp32_ukernel_1x8c8__neonv8_mlal() local
98 vprod0x4 = vmlal_s8(vprod0x4, vb4x1, va0x1); in xnn_qs8_igemm_minmax_fp32_ukernel_1x8c8__neonv8_mlal()
99 vacc0x4 = vpadalq_s16(vacc0x4, vprod0x4); in xnn_qs8_igemm_minmax_fp32_ukernel_1x8c8__neonv8_mlal()
133 const int16x8_t vprod0x4 = vmull_s8(vb4, va0); in xnn_qs8_igemm_minmax_fp32_ukernel_1x8c8__neonv8_mlal() local
134 vacc0x4 = vpadalq_s16(vacc0x4, vprod0x4); in xnn_qs8_igemm_minmax_fp32_ukernel_1x8c8__neonv8_mlal()
D1x8c8-minmax-rndnu-neon-mlal.c96 int16x8_t vprod0x4 = vmull_s8(vb4x0, va0x0); in xnn_qs8_igemm_minmax_rndnu_ukernel_1x8c8__neon_mlal() local
97 vprod0x4 = vmlal_s8(vprod0x4, vb4x1, va0x1); in xnn_qs8_igemm_minmax_rndnu_ukernel_1x8c8__neon_mlal()
98 vacc0x4 = vpadalq_s16(vacc0x4, vprod0x4); in xnn_qs8_igemm_minmax_rndnu_ukernel_1x8c8__neon_mlal()
132 const int16x8_t vprod0x4 = vmull_s8(vb4, va0); in xnn_qs8_igemm_minmax_rndnu_ukernel_1x8c8__neon_mlal() local
133 vacc0x4 = vpadalq_s16(vacc0x4, vprod0x4); in xnn_qs8_igemm_minmax_rndnu_ukernel_1x8c8__neon_mlal()
D1x8c8-minmax-fp32-neon-mlal.c96 int16x8_t vprod0x4 = vmull_s8(vb4x0, va0x0); in xnn_qs8_igemm_minmax_fp32_ukernel_1x8c8__neon_mlal() local
97 vprod0x4 = vmlal_s8(vprod0x4, vb4x1, va0x1); in xnn_qs8_igemm_minmax_fp32_ukernel_1x8c8__neon_mlal()
98 vacc0x4 = vpadalq_s16(vacc0x4, vprod0x4); in xnn_qs8_igemm_minmax_fp32_ukernel_1x8c8__neon_mlal()
132 const int16x8_t vprod0x4 = vmull_s8(vb4, va0); in xnn_qs8_igemm_minmax_fp32_ukernel_1x8c8__neon_mlal() local
133 vacc0x4 = vpadalq_s16(vacc0x4, vprod0x4); in xnn_qs8_igemm_minmax_fp32_ukernel_1x8c8__neon_mlal()
D1x8c16-minmax-rndnu-neon-mlal.c90 int16x8_t vprod0x4 = vmull_s8(vget_low_s8(vb4), vget_low_s8(va0)); in xnn_qs8_igemm_minmax_rndnu_ukernel_1x8c16__neon_mlal() local
91 vprod0x4 = vmlal_s8(vprod0x4, vget_high_s8(vb4), vget_high_s8(va0)); in xnn_qs8_igemm_minmax_rndnu_ukernel_1x8c16__neon_mlal()
92 vacc0x4 = vpadalq_s16(vacc0x4, vprod0x4); in xnn_qs8_igemm_minmax_rndnu_ukernel_1x8c16__neon_mlal()
D2x8c8-minmax-rndnu-neon-mlal.c126 int16x8_t vprod0x4 = vmull_s8(vb4x0, va0x0); in xnn_qs8_igemm_minmax_rndnu_ukernel_2x8c8__neon_mlal() local
128 vprod0x4 = vmlal_s8(vprod0x4, vb4x1, va0x1); in xnn_qs8_igemm_minmax_rndnu_ukernel_2x8c8__neon_mlal()
130 vacc0x4 = vpadalq_s16(vacc0x4, vprod0x4); in xnn_qs8_igemm_minmax_rndnu_ukernel_2x8c8__neon_mlal()
183 const int16x8_t vprod0x4 = vmull_s8(vb4, va0); in xnn_qs8_igemm_minmax_rndnu_ukernel_2x8c8__neon_mlal() local
185 vacc0x4 = vpadalq_s16(vacc0x4, vprod0x4); in xnn_qs8_igemm_minmax_rndnu_ukernel_2x8c8__neon_mlal()
D2x8c8-minmax-fp32-neon-mlal.c126 int16x8_t vprod0x4 = vmull_s8(vb4x0, va0x0); in xnn_qs8_igemm_minmax_fp32_ukernel_2x8c8__neon_mlal() local
128 vprod0x4 = vmlal_s8(vprod0x4, vb4x1, va0x1); in xnn_qs8_igemm_minmax_fp32_ukernel_2x8c8__neon_mlal()
130 vacc0x4 = vpadalq_s16(vacc0x4, vprod0x4); in xnn_qs8_igemm_minmax_fp32_ukernel_2x8c8__neon_mlal()
183 const int16x8_t vprod0x4 = vmull_s8(vb4, va0); in xnn_qs8_igemm_minmax_fp32_ukernel_2x8c8__neon_mlal() local
185 vacc0x4 = vpadalq_s16(vacc0x4, vprod0x4); in xnn_qs8_igemm_minmax_fp32_ukernel_2x8c8__neon_mlal()
D2x8c8-minmax-fp32-neonv8-mlal.c127 int16x8_t vprod0x4 = vmull_s8(vb4x0, va0x0); in xnn_qs8_igemm_minmax_fp32_ukernel_2x8c8__neonv8_mlal() local
129 vprod0x4 = vmlal_s8(vprod0x4, vb4x1, va0x1); in xnn_qs8_igemm_minmax_fp32_ukernel_2x8c8__neonv8_mlal()
131 vacc0x4 = vpadalq_s16(vacc0x4, vprod0x4); in xnn_qs8_igemm_minmax_fp32_ukernel_2x8c8__neonv8_mlal()
184 const int16x8_t vprod0x4 = vmull_s8(vb4, va0); in xnn_qs8_igemm_minmax_fp32_ukernel_2x8c8__neonv8_mlal() local
186 vacc0x4 = vpadalq_s16(vacc0x4, vprod0x4); in xnn_qs8_igemm_minmax_fp32_ukernel_2x8c8__neonv8_mlal()
D1x16c8-minmax-rndnu-neon-mlal.c112 int16x8_t vprod0x4 = vmull_s8(vb4x0, va0x0); in xnn_qs8_igemm_minmax_rndnu_ukernel_1x16c8__neon_mlal() local
113 vprod0x4 = vmlal_s8(vprod0x4, vb4x1, va0x1); in xnn_qs8_igemm_minmax_rndnu_ukernel_1x16c8__neon_mlal()
114 vacc0x4 = vpadalq_s16(vacc0x4, vprod0x4); in xnn_qs8_igemm_minmax_rndnu_ukernel_1x16c8__neon_mlal()
180 const int16x8_t vprod0x4 = vmull_s8(vb4, va0); in xnn_qs8_igemm_minmax_rndnu_ukernel_1x16c8__neon_mlal() local
181 vacc0x4 = vpadalq_s16(vacc0x4, vprod0x4); in xnn_qs8_igemm_minmax_rndnu_ukernel_1x16c8__neon_mlal()
/external/XNNPACK/src/qc8-igemm/gen/
D1x8c8-minmax-fp32-neon-mlal.c96 int16x8_t vprod0x4 = vmull_s8(vb4x0, va0x0); in xnn_qc8_igemm_minmax_fp32_ukernel_1x8c8__neon_mlal() local
97 vprod0x4 = vmlal_s8(vprod0x4, vb4x1, va0x1); in xnn_qc8_igemm_minmax_fp32_ukernel_1x8c8__neon_mlal()
98 vacc0x4 = vpadalq_s16(vacc0x4, vprod0x4); in xnn_qc8_igemm_minmax_fp32_ukernel_1x8c8__neon_mlal()
132 const int16x8_t vprod0x4 = vmull_s8(vb4, va0); in xnn_qc8_igemm_minmax_fp32_ukernel_1x8c8__neon_mlal() local
133 vacc0x4 = vpadalq_s16(vacc0x4, vprod0x4); in xnn_qc8_igemm_minmax_fp32_ukernel_1x8c8__neon_mlal()
D1x8c8-minmax-fp32-neonv8-mlal.c97 int16x8_t vprod0x4 = vmull_s8(vb4x0, va0x0); in xnn_qc8_igemm_minmax_fp32_ukernel_1x8c8__neonv8_mlal() local
98 vprod0x4 = vmlal_s8(vprod0x4, vb4x1, va0x1); in xnn_qc8_igemm_minmax_fp32_ukernel_1x8c8__neonv8_mlal()
99 vacc0x4 = vpadalq_s16(vacc0x4, vprod0x4); in xnn_qc8_igemm_minmax_fp32_ukernel_1x8c8__neonv8_mlal()
133 const int16x8_t vprod0x4 = vmull_s8(vb4, va0); in xnn_qc8_igemm_minmax_fp32_ukernel_1x8c8__neonv8_mlal() local
134 vacc0x4 = vpadalq_s16(vacc0x4, vprod0x4); in xnn_qc8_igemm_minmax_fp32_ukernel_1x8c8__neonv8_mlal()
D2x8c8-minmax-fp32-neonv8-mlal.c127 int16x8_t vprod0x4 = vmull_s8(vb4x0, va0x0); in xnn_qc8_igemm_minmax_fp32_ukernel_2x8c8__neonv8_mlal() local
129 vprod0x4 = vmlal_s8(vprod0x4, vb4x1, va0x1); in xnn_qc8_igemm_minmax_fp32_ukernel_2x8c8__neonv8_mlal()
131 vacc0x4 = vpadalq_s16(vacc0x4, vprod0x4); in xnn_qc8_igemm_minmax_fp32_ukernel_2x8c8__neonv8_mlal()
184 const int16x8_t vprod0x4 = vmull_s8(vb4, va0); in xnn_qc8_igemm_minmax_fp32_ukernel_2x8c8__neonv8_mlal() local
186 vacc0x4 = vpadalq_s16(vacc0x4, vprod0x4); in xnn_qc8_igemm_minmax_fp32_ukernel_2x8c8__neonv8_mlal()
D2x8c8-minmax-fp32-neon-mlal.c126 int16x8_t vprod0x4 = vmull_s8(vb4x0, va0x0); in xnn_qc8_igemm_minmax_fp32_ukernel_2x8c8__neon_mlal() local
128 vprod0x4 = vmlal_s8(vprod0x4, vb4x1, va0x1); in xnn_qc8_igemm_minmax_fp32_ukernel_2x8c8__neon_mlal()
130 vacc0x4 = vpadalq_s16(vacc0x4, vprod0x4); in xnn_qc8_igemm_minmax_fp32_ukernel_2x8c8__neon_mlal()
183 const int16x8_t vprod0x4 = vmull_s8(vb4, va0); in xnn_qc8_igemm_minmax_fp32_ukernel_2x8c8__neon_mlal() local
185 vacc0x4 = vpadalq_s16(vacc0x4, vprod0x4); in xnn_qc8_igemm_minmax_fp32_ukernel_2x8c8__neon_mlal()

123