Home
last modified time | relevance | path

Searched refs:vprod0x6 (Results 1 – 25 of 64) sorted by relevance

123

/external/XNNPACK/src/qs8-gemm/gen/
D1x8c8-minmax-rndnu-neon-mlal.c93 int16x8_t vprod0x6 = vmull_s8(vb6x0, va0x0); in xnn_qs8_gemm_minmax_rndnu_ukernel_1x8c8__neon_mlal() local
94 vprod0x6 = vmlal_s8(vprod0x6, vb6x1, va0x1); in xnn_qs8_gemm_minmax_rndnu_ukernel_1x8c8__neon_mlal()
95 vacc0x6 = vpadalq_s16(vacc0x6, vprod0x6); in xnn_qs8_gemm_minmax_rndnu_ukernel_1x8c8__neon_mlal()
127 const int16x8_t vprod0x6 = vmull_s8(vb6, va0); in xnn_qs8_gemm_minmax_rndnu_ukernel_1x8c8__neon_mlal() local
128 vacc0x6 = vpadalq_s16(vacc0x6, vprod0x6); in xnn_qs8_gemm_minmax_rndnu_ukernel_1x8c8__neon_mlal()
D1x8c8-minmax-fp32-neon-mlal.c93 int16x8_t vprod0x6 = vmull_s8(vb6x0, va0x0); in xnn_qs8_gemm_minmax_fp32_ukernel_1x8c8__neon_mlal() local
94 vprod0x6 = vmlal_s8(vprod0x6, vb6x1, va0x1); in xnn_qs8_gemm_minmax_fp32_ukernel_1x8c8__neon_mlal()
95 vacc0x6 = vpadalq_s16(vacc0x6, vprod0x6); in xnn_qs8_gemm_minmax_fp32_ukernel_1x8c8__neon_mlal()
127 const int16x8_t vprod0x6 = vmull_s8(vb6, va0); in xnn_qs8_gemm_minmax_fp32_ukernel_1x8c8__neon_mlal() local
128 vacc0x6 = vpadalq_s16(vacc0x6, vprod0x6); in xnn_qs8_gemm_minmax_fp32_ukernel_1x8c8__neon_mlal()
D1x8c8-minmax-fp32-neonv8-mlal.c94 int16x8_t vprod0x6 = vmull_s8(vb6x0, va0x0); in xnn_qs8_gemm_minmax_fp32_ukernel_1x8c8__neonv8_mlal() local
95 vprod0x6 = vmlal_s8(vprod0x6, vb6x1, va0x1); in xnn_qs8_gemm_minmax_fp32_ukernel_1x8c8__neonv8_mlal()
96 vacc0x6 = vpadalq_s16(vacc0x6, vprod0x6); in xnn_qs8_gemm_minmax_fp32_ukernel_1x8c8__neonv8_mlal()
128 const int16x8_t vprod0x6 = vmull_s8(vb6, va0); in xnn_qs8_gemm_minmax_fp32_ukernel_1x8c8__neonv8_mlal() local
129 vacc0x6 = vpadalq_s16(vacc0x6, vprod0x6); in xnn_qs8_gemm_minmax_fp32_ukernel_1x8c8__neonv8_mlal()
D1x8c16-minmax-rndnu-neon-mlal.c85 int16x8_t vprod0x6 = vmull_s8(vget_low_s8(vb6), vget_low_s8(va0)); in xnn_qs8_gemm_minmax_rndnu_ukernel_1x8c16__neon_mlal() local
86 vprod0x6 = vmlal_s8(vprod0x6, vget_high_s8(vb6), vget_high_s8(va0)); in xnn_qs8_gemm_minmax_rndnu_ukernel_1x8c16__neon_mlal()
87 vacc0x6 = vpadalq_s16(vacc0x6, vprod0x6); in xnn_qs8_gemm_minmax_rndnu_ukernel_1x8c16__neon_mlal()
D2x8c8-minmax-fp32-neonv8-mlal.c128 int16x8_t vprod0x6 = vmull_s8(vb6x0, va0x0); in xnn_qs8_gemm_minmax_fp32_ukernel_2x8c8__neonv8_mlal() local
130 vprod0x6 = vmlal_s8(vprod0x6, vb6x1, va0x1); in xnn_qs8_gemm_minmax_fp32_ukernel_2x8c8__neonv8_mlal()
132 vacc0x6 = vpadalq_s16(vacc0x6, vprod0x6); in xnn_qs8_gemm_minmax_fp32_ukernel_2x8c8__neonv8_mlal()
181 const int16x8_t vprod0x6 = vmull_s8(vb6, va0); in xnn_qs8_gemm_minmax_fp32_ukernel_2x8c8__neonv8_mlal() local
183 vacc0x6 = vpadalq_s16(vacc0x6, vprod0x6); in xnn_qs8_gemm_minmax_fp32_ukernel_2x8c8__neonv8_mlal()
D2x8c8-minmax-fp32-neon-mlal.c127 int16x8_t vprod0x6 = vmull_s8(vb6x0, va0x0); in xnn_qs8_gemm_minmax_fp32_ukernel_2x8c8__neon_mlal() local
129 vprod0x6 = vmlal_s8(vprod0x6, vb6x1, va0x1); in xnn_qs8_gemm_minmax_fp32_ukernel_2x8c8__neon_mlal()
131 vacc0x6 = vpadalq_s16(vacc0x6, vprod0x6); in xnn_qs8_gemm_minmax_fp32_ukernel_2x8c8__neon_mlal()
180 const int16x8_t vprod0x6 = vmull_s8(vb6, va0); in xnn_qs8_gemm_minmax_fp32_ukernel_2x8c8__neon_mlal() local
182 vacc0x6 = vpadalq_s16(vacc0x6, vprod0x6); in xnn_qs8_gemm_minmax_fp32_ukernel_2x8c8__neon_mlal()
D2x8c8-minmax-rndnu-neon-mlal.c127 int16x8_t vprod0x6 = vmull_s8(vb6x0, va0x0); in xnn_qs8_gemm_minmax_rndnu_ukernel_2x8c8__neon_mlal() local
129 vprod0x6 = vmlal_s8(vprod0x6, vb6x1, va0x1); in xnn_qs8_gemm_minmax_rndnu_ukernel_2x8c8__neon_mlal()
131 vacc0x6 = vpadalq_s16(vacc0x6, vprod0x6); in xnn_qs8_gemm_minmax_rndnu_ukernel_2x8c8__neon_mlal()
180 const int16x8_t vprod0x6 = vmull_s8(vb6, va0); in xnn_qs8_gemm_minmax_rndnu_ukernel_2x8c8__neon_mlal() local
182 vacc0x6 = vpadalq_s16(vacc0x6, vprod0x6); in xnn_qs8_gemm_minmax_rndnu_ukernel_2x8c8__neon_mlal()
D1x16c8-minmax-rndnu-neon-mlal.c109 int16x8_t vprod0x6 = vmull_s8(vb6x0, va0x0); in xnn_qs8_gemm_minmax_rndnu_ukernel_1x16c8__neon_mlal() local
110 vprod0x6 = vmlal_s8(vprod0x6, vb6x1, va0x1); in xnn_qs8_gemm_minmax_rndnu_ukernel_1x16c8__neon_mlal()
111 vacc0x6 = vpadalq_s16(vacc0x6, vprod0x6); in xnn_qs8_gemm_minmax_rndnu_ukernel_1x16c8__neon_mlal()
175 const int16x8_t vprod0x6 = vmull_s8(vb6, va0); in xnn_qs8_gemm_minmax_rndnu_ukernel_1x16c8__neon_mlal() local
176 vacc0x6 = vpadalq_s16(vacc0x6, vprod0x6); in xnn_qs8_gemm_minmax_rndnu_ukernel_1x16c8__neon_mlal()
D2x8c16-minmax-rndnu-neon-mlal.c118 int16x8_t vprod0x6 = vmull_s8(vget_low_s8(vb6), vget_low_s8(va0)); in xnn_qs8_gemm_minmax_rndnu_ukernel_2x8c16__neon_mlal() local
120 vprod0x6 = vmlal_s8(vprod0x6, vget_high_s8(vb6), vget_high_s8(va0)); in xnn_qs8_gemm_minmax_rndnu_ukernel_2x8c16__neon_mlal()
122 vacc0x6 = vpadalq_s16(vacc0x6, vprod0x6); in xnn_qs8_gemm_minmax_rndnu_ukernel_2x8c16__neon_mlal()
/external/XNNPACK/src/qc8-gemm/gen/
D1x8c8-minmax-fp32-neon-mlal.c93 int16x8_t vprod0x6 = vmull_s8(vb6x0, va0x0); in xnn_qc8_gemm_minmax_fp32_ukernel_1x8c8__neon_mlal() local
94 vprod0x6 = vmlal_s8(vprod0x6, vb6x1, va0x1); in xnn_qc8_gemm_minmax_fp32_ukernel_1x8c8__neon_mlal()
95 vacc0x6 = vpadalq_s16(vacc0x6, vprod0x6); in xnn_qc8_gemm_minmax_fp32_ukernel_1x8c8__neon_mlal()
127 const int16x8_t vprod0x6 = vmull_s8(vb6, va0); in xnn_qc8_gemm_minmax_fp32_ukernel_1x8c8__neon_mlal() local
128 vacc0x6 = vpadalq_s16(vacc0x6, vprod0x6); in xnn_qc8_gemm_minmax_fp32_ukernel_1x8c8__neon_mlal()
D1x8c8-minmax-fp32-neonv8-mlal.c94 int16x8_t vprod0x6 = vmull_s8(vb6x0, va0x0); in xnn_qc8_gemm_minmax_fp32_ukernel_1x8c8__neonv8_mlal() local
95 vprod0x6 = vmlal_s8(vprod0x6, vb6x1, va0x1); in xnn_qc8_gemm_minmax_fp32_ukernel_1x8c8__neonv8_mlal()
96 vacc0x6 = vpadalq_s16(vacc0x6, vprod0x6); in xnn_qc8_gemm_minmax_fp32_ukernel_1x8c8__neonv8_mlal()
128 const int16x8_t vprod0x6 = vmull_s8(vb6, va0); in xnn_qc8_gemm_minmax_fp32_ukernel_1x8c8__neonv8_mlal() local
129 vacc0x6 = vpadalq_s16(vacc0x6, vprod0x6); in xnn_qc8_gemm_minmax_fp32_ukernel_1x8c8__neonv8_mlal()
D2x8c8-minmax-fp32-neon-mlal.c127 int16x8_t vprod0x6 = vmull_s8(vb6x0, va0x0); in xnn_qc8_gemm_minmax_fp32_ukernel_2x8c8__neon_mlal() local
129 vprod0x6 = vmlal_s8(vprod0x6, vb6x1, va0x1); in xnn_qc8_gemm_minmax_fp32_ukernel_2x8c8__neon_mlal()
131 vacc0x6 = vpadalq_s16(vacc0x6, vprod0x6); in xnn_qc8_gemm_minmax_fp32_ukernel_2x8c8__neon_mlal()
180 const int16x8_t vprod0x6 = vmull_s8(vb6, va0); in xnn_qc8_gemm_minmax_fp32_ukernel_2x8c8__neon_mlal() local
182 vacc0x6 = vpadalq_s16(vacc0x6, vprod0x6); in xnn_qc8_gemm_minmax_fp32_ukernel_2x8c8__neon_mlal()
D2x8c8-minmax-fp32-neonv8-mlal.c128 int16x8_t vprod0x6 = vmull_s8(vb6x0, va0x0); in xnn_qc8_gemm_minmax_fp32_ukernel_2x8c8__neonv8_mlal() local
130 vprod0x6 = vmlal_s8(vprod0x6, vb6x1, va0x1); in xnn_qc8_gemm_minmax_fp32_ukernel_2x8c8__neonv8_mlal()
132 vacc0x6 = vpadalq_s16(vacc0x6, vprod0x6); in xnn_qc8_gemm_minmax_fp32_ukernel_2x8c8__neonv8_mlal()
181 const int16x8_t vprod0x6 = vmull_s8(vb6, va0); in xnn_qc8_gemm_minmax_fp32_ukernel_2x8c8__neonv8_mlal() local
183 vacc0x6 = vpadalq_s16(vacc0x6, vprod0x6); in xnn_qc8_gemm_minmax_fp32_ukernel_2x8c8__neonv8_mlal()
/external/XNNPACK/src/qs8-igemm/gen/
D1x8c8-minmax-fp32-neonv8-mlal.c105 int16x8_t vprod0x6 = vmull_s8(vb6x0, va0x0); in xnn_qs8_igemm_minmax_fp32_ukernel_1x8c8__neonv8_mlal() local
106 vprod0x6 = vmlal_s8(vprod0x6, vb6x1, va0x1); in xnn_qs8_igemm_minmax_fp32_ukernel_1x8c8__neonv8_mlal()
107 vacc0x6 = vpadalq_s16(vacc0x6, vprod0x6); in xnn_qs8_igemm_minmax_fp32_ukernel_1x8c8__neonv8_mlal()
139 const int16x8_t vprod0x6 = vmull_s8(vb6, va0); in xnn_qs8_igemm_minmax_fp32_ukernel_1x8c8__neonv8_mlal() local
140 vacc0x6 = vpadalq_s16(vacc0x6, vprod0x6); in xnn_qs8_igemm_minmax_fp32_ukernel_1x8c8__neonv8_mlal()
D1x8c8-minmax-rndnu-neon-mlal.c104 int16x8_t vprod0x6 = vmull_s8(vb6x0, va0x0); in xnn_qs8_igemm_minmax_rndnu_ukernel_1x8c8__neon_mlal() local
105 vprod0x6 = vmlal_s8(vprod0x6, vb6x1, va0x1); in xnn_qs8_igemm_minmax_rndnu_ukernel_1x8c8__neon_mlal()
106 vacc0x6 = vpadalq_s16(vacc0x6, vprod0x6); in xnn_qs8_igemm_minmax_rndnu_ukernel_1x8c8__neon_mlal()
138 const int16x8_t vprod0x6 = vmull_s8(vb6, va0); in xnn_qs8_igemm_minmax_rndnu_ukernel_1x8c8__neon_mlal() local
139 vacc0x6 = vpadalq_s16(vacc0x6, vprod0x6); in xnn_qs8_igemm_minmax_rndnu_ukernel_1x8c8__neon_mlal()
D1x8c8-minmax-fp32-neon-mlal.c104 int16x8_t vprod0x6 = vmull_s8(vb6x0, va0x0); in xnn_qs8_igemm_minmax_fp32_ukernel_1x8c8__neon_mlal() local
105 vprod0x6 = vmlal_s8(vprod0x6, vb6x1, va0x1); in xnn_qs8_igemm_minmax_fp32_ukernel_1x8c8__neon_mlal()
106 vacc0x6 = vpadalq_s16(vacc0x6, vprod0x6); in xnn_qs8_igemm_minmax_fp32_ukernel_1x8c8__neon_mlal()
138 const int16x8_t vprod0x6 = vmull_s8(vb6, va0); in xnn_qs8_igemm_minmax_fp32_ukernel_1x8c8__neon_mlal() local
139 vacc0x6 = vpadalq_s16(vacc0x6, vprod0x6); in xnn_qs8_igemm_minmax_fp32_ukernel_1x8c8__neon_mlal()
D1x8c16-minmax-rndnu-neon-mlal.c96 int16x8_t vprod0x6 = vmull_s8(vget_low_s8(vb6), vget_low_s8(va0)); in xnn_qs8_igemm_minmax_rndnu_ukernel_1x8c16__neon_mlal() local
97 vprod0x6 = vmlal_s8(vprod0x6, vget_high_s8(vb6), vget_high_s8(va0)); in xnn_qs8_igemm_minmax_rndnu_ukernel_1x8c16__neon_mlal()
98 vacc0x6 = vpadalq_s16(vacc0x6, vprod0x6); in xnn_qs8_igemm_minmax_rndnu_ukernel_1x8c16__neon_mlal()
D2x8c8-minmax-rndnu-neon-mlal.c140 int16x8_t vprod0x6 = vmull_s8(vb6x0, va0x0); in xnn_qs8_igemm_minmax_rndnu_ukernel_2x8c8__neon_mlal() local
142 vprod0x6 = vmlal_s8(vprod0x6, vb6x1, va0x1); in xnn_qs8_igemm_minmax_rndnu_ukernel_2x8c8__neon_mlal()
144 vacc0x6 = vpadalq_s16(vacc0x6, vprod0x6); in xnn_qs8_igemm_minmax_rndnu_ukernel_2x8c8__neon_mlal()
193 const int16x8_t vprod0x6 = vmull_s8(vb6, va0); in xnn_qs8_igemm_minmax_rndnu_ukernel_2x8c8__neon_mlal() local
195 vacc0x6 = vpadalq_s16(vacc0x6, vprod0x6); in xnn_qs8_igemm_minmax_rndnu_ukernel_2x8c8__neon_mlal()
D2x8c8-minmax-fp32-neon-mlal.c140 int16x8_t vprod0x6 = vmull_s8(vb6x0, va0x0); in xnn_qs8_igemm_minmax_fp32_ukernel_2x8c8__neon_mlal() local
142 vprod0x6 = vmlal_s8(vprod0x6, vb6x1, va0x1); in xnn_qs8_igemm_minmax_fp32_ukernel_2x8c8__neon_mlal()
144 vacc0x6 = vpadalq_s16(vacc0x6, vprod0x6); in xnn_qs8_igemm_minmax_fp32_ukernel_2x8c8__neon_mlal()
193 const int16x8_t vprod0x6 = vmull_s8(vb6, va0); in xnn_qs8_igemm_minmax_fp32_ukernel_2x8c8__neon_mlal() local
195 vacc0x6 = vpadalq_s16(vacc0x6, vprod0x6); in xnn_qs8_igemm_minmax_fp32_ukernel_2x8c8__neon_mlal()
D2x8c8-minmax-fp32-neonv8-mlal.c141 int16x8_t vprod0x6 = vmull_s8(vb6x0, va0x0); in xnn_qs8_igemm_minmax_fp32_ukernel_2x8c8__neonv8_mlal() local
143 vprod0x6 = vmlal_s8(vprod0x6, vb6x1, va0x1); in xnn_qs8_igemm_minmax_fp32_ukernel_2x8c8__neonv8_mlal()
145 vacc0x6 = vpadalq_s16(vacc0x6, vprod0x6); in xnn_qs8_igemm_minmax_fp32_ukernel_2x8c8__neonv8_mlal()
194 const int16x8_t vprod0x6 = vmull_s8(vb6, va0); in xnn_qs8_igemm_minmax_fp32_ukernel_2x8c8__neonv8_mlal() local
196 vacc0x6 = vpadalq_s16(vacc0x6, vprod0x6); in xnn_qs8_igemm_minmax_fp32_ukernel_2x8c8__neonv8_mlal()
D1x16c8-minmax-rndnu-neon-mlal.c120 int16x8_t vprod0x6 = vmull_s8(vb6x0, va0x0); in xnn_qs8_igemm_minmax_rndnu_ukernel_1x16c8__neon_mlal() local
121 vprod0x6 = vmlal_s8(vprod0x6, vb6x1, va0x1); in xnn_qs8_igemm_minmax_rndnu_ukernel_1x16c8__neon_mlal()
122 vacc0x6 = vpadalq_s16(vacc0x6, vprod0x6); in xnn_qs8_igemm_minmax_rndnu_ukernel_1x16c8__neon_mlal()
186 const int16x8_t vprod0x6 = vmull_s8(vb6, va0); in xnn_qs8_igemm_minmax_rndnu_ukernel_1x16c8__neon_mlal() local
187 vacc0x6 = vpadalq_s16(vacc0x6, vprod0x6); in xnn_qs8_igemm_minmax_rndnu_ukernel_1x16c8__neon_mlal()
/external/XNNPACK/src/qc8-igemm/gen/
D1x8c8-minmax-fp32-neon-mlal.c104 int16x8_t vprod0x6 = vmull_s8(vb6x0, va0x0); in xnn_qc8_igemm_minmax_fp32_ukernel_1x8c8__neon_mlal() local
105 vprod0x6 = vmlal_s8(vprod0x6, vb6x1, va0x1); in xnn_qc8_igemm_minmax_fp32_ukernel_1x8c8__neon_mlal()
106 vacc0x6 = vpadalq_s16(vacc0x6, vprod0x6); in xnn_qc8_igemm_minmax_fp32_ukernel_1x8c8__neon_mlal()
138 const int16x8_t vprod0x6 = vmull_s8(vb6, va0); in xnn_qc8_igemm_minmax_fp32_ukernel_1x8c8__neon_mlal() local
139 vacc0x6 = vpadalq_s16(vacc0x6, vprod0x6); in xnn_qc8_igemm_minmax_fp32_ukernel_1x8c8__neon_mlal()
D1x8c8-minmax-fp32-neonv8-mlal.c105 int16x8_t vprod0x6 = vmull_s8(vb6x0, va0x0); in xnn_qc8_igemm_minmax_fp32_ukernel_1x8c8__neonv8_mlal() local
106 vprod0x6 = vmlal_s8(vprod0x6, vb6x1, va0x1); in xnn_qc8_igemm_minmax_fp32_ukernel_1x8c8__neonv8_mlal()
107 vacc0x6 = vpadalq_s16(vacc0x6, vprod0x6); in xnn_qc8_igemm_minmax_fp32_ukernel_1x8c8__neonv8_mlal()
139 const int16x8_t vprod0x6 = vmull_s8(vb6, va0); in xnn_qc8_igemm_minmax_fp32_ukernel_1x8c8__neonv8_mlal() local
140 vacc0x6 = vpadalq_s16(vacc0x6, vprod0x6); in xnn_qc8_igemm_minmax_fp32_ukernel_1x8c8__neonv8_mlal()
D2x8c8-minmax-fp32-neonv8-mlal.c141 int16x8_t vprod0x6 = vmull_s8(vb6x0, va0x0); in xnn_qc8_igemm_minmax_fp32_ukernel_2x8c8__neonv8_mlal() local
143 vprod0x6 = vmlal_s8(vprod0x6, vb6x1, va0x1); in xnn_qc8_igemm_minmax_fp32_ukernel_2x8c8__neonv8_mlal()
145 vacc0x6 = vpadalq_s16(vacc0x6, vprod0x6); in xnn_qc8_igemm_minmax_fp32_ukernel_2x8c8__neonv8_mlal()
194 const int16x8_t vprod0x6 = vmull_s8(vb6, va0); in xnn_qc8_igemm_minmax_fp32_ukernel_2x8c8__neonv8_mlal() local
196 vacc0x6 = vpadalq_s16(vacc0x6, vprod0x6); in xnn_qc8_igemm_minmax_fp32_ukernel_2x8c8__neonv8_mlal()
D2x8c8-minmax-fp32-neon-mlal.c140 int16x8_t vprod0x6 = vmull_s8(vb6x0, va0x0); in xnn_qc8_igemm_minmax_fp32_ukernel_2x8c8__neon_mlal() local
142 vprod0x6 = vmlal_s8(vprod0x6, vb6x1, va0x1); in xnn_qc8_igemm_minmax_fp32_ukernel_2x8c8__neon_mlal()
144 vacc0x6 = vpadalq_s16(vacc0x6, vprod0x6); in xnn_qc8_igemm_minmax_fp32_ukernel_2x8c8__neon_mlal()
193 const int16x8_t vprod0x6 = vmull_s8(vb6, va0); in xnn_qc8_igemm_minmax_fp32_ukernel_2x8c8__neon_mlal() local
195 vacc0x6 = vpadalq_s16(vacc0x6, vprod0x6); in xnn_qc8_igemm_minmax_fp32_ukernel_2x8c8__neon_mlal()

123