Home
last modified time | relevance | path

Searched refs:vnacc0x0123 (Results 1 – 25 of 40) sorted by relevance

12

/external/XNNPACK/src/qu8-igemm/gen/
D1x32c4-minmax-rndnu-neondot.c146 const uint32x4_t vnacc0x0123 = vcombine_u32(vnacc0, vnacc0); in xnn_qu8_igemm_minmax_rndnu_ukernel_1x32c4__neondot() local
147 int32x4_t vacc0x0123 = vreinterpretq_s32_u32(vsubq_u32(vpacc0x0123, vnacc0x0123)); in xnn_qu8_igemm_minmax_rndnu_ukernel_1x32c4__neondot()
148 int32x4_t vacc0x4567 = vreinterpretq_s32_u32(vsubq_u32(vpacc0x4567, vnacc0x0123)); in xnn_qu8_igemm_minmax_rndnu_ukernel_1x32c4__neondot()
149 int32x4_t vacc0x89AB = vreinterpretq_s32_u32(vsubq_u32(vpacc0x89AB, vnacc0x0123)); in xnn_qu8_igemm_minmax_rndnu_ukernel_1x32c4__neondot()
150 int32x4_t vacc0xCDEF = vreinterpretq_s32_u32(vsubq_u32(vpacc0xCDEF, vnacc0x0123)); in xnn_qu8_igemm_minmax_rndnu_ukernel_1x32c4__neondot()
151 int32x4_t vacc0xGHIJ = vreinterpretq_s32_u32(vsubq_u32(vpacc0xGHIJ, vnacc0x0123)); in xnn_qu8_igemm_minmax_rndnu_ukernel_1x32c4__neondot()
152 int32x4_t vacc0xKLMN = vreinterpretq_s32_u32(vsubq_u32(vpacc0xKLMN, vnacc0x0123)); in xnn_qu8_igemm_minmax_rndnu_ukernel_1x32c4__neondot()
153 int32x4_t vacc0xOPQR = vreinterpretq_s32_u32(vsubq_u32(vpacc0xOPQR, vnacc0x0123)); in xnn_qu8_igemm_minmax_rndnu_ukernel_1x32c4__neondot()
154 int32x4_t vacc0xSTUV = vreinterpretq_s32_u32(vsubq_u32(vpacc0xSTUV, vnacc0x0123)); in xnn_qu8_igemm_minmax_rndnu_ukernel_1x32c4__neondot()
D1x16c4-minmax-fp32-neondot.c119 const uint32x4_t vnacc0x0123 = vcombine_u32(vnacc0, vnacc0); in xnn_qu8_igemm_minmax_fp32_ukernel_1x16c4__neondot() local
120 int32x4_t vacc0x0123 = vreinterpretq_s32_u32(vsubq_u32(vpacc0x0123, vnacc0x0123)); in xnn_qu8_igemm_minmax_fp32_ukernel_1x16c4__neondot()
121 int32x4_t vacc0x4567 = vreinterpretq_s32_u32(vsubq_u32(vpacc0x4567, vnacc0x0123)); in xnn_qu8_igemm_minmax_fp32_ukernel_1x16c4__neondot()
122 int32x4_t vacc0x89AB = vreinterpretq_s32_u32(vsubq_u32(vpacc0x89AB, vnacc0x0123)); in xnn_qu8_igemm_minmax_fp32_ukernel_1x16c4__neondot()
123 int32x4_t vacc0xCDEF = vreinterpretq_s32_u32(vsubq_u32(vpacc0xCDEF, vnacc0x0123)); in xnn_qu8_igemm_minmax_fp32_ukernel_1x16c4__neondot()
D1x16c4-minmax-rndnu-neondot.c118 const uint32x4_t vnacc0x0123 = vcombine_u32(vnacc0, vnacc0); in xnn_qu8_igemm_minmax_rndnu_ukernel_1x16c4__neondot() local
119 int32x4_t vacc0x0123 = vreinterpretq_s32_u32(vsubq_u32(vpacc0x0123, vnacc0x0123)); in xnn_qu8_igemm_minmax_rndnu_ukernel_1x16c4__neondot()
120 int32x4_t vacc0x4567 = vreinterpretq_s32_u32(vsubq_u32(vpacc0x4567, vnacc0x0123)); in xnn_qu8_igemm_minmax_rndnu_ukernel_1x16c4__neondot()
121 int32x4_t vacc0x89AB = vreinterpretq_s32_u32(vsubq_u32(vpacc0x89AB, vnacc0x0123)); in xnn_qu8_igemm_minmax_rndnu_ukernel_1x16c4__neondot()
122 int32x4_t vacc0xCDEF = vreinterpretq_s32_u32(vsubq_u32(vpacc0xCDEF, vnacc0x0123)); in xnn_qu8_igemm_minmax_rndnu_ukernel_1x16c4__neondot()
D2x32c4-minmax-rndnu-neondot.c191 const uint32x4_t vnacc0x0123 = vcombine_u32(vnacc0, vnacc0); in xnn_qu8_igemm_minmax_rndnu_ukernel_2x32c4__neondot() local
192 int32x4_t vacc0x0123 = vreinterpretq_s32_u32(vsubq_u32(vpacc0x0123, vnacc0x0123)); in xnn_qu8_igemm_minmax_rndnu_ukernel_2x32c4__neondot()
193 int32x4_t vacc0x4567 = vreinterpretq_s32_u32(vsubq_u32(vpacc0x4567, vnacc0x0123)); in xnn_qu8_igemm_minmax_rndnu_ukernel_2x32c4__neondot()
194 int32x4_t vacc0x89AB = vreinterpretq_s32_u32(vsubq_u32(vpacc0x89AB, vnacc0x0123)); in xnn_qu8_igemm_minmax_rndnu_ukernel_2x32c4__neondot()
195 int32x4_t vacc0xCDEF = vreinterpretq_s32_u32(vsubq_u32(vpacc0xCDEF, vnacc0x0123)); in xnn_qu8_igemm_minmax_rndnu_ukernel_2x32c4__neondot()
196 int32x4_t vacc0xGHIJ = vreinterpretq_s32_u32(vsubq_u32(vpacc0xGHIJ, vnacc0x0123)); in xnn_qu8_igemm_minmax_rndnu_ukernel_2x32c4__neondot()
197 int32x4_t vacc0xKLMN = vreinterpretq_s32_u32(vsubq_u32(vpacc0xKLMN, vnacc0x0123)); in xnn_qu8_igemm_minmax_rndnu_ukernel_2x32c4__neondot()
198 int32x4_t vacc0xOPQR = vreinterpretq_s32_u32(vsubq_u32(vpacc0xOPQR, vnacc0x0123)); in xnn_qu8_igemm_minmax_rndnu_ukernel_2x32c4__neondot()
199 int32x4_t vacc0xSTUV = vreinterpretq_s32_u32(vsubq_u32(vpacc0xSTUV, vnacc0x0123)); in xnn_qu8_igemm_minmax_rndnu_ukernel_2x32c4__neondot()
D2x16c4-minmax-fp32-neondot.c148 const uint32x4_t vnacc0x0123 = vcombine_u32(vnacc0, vnacc0); in xnn_qu8_igemm_minmax_fp32_ukernel_2x16c4__neondot() local
149 int32x4_t vacc0x0123 = vreinterpretq_s32_u32(vsubq_u32(vpacc0x0123, vnacc0x0123)); in xnn_qu8_igemm_minmax_fp32_ukernel_2x16c4__neondot()
150 int32x4_t vacc0x4567 = vreinterpretq_s32_u32(vsubq_u32(vpacc0x4567, vnacc0x0123)); in xnn_qu8_igemm_minmax_fp32_ukernel_2x16c4__neondot()
151 int32x4_t vacc0x89AB = vreinterpretq_s32_u32(vsubq_u32(vpacc0x89AB, vnacc0x0123)); in xnn_qu8_igemm_minmax_fp32_ukernel_2x16c4__neondot()
152 int32x4_t vacc0xCDEF = vreinterpretq_s32_u32(vsubq_u32(vpacc0xCDEF, vnacc0x0123)); in xnn_qu8_igemm_minmax_fp32_ukernel_2x16c4__neondot()
D1x8c4-minmax-rndnu-neondot.c104 const uint32x4_t vnacc0x0123 = vcombine_u32(vnacc0, vnacc0); in xnn_qu8_igemm_minmax_rndnu_ukernel_1x8c4__neondot() local
105 int32x4_t vacc0x0123 = vreinterpretq_s32_u32(vsubq_u32(vpacc0x0123, vnacc0x0123)); in xnn_qu8_igemm_minmax_rndnu_ukernel_1x8c4__neondot()
106 int32x4_t vacc0x4567 = vreinterpretq_s32_u32(vsubq_u32(vpacc0x4567, vnacc0x0123)); in xnn_qu8_igemm_minmax_rndnu_ukernel_1x8c4__neondot()
D2x16c4-minmax-rndnu-neondot.c147 const uint32x4_t vnacc0x0123 = vcombine_u32(vnacc0, vnacc0); in xnn_qu8_igemm_minmax_rndnu_ukernel_2x16c4__neondot() local
148 int32x4_t vacc0x0123 = vreinterpretq_s32_u32(vsubq_u32(vpacc0x0123, vnacc0x0123)); in xnn_qu8_igemm_minmax_rndnu_ukernel_2x16c4__neondot()
149 int32x4_t vacc0x4567 = vreinterpretq_s32_u32(vsubq_u32(vpacc0x4567, vnacc0x0123)); in xnn_qu8_igemm_minmax_rndnu_ukernel_2x16c4__neondot()
150 int32x4_t vacc0x89AB = vreinterpretq_s32_u32(vsubq_u32(vpacc0x89AB, vnacc0x0123)); in xnn_qu8_igemm_minmax_rndnu_ukernel_2x16c4__neondot()
151 int32x4_t vacc0xCDEF = vreinterpretq_s32_u32(vsubq_u32(vpacc0xCDEF, vnacc0x0123)); in xnn_qu8_igemm_minmax_rndnu_ukernel_2x16c4__neondot()
D3x32c4-minmax-rndnu-neondot.c236 const uint32x4_t vnacc0x0123 = vcombine_u32(vnacc0, vnacc0); in xnn_qu8_igemm_minmax_rndnu_ukernel_3x32c4__neondot() local
237 int32x4_t vacc0x0123 = vreinterpretq_s32_u32(vsubq_u32(vpacc0x0123, vnacc0x0123)); in xnn_qu8_igemm_minmax_rndnu_ukernel_3x32c4__neondot()
238 int32x4_t vacc0x4567 = vreinterpretq_s32_u32(vsubq_u32(vpacc0x4567, vnacc0x0123)); in xnn_qu8_igemm_minmax_rndnu_ukernel_3x32c4__neondot()
239 int32x4_t vacc0x89AB = vreinterpretq_s32_u32(vsubq_u32(vpacc0x89AB, vnacc0x0123)); in xnn_qu8_igemm_minmax_rndnu_ukernel_3x32c4__neondot()
240 int32x4_t vacc0xCDEF = vreinterpretq_s32_u32(vsubq_u32(vpacc0xCDEF, vnacc0x0123)); in xnn_qu8_igemm_minmax_rndnu_ukernel_3x32c4__neondot()
241 int32x4_t vacc0xGHIJ = vreinterpretq_s32_u32(vsubq_u32(vpacc0xGHIJ, vnacc0x0123)); in xnn_qu8_igemm_minmax_rndnu_ukernel_3x32c4__neondot()
242 int32x4_t vacc0xKLMN = vreinterpretq_s32_u32(vsubq_u32(vpacc0xKLMN, vnacc0x0123)); in xnn_qu8_igemm_minmax_rndnu_ukernel_3x32c4__neondot()
243 int32x4_t vacc0xOPQR = vreinterpretq_s32_u32(vsubq_u32(vpacc0xOPQR, vnacc0x0123)); in xnn_qu8_igemm_minmax_rndnu_ukernel_3x32c4__neondot()
244 int32x4_t vacc0xSTUV = vreinterpretq_s32_u32(vsubq_u32(vpacc0xSTUV, vnacc0x0123)); in xnn_qu8_igemm_minmax_rndnu_ukernel_3x32c4__neondot()
D2x8c4-minmax-rndnu-neondot.c125 const uint32x4_t vnacc0x0123 = vcombine_u32(vnacc0, vnacc0); in xnn_qu8_igemm_minmax_rndnu_ukernel_2x8c4__neondot() local
126 int32x4_t vacc0x0123 = vreinterpretq_s32_u32(vsubq_u32(vpacc0x0123, vnacc0x0123)); in xnn_qu8_igemm_minmax_rndnu_ukernel_2x8c4__neondot()
127 int32x4_t vacc0x4567 = vreinterpretq_s32_u32(vsubq_u32(vpacc0x4567, vnacc0x0123)); in xnn_qu8_igemm_minmax_rndnu_ukernel_2x8c4__neondot()
D3x16c4-minmax-rndnu-neondot.c176 const uint32x4_t vnacc0x0123 = vcombine_u32(vnacc0, vnacc0); in xnn_qu8_igemm_minmax_rndnu_ukernel_3x16c4__neondot() local
177 int32x4_t vacc0x0123 = vreinterpretq_s32_u32(vsubq_u32(vpacc0x0123, vnacc0x0123)); in xnn_qu8_igemm_minmax_rndnu_ukernel_3x16c4__neondot()
178 int32x4_t vacc0x4567 = vreinterpretq_s32_u32(vsubq_u32(vpacc0x4567, vnacc0x0123)); in xnn_qu8_igemm_minmax_rndnu_ukernel_3x16c4__neondot()
179 int32x4_t vacc0x89AB = vreinterpretq_s32_u32(vsubq_u32(vpacc0x89AB, vnacc0x0123)); in xnn_qu8_igemm_minmax_rndnu_ukernel_3x16c4__neondot()
180 int32x4_t vacc0xCDEF = vreinterpretq_s32_u32(vsubq_u32(vpacc0xCDEF, vnacc0x0123)); in xnn_qu8_igemm_minmax_rndnu_ukernel_3x16c4__neondot()
D3x8c4-minmax-rndnu-neondot.c146 const uint32x4_t vnacc0x0123 = vcombine_u32(vnacc0, vnacc0); in xnn_qu8_igemm_minmax_rndnu_ukernel_3x8c4__neondot() local
147 int32x4_t vacc0x0123 = vreinterpretq_s32_u32(vsubq_u32(vpacc0x0123, vnacc0x0123)); in xnn_qu8_igemm_minmax_rndnu_ukernel_3x8c4__neondot()
148 int32x4_t vacc0x4567 = vreinterpretq_s32_u32(vsubq_u32(vpacc0x4567, vnacc0x0123)); in xnn_qu8_igemm_minmax_rndnu_ukernel_3x8c4__neondot()
D4x16c4-minmax-fp32-neondot.c206 const uint32x4_t vnacc0x0123 = vcombine_u32(vnacc0, vnacc0); in xnn_qu8_igemm_minmax_fp32_ukernel_4x16c4__neondot() local
207 int32x4_t vacc0x0123 = vreinterpretq_s32_u32(vsubq_u32(vpacc0x0123, vnacc0x0123)); in xnn_qu8_igemm_minmax_fp32_ukernel_4x16c4__neondot()
208 int32x4_t vacc0x4567 = vreinterpretq_s32_u32(vsubq_u32(vpacc0x4567, vnacc0x0123)); in xnn_qu8_igemm_minmax_fp32_ukernel_4x16c4__neondot()
209 int32x4_t vacc0x89AB = vreinterpretq_s32_u32(vsubq_u32(vpacc0x89AB, vnacc0x0123)); in xnn_qu8_igemm_minmax_fp32_ukernel_4x16c4__neondot()
210 int32x4_t vacc0xCDEF = vreinterpretq_s32_u32(vsubq_u32(vpacc0xCDEF, vnacc0x0123)); in xnn_qu8_igemm_minmax_fp32_ukernel_4x16c4__neondot()
D4x16c4-minmax-rndnu-neondot.c205 const uint32x4_t vnacc0x0123 = vcombine_u32(vnacc0, vnacc0); in xnn_qu8_igemm_minmax_rndnu_ukernel_4x16c4__neondot() local
206 int32x4_t vacc0x0123 = vreinterpretq_s32_u32(vsubq_u32(vpacc0x0123, vnacc0x0123)); in xnn_qu8_igemm_minmax_rndnu_ukernel_4x16c4__neondot()
207 int32x4_t vacc0x4567 = vreinterpretq_s32_u32(vsubq_u32(vpacc0x4567, vnacc0x0123)); in xnn_qu8_igemm_minmax_rndnu_ukernel_4x16c4__neondot()
208 int32x4_t vacc0x89AB = vreinterpretq_s32_u32(vsubq_u32(vpacc0x89AB, vnacc0x0123)); in xnn_qu8_igemm_minmax_rndnu_ukernel_4x16c4__neondot()
209 int32x4_t vacc0xCDEF = vreinterpretq_s32_u32(vsubq_u32(vpacc0xCDEF, vnacc0x0123)); in xnn_qu8_igemm_minmax_rndnu_ukernel_4x16c4__neondot()
/external/XNNPACK/src/qu8-gemm/gen/
D1x32c4-minmax-rndnu-neondot.c134 const uint32x4_t vnacc0x0123 = vcombine_u32(vnacc0, vnacc0); in xnn_qu8_gemm_minmax_rndnu_ukernel_1x32c4__neondot() local
135 int32x4_t vacc0x0123 = vreinterpretq_s32_u32(vsubq_u32(vpacc0x0123, vnacc0x0123)); in xnn_qu8_gemm_minmax_rndnu_ukernel_1x32c4__neondot()
136 int32x4_t vacc0x4567 = vreinterpretq_s32_u32(vsubq_u32(vpacc0x4567, vnacc0x0123)); in xnn_qu8_gemm_minmax_rndnu_ukernel_1x32c4__neondot()
137 int32x4_t vacc0x89AB = vreinterpretq_s32_u32(vsubq_u32(vpacc0x89AB, vnacc0x0123)); in xnn_qu8_gemm_minmax_rndnu_ukernel_1x32c4__neondot()
138 int32x4_t vacc0xCDEF = vreinterpretq_s32_u32(vsubq_u32(vpacc0xCDEF, vnacc0x0123)); in xnn_qu8_gemm_minmax_rndnu_ukernel_1x32c4__neondot()
139 int32x4_t vacc0xGHIJ = vreinterpretq_s32_u32(vsubq_u32(vpacc0xGHIJ, vnacc0x0123)); in xnn_qu8_gemm_minmax_rndnu_ukernel_1x32c4__neondot()
140 int32x4_t vacc0xKLMN = vreinterpretq_s32_u32(vsubq_u32(vpacc0xKLMN, vnacc0x0123)); in xnn_qu8_gemm_minmax_rndnu_ukernel_1x32c4__neondot()
141 int32x4_t vacc0xOPQR = vreinterpretq_s32_u32(vsubq_u32(vpacc0xOPQR, vnacc0x0123)); in xnn_qu8_gemm_minmax_rndnu_ukernel_1x32c4__neondot()
142 int32x4_t vacc0xSTUV = vreinterpretq_s32_u32(vsubq_u32(vpacc0xSTUV, vnacc0x0123)); in xnn_qu8_gemm_minmax_rndnu_ukernel_1x32c4__neondot()
D1x16c4-minmax-rndnu-neondot.c106 const uint32x4_t vnacc0x0123 = vcombine_u32(vnacc0, vnacc0); in xnn_qu8_gemm_minmax_rndnu_ukernel_1x16c4__neondot() local
107 int32x4_t vacc0x0123 = vreinterpretq_s32_u32(vsubq_u32(vpacc0x0123, vnacc0x0123)); in xnn_qu8_gemm_minmax_rndnu_ukernel_1x16c4__neondot()
108 int32x4_t vacc0x4567 = vreinterpretq_s32_u32(vsubq_u32(vpacc0x4567, vnacc0x0123)); in xnn_qu8_gemm_minmax_rndnu_ukernel_1x16c4__neondot()
109 int32x4_t vacc0x89AB = vreinterpretq_s32_u32(vsubq_u32(vpacc0x89AB, vnacc0x0123)); in xnn_qu8_gemm_minmax_rndnu_ukernel_1x16c4__neondot()
110 int32x4_t vacc0xCDEF = vreinterpretq_s32_u32(vsubq_u32(vpacc0xCDEF, vnacc0x0123)); in xnn_qu8_gemm_minmax_rndnu_ukernel_1x16c4__neondot()
D1x16c4-minmax-fp32-neondot.c107 const uint32x4_t vnacc0x0123 = vcombine_u32(vnacc0, vnacc0); in xnn_qu8_gemm_minmax_fp32_ukernel_1x16c4__neondot() local
108 int32x4_t vacc0x0123 = vreinterpretq_s32_u32(vsubq_u32(vpacc0x0123, vnacc0x0123)); in xnn_qu8_gemm_minmax_fp32_ukernel_1x16c4__neondot()
109 int32x4_t vacc0x4567 = vreinterpretq_s32_u32(vsubq_u32(vpacc0x4567, vnacc0x0123)); in xnn_qu8_gemm_minmax_fp32_ukernel_1x16c4__neondot()
110 int32x4_t vacc0x89AB = vreinterpretq_s32_u32(vsubq_u32(vpacc0x89AB, vnacc0x0123)); in xnn_qu8_gemm_minmax_fp32_ukernel_1x16c4__neondot()
111 int32x4_t vacc0xCDEF = vreinterpretq_s32_u32(vsubq_u32(vpacc0xCDEF, vnacc0x0123)); in xnn_qu8_gemm_minmax_fp32_ukernel_1x16c4__neondot()
D2x32c4-minmax-rndnu-neondot.c177 const uint32x4_t vnacc0x0123 = vcombine_u32(vnacc0, vnacc0); in xnn_qu8_gemm_minmax_rndnu_ukernel_2x32c4__neondot() local
178 int32x4_t vacc0x0123 = vreinterpretq_s32_u32(vsubq_u32(vpacc0x0123, vnacc0x0123)); in xnn_qu8_gemm_minmax_rndnu_ukernel_2x32c4__neondot()
179 int32x4_t vacc0x4567 = vreinterpretq_s32_u32(vsubq_u32(vpacc0x4567, vnacc0x0123)); in xnn_qu8_gemm_minmax_rndnu_ukernel_2x32c4__neondot()
180 int32x4_t vacc0x89AB = vreinterpretq_s32_u32(vsubq_u32(vpacc0x89AB, vnacc0x0123)); in xnn_qu8_gemm_minmax_rndnu_ukernel_2x32c4__neondot()
181 int32x4_t vacc0xCDEF = vreinterpretq_s32_u32(vsubq_u32(vpacc0xCDEF, vnacc0x0123)); in xnn_qu8_gemm_minmax_rndnu_ukernel_2x32c4__neondot()
182 int32x4_t vacc0xGHIJ = vreinterpretq_s32_u32(vsubq_u32(vpacc0xGHIJ, vnacc0x0123)); in xnn_qu8_gemm_minmax_rndnu_ukernel_2x32c4__neondot()
183 int32x4_t vacc0xKLMN = vreinterpretq_s32_u32(vsubq_u32(vpacc0xKLMN, vnacc0x0123)); in xnn_qu8_gemm_minmax_rndnu_ukernel_2x32c4__neondot()
184 int32x4_t vacc0xOPQR = vreinterpretq_s32_u32(vsubq_u32(vpacc0xOPQR, vnacc0x0123)); in xnn_qu8_gemm_minmax_rndnu_ukernel_2x32c4__neondot()
185 int32x4_t vacc0xSTUV = vreinterpretq_s32_u32(vsubq_u32(vpacc0xSTUV, vnacc0x0123)); in xnn_qu8_gemm_minmax_rndnu_ukernel_2x32c4__neondot()
D1x8c4-minmax-rndnu-neondot.c92 const uint32x4_t vnacc0x0123 = vcombine_u32(vnacc0, vnacc0); in xnn_qu8_gemm_minmax_rndnu_ukernel_1x8c4__neondot() local
93 int32x4_t vacc0x0123 = vreinterpretq_s32_u32(vsubq_u32(vpacc0x0123, vnacc0x0123)); in xnn_qu8_gemm_minmax_rndnu_ukernel_1x8c4__neondot()
94 int32x4_t vacc0x4567 = vreinterpretq_s32_u32(vsubq_u32(vpacc0x4567, vnacc0x0123)); in xnn_qu8_gemm_minmax_rndnu_ukernel_1x8c4__neondot()
D2x16c4-minmax-fp32-neondot.c134 const uint32x4_t vnacc0x0123 = vcombine_u32(vnacc0, vnacc0); in xnn_qu8_gemm_minmax_fp32_ukernel_2x16c4__neondot() local
135 int32x4_t vacc0x0123 = vreinterpretq_s32_u32(vsubq_u32(vpacc0x0123, vnacc0x0123)); in xnn_qu8_gemm_minmax_fp32_ukernel_2x16c4__neondot()
136 int32x4_t vacc0x4567 = vreinterpretq_s32_u32(vsubq_u32(vpacc0x4567, vnacc0x0123)); in xnn_qu8_gemm_minmax_fp32_ukernel_2x16c4__neondot()
137 int32x4_t vacc0x89AB = vreinterpretq_s32_u32(vsubq_u32(vpacc0x89AB, vnacc0x0123)); in xnn_qu8_gemm_minmax_fp32_ukernel_2x16c4__neondot()
138 int32x4_t vacc0xCDEF = vreinterpretq_s32_u32(vsubq_u32(vpacc0xCDEF, vnacc0x0123)); in xnn_qu8_gemm_minmax_fp32_ukernel_2x16c4__neondot()
D2x16c4-minmax-rndnu-neondot.c133 const uint32x4_t vnacc0x0123 = vcombine_u32(vnacc0, vnacc0); in xnn_qu8_gemm_minmax_rndnu_ukernel_2x16c4__neondot() local
134 int32x4_t vacc0x0123 = vreinterpretq_s32_u32(vsubq_u32(vpacc0x0123, vnacc0x0123)); in xnn_qu8_gemm_minmax_rndnu_ukernel_2x16c4__neondot()
135 int32x4_t vacc0x4567 = vreinterpretq_s32_u32(vsubq_u32(vpacc0x4567, vnacc0x0123)); in xnn_qu8_gemm_minmax_rndnu_ukernel_2x16c4__neondot()
136 int32x4_t vacc0x89AB = vreinterpretq_s32_u32(vsubq_u32(vpacc0x89AB, vnacc0x0123)); in xnn_qu8_gemm_minmax_rndnu_ukernel_2x16c4__neondot()
137 int32x4_t vacc0xCDEF = vreinterpretq_s32_u32(vsubq_u32(vpacc0xCDEF, vnacc0x0123)); in xnn_qu8_gemm_minmax_rndnu_ukernel_2x16c4__neondot()
D3x32c4-minmax-rndnu-neondot.c220 const uint32x4_t vnacc0x0123 = vcombine_u32(vnacc0, vnacc0); in xnn_qu8_gemm_minmax_rndnu_ukernel_3x32c4__neondot() local
221 int32x4_t vacc0x0123 = vreinterpretq_s32_u32(vsubq_u32(vpacc0x0123, vnacc0x0123)); in xnn_qu8_gemm_minmax_rndnu_ukernel_3x32c4__neondot()
222 int32x4_t vacc0x4567 = vreinterpretq_s32_u32(vsubq_u32(vpacc0x4567, vnacc0x0123)); in xnn_qu8_gemm_minmax_rndnu_ukernel_3x32c4__neondot()
223 int32x4_t vacc0x89AB = vreinterpretq_s32_u32(vsubq_u32(vpacc0x89AB, vnacc0x0123)); in xnn_qu8_gemm_minmax_rndnu_ukernel_3x32c4__neondot()
224 int32x4_t vacc0xCDEF = vreinterpretq_s32_u32(vsubq_u32(vpacc0xCDEF, vnacc0x0123)); in xnn_qu8_gemm_minmax_rndnu_ukernel_3x32c4__neondot()
225 int32x4_t vacc0xGHIJ = vreinterpretq_s32_u32(vsubq_u32(vpacc0xGHIJ, vnacc0x0123)); in xnn_qu8_gemm_minmax_rndnu_ukernel_3x32c4__neondot()
226 int32x4_t vacc0xKLMN = vreinterpretq_s32_u32(vsubq_u32(vpacc0xKLMN, vnacc0x0123)); in xnn_qu8_gemm_minmax_rndnu_ukernel_3x32c4__neondot()
227 int32x4_t vacc0xOPQR = vreinterpretq_s32_u32(vsubq_u32(vpacc0xOPQR, vnacc0x0123)); in xnn_qu8_gemm_minmax_rndnu_ukernel_3x32c4__neondot()
228 int32x4_t vacc0xSTUV = vreinterpretq_s32_u32(vsubq_u32(vpacc0xSTUV, vnacc0x0123)); in xnn_qu8_gemm_minmax_rndnu_ukernel_3x32c4__neondot()
D2x8c4-minmax-rndnu-neondot.c111 const uint32x4_t vnacc0x0123 = vcombine_u32(vnacc0, vnacc0); in xnn_qu8_gemm_minmax_rndnu_ukernel_2x8c4__neondot() local
112 int32x4_t vacc0x0123 = vreinterpretq_s32_u32(vsubq_u32(vpacc0x0123, vnacc0x0123)); in xnn_qu8_gemm_minmax_rndnu_ukernel_2x8c4__neondot()
113 int32x4_t vacc0x4567 = vreinterpretq_s32_u32(vsubq_u32(vpacc0x4567, vnacc0x0123)); in xnn_qu8_gemm_minmax_rndnu_ukernel_2x8c4__neondot()
D3x16c4-minmax-rndnu-neondot.c160 const uint32x4_t vnacc0x0123 = vcombine_u32(vnacc0, vnacc0); in xnn_qu8_gemm_minmax_rndnu_ukernel_3x16c4__neondot() local
161 int32x4_t vacc0x0123 = vreinterpretq_s32_u32(vsubq_u32(vpacc0x0123, vnacc0x0123)); in xnn_qu8_gemm_minmax_rndnu_ukernel_3x16c4__neondot()
162 int32x4_t vacc0x4567 = vreinterpretq_s32_u32(vsubq_u32(vpacc0x4567, vnacc0x0123)); in xnn_qu8_gemm_minmax_rndnu_ukernel_3x16c4__neondot()
163 int32x4_t vacc0x89AB = vreinterpretq_s32_u32(vsubq_u32(vpacc0x89AB, vnacc0x0123)); in xnn_qu8_gemm_minmax_rndnu_ukernel_3x16c4__neondot()
164 int32x4_t vacc0xCDEF = vreinterpretq_s32_u32(vsubq_u32(vpacc0xCDEF, vnacc0x0123)); in xnn_qu8_gemm_minmax_rndnu_ukernel_3x16c4__neondot()
D3x8c4-minmax-rndnu-neondot.c130 const uint32x4_t vnacc0x0123 = vcombine_u32(vnacc0, vnacc0); in xnn_qu8_gemm_minmax_rndnu_ukernel_3x8c4__neondot() local
131 int32x4_t vacc0x0123 = vreinterpretq_s32_u32(vsubq_u32(vpacc0x0123, vnacc0x0123)); in xnn_qu8_gemm_minmax_rndnu_ukernel_3x8c4__neondot()
132 int32x4_t vacc0x4567 = vreinterpretq_s32_u32(vsubq_u32(vpacc0x4567, vnacc0x0123)); in xnn_qu8_gemm_minmax_rndnu_ukernel_3x8c4__neondot()
D4x16c4-minmax-fp32-neondot.c188 const uint32x4_t vnacc0x0123 = vcombine_u32(vnacc0, vnacc0); in xnn_qu8_gemm_minmax_fp32_ukernel_4x16c4__neondot() local
189 int32x4_t vacc0x0123 = vreinterpretq_s32_u32(vsubq_u32(vpacc0x0123, vnacc0x0123)); in xnn_qu8_gemm_minmax_fp32_ukernel_4x16c4__neondot()
190 int32x4_t vacc0x4567 = vreinterpretq_s32_u32(vsubq_u32(vpacc0x4567, vnacc0x0123)); in xnn_qu8_gemm_minmax_fp32_ukernel_4x16c4__neondot()
191 int32x4_t vacc0x89AB = vreinterpretq_s32_u32(vsubq_u32(vpacc0x89AB, vnacc0x0123)); in xnn_qu8_gemm_minmax_fp32_ukernel_4x16c4__neondot()
192 int32x4_t vacc0xCDEF = vreinterpretq_s32_u32(vsubq_u32(vpacc0xCDEF, vnacc0x0123)); in xnn_qu8_gemm_minmax_fp32_ukernel_4x16c4__neondot()

12