Home
last modified time | relevance | path

Searched refs:vpsum0x4 (Results 1 – 25 of 64) sorted by relevance

123

/external/XNNPACK/src/qs8-gemm/gen/
D1x8c8-minmax-rndnu-neon-mull.c103 const int32x2_t vpsum0x4 = vadd_s32(vget_low_s32(vacc0x4), vget_high_s32(vacc0x4)); in xnn_qs8_gemm_minmax_rndnu_ukernel_1x8c8__neon_mull() local
107 const int32x2_t vsum0x45 = vpadd_s32(vpsum0x4, vpsum0x5); in xnn_qs8_gemm_minmax_rndnu_ukernel_1x8c8__neon_mull()
D1x8c16-minmax-rndnu-neon-mlal.c110 const int32x2_t vpsum0x4 = vadd_s32(vget_low_s32(vacc0x4), vget_high_s32(vacc0x4)); in xnn_qs8_gemm_minmax_rndnu_ukernel_1x8c16__neon_mlal() local
114 const int32x2_t vsum0x45 = vpadd_s32(vpsum0x4, vpsum0x5); in xnn_qs8_gemm_minmax_rndnu_ukernel_1x8c16__neon_mlal()
D1x8c8-minmax-rndnu-neon-mlal.c152 const int32x2_t vpsum0x4 = vadd_s32(vget_low_s32(vacc0x4), vget_high_s32(vacc0x4)); in xnn_qs8_gemm_minmax_rndnu_ukernel_1x8c8__neon_mlal() local
156 const int32x2_t vsum0x45 = vpadd_s32(vpsum0x4, vpsum0x5); in xnn_qs8_gemm_minmax_rndnu_ukernel_1x8c8__neon_mlal()
D1x8c8-minmax-fp32-neon-mlal.c152 const int32x2_t vpsum0x4 = vadd_s32(vget_low_s32(vacc0x4), vget_high_s32(vacc0x4)); in xnn_qs8_gemm_minmax_fp32_ukernel_1x8c8__neon_mlal() local
156 const int32x2_t vsum0x45 = vpadd_s32(vpsum0x4, vpsum0x5); in xnn_qs8_gemm_minmax_fp32_ukernel_1x8c8__neon_mlal()
D1x8c8-minmax-fp32-neonv8-mlal.c153 const int32x2_t vpsum0x4 = vadd_s32(vget_low_s32(vacc0x4), vget_high_s32(vacc0x4)); in xnn_qs8_gemm_minmax_fp32_ukernel_1x8c8__neonv8_mlal() local
157 const int32x2_t vsum0x45 = vpadd_s32(vpsum0x4, vpsum0x5); in xnn_qs8_gemm_minmax_fp32_ukernel_1x8c8__neonv8_mlal()
D2x8c8-minmax-rndnu-neon-mull.c140 const int32x2_t vpsum0x4 = vadd_s32(vget_low_s32(vacc0x4), vget_high_s32(vacc0x4)); in xnn_qs8_gemm_minmax_rndnu_ukernel_2x8c8__neon_mull() local
144 const int32x2_t vsum0x45 = vpadd_s32(vpsum0x4, vpsum0x5); in xnn_qs8_gemm_minmax_rndnu_ukernel_2x8c8__neon_mull()
D1x16c8-minmax-rndnu-neon-mull.c141 const int32x2_t vpsum0x4 = vadd_s32(vget_low_s32(vacc0x4), vget_high_s32(vacc0x4)); in xnn_qs8_gemm_minmax_rndnu_ukernel_1x16c8__neon_mull() local
145 const int32x2_t vsum0x45 = vpadd_s32(vpsum0x4, vpsum0x5); in xnn_qs8_gemm_minmax_rndnu_ukernel_1x16c8__neon_mull()
D2x8c16-minmax-rndnu-neon-mlal.c155 const int32x2_t vpsum0x4 = vadd_s32(vget_low_s32(vacc0x4), vget_high_s32(vacc0x4)); in xnn_qs8_gemm_minmax_rndnu_ukernel_2x8c16__neon_mlal() local
159 const int32x2_t vsum0x45 = vpadd_s32(vpsum0x4, vpsum0x5); in xnn_qs8_gemm_minmax_rndnu_ukernel_2x8c16__neon_mlal()
D3x8c8-minmax-rndnu-neon-mull.c177 const int32x2_t vpsum0x4 = vadd_s32(vget_low_s32(vacc0x4), vget_high_s32(vacc0x4)); in xnn_qs8_gemm_minmax_rndnu_ukernel_3x8c8__neon_mull() local
181 const int32x2_t vsum0x45 = vpadd_s32(vpsum0x4, vpsum0x5); in xnn_qs8_gemm_minmax_rndnu_ukernel_3x8c8__neon_mull()
D1x16c16-minmax-rndnu-neon-mlal.c156 const int32x2_t vpsum0x4 = vadd_s32(vget_low_s32(vacc0x4), vget_high_s32(vacc0x4)); in xnn_qs8_gemm_minmax_rndnu_ukernel_1x16c16__neon_mlal() local
160 const int32x2_t vsum0x45 = vpadd_s32(vpsum0x4, vpsum0x5); in xnn_qs8_gemm_minmax_rndnu_ukernel_1x16c16__neon_mlal()
D2x8c8-minmax-fp32-neonv8-mlal.c216 const int32x2_t vpsum0x4 = vadd_s32(vget_low_s32(vacc0x4), vget_high_s32(vacc0x4)); in xnn_qs8_gemm_minmax_fp32_ukernel_2x8c8__neonv8_mlal() local
220 const int32x2_t vsum0x45 = vpadd_s32(vpsum0x4, vpsum0x5); in xnn_qs8_gemm_minmax_fp32_ukernel_2x8c8__neonv8_mlal()
/external/XNNPACK/src/qs8-igemm/gen/
D1x8c8-minmax-rndnu-neon-mull.c117 const int32x2_t vpsum0x4 = vadd_s32(vget_low_s32(vacc0x4), vget_high_s32(vacc0x4)); in xnn_qs8_igemm_minmax_rndnu_ukernel_1x8c8__neon_mull() local
121 const int32x2_t vsum0x45 = vpadd_s32(vpsum0x4, vpsum0x5); in xnn_qs8_igemm_minmax_rndnu_ukernel_1x8c8__neon_mull()
D1x8c16-minmax-rndnu-neon-mlal.c124 const int32x2_t vpsum0x4 = vadd_s32(vget_low_s32(vacc0x4), vget_high_s32(vacc0x4)); in xnn_qs8_igemm_minmax_rndnu_ukernel_1x8c16__neon_mlal() local
128 const int32x2_t vsum0x45 = vpadd_s32(vpsum0x4, vpsum0x5); in xnn_qs8_igemm_minmax_rndnu_ukernel_1x8c16__neon_mlal()
D1x8c8-minmax-fp32-neonv8-mlal.c167 const int32x2_t vpsum0x4 = vadd_s32(vget_low_s32(vacc0x4), vget_high_s32(vacc0x4)); in xnn_qs8_igemm_minmax_fp32_ukernel_1x8c8__neonv8_mlal() local
171 const int32x2_t vsum0x45 = vpadd_s32(vpsum0x4, vpsum0x5); in xnn_qs8_igemm_minmax_fp32_ukernel_1x8c8__neonv8_mlal()
D1x8c8-minmax-rndnu-neon-mlal.c166 const int32x2_t vpsum0x4 = vadd_s32(vget_low_s32(vacc0x4), vget_high_s32(vacc0x4)); in xnn_qs8_igemm_minmax_rndnu_ukernel_1x8c8__neon_mlal() local
170 const int32x2_t vsum0x45 = vpadd_s32(vpsum0x4, vpsum0x5); in xnn_qs8_igemm_minmax_rndnu_ukernel_1x8c8__neon_mlal()
D1x8c8-minmax-fp32-neon-mlal.c166 const int32x2_t vpsum0x4 = vadd_s32(vget_low_s32(vacc0x4), vget_high_s32(vacc0x4)); in xnn_qs8_igemm_minmax_fp32_ukernel_1x8c8__neon_mlal() local
170 const int32x2_t vsum0x45 = vpadd_s32(vpsum0x4, vpsum0x5); in xnn_qs8_igemm_minmax_fp32_ukernel_1x8c8__neon_mlal()
D2x8c8-minmax-rndnu-neon-mull.c156 const int32x2_t vpsum0x4 = vadd_s32(vget_low_s32(vacc0x4), vget_high_s32(vacc0x4)); in xnn_qs8_igemm_minmax_rndnu_ukernel_2x8c8__neon_mull() local
160 const int32x2_t vsum0x45 = vpadd_s32(vpsum0x4, vpsum0x5); in xnn_qs8_igemm_minmax_rndnu_ukernel_2x8c8__neon_mull()
D1x16c8-minmax-rndnu-neon-mull.c155 const int32x2_t vpsum0x4 = vadd_s32(vget_low_s32(vacc0x4), vget_high_s32(vacc0x4)); in xnn_qs8_igemm_minmax_rndnu_ukernel_1x16c8__neon_mull() local
159 const int32x2_t vsum0x45 = vpadd_s32(vpsum0x4, vpsum0x5); in xnn_qs8_igemm_minmax_rndnu_ukernel_1x16c8__neon_mull()
D2x8c16-minmax-rndnu-neon-mlal.c171 const int32x2_t vpsum0x4 = vadd_s32(vget_low_s32(vacc0x4), vget_high_s32(vacc0x4)); in xnn_qs8_igemm_minmax_rndnu_ukernel_2x8c16__neon_mlal() local
175 const int32x2_t vsum0x45 = vpadd_s32(vpsum0x4, vpsum0x5); in xnn_qs8_igemm_minmax_rndnu_ukernel_2x8c16__neon_mlal()
D1x16c16-minmax-rndnu-neon-mlal.c170 const int32x2_t vpsum0x4 = vadd_s32(vget_low_s32(vacc0x4), vget_high_s32(vacc0x4)); in xnn_qs8_igemm_minmax_rndnu_ukernel_1x16c16__neon_mlal() local
174 const int32x2_t vsum0x45 = vpadd_s32(vpsum0x4, vpsum0x5); in xnn_qs8_igemm_minmax_rndnu_ukernel_1x16c16__neon_mlal()
D2x8c8-minmax-rndnu-neon-mlal.c231 const int32x2_t vpsum0x4 = vadd_s32(vget_low_s32(vacc0x4), vget_high_s32(vacc0x4)); in xnn_qs8_igemm_minmax_rndnu_ukernel_2x8c8__neon_mlal() local
235 const int32x2_t vsum0x45 = vpadd_s32(vpsum0x4, vpsum0x5); in xnn_qs8_igemm_minmax_rndnu_ukernel_2x8c8__neon_mlal()
/external/XNNPACK/src/qc8-gemm/gen/
D1x8c8-minmax-fp32-neon-mlal.c152 const int32x2_t vpsum0x4 = vadd_s32(vget_low_s32(vacc0x4), vget_high_s32(vacc0x4)); in xnn_qc8_gemm_minmax_fp32_ukernel_1x8c8__neon_mlal() local
156 const int32x2_t vsum0x45 = vpadd_s32(vpsum0x4, vpsum0x5); in xnn_qc8_gemm_minmax_fp32_ukernel_1x8c8__neon_mlal()
D1x8c8-minmax-fp32-neonv8-mlal.c153 const int32x2_t vpsum0x4 = vadd_s32(vget_low_s32(vacc0x4), vget_high_s32(vacc0x4)); in xnn_qc8_gemm_minmax_fp32_ukernel_1x8c8__neonv8_mlal() local
157 const int32x2_t vsum0x45 = vpadd_s32(vpsum0x4, vpsum0x5); in xnn_qc8_gemm_minmax_fp32_ukernel_1x8c8__neonv8_mlal()
/external/XNNPACK/src/qc8-igemm/gen/
D1x8c8-minmax-fp32-neon-mlal.c166 const int32x2_t vpsum0x4 = vadd_s32(vget_low_s32(vacc0x4), vget_high_s32(vacc0x4)); in xnn_qc8_igemm_minmax_fp32_ukernel_1x8c8__neon_mlal() local
170 const int32x2_t vsum0x45 = vpadd_s32(vpsum0x4, vpsum0x5); in xnn_qc8_igemm_minmax_fp32_ukernel_1x8c8__neon_mlal()
D1x8c8-minmax-fp32-neonv8-mlal.c167 const int32x2_t vpsum0x4 = vadd_s32(vget_low_s32(vacc0x4), vget_high_s32(vacc0x4)); in xnn_qc8_igemm_minmax_fp32_ukernel_1x8c8__neonv8_mlal() local
171 const int32x2_t vsum0x45 = vpadd_s32(vpsum0x4, vpsum0x5); in xnn_qc8_igemm_minmax_fp32_ukernel_1x8c8__neonv8_mlal()

123