Home
last modified time | relevance | path

Searched refs:vpsum1x2 (Results 1 – 25 of 36) sorted by relevance

12

/external/XNNPACK/src/qs8-gemm/gen/
D2x8c8-minmax-neon-mull-padal.c148 const int32x2_t vpsum1x2 = vadd_s32(vget_low_s32(vacc1x2), vget_high_s32(vacc1x2)); in xnn_qs8_gemm_minmax_ukernel_2x8c8__neon_mull_padal() local
151 const int32x2_t vsum1x23 = vpadd_s32(vpsum1x2, vpsum1x3); in xnn_qs8_gemm_minmax_ukernel_2x8c8__neon_mull_padal()
D2x8c16-minmax-neon-mlal-padal.c164 const int32x2_t vpsum1x2 = vadd_s32(vget_low_s32(vacc1x2), vget_high_s32(vacc1x2)); in xnn_qs8_gemm_minmax_ukernel_2x8c16__neon_mlal_padal() local
167 const int32x2_t vsum1x23 = vpadd_s32(vpsum1x2, vpsum1x3); in xnn_qs8_gemm_minmax_ukernel_2x8c16__neon_mlal_padal()
D2x8c8-minmax-neon-mlal-padal.c223 const int32x2_t vpsum1x2 = vadd_s32(vget_low_s32(vacc1x2), vget_high_s32(vacc1x2)); in xnn_qs8_gemm_minmax_ukernel_2x8c8__neon_mlal_padal() local
226 const int32x2_t vsum1x23 = vpadd_s32(vpsum1x2, vpsum1x3); in xnn_qs8_gemm_minmax_ukernel_2x8c8__neon_mlal_padal()
D3x8c8-minmax-neon-mull-padal.c185 const int32x2_t vpsum1x2 = vadd_s32(vget_low_s32(vacc1x2), vget_high_s32(vacc1x2)); in xnn_qs8_gemm_minmax_ukernel_3x8c8__neon_mull_padal() local
188 const int32x2_t vsum1x23 = vpadd_s32(vpsum1x2, vpsum1x3); in xnn_qs8_gemm_minmax_ukernel_3x8c8__neon_mull_padal()
D3x8c16-minmax-neon-mlal-padal.c209 const int32x2_t vpsum1x2 = vadd_s32(vget_low_s32(vacc1x2), vget_high_s32(vacc1x2)); in xnn_qs8_gemm_minmax_ukernel_3x8c16__neon_mlal_padal() local
212 const int32x2_t vsum1x23 = vpadd_s32(vpsum1x2, vpsum1x3); in xnn_qs8_gemm_minmax_ukernel_3x8c16__neon_mlal_padal()
D4x8c8-minmax-neon-mull-padal.c222 const int32x2_t vpsum1x2 = vadd_s32(vget_low_s32(vacc1x2), vget_high_s32(vacc1x2)); in xnn_qs8_gemm_minmax_ukernel_4x8c8__neon_mull_padal() local
225 const int32x2_t vsum1x23 = vpadd_s32(vpsum1x2, vpsum1x3); in xnn_qs8_gemm_minmax_ukernel_4x8c8__neon_mull_padal()
D2x16c8-minmax-neon-mull-padal.c230 const int32x2_t vpsum1x2 = vadd_s32(vget_low_s32(vacc1x2), vget_high_s32(vacc1x2)); in xnn_qs8_gemm_minmax_ukernel_2x16c8__neon_mull_padal() local
233 const int32x2_t vsum1x23 = vpadd_s32(vpsum1x2, vpsum1x3); in xnn_qs8_gemm_minmax_ukernel_2x16c8__neon_mull_padal()
D3x8c8-minmax-neon-mlal-padal.c286 const int32x2_t vpsum1x2 = vadd_s32(vget_low_s32(vacc1x2), vget_high_s32(vacc1x2)); in xnn_qs8_gemm_minmax_ukernel_3x8c8__neon_mlal_padal() local
289 const int32x2_t vsum1x23 = vpadd_s32(vpsum1x2, vpsum1x3); in xnn_qs8_gemm_minmax_ukernel_3x8c8__neon_mlal_padal()
D2x16c16-minmax-neon-mlal-padal.c262 const int32x2_t vpsum1x2 = vadd_s32(vget_low_s32(vacc1x2), vget_high_s32(vacc1x2)); in xnn_qs8_gemm_minmax_ukernel_2x16c16__neon_mlal_padal() local
265 const int32x2_t vsum1x23 = vpadd_s32(vpsum1x2, vpsum1x3); in xnn_qs8_gemm_minmax_ukernel_2x16c16__neon_mlal_padal()
D4x8c16-minmax-neon-mlal-padal.c254 const int32x2_t vpsum1x2 = vadd_s32(vget_low_s32(vacc1x2), vget_high_s32(vacc1x2)); in xnn_qs8_gemm_minmax_ukernel_4x8c16__neon_mlal_padal() local
257 const int32x2_t vsum1x23 = vpadd_s32(vpsum1x2, vpsum1x3); in xnn_qs8_gemm_minmax_ukernel_4x8c16__neon_mlal_padal()
D4x8c8-minmax-neon-mlal-padal.c349 const int32x2_t vpsum1x2 = vadd_s32(vget_low_s32(vacc1x2), vget_high_s32(vacc1x2)); in xnn_qs8_gemm_minmax_ukernel_4x8c8__neon_mlal_padal() local
352 const int32x2_t vsum1x23 = vpadd_s32(vpsum1x2, vpsum1x3); in xnn_qs8_gemm_minmax_ukernel_4x8c8__neon_mlal_padal()
D3x16c8-minmax-neon-mull-padal.c297 const int32x2_t vpsum1x2 = vadd_s32(vget_low_s32(vacc1x2), vget_high_s32(vacc1x2)); in xnn_qs8_gemm_minmax_ukernel_3x16c8__neon_mull_padal() local
300 const int32x2_t vsum1x23 = vpadd_s32(vpsum1x2, vpsum1x3); in xnn_qs8_gemm_minmax_ukernel_3x16c8__neon_mull_padal()
D2x16c8-minmax-neon-mlal-padal.c369 const int32x2_t vpsum1x2 = vadd_s32(vget_low_s32(vacc1x2), vget_high_s32(vacc1x2)); in xnn_qs8_gemm_minmax_ukernel_2x16c8__neon_mlal_padal() local
372 const int32x2_t vsum1x23 = vpadd_s32(vpsum1x2, vpsum1x3); in xnn_qs8_gemm_minmax_ukernel_2x16c8__neon_mlal_padal()
/external/XNNPACK/src/qs8-igemm/gen/
D2x8c8-minmax-neon-mull-padal.c164 const int32x2_t vpsum1x2 = vadd_s32(vget_low_s32(vacc1x2), vget_high_s32(vacc1x2)); in xnn_qs8_igemm_minmax_ukernel_2x8c8__neon_mull_padal() local
167 const int32x2_t vsum1x23 = vpadd_s32(vpsum1x2, vpsum1x3); in xnn_qs8_igemm_minmax_ukernel_2x8c8__neon_mull_padal()
D2x8c16-minmax-neon-mlal-padal.c180 const int32x2_t vpsum1x2 = vadd_s32(vget_low_s32(vacc1x2), vget_high_s32(vacc1x2)); in xnn_qs8_igemm_minmax_ukernel_2x8c16__neon_mlal_padal() local
183 const int32x2_t vsum1x23 = vpadd_s32(vpsum1x2, vpsum1x3); in xnn_qs8_igemm_minmax_ukernel_2x8c16__neon_mlal_padal()
D2x8c8-minmax-neon-mlal-padal.c239 const int32x2_t vpsum1x2 = vadd_s32(vget_low_s32(vacc1x2), vget_high_s32(vacc1x2)); in xnn_qs8_igemm_minmax_ukernel_2x8c8__neon_mlal_padal() local
242 const int32x2_t vsum1x23 = vpadd_s32(vpsum1x2, vpsum1x3); in xnn_qs8_igemm_minmax_ukernel_2x8c8__neon_mlal_padal()
D3x8c8-minmax-neon-mull-padal.c203 const int32x2_t vpsum1x2 = vadd_s32(vget_low_s32(vacc1x2), vget_high_s32(vacc1x2)); in xnn_qs8_igemm_minmax_ukernel_3x8c8__neon_mull_padal() local
206 const int32x2_t vsum1x23 = vpadd_s32(vpsum1x2, vpsum1x3); in xnn_qs8_igemm_minmax_ukernel_3x8c8__neon_mull_padal()
D3x8c16-minmax-neon-mlal-padal.c227 const int32x2_t vpsum1x2 = vadd_s32(vget_low_s32(vacc1x2), vget_high_s32(vacc1x2)); in xnn_qs8_igemm_minmax_ukernel_3x8c16__neon_mlal_padal() local
230 const int32x2_t vsum1x23 = vpadd_s32(vpsum1x2, vpsum1x3); in xnn_qs8_igemm_minmax_ukernel_3x8c16__neon_mlal_padal()
D4x8c8-minmax-neon-mull-padal.c242 const int32x2_t vpsum1x2 = vadd_s32(vget_low_s32(vacc1x2), vget_high_s32(vacc1x2)); in xnn_qs8_igemm_minmax_ukernel_4x8c8__neon_mull_padal() local
245 const int32x2_t vsum1x23 = vpadd_s32(vpsum1x2, vpsum1x3); in xnn_qs8_igemm_minmax_ukernel_4x8c8__neon_mull_padal()
D2x16c8-minmax-neon-mull-padal.c246 const int32x2_t vpsum1x2 = vadd_s32(vget_low_s32(vacc1x2), vget_high_s32(vacc1x2)); in xnn_qs8_igemm_minmax_ukernel_2x16c8__neon_mull_padal() local
249 const int32x2_t vsum1x23 = vpadd_s32(vpsum1x2, vpsum1x3); in xnn_qs8_igemm_minmax_ukernel_2x16c8__neon_mull_padal()
D3x8c8-minmax-neon-mlal-padal.c304 const int32x2_t vpsum1x2 = vadd_s32(vget_low_s32(vacc1x2), vget_high_s32(vacc1x2)); in xnn_qs8_igemm_minmax_ukernel_3x8c8__neon_mlal_padal() local
307 const int32x2_t vsum1x23 = vpadd_s32(vpsum1x2, vpsum1x3); in xnn_qs8_igemm_minmax_ukernel_3x8c8__neon_mlal_padal()
D4x8c16-minmax-neon-mlal-padal.c274 const int32x2_t vpsum1x2 = vadd_s32(vget_low_s32(vacc1x2), vget_high_s32(vacc1x2)); in xnn_qs8_igemm_minmax_ukernel_4x8c16__neon_mlal_padal() local
277 const int32x2_t vsum1x23 = vpadd_s32(vpsum1x2, vpsum1x3); in xnn_qs8_igemm_minmax_ukernel_4x8c16__neon_mlal_padal()
D2x16c16-minmax-neon-mlal-padal.c278 const int32x2_t vpsum1x2 = vadd_s32(vget_low_s32(vacc1x2), vget_high_s32(vacc1x2)); in xnn_qs8_igemm_minmax_ukernel_2x16c16__neon_mlal_padal() local
281 const int32x2_t vsum1x23 = vpadd_s32(vpsum1x2, vpsum1x3); in xnn_qs8_igemm_minmax_ukernel_2x16c16__neon_mlal_padal()
D3x16c8-minmax-neon-mull-padal.c315 const int32x2_t vpsum1x2 = vadd_s32(vget_low_s32(vacc1x2), vget_high_s32(vacc1x2)); in xnn_qs8_igemm_minmax_ukernel_3x16c8__neon_mull_padal() local
318 const int32x2_t vsum1x23 = vpadd_s32(vpsum1x2, vpsum1x3); in xnn_qs8_igemm_minmax_ukernel_3x16c8__neon_mull_padal()
D2x16c8-minmax-neon-mlal-padal.c385 const int32x2_t vpsum1x2 = vadd_s32(vget_low_s32(vacc1x2), vget_high_s32(vacc1x2)); in xnn_qs8_igemm_minmax_ukernel_2x16c8__neon_mlal_padal() local
388 const int32x2_t vsum1x23 = vpadd_s32(vpsum1x2, vpsum1x3); in xnn_qs8_igemm_minmax_ukernel_2x16c8__neon_mlal_padal()

12