Lines Matching refs:vpaddq_s32
278 const int32x4_t vsum0x01 = vpaddq_s32(vacc0x0, vacc0x1); in xnn_qs8_gemm_minmax_ukernel_3x16c16__neon_mlal_padal()
279 const int32x4_t vsum0x23 = vpaddq_s32(vacc0x2, vacc0x3); in xnn_qs8_gemm_minmax_ukernel_3x16c16__neon_mlal_padal()
280 const int32x4_t vsum0x45 = vpaddq_s32(vacc0x4, vacc0x5); in xnn_qs8_gemm_minmax_ukernel_3x16c16__neon_mlal_padal()
281 const int32x4_t vsum0x67 = vpaddq_s32(vacc0x6, vacc0x7); in xnn_qs8_gemm_minmax_ukernel_3x16c16__neon_mlal_padal()
282 const int32x4_t vsum0x89 = vpaddq_s32(vacc0x8, vacc0x9); in xnn_qs8_gemm_minmax_ukernel_3x16c16__neon_mlal_padal()
283 const int32x4_t vsum0xAB = vpaddq_s32(vacc0x10, vacc0x11); in xnn_qs8_gemm_minmax_ukernel_3x16c16__neon_mlal_padal()
284 const int32x4_t vsum0xCD = vpaddq_s32(vacc0x12, vacc0x13); in xnn_qs8_gemm_minmax_ukernel_3x16c16__neon_mlal_padal()
285 const int32x4_t vsum0xEF = vpaddq_s32(vacc0x14, vacc0x15); in xnn_qs8_gemm_minmax_ukernel_3x16c16__neon_mlal_padal()
286 const int32x4_t vsum1x01 = vpaddq_s32(vacc1x0, vacc1x1); in xnn_qs8_gemm_minmax_ukernel_3x16c16__neon_mlal_padal()
287 const int32x4_t vsum1x23 = vpaddq_s32(vacc1x2, vacc1x3); in xnn_qs8_gemm_minmax_ukernel_3x16c16__neon_mlal_padal()
288 const int32x4_t vsum1x45 = vpaddq_s32(vacc1x4, vacc1x5); in xnn_qs8_gemm_minmax_ukernel_3x16c16__neon_mlal_padal()
289 const int32x4_t vsum1x67 = vpaddq_s32(vacc1x6, vacc1x7); in xnn_qs8_gemm_minmax_ukernel_3x16c16__neon_mlal_padal()
290 const int32x4_t vsum1x89 = vpaddq_s32(vacc1x8, vacc1x9); in xnn_qs8_gemm_minmax_ukernel_3x16c16__neon_mlal_padal()
291 const int32x4_t vsum1xAB = vpaddq_s32(vacc1x10, vacc1x11); in xnn_qs8_gemm_minmax_ukernel_3x16c16__neon_mlal_padal()
292 const int32x4_t vsum1xCD = vpaddq_s32(vacc1x12, vacc1x13); in xnn_qs8_gemm_minmax_ukernel_3x16c16__neon_mlal_padal()
293 const int32x4_t vsum1xEF = vpaddq_s32(vacc1x14, vacc1x15); in xnn_qs8_gemm_minmax_ukernel_3x16c16__neon_mlal_padal()
294 const int32x4_t vsum2x01 = vpaddq_s32(vacc2x0, vacc2x1); in xnn_qs8_gemm_minmax_ukernel_3x16c16__neon_mlal_padal()
295 const int32x4_t vsum2x23 = vpaddq_s32(vacc2x2, vacc2x3); in xnn_qs8_gemm_minmax_ukernel_3x16c16__neon_mlal_padal()
296 const int32x4_t vsum2x45 = vpaddq_s32(vacc2x4, vacc2x5); in xnn_qs8_gemm_minmax_ukernel_3x16c16__neon_mlal_padal()
297 const int32x4_t vsum2x67 = vpaddq_s32(vacc2x6, vacc2x7); in xnn_qs8_gemm_minmax_ukernel_3x16c16__neon_mlal_padal()
298 const int32x4_t vsum2x89 = vpaddq_s32(vacc2x8, vacc2x9); in xnn_qs8_gemm_minmax_ukernel_3x16c16__neon_mlal_padal()
299 const int32x4_t vsum2xAB = vpaddq_s32(vacc2x10, vacc2x11); in xnn_qs8_gemm_minmax_ukernel_3x16c16__neon_mlal_padal()
300 const int32x4_t vsum2xCD = vpaddq_s32(vacc2x12, vacc2x13); in xnn_qs8_gemm_minmax_ukernel_3x16c16__neon_mlal_padal()
301 const int32x4_t vsum2xEF = vpaddq_s32(vacc2x14, vacc2x15); in xnn_qs8_gemm_minmax_ukernel_3x16c16__neon_mlal_padal()
302 int32x4_t vacc0x0123 = vpaddq_s32(vsum0x01, vsum0x23); in xnn_qs8_gemm_minmax_ukernel_3x16c16__neon_mlal_padal()
303 int32x4_t vacc0x4567 = vpaddq_s32(vsum0x45, vsum0x67); in xnn_qs8_gemm_minmax_ukernel_3x16c16__neon_mlal_padal()
304 int32x4_t vacc0x89AB = vpaddq_s32(vsum0x89, vsum0xAB); in xnn_qs8_gemm_minmax_ukernel_3x16c16__neon_mlal_padal()
305 int32x4_t vacc0xCDEF = vpaddq_s32(vsum0xCD, vsum0xEF); in xnn_qs8_gemm_minmax_ukernel_3x16c16__neon_mlal_padal()
306 int32x4_t vacc1x0123 = vpaddq_s32(vsum1x01, vsum1x23); in xnn_qs8_gemm_minmax_ukernel_3x16c16__neon_mlal_padal()
307 int32x4_t vacc1x4567 = vpaddq_s32(vsum1x45, vsum1x67); in xnn_qs8_gemm_minmax_ukernel_3x16c16__neon_mlal_padal()
308 int32x4_t vacc1x89AB = vpaddq_s32(vsum1x89, vsum1xAB); in xnn_qs8_gemm_minmax_ukernel_3x16c16__neon_mlal_padal()
309 int32x4_t vacc1xCDEF = vpaddq_s32(vsum1xCD, vsum1xEF); in xnn_qs8_gemm_minmax_ukernel_3x16c16__neon_mlal_padal()
310 int32x4_t vacc2x0123 = vpaddq_s32(vsum2x01, vsum2x23); in xnn_qs8_gemm_minmax_ukernel_3x16c16__neon_mlal_padal()
311 int32x4_t vacc2x4567 = vpaddq_s32(vsum2x45, vsum2x67); in xnn_qs8_gemm_minmax_ukernel_3x16c16__neon_mlal_padal()
312 int32x4_t vacc2x89AB = vpaddq_s32(vsum2x89, vsum2xAB); in xnn_qs8_gemm_minmax_ukernel_3x16c16__neon_mlal_padal()
313 int32x4_t vacc2xCDEF = vpaddq_s32(vsum2xCD, vsum2xEF); in xnn_qs8_gemm_minmax_ukernel_3x16c16__neon_mlal_padal()