note problematic latency of dpbusd/dpwssd

This commit is contained in:
dzaima 2025-04-03 00:37:11 +03:00
parent 7ed5ba3b2f
commit 5928bd91bc

View File

@ -140,7 +140,7 @@ def sum_accumulator{DE, unr, VT=[_]SE if int_els{DE,SE} and DE>SE and hasarch{'A
# x86_64-specific accumulators; TODO: AVX-512 could use dpbusd/dpwssd # x86_64-specific accumulators; TODO: AVX-512 could use dpbusd/dpwssd, though with the higher latency that should be behind a high unroll amount
def sum_accumulator{DE, unr, VT=[k]SE if int_els{DE,SE} and width{SE}==8 and hasarch{'X86_64'}} = { def sum_accumulator{DE, unr, VT=[k]SE if int_els{DE,SE} and width{SE}==8 and hasarch{'X86_64'}} = {
def VM = [k/8]u64 def VM = [k/8]u64
def VU = [k]u8 def VU = [k]u8