Repository navigation
mulhsu instruction actually returns the lower bits of the multiplication result, not the higher bits. #101
Description
Activity
The following patch may help.
diff --git a/ventus/src/pipeline/Multiplier.scala b/ventus/src/pipeline/Multiplier.scala index 8a519d4..1a18cc1 100644 --- a/ventus/src/pipeline/Multiplier.scala +++ b/ventus/src/pipeline/Multiplier.scala @@ -202,7 +202,7 @@ class ArrayMultiplier(num_thread: Int, len: Int) val xlen = len - 1 val res = Mux(isMAC(func),mac_out(xlen,0), - Mux(func===FN_MULH | func===FN_MULHU, result(2*len-1, len),result(xlen,0)))//Mux(func===FN_MULH | func===FN_MULHU, result(2*xlen-1, xlen),result(xlen,0))) + Mux(func===FN_MULH | func===FN_MULHU | func===FN_MULHSU, result(2*len-1, len),result(xlen,0)))//Mux(func===FN_MULH | func===FN_MULHU, result(2*xlen-1, xlen),result(xlen,0))) io.out.bits.result:=PipelineReg(latency)(res) io.out.bits.ctrl:=ctrlVec.last io.out.bits.mask:=maskVec.lastThanks agin for your help!
The mulhsu instruction actually returns the lower bits of the multiplication result, not the higher bits.
For the operation:
Input:
x1 = 0xabababab, x3 = 0xa36566c8
Output:
x20 = 0x1ffb3f98
Expected result (from Spike reference):
x20 = 0xca2ce72d
Mathematically correct result: (signed)0xabababab * (unsigned)0xa36566c8 = 0xca2ce72d_1ffb3f98
Please refer to our PoC for reproduction: https://github.com/gpu-experimental/poc-ventus-b30
It appears that the GVM has crashed.
Thank you for your hard work!