search

riscv-software-src / riscv-isa-sim

Spike, a RISC-V ISA Simulator
Other
2.43k stars 855 forks source link

issue while using Half-precision Floats for riscv-v extension #487

Open soniab opened 4 years ago

soniab commented 4 years ago

While trying to use Half-precision Floating point operations for riscv-v extension instruction set architecture with spike. My assembly code is giving me an error "An illegal instruction was executed!", when I am using vsetvli t0, a0, e32,m1,d2 followed by some operation like vfdot.vv or vfadd.vv in my assembly.

As per my understanding, Spike supports Half-Precision Floating point operations for riscv-v extension but I need to program using assembly programming at this stage. "Please correct me If my understanding is wrong".

I will highly appreciate any help or pointer on the above issue.

chihminchao commented 4 years ago

@soniab Do you specify half precision isa extension such as "--isa=rv64gcv_zfh" ?

soniab commented 4 years ago

Yes, I have specified "--isa=rv64gcv_zfh" but still getting the same error. Do I also need to specify isa extension as half precision while installing spike?

chihminchao commented 4 years ago

The other possible cause is that mstatus.vs is not setup correctly.

soniab commented 4 years ago

Is there any specific guidelines that you want me to follow while doing setup for mstatus.vs?

chihminchao commented 4 years ago

Just initialize it as non-zero value to enable vector unit such what we have to do for floating unit

soniab commented 4 years ago

I have verified value of mstatus_vs and it is set as non-zero value already. It is defined as "#define MSTATUS_VS 0x00000600" in riscv/encoding.h.

But still I am getting the same error.

In my assembly, If I have used following vector instructions: vsetvli a5, t0, e32,m1,d2 vfadd.vv v1, v1, v2

After running with spike, I am getting error "An illegal instruction was executed!"

Do I need to set any other value or I am missing something.

I will really appreciate your help in this regard.

chihminchao commented 4 years ago

:vsetvli a5, t0, e32,m1,d2" d2 means "EDIV 2" but current spike doesn't support ediv extension.

soniab commented 4 years ago

Thanks for this update. This is very helpful.

Would you please help me, which instructions I can use for half-precision floating point operations in my assembly for riscv-v extension. Because float16_t, I am not able to use it as data type in my test program. With float (32b) as data type in my test program, I thought only this way to use for half-precision floating point instructions in assembly.

chihminchao commented 4 years ago

You may ask for correct gcc and binutils version in compiler or riscv sw-dev google group. I am not sure about their status.