# Assignment2: RISC-V Toolchain
Contributed by [andy891023](https://github.com/andy891023)

## Rewrite Implementing FP32 Operations by Applying FP32 to Bfloat16 Conversion Algorithm
I rewrote the assembly code Applying FP32 to Bfloat16 Conversion Algorithm from [brain049](https://github.com/brian049), the reason I chose this project is because the algorithm can reduce the size of the memory storing.While the accuracy decreases, what I gain is a reduction in storage space.

## Makefile
###

$ make

You shall see the following messages:

riscv-none-elf-gcc -march=rv32i_zicsr_zifencei -mabi=ilp32 -Ofast -Wall -c -o getcycles.o getcycles.S
riscv-none-elf-gcc -march=rv32i_zicsr_zifencei -mabi=ilp32 -Ofast -Wall -c -o getinstret.o getinstret.S
riscv-none-elf-gcc -march=rv32i_zicsr_zifencei -mabi=ilp32 -Ofast -Wall -c -o hello.o hello.s
riscv-none-elf-gcc -march=rv32i_zicsr_zifencei -mabi=ilp32 -Ofast -Wall -c -o main.o main.c
riscv-none-elf-gcc -o perfcount.elf getcycles.o getinstret.o hello.o main.o

The ==perfcount.elf== will produce after that, and you can run it by the follwing command.

$ ../../build/rv32emu perfcount.elf

Expected output:

Input FP32:3f99999a
Output bfloat16:3fd90000
cycle count: 3443
instret: 2be
inferior exit code 0

## Modified RISC-V code
The modified RISC-V code which can run in [rv32emu](https://github.com/sysprog21/rv32emu) is shown below, input 32-bits number and the output is the 16-bits number.

.data
    #.align 8
    .global start
    .set SYSEXIT, 93
    .set SYSWRITE, 64
    sign_mask: .word 0x80000000
    exp_mask: .word 0x7F800000
    man_mask: .word 0x007FFFFF
    man16_mask: .word 0x007F0000
    r_mask: .word 0xFF800000
    divisor: .word 0x100
    mul_use: .word 0x00800000
    mul_use2: .word 0x01000000
    mul_use3: .word 0x3F800000
   
    str1:     .ascii "zero\n"
              .set str1_size, .-str1
    str2:     .ascii "infinity or NaN\n"
              .set str2_size, .-str2

.text
start:
    add x5, a0, x0 # x5=a0
    addi x30, x0, 1
is_it_zero_or_infinity_or_NaN:
    # Load exp and man into a0 and a1
    lw a0, exp_mask
    lw a1, man_mask

    and x6, x5, a0     # exp
    and x7, x5, a1     # man
    
    # Check for zero
    beqz x6, zero_case
    beqz x7, zero_case

Normalize:
    # Check for infinity or NaN
    li a0, 0x7F800000
    beq x6, a0, infinity_nan_case

    # Normalized number
    add a0, x0, x5
    add x6, a0, x0
    
    lw a0, r_mask
    and x6, x6, a0     # r_mask

    # r /= 0x100
    srli x6, x6, 8
    
    add a0, x0, x5
    add x5, a0, x6     # y = x + r

    # Mask the lower 16 bits of y
    li t6, 0xFFFF0000
    and x5, x5, t6

    sw x5, 0(x8)
    
    addi x8, x8, 8
    
    j done

zero_case:
    sw x5, 0(x8)
    addi x8, x8, 8
    
    li a7, SYSWRITE
    li a0, 1
    la a1, str1
    li a2, str1_size
    ecall
    
    j done

infinity_nan_case:
    
    sw x5, 0(x8)
    addi x8, x8, 8
    
    li a7, SYSWRITE
    li a0, 1
    la a1, str2
    li a2, str2_size
    ecall
    j done

done:
    addi x30, x30, -1
    bnez x30, is_it_zero_or_infinity_or_NaN
  
    add t1, x0, x0    # initialize t1
    addi x8, x8, -8
    mv a0, x5
    ret

## Compare performance
| Level | cycle count | instret | text | data | bss | dec | hex |
| ------ | ----------- | ------- | ----- | ---- | ---- | ----- | ---- |
| -O0 | 3446 | 2be | 51418 | 1932 | 1528 | 54878 | d65e |
| -O1 | 3443 | 2be | 51346 | 1932 | 1528 | 54806 | d616 |
| -O2 | 3443 | 2be | 51346 | 1932 | 1528 | 54806 | d616 |
| -O3 | 3443 | 2be | 51346 | 1932 | 1528 | 54806 | d616 |
| -Ofast | 3443 | 2be | 51346 | 1932 | 1528 | 54806 | d616|

:::warning
You should improve the assembly implementation and then compare instead of changing optimization levels which affect C code.

:notes: jserv
:::