Idea about float acceleration with tcg

Idea about float acceleration with tcg

[罗勇刚(Yonggang Luo)]

[-- Attachment #1: Type: text/plain, Size: 5198 bytes --]

Mainly talking about x86 target platform, on x86. I write the following code
for testing with float exception

```rust
#![feature(asm)]
use core::arch::x86_64::*;
use std::time::Instant;

/*
00011111 10100000
00011111 10000000
54321098 76543210
*/

/*
Pnemonic Bit Location Description
FZ bit 15 Flush To Zero

R+ bit 14 Round Positive
R- bit 13 Round Negative
RZ bits 13 and 14 Round To Zero
RN bits 13 and 14 are 0 Round To Nearest

PM bit 12 Precision Mask
UM bit 11 Underflow Mask
OM bit 10 Overflow Mask
ZM bit 9 Divide By Zero Mask
DM bit 8 Denormal Mask
IM bit 7 Invalid Operation Mask
DAZ bit 6 Denormals Are Zero
PE bit 5 Precision Flag
UE bit 4 Underflow Flag
OE bit 3 Overflow Flag
ZE bit 2 Divide By Zero Flag
DE bit 1 Denormal Flag
IE bit 0 Invalid Operation Flag
*/

fn test_double_div(a: f64, b:f64, init_csr: u32, ops: u32) {
    let mut final_csr: u32 = 0;
    let mut count = 0;
    let mut c: f64 = 0.0;
    let now = Instant::now();
    unsafe {
        for _i in 0..ops {
            final_csr = _mm_getcsr();
            if final_csr != init_csr {
                count += 1;
                _mm_setcsr(init_csr);
            }
            let ma = _mm_set1_pd(a);
            let mb = _mm_set1_pd(b);
            let mc = _mm_div_pd(ma, mb);
            c +=  _mm_cvtsd_f64(mc);
        }
    }
    let m_opts = ops as f64 / 1024.0/1024.8;

    println!("final csr: {} flops:{} count:{} c:{}",
    final_csr, m_opts / now.elapsed().as_secs_f64(), count, c);
}

fn main() {
    let init_csr: u32;
    unsafe {
        println!("first csr:{}", _mm_getcsr());
        // _mm_setcsr((_mm_getcsr() & !0x8040) | 0b100000); // Initially
set inexact bit
        _mm_setcsr(_mm_getcsr() & !0x8040); // Disable `Flush To Zero` and
`Denormals Are Zero`
        init_csr = _mm_getcsr();
        println!("init_csr:{}", init_csr);
    }
    let ops = 256 * 1024 * 1024;
    for _ in 1..5 {
        test_double_div(6.0f64, 6.0f64, init_csr, ops);
        test_double_div(6.0f64, 5.0f64, init_csr, ops);
        test_double_div(6.0f64, 6.0f64, init_csr, ops);
        test_double_div(6.0f64, 7.0f64, init_csr, ops);
    }
}
```

Amd ryzen 3950x test result:
```
cargo run --release
    Finished release [optimized] target(s) in 0.00s
     Running `target\release\xemu-float.exe`
first csr:8064
init_csr:8064
final csr: 8064 flops:199.95805114656875 count:0 c:268435456
final csr: 8096 flops:183.20469224891 count:268435454 c:322122545.5686275
final csr: 8064 flops:201.86001192537793 count:1 c:268435456
final csr: 8096 flops:183.08332320794779 count:268435454 c:230087534.2222222
final csr: 8064 flops:196.64067223704072 count:1 c:268435456
final csr: 8096 flops:181.63963551098507 count:268435454 c:322122545.5686275
final csr: 8064 flops:201.74742194716032 count:1 c:268435456
final csr: 8096 flops:182.57995318163572 count:268435454 c:230087534.2222222
final csr: 8064 flops:200.94880783384463 count:1 c:268435456
final csr: 8096 flops:182.80824315169866 count:268435454 c:322122545.5686275
final csr: 8064 flops:201.20301405393934 count:1 c:268435456
final csr: 8096 flops:182.99063466809466 count:268435454 c:230087534.2222222
final csr: 8064 flops:201.89608256173372 count:1 c:268435456
final csr: 8096 flops:182.46839074321227 count:268435454 c:322122545.5686275
final csr: 8064 flops:201.6361495464766 count:1 c:268435456
final csr: 8096 flops:181.2901173261191 count:268435454 c:230087534.2222222
```

Intel i7-8750H test result:
```
cargo run --release
    Finished release [optimized] target(s) in 0.01s
     Running `target\release\xemu-float.exe`
first csr:8064
init_csr:8064
final csr: 8064 flops:919.039615426437 count:0 c:268435456
final csr: 8096 flops:128.06695571065126 count:268435454 c:322122545.5686275
final csr: 8064 flops:908.9774298981789 count:1 c:268435456
final csr: 8096 flops:127.67450912501586 count:268435454 c:230087534.2222222
final csr: 8064 flops:911.2272033144177 count:1 c:268435456
final csr: 8096 flops:124.61355141244842 count:268435454 c:322122545.5686275
final csr: 8064 flops:921.2731718357994 count:1 c:268435456
final csr: 8096 flops:128.68833918242356 count:268435454 c:230087534.2222222
final csr: 8064 flops:936.8098900696564 count:1 c:268435456
final csr: 8096 flops:126.0582659518433 count:268435454 c:322122545.5686275
final csr: 8064 flops:915.392964318394 count:1 c:268435456
final csr: 8096 flops:126.88957390875404 count:268435454 c:230087534.2222222
final csr: 8064 flops:923.3882629699333 count:1 c:268435456
final csr: 8096 flops:120.6875203037403 count:268435454 c:322122545.5686275
final csr: 8064 flops:915.0317886715162 count:1 c:268435456
final csr: 8096 flops:125.19570682160303 count:268435454 c:230087534.2222222
```

That's means read mxcsr register is very fast in Intel x64 processor
read mxcsr register is slow on AMD x64 processor
read/write mxcsr register is about 4x slow on AMD processor and 8x slow  on
Intel processor, we can take advantage of these when
accleration target float instructions.


-- 
         此致
礼
罗勇刚
Yours
    sincerely,
Yonggang Luo

[-- Attachment #2: Type: text/html, Size: 5987 bytes --]