U++ framework
Do not panic. Ask here before giving up.

Home » Developing U++ » U++ Developers corner » SSE2 and SVO optimization (Painter, memcpy....)
Re: BufferPainter::Clear() optimization [message #53989 is a reply to message #53986] Wed, 20 May 2020 01:34 Go to previous messageGo to previous message
Tom1
Messages: 1319
Registered: March 2007
Ultimate Contributor
Hi Mirek,

Yes, I'm nuts... still working at this hour.

Anyway, here's a new version - Fill3T3 - that can actually handle all alignment variations (even those not handled by 7a). Please benchmark and check for correctness:

never_inline void FillStream(dword *b, dword data, int len){
	
	while((uintptr_t)b & 15){ // Try to align
		*b++=data;
		len--;
	};
	__m128i *w = (__m128i *)b;
	__m128i q = _mm_set1_epi32((int)data);
	if(len>=16){
		__m128i *e = w + (len>>2) - 3;
		do{
			_mm_stream_si128(w++, q);
			_mm_stream_si128(w++, q);
			_mm_stream_si128(w++, q);
			_mm_stream_si128(w++, q);
		}while(w<e);
	}
	if(len & 8) {
		_mm_stream_si128(w++, q);
		_mm_stream_si128(w++, q);
	}
	if(len & 4) {
		_mm_stream_si128(w++, q);
	}
	_mm_sfence();
	_mm_storeu_si128((__m128i*)(b + len - 4), q); // Tail align
}

void inline Fill3T3(dword *b, dword data, int len){
	if(len<4){
		if(len&1) *b++ = data;
		if(len&2){ *b++ = data; *b++ = data; }
		return;
	}

	__m128i *w = (__m128i *)b;
	__m128i q = _mm_set1_epi32((int)data);

	if(len >= 32) {
		if(len>1024*1024 && (((uintptr_t)b & 3)==0)){
			FillStream(b,data,len);
			return;
		}
		
		__m128i *e = w + (len>>2) - 7;
		do{
			_mm_storeu_si128(w++, q);
			_mm_storeu_si128(w++, q);
			_mm_storeu_si128(w++, q);
			_mm_storeu_si128(w++, q);
			_mm_storeu_si128(w++, q);
			_mm_storeu_si128(w++, q);
			_mm_storeu_si128(w++, q);
			_mm_storeu_si128(w++, q);
		}while(w<e);
	}
	if(len & 16) {
		_mm_storeu_si128(w++, q);
		_mm_storeu_si128(w++, q);
		_mm_storeu_si128(w++, q);
		_mm_storeu_si128(w++, q);
	}
	if(len & 8) {
		_mm_storeu_si128(w++, q);
		_mm_storeu_si128(w++, q);
	}
	if(len & 4) {
		_mm_storeu_si128(w++, q);
	}
	_mm_storeu_si128((__m128i*)(b + len - 4), q); // Tail align

}


Best regards,

Tom
 
Read Message
Read Message
Read Message
Read Message
Read Message
Read Message
Read Message
Read Message
Read Message
Read Message
Read Message
Read Message
Read Message
Read Message
Read Message
Read Message
Read Message
Read Message
Read Message
Read Message
Read Message
Read Message
Read Message
Read Message
Read Message
Read Message
Read Message
Read Message
Read Message
Read Message
Read Message
Read Message
Read Message
Read Message
Read Message
Read Message
Read Message
Read Message
Read Message
Read Message
Read Message
Read Message
Read Message
Read Message
Read Message
Read Message
Read Message
Read Message
Read Message
Read Message
Read Message
Read Message
Read Message
Read Message
Read Message
Read Message
Read Message
Read Message
Read Message
Read Message
Read Message
Read Message
Read Message
Read Message
Read Message
Read Message
Read Message
Read Message
Read Message
Read Message
Read Message
Read Message
Read Message
Read Message
Read Message
Read Message
Read Message
Read Message
Read Message
Read Message
Read Message
Read Message
Read Message
Read Message
Read Message
Read Message
Read Message
Read Message
Read Message
Read Message
Read Message
Read Message
Read Message
Read Message
Read Message
Read Message
Read Message
Read Message
Read Message
Read Message
Read Message
Read Message
Read Message
Read Message
Read Message
Read Message
Read Message
Read Message
Read Message
Read Message
Read Message
Read Message
Read Message
Read Message
Read Message
Read Message
Read Message
Read Message
Read Message
Read Message
Previous Topic: Should we still care about big-endian CPUs?
Next Topic: TheIDE crash after switching package
Goto Forum:
  


Current Time: Tue Sep 15 23:33:17 GMT+2 2026

Total time taken to generate the page: 0.02154 seconds