void Fill3(RGBA *t, RGBA c, int len)
{
	__m128i val4 = _mm_set1_epi32(*(int*)&c);
	auto Set4 = [&](int at) { _mm_storeu_si128((__m128i *)(t + at), val4); };
	auto Set4S = [&](int at) { _mm_stream_si128((__m128i *)(t + at), val4); };
	if(len >= 16) {
		if(len > 1024*1024 / 16 && ((uintptr_t)t & 3) == 0) { // for really huge data, bypass the cache
			while((uintptr_t)t & 15) { // align to 16 bytes for SSE
				*t++ = c;
				len--;
			}
			do {
				Set4S(0);
				Set4S(4);
				Set4S(8);
				Set4S(12);
				t += 16;
				len -= 16;
			}
			while(len >= 16);
			_mm_sfence();
		}
		else
			do {
				Set4(0);
				Set4(4);
				Set4(8);
				Set4(12);
				t += 16;
				len -= 16;
			}
			while(len >= 16);
	}
	if(len & 8) {
		Set4(0);
		Set4(4);
		t += 8;
	}
	if(len & 4) {
		Set4(0);
		t += 4;
	}
	if(len & 2) {
		t[0] = t[1] = c;
		t += 2;
	}
	if(len & 1)
		t[0] = c;
}