Pricinou je redukcia ybztocnych operacii s pamatou
Lebo prechod dat cez CPU je uzasne draha operacia v dnesnych CPU (nase mernaia y roku 2010/2011) poskytnute AMD =ta ich pouzila Mantle a HSA a teraz vsteci opakuju resp. rusia nadbytocnu vymenu dat...for(....) {
C[i]=A[i]+B[i];
}trval
stredna honota: 6,271s
smerodajna odchyla : 2,000s
media: 5,015for(....) {
C[i]=A[i]+B[i];
B[i]=A[i];
}Stredna hodnota: 14,530
smerapdajna odchylka :3,662
median: 12,180A v pripade prenosu dat do GPU a z nej sa prenasuaju data RAM-> CPU->GPUakcelaracia v GPU nefunguje ako DMA pre GPU...Program v~OpenCL bežiaci na~GPU\\ bez~prenosov dát
Stredna hodnota: 0,940
Smerodjana odhcylka:0,100
Median: 0,940Program v~OpenCL bežiaci na~GPU\\ s~prenosmi dát CPU -GPU
Stredna hodnota 1268,060
Smerodajna na odchylka 59,470
Median:1276,820Ak sa minimlaizuje prenos dat, (HSA ma za ciel ho znmazat) a V calcu to ide velm podobne s HSA, co de facto je openCL bez prenosov dat
http://www.extremetech.com/computing/174632-amd-kave... ... A preto MS a vsetci ostani minimlaizuju dtaove prenosy cez CPUDirectX 12 shows its muscle in early benchmarks
By Matt Smith — February 6, 2015
The largest improvement occurred alongside AMD’s Radeon 290X, which leaped from an average framerate of 8.3 per second at Extreme quality (with DirectX 11) to an average of 42.9FPS.
http://www.digitaltrends.com/computing/directx-12-show... ...
