Variable in OpenCL kernel 'for-loop' reduces performance -


There is a loop for me in my colonel, my code was rigid-coded to iterate me for a certain number of : 

 for  (int kk = 0; kk & lt; 50000; kk ++) {& lt; ... my code here ... & gt; }   
 I do not think the code in the loop is relevant to my question, it is a very simple table showing and is integer math.  
 I wanted to make my colonel code a bit more flexible so that I could modify the loop so that the number of iterations of my loop (50000) be replaced by the kernel input parameter 'num_loops'.  
 for  (int kk = 0; kk & lt; num_loops; kk ++) {& lt; ... and code here ... & gt; }   
 The thing I found is that whenever my host program Kernel is  
  num_loops = 50000   
 which There is already a value similar to the hard coded value, the performance of my kernel is cut in about half.  
 I am trying to figure out what is going on in the performance, I think there is something to do with the OpenCL compiler which is not able to unlock the loop efficiently?  
 Am I trying to do without imposing a penalty?  
  Update:  There are some results by playing with "#pragma unroll"  
 Here, it seems that unopened loops to my display problems Does not solve.  
 Unorrolling even the hard-working Loop Degrades display  
 Here is a general loop with hard-coded value (best performance):  
 For  (int kk = 0; kk & letocho; 50000; kk ++) // Execution time = 0.18 (40180 MI ops / sec)   
 If I unlock the loop, then things get worse:  
  Unblock 50000 for #pragma unroll // or #pragma (int kk = 0; kk & lt; 50000; kk ++) Unlock / time to take time = 0.22 (33000 MI ops / sec)   
 There is a loop which uses the variable, num_loops = 50000:  
 for  (int kk = 0; kk & lt; num_loops; kk ++) // execution time = 0.26 (27760 MI ops / sec) #pragma u for nomination 50000 (int Time = 0.26 (27760 MI ops / second) for #pragma (int kk = 0; kk & lt; num_loops; kk ++) to execute kike = 0; kk & lt; num_loops; kk ++) Time to unlock // Time to execute = 0.24 (30280 MI ops / sec)   
 Straighten with "#Proma Anol" becomes slightly better when using the num_loops variable , Although this performance is still difficult-to Injury, is 25% slower than Anolld version.  
 Any other ideas on how to use num_loops as a loop variable without the hit of the display?  
 
 Yes, most of the reasons for the decline in performance is that the compiler can not unlock the loop. There are some things that you can try to improve the situation.  
 
 You can define parameters as a preprocessor macro passed through your program build option. This is a common move that is used to create values ​​which are known only as run-time skeletal time constants. For example:  
  Clubeild program (Program, 1, and Device, "-Nan_Lops = 50000", zero, zero);   
 You can dynamically create build options using the  sprintf  to make it more convenient. It is obviously worth it only if you often get the parameter There is no need to change, so that the overhead of the reprint is not a problem.  
 
 You can check that your OpenCL platform can be used by any programmers who can provide compiler signs about loop-anololing. For example, some OpenCL compilers recognize  #pragma unroll  (or similar). OpenCL 2.0 has a feature for:  __feature__ ((opencl_unroll_hint)) .  
 
 You can manually unlock the loop depending on how it looks, depending on what you can guess about the  num_loops  parameter. For example, if you know (or can be sure) that it will always be a valuable value of 4, then you can do something like:  
 for  (int kk = 0 ; Kk & lt; num_loops;) {& lt; ... and more code here ... & gt; KK ++; & Lt; ... and more code here ... & gt; KK ++; & Lt; ... and more code here ... & gt; KK ++; & Lt; ... and more code here ... & gt; KK ++; }   
 Even if you can not create such assumptions, you should still be able to unlock the manual, but it may require some extra work (For example, to eliminate the remaining iterations).

Comments

Post a Comment

Popular posts from this blog

Verilog Error: output or inout port "Q" must be connected to a structural net expression -

I get the error every time, I try to compile. I'm not sure why anyone can help is? I'm new to verilog. Module D_FF (clerk, d, reset_n, q); Input D, Clack, reset_en; Output Q; reg Q; lab4_gdl f1 (.lk (~ clk), d (d), .q.m (qm)); lab4_GDL f2 (.Clk (Clk), D (QM), .Q (Q)); Always start at @ (posedge clk, neggeous reset_n) (Reset_n == 0) Q & lt; = 0; Other questions & lt; = D; Edit End: The problem is what we are asking to do: In this section, you apply the memory / registration circuit on the AlterEdie 2 board. will do. The circuit has the following specifications: The present value of swift SW15-0 on the D2 board should be shown in four hexadecimal four sections of HEX3-0. This part of the circuit will be combination logic. To use an active-less asynchronous reset and KEY1 to use KEY0 as the clock input, you must store SW15-0in in a 16-bit register in the value The register should be enabled to have a 16-bit positive edge, which uses the embedded D flip-flo
Read more

jasper reports - How to center align barcode using jasperreports and barcode4j -

I use iReport 5.5.0 . How to align a barcode component to the center The relevant piece of my jrxml is below & Lt; JR: Code 128 xmlns: JR = "http://jasperreports.sourceforge.net/jasperreports/components" xsi: schemaLocation = "http://jasperreports.sourceforge.net/jasperreports/components http: //jasperreports.sourceforge Purge / xsd / components.xsd "textPosition =" bottom "& gt; & Lt; Jr: codeExpression & gt; & Lt; [CDATA [$ F {number}]] & gt; & Lt; / Jr: codeExpression & gt; & Lt; / Jr: Code 128 & gt; & Lt; / componentElement & gt; "Text" itemprop = "text"> Create a style after this, it is called "barcode", like: & lt; Style name = "barcode" fontName = "Helvetica" fONTSIZE = "10" hAlign = "center" valign = "center" /> and set the property of your barcode element to "barcode".
Read more

c# - ASP.NET MVC - Attaching an entity of type 'MODELNAME' failed because another entity of the same type already has the same primary key value -

In brief, the wrapper module is inserted during posting and the status of an entry is changed to 'modified' . Before changing the state, the state is set to 'separate', but calling attachments () causes the same error to be thrown away. I am using EF 6. Model / wrapper class public class avi model {public A is a {get; Set; } Public listing & lt; B & gt; B {received; Set; } Public cc {receipt; Set; }} Administrator edit public functioning (int? id) {if (id == null) {new HttpStatusCodeResult (HttpStatusCode.BadRequest); } If (! Conveyor Access (id.Value)) new HTTPTitus code result (HTTP status code. Forbidden); Var aViewModel = new AViewModel (); aViewModel.A = db.As.Find (id); If (aViewModel.Receipt == zero) {return HttpNotFound (); } aViewModel.b = db.Bs.Where (x => x.aid == id.Value) .Oolist (); AViewModel.Vendor = db.Cs.Where (x => x.cid == aViewModel.a.cID) .FirstOrDefault (); See Return (aViewModel); } [HTPFost] [Valid AntitiferousTeacon
Read more