From 6b18bca7f5daf641b4171aff1150bb8517ecd920 Mon Sep 17 00:00:00 2001
From: Mohamed Ibrahim <mibrahim@mibrahim.net>
Date: Sat, 29 Aug 2015 08:08:41 -0400
Subject: [PATCH] AddingLambdasToAparapi.md
---
doc/AddingLambdasToAparapi.md | 9 ++++++---
1 file changed, 6 insertions(+), 3 deletions(-)
diff --git a/doc/AddingLambdasToAparapi.md b/doc/AddingLambdasToAparapi.md
index 4d8878ee..07e9ab9f 100644
--- a/doc/AddingLambdasToAparapi.md
+++ b/doc/AddingLambdasToAparapi.md
@@ -29,17 +29,19 @@ So you can run the familiar 'squares' kernel using
int in[] = ..//
int out[] = .../
Device.bestGPU().forEach(in.length, (i)->{
- out[i] = in[i]*in[i];
+ out[i] = in[i] * in[i];
});
+
Instead of
int in[] = ..//
int out[] = .../
Device.bestGPU().forEach(in.length, new IntConsumer(){
public void accept(int i){
- out[i] = in[i]*in[i];
+ out[i] = in[i] * in[i];
}
});
+
To accomodate lambda's we created Device.forEach(int range, IntConsumer ic) which converts the bytecode of the ic parameter to OpenCL at runtime. The captured args (in, out and i - in this case) are passed to the GPU and the kernel executed.
During our early experiments we encountered an interesting issue. The new 'lambdafied' javac uses Java 7 method handles and invoke dynamic instructions to dispatch the lambda code. It does this by injecting a call to a MethodHandle factory into the call site. At runtime, this factory creates a synthetic class (to capture call-site args) and passes this to our Device.forEach().
@@ -53,6 +55,7 @@ This will mean that in future we will change how Aparapi is launched.
Instead of
$ java -Djava.library.path=path/to/aparapi -classpath path/to/aparapi/aparapi.jar:your.jar YourClass
+
We will use
$ java -agentlib=path/to/aparapi/aparapi.dll -classpath path/to/aparapi/aparapi.jar:your.jar YourClass
@@ -100,4 +103,4 @@ or the range can be a lambda itself, here we specify a start and end value for t
Device.bestGPU().forEach(1024, 1, r->{return(r/2);},(pass, r, id)->{lambda});
// or
- Device.bestGPU().forEach(1, 1024, r->{return(r*2);},(pass, r, id)->{lambda});
\ No newline at end of file
+ Device.bestGPU().forEach(1, 1024, r->{return(r*2);},(pass, r, id)->{lambda});
--
GitLab