Upgrade to CUDA 9.0 (#2161)

* Upgrade to CUDA 9.0 * sparse reallocate logic micro fix * Change javacpp-presets back to release version for CUDA
sascha08-15 · Oct 4, 2017 · febf2ea · febf2ea
1 parent e11435e
commit febf2ea
Show file tree

Hide file tree

Showing 8 changed files with 48 additions and 29 deletions.
diff --git a/change-cuda-versions.sh b/change-cuda-versions.sh
@@ -21,9 +21,11 @@
 
 set -e
 
-VALID_VERSIONS=( 7.5 8.0 )
-CUDA_75_VERSION="7\.5"
+VALID_VERSIONS=( 8.0 9.0 )
 CUDA_80_VERSION="8\.0"
+CUDA_90_VERSION="9\.0"
+CUDNN_60_VERSION="6\.0"
+CUDNN_70_VERSION="7\.0"
 
 usage() {
   echo "Usage: $(basename $0) [-h|--help] <cuda version to be used>
@@ -49,16 +51,20 @@ check_cuda_version() {
 
 check_cuda_version "$TO_VERSION"
 
-if [ $TO_VERSION = "8.0" ]; then
-  FROM_BINARY="-7\.5"
-  TO_BINARY="-8\.0"
-  FROM_VERSION=$CUDA_75_VERSION
-  TO_VERSION=$CUDA_80_VERSION
-else
+if [ $TO_VERSION = "9.0" ]; then
   FROM_BINARY="-8\.0"
-  TO_BINARY="-7\.5"
+  TO_BINARY="-9\.0"
   FROM_VERSION=$CUDA_80_VERSION
-  TO_VERSION=$CUDA_75_VERSION
+  TO_VERSION=$CUDA_90_VERSION
+  FROM_VERSION2=$CUDNN_60_VERSION
+  TO_VERSION2=$CUDNN_70_VERSION
+else
+  FROM_BINARY="-9\.0"
+  TO_BINARY="-8\.0"
+  FROM_VERSION=$CUDA_90_VERSION
+  TO_VERSION=$CUDA_80_VERSION
+  FROM_VERSION2=$CUDNN_70_VERSION
+  TO_VERSION2=$CUDNN_60_VERSION
 fi
 
 sed_i() {
@@ -71,16 +77,20 @@ echo "Updating CUDA versions in pom.xml files to CUDA $1";
 
 BASEDIR=$(dirname $0)
 
-#Artifact ids, ending with "-7.5" or "-8.0". nd4j-cuda, deeplearning4j-cuda, etc.
+#Artifact ids, ending with "-8.0" or "-9.0". nd4j-cuda, deeplearning4j-cuda, etc.
 find "$BASEDIR" -name 'pom.xml' -not -path '*target*' \
   -exec bash -c "sed_i 's/\(artifactId>.*\)'$FROM_BINARY'<\/artifactId>/\1'$TO_BINARY'<\/artifactId>/g' {}" \;
 
-#Artifact ids, ending with "-7.5-platform" or "-8.0-platform". nd4j-cuda-platform, etc.
+#Artifact ids, ending with "-8.0-platform" or "-9.0-platform". nd4j-cuda-platform, etc.
 find "$BASEDIR" -name 'pom.xml' -not -path '*target*' \
   -exec bash -c "sed_i 's/\(artifactId>.*\)'$FROM_BINARY'-platform<\/artifactId>/\1'$TO_BINARY'-platform<\/artifactId>/g' {}" \;
 
-#CUDA versions, like <cuda.version>7.5</cuda.version>
+#CUDA versions, like <cuda.version>9.0</cuda.version>
 find "$BASEDIR" -name 'pom.xml' -not -path '*target*' \
   -exec bash -c "sed_i 's/\(cuda.version>\)'$FROM_VERSION'<\/cuda.version>/\1'$TO_VERSION'<\/cuda.version>/g' {}" \;
 
+#cuDNN versions, like <cudnn.version>7.0</cudnn.version>
+find "$BASEDIR" -name 'pom.xml' -not -path '*target*' \
+  -exec bash -c "sed_i 's/\(cudnn.version>\)'$FROM_VERSION2'<\/cudnn.version>/\1'$TO_VERSION2'<\/cudnn.version>/g' {}" \;
+
 echo "Done updating CUDA versions.";
diff --git a/nd4j-backends/nd4j-backend-impls/nd4j-cuda-platform/pom.xml b/nd4j-backends/nd4j-backend-impls/nd4j-cuda-platform/pom.xml
@@ -6,12 +6,12 @@
     </parent>
     <modelVersion>4.0.0</modelVersion>
 
-    <artifactId>nd4j-cuda-8.0-platform</artifactId>
+    <artifactId>nd4j-cuda-9.0-platform</artifactId>
     <name>nd4j-cuda-platform</name>
 
     <properties>
-        <cuda.version>8.0</cuda.version>
-        <cudnn.version>6.0</cudnn.version>
+        <cuda.version>9.0</cuda.version>
+        <cudnn.version>7.0</cudnn.version>
         <nd4j.backend>nd4j-cuda-${cuda.version}</nd4j.backend>
     </properties>
 

diff --git a/nd4j-backends/nd4j-backend-impls/nd4j-cuda/pom.xml b/nd4j-backends/nd4j-backend-impls/nd4j-cuda/pom.xml
@@ -6,12 +6,12 @@
     </parent>
     <modelVersion>4.0.0</modelVersion>
 
-    <artifactId>nd4j-cuda-8.0</artifactId>
+    <artifactId>nd4j-cuda-9.0</artifactId>
     <name>nd4j-cuda</name>
 
     <properties>
-        <cuda.version>8.0</cuda.version>
-        <cudnn.version>6.0</cudnn.version>
+        <cuda.version>9.0</cuda.version>
+        <cudnn.version>7.0</cudnn.version>
         <env.LIBND4J_HOME>${basedir}/../../../../libnd4j/</env.LIBND4J_HOME>
     </properties>
 

diff --git a/...d4j-backend-impls/nd4j-cuda/src/main/java/org/nd4j/linalg/jcublas/blas/JcublasLevel3.java b/...d4j-backend-impls/nd4j-cuda/src/main/java/org/nd4j/linalg/jcublas/blas/JcublasLevel3.java
@@ -62,12 +62,16 @@ protected void hgemm(char Order, char TransA, char TransB, int M, int N, int K,
 
             int arch = CudaEnvironment.getInstance().getCurrentDeviceArchitecture();
 
-            if (CUDA_VERSION == 8000 && (arch == 53 || arch == 60)) {
+            if (CUDA_VERSION >= 8000 && (arch == 53 || arch == 60)) {
                 // on these selected archs we run with cublasHgemm
+                __half alphaHalf = new __half();
+                __half betaHalf = new __half();
+                new ShortPointer(alphaHalf).put((short) HalfIndexer.fromFloat(alpha));
+                new ShortPointer(betaHalf).put((short) HalfIndexer.fromFloat(beta));
+
                 cublasHgemm(new cublasContext(handle), convertTranspose(TransA), convertTranspose(TransB), M, N, K,
-                                new __half().x((short) HalfIndexer.fromFloat(alpha)),
-                                new __half(cAPointer.getDevicePointer()), lda, new __half(cBPointer.getDevicePointer()),
-                                ldb, new __half().x((short) HalfIndexer.fromFloat(beta)),
+                                alphaHalf, new __half(cAPointer.getDevicePointer()), lda,
+                                new __half(cBPointer.getDevicePointer()), ldb, betaHalf,
                                 new __half(cCPointer.getDevicePointer()), ldc);
             } else {
                 // CUDA_R_16F == 2 for CUDA 8

diff --git a/...kend-impls/nd4j-cuda/src/main/java/org/nd4j/linalg/jcublas/buffer/BaseCudaDataBuffer.java b/...kend-impls/nd4j-cuda/src/main/java/org/nd4j/linalg/jcublas/buffer/BaseCudaDataBuffer.java
@@ -1124,7 +1124,7 @@ public DataBuffer reallocate(long length) {
             pointer.address();
 
             // we need to update length with new value now
-            this.length = length;
+            //this.length = length;
         if(isAttached()){
             // do nothing here, that's workspaces
         } else{
@@ -1134,6 +1134,11 @@ public DataBuffer reallocate(long length) {
         return this;
     }
 
+    @Override
+    public long capacity() {
+        return pointer.capacity();
+    }
+
     /*
     protected short fromFloat( float fval ) {
         int fbits = Float.floatToIntBits( fval );

diff --git a/nd4j-backends/nd4j-tests/pom.xml b/nd4j-backends/nd4j-tests/pom.xml
@@ -89,13 +89,13 @@
 
         <!--<dependency>
             <groupId>org.nd4j</groupId>
-            <artifactId>nd4j-cuda-8.0</artifactId>
+            <artifactId>nd4j-cuda-9.0</artifactId>
             <version>${project.version}</version>
         </dependency>-->
 
         <dependency>
             <groupId>org.nd4j</groupId>
-            <artifactId>nd4j-native</artifactId>
+            <artifactId>nd4j-cuda-9.0</artifactId>
             <version>${project.version}</version>
         </dependency>
 

diff --git a/nd4j-buffer/src/main/java/org/nd4j/linalg/api/buffer/BaseDataBuffer.java b/nd4j-buffer/src/main/java/org/nd4j/linalg/api/buffer/BaseDataBuffer.java
@@ -1577,7 +1577,7 @@ public DataBuffer reallocate(long length) {
             }
         }
         Pointer.memcpy(pointer, oldPointer, this.length() * getElementSize());
-        this.underlyingLength = length;
+        //this.underlyingLength = length;
         return this;
     }
 

diff --git a/nd4j-uberjar/pom.xml b/nd4j-uberjar/pom.xml
@@ -264,12 +264,12 @@
     <dependencies>
       <dependency>
         <groupId>org.nd4j</groupId>
-        <artifactId>nd4j-cuda-8.0</artifactId>
+        <artifactId>nd4j-cuda-9.0</artifactId>
         <version>${project.version}</version>
       </dependency>
       <dependency>
         <groupId>org.nd4j</groupId>
-        <artifactId>nd4j-cuda-8.0-platform</artifactId>
+        <artifactId>nd4j-cuda-9.0-platform</artifactId>
         <version>${project.version}</version>
       </dependency>
     </dependencies>