public void releaseMatrixOutputForGPUInstruction(String varName) throws DMLRuntimeException {
MatrixObject mo = getMatrixObject(varName);
if (mo.getGPUObject() == null || !mo.getGPUObject().isAllocated()) {
throw new DMLRuntimeException("No output is allocated on GPU");
}
mo.getGPUObject().releaseOutput();
}
/**
* Allocates the {@link GPUObject} for a given LOPS Variable (eg. _mVar3)
*
* @param varName variable name
* @return matrix object
* @throws DMLRuntimeException if DMLRuntimeException occurs
*/
public MatrixObject allocateGPUMatrixObject(String varName) throws DMLRuntimeException {
MatrixObject mo = getMatrixObject(varName);
if (mo.getGPUObject() == null) {
mo.setGPUObject(GPUContext.createGPUObject(mo));
}
return mo;
}
public MatrixObject getDenseMatrixOutputForGPUInstruction(String varName)
throws DMLRuntimeException {
MatrixObject mo = allocateGPUMatrixObject(varName);
mo.getGPUObject().acquireDeviceModifyDense();
mo.getMatrixCharacteristics().setNonZeros(-1);
return mo;
}
public static void matmult(
MatrixObject left1,
MatrixObject right1,
MatrixObject output,
boolean isLeftTransposed1,
boolean isRightTransposed1)
throws DMLRuntimeException {
if (isInSparseFormat(left1) || isInSparseFormat(right1)) {
throw new DMLRuntimeException("Sparse GPU matrix multiplication is not implemented");
}
// Since CuBLAS expects inputs in column-major format,
// reverse the order of matrix-multiplication and take care of dimension mismatch.
MatrixObject left = right1;
MatrixObject right = left1;
boolean isLeftTransposed = isRightTransposed1;
boolean isRightTransposed = isLeftTransposed1;
char transa = isLeftTransposed ? 'T' : 'N';
char transb = isRightTransposed ? 'T' : 'N';
// Note: the dimensions are swapped
int m = (int) (isLeftTransposed ? left.getNumRows() : left.getNumColumns());
int n = (int) (isRightTransposed ? right.getNumColumns() : right.getNumRows());
int k = (int) (isLeftTransposed ? left.getNumColumns() : left.getNumRows());
int k1 = (int) (isRightTransposed ? right.getNumRows() : right.getNumColumns());
if (k != k1) throw new DMLRuntimeException("Dimension mismatch: " + k + " != " + k1);
if (m == -1 || n == -1 || k == -1) throw new DMLRuntimeException("Incorrect dimensions");
double alpha = 1;
double beta = 0;
int lda = isLeftTransposed ? k : m;
int ldb = isRightTransposed ? n : k;
int ldc = m;
if (!left.getGPUObject().isAllocated() || !right.getGPUObject().isAllocated())
throw new DMLRuntimeException(
"One of input is not allocated:"
+ left.getGPUObject().isAllocated()
+ " "
+ right.getGPUObject().isAllocated());
if (!output.getGPUObject().isAllocated())
throw new DMLRuntimeException(
"Output is not allocated:" + output.getGPUObject().isAllocated());
Pointer A = ((JCudaObject) left.getGPUObject()).jcudaPointer;
Pointer B = ((JCudaObject) right.getGPUObject()).jcudaPointer;
Pointer C = ((JCudaObject) output.getGPUObject()).jcudaPointer;
JCublas.cublasDgemm(transa, transb, m, n, k, alpha, A, lda, B, ldb, beta, C, ldc);
}
public MatrixObject getMatrixInputForGPUInstruction(String varName) throws DMLRuntimeException {
MatrixObject mo = getMatrixObject(varName);
if (mo == null) {
throw new DMLRuntimeException("No matrix object available for variable:" + varName);
}
if (mo.getGPUObject() == null) {
mo.setGPUObject(GPUContext.createGPUObject(mo));
}
boolean acquired = false;
if (!mo.getGPUObject().isAllocated()) {
mo.acquireRead();
acquired = true;
}
mo.getGPUObject().acquireDeviceRead();
if (acquired) {
mo.release();
}
return mo;
}
public static boolean isInSparseFormat(MatrixObject mo) {
if (mo.getGPUObject() != null && mo.getGPUObject().isAllocated())
return mo.getGPUObject().isInSparseFormat();
return MatrixBlock.evalSparseFormatInMemory(mo.getNumRows(), mo.getNumColumns(), mo.getNnz());
}
public void releaseMatrixInputForGPUInstruction(String varName) throws DMLRuntimeException {
MatrixObject mo = getMatrixObject(varName);
mo.getGPUObject().releaseInput();
}