Align Buffer instances

This works around TooTallNate/ref#28 and may result in a general performance increase.
addaleax · Mar 28, 2015 · e5b7528 · e5b7528
1 parent 7fed565
commit e5b7528
Show file tree

Hide file tree

Showing 2 changed files with 17 additions and 9 deletions.
diff --git a/lib/node-lapack/fortranArray.js b/lib/node-lapack/fortranArray.js
@@ -46,7 +46,7 @@ function jsMatrixToFortranArray(array, elementSize) {
     var op = elementSize == 8 ? 'writeDoubleLE' : 'writeFloatLE';
     var m = array.length;
     var n = array[0].length;
-    var fortranArrayStart = fortranArray = new Buffer(m * n * elementSize);
+    var fortranArrayStart = fortranArray = alignedBuffer(m * n * elementSize);
     for(var j = 0; j < n; j++) {
         for(var i = 0; i < m; i++) {
             fortranArray[op](array[i][j], elementSize * (j * m + i));
@@ -66,6 +66,13 @@ function fortranArrayToJSArray(fortranArray, n, op, elementSize) {
     return array;
 }
 
+// return an buffer suitably aligned for SSE operations
+function alignedBuffer(size) {
+    var buffer = new Buffer(size + 16);
+    return buffer.slice(16 - (buffer.address() % 16));
+}
+
 module.exports.fortranArrayToJSMatrix = fortranArrayToJSMatrix;
 module.exports.jsMatrixToFortranArray = jsMatrixToFortranArray;
 module.exports.fortranArrayToJSArray = fortranArrayToJSArray;
+module.exports.alignedBuffer = alignedBuffer;
diff --git a/lib/node-lapack/lapack.js b/lib/node-lapack/lapack.js
@@ -21,6 +21,7 @@ THE SOFTWARE.
 */
 
 var fortranArray = require('./fortranArray');
+var alignedBuffer = fortranArray.alignedBuffer;
 var FFI = require('ffi');
 
 var LAPACK;
@@ -138,19 +139,19 @@ function geqrf(matrix, op, lapackFunc, elementSize, callback) {
     var qr;
 
     matrixOp(matrix, elementSize, function(m, n, f_m, f_n, f_a, f_lda) {
-	var f_tau = new Buffer(m * n * elementSize);
+	var f_tau = alignedBuffer(m * n * elementSize);
 	var f_info = new Buffer(FORTRAN_INT);
 	var f_lwork = new Buffer(FORTRAN_INT);
 	var f_work;
 	f_lwork.writeInt32LE(-1, 0);
 
 	// get optimal size of workspace
-	f_work = new Buffer(FORTRAN_DOUBLE);
+	f_work = alignedBuffer(FORTRAN_DOUBLE);
 	lapackFunc(f_m, f_n, f_a, f_lda, f_tau, f_work, f_lwork, f_info);
 	lwork = f_work[op](0);
 
 	// allocate workspace
-	f_work = new Buffer(lwork * elementSize);
+	f_work = alignedBuffer(lwork * elementSize);
 	f_lwork.writeInt32LE(lwork, 0);
 
 	// perform QR decomp
@@ -257,18 +258,18 @@ function sgesvd(jobu, jobvt, matrix) {
     var svd;
 
     matrixOp(matrix, FORTRAN_FLOAT, function(m, n, f_m, f_n, f_a, f_lda) {
-	var f_s = new Buffer(Math.pow(Math.min(m, n), 2) * FORTRAN_FLOAT);
-	var f_u = new Buffer(Math.pow(m, 2) * FORTRAN_FLOAT);
+	var f_s = alignedBuffer(Math.pow(Math.min(m, n), 2) * FORTRAN_FLOAT);
+	var f_u = alignedBuffer(Math.pow(m, 2) * FORTRAN_FLOAT);
 	var f_ldu = new Buffer(FORTRAN_INT);
 	f_ldu.writeInt32LE(m, 0);
 
 	// TODO: punting on dims for now. revisit with http://www.netlib.org/lapack/single/sgesvd.f
-	var f_vt = new Buffer(Math.pow(n, 2) * FORTRAN_FLOAT);
+	var f_vt = alignedBuffer(Math.pow(n, 2) * FORTRAN_FLOAT);
 	var f_ldvt = new Buffer(FORTRAN_INT);
 	f_ldvt.writeInt32LE(n, 0);
 
 	var lwork = -1;
-	var f_work = new Buffer(FORTRAN_FLOAT);
+	var f_work = alignedBuffer(FORTRAN_FLOAT);
 	var f_lwork = new Buffer(FORTRAN_INT);
 	f_lwork.writeInt32LE(lwork, 0);
 	var f_info = new Buffer(FORTRAN_INT);
@@ -277,7 +278,7 @@ function sgesvd(jobu, jobvt, matrix) {
 		      f_work, f_lwork, f_info);
 
 	lwork = f_work.readFloatLE(0);
-	f_work = new Buffer(lwork * FORTRAN_FLOAT);
+	f_work = alignedBuffer(lwork * FORTRAN_FLOAT);
 	f_lwork.writeInt32LE(lwork, 0);
 
 	LAPACK.sgesvd_(f_jobu, f_jobvt, f_m, f_n, f_a, f_lda, f_s, f_u, f_ldu, f_vt, f_ldvt,