fix: return β*C when K is zero

lib/node_modules/@stdlib/blas/base/dgemm/lib/base.js

@@ -29,7 +29,7 @@
 
 // VARIABLES //
 
 var bsize = blockSize( 'float64', 'float64' ); // TODO: consider using a larger block size
 
 
 // FUNCTIONS //
@@ -53,7 +53,7 @@
 * var bool = isTransposed( 'no-transpose' );
 * // returns false
 */
 function isTransposed( str ) { // TODO: consider moving to a separate helper utility package
 	return ( str !== 'no-transpose' );
 }
 
@@ -85,7 +85,7 @@
 * zeros( 2, 3, X, 1, 2, 0 );
 * // X => <Float64Array>[ 0.0, 0.0, 0.0, 0.0, 0.0, 0.0 ]
 */
 function zeros( M, N, X, strideX1, strideX2, offsetX ) { // TODO: consider moving to a separate package
 	var dx0;
 	var dx1;
 	var S0;
@@ -147,7 +147,7 @@
 * scal( 2, 3, 5.0, X, 1, 2, 0 );
 * // X => <Float64Array>[ 5.0, 10.0, 15.0, 20.0, 25.0, 30.0 ]
 */
 function scal( M, N, beta, X, strideX1, strideX2, offsetX ) { // TODO: consider moving to a separate package
 	var dx0;
 	var dx1;
 	var S0;
@@ -418,7 +418,7 @@
 		C = scal( M, N, beta, C, strideC1, strideC2, offsetC );
 	}
 	// Check whether we can early return...
-	if ( alpha === 0.0 ) {
+	if ( alpha === 0.0 || K === 0 ) {
 		return C;
 	}
 	// Determine the memory layouts of `A` and `B`...

lib/node_modules/@stdlib/blas/base/dgemm/test/test.dgemm.js

@@ -26,6 +26,8 @@ var tape = require( 'tape' );
 var Float64Array = require( '@stdlib/array/float64' );
 var dscal = require( '@stdlib/blas/base/dscal' );
 var resolveEnum = require( '@stdlib/blas/base/transpose-operation-resolve-enum' );
+var PINF = require( '@stdlib/constants/float64/pinf' );
+var NINF = require( '@stdlib/constants/float64/ninf' );
 var dgemm = require( './../lib/dgemm.js' );
 
 
@@ -878,3 +880,61 @@ tape( 'the function supports `conjugate-transpose`, which is equivalent to `tran
 	t.deepEqual( out, expected, 'returns expected value' );
 	t.end();
 });
+
+tape( 'if `K` is `0`, the function returns the third input matrix scaled by `β`, regardless of the value of `α` and the memory layout', function test( t ) {
+	var expected;
+	var alphas;
+	var out;
+	var c;
+	var i;
+
+	alphas = [ NaN, PINF, NINF, 2.0 ];
+	expected = new Float64Array( [ 0.5, 1.0, 1.5, 2.0 ] );
+
+	for ( i = 0; i < alphas.length; i++ ) {
+		c = new Float64Array( [ 1.0, 2.0, 3.0, 4.0 ] );
+		out = dgemm( 'row-major', 'no-transpose', 'no-transpose', 2, 2, 0, alphas[ i ], new Float64Array( 0 ), 1, new Float64Array( 0 ), 2, 0.5, c, 2 );
+		t.strictEqual( out, c, 'returns expected value' );
+		t.deepEqual( out, expected, 'returns expected value' );
+
+		c = new Float64Array( [ 1.0, 2.0, 3.0, 4.0 ] );
+		out = dgemm( 'row-major', 'no-transpose', 'transpose', 2, 2, 0, alphas[ i ], new Float64Array( 0 ), 1, new Float64Array( 0 ), 1, 0.5, c, 2 );
+		t.strictEqual( out, c, 'returns expected value' );
+		t.deepEqual( out, expected, 'returns expected value' );
+
+		c = new Float64Array( [ 1.0, 2.0, 3.0, 4.0 ] );
+		out = dgemm( 'column-major', 'transpose', 'no-transpose', 2, 2, 0, alphas[ i ], new Float64Array( 0 ), 2, new Float64Array( 0 ), 2, 0.5, c, 2 );
+		t.strictEqual( out, c, 'returns expected value' );
+		t.deepEqual( out, expected, 'returns expected value' );
+	}
+	t.end();
+});
+
+tape( 'if `K` is `0` and `β` is `0`, the function returns the third input matrix filled with zeros, regardless of the value of `α` and the memory layout', function test( t ) {
+	var expected;
+	var alphas;
+	var out;
+	var c;
+	var i;
+
+	alphas = [ NaN, PINF, NINF, 2.0 ];
+	expected = new Float64Array( [ 0.0, 0.0, 0.0, 0.0 ] );
+
+	for ( i = 0; i < alphas.length; i++ ) {
+		c = new Float64Array( [ 1.0, 2.0, 3.0, 4.0 ] );
+		out = dgemm( 'row-major', 'no-transpose', 'no-transpose', 2, 2, 0, alphas[ i ], new Float64Array( 0 ), 1, new Float64Array( 0 ), 2, 0.0, c, 2 );
+		t.strictEqual( out, c, 'returns expected value' );
+		t.deepEqual( out, expected, 'returns expected value' );
+
+		c = new Float64Array( [ 1.0, 2.0, 3.0, 4.0 ] );
+		out = dgemm( 'row-major', 'no-transpose', 'transpose', 2, 2, 0, alphas[ i ], new Float64Array( 0 ), 1, new Float64Array( 0 ), 1, 0.0, c, 2 );
+		t.strictEqual( out, c, 'returns expected value' );
+		t.deepEqual( out, expected, 'returns expected value' );
+
+		c = new Float64Array( [ 1.0, 2.0, 3.0, 4.0 ] );
+		out = dgemm( 'column-major', 'transpose', 'no-transpose', 2, 2, 0, alphas[ i ], new Float64Array( 0 ), 2, new Float64Array( 0 ), 2, 0.0, c, 2 );
+		t.strictEqual( out, c, 'returns expected value' );
+		t.deepEqual( out, expected, 'returns expected value' );
+	}
+	t.end();
+});

lib/node_modules/@stdlib/blas/base/ggemm/lib/accessors.js

@@ -29,7 +29,7 @@
 
 // VARIABLES //
 
 var bsize = blockSize( 'float64', 'float64' ); // TODO: consider using a larger block size
 
 
 // FUNCTIONS //
@@ -53,7 +53,7 @@
 * var bool = isTransposed( 'no-transpose' );
 * // returns false
 */
 function isTransposed( str ) { // TODO: consider moving to a separate helper utility package
 	return ( str !== 'no-transpose' );
 }
 
@@ -89,7 +89,7 @@
 * zeros( 2, 3, arraylike2object( toAccessorArray( X ) ), 1, 2, 0 );
 * // X => [ 0.0, 0.0, 0.0, 0.0, 0.0, 0.0 ]
 */
 function zeros( M, N, X, strideX1, strideX2, offsetX ) { // TODO: consider moving to a separate package
 	var xbuf;
 	var set;
 	var dx0;
@@ -163,7 +163,7 @@
 * scal( 2, 3, 5.0, arraylike2object( toAccessorArray( X ) ), 1, 2, 0 );
 * // X => [ 5.0, 10.0, 15.0, 20.0, 25.0, 30.0 ]
 */
 function scal( M, N, beta, X, strideX1, strideX2, offsetX ) { // TODO: consider moving to a separate package
 	var xbuf;
 	var get;
 	var set;
@@ -497,7 +497,7 @@
 		C = scal( M, N, beta, C, strideC1, strideC2, offsetC );
 	}
 	// Check whether we can early return...
-	if ( alpha === 0.0 ) {
+	if ( alpha === 0.0 || K === 0 ) {
 		return C;
 	}
 	// Determine the memory layouts of `A` and `B`...

lib/node_modules/@stdlib/blas/base/ggemm/lib/base.js

@@ -31,7 +31,7 @@
 
 // VARIABLES //
 
 var bsize = blockSize( 'float64', 'float64' ); // TODO: consider using a larger block size
 
 
 // FUNCTIONS //
@@ -55,7 +55,7 @@
 * var bool = isTransposed( 'no-transpose' );
 * // returns false
 */
 function isTransposed( str ) { // TODO: consider moving to a separate helper utility package
 	return ( str !== 'no-transpose' );
 }
 
@@ -416,7 +416,7 @@
 		C = scal( M, N, beta, C, strideC1, strideC2, offsetC );
 	}
 	// Check whether we can early return...
-	if ( alpha === 0.0 ) {
+	if ( alpha === 0.0 || K === 0 ) {
 		return C;
 	}
 	// Determine the memory layouts of `A` and `B`...

lib/node_modules/@stdlib/blas/base/ggemm/test/test.main.js

@@ -28,6 +28,8 @@ var gfill = require( '@stdlib/blas/ext/base/gfill' );
 var toAccessorArray = require( '@stdlib/array/base/to-accessor-array' );
 var copy = require( '@stdlib/array/base/copy' );
 var resolveEnum = require( '@stdlib/blas/base/transpose-operation-resolve-enum' );
+var PINF = require( '@stdlib/constants/float64/pinf' );
+var NINF = require( '@stdlib/constants/float64/ninf' );
 var ggemm = require( './../lib/main.js' );
 
 
@@ -1591,3 +1593,127 @@ tape( 'the function supports `conjugate-transpose`, which is equivalent to `tran
 	t.deepEqual( out, expected, 'returns expected value' );
 	t.end();
 });
+
+tape( 'if `K` is `0`, the function returns the third input matrix scaled by `β`, regardless of the value of `α` and the memory layout', function test( t ) {
+	var expected;
+	var alphas;
+	var out;
+	var c;
+	var i;
+
+	alphas = [ NaN, PINF, NINF, 2.0 ];
+	expected = [ 0.5, 1.0, 1.5, 2.0 ];
+
+	for ( i = 0; i < alphas.length; i++ ) {
+		c = [ 1.0, 2.0, 3.0, 4.0 ];
+		out = ggemm( 'row-major', 'no-transpose', 'no-transpose', 2, 2, 0, alphas[ i ], [], 1, [], 2, 0.5, c, 2 );
+		t.strictEqual( out, c, 'returns expected value' );
+		t.deepEqual( out, expected, 'returns expected value' );
+
+		c = [ 1.0, 2.0, 3.0, 4.0 ];
+		out = ggemm( 'row-major', 'no-transpose', 'transpose', 2, 2, 0, alphas[ i ], [], 1, [], 1, 0.5, c, 2 );
+		t.strictEqual( out, c, 'returns expected value' );
+		t.deepEqual( out, expected, 'returns expected value' );
+
+		c = [ 1.0, 2.0, 3.0, 4.0 ];
+		out = ggemm( 'column-major', 'transpose', 'no-transpose', 2, 2, 0, alphas[ i ], [], 2, [], 2, 0.5, c, 2 );
+		t.strictEqual( out, c, 'returns expected value' );
+		t.deepEqual( out, expected, 'returns expected value' );
+	}
+	t.end();
+});
+
+tape( 'if `K` is `0` and `β` is `0`, the function returns the third input matrix filled with zeros, regardless of the value of `α` and the memory layout', function test( t ) {
+	var expected;
+	var alphas;
+	var out;
+	var c;
+	var i;
+
+	alphas = [ NaN, PINF, NINF, 2.0 ];
+	expected = [ 0.0, 0.0, 0.0, 0.0 ];
+
+	for ( i = 0; i < alphas.length; i++ ) {
+		c = [ 1.0, 2.0, 3.0, 4.0 ];
+		out = ggemm( 'row-major', 'no-transpose', 'no-transpose', 2, 2, 0, alphas[ i ], [], 1, [], 2, 0.0, c, 2 );
+		t.strictEqual( out, c, 'returns expected value' );
+		t.deepEqual( out, expected, 'returns expected value' );
+
+		c = [ 1.0, 2.0, 3.0, 4.0 ];
+		out = ggemm( 'row-major', 'no-transpose', 'transpose', 2, 2, 0, alphas[ i ], [], 1, [], 1, 0.0, c, 2 );
+		t.strictEqual( out, c, 'returns expected value' );
+		t.deepEqual( out, expected, 'returns expected value' );
+
+		c = [ 1.0, 2.0, 3.0, 4.0 ];
+		out = ggemm( 'column-major', 'transpose', 'no-transpose', 2, 2, 0, alphas[ i ], [], 2, [], 2, 0.0, c, 2 );
+		t.strictEqual( out, c, 'returns expected value' );
+		t.deepEqual( out, expected, 'returns expected value' );
+	}
+	t.end();
+});
+
+tape( 'if `K` is `0`, the function returns the third input matrix scaled by `β`, regardless of the value of `α` and the memory layout (accessors)', function test( t ) {
+	var expected;
+	var alphas;
+	var cbuf;
+	var out;
+	var c;
+	var i;
+
+	alphas = [ NaN, PINF, NINF, 2.0 ];
+	expected = [ 0.5, 1.0, 1.5, 2.0 ];
+
+	for ( i = 0; i < alphas.length; i++ ) {
+		cbuf = [ 1.0, 2.0, 3.0, 4.0 ];
+		c = toAccessorArray( cbuf );
+		out = ggemm( 'row-major', 'no-transpose', 'no-transpose', 2, 2, 0, alphas[ i ], toAccessorArray( [] ), 1, toAccessorArray( [] ), 2, 0.5, c, 2 );
+		t.strictEqual( out, c, 'returns expected value' );
+		t.deepEqual( cbuf, expected, 'returns expected value' );
+
+		cbuf = [ 1.0, 2.0, 3.0, 4.0 ];
+		c = toAccessorArray( cbuf );
+		out = ggemm( 'row-major', 'no-transpose', 'transpose', 2, 2, 0, alphas[ i ], toAccessorArray( [] ), 1, toAccessorArray( [] ), 1, 0.5, c, 2 );
+		t.strictEqual( out, c, 'returns expected value' );
+		t.deepEqual( cbuf, expected, 'returns expected value' );
+
+		cbuf = [ 1.0, 2.0, 3.0, 4.0 ];
+		c = toAccessorArray( cbuf );
+		out = ggemm( 'column-major', 'transpose', 'no-transpose', 2, 2, 0, alphas[ i ], toAccessorArray( [] ), 2, toAccessorArray( [] ), 2, 0.5, c, 2 );
+		t.strictEqual( out, c, 'returns expected value' );
+		t.deepEqual( cbuf, expected, 'returns expected value' );
+	}
+	t.end();
+});
+
+tape( 'if `K` is `0` and `β` is `0`, the function returns the third input matrix filled with zeros, regardless of the value of `α` and the memory layout (accessors)', function test( t ) {
+	var expected;
+	var alphas;
+	var cbuf;
+	var out;
+	var c;
+	var i;
+
+	alphas = [ NaN, PINF, NINF, 2.0 ];
+	expected = [ 0.0, 0.0, 0.0, 0.0 ];
+
+	for ( i = 0; i < alphas.length; i++ ) {
+		cbuf = [ 1.0, 2.0, 3.0, 4.0 ];
+		c = toAccessorArray( cbuf );
+		out = ggemm( 'row-major', 'no-transpose', 'no-transpose', 2, 2, 0, alphas[ i ], toAccessorArray( [] ), 1, toAccessorArray( [] ), 2, 0.0, c, 2 );
+		t.strictEqual( out, c, 'returns expected value' );
+		t.deepEqual( cbuf, expected, 'returns expected value' );
+
+		cbuf = [ 1.0, 2.0, 3.0, 4.0 ];
+		c = toAccessorArray( cbuf );
+		out = ggemm( 'row-major', 'no-transpose', 'transpose', 2, 2, 0, alphas[ i ], toAccessorArray( [] ), 1, toAccessorArray( [] ), 1, 0.0, c, 2 );
+		t.strictEqual( out, c, 'returns expected value' );
+		t.deepEqual( cbuf, expected, 'returns expected value' );
+
+		cbuf = [ 1.0, 2.0, 3.0, 4.0 ];
+		c = toAccessorArray( cbuf );
+		out = ggemm( 'column-major', 'transpose', 'no-transpose', 2, 2, 0, alphas[ i ], toAccessorArray( [] ), 2, toAccessorArray( [] ), 2, 0.0, c, 2 );
+		t.strictEqual( out, c, 'returns expected value' );
+		t.deepEqual( cbuf, expected, 'returns expected value' );
+	}
+	t.end();
+});

lib/node_modules/@stdlib/blas/base/sgemm/lib/base.js

@@ -419,7 +419,7 @@ function sgemm( transA, transB, M, N, K, alpha, A, strideA1, strideA2, offsetA,
 		C = scal( M, N, beta, C, strideC1, strideC2, offsetC );
 	}
 	// Check whether we can early return...
-	if ( alpha === 0.0 ) {
+	if ( alpha === 0.0 || K === 0 ) {
 		return C;
 	}
 	// Determine the memory layouts of `A` and `B`...

lib/node_modules/@stdlib/blas/base/sgemm/test/test.sgemm.js

@@ -26,6 +26,8 @@ var tape = require( 'tape' );
 var Float32Array = require( '@stdlib/array/float32' );
 var sscal = require( '@stdlib/blas/base/sscal' );
 var resolveEnum = require( '@stdlib/blas/base/transpose-operation-resolve-enum' );
+var PINF = require( '@stdlib/constants/float32/pinf' );
+var NINF = require( '@stdlib/constants/float32/ninf' );
 var sgemm = require( './../lib/sgemm.js' );
 
 
@@ -878,3 +880,61 @@ tape( 'the function supports `conjugate-transpose`, which is equivalent to `tran
 	t.deepEqual( out, expected, 'returns expected value' );
 	t.end();
 });
+
+tape( 'if `K` is `0`, the function returns the third input matrix scaled by `β`, regardless of the value of `α` and the memory layout', function test( t ) {
+	var expected;
+	var alphas;
+	var out;
+	var c;
+	var i;
+
+	alphas = [ NaN, PINF, NINF, 2.0 ];
+	expected = new Float32Array( [ 0.5, 1.0, 1.5, 2.0 ] );
+
+	for ( i = 0; i < alphas.length; i++ ) {
+		c = new Float32Array( [ 1.0, 2.0, 3.0, 4.0 ] );
+		out = sgemm( 'row-major', 'no-transpose', 'no-transpose', 2, 2, 0, alphas[ i ], new Float32Array( 0 ), 1, new Float32Array( 0 ), 2, 0.5, c, 2 );
+		t.strictEqual( out, c, 'returns expected value' );
+		t.deepEqual( out, expected, 'returns expected value' );
+
+		c = new Float32Array( [ 1.0, 2.0, 3.0, 4.0 ] );
+		out = sgemm( 'row-major', 'no-transpose', 'transpose', 2, 2, 0, alphas[ i ], new Float32Array( 0 ), 1, new Float32Array( 0 ), 1, 0.5, c, 2 );
+		t.strictEqual( out, c, 'returns expected value' );
+		t.deepEqual( out, expected, 'returns expected value' );
+
+		c = new Float32Array( [ 1.0, 2.0, 3.0, 4.0 ] );
+		out = sgemm( 'column-major', 'transpose', 'no-transpose', 2, 2, 0, alphas[ i ], new Float32Array( 0 ), 2, new Float32Array( 0 ), 2, 0.5, c, 2 );
+		t.strictEqual( out, c, 'returns expected value' );
+		t.deepEqual( out, expected, 'returns expected value' );
+	}
+	t.end();
+});
+
+tape( 'if `K` is `0` and `β` is `0`, the function returns the third input matrix filled with zeros, regardless of the value of `α` and the memory layout', function test( t ) {
+	var expected;
+	var alphas;
+	var out;
+	var c;
+	var i;
+
+	alphas = [ NaN, PINF, NINF, 2.0 ];
+	expected = new Float32Array( [ 0.0, 0.0, 0.0, 0.0 ] );
+
+	for ( i = 0; i < alphas.length; i++ ) {
+		c = new Float32Array( [ 1.0, 2.0, 3.0, 4.0 ] );
+		out = sgemm( 'row-major', 'no-transpose', 'no-transpose', 2, 2, 0, alphas[ i ], new Float32Array( 0 ), 1, new Float32Array( 0 ), 2, 0.0, c, 2 );
+		t.strictEqual( out, c, 'returns expected value' );
+		t.deepEqual( out, expected, 'returns expected value' );
+
+		c = new Float32Array( [ 1.0, 2.0, 3.0, 4.0 ] );
+		out = sgemm( 'row-major', 'no-transpose', 'transpose', 2, 2, 0, alphas[ i ], new Float32Array( 0 ), 1, new Float32Array( 0 ), 1, 0.0, c, 2 );
+		t.strictEqual( out, c, 'returns expected value' );
+		t.deepEqual( out, expected, 'returns expected value' );
+
+		c = new Float32Array( [ 1.0, 2.0, 3.0, 4.0 ] );
+		out = sgemm( 'column-major', 'transpose', 'no-transpose', 2, 2, 0, alphas[ i ], new Float32Array( 0 ), 2, new Float32Array( 0 ), 2, 0.0, c, 2 );
+		t.strictEqual( out, c, 'returns expected value' );
+		t.deepEqual( out, expected, 'returns expected value' );
+	}
+	t.end();
+});