Fix bounds checking in dgemm and sgemm, add offset test cases

Co-authored-by: Quafadas <[email protected]>

Author: Copilot

blas/src/main/java/dev/ludovic/netlib/blas/AbstractBLAS.java

@@ -291,9 +291,9 @@ public void dgemm(String transa, String transb, int m, int n, int k, double alph
     requireNonNull(a);
     requireNonNull(b);
     requireNonNull(c);
-    checkIndex(offseta + (lsame("N", transa) ? k : m) * lda - 1, a.length);
-    checkIndex(offsetb + (lsame("N", transb) ? n : k) * ldb - 1, b.length);
-    checkIndex(offsetc + m * n - 1, c.length);
+    checkIndex(offseta + (lsame("N", transa) ? (k - 1) * lda + (m - 1) : (m - 1) * lda + (k - 1)), a.length);
+    checkIndex(offsetb + (lsame("N", transb) ? (n - 1) * ldb + (k - 1) : (k - 1) * ldb + (n - 1)), b.length);
+    checkIndex(offsetc + (n - 1) * ldc + (m - 1), c.length);
     dgemmK(transa, transb, m, n, k, alpha, a, offseta, lda, b, offsetb, ldb, beta, c, offsetc, ldc);
   }
 
@@ -321,9 +321,9 @@ public void sgemm(String transa, String transb, int m, int n, int k, float alpha
     requireNonNull(a);
     requireNonNull(b);
     requireNonNull(c);
-    checkIndex(offseta + (lsame("N", transa) ? k : m) * lda - 1, a.length);
-    checkIndex(offsetb + (lsame("N", transb) ? n : k) * ldb - 1, b.length);
-    checkIndex(offsetc + m * n - 1, c.length);
+    checkIndex(offseta + (lsame("N", transa) ? (k - 1) * lda + (m - 1) : (m - 1) * lda + (k - 1)), a.length);
+    checkIndex(offsetb + (lsame("N", transb) ? (n - 1) * ldb + (k - 1) : (k - 1) * ldb + (n - 1)), b.length);
+    checkIndex(offsetc + (n - 1) * ldc + (m - 1), c.length);
     sgemmK(transa, transb, m, n, k, alpha, a, offseta, lda, b, offsetb, ldb, beta, c, offsetc, ldc);
   }
 

blas/src/test/java/dev/ludovic/netlib/blas/DgemmTest.java

@@ -185,4 +185,25 @@ void testSanity(BLAS blas) {
         blas.dgemm("T", "T", M, N/2, K, 0.0, dgeAT, K, dgeBT, N/2, 1.0, dgeCcopy = dgeC.clone(), M);
         assertArrayEquals(expected, dgeCcopy, depsilon);
     }
+
+    @ParameterizedTest
+    @MethodSource("BLASImplementations")
+    void testOffsetBoundsChecking(BLAS blas) {
+        // Test case that reproduces the original bounds checking issue
+        // Matrix A (2x3) with offset=1, stored in array of length 9
+        double[] a = {1.0, 4.0, 7.0, 2.0, 5.0, 8.0, 3.0, 6.0, 9.0};
+        double[] b = {1.0, 2.0, 3.0};
+        double[] c = new double[2];
+        double[] cExpected = new double[2];
+
+        // This should not throw IndexOutOfBoundsException
+        assertDoesNotThrow(() -> {
+            blas.dgemm("N", "N", 2, 1, 3, 1.0, a, 1, 3, b, 0, 3, 0.0, c, 0, 2);
+        });
+
+        // Compare with F2J result
+        f2j.dgemm("N", "N", 2, 1, 3, 1.0, a, 1, 3, b, 0, 3, 0.0, cExpected, 0, 2);
+        blas.dgemm("N", "N", 2, 1, 3, 1.0, a, 1, 3, b, 0, 3, 0.0, c, 0, 2);
+        assertArrayEquals(cExpected, c, depsilon);
+    }
 }

blas/src/test/java/dev/ludovic/netlib/blas/SgemmTest.java

@@ -181,4 +181,25 @@ void testSanity(BLAS blas) {
         blas.sgemm("T", "T", M, N/2, K, 0.0f, sgeAT, K, sgeBT, N/2, 1.0f, sgeCcopy = sgeC.clone(), M);
         assertArrayEquals(expected, sgeCcopy, sepsilon);
     }
+
+    @ParameterizedTest
+    @MethodSource("BLASImplementations")
+    void testOffsetBoundsChecking(BLAS blas) {
+        // Test case that reproduces the original bounds checking issue for sgemm
+        // Matrix A (2x3) with offset=1, stored in array of length 9
+        float[] a = {1.0f, 4.0f, 7.0f, 2.0f, 5.0f, 8.0f, 3.0f, 6.0f, 9.0f};
+        float[] b = {1.0f, 2.0f, 3.0f};
+        float[] c = new float[2];
+        float[] cExpected = new float[2];
+
+        // This should not throw IndexOutOfBoundsException
+        assertDoesNotThrow(() -> {
+            blas.sgemm("N", "N", 2, 1, 3, 1.0f, a, 1, 3, b, 0, 3, 0.0f, c, 0, 2);
+        });
+
+        // Compare with F2J result
+        f2j.sgemm("N", "N", 2, 1, 3, 1.0f, a, 1, 3, b, 0, 3, 0.0f, cExpected, 0, 2);
+        blas.sgemm("N", "N", 2, 1, 3, 1.0f, a, 1, 3, b, 0, 3, 0.0f, c, 0, 2);
+        assertArrayEquals(cExpected, c, sepsilon);
+    }
 }