FIX: matrix_multiply now works when given a column matrix

numpy · ovillellas · Oct 19, 2012 · Oct 19, 2012 · Oct 24, 2012 · Nov 7, 2012
commit 0e23dbbf7ca48a1fe927043cd7006096751e315d
diff --git a/numpy/core/src/umath/umath_linalg.c.src b/numpy/core/src/umath/umath_linalg.c.src
@@ -1112,63 +1112,101 @@ dump_gemm_params(const char* name, const GEMM_PARAMS_t* params)
               "TRANSA", ' ', params->TRANSA, "TRANSB", ' ', params->TRANSB);
 }
 
+
+typedef struct _matrix_desc {
+    int need_buff;
+    fortran_int lead;
+    size_t size;
+    void *buff;
+} matrix_desc;
+
+static inline void
+matrix_desc_init(matrix_desc *mtxd,
+                 npy_intp* steps,
+                 size_t sot,
+                 fortran_int rows,
+                 fortran_int columns)
+/* initialized a matrix desc based on the information in steps
+   it will fill the matrix desc with the values needed.
+   steps[0] contains column step
+   steps[1] contains row step
+*/
+{
+    /* if column step is not element step, a copy will be needed
+       to arrange the array in a way compatible with blas
+    */
+    mtxd->need_buff = steps[0] != sot;
+
    if (!mtxd->need_buff) {
+        if (steps[1]) {
+            mtxd->lead = (fortran_int) steps[1] / sot;
+        } else {
+            /* step is 0, this means either it is only 1 column or
+               there is a step trick around*/
+            if (columns > 1) {
+                /* step tricks not supported in gemm... make a copy */
+                mtxd->need_buff = 1;
+            } else {
+                /* lead must be at least 1 */
+                mtxd->lead = rows;
+            }
+        }
+    }
+
+    /* if a copy is performed, the lead is the number of columns */
+    if (mtxd->need_buff) {
+        mtxd->lead = rows;
+    }
+
+    mtxd->size = rows*columns*sot*mtxd->need_buff;
+    mtxd->buff = NULL;
+}
+
+static inline npy_int8 *
+matrix_desc_assign_buff(matrix_desc* mtxd, npy_int8 *p)
+{
+    if (mtxd->need_buff) {
+        mtxd->buff = p;
+        return p + mtxd->size;
+    } else {
+        return p;
+    }
+}
+
 static inline int
 init_gemm_params(GEMM_PARAMS_t *params,
                  fortran_int m,
                  fortran_int k,
                  fortran_int n,
                  npy_intp* steps,
-                 size_t size_of_type)
+                 size_t sot)
 {
+    matrix_desc a, b, c;
+    matrix_desc_init(&a, steps + 0, sot, m, k);
+    matrix_desc_init(&b, steps + 2, sot, k, n);
+    matrix_desc_init(&c, steps + 4, sot, m, n);
     npy_uint8 *mem_buff = NULL;
-    int needs_a, needs_b, needs_c;
-    size_t a_size, b_size, c_size;
-    void *a, *b, *c;
-
-    needs_a = steps[0] != size_of_type;
-    needs_b = steps[2] != size_of_type;
-    needs_c = steps[4] != size_of_type;
 
-    a_size = m*k*size_of_type*needs_a;
-    b_size = k*n*size_of_type*needs_b;
-    c_size = m*n*size_of_type*needs_c;
-
-    if (a_size + b_size + c_size)
+    if (a.size + b.size + c.size)
     {
         npy_uint8 *current;
         /* need to allocate some buffer */
-        mem_buff = malloc(a_size + b_size + c_size);
+        mem_buff = malloc(a.size + b.size + c.size);
         if (!mem_buff)
             goto error;
 
         current = mem_buff;
-        if (needs_a) {
-            a = current;
-            current += a_size;
-        } else {
-            a = NULL;
-        }
-
-        if (needs_b) {
-            b = current;
-            current += b_size;
-        } else {
-            b = NULL;
-        }
-
-        if (needs_c) {
-            c = current;
-        } else {
-            c = NULL;
-        }
+        current = matrix_desc_assign_buff(&a, current);
+        current = matrix_desc_assign_buff(&b, current);
+        current = matrix_desc_assign_buff(&c, current);
     }
 
-    params->A = a;
-    params->B = b;
-    params->C = c;
-    params->LDA = (fortran_int)(needs_a ? m : steps[1]/size_of_type);
-    params->LDB = (fortran_int)(needs_b ? k : steps[3]/size_of_type);
-    params->LDC = (fortran_int)(needs_c ? m : steps[5]/size_of_type);
+    params->A = a.buff;
+    params->B = b.buff;
+    params->C = c.buff;
+    params->LDA = a.lead;
+    params->LDB = b.lead;
+    params->LDC = c.lead;
     params->M = m;
     params->N = n;
     params->K = k;

diff --git a/numpy/core/tests/test_gufuncs_linalg.py b/numpy/core/tests/test_gufuncs_linalg.py
@@ -271,6 +271,11 @@ def do(self, a, b):
         else:
             assert_almost_equal(res[0], np.dot(a[0],b[0]))
 
+    def test_column_matrix(self):
+        A = np.arange(2*2).reshape((2,2))
+        B = np.arange(2*1).reshape((2,1))
+        res = gula.matrix_multiply(A,B)
+        assert_almost_equal(res, np.dot(A,B))
 
 class TestInv(GeneralTestCase, TestCase):
     def do(self, a, b):