[PATCH v2 0/2] MR1618: Draft: gdiplus: improve performance of GdipScaleMatrix

[Bartosz Kosiorek (＠gang65)]

--
  v2: gdiplus: improve performance of matrix multiplication by unrolling loop
      gdiplus: improve performance of GdipScaleMatrix

https://gitlab.winehq.org/wine/wine/-/merge_requests/1618

[PATCH v2 1/2] gdiplus: improve performance of GdipScaleMatrix

[Bartosz Kosiorek]

From: Bartosz Kosiorek

Wine-Bug: https://bugs.winehq.org/show_bug.cgi?id=53947
---
 dlls/gdiplus/matrix.c       | 27 ++++++++++++++++---------
 dlls/gdiplus/tests/matrix.c | 40 +++++++++++++++++++++++++++++++++++++
 2 files changed, 58 insertions(+), 9 deletions(-)

diff --git a/dlls/gdiplus/matrix.c b/dlls/gdiplus/matrix.c
index 40abbc93e21..d96f425aeef 100644
--- a/dlls/gdiplus/matrix.c
+++ b/dlls/gdiplus/matrix.c
@@ -294,17 +294,26 @@ GpStatus WINGDIPAPI GdipScaleMatrix(GpMatrix *matrix, REAL scaleX, REAL scaleY,
     if(!matrix)
         return InvalidParameter;

-    scale[0] = scaleX;
-    scale[1] = 0.0;
-    scale[2] = 0.0;
-    scale[3] = scaleY;
-    scale[4] = 0.0;
-    scale[5] = 0.0;
-
     if(order == MatrixOrderAppend)
-        matrix_multiply(matrix->matrix, scale, matrix->matrix);
+    {
+        scale[0] = matrix->matrix[0] * scaleX;
+        scale[1] = matrix->matrix[1] * scaleY;
+        scale[2] = matrix->matrix[2] * scaleX;
+        scale[3] = matrix->matrix[3] * scaleY;
+        scale[4] = matrix->matrix[4] * scaleX;
+        scale[5] = matrix->matrix[5] * scaleY;
+        memcpy(matrix->matrix, scale, 6 * sizeof(REAL));
+    }
     else if (order == MatrixOrderPrepend)
-        matrix_multiply(scale, matrix->matrix, matrix->matrix);
+    {
+        scale[0] = scaleX * matrix->matrix[0];
+        scale[1] = scaleX * matrix->matrix[1];
+        scale[2] = scaleY * matrix->matrix[2];
+        scale[3] = scaleY * matrix->matrix[3];
+        scale[4] = matrix->matrix[4];
+        scale[5] = matrix->matrix[5];
+        memcpy(matrix->matrix, scale, 6 * sizeof(REAL));
+    }
     else
         return InvalidParameter;

diff --git a/dlls/gdiplus/tests/matrix.c b/dlls/gdiplus/tests/matrix.c
index 5ca0209f6a9..5f219dafff1 100644
--- a/dlls/gdiplus/tests/matrix.c
+++ b/dlls/gdiplus/tests/matrix.c
@@ -110,6 +110,45 @@ static void test_transform(void)
     GdipDeleteMatrix(matrix);
 }

+static void test_scale(void)
+{
+    GpStatus status;
+    GpMatrix *matrix = NULL;
+    REAL elems[6];
+
+    REAL expected_elem[] = {3.0, -4.0, 90.0, 80.0, -1500.0, 1200.0};
+    REAL expected_elem2[] = {3.0, -6.0, 60.0, 80.0, -500.0, 600.0};
+
+    GdipCreateMatrix2(1.0, -2.0, 30.0, 40.0, -500.0, 600.0, &matrix);
+
+    status = GdipScaleMatrix(NULL, 3, 2, MatrixOrderAppend);
+    expect(InvalidParameter, status);
+    status = GdipScaleMatrix(matrix, 3, 2, MatrixOrderAppend);
+    expect(Ok, status);
+
+    status = GdipGetMatrixElements(matrix, elems);
+    expect(Ok, status);
+    GdipDeleteMatrix(matrix);
+
+    for(INT i = 0; i < 6; i++)
+        ok(expected_elem[i] == elems[i], "Expected #%d to be (%.1f) but got (%.1f)\n", i,
+           expected_elem[i], elems[i]);
+
+    GdipCreateMatrix2(1.0, -2.0, 30.0, 40.0, -500.0, 600.0, &matrix);
+
+    status = GdipScaleMatrix(matrix, 3, 2, MatrixOrderPrepend);
+    expect(Ok, status);
+
+    status = GdipGetMatrixElements(matrix, elems);
+    expect(Ok, status);
+    GdipDeleteMatrix(matrix);
+
+    for(INT i = 0; i < 6; i++)
+        ok(expected_elem2[i] == elems[i], "Expected #%d to be (%.1f) but got (%.1f)\n", i,
+           expected_elem2[i], elems[i]);
+
+}
+
 static void test_isinvertible(void)
 {
     GpStatus status;
@@ -390,6 +429,7 @@ START_TEST(matrix)
     GdiplusStartup(&gdiplusToken, &gdiplusStartupInput, NULL);

test_constructor_destructor();
+    test_scale();
     test_transform();
     test_isinvertible();
     test_invert();

[PATCH v2 2/2] gdiplus: improve performance of matrix multiplication by unrolling loop

[Bartosz Kosiorek]

From: Bartosz Kosiorek

---
 dlls/gdiplus/matrix.c | 16 +++++++++-------
 1 file changed, 9 insertions(+), 7 deletions(-)

diff --git a/dlls/gdiplus/matrix.c b/dlls/gdiplus/matrix.c
index d96f425aeef..fa287e6cb00 100644
--- a/dlls/gdiplus/matrix.c
+++ b/dlls/gdiplus/matrix.c
@@ -42,13 +42,15 @@ WINE_DEFAULT_DEBUG_CHANNEL(gdiplus);
 static void matrix_multiply(GDIPCONST REAL * left, GDIPCONST REAL * right, REAL * out)
 {
     REAL temp[6];
-    int i, odd;
-
-    for(i = 0; i < 6; i++){
-        odd = i % 2;
-        temp[i] = left[i - odd] * right[odd] + left[i - odd + 1] * right[odd + 2] +
-                  (i >= 4 ? right[odd + 4] : 0.0);
-    }
+    // Unroll loop to optimize execution speed
+    temp[0] = left[0] * right[0] + left[1] * right[2];
+    temp[1] = left[0] * right[1] + left[1] * right[3];
+    temp[2] = left[2] * right[0] + left[3] * right[2];
+    temp[3] = left[2] * right[1] + left[3] * right[3];
+    temp[4] = left[4] * right[0] + left[5] * right[2] +
+                right[4];
+    temp[5] = left[4] * right[1] + left[5] * right[3] +
+                right[5];

memcpy(out, temp, 6 * sizeof(REAL));
 }