Re: [PATCH v3 0/1] MR3971: gdiplus: Opimize GDI32_GdipDrawPath

[Esme Povirk (＠madewokherd)]

On Sat Oct 14 22:06:36 2023 +0000, Bartosz Kosiorek wrote:
> With previous implementation we have 2*4 float multiplication inside
> `GdipTransformMatrixPoints` function:
>         pts[i].X = x * matrix->matrix[0] + y * matrix->matrix[2] + matrix->matrix[4];
>         pts[i].Y = x * matrix->matrix[1] + y * matrix->matrix[3] + matrix->matrix[5];
>
> Here is the full source code:
> GpStatus WINGDIPAPI GdipTransformMatrixPoints(GpMatrix *matrix, GpPointF *pts,
>                                               INT count)
> {
>     REAL x, y;
>     INT i;
>     TRACE("(%s, %p, %d)\n", debugstr_matrix(matrix), pts, count);
>     if(!matrix || !pts || count <= 0)
>         return InvalidParameter;
>     for(i = 0; i < count; i++)
>     {
>         x = pts[i].X;
>         y = pts[i].Y;
>         pts[i].X = x * matrix->matrix[0] + y * matrix->matrix[2] + matrix->matrix[4];
>         pts[i].Y = x * matrix->matrix[1] + y * matrix->matrix[3] + matrix->matrix[5];
>     }
>     return Ok;
> }
>
> As the vector is (0,0) and (1, 1), we could replace invocation of this
> function by simple assigment (no need to multiple anything):
>         scale_x = graphics->worldtrans.matrix[0] + graphics->worldtrans.matrix[2];
>         scale_y = graphics->worldtrans.matrix[1] + graphics->worldtrans.matrix[3];
>
> We could also optimize it more and calculate width via `sqrt` only once
> (I would like to leave it for next MR).

This seems simple enough that I have no problem changing it without performance numbers. Note that we immediately turn around and do the same thing with graphics->gdi_transform (with some indirection through gdip_transform_points and get_graphics_transform that could be eliminated).