ddraw: Use system memory for version 4 vertex buffers.

dlls/ddraw/ddraw_private.h

@@ -595,6 +595,7 @@ struct d3d_vertex_buffer
     DWORD                size;
     BOOL                 dynamic;
     bool discarded;
+    bool sysmem;
 };
 
 HRESULT d3d_vertex_buffer_create(struct d3d_vertex_buffer **buffer, struct ddraw *ddraw,

dlls/ddraw/device.c

@@ -4128,7 +4128,7 @@ static HRESULT d3d_device7_DrawPrimitiveVB(IDirect3DDevice7 *iface, D3DPRIMITIVE
 
     stride = get_flexible_vertex_size(vb_impl->fvf);
 
-    if (vb_impl->Caps & D3DVBCAPS_SYSTEMMEMORY)
+    if (vb_impl->sysmem)
     {
         TRACE("Drawing from D3DVBCAPS_SYSTEMMEMORY vertex buffer, forwarding to DrawPrimitive().\n");
         wined3d_mutex_lock();
@@ -4246,7 +4246,7 @@ static HRESULT d3d_device7_DrawIndexedPrimitiveVB(IDirect3DDevice7 *iface,
 
     vb_impl->discarded = false;
 
-    if (vb_impl->Caps & D3DVBCAPS_SYSTEMMEMORY)
+    if (vb_impl->sysmem)
     {
         TRACE("Drawing from D3DVBCAPS_SYSTEMMEMORY vertex buffer, forwarding to DrawIndexedPrimitive().\n");
         wined3d_mutex_lock();

dlls/ddraw/vertexbuffer.c

@@ -120,7 +120,7 @@ static HRESULT d3d_vertex_buffer_create_wined3d_buffer(struct d3d_vertex_buffer
     if (dynamic)
         desc.usage |= WINED3DUSAGE_DYNAMIC;
     desc.bind_flags = WINED3D_BIND_VERTEX_BUFFER;
-    if (buffer->Caps & D3DVBCAPS_SYSTEMMEMORY)
+    if (buffer->sysmem)
         desc.access = WINED3D_RESOURCE_ACCESS_CPU | WINED3D_RESOURCE_ACCESS_MAP_R | WINED3D_RESOURCE_ACCESS_MAP_W;
     else
         desc.access = WINED3D_RESOURCE_ACCESS_GPU | WINED3D_RESOURCE_ACCESS_MAP_R | WINED3D_RESOURCE_ACCESS_MAP_W;
@@ -466,6 +466,19 @@ HRESULT d3d_vertex_buffer_create(struct d3d_vertex_buffer **vertex_buf,
     buffer->fvf = desc->dwFVF;
     buffer->size = get_flexible_vertex_size(desc->dwFVF) * desc->dwNumVertices;
 
+    /* ddraw4 vertex buffers ignore DISCARD and NOOVERWRITE, even on
+     * pretransformed geometry, which means that a GPU-based buffer cannot
+     * perform well.
+     *
+     * While at least one contemporaneous card (Geforce 4) does seem to show a
+     * difference in its performance characteristics based on whether
+     * D3DVBCAPS_SYSTEMMEMORY is set, it also doesn't *improve* performance to
+     * use a non-SYSTEMMEMORY buffer with ddraw4. For wined3d it should always
+     * be better to use sysmem.
+     *
+     * This improves performance in Prince of Persia 3D. */
+    buffer->sysmem = ((buffer->Caps & D3DVBCAPS_SYSTEMMEMORY) || buffer->version < 7);
+
     wined3d_mutex_lock();
 
     if (FAILED(hr = d3d_vertex_buffer_create_wined3d_buffer(buffer, FALSE, &buffer->wined3d_buffer)))