wiiu/render: Improve SDL_UpdateTexture performance

When possible, use the DMA to copy the whole texture data at once, this
improves performance by nearly 15x in the best case.
In all other cases, use OSBlockMove instead of the base memcpy
implementation for a ~1.5x speedup in the slowest path.

Author: Mefiresu

src/render/wiiu/SDL_rtexture_wiiu.c

@@ -33,6 +33,10 @@
 #include <gx2/mem.h>
 #include <gx2r/surface.h>
 #include <gx2r/resource.h>
+#include <coreinit/cache.h>
+#include <coreinit/memory.h>
+#include <dmae/mem.h>
+#include <dmae/sync.h>
 
 #include <malloc.h>
 #include <stdarg.h>
@@ -170,8 +174,10 @@ int WIIU_SDL_UpdateTexture(SDL_Renderer * renderer, SDL_Texture * texture,
     WIIU_VideoData *videodata = (WIIU_VideoData *) SDL_GetVideoDevice()->driverdata;
     Uint32 BytesPerPixel = SDL_BYTESPERPIXEL(texture->format);
     size_t length = rect->w * BytesPerPixel;
+    size_t total_size = length * rect->h;
     Uint8 *src = (Uint8 *) pixels, *dst;
     int row, dst_pitch;
+    bool src_aligned, dst_aligned;
 
     if (!videodata->hasForeground) {
         return 0;
@@ -180,10 +186,30 @@ int WIIU_SDL_UpdateTexture(SDL_Renderer * renderer, SDL_Texture * texture,
     /* We write the rules, and we say all textures are streaming */
     WIIU_SDL_LockTexture(renderer, texture, rect, (void**)&dst, &dst_pitch);
 
-    for (row = 0; row < rect->h; ++row) {
-        SDL_memcpy(dst, src, length);
-        src += pitch;
-        dst += dst_pitch;
+    if (length == pitch && length == dst_pitch) {
+        /* DMA requires 8 byte alignment */
+        src_aligned = ((uintptr_t)src & 7U) == 0;
+        dst_aligned = ((uintptr_t)dst & 7U) == 0;
+        /* DMA works best on textures bigger than 5KiB */
+        if ((total_size > 5120) && (src_aligned && dst_aligned)) {
+            DCFlushRange(src, total_size);
+            /* Run a single DMA transfer and wait until transfer is done */
+            while (!DMAEWaitDone(DMAECopyMem(dst, src, total_size >> 2, DMAE_SWAP_NONE)));
+        } else {
+            /* Otherwise, fallback to a single memory copy */
+            OSBlockMove(dst, src, total_size, true);
+        }
+    } else {
+        /* Flush source first */
+        DCFlushRange(src, total_size);
+        for (row = 0; row < rect->h; ++row) {
+            /* Do not flush per line here */
+            OSBlockMove(dst, src, length, false);
+            src += pitch;
+            dst += dst_pitch;
+        }
+        /* Now that we're done, we can flush the entire destination at once */
+        DCFlushRange(dst, dst_pitch * rect->h);
     }
 
     WIIU_SDL_UnlockTexture(renderer, texture);