GPU byte swapping.

A lot of other packets likely do this.

src/xenia/gpu/graphics_system.cc

@@ -161,7 +161,7 @@ void GraphicsSystem::WriteRegister(uint32_t r, uint64_t value) {
 }
 
 void GraphicsSystem::DispatchInterruptCallback(
-  uint32_t source, uint32_t cpu) {
+    uint32_t source, uint32_t cpu) {
   // Pick a CPU, if needed. We're going to guess 2. Because.
   if (cpu == 0xFFFFFFFF) {
     cpu = 2;

src/xenia/gpu/ring_buffer_worker.cc

@@ -483,13 +483,10 @@ uint32_t RingBufferWorker::ExecutePacket(PacketArgs& args) {
             // Write value.
             data_value = value;
           }
-          if (!(address & 0xC0000000)) {
-            XESETUINT32BE(p + TRANSLATE_ADDR(address), data_value);
-          } else {
-            // TODO(benvanik): read up on PM4_EVENT_WRITE_SHD.
-            // No clue. Maybe relative write based on a register base?
-            XELOGE("UNKNOWN FORM OF PM4_EVENT_WRITE_SHD");
-          }
+          uint32_t endianness = address >> 29;
+          address &= ~0xC0000000;
+          data_value = GpuSwap(data_value, endianness);
+          XESETUINT32BE(p + TRANSLATE_ADDR(address), data_value);
         }
         break;
 

src/xenia/gpu/xenos/xenos.h

@@ -42,6 +42,22 @@ typedef enum {
   XE_GPU_PRIMITIVE_TYPE_LINE_LOOP         = 0x0C,
 } XE_GPU_PRIMITIVE_TYPE;
 
+XEFORCEINLINE uint32_t GpuSwap(uint32_t value, uint32_t endianness) {
+  switch (endianness) {
+  default:
+  case 0x0: // No swap.
+    return value;
+  case 0x1: // Swap bytes in half words.
+    return ((value << 8) & 0xFF00FF00) |
+           ((value >> 8) & 0x00FF00FF);
+  case 0x2: // Swap bytes.
+    // NOTE: we are likely doing two swaps here. Wasteful. Oh well.
+    return XESWAP32(value);
+  case 0x3: // Swap half words.
+    return ((value >> 16) & 0xFFFF) | (value << 16);
+  }
+}
+
 // XE_GPU_REG_SQ_PROGRAM_CNTL
 typedef union {
   XEPACKEDSTRUCTANONYMOUS({