vector_alu: Improve handling of mbcnt append/consume patterns (#3184)

* vector_alu: Improve handling of mbcnt append/consume patterns

The existing implementation was written to handle a single pattern of mbcnt before the DS_APPEND instruction

v_mbcnt_hi_u32_b32 vX, exec_hi, 0
v_mbcnt_lo_u32_b32 vX, exec_lo, vX
ds_append       vY offset:4 gds
v_add_i32       vX, vcc, vY, vX

In this case however the DS_APPEND is before the mbcnt pattern

ds_append       vX gds
v_mbcnt_hi_u32_b32 vY, exec_hi, vX
v_mbcnt_lo_u32_b32 vZ, exec_lo, vY

The mbcnt instructions are always in pairs of hi/lo and in general are quite flexible. But they assume the subgroup size is 64 so they are not recompiled literally. Together with DS_APPEND they are used to derive a unique per thread index in a buffer (different from using thread_id as order could be random). DS_APPEND instruction works on per subgroup level, by adding number of active threads of subgroup to the GDS counter, essentially giving a multiple-of-64 base index to all threads. Then each thread executes the mbcnt pair which returns the number of active threads with id less than the itself and adds it with the base.

The recompiler translates DS_APPEND into an atomic increment of a storage buffer counter, which already gives the desired unique index, so this pattern is a no-op. On main it was set to zero as per the first pattern to avoid altering the DS_APPEND result. The new handling passes through the initial value of the pattern instead, which has the same effect but works on either case.

* vk_rasterizer: Always sync DMA buffers

src/shader_recompiler/frontend/translate/vector_alu.cpp

@@ -558,27 +558,31 @@ void Translator::V_BCNT_U32_B32(const GcnInst& inst) {
 
 void Translator::V_MBCNT_U32_B32(bool is_low, const GcnInst& inst) {
     if (!is_low) {
-        // v_mbcnt_hi_u32_b32 v2, -1, 0
+        // v_mbcnt_hi_u32_b32 vX, -1, 0
         if (inst.src[0].field == OperandField::SignedConstIntNeg && inst.src[0].code == 193 &&
             inst.src[1].field == OperandField::ConstZero) {
             return;
         }
-        // v_mbcnt_hi_u32_b32 vX, exec_hi, 0
-        if (inst.src[0].field == OperandField::ExecHi &&
-            inst.src[1].field == OperandField::ConstZero) {
-            return;
+        // v_mbcnt_hi_u32_b32 vX, exec_hi, 0/vZ
+        if ((inst.src[0].field == OperandField::ExecHi ||
+             inst.src[0].field == OperandField::VccHi) &&
+            (inst.src[1].field == OperandField::ConstZero ||
+             inst.src[1].field == OperandField::VectorGPR)) {
+            return SetDst(inst.dst[0], GetSrc(inst.src[1]));
         }
+        UNREACHABLE();
     } else {
-        // v_mbcnt_lo_u32_b32 v2, -1, vX
+        // v_mbcnt_lo_u32_b32 vY, -1, vX
         // used combined with above to fetch lane id in non-compute stages
         if (inst.src[0].field == OperandField::SignedConstIntNeg && inst.src[0].code == 193) {
-            SetDst(inst.dst[0], ir.LaneId());
+            return SetDst(inst.dst[0], ir.LaneId());
         }
-        // v_mbcnt_lo_u32_b32 v20, exec_lo, vX
-        // used combined in above for append buffer indexing.
-        if (inst.src[0].field == OperandField::ExecLo) {
-            SetDst(inst.dst[0], ir.Imm32(0));
+        // v_mbcnt_lo_u32_b32 vY, exec_lo, vX
+        // used combined with above for append buffer indexing.
+        if (inst.src[0].field == OperandField::ExecLo || inst.src[0].field == OperandField::VccLo) {
+            return SetDst(inst.dst[0], GetSrc(inst.src[1]));
         }
+        UNREACHABLE();
     }
 }
 

src/video_core/renderer_vulkan/vk_rasterizer.cpp

@@ -471,7 +471,7 @@ bool Rasterizer::BindResources(const Pipeline* pipeline) {
         uses_dma |= stage->uses_dma;
     }
 
-    if (uses_dma && !fault_process_pending) {
+    if (uses_dma) {
         // We only use fault buffer for DMA right now.
         {
             Common::RecursiveSharedLock lock{mapped_ranges_mutex};