graphics: derive compute wave size from dispatch mode

src/graphics/guest_gpu/command_processor/pm4Handlers.cpp

@@ -1360,10 +1360,6 @@ static void HwShSetCsRegister(CommandProcessor& cp, uint32_t cmd_offset, uint32_
 			                         Pm4::COMPUTE_PGM_RSRC1_BULKY_MASK) != 0u;
 			cs_regs.threadgroup_configuration = ((value >> Pm4::COMPUTE_PGM_RSRC1_WGP_MODE_SHIFT) &
 			                                     Pm4::COMPUTE_PGM_RSRC1_WGP_MODE_MASK) != 0u;
-			cs_regs.wave_size                 = (((value >> Pm4::COMPUTE_PGM_RSRC1_W32_EN_SHIFT) &
-			                                      Pm4::COMPUTE_PGM_RSRC1_W32_EN_MASK) != 0u)
-			                                        ? 32u
-			                                        : 64u;
 			break;
 		case Pm4::COMPUTE_PGM_RSRC2:
 			cs_regs.scratch_en     = ((value >> Pm4::COMPUTE_PGM_RSRC2_SCRATCH_EN_SHIFT) &

src/graphics/guest_gpu/graphicsRun.cpp

@@ -1108,6 +1108,8 @@ void CommandProcessor::DrawIndirectMulti(uint32_t data_offset, uint32_t max_coun
 
 void CommandProcessor::DispatchDirect(uint32_t thread_group_x, uint32_t thread_group_y,
                                       uint32_t thread_group_z, uint32_t mode) {
+	m_sh_ctx.SetCsWaveSize(Pm4::ComputeWaveSize(mode));
+
 	uint32_t frame_num = 0;
 	// uint32_t local_x   = 1;
 	// uint32_t local_y   = 1;

src/graphics/guest_gpu/hardwareContext.h

@@ -1089,6 +1089,7 @@ public:
 	void SetPsShaderChksum(uint32_t value) { m_ps.ps_regs.chksum = value; }
 
 	void SetCsShader(const CsStageRegisters& cs_regs) { m_cs.cs_regs = cs_regs; }
+	void SetCsWaveSize(uint8_t wave_size) { m_cs.cs_regs.wave_size = wave_size; }
 
 	void SetPsUserSgpr(uint32_t id, uint32_t value, UserSgprType type) {
 		m_ps.ps_user_sgpr.value[id] = value;

src/graphics/guest_gpu/pm4.h

@@ -984,8 +984,16 @@ constexpr uint32_t COMPUTE_PGM_RSRC1_FP16_OVFL_SHIFT  = 26;
 constexpr uint32_t COMPUTE_PGM_RSRC1_FP16_OVFL_MASK   = 0x1;
 constexpr uint32_t COMPUTE_PGM_RSRC1_WGP_MODE_SHIFT   = 29;
 constexpr uint32_t COMPUTE_PGM_RSRC1_WGP_MODE_MASK    = 0x1;
-constexpr uint32_t COMPUTE_PGM_RSRC1_W32_EN_SHIFT     = 30;
+
-constexpr uint32_t COMPUTE_PGM_RSRC1_W32_EN_MASK      = 0x1;
+constexpr uint32_t COMPUTE_DISPATCH_INITIATOR_CS_W32_EN_SHIFT = 15;
+constexpr uint32_t COMPUTE_DISPATCH_INITIATOR_CS_W32_EN_MASK  = 0x1;
+
+[[nodiscard]] constexpr uint8_t ComputeWaveSize(uint32_t dispatch_initiator) {
+	return ((dispatch_initiator >> COMPUTE_DISPATCH_INITIATOR_CS_W32_EN_SHIFT) &
+	        COMPUTE_DISPATCH_INITIATOR_CS_W32_EN_MASK) != 0u
+	           ? 32u
+	           : 64u;
+}
 
 constexpr uint32_t COMPUTE_PGM_RSRC2                      = 0x213;
 constexpr uint32_t COMPUTE_PGM_RSRC2_SCRATCH_EN_SHIFT     = 0;

src/graphics/host_gpu/renderer/renderCompute.cpp

@@ -115,7 +115,7 @@ bool ResolveComputeImageClear(const ShaderComputeInputInfo& input, uint32_t grou
 	    group_z == 1 && input.dispatch_threads_num[0] == group_x &&
 	    input.dispatch_threads_num[1] == 1 && input.dispatch_threads_num[2] == 1 &&
 	    input.group_id[0] && !input.group_id[1] && !input.group_id[2] &&
-	    input.thread_ids_num == 1 && input.wave_size == 32 && !input.tg_size_en && mode == 0x61u &&
+	    input.thread_ids_num == 1 && input.wave_size == 64 && !input.tg_size_en && mode == 0x61u &&
 	    group_x % input.threads_num[0] == 0 && descriptor.NumRecords() == group_x;
 	const auto size = BufferDescriptorSize(descriptor);
 	if (!full_dispatch || size == 0) {

tests/ShaderRecompilerComputeTests.cpp

@@ -18780,14 +18780,14 @@ void CheckPs5GameExampleImageClearRuntimeShape() {
   compute.dispatch_threads_num[2] = 1;
   compute.group_id[0] = true;
   compute.dispatch_thread_dimensions = true;
-  compute.wave_size = 32;
+  compute.wave_size = 64;
   compute.thread_ids_num = 1;
   compute.workgroup_register = 8;
 
   const auto Compile = [&](const char *stage, const std::vector<u32> &code) {
     ShaderRecompiler::CompileOptions options;
     options.stage = ShaderType::Compute;
-    options.wave_size = 32;
+    options.wave_size = 64;
     options.user_data_base = 0;
     options.user_data_count = static_cast<u32>(user_data.size());
     options.user_data = user_data.data();

tests/shaderCfgTests.cpp

@@ -8137,23 +8137,11 @@ void TestNewShaderRecompilerU64PairLowering() {
         "dispatcher did not use one canonical U64 vector spill slot");
 }
 
-void TestComputeShaderInputWaveSize() {
+void TestComputeDispatchWaveSize() {
-  const auto decode_wave_size = [](uint32_t rsrc1) {
+  Check(Pm4::ComputeWaveSize(0x00000041u) == 64u,
-    const bool wave32 = (((rsrc1 >> Pm4::COMPUTE_PGM_RSRC1_W32_EN_SHIFT) &
+        "dispatch without CS_W32_EN did not select wave64");
-                          Pm4::COMPUTE_PGM_RSRC1_W32_EN_MASK) != 0u);
+  Check(Pm4::ComputeWaveSize(0x00008041u) == 32u,
-    return wave32 ? 32u : 64u;
+        "dispatch with CS_W32_EN did not select wave32");
-  };
-
-  constexpr uint32_t observed_wave32_rsrc1_a = 0x402c0146u;
-  constexpr uint32_t observed_wave32_rsrc1_b = 0x402c00c1u;
-  Check(decode_wave_size(observed_wave32_rsrc1_a) == 32u,
-        "COMPUTE_PGM_RSRC1 W32_EN bit did not match observed PS5 value A");
-  Check(decode_wave_size(observed_wave32_rsrc1_b) == 32u,
-        "COMPUTE_PGM_RSRC1 W32_EN bit did not match observed PS5 value B");
-
-  constexpr uint32_t w32_en_bit = 1u << Pm4::COMPUTE_PGM_RSRC1_W32_EN_SHIFT;
-  Check(decode_wave_size(observed_wave32_rsrc1_a & ~w32_en_bit) == 64u,
-        "cleared COMPUTE_PGM_RSRC1 W32_EN bit did not decode as wave64");
 }
 
 void TestNewShaderRecompilerBufferLoadsGuardedByExec() {
@@ -11581,7 +11569,7 @@ int main() {
   TestNewShaderRecompilerCfgIrreducibleDispatcher();
   TestNewShaderRecompilerDispatcherSpillsU32x3();
   TestNewShaderRecompilerU64PairLowering();
-  TestComputeShaderInputWaveSize();
+  TestComputeDispatchWaveSize();
   TestNewShaderRecompilerBufferLoadsGuardedByExec();
   TestNewShaderRecompilerBufferAtomicsGuardedByBounds();
   TestNewShaderRecompilerPixelImageSampleLodSelection();