renderer: skip inactive pixel shaders

src/graphics/host_gpu/renderer/renderDraw.cpp

@@ -575,12 +575,17 @@ RenderState RenderExecutor::AcquireRenderTargets(CommandBuffer& buffer, RenderCo
 	return state;
 }
 
-static bool DrawHasActivePixelShader(const CommandBuffer& buffer) {
+static uint32_t DrawColorOutputMask(const HW::Context& ctx) {
-	const auto& ctx              = buffer.GetRegisters();
+	const auto& sh_regs     = ctx.GetShaderRegisters();
-	const auto& sh_regs          = ctx.GetShaderRegisters();
+	const auto  write_mask  = ctx.GetRenderTargetMask() & sh_regs.m_cbShaderMask;
-	const bool  has_color_output = (ctx.GetRenderTargetMask() & sh_regs.m_cbShaderMask) != 0;
+	uint32_t    output_mask = 0;
-	return buffer.GetShaders().GetPs().ps_regs.data_addr != 0 &&
+	for (uint32_t slot = 0; slot < RENDER_COLOR_ATTACHMENTS_MAX; slot++) {
-	       (has_color_output || PixelShaderHasDepthOrCoverageSideEffects(sh_regs));
+		if (sh_regs.target_output_mode[slot] != 0 &&
+		    render_target_mask_slot(write_mask, slot) != 0) {
+			output_mask |= 1u << slot;
+		}
+	}
+	return output_mask;
 }
 
 enum class CbColorMode : uint8_t {
@@ -847,7 +852,7 @@ static bool ResolvePrimitiveRestart(const CommandBuffer& buffer,
 }
 
 static void RefreshShaders(CommandBuffer& buffer, const DrawCallInfo& draw,
-                           DrawRenderState& state) {
+                           uint32_t color_output_mask, DrawRenderState& state) {
 	auto& ctx    = buffer.GetRegisters();
 	auto& sh_ctx = buffer.GetShaders();
 
@@ -861,7 +866,7 @@ static void RefreshShaders(CommandBuffer& buffer, const DrawCallInfo& draw,
 	    target_export_mapping {};
 	for (uint32_t slot = 0; slot < RENDER_COLOR_ATTACHMENTS_MAX; slot++) {
 		const auto& rt = ctx.GetRenderTarget(slot);
-		if (rt.base.addr != 0 && render_target_mask_slot(ctx.GetRenderTargetMask(), slot) != 0) {
+		if ((color_output_mask & (1u << slot)) != 0 && rt.base.addr != 0) {
 			target_export_mapping[slot] =
 			    TextureGetRenderTargetFormat(rt.info.format, rt.info.channel_type,
 			                                 rt.info.channel_order)
@@ -880,8 +885,12 @@ static void RefreshShaders(CommandBuffer& buffer, const DrawCallInfo& draw,
 bool RenderExecutor::PrepareDrawRenderState(CommandBuffer& buffer, const DrawCallInfo& draw,
                                             uint32_t            render_target_slice_offset,
 	                                        DrawRenderState& state) {
-	state.ps_active = DrawHasActivePixelShader(buffer);
+	const auto& shader_regs       = buffer.GetRegisters().GetShaderRegisters();
-	RefreshShaders(buffer, draw, state);
+	const auto  color_output_mask = DrawColorOutputMask(buffer.GetRegisters());
+	state.ps_active = buffer.GetShaders().GetPs().ps_regs.data_addr != 0 &&
+	                  (color_output_mask != 0 ||
+	                   PixelShaderHasDepthOrCoverageSideEffects(shader_regs));
+	RefreshShaders(buffer, draw, color_output_mask, state);
 	uint32_t mrt_mask = 0;
 	if (state.ps_active) {
 		for (const auto& output: state.ps_input_info.stage.program->info.outputs) {
@@ -890,6 +899,7 @@ bool RenderExecutor::PrepareDrawRenderState(CommandBuffer& buffer, const DrawCal
 			}
 		}
 	}
+	mrt_mask &= color_output_mask;
 	if (draw.IsIndexed()) {
 		LogDrawPhase(draw.Name(), "ResolveRenderColorTarget");
 	}

tests/ShaderRecompilerComputeTests.cpp

@@ -14373,6 +14373,40 @@ public:
                 "a stale color target clipped stencil coverage or the NULL-export shader was skipped");
       }
 
+      // A depth-only draw retains the previous PS address and write masks, but
+      // SPI_SHADER_COL_FORMAT disables its exports. Compiling this stale PS
+      // with the new zero-input metadata would fail before the depth draw.
+      static const auto stale_pixel = [] {
+        std::vector<u32> code{EncodeVintrp(0, 0, 0, 0, 0),
+                              EncodeVintrp(1, 0, 0, 0, 1),
+                              EncodeExp0(0, 0xf), EncodeExp1(0, 0, 0, 0)};
+        AppendEnd(&code);
+        return code;
+      }();
+      const auto stale_address = reinterpret_cast<uint64_t>(stale_pixel.data());
+      ShaderMapUserData(stale_address,
+          {.type = Prospero::ShaderBinaryType::kPs,
+           .code_size_bytes = static_cast<u32>(stale_pixel.size() * sizeof(u32))});
+      shaders.SetPsShaderBase(stale_address);
+      registers.SetPsInControl(0);
+      registers.SetDepthShaderControl({});
+      registers.SetDepthControl({.z_enable = true, .z_write_enable = true,
+                                .zfunc = static_cast<uint8_t>(vk::CompareOp::eAlways)});
+      for (const auto slot : {0u, 3u}) {
+        registers.SetTargetOutputMode(slot, 0);
+      }
+      RenderExecutorTestAccess::DrawAuto(
+          executor, scheduler.Current(), {.vertex_count = 3, .instance_count = 1});
+      const auto depth_only_pixels = ReadCachedTexel(
+          name, context, depth.image_id, {}, {extent, extent, 1});
+      Require(name, "disabled stale PS preserves fixed-function depth",
+              std::ranges::all_of(depth_only_pixels, [](u32 v) { return v == 0; }),
+              "disabled color exports invoked the stale PS or clipped fixed-function depth");
+      registers.SetPsInControl(0x8000);
+      for (const auto slot : {0u, 3u}) {
+        registers.SetTargetOutputMode(slot, 4);
+      }
+
       // With MRT0 still bound, an MRT3-only export must retain location3 in
       // rendering attachments, pipeline formats, blend masks and dynamic write enables.
       static const auto sparse_pixel = [] {