fix fetch shader V_SAD_U32, video-out buffer overlap

src/graphics/guest_gpu/graphicsRun.cpp

@@ -1142,23 +1142,6 @@ void CommandProcessor::DrawIndex(uint32_t index_count, const void* index_addr, u
 		     "\n",
 		     vertex_offset_add, first_instance);
 	}
-	const auto frame_num = GraphicsRunGetFrameNum();
-	if (frame_num >= 480) {
-		static std::atomic<uint32_t> log_count {0};
-		if (log_count.fetch_add(1, std::memory_order_relaxed) < 512) {
-			const auto& oa = m_ucfg.GetGdsOaCounter(m_ucfg.GetGdsOaState().GetIndex());
-			LOGF("QueuePoint DrawIndex: frame=%u submit=%" PRIu64
-			     " queue=%d index_count=%u instances=%u prim=%u "
-			     "es=0x%016" PRIx64 " ps=0x%016" PRIx64
-			     " oa_index=%u oa_enabled=%s oa_addr=0x%04" PRIx32 " oa_space=0x%08" PRIx32 "\n",
-			     frame_num, m_submit_id, m_scheduler.Queue(), index_count, instance_count,
-			     m_ucfg.GetPrimType(), m_sh_ctx.GetVs().es_regs.data_addr,
-			     m_sh_ctx.GetPs().ps_regs.data_addr, m_ucfg.GetGdsOaState().GetIndex(),
-			     oa.IsCounterEnabled() ? "true" : "false", oa.GetAddressBytes(),
-			     oa.GetSpaceAvailable());
-		}
-	}
-
 	RenderDrawIndex(m_submit_id, CurrentBuffer(), &m_ctx, &m_ucfg, &m_sh_ctx, m_index_type_and_size,
 	                index_count, index_addr, flags, type, instance_count,
 	                render_target_slice_offset, vertex_offset_add, first_instance);
@@ -1458,22 +1441,6 @@ void CommandProcessor::DrawIndexAuto(uint32_t index_count, uint32_t flags,
 	Common::LockGuard lock(m_mutex);
 
 	CheckBuffer();
-	const auto frame_num = GraphicsRunGetFrameNum();
-	if (frame_num >= 480) {
-		static std::atomic<uint32_t> log_count {0};
-		if (log_count.fetch_add(1, std::memory_order_relaxed) < 512) {
-			const auto& oa = m_ucfg.GetGdsOaCounter(m_ucfg.GetGdsOaState().GetIndex());
-			LOGF("QueuePoint DrawIndexAuto: frame=%u submit=%" PRIu64
-			     " queue=%d index_count=%u instances=%u prim=%u "
-			     "first_vertex=%u first_instance=%u es=0x%016" PRIx64 " ps=0x%016" PRIx64
-			     " oa_index=%u oa_enabled=%s oa_addr=0x%04" PRIx32 " oa_space=0x%08" PRIx32 "\n",
-			     frame_num, m_submit_id, m_scheduler.Queue(), index_count, instance_count,
-			     m_ucfg.GetPrimType(), first_vertex, first_instance,
-			     m_sh_ctx.GetVs().es_regs.data_addr, m_sh_ctx.GetPs().ps_regs.data_addr,
-			     m_ucfg.GetGdsOaState().GetIndex(), oa.IsCounterEnabled() ? "true" : "false",
-			     oa.GetAddressBytes(), oa.GetSpaceAvailable());
-		}
-	}
 
 	RenderDrawIndexAuto(m_submit_id, CurrentBuffer(), &m_ctx, &m_ucfg, &m_sh_ctx, index_count,
 	                    flags, render_target_slice_offset, instance_count, first_vertex,

src/graphics/host_gpu/renderer/textureCache.cpp

@@ -2578,11 +2578,13 @@ void TextureCache::RefreshVideoOut(VideoOutVulkanImage* image, bool render_targe
 	if (cached.gpu_modified) {
 		return;
 	}
-	const auto& info         = cached.video_out;
+	const auto& info           = cached.video_out;
-	const bool  image_dirty  = m_memory_tracker.IsRegionCpuModified(info.address, info.size);
+	const bool  image_dirty    = m_memory_tracker.IsRegionCpuModified(info.address, info.size);
-	const bool  buffer_dirty = cached.buffer_modified ||
+	const bool  buffer_overlap = m_buffer_cache.HasPageOverlap(info.address, info.size);
-	                           m_buffer_cache.IsRegionCpuModified(info.address, info.size) ||
+	const bool  buffer_dirty =
-	                           m_buffer_cache.IsRegionGpuModified(info.address, info.size);
+	    cached.buffer_modified ||
+	    (buffer_overlap && (m_buffer_cache.IsRegionCpuModified(info.address, info.size) ||
+	                        m_buffer_cache.IsRegionGpuModified(info.address, info.size)));
 	if (!image_dirty && !buffer_dirty) {
 		if (info.compression == VideoOutCompression::Uncompressed ||
 		    CanUseVideoOutNativeWithoutUpload(info.compression, render_target, false, false)) {

src/graphics/shader/recompiler/ShaderRecompiler.cpp

@@ -336,9 +336,21 @@ EmbeddedFetchData DetectEmbeddedVertexFetch(const Decoder::Program&      decoded
 	}
 
 	for (const auto& inst: decoded.instructions) {
-		if (data.loads.empty() && inst.opcode == Decoder::Opcode::VAddI32 &&
+		// Fetch shaders accumulate the draw's vertex offset in v0. The PS5 NGG ABI
-		    IsDecodedVgpr(inst.dst) && inst.dst.reg == 0 && IsDecodedSgpr(inst.src0) &&
+		// seeds S_NGG_VERTEX_INDEX in v5 and applies the same offset there before fetching.
-		    IsDecodedVgpr(inst.src1) && inst.src1.reg == 0) {
+		const bool vertex_index_accumulator =
+		    IsDecodedVgpr(inst.dst) &&
+		    (inst.dst.reg == 0 || (user_data_base == 8 && inst.dst.reg == 5));
+		uint32_t sad_zero = 0;
+		const bool vertex_offset_add =
+		    vertex_index_accumulator && IsDecodedSgpr(inst.src0) &&
+		    ((inst.opcode == Decoder::Opcode::VAddI32 && IsDecodedVgpr(inst.src1) &&
+		      inst.src1.reg == inst.dst.reg) ||
+		     (user_data_base == 8 && inst.dst.reg == 5 &&
+		      inst.opcode == Decoder::Opcode::VSadU32 && IsDecodedVgpr(inst.src2) &&
+		      inst.src2.reg == inst.dst.reg &&
+		      TryDecodedOperandConstant(sgprs, inst.src1, sad_zero) && sad_zero == 0));
+		if (data.loads.empty() && vertex_offset_add) {
 			const auto reg = DecodedSgprReg(inst.src0);
 			if (reg >= user_data_base && reg - user_data_base < user_data_count) {
 				if (offset_candidate >= 0 && offset_candidate != static_cast<int32_t>(reg)) {

tests/ShaderRecompilerComputeTests.cpp

@@ -9009,14 +9009,29 @@ void CheckPs5GameExampleImageClearRuntimeShape() {
 
 void CheckEmbeddedFetchVertexOffset() {
   const auto MakeFetch = [](std::initializer_list<std::pair<u32, u32>> adds,
-                            std::optional<std::pair<u32, u32>> late_add = {}) {
+                            std::optional<std::pair<u32, u32>> late_add = {},
+                            u32 accumulator_vgpr = 0, bool ngg_sad = false) {
     std::vector<u32> code;
     code.push_back(EncodeSMovB32(0, InlineU32(0)));
     code.push_back(EncodeSmem0(0x02u, 20, 4));
     code.push_back(EncodeSmem1(0));
+    const auto AppendOffsets = [&]() {
+      for (const auto [sgpr, index_vgpr] : adds) {
+        if (ngg_sad) {
+          AppendVop3(&code, 0x15du, accumulator_vgpr, sgpr, InlineU32(0),
+                     Vgpr(index_vgpr));
+        } else {
+          AppendVop3B(&code, 0x30fu, accumulator_vgpr, 0, sgpr,
+                      Vgpr(index_vgpr));
+        }
+      }
+    };
+    if (ngg_sad) {
+      AppendOffsets();
+    }
     code.push_back(EncodeVop2(0x01u, 0, Vgpr(8), 5));
-    for (const auto [sgpr, index_vgpr] : adds) {
+    if (!ngg_sad) {
-      AppendVop3B(&code, 0x30fu, 0, 0, sgpr, Vgpr(index_vgpr));
+      AppendOffsets();
     }
     code.push_back(EncodeMubuf0(0x03u, 0, true));
     code.push_back(EncodeMubuf1(9, 5, 0));
@@ -9076,6 +9091,18 @@ void CheckEmbeddedFetchVertexOffset() {
           Resolve(valid, 5) == 5,
       "canonical fetch offset or register index-offset precedence is wrong");
 
+  const auto ngg_code = MakeFetch({{18, 5}}, {}, 5, true);
+  Require("EmbeddedFetchNggVertexOffset", "encoding",
+          ngg_code[3] == 0xd55d0005u && ngg_code[4] == 0x04150012u,
+          "test does not encode the PS5 V_SAD_U32 vertex-offset prolog");
+  const auto ngg =
+      Compile("EmbeddedFetchNggVertexOffset", ngg_code, 8);
+  Require("EmbeddedFetchNggVertexOffset", "parse",
+          ngg.program.info.vertex_offset_sgpr == 18 && Resolve(ngg, 0) == 8 &&
+              Resolve(ngg, 5) == 5,
+          "PS5 NGG vertex-index offset or register index-offset precedence is "
+          "wrong");
+
   const auto pointer = 0x5b7c5100u;
   const auto late = Compile("EmbeddedFetchLateOffset",
                             MakeFetch({}, std::pair<u32, u32>{18, 0}), pointer);