From 9c00ce93661566c3ed25967fe0f0e9197e9d48d2 Mon Sep 17 00:00:00 2001
From: goldislead <69987043+goldislead@users.noreply.github.com>
Date: Fri, 3 Apr 2026 22:17:43 -0700
Subject: [PATCH] [GPU] EVENT_WRITE_ZPD batched sample accumulation cvar

---
 src/xenia/gpu/gpu_flags.cc                      |  5 +++++
 src/xenia/gpu/gpu_flags.h                       |  2 ++
 src/xenia/gpu/pm4_command_processor_implement.h | 15 ++++++++++++++-
 3 files changed, 21 insertions(+), 1 deletion(-)

diff --git a/src/xenia/gpu/gpu_flags.cc b/src/xenia/gpu/gpu_flags.cc
index 8f99e4e06..9ed9335fa 100644
--- a/src/xenia/gpu/gpu_flags.cc
+++ b/src/xenia/gpu/gpu_flags.cc
@@ -59,6 +59,11 @@ DEFINE_bool(
     "when MSAA is used with fullscreen passes.",
     "GPU");
 
+DEFINE_bool(query_occlusion_batched, false,
+            "Increment the sample count total by a fixed amount on every "
+            "EVENT_WRITE_ZPD. This provides an approximate batched occlusion "
+            "query implementation for many titles.",
+            "GPU");
 DEFINE_int32(query_occlusion_sample_lower_threshold, 80,
              "If set to -1 no sample counts are written, games may hang. Else, "
              "the sample count of every tile will be incremented on every "
diff --git a/src/xenia/gpu/gpu_flags.h b/src/xenia/gpu/gpu_flags.h
index 5dbc34f09..1778371eb 100644
--- a/src/xenia/gpu/gpu_flags.h
+++ b/src/xenia/gpu/gpu_flags.h
@@ -26,6 +26,8 @@ DECLARE_bool(non_seamless_cube_map);
 
 DECLARE_bool(half_pixel_offset);
 
+DECLARE_bool(query_occlusion_batched);
+
 DECLARE_int32(query_occlusion_sample_lower_threshold);
 
 DECLARE_int32(query_occlusion_sample_upper_threshold);
diff --git a/src/xenia/gpu/pm4_command_processor_implement.h b/src/xenia/gpu/pm4_command_processor_implement.h
index 0072ecc7e..1302e13ab 100644
--- a/src/xenia/gpu/pm4_command_processor_implement.h
+++ b/src/xenia/gpu/pm4_command_processor_implement.h
@@ -955,6 +955,7 @@ bool COMMAND_PROCESSOR::ExecutePacketType3_EVENT_WRITE_EXT(
 }
 
 static uint32_t samples = cvars::query_occlusion_sample_upper_threshold;
+static uint32_t batched_samples = 0;
 
 XE_NOINLINE
 bool COMMAND_PROCESSOR::ExecutePacketType3_EVENT_WRITE_ZPD(
@@ -982,7 +983,19 @@ bool COMMAND_PROCESSOR::ExecutePacketType3_EVENT_WRITE_ZPD(
   bool is_end_via_z_fail = pSampleCounts->ZFail_A == kQueryFinished &&
                            pSampleCounts->ZFail_B == kQueryFinished;
   std::memset(pSampleCounts, 0, sizeof(xe_gpu_depth_sample_counts));
-  if (is_end_via_z_pass || is_end_via_z_fail) {
+
+  // Titles that use QueryBatch (4D5309B1, 4E4D0801) don't use an END result for
+  // every query. Each ISSUE snapshots the sample count into a new slot and
+  // recovers the total by looking at differences between slots.
+  // END detection isn't sufficient here.
+  if (cvars::query_occlusion_batched) {
+    // Mimic batched behavior by reporting a running count and writing it on
+    // every event.
+    const uint32_t step = std::max(uint32_t(1), samples);
+    batched_samples = std::min(batched_samples, UINT32_MAX - step) + step;
+    pSampleCounts->ZPass_A = batched_samples;
+    pSampleCounts->Total_A = batched_samples;
+  } else if (is_end_via_z_pass || is_end_via_z_fail) {
     pSampleCounts->ZPass_A = samples;
     pSampleCounts->Total_A = samples;
   }