From 8e42e99788faf3b8443e479b3ab7e5f6201358d6 Mon Sep 17 00:00:00 2001
From: nmzik <Nmzik@mail.ru>
Date: Tue, 29 Sep 2026 03:00:57 +0400
Subject: [PATCH] graphics: batch release labels without redundant submissions

---
 .../command_processor/pm4Handlers.cpp         | 10 ++++----
 tests/ShaderRecompilerComputeTests.cpp        | 24 ++++++++++++-------
 2 files changed, 19 insertions(+), 15 deletions(-)

diff --git a/src/graphics/guest_gpu/command_processor/pm4Handlers.cpp b/src/graphics/guest_gpu/command_processor/pm4Handlers.cpp
index c95983b69..8df48e30a 100644
--- a/src/graphics/guest_gpu/command_processor/pm4Handlers.cpp
+++ b/src/graphics/guest_gpu/command_processor/pm4Handlers.cpp
@@ -2266,7 +2266,6 @@ KYTY_CP_OP_PARSER(CpOpReleaseMem) {
 	const bool gl2_writeback = ((gcr_cntl & GcrGl2Writeback) != 0);
 
 	auto trigger_interrupt = [&]() {
-		bool queued = false;
 		switch (interrupt_selector) {
 			case 0x00:
 			case 0x03: break;
@@ -2274,13 +2273,10 @@ KYTY_CP_OP_PARSER(CpOpReleaseMem) {
 			case 0x02:
 			case 0x04:
 				cp.TriggerEopEventAtEndOfPipe(interrupt_context_id);
-				queued = true;
+				cp.BufferFlush();
 				break;
 			default: EXIT("unknown release_mem interrupt selector\n");
 		}
-		if (queued) {
-			cp.BufferFlush();
-		}
 	};
 
 	if (data_sel == 0 || interrupt_selector == 4) {
@@ -2318,7 +2314,9 @@ KYTY_CP_OP_PARSER(CpOpReleaseMem) {
 		cp.WriteAtEndOfPipe32(cache_policy, event_write_dest, eop_event_type, cache_action,
 		                      event_index, event_source, dst_gpu_addr, static_cast<uint32_t>(value),
 		                      interrupt_selector, interrupt_context_id);
-		cp.BufferFlush();
+		if (interrupt_selector == 0x01 || interrupt_selector == 0x02) {
+			cp.BufferFlush();
+		}
 
 		return 7;
 	}
diff --git a/tests/ShaderRecompilerComputeTests.cpp b/tests/ShaderRecompilerComputeTests.cpp
index e070b2d7c..cf6c445ba 100644
--- a/tests/ShaderRecompilerComputeTests.cpp
+++ b/tests/ShaderRecompilerComputeTests.cpp
@@ -2565,13 +2565,20 @@ public:
                     release_label <= Sync::ReadReferenceClock(),
                 "clock write with writeback and interrupt lost its data");
 
-        auto immediate = make_release_mem(1, 0, &release_label, 0x11223344u);
-        Pm4Execution immediate_execution;
-        const auto immediate_tick = gpu_scheduler.CurrentTick();
-        const auto immediate_result =
-            processor->Process(immediate_execution, immediate);
-        const bool immediate_split_once =
-            gpu_scheduler.CurrentTick() == immediate_tick + 1;
+        for (const auto interrupt : {0u, 3u, 1u, 2u}) {
+          release_label = 0;
+          auto immediate = make_release_mem(1, interrupt, &release_label, 0x11223344u);
+          Pm4Execution immediate_execution;
+          const auto immediate_tick = gpu_scheduler.CurrentTick();
+          const auto immediate_result =
+              processor->Process(immediate_execution, immediate);
+          Require("GpuCommandLane", "32-bit release boundary",
+                  immediate_result == Pm4ProcessResult::Complete &&
+                      gpu_scheduler.CurrentTick() ==
+                          immediate_tick + (interrupt == 1 || interrupt == 2) &&
+                      release_label == (interrupt == 1 ? 0 : 0x11223344u),
+                  "label-only release submitted work or interrupt release lost its boundary");
+        }
 
         auto gds = make_release_mem(5, 0, &gds_label, 1ull << 16u);
         Pm4Execution gds_execution;
@@ -2597,8 +2604,7 @@ public:
             gpu_scheduler.CurrentTick() == gds_interrupt_tick + 1;
 
         release_mem_submission_counts =
-            immediate_result == Pm4ProcessResult::Complete &&
-            immediate_split_once && gds_result == Pm4ProcessResult::Complete &&
+            gds_result == Pm4ProcessResult::Complete &&
             gds_waited_once && interrupt_result == Pm4ProcessResult::Complete &&
             interrupt_split_once &&
             gds_interrupt_result == Pm4ProcessResult::Complete &&