From b244864a61d5a5b40eefaa8f9b7b4d46361ae056 Mon Sep 17 00:00:00 2001
From: notaz <notasas@gmail.com>
Date: Fri, 20 Dec 2024 01:04:28 +0200
Subject: [PATCH] gpu: some timing improvements

libretro/pcsx_rearmed#857
notaz/pcsx_rearmed#217
---
 frontend/plugin_lib.h |  2 +-
 libpcsxcore/gpu.c     |  9 ++++++---
 libpcsxcore/gpu.h     |  2 +-
 plugins/gpulib/gpu.c  | 11 ++++++-----
 plugins/gpulib/gpu.h  |  2 +-
 5 files changed, 15 insertions(+), 11 deletions(-)

diff --git a/frontend/plugin_lib.h b/frontend/plugin_lib.h
index 420f89f9..31acae97 100644
--- a/frontend/plugin_lib.h
+++ b/frontend/plugin_lib.h
@@ -63,7 +63,7 @@ struct rearmed_cbs {
 	void  (*pl_vout_set_raw_vram)(void *vram);
 	void  (*pl_set_gpu_caps)(int caps);
 	// emulation related
-	void  (*gpu_state_change)(int what);
+	void  (*gpu_state_change)(int what, int cycles);
 	// some stats, for display by some plugins
 	int flips_per_sec, cpu_usage;
 	float vsps_cur; // currect vsync/s
diff --git a/libpcsxcore/gpu.c b/libpcsxcore/gpu.c
index 425537d4..b86cfe65 100644
--- a/libpcsxcore/gpu.c
+++ b/libpcsxcore/gpu.c
@@ -13,7 +13,7 @@
 #include "gpu.h"
 #include "psxdma.h"
 
-void gpu_state_change(int what)
+void gpu_state_change(int what, int cycles)
 {
 	enum psx_gpu_state state = what;
 	switch (state)
@@ -22,10 +22,13 @@ void gpu_state_change(int what)
 		psxRegs.gpuIdleAfter = psxRegs.cycle + PSXCLK / 50;
 		break;
 	case PGS_VRAM_TRANSFER_END:
-		psxRegs.gpuIdleAfter = psxRegs.cycle;
+		psxRegs.gpuIdleAfter = psxRegs.cycle - 1;
 		break;
 	case PGS_PRIMITIVE_START:
-		psxRegs.gpuIdleAfter = psxRegs.cycle + 200;
+		// limit because gpulib delays things with it's buffering...
+		if (cycles > 512)
+			cycles = 512;
+		psxRegs.gpuIdleAfter = psxRegs.cycle + cycles - 1;
 		break;
 	}
 }
diff --git a/libpcsxcore/gpu.h b/libpcsxcore/gpu.h
index ec5019c9..69474b45 100644
--- a/libpcsxcore/gpu.h
+++ b/libpcsxcore/gpu.h
@@ -49,6 +49,6 @@ enum psx_gpu_state {
   PGS_PRIMITIVE_START, // for non-dma only
 };
 
-void gpu_state_change(int what);
+void gpu_state_change(int what, int cycles);
 
 #endif /* __GPU_H__ */
diff --git a/plugins/gpulib/gpu.c b/plugins/gpulib/gpu.c
index 677e18b5..ac9b86a3 100644
--- a/plugins/gpulib/gpu.c
+++ b/plugins/gpulib/gpu.c
@@ -514,7 +514,7 @@ static noinline void start_vram_transfer(struct psx_gpu *gpu, uint32_t pos_word,
   log_io(gpu, "start_vram_transfer %c (%d, %d) %dx%d\n", is_read ? 'r' : 'w',
     gpu->dma.x, gpu->dma.y, gpu->dma.w, gpu->dma.h);
   if (gpu->gpu_state_change)
-    gpu->gpu_state_change(PGS_VRAM_TRANSFER_START);
+    gpu->gpu_state_change(PGS_VRAM_TRANSFER_START, 0);
 }
 
 static void finish_vram_transfer(struct psx_gpu *gpu, int is_read)
@@ -540,7 +540,7 @@ static void finish_vram_transfer(struct psx_gpu *gpu, int is_read)
                            gpu->dma_start.w, gpu->dma_start.h, 0);
   }
   if (gpu->gpu_state_change)
-    gpu->gpu_state_change(PGS_VRAM_TRANSFER_END);
+    gpu->gpu_state_change(PGS_VRAM_TRANSFER_END, 0);
 }
 
 static void do_vram_copy(struct psx_gpu *gpu, const uint32_t *params, int *cpu_cycles)
@@ -749,14 +749,15 @@ static noinline int do_cmd_buffer(struct psx_gpu *gpu, uint32_t *data, int count
 
 static noinline void flush_cmd_buffer(struct psx_gpu *gpu)
 {
+  int cycles_last = 0;
   int dummy = 0, left;
-  left = do_cmd_buffer(gpu, gpu->cmd_buffer, gpu->cmd_len, &dummy, &dummy);
+  left = do_cmd_buffer(gpu, gpu->cmd_buffer, gpu->cmd_len, &dummy, &cycles_last);
   if (left > 0)
     memmove(gpu->cmd_buffer, gpu->cmd_buffer + gpu->cmd_len - left, left * 4);
   if (left != gpu->cmd_len) {
-    if (!gpu->dma.h && gpu->gpu_state_change)
-      gpu->gpu_state_change(PGS_PRIMITIVE_START);
     gpu->cmd_len = left;
+    if (!gpu->dma.h && gpu->gpu_state_change)
+      gpu->gpu_state_change(PGS_PRIMITIVE_START, cycles_last);
   }
 }
 
diff --git a/plugins/gpulib/gpu.h b/plugins/gpulib/gpu.h
index 7a84a254..893a7512 100644
--- a/plugins/gpulib/gpu.h
+++ b/plugins/gpulib/gpu.h
@@ -117,7 +117,7 @@ struct psx_gpu {
     (int *x, int *y, int *w, int *h, int *vram_h);
   void *(*mmap)(unsigned int size);
   void  (*munmap)(void *ptr, unsigned int size);
-  void  (*gpu_state_change)(int what); // psx_gpu_state
+  void  (*gpu_state_change)(int what, int cycles); // psx_gpu_state
 };
 
 extern struct psx_gpu gpu;
-- 
2.39.5