pvr: auto frame skip to replace current and previous synchronous rendering

Restore previous synchronous rendering behaviour (normal auto frame skip)
No frame skipping when disabled
better cpu speed measure over 4 vblanks

core/hw/pvr/spg.cpp

@@ -1,3 +1,4 @@
+#include <array>
 #include "spg.h"
 #include "hw/holly/holly_intc.h"
 #include "hw/holly/sb.h"
@@ -23,6 +24,11 @@ static u32 Frame_Cycles;
 int render_end_schid;
 int vblank_schid;
 
+static std::array<double, 4> real_times;
+static std::array<u64, 4> cpu_cycles;
+static u32 cpu_time_idx;
+bool SH4FastEnough;
+
 void CalculateSync()
 {
 	u32 pixel_clock = PIXEL_CLOCK / (FB_R_CTRL.vclk_div ? 1 : 2);
@@ -127,13 +133,28 @@ int spg_line_sched(int tag, int cycl, int jit)
 			else
 				SPG_STATUS.fieldnum=0;
 
-			vblk_cnt++;
 			rend_vblank();
 
+			double now = os_GetSeconds() * 1000000.0;
+			cpu_time_idx = (cpu_time_idx + 1) % cpu_cycles.size();
+			if (cpu_cycles[cpu_time_idx] != 0)
+			{
+				u32 cycle_span = (u32)(sh4_sched_now64() - cpu_cycles[cpu_time_idx]);
+				double time_span = now - real_times[cpu_time_idx];
+				double cpu_speed = ((double)cycle_span / time_span) / (SH4_MAIN_CLOCK / 100000000);
+				SH4FastEnough = cpu_speed >= 85.0;
+			}
+			else
+				SH4FastEnough = false;
+			cpu_cycles[cpu_time_idx] = sh4_sched_now64();
+			real_times[cpu_time_idx] = now;
+
 #ifdef TEST_AUTOMATION
 			replay_input();
 #endif
-			
+
+#if !defined(NDEBUG) || defined(DEBUGFAST)
+			vblk_cnt++;
 			if ((os_GetSeconds()-last_fps)>2)
 			{
 				static int Last_FC;
@@ -190,6 +211,7 @@ int spg_line_sched(int tag, int cycl, int jit)
 				fskip=0;
 				last_fps=os_GetSeconds();
 			}
+#endif
 		}
 		if (lightgun_line != 0xffff && lightgun_line == prv_cur_scanline)
 		{
@@ -273,6 +295,11 @@ void spg_Term()
 void spg_Reset(bool hard)
 {
 	CalculateSync();
+
+	SH4FastEnough = false;
+	cpu_time_idx = 0;
+	cpu_cycles.fill(0);
+	real_times.fill(0.0);
 }
 
 void SetREP(TA_context* cntx)

core/hw/pvr/spg.h

@@ -1,6 +1,8 @@
 #pragma once
 #include "ta_ctx.h"
 
+extern bool SH4FastEnough;
+
 bool spg_Init();
 void spg_Term();
 void spg_Reset(bool Manual);

core/hw/pvr/ta_ctx.cpp

@@ -1,8 +1,10 @@
 #include "ta_ctx.h"
+#include "spg.h"
 #include "oslib/oslib.h"
 
 extern u32 fskip;
 extern u32 FrameCount;
+static int RenderCount;
 
 TA_context* ta_ctx;
 tad_context ta_tad;
@@ -73,13 +75,18 @@ bool QueueRender(TA_context* ctx)
 {
 	verify(ctx != 0);
 	
-	if (rqueue && settings.pvr.SynchronousRender)
+	bool skipFrame = false;
-		//wait for a frame if
+	RenderCount++;
-		//  we have another one queue'd
+	if (RenderCount % (settings.pvr.ta_skip + 1) != 0)
-		//  and SynchronousRender is enabled
+		skipFrame = true;
+	else if (rqueue && (settings.pvr.AutoSkipFrame == 0
+				|| (settings.pvr.AutoSkipFrame == 1 && SH4FastEnough)))
+		// The previous render hasn't completed yet so we wait.
+		// If autoskipframe is enabled (normal level), we only do so if the CPU is running
+		// fast enough over the last frames
 		frame_finished.Wait();
 
-	if (rqueue)
+	if (skipFrame || rqueue)
 	{
 		tactx_Recycle(ctx);
 		fskip++;

core/hw/pvr/ta_vtx.cpp

@@ -1432,8 +1432,6 @@ static void getRegionTileClipping(u32& xmin, u32& xmax, u32& ymin, u32& ymax);
 
 FifoSplitter<0> TAFifo0;
 
-int ta_parse_cnt = 0;
-
 //
 // Check if a vertex has huge x,y,z values or negative z
 //
@@ -1581,69 +1579,66 @@ bool ta_parse_vdrc(TA_context* ctx)
 	vd_ctx = ctx;
 	vd_rc = vd_ctx->rend;
 
-	ta_parse_cnt++;
+	TAFifo0.vdec_init();
-	if (ctx->rend.isRTT || 0 == (ta_parse_cnt %  ( settings.pvr.ta_skip + 1)))
+
+	bool empty_context = true;
+	int op_poly_count = 0;
+	int pt_poly_count = 0;
+	int tr_poly_count = 0;
+
+	PolyParam *bgpp = vd_rc.global_param_op.head();
+	if (bgpp->pcw.Texture)
 	{
-		TAFifo0.vdec_init();
+		bgpp->texid = renderer->GetTexture(bgpp->tsp, bgpp->tcw);
-		
+		empty_context = false;
-		bool empty_context = true;
-		int op_poly_count = 0;
-		int pt_poly_count = 0;
-		int tr_poly_count = 0;
-
-		PolyParam *bgpp = vd_rc.global_param_op.head();
-		if (bgpp->pcw.Texture)
-		{
-			bgpp->texid = renderer->GetTexture(bgpp->tsp, bgpp->tcw);
-			empty_context = false;
-		}
-
-		for (u32 pass = 0; pass <= ctx->tad.render_pass_count; pass++)
-		{
-			ctx->MarkRend(pass);
-			vd_rc.proc_start = ctx->rend.proc_start;
-			vd_rc.proc_end = ctx->rend.proc_end;
-
-			Ta_Dma* ta_data=(Ta_Dma*)vd_rc.proc_start;
-			Ta_Dma* ta_data_end=((Ta_Dma*)vd_rc.proc_end)-1;
-
-			do
-			{
-				ta_data =TaCmd(ta_data,ta_data_end);
-			}
-			while(ta_data<=ta_data_end);
-
-			if (ctx->rend.Overrun)
-				break;
-
-			bool empty_pass = vd_rc.global_param_op.used() == (pass == 0 ? 0 : (int)vd_rc.render_passes.LastPtr()->op_count)
-					&& vd_rc.global_param_pt.used() == (pass == 0 ? 0 : (int)vd_rc.render_passes.LastPtr()->pt_count)
-					&& vd_rc.global_param_tr.used() == (pass == 0 ? 0 : (int)vd_rc.render_passes.LastPtr()->tr_count);
-			empty_context = empty_context && empty_pass;
-
-			if (pass == 0 || !empty_pass)
-			{
-				RenderPass *render_pass = vd_rc.render_passes.Append();
-				render_pass->op_count = vd_rc.global_param_op.used();
-				make_index(&vd_rc.global_param_op, op_poly_count,
-						render_pass->op_count, true, &vd_rc);
-				op_poly_count = render_pass->op_count;
-				render_pass->mvo_count = vd_rc.global_param_mvo.used();
-				render_pass->pt_count = vd_rc.global_param_pt.used();
-				make_index(&vd_rc.global_param_pt, pt_poly_count,
-						render_pass->pt_count, true, &vd_rc);
-				pt_poly_count = render_pass->pt_count;
-				render_pass->tr_count = vd_rc.global_param_tr.used();
-				make_index(&vd_rc.global_param_tr, tr_poly_count,
-						render_pass->tr_count, false, &vd_rc);
-				tr_poly_count = render_pass->tr_count;
-				render_pass->mvo_tr_count = vd_rc.global_param_mvo_tr.used();
-				render_pass->autosort = UsingAutoSort(pass);
-				render_pass->z_clear = ClearZBeforePass(pass);
-			}
-		}
-		rv = !empty_context;
 	}
+
+	for (u32 pass = 0; pass <= ctx->tad.render_pass_count; pass++)
+	{
+		ctx->MarkRend(pass);
+		vd_rc.proc_start = ctx->rend.proc_start;
+		vd_rc.proc_end = ctx->rend.proc_end;
+
+		Ta_Dma* ta_data=(Ta_Dma*)vd_rc.proc_start;
+		Ta_Dma* ta_data_end=((Ta_Dma*)vd_rc.proc_end)-1;
+
+		do
+		{
+			ta_data =TaCmd(ta_data,ta_data_end);
+		}
+		while(ta_data<=ta_data_end);
+
+		if (ctx->rend.Overrun)
+			break;
+
+		bool empty_pass = vd_rc.global_param_op.used() == (pass == 0 ? 0 : (int)vd_rc.render_passes.LastPtr()->op_count)
+				&& vd_rc.global_param_pt.used() == (pass == 0 ? 0 : (int)vd_rc.render_passes.LastPtr()->pt_count)
+				&& vd_rc.global_param_tr.used() == (pass == 0 ? 0 : (int)vd_rc.render_passes.LastPtr()->tr_count);
+		empty_context = empty_context && empty_pass;
+
+		if (pass == 0 || !empty_pass)
+		{
+			RenderPass *render_pass = vd_rc.render_passes.Append();
+			render_pass->op_count = vd_rc.global_param_op.used();
+			make_index(&vd_rc.global_param_op, op_poly_count,
+					render_pass->op_count, true, &vd_rc);
+			op_poly_count = render_pass->op_count;
+			render_pass->mvo_count = vd_rc.global_param_mvo.used();
+			render_pass->pt_count = vd_rc.global_param_pt.used();
+			make_index(&vd_rc.global_param_pt, pt_poly_count,
+					render_pass->pt_count, true, &vd_rc);
+			pt_poly_count = render_pass->pt_count;
+			render_pass->tr_count = vd_rc.global_param_tr.used();
+			make_index(&vd_rc.global_param_tr, tr_poly_count,
+					render_pass->tr_count, false, &vd_rc);
+			tr_poly_count = render_pass->tr_count;
+			render_pass->mvo_tr_count = vd_rc.global_param_mvo_tr.used();
+			render_pass->autosort = UsingAutoSort(pass);
+			render_pass->z_clear = ClearZBeforePass(pass);
+		}
+	}
+	rv = !empty_context;
+
 	bool overrun = ctx->rend.Overrun;
 	if (overrun)
 		WARN_LOG(PVR, "ERROR: TA context overrun");

core/nullDC.cpp

@@ -760,7 +760,7 @@ void InitSettings()
 	settings.pvr.ta_skip			= 0;
 
 	settings.pvr.MaxThreads		    = 3;
-	settings.pvr.SynchronousRender	= true;
+	settings.pvr.AutoSkipFrame		= 0;
 
 	settings.debug.SerialConsole	= false;
 	settings.debug.SerialPTY        = false;
@@ -860,7 +860,16 @@ void LoadSettings(bool game_specific)
 	settings.pvr.ta_skip			= cfgLoadInt(config_section, "ta.skip", settings.pvr.ta_skip);
 
 	settings.pvr.MaxThreads		    = cfgLoadInt(config_section, "pvr.MaxThreads", settings.pvr.MaxThreads);
-	settings.pvr.SynchronousRender	= cfgLoadBool(config_section, "pvr.SynchronousRendering", settings.pvr.SynchronousRender);
+	if (game_specific)
+		settings.pvr.AutoSkipFrame = cfgLoadInt(config_section, "pvr.AutoSkipFrame", settings.pvr.AutoSkipFrame);
+	else
+	{
+		// compatibility with previous SynchronousRendering option
+		int autoskip = cfgLoadInt(config_section, "pvr.AutoSkipFrame", 99);
+		if (autoskip == 99)
+			autoskip = cfgLoadBool(config_section, "pvr.SynchronousRendering", true) ? 1 : 2;
+		settings.pvr.AutoSkipFrame = autoskip;
+	}
 
 	settings.debug.SerialConsole	= cfgLoadBool(config_section, "Debug.SerialConsoleEnabled", settings.debug.SerialConsole);
 	settings.debug.SerialPTY		= cfgLoadBool(config_section, "Debug.SerialPTY", settings.debug.SerialPTY);
@@ -1017,7 +1026,7 @@ void SaveSettings()
 	cfgSaveBool("config", "rend.WidescreenGameHacks", settings.rend.WidescreenGameHacks);
 
 	cfgSaveInt("config", "pvr.MaxThreads", settings.pvr.MaxThreads);
-	cfgSaveBool("config", "pvr.SynchronousRendering", settings.pvr.SynchronousRender);
+	cfgSaveInt("config", "pvr.AutoSkipFrame", settings.pvr.AutoSkipFrame);
 
 	cfgSaveBool("config", "Debug.SerialConsoleEnabled", settings.debug.SerialConsole);
 	cfgSaveBool("config", "Debug.SerialPTY", settings.debug.SerialPTY);

core/rend/gui.cpp

@@ -1151,9 +1151,21 @@ static void gui_display_settings()
 		    }
 		    if (ImGui::CollapsingHeader("Rendering Options", ImGuiTreeNodeFlags_DefaultOpen))
 		    {
-		    	ImGui::Checkbox("Synchronous Rendering", &settings.pvr.SynchronousRender);
+		    	ImGui::Text("Automatic Frame Skipping:");
+		    	ImGui::Columns(3, "autoskip", false);
+		    	ImGui::RadioButton("Disabled", &settings.pvr.AutoSkipFrame, 0);
 	            ImGui::SameLine();
-	            ShowHelpMarker("Reduce frame skipping by pausing the CPU when possible. Recommended for most platforms");
+	            ShowHelpMarker("No frame skipping");
+            	ImGui::NextColumn();
+		    	ImGui::RadioButton("Normal", &settings.pvr.AutoSkipFrame, 1);
+	            ImGui::SameLine();
+	            ShowHelpMarker("Skip a frame when the GPU and CPU are both running slow");
+            	ImGui::NextColumn();
+		    	ImGui::RadioButton("Maximum", &settings.pvr.AutoSkipFrame, 2);
+	            ImGui::SameLine();
+	            ShowHelpMarker("Skip a frame when the GPU is running slow");
+		    	ImGui::Columns(1, nullptr, false);
+
 		    	ImGui::Checkbox("Clipping", &settings.rend.Clipping);
 	            ImGui::SameLine();
 	            ShowHelpMarker("Enable clipping. May produce graphical errors when disabled");

core/types.h

@@ -475,7 +475,7 @@ struct settings_t
 		RenderType rend;
 
 		u32 MaxThreads;
-		bool SynchronousRender;
+		int AutoSkipFrame;		// 0: none, 1: some, 2: more
 
 		bool IsOpenGL() { return rend == RenderType::OpenGL || rend == RenderType::OpenGL_OIT; }
 	} pvr;