rework gpu busy timing

previous implementation caused complications with dynarecs
notaz · Oct 16, 2023 · f8896d1 · f8896d1
1 parent 7da5c7a
commit f8896d1
Show file tree

Hide file tree

Showing 7 changed files with 18 additions and 17 deletions.
diff --git a/libpcsxcore/gpu.c b/libpcsxcore/gpu.c
@@ -19,14 +19,13 @@ void gpu_state_change(int what)
 	switch (state)
 	{
 	case PGS_VRAM_TRANSFER_START:
-		HW_GPU_STATUS &= ~SWAP32(PSXGPU_nBUSY);
+		psxRegs.gpuIdleAfter = psxRegs.cycle + PSXCLK / 50;
 		break;
 	case PGS_VRAM_TRANSFER_END:
-		HW_GPU_STATUS |= SWAP32(PSXGPU_nBUSY);
+		psxRegs.gpuIdleAfter = psxRegs.cycle;
 		break;
 	case PGS_PRIMITIVE_START:
-		HW_GPU_STATUS &= ~SWAP32(PSXGPU_nBUSY);
-		set_event(PSXINT_GPUDMA, 200); // see gpuInterrupt
+		psxRegs.gpuIdleAfter = psxRegs.cycle + 200;
 		break;
 	}
 }
diff --git a/libpcsxcore/misc.c b/libpcsxcore/misc.c
@@ -738,6 +738,8 @@ int LoadState(const char *file) {
 	SaveFuncs.read(f, &psxRegs, offsetof(psxRegisters, gteBusyCycle));
 	psxRegs.gteBusyCycle = psxRegs.cycle;
 	psxRegs.biosBranchCheck = ~0;
+	psxRegs.gpuIdleAfter = psxRegs.cycle - 1;
+	HW_GPU_STATUS &= SWAP32(~PSXGPU_nBUSY);
 
 	psxCpu->Notify(R3000ACPU_NOTIFY_AFTER_LOAD, NULL);
 
@@ -750,8 +752,7 @@ int LoadState(const char *file) {
 	SaveFuncs.read(f, gpufP, sizeof(GPUFreeze_t));
 	GPU_freeze(0, gpufP);
 	free(gpufP);
-	if (HW_GPU_STATUS == 0)
-		HW_GPU_STATUS = SWAP32(GPU_readStatus());
+	gpuSyncPluginSR();
 
 	// spu
 	SaveFuncs.read(f, &Size, 4);

diff --git a/libpcsxcore/psxbios.c b/libpcsxcore/psxbios.c
@@ -3322,7 +3322,6 @@ void psxBiosSetupBootState(void)
 		GPU_writeStatus(gpu_ctl_def[i]);
 	for (i = 0; i < sizeof(gpu_data_def) / sizeof(gpu_data_def[0]); i++)
 		GPU_writeData(gpu_data_def[i]);
-	HW_GPU_STATUS |= SWAP32(PSXGPU_nBUSY);
 
 	// spu
 	for (i = 0x1f801d80; i < sizeof(spu_config) / sizeof(spu_config[0]); i++)

diff --git a/libpcsxcore/psxcounters.c b/libpcsxcore/psxcounters.c
@@ -397,6 +397,8 @@ void psxRcntUpdate()
             }
             HW_GPU_STATUS = SWAP32(status);
             GPU_vBlank(0, field);
+            if ((s32)(psxRegs.gpuIdleAfter - psxRegs.cycle) < 0)
+                psxRegs.gpuIdleAfter = psxRegs.cycle - 1; // prevent overflow
 
             if ((rcnts[0].mode & 7) == (RcSyncModeEnable | Rc01UnblankReset) ||
                 (rcnts[0].mode & 7) == (RcSyncModeEnable | Rc01UnblankReset2))

diff --git a/libpcsxcore/psxdma.c b/libpcsxcore/psxdma.c
@@ -157,7 +157,7 @@ void psxDma2(u32 madr, u32 bcr, u32 chcr) { // GPU
 			HW_DMA2_MADR = SWAPu32(madr + words_copy * 4);
 
 			// careful: gpu_state_change() also messes with this
-			HW_GPU_STATUS &= SWAP32(~PSXGPU_nBUSY);
+			psxRegs.gpuIdleAfter = psxRegs.cycle + words / 4 + 16;
 			// already 32-bit word size ((size * 4) / 4)
 			set_event(PSXINT_GPUDMA, words / 4);
 			return;
@@ -180,7 +180,7 @@ void psxDma2(u32 madr, u32 bcr, u32 chcr) { // GPU
 			HW_DMA2_MADR = SWAPu32(madr);
 
 			// careful: gpu_state_change() also messes with this
-			HW_GPU_STATUS &= SWAP32(~PSXGPU_nBUSY);
+			psxRegs.gpuIdleAfter = psxRegs.cycle + words / 4 + 16;
 			// already 32-bit word size ((size * 4) / 4)
 			set_event(PSXINT_GPUDMA, words / 4);
 			return;
@@ -199,13 +199,13 @@ void psxDma2(u32 madr, u32 bcr, u32 chcr) { // GPU
 			if ((int)size <= 0)
 				size = gpuDmaChainSize(madr);
 
-			HW_GPU_STATUS &= SWAP32(~PSXGPU_nBUSY);
 			HW_DMA2_MADR = SWAPu32(madr_next);
 
 			// Tekken 3 = use 1.0 only (not 1.5x)
 
 			// Einhander = parse linked list in pieces (todo)
 			// Rebel Assault 2 = parse linked list in pieces (todo)
+			psxRegs.gpuIdleAfter = psxRegs.cycle + size + 16;
 			set_event(PSXINT_GPUDMA, size);
 			return;
 
@@ -218,14 +218,13 @@ void psxDma2(u32 madr, u32 bcr, u32 chcr) { // GPU
 	DMA_INTERRUPT(2);
 }
 
-// note: this is also (ab)used for non-dma prim command
-// to delay gpu returning to idle state, see gpu_state_change()
 void gpuInterrupt() {
 	if (HW_DMA2_CHCR == SWAP32(0x01000401) && !(HW_DMA2_MADR & SWAP32(0x800000)))
 	{
 		u32 size, madr_next = 0xffffff;
 		size = GPU_dmaChain((u32 *)psxM, HW_DMA2_MADR & 0x1fffff, &madr_next);
 		HW_DMA2_MADR = SWAPu32(madr_next);
+		psxRegs.gpuIdleAfter = psxRegs.cycle + size + 64;
 		set_event(PSXINT_GPUDMA, size);
 		return;
 	}
@@ -234,7 +233,6 @@ void gpuInterrupt() {
 		HW_DMA2_CHCR &= SWAP32(~0x01000000);
 		DMA_INTERRUPT(2);
 	}
-	HW_GPU_STATUS |= SWAP32(PSXGPU_nBUSY); // GPU no longer busy
 }
 
 void psxDma6(u32 madr, u32 bcr, u32 chcr) {

diff --git a/libpcsxcore/psxhw.c b/libpcsxcore/psxhw.c
@@ -41,7 +41,7 @@ void psxHwReset() {
 	mdecInit(); // initialize mdec decoder
 	cdrReset();
 	psxRcntInit();
-	HW_GPU_STATUS = SWAP32(0x14802000);
+	HW_GPU_STATUS = SWAP32(0x10802000);
 	psxHwReadGpuSRptr = Config.hacks.gpu_busy_hack
 		? psxHwReadGpuSRbusyHack : psxHwReadGpuSR;
 }
@@ -91,16 +91,17 @@ void psxHwWriteGpuSR(u32 value)
 
 u32 psxHwReadGpuSR(void)
 {
-	u32 v;
+	u32 v, c = psxRegs.cycle;
 
 	// meh2, syncing for img bit, might want to avoid it..
 	gpuSyncPluginSR();
 	v = SWAP32(HW_GPU_STATUS);
+	v |= ((s32)(psxRegs.gpuIdleAfter - c) >> 31) & PSXGPU_nBUSY;
 
 	// XXX: because of large timeslices can't use hSyncCount, using rough
 	// approximization instead. Perhaps better use hcounter code here or something.
 	if (hSyncCount < 240 && (v & PSXGPU_ILACE_BITS) != PSXGPU_ILACE_BITS)
-		v |= PSXGPU_LCF & (psxRegs.cycle << 20);
+		v |= PSXGPU_LCF & (c << 20);
 	return v;
 }
 

diff --git a/libpcsxcore/r3000a.h b/libpcsxcore/r3000a.h
@@ -204,7 +204,8 @@ typedef struct {
 	u32 dloadVal[2];
 	u32 biosBranchCheck;
 	u32 cpuInRecursion;
-	u32 reserved[2];
+	u32 gpuIdleAfter;
+	u32 reserved[1];
 	// warning: changing anything in psxRegisters requires update of all
 	// asm in libpcsxcore/new_dynarec/
 } psxRegisters;