naomi: improve performance of M1 and M4 carts dma

Limit buffer size to 1 or 2 KB instead of 32 KB.
Only decrypt right before dma transfer.
Fixes frame drops and audio underruns in vf4evoct attract mode.
Issue #1717

core/hw/naomi/m1cartridge.cpp

@@ -39,7 +39,6 @@ void M1Cartridge::AdvancePtr(u32 size)
 			has_history = true;
 			buffer_actual_size = 0;
 		}
-		enc_fill();
 	}
 	else
 		NaomiCartridge::AdvancePtr(size);
@@ -147,11 +146,13 @@ void M1Cartridge::Serialize(Serializer& ser) const
 void M1Cartridge::Deserialize(Deserializer& deser)
 {
 	deser >> buffer;
+	deser.skip(32768 - sizeof(buffer), Deserializer::V52);
 	deser >> dict;
 	deser >> hist;
 	deser >> avail_val;
 	deser >> rom_cur_address;
 	deser >> buffer_actual_size;
+	buffer_actual_size = std::min<u32>(buffer_actual_size, sizeof(buffer));
 	deser >> avail_bits;
 	deser >> stream_ended;
 	deser >> has_history;

core/hw/naomi/m1cartridge.h

@@ -26,6 +26,7 @@ class M1Cartridge : public NaomiCartridge
 	{
 		if (encryption)
 		{
+			enc_fill();
 			size = std::min(size, (u32)sizeof(buffer));
 			return buffer;
 		}
@@ -48,7 +49,6 @@ class M1Cartridge : public NaomiCartridge
 			//printf("M1 ENCRYPTION ON @ %08x\n", dma_offset);
 			encryption = true;
 			enc_reset();
-			enc_fill();
 		}
 		else
 			encryption = false;
@@ -100,7 +100,7 @@ class M1Cartridge : public NaomiCartridge
 
 	u16 actel_id;
 
-	u8 buffer[32768];
+	u8 buffer[1024];
 	u8 dict[111], hist[2];
 	u64 avail_val;
 	u32 rom_cur_address, buffer_actual_size, avail_bits;

core/hw/naomi/m4cartridge.cpp

@@ -321,8 +321,10 @@ void M4Cartridge::Serialize(Serializer& ser) const
 void M4Cartridge::Deserialize(Deserializer& deser)
 {
 	deser >> buffer;
+	deser.skip(32768 - sizeof(buffer), Deserializer::V52);
 	deser >> rom_cur_address;
 	deser >> buffer_actual_size;
+	buffer_actual_size = std::min<u32>(buffer_actual_size, sizeof(buffer));
 	deser >> iv;
 	deser >> counter;
 	deser >> encryption;

core/hw/naomi/m4cartridge.h

@@ -69,7 +69,7 @@ class M4Cartridge: public NaomiCartridge {
 	u16 subkey2 = 0;
 	u16 one_round[0x10000];
 
-	u8 buffer[32768];
+	u8 buffer[2048];
 	u32 rom_cur_address, buffer_actual_size;
 	u16 iv;
 	u8 counter;
@@ -80,6 +80,8 @@ class M4Cartridge: public NaomiCartridge {
 	void enc_init();
 	void enc_fill();
 	u16 decrypt_one_round(u16 word, u16 subkey);
+
+	static_assert(sizeof(RomBootID) <= sizeof(buffer));
 };
 
 #endif /* CORE_HW_NAOMI_M4CARTRIDGE_H_ */