Add async CD access

7 files changed, 371 insertions, 4 deletions

diff --git a/Makefile b/Makefile
index f633665..9866050 100644
--- a/Makefile
+++ b/Makefile
@@ -6,7 +6,7 @@ CFLAGS += -Wall -Iinclude -ffast-math
 ifeq ($(DEBUG), 1)
 CFLAGS += -O0 -ggdb
 else
-ifeq ($(platform), vita)
+ifeq ($(platform), $(filter $(platform), vita ctr))
 CFLAGS += -O3 -DNDEBUG
 else
 CFLAGS += -O2 -DNDEBUG
diff --git a/frontend/3ds/3ds_utils.h b/frontend/3ds/3ds_utils.h
index cad4cbe..c9b9d7f 100644
--- a/frontend/3ds/3ds_utils.h
+++ b/frontend/3ds/3ds_utils.h
@@ -24,6 +24,20 @@ void threadFree(int32_t thread);
 void threadExit(int32_t rc)  __attribute__((noreturn));
 
 int32_t svcGetSystemInfo(int64_t* out, uint32_t type, int32_t param);
+
+int32_t svcCreateSemaphore(uint32_t *sem, int32_t initial_count, uint32_t max_count);
+int32_t svcReleaseSemaphore(int32_t *count, uint32_t sem, int32_t release_count);
+int32_t svcWaitSynchronization(uint32_t handle, int64_t nanoseconds);
+
+typedef int32_t LightLock;
+
+void LightLock_Init(LightLock* lock);
+void LightLock_Lock(LightLock* lock);
+int LightLock_TryLock(LightLock* lock);
+void LightLock_Unlock(LightLock* lock);
+
+int32_t APT_CheckNew3DS(bool *out);
+
 int32_t svcBackdoor(int32_t (*callback)(void));
 
 #define DEBUG_HOLD() do{printf("%s@%s:%d.\n",__FUNCTION__, __FILE__, __LINE__);fflush(stdout);wait_for_input();}while(0)
diff --git a/frontend/3ds/pthread.h b/frontend/3ds/pthread.h
index 9f43707..cc3c965 100644
--- a/frontend/3ds/pthread.h
+++ b/frontend/3ds/pthread.h
@@ -9,15 +9,34 @@
 #include "3ds_utils.h"
 
 #define CTR_PTHREAD_STACK_SIZE 0x10000
+#define FALSE 0
 
 typedef int32_t pthread_t;
 typedef int pthread_attr_t;
 
+typedef LightLock pthread_mutex_t;
+typedef int pthread_mutexattr_t;
+
+typedef struct {
+  uint32_t semaphore;
+  LightLock lock;
+  uint32_t waiting;
+} pthread_cond_t;
+
+typedef int pthread_condattr_t;
+
 static inline int pthread_create(pthread_t *thread,
       const pthread_attr_t *attr, void *(*start_routine)(void*), void *arg)
 {
-   thread = threadCreate(start_routine, arg, CTR_PTHREAD_STACK_SIZE, 0x25, -2, FALSE);
-   return 1;
+   int procnum = -2; // use default cpu
+   bool isNew3DS;
+   APT_CheckNew3DS(&isNew3DS);
+
+   if (isNew3DS)
+     procnum = 2;
+
+   *thread = threadCreate(start_routine, arg, CTR_PTHREAD_STACK_SIZE, 0x25, procnum, FALSE);
+   return 0;
 }
 
 
@@ -41,6 +60,64 @@ static inline void pthread_exit(void *retval)
    threadExit(0);
 }
 
+static inline int pthread_mutex_init(pthread_mutex_t *mutex, const pthread_mutexattr_t *attr) {
+  LightLock_Init(mutex);
+  return 0;
+}
+
+static inline int pthread_mutex_lock(pthread_mutex_t *mutex) {
+  LightLock_Lock(mutex);
+  return 0;
+}
+
+static inline int pthread_mutex_unlock(pthread_mutex_t *mutex) {
+  LightLock_Unlock(mutex);
+  return 0;
+}
+
+static inline int pthread_mutex_destroy(pthread_mutex_t *mutex) {
+  return 0;
+}
+
+static inline int pthread_cond_init(pthread_cond_t *cond, const pthread_condattr_t *attr) {
+  if (svcCreateSemaphore(&cond->semaphore, 0, 1))
+    goto error;
+
+  LightLock_Init(&cond->lock);
+  cond->waiting = 0;
+  return 0;
+
+ error:
+  svcCloseHandle(cond->semaphore);
+  return -1;
+}
+
+static inline int pthread_cond_signal(pthread_cond_t *cond) {
+  int32_t count;
+  LightLock_Lock(&cond->lock);
+  if (cond->waiting) {
+    cond->waiting--;
+    svcReleaseSemaphore(&count, cond->semaphore, 1);
+  }
+  LightLock_Unlock(&cond->lock);
+  return 0;
+}
+
+static inline int pthread_cond_wait(pthread_cond_t *cond, pthread_mutex_t *lock) {
+  LightLock_Lock(&cond->lock);
+  cond->waiting++;
+  LightLock_Unlock(lock);
+  LightLock_Unlock(&cond->lock);
+  svcWaitSynchronization(cond->semaphore, INT64_MAX);
+  LightLock_Lock(lock);
+  return 0;
+}
+
+static inline int pthread_cond_destroy(pthread_cond_t *cond) {
+  svcCloseHandle(cond->semaphore);
+  return 0;
+}
+
 
 #endif //_3DS_PTHREAD_WRAP__
 
diff --git a/frontend/libretro.c b/frontend/libretro.c
index cc443fe..bd7dc43 100644
--- a/frontend/libretro.c
+++ b/frontend/libretro.c
@@ -1852,6 +1852,18 @@ static void update_variables(bool in_flight)
          Config.VSyncWA = 1;
    }
 
+#ifndef _WIN32
+   var.value = NULL;
+   var.key = "pcsx_rearmed_async_cd";
+   if (environ_cb(RETRO_ENVIRONMENT_GET_VARIABLE, &var) || var.value)
+   {
+      if (strcmp(var.value, "async") == 0)
+        Config.AsyncCD = 1;
+      else
+        Config.AsyncCD = 0;
+   }
+#endif
+
    var.value = NULL;
    var.key = "pcsx_rearmed_noxadecoding";
    if (environ_cb(RETRO_ENVIRONMENT_GET_VARIABLE, &var) || var.value)
diff --git a/frontend/libretro_core_options.h b/frontend/libretro_core_options.h
index badd856..6f26c71 100644
--- a/frontend/libretro_core_options.h
+++ b/frontend/libretro_core_options.h
@@ -961,7 +961,19 @@ struct retro_core_option_definition option_defs_us[] = {
       },
       "disabled",
    },
-
+#ifndef _WIN32
+   {
+      "pcsx_rearmed_async_cd",
+      "CD Access Method (Restart)",
+      "Select method used to read data from content disk images. 'Synchronous' mimics original hardware. 'Asynchronous' can reduce stuttering on devices with slow storage.",
+      {
+         { "sync", "Synchronous" },
+         { "async",  "Asynchronous" },
+         { NULL, NULL},
+      },
+      "sync",
+   },
+#endif
    /* ADVANCED OPTIONS */
    {
       "pcsx_rearmed_noxadecoding",
diff --git a/libpcsxcore/cdriso.c b/libpcsxcore/cdriso.c
index 8171674..07adbe6 100644
--- a/libpcsxcore/cdriso.c
+++ b/libpcsxcore/cdriso.c
@@ -1163,6 +1163,194 @@ static int opensbifile(const char *isoname) {
 	return LoadSBI(sbiname, s);
 }
 
+#ifdef _WIN32
+static void readThreadStop() {}
+static void readThreadStart() {}
+#else
+static pthread_t read_thread_id;
+
+static pthread_cond_t read_thread_msg_avail;
+static pthread_cond_t read_thread_msg_done;
+static pthread_mutex_t read_thread_msg_lock;
+
+static pthread_cond_t sectorbuffer_cond;
+static pthread_mutex_t sectorbuffer_lock;
+
+static boolean read_thread_running = FALSE;
+static int read_thread_sector_start = -1;
+static int read_thread_sector_end = -1;
+
+typedef struct {
+  int sector;
+  long ret;
+  unsigned char data[CD_FRAMESIZE_RAW];
+} SectorBufferEntry;
+
+#define SECTOR_BUFFER_SIZE 4096
+
+static SectorBufferEntry *sectorbuffer;
+static size_t sectorbuffer_index;
+
+int (*sync_cdimg_read_func)(FILE *f, unsigned int base, void *dest, int sector);
+unsigned char *(*sync_CDR_getBuffer)(void);
+
+static unsigned char * CALLBACK ISOgetBuffer_async(void);
+static int cdread_async(FILE *f, unsigned int base, void *dest, int sector);
+
+static void *readThreadMain(void *param) {
+  int max_sector = -1;
+  int requested_sector_start = -1;
+  int requested_sector_end = -1;
+  int last_read_sector = -1;
+  int index = 0;
+
+  int ra_sector = -1;
+  int max_ra = 128;
+  int initial_ra = 1;
+  int speedmult_ra = 4;
+
+  int ra_count = 0;
+  int how_far_ahead = 0;
+
+  unsigned char tmpdata[CD_FRAMESIZE_RAW];
+  long ret;
+
+  max_sector = msf2sec(ti[numtracks].start) + msf2sec(ti[numtracks].length);
+
+  while(1) {
+    pthread_mutex_lock(&read_thread_msg_lock);
+
+    // If we don't have readahead and we don't have a sector request, wait for one.
+    // If we still have readahead to go, don't block, just keep going.
+    // And if we ever have a sector request pending, acknowledge and reset it.
+
+    if (!ra_count) {
+      if (read_thread_sector_start == -1 && read_thread_running) {
+        pthread_cond_wait(&read_thread_msg_avail, &read_thread_msg_lock);
+      }
+    }
+
+    if (read_thread_sector_start != -1) {
+      requested_sector_start = read_thread_sector_start;
+      requested_sector_end = read_thread_sector_end;
+      read_thread_sector_start = -1;
+      read_thread_sector_end = -1;
+      pthread_cond_signal(&read_thread_msg_done);
+    }
+
+    pthread_mutex_unlock(&read_thread_msg_lock);
+
+    if (!read_thread_running)
+      break;
+
+    // Readahead code, based on the implementation in mednafen psx's cdromif.cpp
+    if (requested_sector_start != -1) {
+      if (last_read_sector != -1 && last_read_sector == (requested_sector_start - 1)) {
+        how_far_ahead = ra_sector - requested_sector_end;
+
+        if(how_far_ahead <= max_ra)
+          ra_count = (max_ra - how_far_ahead + 1 ? max_ra - how_far_ahead + 1 : speedmult_ra);
+        else
+          ra_count++;
+      } else if (requested_sector_end != last_read_sector) {
+        ra_sector = requested_sector_end;
+        ra_count = initial_ra;
+      }
+
+      last_read_sector = requested_sector_end;
+    }
+
+    // check for end of CD
+    if (ra_count && ra_sector >= max_sector) {
+      ra_count = 0;
+    }
+
+    if (ra_count) {
+
+      index = ra_sector % SECTOR_BUFFER_SIZE;
+      pthread_mutex_lock(&sectorbuffer_lock);
+      if (sectorbuffer[index].sector != ra_sector) {
+        pthread_mutex_unlock(&sectorbuffer_lock);
+
+        ret = sync_cdimg_read_func(cdHandle, 0, tmpdata, ra_sector);
+
+        pthread_mutex_lock(&sectorbuffer_lock);
+        sectorbuffer[index].ret = ret;
+        sectorbuffer[index].sector = ra_sector;
+        memcpy(sectorbuffer[index].data, tmpdata, CD_FRAMESIZE_RAW);
+      }
+      pthread_cond_signal(&sectorbuffer_cond);
+      pthread_mutex_unlock(&sectorbuffer_lock);
+
+      ra_sector++;
+      ra_count--;
+    }
+  }
+
+  return NULL;
+}
+
+static void readThreadStop() {
+  if (read_thread_running == TRUE) {
+    read_thread_running = FALSE;
+    pthread_cond_signal(&read_thread_msg_avail);
+    pthread_join(read_thread_id, NULL);
+  }
+
+  pthread_cond_destroy(&read_thread_msg_done);
+  pthread_cond_destroy(&read_thread_msg_avail);
+  pthread_mutex_destroy(&read_thread_msg_lock);
+
+  pthread_cond_destroy(&sectorbuffer_cond);
+  pthread_mutex_destroy(&sectorbuffer_lock);
+
+  CDR_getBuffer = sync_CDR_getBuffer;
+  cdimg_read_func = sync_cdimg_read_func;
+
+  free(sectorbuffer);
+  sectorbuffer = NULL;
+}
+
+static void readThreadStart() {
+  SysPrintf("Starting async CD thread\n");
+
+  if (read_thread_running == TRUE)
+    return;
+
+  read_thread_running = TRUE;
+  read_thread_sector_start = -1;
+  read_thread_sector_end = -1;
+  sectorbuffer_index = 0;
+
+  sectorbuffer = calloc(SECTOR_BUFFER_SIZE, sizeof(SectorBufferEntry));
+  if(!sectorbuffer)
+    goto error;
+
+  sectorbuffer[0].sector = -1; // Otherwise we might think we've already fetched sector 0!
+
+  sync_CDR_getBuffer = CDR_getBuffer;
+  CDR_getBuffer = ISOgetBuffer_async;
+  sync_cdimg_read_func = cdimg_read_func;
+  cdimg_read_func = cdread_async;
+
+  if (pthread_cond_init(&read_thread_msg_avail, NULL) ||
+      pthread_cond_init(&read_thread_msg_done, NULL) ||
+      pthread_mutex_init(&read_thread_msg_lock, NULL) ||
+      pthread_cond_init(&sectorbuffer_cond, NULL) ||
+      pthread_mutex_init(&sectorbuffer_lock, NULL) ||
+      pthread_create(&read_thread_id, NULL, readThreadMain, NULL))
+    goto error;
+
+  return;
+
+ error:
+  SysPrintf("Error starting async CD thread\n");
+  SysPrintf("Falling back to sync\n");
+
+  readThreadStop();
+}
+#endif
+
 static int cdread_normal(FILE *f, unsigned int base, void *dest, int sector)
 {
 	fseek(f, base + sector * CD_FRAMESIZE_RAW, SEEK_SET);
@@ -1323,6 +1511,54 @@ static int cdread_2048(FILE *f, unsigned int base, void *dest, int sector)
 	return ret;
 }
 
+#ifndef _WIN32
+
+static int cdread_async(FILE *f, unsigned int base, void *dest, int sector) {
+  boolean found = FALSE;
+  int i = sector % SECTOR_BUFFER_SIZE;
+  long ret;
+
+  if (f != cdHandle || base != 0 || dest != cdbuffer) {
+    // Async reads are only supported for cdbuffer, so call the sync
+    // function directly.
+    return sync_cdimg_read_func(f, base, dest, sector);
+  }
+
+  pthread_mutex_lock(&read_thread_msg_lock);
+
+  // Only wait if we're not trying to read the next sector and
+  // sector_start is set (meaning the last request hasn't been
+  // processed yet)
+  while(read_thread_sector_start != -1 && read_thread_sector_end + 1 != sector) {
+    pthread_cond_wait(&read_thread_msg_done, &read_thread_msg_lock);
+  }
+
+  if (read_thread_sector_start == -1)
+    read_thread_sector_start = sector;
+
+  read_thread_sector_end = sector;
+  pthread_cond_signal(&read_thread_msg_avail);
+  pthread_mutex_unlock(&read_thread_msg_lock);
+
+  do {
+    pthread_mutex_lock(&sectorbuffer_lock);
+    if (sectorbuffer[i].sector == sector) {
+      sectorbuffer_index = i;
+      ret = sectorbuffer[i].ret;
+      found = TRUE;
+    }
+
+    if (!found) {
+      pthread_cond_wait(&sectorbuffer_cond, &sectorbuffer_lock);
+    }
+    pthread_mutex_unlock(&sectorbuffer_lock);
+  } while (!found);
+
+  return ret;
+}
+
+#endif
+
 static unsigned char * CALLBACK ISOgetBuffer_compr(void) {
 	return compr_img->buff_raw[compr_img->sector_in_blk] + 12;
 }
@@ -1333,6 +1569,17 @@ static unsigned char * CALLBACK ISOgetBuffer_chd(void) {
 }
 #endif
 
+#ifndef _WIN32
+static unsigned char * CALLBACK ISOgetBuffer_async(void) {
+  unsigned char *buffer;
+  pthread_mutex_lock(&sectorbuffer_lock);
+  buffer = sectorbuffer[sectorbuffer_index].data;
+  pthread_mutex_unlock(&sectorbuffer_lock);
+  return buffer + 12;
+}
+
+#endif
+
 static unsigned char * CALLBACK ISOgetBuffer(void) {
 	return cdbuffer + 12;
 }
@@ -1473,6 +1720,9 @@ static long CALLBACK ISOopen(void) {
 	cdda_cur_sector = 0;
 	cdda_file_offset = 0;
 
+  if (Config.AsyncCD) {
+    readThreadStart();
+  }
 	return 0;
 }
 
@@ -1518,6 +1768,7 @@ static long CALLBACK ISOclose(void) {
 	memset(cdbuffer, 0, sizeof(cdbuffer));
 	CDR_getBuffer = ISOgetBuffer;
 
+  readThreadStop();
 	return 0;
 }
 
diff --git a/libpcsxcore/psxcommon.h b/libpcsxcore/psxcommon.h
index 19c8d77..cb417d4 100644
--- a/libpcsxcore/psxcommon.h
+++ b/libpcsxcore/psxcommon.h
@@ -120,6 +120,7 @@ typedef struct {
 	boolean Mdec;
 	boolean PsxAuto;
 	boolean Cdda;
+	boolean AsyncCD;
 	boolean HLE;
 	boolean SlowBoot;
 	boolean Debug;