vatrace/vafool: refine it

remove va_fool_getframe.c which looks strange
correct some coding style for file va.c
simplify va_fool.c for both decode and encode
refine the surface dump of va_trace.c

Signed-off-by: Austin Yuan <shengquan.yuan@gmail.com>

10 files changed, 469 insertions, 888 deletions

diff --git a/va/Android.mk b/va/Android.mk
index 0f73a3e..d031814 100755
--- a/va/Android.mk
+++ b/va/Android.mk
@@ -35,8 +35,7 @@ include $(CLEAR_VARS)
 LOCAL_SRC_FILES := \
 	va.c \
 	va_trace.c \
-	va_fool.c \
-	va_fool_getframe.c
+	va_fool.c
 
 LOCAL_CFLAGS += \
 	-DANDROID \
@@ -59,7 +58,7 @@ LOCAL_COPY_HEADERS_TO := libva/va
 LOCAL_MODULE_TAGS := optional
 LOCAL_MODULE := libva
 
-LOCAL_SHARED_LIBRARIES := libdl libdrm libcutils
+LOCAL_SHARED_LIBRARIES := libdl libdrm libcutils liblog
 
 include $(BUILD_SHARED_LIBRARY)
 
diff --git a/va/Makefile.am b/va/Makefile.am
index 2130eb1..16a1e8f 100644
--- a/va/Makefile.am
+++ b/va/Makefile.am
@@ -34,7 +34,6 @@ LDADD = \
 libva_source_c = \
 	va.c			\
 	va_fool.c		\
-	va_fool_getframe.c	\
 	va_trace.c		\
 	$(NULL)
 
diff --git a/va/android/va_android.cpp b/va/android/va_android.cpp
index 869e4af..1f114d7 100644
--- a/va/android/va_android.cpp
+++ b/va/android/va_android.cpp
@@ -25,6 +25,8 @@
 #define _GNU_SOURCE 1
 #include "va.h"
 #include "va_backend.h"
+#include "va_trace.h"
+#include "va_fool.h"
 #include "va_android.h"
 #include "va_dricommon.h" /* needs some helper functions from this file */
 #include <stdio.h>
@@ -267,11 +269,6 @@ extern "C"  {
         );
 }
 
-#define VA_TRACE(trace_func,...)                \
-    if (trace_flag) {                           \
-        trace_func(__VA_ARGS__);                \
-    }
-
 VAStatus vaPutSurface (
     VADisplay dpy,
     VASurfaceID surface,
@@ -300,9 +297,9 @@ VAStatus vaPutSurface (
     CHECK_DISPLAY(dpy);
     ctx = CTX(dpy);
 
-    VA_TRACE(va_TracePutSurface, dpy, surface, static_cast<void*>(&draw), srcx, srcy, srcw, srch,
-             destx, desty, destw, desth,
-             cliprects, number_cliprects, flags );
+    VA_TRACE_LOG(va_TracePutSurface, dpy, surface, static_cast<void*>(&draw), srcx, srcy, srcw, srch,
+                 destx, desty, destw, desth,
+                 cliprects, number_cliprects, flags );
     
     return ctx->vtable->vaPutSurface( ctx, surface, static_cast<void*>(&draw), srcx, srcy, srcw, srch, 
                                      destx, desty, destw, desth,
diff --git a/va/va.c b/va/va.c
index 1fca874..3741d21 100644
--- a/va/va.c
+++ b/va/va.c
@@ -48,18 +48,6 @@
 #define CHECK_MAXIMUM(s, ctx, var) if (!va_checkMaximum(ctx->max_##var, #var)) s = VA_STATUS_ERROR_UNKNOWN;
 #define CHECK_STRING(s, ctx, var) if (!va_checkString(ctx->str_##var, #var)) s = VA_STATUS_ERROR_UNKNOWN;
 
-extern int trace_flag;
-#define VA_TRACE(trace_func,...)                \
-    if (trace_flag) {                           \
-        trace_func(__VA_ARGS__);                \
-    }
-
-extern int fool_decode;
-extern int fool_encode;
-#define VA_FOOL(fool_func,...)                 \
-    if (fool_decode || fool_encode) {          \
-        ret = fool_func(__VA_ARGS__);          \
-    }
 
 /*
  * read a config "env" for libva.conf or from environment setting
@@ -73,7 +61,6 @@ int va_parseConfig(char *env, char *env_value)
     char oneline[1024];
     FILE *fp=NULL;
 
-
     if (env == NULL)
         return 1;
     
@@ -138,8 +125,7 @@ void va_infoMessage(const char *msg, ...)
 
 static Bool va_checkVtable(void *ptr, char *function)
 {
-    if (!ptr)
-    {
+    if (!ptr) {
         va_errorMessage("No valid vtable entry for va%s\n", function);
         return False;
     }
@@ -148,8 +134,7 @@ static Bool va_checkVtable(void *ptr, char *function)
 
 static Bool va_checkMaximum(int value, char *variable)
 {
-    if (!value)
-    {
+    if (!value) {
         va_errorMessage("Failed to define max_%s in init\n", variable);
         return False;
     }
@@ -158,8 +143,7 @@ static Bool va_checkMaximum(int value, char *variable)
 
 static Bool va_checkString(const char* value, char *variable)
 {
-    if (!value)
-    {
+    if (!value) {
         va_errorMessage("Failed to define str_%s in init\n", variable);
         return False;
     }
@@ -182,19 +166,14 @@ static VAStatus va_openDriver(VADisplay dpy, char *driver_name)
     char *driver_dir;
     
     if (geteuid() == getuid())
-    {
         /* don't allow setuid apps to use LIBVA_DRIVERS_PATH */
         search_path = getenv("LIBVA_DRIVERS_PATH");
-    }
     if (!search_path)
-    {
         search_path = VA_DRIVERS_PATH;
-    }
 
     search_path = strdup((const char *)search_path);
-    driver_dir = strtok_r((const char *)search_path, ":", &saveptr);
-    while(driver_dir)
-    {
+    driver_dir = strtok_r(search_path, ":", &saveptr);
+    while (driver_dir) {
         void *handle = NULL;
         char *driver_path = (char *) malloc( strlen(driver_dir) +
                                              strlen(driver_name) +
@@ -210,25 +189,17 @@ static VAStatus va_openDriver(VADisplay dpy, char *driver_name)
 #else
         handle = dlopen( driver_path, RTLD_NOW| RTLD_GLOBAL);
 #endif
-        if (!handle)
-        {
+        if (!handle) {
             /* Don't give errors for non-existing files */
             if (0 == access( driver_path, F_OK))
-            {	
                 va_errorMessage("dlopen of %s failed: %s\n", driver_path, dlerror());
-            }
-        }
-        else
-        {
+        } else {
             VADriverInit init_func;
             init_func = (VADriverInit) dlsym(handle, VA_DRIVER_INIT_FUNC_S);
-            if (!init_func)
-            {
+            if (!init_func) {
                 va_errorMessage("%s has no function %s\n", driver_path, VA_DRIVER_INIT_FUNC_S);
                 dlclose(handle);
-            }
-            else
-            {
+            } else {
                 struct VADriverVTable *vtable = ctx->vtable;
 
                 vaStatus = VA_STATUS_SUCCESS;
@@ -242,8 +213,7 @@ static VAStatus va_openDriver(VADisplay dpy, char *driver_name)
                 if (VA_STATUS_SUCCESS == vaStatus)
                     vaStatus = (*init_func)(ctx);
 
-                if (VA_STATUS_SUCCESS == vaStatus)
-                {
+                if (VA_STATUS_SUCCESS == vaStatus) {
                     CHECK_MAXIMUM(vaStatus, ctx, profiles);
                     CHECK_MAXIMUM(vaStatus, ctx, entrypoints);
                     CHECK_MAXIMUM(vaStatus, ctx, attributes);
@@ -292,15 +262,12 @@ static VAStatus va_openDriver(VADisplay dpy, char *driver_name)
                     CHECK_VTABLE(vaStatus, ctx, GetDisplayAttributes);
                     CHECK_VTABLE(vaStatus, ctx, SetDisplayAttributes);
                 }
-                if (VA_STATUS_SUCCESS != vaStatus)
-                {
+                if (VA_STATUS_SUCCESS != vaStatus) {
                     va_errorMessage("%s init failed\n", driver_path);
                     dlclose(handle);
                 }
                 if (VA_STATUS_SUCCESS == vaStatus)
-                {
                     ctx->handle = handle;
-                }
                 free(driver_path);
                 break;
             }
@@ -318,7 +285,7 @@ static VAStatus va_openDriver(VADisplay dpy, char *driver_name)
 VAPrivFunc vaGetLibFunc(VADisplay dpy, const char *func)
 {
     VADriverContextP ctx;
-    if( !vaDisplayIsValid(dpy) )
+    if (!vaDisplayIsValid(dpy))
         return NULL;
     ctx = CTX(dpy);
 
@@ -334,8 +301,7 @@ VAPrivFunc vaGetLibFunc(VADisplay dpy, const char *func)
  */
 const char *vaErrorStr(VAStatus error_status)
 {
-    switch(error_status)
-    {
+    switch(error_status) {
         case VA_STATUS_SUCCESS:
             return "success (no error)";
         case VA_STATUS_ERROR_OPERATION_FAILED:
@@ -407,21 +373,17 @@ VAStatus vaInitialize (
     va_infoMessage("libva version %s\n", VA_VERSION_S);
 
     driver_name_env = getenv("LIBVA_DRIVER_NAME");
-    if (driver_name_env && geteuid() == getuid())
-    {
+    if (driver_name_env && geteuid() == getuid()) {
         /* Don't allow setuid apps to use LIBVA_DRIVER_NAME */
         driver_name = strdup(driver_name_env);
         vaStatus = VA_STATUS_SUCCESS;
         va_infoMessage("User requested driver '%s'\n", driver_name);
-    }
-    else
-    {
+    } else {
         vaStatus = va_getDriverName(dpy, &driver_name);
         va_infoMessage("va_getDriverName() returns %d\n", vaStatus);
     }
 
-    if (VA_STATUS_SUCCESS == vaStatus)
-    {
+    if (VA_STATUS_SUCCESS == vaStatus) {
         vaStatus = va_openDriver(dpy, driver_name);
         va_infoMessage("va_openDriver() returns %d\n", vaStatus);
 
@@ -432,7 +394,7 @@ VAStatus vaInitialize (
     if (driver_name)
         free(driver_name);
     
-    VA_TRACE(va_TraceInitialize, dpy, major_version, minor_version);
+    VA_TRACE_LOG(va_TraceInitialize, dpy, major_version, minor_version);
 
     return vaStatus;
 }
@@ -463,7 +425,7 @@ VAStatus vaTerminate (
   if (VA_STATUS_SUCCESS == vaStatus)
       pDisplayContext->vaDestroy(pDisplayContext);
 
-  VA_TRACE(va_TraceTerminate, dpy);
+  VA_TRACE_LOG(va_TraceTerminate, dpy);
 
   va_TraceEnd(dpy);
 
@@ -484,7 +446,7 @@ const char *vaQueryVendorString (
     VADisplay dpy
 )
 {
-  if( !vaDisplayIsValid(dpy) )
+  if (!vaDisplayIsValid(dpy))
       return NULL;
   
   return CTX(dpy)->str_vendor;
@@ -496,7 +458,7 @@ int vaMaxNumProfiles (
     VADisplay dpy
 )
 {
-  if( !vaDisplayIsValid(dpy) )
+  if (!vaDisplayIsValid(dpy))
       return 0;
   
   return CTX(dpy)->max_profiles;
@@ -507,7 +469,7 @@ int vaMaxNumEntrypoints (
     VADisplay dpy
 )
 {
-  if( !vaDisplayIsValid(dpy) )
+  if (!vaDisplayIsValid(dpy))
       return 0;
   
   return CTX(dpy)->max_entrypoints;
@@ -519,7 +481,7 @@ int vaMaxNumConfigAttributes (
     VADisplay dpy
 )
 {
-  if( !vaDisplayIsValid(dpy) )
+  if (!vaDisplayIsValid(dpy))
       return 0;
   
   return CTX(dpy)->max_attributes;
@@ -583,11 +545,11 @@ VAStatus vaCreateConfig (
   CHECK_DISPLAY(dpy);
   ctx = CTX(dpy);
 
-  VA_FOOL(va_FoolCreateConfig, dpy, profile, entrypoint, attrib_list, num_attribs, config_id);
+  vaStatus = ctx->vtable->vaCreateConfig ( ctx, profile, entrypoint, attrib_list, num_attribs, config_id );
 
-  vaStatus =  ctx->vtable->vaCreateConfig ( ctx, profile, entrypoint, attrib_list, num_attribs, config_id );
-
-  VA_TRACE(va_TraceCreateConfig, dpy, profile, entrypoint, attrib_list, num_attribs, config_id);
+  /* record the current entrypoint for further trace/fool determination */
+  VA_TRACE_FUNC(va_TraceCreateConfig, dpy, profile, entrypoint, attrib_list, num_attribs, config_id);
+  VA_FOOL_FUNC(va_FoolCreateConfig, dpy, profile, entrypoint, attrib_list, num_attribs, config_id);
   
   return vaStatus;
 }
@@ -638,9 +600,7 @@ VAStatus vaCreateSurfaces (
 
   vaStatus = ctx->vtable->vaCreateSurfaces( ctx, width, height, format, num_surfaces, surfaces );
 
-  VA_TRACE(va_TraceCreateSurface, dpy, width, height, format, num_surfaces, surfaces);
-
-  VA_FOOL(va_FoolCreateSurfaces, dpy, width, height, format, num_surfaces, surfaces);
+  VA_TRACE_LOG(va_TraceCreateSurface, dpy, width, height, format, num_surfaces, surfaces);
   
   return vaStatus;
 }
@@ -679,7 +639,8 @@ VAStatus vaCreateContext (
   vaStatus = ctx->vtable->vaCreateContext( ctx, config_id, picture_width, picture_height,
                                       flag, render_targets, num_render_targets, context );
 
-  VA_TRACE(va_TraceCreateContext, dpy, config_id, picture_width, picture_height, flag, render_targets, num_render_targets, context);
+  /* keep current encode/decode resoluton */
+  VA_TRACE_FUNC(va_TraceCreateContext, dpy, config_id, picture_width, picture_height, flag, render_targets, num_render_targets, context);
 
   return vaStatus;
 }
@@ -711,7 +672,7 @@ VAStatus vaCreateBuffer (
   ctx = CTX(dpy);
   int ret = 0;
 
-  VA_FOOL(va_FoolCreateBuffer, dpy, context, type, size, num_elements, data, buf_id);
+  VA_FOOL_FUNC(va_FoolCreateBuffer, dpy, context, type, size, num_elements, data, buf_id);
   if (ret)
       return VA_STATUS_SUCCESS;
 
@@ -727,7 +688,9 @@ VAStatus vaBufferSetNumElements (
   VADriverContextP ctx;
   CHECK_DISPLAY(dpy);
   ctx = CTX(dpy);
-
+  
+  VA_FOOL_RETURN();
+  
   return ctx->vtable->vaBufferSetNumElements( ctx, buf_id, num_elements );
 }
 
@@ -744,15 +707,14 @@ VAStatus vaMapBuffer (
   
   CHECK_DISPLAY(dpy);
   ctx = CTX(dpy);
-  
-  VA_FOOL(va_FoolMapBuffer, dpy, buf_id, pbuf);
+
+  VA_FOOL_FUNC(va_FoolMapBuffer, dpy, buf_id, pbuf);
   if (ret)
       return VA_STATUS_SUCCESS;
-
+  
   va_status = ctx->vtable->vaMapBuffer( ctx, buf_id, pbuf );
 
-  if (va_status == VA_STATUS_SUCCESS)
-      VA_TRACE(va_TraceMapBuffer, dpy, buf_id, pbuf);
+  VA_TRACE_LOG(va_TraceMapBuffer, dpy, buf_id, pbuf);
   
   return va_status;
 }
@@ -767,7 +729,7 @@ VAStatus vaUnmapBuffer (
   ctx = CTX(dpy);
   int ret = 0;
 
-  VA_FOOL(va_FoolUnmapBuffer, dpy, buf_id);
+  VA_FOOL_FUNC(va_FoolUnmapBuffer, dpy, buf_id);
   if (ret)
       return VA_STATUS_SUCCESS;
 
@@ -783,6 +745,8 @@ VAStatus vaDestroyBuffer (
   CHECK_DISPLAY(dpy);
   ctx = CTX(dpy);
 
+  VA_FOOL_RETURN();
+  
   return ctx->vtable->vaDestroyBuffer( ctx, buffer_id );
 }
 
@@ -796,9 +760,15 @@ VAStatus vaBufferInfo (
 )
 {
   VADriverContextP ctx;
+  int ret = 0;
+  
   CHECK_DISPLAY(dpy);
   ctx = CTX(dpy);
 
+  VA_FOOL_FUNC(va_FoolBufferInfo, dpy, buf_id, type, size, num_elements);
+  if (ret)
+      return VA_STATUS_SUCCESS;
+  
   return ctx->vtable->vaBufferInfo( ctx, buf_id, type, size, num_elements );
 }
 
@@ -809,18 +779,18 @@ VAStatus vaBeginPicture (
 )
 {
   VADriverContextP ctx;
+  VAStatus va_status;
   int ret = 0;
 
   CHECK_DISPLAY(dpy);
   ctx = CTX(dpy);
 
-  VA_TRACE(va_TraceBeginPicture, dpy, context, render_target);
-
-  VA_FOOL(va_FoolBeginPicture, dpy, context, render_target);
-  if (ret)
-      return VA_STATUS_SUCCESS;
-
-  return ctx->vtable->vaBeginPicture( ctx, context, render_target );
+  VA_TRACE_FUNC(va_TraceBeginPicture, dpy, context, render_target);
+  VA_FOOL_RETURN();
+  
+  va_status = ctx->vtable->vaBeginPicture( ctx, context, render_target );
+  
+  return va_status;
 }
 
 VAStatus vaRenderPicture (
@@ -836,11 +806,8 @@ VAStatus vaRenderPicture (
   CHECK_DISPLAY(dpy);
   ctx = CTX(dpy);
 
-  VA_FOOL(va_FoolRenderPicture, dpy, context, buffers, num_buffers);
-  if (ret)
-      return VA_STATUS_SUCCESS;
-
-  VA_TRACE(va_TraceRenderPicture, dpy, context, buffers, num_buffers);
+  VA_TRACE_LOG(va_TraceRenderPicture, dpy, context, buffers, num_buffers);
+  VA_FOOL_RETURN();
 
   return ctx->vtable->vaRenderPicture( ctx, context, buffers, num_buffers );
 }
@@ -857,15 +824,14 @@ VAStatus vaEndPicture (
   CHECK_DISPLAY(dpy);
   ctx = CTX(dpy);
 
-  VA_FOOL(va_FoolEndPicture, dpy, context);
-  if (ret) {
-      VA_TRACE(va_TraceEndPicture, dpy, context);
-      return VA_STATUS_SUCCESS;
-  }
+  /* dump encode source surface */
+  VA_TRACE_SURFACE(va_TraceEndPicture, dpy, context, 0);
+  /* return directly if do dummy operation */
+  VA_FOOL_RETURN();
   
   va_status = ctx->vtable->vaEndPicture( ctx, context );
-  
-  VA_TRACE(va_TraceEndPicture, dpy, context);
+  /* dump decode dest surface */
+  VA_TRACE_SURFACE(va_TraceEndPicture, dpy, context, 1);
 
   return va_status;
 }
@@ -882,12 +848,8 @@ VAStatus vaSyncSurface (
   CHECK_DISPLAY(dpy);
   ctx = CTX(dpy);
 
-  VA_FOOL(va_FoolSyncSurface, dpy, render_target);
-  if (ret)
-      return VA_STATUS_SUCCESS;
-  
   va_status = ctx->vtable->vaSyncSurface( ctx, render_target );
-  VA_TRACE(va_TraceSyncSurface, dpy, render_target);
+  VA_TRACE_LOG(va_TraceSyncSurface, dpy, render_target);
 
   return va_status;
 }
@@ -905,7 +867,7 @@ VAStatus vaQuerySurfaceStatus (
 
   va_status = ctx->vtable->vaQuerySurfaceStatus( ctx, render_target, status );
 
-  VA_TRACE(va_TraceQuerySurfaceStatus, dpy, render_target, status);
+  VA_TRACE_LOG(va_TraceQuerySurfaceStatus, dpy, render_target, status);
 
   return va_status;
 }
@@ -924,7 +886,7 @@ VAStatus vaQuerySurfaceError (
 
   va_status = ctx->vtable->vaQuerySurfaceError( ctx, surface, error_status, error_info );
 
-  VA_TRACE(va_TraceQuerySurfaceError, dpy, surface, error_status, error_info);
+  VA_TRACE_LOG(va_TraceQuerySurfaceError, dpy, surface, error_status, error_info);
 
   return va_status;
 }
@@ -934,7 +896,7 @@ int vaMaxNumImageFormats (
     VADisplay dpy
 )
 {
-  if( !vaDisplayIsValid(dpy) )
+  if (!vaDisplayIsValid(dpy))
       return 0;
   
   return CTX(dpy)->max_image_formats;
@@ -1100,7 +1062,7 @@ int vaMaxNumSubpictureFormats (
     VADisplay dpy
 )
 {
-  if( !vaDisplayIsValid(dpy) )
+  if (!vaDisplayIsValid(dpy))
       return 0;
   
   return CTX(dpy)->max_subpic_formats;
@@ -1126,10 +1088,6 @@ VAStatus vaQuerySubpictureFormats (
   CHECK_DISPLAY(dpy);
   ctx = CTX(dpy);
 
-  VA_FOOL(va_FoolQuerySubpictureFormats, dpy, format_list, flags, num_formats);
-  if (ret)
-      return VA_STATUS_SUCCESS;
-
   return ctx->vtable->vaQuerySubpictureFormats ( ctx, format_list, flags, num_formats);
 }
 
@@ -1275,12 +1233,12 @@ int vaMaxNumDisplayAttributes (
 {
   int tmp;
     
-  if( !vaDisplayIsValid(dpy) )
+  if (!vaDisplayIsValid(dpy))
       return 0;
   
   tmp = CTX(dpy)->max_display_attributes;
 
-  VA_TRACE(va_TraceMaxNumDisplayAttributes, dpy, tmp);
+  VA_TRACE_LOG(va_TraceMaxNumDisplayAttributes, dpy, tmp);
   
   return tmp;
 }
@@ -1298,14 +1256,13 @@ VAStatus vaQueryDisplayAttributes (
 )
 {
   VADriverContextP ctx;
-  CHECK_DISPLAY(dpy);
-  ctx = CTX(dpy);
-
   VAStatus va_status;
   
+  CHECK_DISPLAY(dpy);
+  ctx = CTX(dpy);
   va_status = ctx->vtable->vaQueryDisplayAttributes ( ctx, attr_list, num_attributes );
 
-  VA_TRACE(va_TraceQueryDisplayAttributes, dpy, attr_list, num_attributes);
+  VA_TRACE_LOG(va_TraceQueryDisplayAttributes, dpy, attr_list, num_attributes);
 
   return va_status;
   
@@ -1324,14 +1281,13 @@ VAStatus vaGetDisplayAttributes (
 )
 {
   VADriverContextP ctx;
+  VAStatus va_status;
+
   CHECK_DISPLAY(dpy);
   ctx = CTX(dpy);
-
-  VAStatus va_status;
-  
   va_status = ctx->vtable->vaGetDisplayAttributes ( ctx, attr_list, num_attributes );
 
-  VA_TRACE(va_TraceGetDisplayAttributes, dpy, attr_list, num_attributes);
+  VA_TRACE_LOG(va_TraceGetDisplayAttributes, dpy, attr_list, num_attributes);
   
   return va_status;
 }
@@ -1349,13 +1305,14 @@ VAStatus vaSetDisplayAttributes (
 )
 {
   VADriverContextP ctx;
+  VAStatus va_status;
   CHECK_DISPLAY(dpy);
   ctx = CTX(dpy);
 
-  VA_TRACE(va_TraceSetDisplayAttributes, dpy, attr_list, num_attributes);
-
+  va_status = ctx->vtable->vaSetDisplayAttributes ( ctx, attr_list, num_attributes );
+  VA_TRACE_LOG(va_TraceSetDisplayAttributes, dpy, attr_list, num_attributes);
   
-  return ctx->vtable->vaSetDisplayAttributes ( ctx, attr_list, num_attributes );
+  return va_status;
 }
 
 VAStatus vaLockSurface(VADisplay dpy,
diff --git a/va/va_fool.c b/va/va_fool.c
index 53b8f7a..807307e 100644
--- a/va/va_fool.c
+++ b/va/va_fool.c
@@ -39,6 +39,7 @@
 #include <sys/stat.h>
 #include <unistd.h>
 #include <time.h>
+#include <fcntl.h>
 
 /*
  * Do dummy decode/encode, ignore the input data
@@ -46,42 +47,37 @@
  * We export env "VA_FOOL", with which, we can do fake decode/encode:
  *
  * LIBVA_FOOL_DECODE:
- * . if set, decode does nothing, but fill in some YUV data
- * LIBVA_FOOL_ENCODE=<clip name>:
- * . if set, encode does nothing, but fill in the coded buffer from a H264 clip.
- * . VA CONTEXT/CONFIG/SURFACE will call into drivers, but VA Buffer creation is done in here
- * . Bypass all "vaBeginPic/vaRenderPic/vaEndPic"
+ * . if set, decode does nothing
+ * LIBVA_FOOL_ENCODE=<framename>:
+ * . if set, encode does nothing, but fill in the coded buffer from the content of files with
+ *   name framename.0,framename.1,framename.2, ..., framename.N, framename.N,framename.N,...
+ * LIBVA_FOOL_JPEG=<framename>:fill the content of filename to codedbuf for jpeg encoding
  * LIBVA_FOOL_POSTP:
  * . if set, do nothing for vaPutSurface
  */
 
 
 /* global settings */
-
-/* LIBVA_FOOL_DECODE/LIBVA_FOOL_ENCODE/LIBVA_FOOL_POSTP */
-int fool_decode = 0;
-int fool_encode = 0;
+int fool_codec = 0;
 int fool_postp  = 0;
 
-
-
-#define NAL_BUF_SIZE  65536  // maximum NAL unit size
-#define RING_BUF_SIZE  8192  // input ring buffer size, MUST be a power of two!
-#define MAX_FRAME 16
-#define SLICE_NUM 4
-
 #define FOOL_CONTEXT_MAX 4
+
+#define FOOL_BUFID_MAGIC   0x12345600
+#define FOOL_BUFID_MASK    0xffffff00
 /* per context settings */
 static struct _fool_context {
     VADisplay dpy; /* should use context as the key */
 
-    VAProfile fool_profile; /* current profile for buffers */
-    VAEntrypoint fool_entrypoint; /* current entrypoint */
+    char *fn_enc;/* file pattern with codedbuf content for encode */
+    char *segbuf_enc; /* the segment buffer of coded buffer, load frome fn_enc */
+    int file_count;
 
-    FILE *fool_fp_codedclip; /* load a clip from disk for fooling encode*/
-    char *frame_buf;
-    VACodedBufferSegment *codebuf;
+    char *fn_jpg;/* file name of JPEG fool with codedbuf content */
+    char *segbuf_jpg; /* the segment buffer of coded buffer, load frome fn_jpg */
 
+    VAEntrypoint entrypoint; /* current entrypoint */
+    
     /* all buffers with same type share one malloc-ed memory
      * bufferID = (buffer numbers with the same type << 8) || type
      * the malloc-ed memory can be find by fool_buf[bufferID & 0xff]
@@ -89,18 +85,10 @@ static struct _fool_context {
      */
     char *fool_buf[VABufferTypeMax]; /* memory of fool buffers */
     unsigned int fool_buf_size[VABufferTypeMax]; /* size of memory of fool buffers */
+    unsigned int fool_buf_element[VABufferTypeMax]; /* element count of created buffers */
     unsigned int fool_buf_count[VABufferTypeMax]; /* count of created buffers */
     VAContextID context;
-} fool_context[FOOL_CONTEXT_MAX] = { {0} }; /* trace five context at the same time */
-
-#define FOOL_DECODE(idx) (fool_decode && (fool_context[idx].fool_entrypoint == VAEntrypointVLD))
-#define FOOL_ENCODE(idx)                                                \
-    (fool_encode                                                        \
-     && (fool_context[idx].fool_entrypoint == VAEntrypointEncSlice)     \
-     && (fool_context[idx].fool_profile >= VAProfileH264Baseline)       \
-     && (fool_context[idx].fool_profile <= VAProfileH264High))
-
-
+} fool_context[FOOL_CONTEXT_MAX]; /* trace five context at the same time */
 
 #define DPY2INDEX(dpy)                                  \
     int idx;                                            \
@@ -112,40 +100,12 @@ static struct _fool_context {
     if (idx == FOOL_CONTEXT_MAX)                        \
         return 0;  /* let driver go */
 
-
 /* Prototype declarations (functions defined in va.c) */
 
 void va_errorMessage(const char *msg, ...);
 void va_infoMessage(const char *msg, ...);
 
-int va_parseConfig(char *env, char *env_value);
-
-VAStatus vaBufferInfo(
-        VADisplay dpy,
-        VAContextID context,	/* in */
-        VABufferID buf_id,		/* in */
-        VABufferType *type,		/* out */
-        unsigned int *size,		/* out */
-        unsigned int *num_elements	/* out */
-);
-
-VAStatus vaLockSurface(VADisplay dpy,
-        VASurfaceID surface,
-        unsigned int *fourcc, /* following are output argument */
-        unsigned int *luma_stride,
-        unsigned int *chroma_u_stride,
-        unsigned int *chroma_v_stride,
-        unsigned int *luma_offset,
-        unsigned int *chroma_u_offset,
-        unsigned int *chroma_v_offset,
-        unsigned int *buffer_name,
-        void **buffer
-);
-
-VAStatus vaUnlockSurface(VADisplay dpy,
-        VASurfaceID surface
-);
-
+int  va_parseConfig(char *env, char *env_value);
 
 void va_FoolInit(VADisplay dpy)
 {
@@ -159,41 +119,30 @@ void va_FoolInit(VADisplay dpy)
     if (fool_index == FOOL_CONTEXT_MAX)
         return;
 
+    memset(&fool_context[fool_index], 0, sizeof(struct _fool_context));
     if (va_parseConfig("LIBVA_FOOL_POSTP", NULL) == 0) {
         fool_postp = 1;
         va_infoMessage("LIBVA_FOOL_POSTP is on, dummy vaPutSurface\n");
     }
-
-
+    
     if (va_parseConfig("LIBVA_FOOL_DECODE", NULL) == 0) {
-        fool_decode = 1;
+        fool_codec  |= VA_FOOL_FLAG_DECODE;
         va_infoMessage("LIBVA_FOOL_DECODE is on, dummy decode\n");
     }
-
-
     if (va_parseConfig("LIBVA_FOOL_ENCODE", &env_value[0]) == 0) {
-        fool_context[fool_index].fool_fp_codedclip = fopen(env_value, "r");
-
-        if (fool_context[fool_index].fool_fp_codedclip) {
-            fool_encode = 1;
-        } else
-            fool_encode = 0;
-
-        if (fool_encode) /* malloc the buffer for fake clip */
-        {
-            fool_context[fool_index].frame_buf = malloc(MAX_FRAME*SLICE_NUM*NAL_BUF_SIZE*sizeof(char));
-            fool_context[fool_index].codebuf = malloc(sizeof(VACodedBufferSegment));
-        }
-
-        if (fool_context[fool_index].frame_buf == NULL)
-            fool_encode = 0;
-
-        if (fool_encode)
-            va_infoMessage("LIBVA_FOOL_ENCODE is on, dummy encode\n");
-
+        fool_codec  |= VA_FOOL_FLAG_ENCODE;
+        fool_context[fool_index].fn_enc = strdup(env_value);
+        va_infoMessage("LIBVA_FOOL_ENCODE is on, load encode data from file with patten %s\n",
+                       fool_context[fool_index].fn_enc);
     }
-
-    if (fool_encode || fool_decode)
+    if (va_parseConfig("LIBVA_FOOL_JPEG", &env_value[0]) == 0) {
+        fool_codec  |= VA_FOOL_FLAG_JPEG;
+        fool_context[fool_index].fn_jpg = strdup(env_value);
+        va_infoMessage("LIBVA_FOOL_JPEG is on, load encode data from file with patten %s\n",
+                       fool_context[fool_index].fn_jpg);
+    }
+    
+    if (fool_codec)
         fool_context[fool_index].dpy = dpy;
 }
 
@@ -201,29 +150,23 @@ void va_FoolInit(VADisplay dpy)
 int va_FoolEnd(VADisplay dpy)
 {
     int i;
-
     DPY2INDEX(dpy);
 
     for (i = 0; i < VABufferTypeMax; i++) {/* free memory */
         if (fool_context[idx].fool_buf[i])
             free(fool_context[idx].fool_buf[i]);
     }
-    if (fool_context[idx].fool_fp_codedclip)
-        fclose(fool_context[idx].fool_fp_codedclip);
-
-    if (fool_context[idx].frame_buf)
-        free(fool_context[idx].frame_buf);
+    if (fool_context[idx].segbuf_enc)
+        free(fool_context[idx].segbuf_enc);
+    if (fool_context[idx].segbuf_jpg)
+        free(fool_context[idx].segbuf_jpg);
+    if (fool_context[idx].fn_enc)
+        free(fool_context[idx].fn_enc);
+    if (fool_context[idx].fn_jpg)
+        free(fool_context[idx].fn_jpg);
     
-    if (fool_context[idx].codebuf)
-        free(fool_context[idx].codebuf);
-
     memset(&fool_context[idx], 0, sizeof(struct _fool_context));
-    return 0;
-}
-
-int va_FoolCodedBuf(VADisplay dpy)
-{
-    /* do nothing */
+    
     return 0;
 }
 
@@ -239,310 +182,175 @@ int va_FoolCreateConfig(
 {
     DPY2INDEX(dpy);
 
-    /* call into driver level to allocate real context/surface/buffers, etc */
-    fool_context[idx].fool_profile = profile;
-    fool_context[idx].fool_entrypoint = entrypoint;
-    return 0;
-}
-
-static int yuvgen_planar(
-        int width, int height,
-        unsigned char *Y_start, int Y_pitch,
-        unsigned char *U_start, int U_pitch,
-        unsigned char *V_start, int V_pitch,
-        int UV_interleave, int box_width, int row_shift,
-        int field
-)
-{
-    int row;
-
-    /* copy Y plane */
-    for (row=0;row<height;row++) {
-        unsigned char *Y_row = Y_start + row * Y_pitch;
-        int jj, xpos, ypos;
-
-        ypos = (row / box_width) & 0x1;
-
-        /* fill garbage data into the other field */
-        if (((field == VA_TOP_FIELD) && (row &1))
-                || ((field == VA_BOTTOM_FIELD) && ((row &1)==0))) {
-            memset(Y_row, 0xff, width);
-            continue;
-        }
-
-        for (jj=0; jj<width; jj++) {
-            xpos = ((row_shift + jj) / box_width) & 0x1;
-
-            if ((xpos == 0) && (ypos == 0))
-                Y_row[jj] = 0xeb;
-            if ((xpos == 1) && (ypos == 1))
-                Y_row[jj] = 0xeb;
-
-            if ((xpos == 1) && (ypos == 0))
-                Y_row[jj] = 0x10;
-            if ((xpos == 0) && (ypos == 1))
-                Y_row[jj] = 0x10;
-        }
-    }
-
-    /* copy UV data */
-    for( row =0; row < height/2; row++) {
-        unsigned short value = 0x80;
-
-        /* fill garbage data into the other field */
-        if (((field == VA_TOP_FIELD) && (row &1))
-                || ((field == VA_BOTTOM_FIELD) && ((row &1)==0))) {
-            value = 0xff;
-        }
-
-        if (UV_interleave) {
-            unsigned short *UV_row = (unsigned short *)(U_start + row * U_pitch);
-
-            memset(UV_row, value, width);
-        } else {
-            unsigned char *U_row = U_start + row * U_pitch;
-            unsigned char *V_row = V_start + row * V_pitch;
-
-            memset (U_row,value,width/2);
-            memset (V_row,value,width/2);
-        }
-    }
-
-    return 0;
+    fool_context[idx].entrypoint = entrypoint;
+    
+    /*
+     * check fool_codec to align with current context
+     * e.g. fool_codec = decode then for encode, the
+     * vaBegin/vaRender/vaEnd also run into fool path
+     * which is not desired
+     */
+    if (((fool_codec & VA_FOOL_FLAG_DECODE) && (entrypoint == VAEntrypointVLD)) ||
+        ((fool_codec & VA_FOOL_FLAG_ENCODE) && (entrypoint == VAEntrypointEncSlice)) ||
+        ((fool_codec & VA_FOOL_FLAG_JPEG) && (entrypoint == VAEntrypointEncPicture)))
+        ; /* the fool_codec is meaningful */
+    else
+        fool_codec = 0;
+
+    return 0; /* driver continue */
 }
 
 
-int va_FoolCreateSurfaces(
-        VADisplay dpy,
-        int width,
-        int height,
-        int format,
-        int num_surfaces,
-        VASurfaceID *surfaces	/* out */
+VAStatus va_FoolCreateBuffer(
+    VADisplay dpy,
+    VAContextID context,	/* in */
+    VABufferType type,		/* in */
+    unsigned int size,		/* in */
+    unsigned int num_elements,	/* in */
+    void *data,			/* in */
+    VABufferID *buf_id		/* out */
 )
 {
-    int i;
-    unsigned int fourcc; /* following are output argument */
-    unsigned int luma_stride;
-    unsigned int chroma_u_stride;
-    unsigned int chroma_v_stride;
-    unsigned int luma_offset;
-    unsigned int chroma_u_offset;
-    unsigned int chroma_v_offset;
-    unsigned int buffer_name;
-    void *buffer = NULL;
-    unsigned char *Y_data, *U_data, *V_data;
-
-    int box_width = num_surfaces/2;
-    int row_shift = 0;
-    VAStatus va_status;
-
+    unsigned int new_size = size * num_elements;
+    unsigned int old_size;
     DPY2INDEX(dpy);
 
-    if (FOOL_DECODE(idx)) { 
-        /* call into driver level to allocate real context/surface/buffers, etc
-         * fill in the YUV data, will be overwrite if it is encode context
-         */
-        for (i = 0; i < num_surfaces; i++) {
-            /* fool decoder: fill with auto-generated YUV data */
-            va_status = vaLockSurface(dpy, surfaces[i], &fourcc,
-                    &luma_stride, &chroma_u_stride, &chroma_v_stride,
-                    &luma_offset, &chroma_u_offset, &chroma_v_offset,
-                    &buffer_name, &buffer);
-
-            if (va_status != VA_STATUS_SUCCESS)
-                return 0;
-
-            if (!buffer) {
-                vaUnlockSurface(dpy, surfaces[i]);
-                return 0;
-            }
-
-            Y_data = buffer;
-
-            /* UV should be same for NV12 */
-            U_data = buffer + chroma_u_offset;
-            V_data = buffer + chroma_v_offset;
-
-            yuvgen_planar(width, height,
-                    Y_data, luma_stride,
-                    U_data, chroma_v_stride,
-                    V_data, chroma_v_stride,
-                    (fourcc==VA_FOURCC_NV12),
-                    box_width, row_shift, 0);
-
-            vaUnlockSurface(dpy, surfaces[i]);
-
-            row_shift++;
-            if (row_shift==(2*box_width))
-                row_shift= 0;
-        }
-        return 0; /* the return value is ignored */
-    }
-    return 0; /* the return value is ignored */
+    old_size = fool_context[idx].fool_buf_size[type] * fool_context[idx].fool_buf_element[type];
+
+    if (old_size < new_size)
+        fool_context[idx].fool_buf[type] = realloc(fool_context[idx].fool_buf[type], new_size);
+    
+    fool_context[idx].fool_buf_size[type] = size;
+    fool_context[idx].fool_buf_element[type] = num_elements;
+    fool_context[idx].fool_buf_count[type]++;
+    /* because we ignore the vaRenderPicture, 
+     * all buffers with same type share same real memory
+     * bufferID = (magic number) | type
+     */
+    *buf_id = FOOL_BUFID_MAGIC | type;
+
+    return 1; /* don't call into driver */
 }
 
-VAStatus va_FoolCreateBuffer (
-        VADisplay dpy,
-        VAContextID context,	/* in */
-        VABufferType type,		/* in */
-        unsigned int size,		/* in */
-        unsigned int num_elements,	/* in */
-        void *data,			/* in */
-        VABufferID *buf_id		/* out */
+VAStatus va_FoolBufferInfo(
+    VADisplay dpy,
+    VABufferID buf_id,  /* in */
+    VABufferType *type, /* out */
+    unsigned int *size,         /* out */
+    unsigned int *num_elements /* out */
 )
 {
+    unsigned int magic = buf_id & FOOL_BUFID_MASK;
     DPY2INDEX(dpy);
 
-    if (FOOL_ENCODE(idx) || FOOL_DECODE(idx)) { /* fool buffer creation */
-        int new_size = size * num_elements;
-
-        if (type == VAEncCodedBufferType) /* only a VACodedBufferSegment */
-            new_size = sizeof(VACodedBufferSegment);
-
-        if (fool_context[idx].fool_buf_size[type] == 0)
-            fool_context[idx].fool_buf[type] = calloc(1, new_size);
-        else if (fool_context[idx].fool_buf_size[type] <= new_size)
-            fool_context[idx].fool_buf[type] = realloc(fool_context[idx].fool_buf, new_size);
-
-        if (fool_context[idx].fool_buf[type] == NULL) {
-            va_FoolEnd(dpy);
-            return 0; /* let driver go */
-        }
-
-        /* because we ignore the vaRenderPicture, 
-         * all buffers with same type share same real memory
-         * bufferID = (buffer count << 8) | type
-         */
-        fool_context[idx].fool_buf_count[type]++;
-        *buf_id = (fool_context[idx].fool_buf_count[type] << 8) | type;
-
-        return 1; /* don't call into driver */
-    }
+    if (magic != FOOL_BUFID_MAGIC)
+        return 0;
 
-    return 0; /* let driver go ... */
+    *type = buf_id & 0xff;
+    *size = fool_context[idx].fool_buf_size[*type];
+    *num_elements = fool_context[idx].fool_buf_element[*type];;
+    
+    return 1; /* don't call into driver */
 }
 
-VAStatus va_FoolMapBuffer (
-        VADisplay dpy,
-        VABufferID buf_id,	/* in */
-        void **pbuf 	/* out */
-)
+static int va_FoolFillCodedBufEnc(int idx)
 {
-    VABufferType type;
-    unsigned int size,frame_size = 0;
-    unsigned int num_elements;
-    DPY2INDEX(dpy);
-
-    if (FOOL_ENCODE(idx) || FOOL_DECODE(idx)) { /* fool buffer creation */
-        unsigned int buf_idx = buf_id & 0xff;
-
-        /* Image buffer? */
-        vaBufferInfo(dpy, fool_context[idx].context, buf_id, &type, &size, &num_elements);
-        if (type == VAImageBufferType  && FOOL_ENCODE(idx))
-            return 0;
-
-        /* buf_id is the buffer type */
-        if (fool_context[idx].fool_buf[buf_idx] != NULL)
-            *pbuf = fool_context[idx].fool_buf[buf_idx];
-        else
-            *pbuf = NULL;
-
-        /* expect APP to MapBuffer when get the the coded data */
-        if (*pbuf && (buf_idx == VAEncCodedBufferType)) { /* it is coded buffer */
-            /* read from a clip */
-            frame_size = va_FoolGetFrame(fool_context[idx].fool_fp_codedclip,
-                    fool_context[idx].frame_buf);
-
-            memset(fool_context[idx].codebuf,0,sizeof(VACodedBufferSegment));
-            fool_context[idx].codebuf->size = frame_size;
-            fool_context[idx].codebuf->bit_offset = 0;
-            fool_context[idx].codebuf->status = 0;
-            fool_context[idx].codebuf->reserved = 0;
-            fool_context[idx].codebuf->buf = fool_context[idx].frame_buf;
-            fool_context[idx].codebuf->next = NULL;
-            *pbuf = fool_context[idx].codebuf;
+    char file_name[1024];
+    struct stat file_stat;
+    VACodedBufferSegment *codedbuf;
+    int i, fd = -1;
+
+    /* try file_name.file_count, if fail, try file_name.file_count-- */
+    for (i=0; i<=1; i++) {
+        sprintf(file_name, "%s.%d",
+                fool_context[idx].fn_enc,
+                fool_context[idx].file_count);
+
+        if ((fd = open(file_name, O_RDONLY)) != -1) {
+            fstat(fd, &file_stat);
+            fool_context[idx].file_count++; /* open next file */
+            break;
         }
-        return 1; /* don't call into driver */
+        
+        fool_context[idx].file_count--; /* fall back to previous file */
+        if (fool_context[idx].file_count < 0)
+            fool_context[idx].file_count = 0;
     }
+    if (fd != -1) {
+        fool_context[idx].segbuf_enc = realloc(fool_context[idx].segbuf_enc, file_stat.st_size);
+        read(fd, fool_context[idx].segbuf_enc, file_stat.st_size);
+        close(fd);
+    }
+    codedbuf = (VACodedBufferSegment *)fool_context[idx].fool_buf[VAEncCodedBufferType];
+    codedbuf->size = file_stat.st_size;
+    codedbuf->bit_offset = 0;
+    codedbuf->status = 0;
+    codedbuf->reserved = 0;
+    codedbuf->buf = fool_context[idx].segbuf_enc;
+    codedbuf->next = NULL;
 
-    return 0; /* let driver go ... */
+    return 0;
 }
 
 
-int va_FoolBeginPicture(
-        VADisplay dpy,
-        VAContextID context,
-        VASurfaceID render_target
-)
+static int va_FoolFillCodedBufJPG(int idx)
 {
-    DPY2INDEX(dpy);
-
-    if (FOOL_ENCODE(idx) || FOOL_DECODE(idx)) {
-        if (fool_context[idx].context == 0)
-            fool_context[idx].context = context;
-        return 1; /* don't call into driver level */
+    struct stat file_stat;
+    VACodedBufferSegment *codedbuf;
+    int i, fd = -1;
+
+    if ((fd = open(fool_context[idx].fn_jpg, O_RDONLY)) != -1)
+        fstat(fd, &file_stat);
+        
+    if (fd != -1) {
+        fool_context[idx].segbuf_jpg = realloc(fool_context[idx].segbuf_jpg, file_stat.st_size);
+        read(fd, fool_context[idx].segbuf_jpg, file_stat.st_size);
+        close(fd);
     }
+    codedbuf = (VACodedBufferSegment *)fool_context[idx].fool_buf[VAEncCodedBufferType];
+    codedbuf->size = file_stat.st_size;
+    codedbuf->bit_offset = 0;
+    codedbuf->status = 0;
+    codedbuf->reserved = 0;
+    codedbuf->buf = fool_context[idx].segbuf_jpg;
+    codedbuf->next = NULL;
 
-    return 0; /* let driver go ... */
-}
-
-int va_FoolRenderPicture(
-        VADisplay dpy,
-        VAContextID context,
-        VABufferID *buffers,
-        int num_buffers
-)
-{
-    DPY2INDEX(dpy);
-
-    if (FOOL_ENCODE(idx) || FOOL_DECODE(idx))
-        return 1; /* don't call into driver level */
-
-    return 0;  /* let driver go ... */
+    return 0;
 }
 
 
-int va_FoolEndPicture(
-        VADisplay dpy,
-        VAContextID context
-)
+static int va_FoolFillCodedBuf(int idx)
 {
-    DPY2INDEX(dpy);
-
-    /* don't call into driver level */
-
-    /* do real fooling operation here */
-
-    /* only support H264 encoding currently */
-    if (FOOL_ENCODE(idx)) {
-        /* expect vaMapBuffer will handle it
-         * or else, need to save the codedbuf ID,
-         * and fool encode it here
-         */
-        /* va_FoolCodedBuf(dpy); */
-        return 1; /* don't call into driver level */
-    }
-
-    if (FOOL_DECODE(idx))
-        return 1;  /* don't call into driver level */
-
-    return 0; /* let driver go ... */
+    if (fool_context[idx].entrypoint == VAEntrypointEncSlice)
+        va_FoolFillCodedBufEnc(idx);
+    else if (fool_context[idx].entrypoint == VAEntrypointEncPicture)
+        va_FoolFillCodedBufJPG(idx);
+        
+    return 0;
 }
 
-int va_FoolSyncSurface(
-        VADisplay dpy, 
-        VASurfaceID render_target
+
+VAStatus va_FoolMapBuffer(
+    VADisplay dpy,
+    VABufferID buf_id,	/* in */
+    void **pbuf 	/* out */
 )
 {
+    unsigned int buftype = buf_id & 0xff;
+    unsigned int magic = buf_id & FOOL_BUFID_MASK;
     DPY2INDEX(dpy);
 
-    /*Fill in black and white squares. */
-    if (FOOL_DECODE(idx) || FOOL_DECODE(idx))
-        return 1;
+    if (magic != FOOL_BUFID_MAGIC)
+        return 0;
 
-    return 0;
+    /* buf_id is the buffer type */
+    *pbuf = fool_context[idx].fool_buf[buftype];
 
+    /* it is coded buffer, fill the fake segment buf from file */
+    if (*pbuf && (buftype == VAEncCodedBufferType))
+        va_FoolFillCodedBuf(idx);
+    
+    return 1; /* don't call into driver */
 }
 
 VAStatus va_FoolUnmapBuffer(
@@ -550,28 +358,10 @@ VAStatus va_FoolUnmapBuffer(
         VABufferID buf_id	/* in */
 )
 {
-    DPY2INDEX(dpy);
+    unsigned int magic = buf_id & FOOL_BUFID_MASK;
 
-    if (FOOL_ENCODE(idx) || FOOL_DECODE(idx))
-        return 1; /* fool buffer creation */
+    if (magic != FOOL_BUFID_MAGIC)
+        return 0;
 
-    return 0;
+    return 1;
 }
-
-VAStatus va_FoolQuerySubpictureFormats(
-        VADisplay dpy,
-        VAImageFormat *format_list,
-        unsigned int *flags,
-        unsigned int *num_formats
-)
-{
-    DPY2INDEX(dpy);
-
-    if (FOOL_ENCODE(idx) || FOOL_DECODE(idx)) { 
-        if (num_formats)
-            *num_formats = 0;
-        return 1;
-    }
-    return 0;
-}
-
diff --git a/va/va_fool.h b/va/va_fool.h
index ba11b50..5ea4830 100644
--- a/va/va_fool.h
+++ b/va/va_fool.h
@@ -28,35 +28,40 @@
 
 #include <stdio.h>
 
-void va_FoolInit(VADisplay dpy);
-
-int va_FoolEnd(VADisplay dpy);
+#ifdef __cplusplus
+extern "C" {
+#endif
 
+extern int fool_codec;
+extern int fool_postp;
 
-int va_FoolGetFrame(FILE *input_fp, char *frame_buf);
+#define VA_FOOL_FLAG_DECODE  0x1
+#define VA_FOOL_FLAG_ENCODE  0x2
+#define VA_FOOL_FLAG_JPEG    0x4
 
-int va_FoolCodedBuf(VADisplay dpy);
+#define VA_FOOL_FUNC(fool_func,...)            \
+    if (fool_codec) {                          \
+        ret = fool_func(__VA_ARGS__);          \
+    }
+#define VA_FOOL_RETURN()                       \
+    if (fool_codec) {                          \
+        return VA_STATUS_SUCCESS;              \
+    }
 
+void va_FoolInit(VADisplay dpy);
+int va_FoolEnd(VADisplay dpy);
 
 int va_FoolCreateConfig(
-    VADisplay dpy,
-    VAProfile profile, 
-    VAEntrypoint entrypoint, 
-    VAConfigAttrib *attrib_list,
-    int num_attribs,
-    VAConfigID *config_id /* out */
+        VADisplay dpy,
+        VAProfile profile, 
+        VAEntrypoint entrypoint, 
+        VAConfigAttrib *attrib_list,
+        int num_attribs,
+        VAConfigID *config_id /* out */
 );
 
-int va_FoolCreateSurfaces(
-    VADisplay dpy,
-    int width,
-    int height,
-    int format,
-    int num_surfaces,
-    VASurfaceID *surfaces	/* out */
-);
 
-VAStatus va_FoolCreateBuffer (
+VAStatus va_FoolCreateBuffer(
     VADisplay dpy,
     VAContextID context,	/* in */
     VABufferType type,		/* in */
@@ -72,41 +77,22 @@ VAStatus va_FoolMapBuffer (
     void **pbuf 	/* out */
 );
 
-int va_FoolBeginPicture(
-    VADisplay dpy,
-    VAContextID context,
-    VASurfaceID render_target
-);
-
-int va_FoolRenderPicture(
-    VADisplay dpy,
-    VAContextID context,
-    VABufferID *buffers,
-    int num_buffers
-);
-
-int va_FoolEndPicture(
-    VADisplay dpy,
-    VAContextID context
-);
-
-VAStatus va_FoolUnmapBuffer (
-    VADisplay dpy,
-    VABufferID buf_id  /* in */
+VAStatus va_FoolUnmapBuffer(
+        VADisplay dpy,
+        VABufferID buf_id	/* in */
 );
 
-
-VAStatus va_FoolQuerySubpictureFormats (
+VAStatus va_FoolBufferInfo (
     VADisplay dpy,
-    VAImageFormat *format_list,
-    unsigned int *flags,
-    unsigned int *num_formats
-);
-int va_FoolSyncSurface(
-    VADisplay dpy, 
-    VASurfaceID render_target
+    VABufferID buf_id,  /* in */
+    VABufferType *type, /* out */
+    unsigned int *size,         /* out */
+    unsigned int *num_elements /* out */
 );
-
-
+    
+    
+#ifdef __cplusplus
+}
+#endif
 
 #endif
diff --git a/va/va_fool_getframe.c b/va/va_fool_getframe.c
deleted file mode 100644
index 045bfb2..0000000
--- a/va/va_fool_getframe.c
+++ /dev/null
@@ -1,210 +0,0 @@
-/* The code refers to
- * http://keyj.s2000.at/files/projects/h264-src.tar.gz
- */
-#include <string.h>
-#include <stdio.h>
-
-#define SLICE_NUM 4 
-#define NAL_BUF_SIZE  65536  // maximum NAL unit size
-#define RING_BUF_SIZE  8192  // input ring buffer size, MUST be a power of two!
-
-typedef struct _nal_unit {
-  int NumBytesInNALunit;
-  int forbidden_zero_bit;
-  int nal_ref_idc;
-  int nal_unit_type;
-  unsigned char *last_rbsp_byte;
-} nal_unit;
-
- typedef struct _slice_header {
-  int first_mb_in_slice;
-} slice_header;
- 
-static int get_next_nal_unit(FILE *input_fp, nal_unit *nalu); 
-static int get_unsigned_exp_golomb();
-static void decode_slice_header(slice_header *sh);
-static void input_read(FILE *input_fp, unsigned char *dest, int size);
-static int input_get_bits(int bit_count);
-
-
-static unsigned char nal_buf[NAL_BUF_SIZE];
-static unsigned char ring_buf[RING_BUF_SIZE];
-static int input_remain = 0;
-static int ring_pos = 0;
-static int nal_pos;
-static int nal_bit;
-static int frame_no = 0;
-
-#define RING_MOD  ((RING_BUF_SIZE)-1)
-#define HALF_RING ((RING_BUF_SIZE)/2)
-
-#define gnn_advance() do { \
-	ring_pos = (ring_pos+1)&RING_MOD; \
-	--input_remain; \
-	if (ring_pos==0) input_read(input_fp, &ring_buf[HALF_RING],HALF_RING); \
-	if (ring_pos==HALF_RING) input_read(input_fp, &ring_buf[0],HALF_RING); \
-} while(0)
-
-#define gnn_add_segment(end) do { \
-	int size = end-segment_start; \
-	if (size>0) { \
-		memcpy(&nal_buf[nalu_size],&ring_buf[segment_start],size); \
-		nalu_size += size; \
-	} \
-	segment_start = end&RING_MOD; \
-} while(0)
-
-static int input_get_bits(int bit_count) 
-{
-    int res = 0;
-    register unsigned int x = 
-        (nal_buf[nal_pos]<<24)|
-        (nal_buf[nal_pos+1]<<16)|
-        (nal_buf[nal_pos+2]<<8)|
-        nal_buf[nal_pos+3];
-
-    res = (x>>(32-bit_count-nal_bit))&((1<<bit_count)-1);
-    nal_bit += bit_count;
-    nal_pos += nal_bit>>3;
-    nal_bit &= 7;
-
-    return res;
-}
-
-static int input_get_one_bit() 
-{
-    int res = (nal_buf[nal_pos]>>(7-nal_bit))&1;
-
-    if (++nal_bit>7) {
-        ++nal_pos;
-        nal_bit = 0;
-    }
-    return res;
-}
-
-static int get_unsigned_exp_golomb() 
-{
-    int exp;
-
-    for(exp = 0; !input_get_one_bit(); ++exp);
-    
-    if (exp)
-        return (1<<exp) - 1 + input_get_bits(exp);
-    else
-        return 0;
-}
-
-static void decode_slice_header(slice_header *sh ) 
-{
-    memset((void*)sh,0,sizeof(slice_header));
-    sh->first_mb_in_slice = get_unsigned_exp_golomb(); 
-}
-
-static void input_read(FILE *input_fp, unsigned char *dest, int size) 
-{
-    int count = fread(dest, 1, size, input_fp);
-
-    input_remain += count;
-}
-
-static int get_next_nal_unit(FILE *input_fp, nal_unit *nalu)
-{
-    int i,segment_start;
-    int nalu_size = 0;
-    int NumBytesInRbsp = 0;
-
-    /* search for the next NALU start
-     * here is the sync that the start of the NALU is 0x00000001
-     */
-    for (;;) {
-        if (input_remain<= 4) {
-            /* clip restart */
-            memset(ring_buf,0,sizeof(char)*RING_BUF_SIZE);
-            memset(nal_buf,0,sizeof(char)*NAL_BUF_SIZE);
-
-            fseek(input_fp,0,SEEK_SET);
-            input_remain = 0;
-            input_read(input_fp, ring_buf, RING_BUF_SIZE);
-            ring_pos = 0;
-            return 1;
-        }
-        if ((!ring_buf[ring_pos]) &&
-           (!ring_buf[(ring_pos+1)&RING_MOD]) &&
-           (!ring_buf[(ring_pos+2)&RING_MOD]) &&
-           ( ring_buf[(ring_pos+3)&RING_MOD]==1))
-            break;
-        gnn_advance();
-    }
-    
-    for(i=0;i<4;++i)
-        gnn_advance();
-
-    /* add bytes to the NALU until the end is found */
-    segment_start = ring_pos;
-    while (input_remain) {
-        if ((!ring_buf[ring_pos]) &&
-           (!ring_buf[(ring_pos+1)&RING_MOD]) &&
-           (!ring_buf[(ring_pos+2)&RING_MOD]))
-            break;
-        ring_pos = (ring_pos+1)&RING_MOD;
-        --input_remain;
-        
-        if (ring_pos==0) {
-            gnn_add_segment(RING_BUF_SIZE);
-            input_read(input_fp, &ring_buf[HALF_RING],HALF_RING);
-        }
-
-        if (ring_pos==HALF_RING) {
-            gnn_add_segment(HALF_RING);
-            input_read(input_fp, &ring_buf[0], HALF_RING);
-        }
-    }
-
-    gnn_add_segment(ring_pos);
-
-    /* read the NAL unit */
-    nal_pos = 0; nal_bit = 0;
-    nalu->forbidden_zero_bit = input_get_bits(1);
-    nalu->nal_ref_idc = input_get_bits(2);
-    nalu->nal_unit_type = input_get_bits(5);
-    nalu->last_rbsp_byte = &nal_buf[nalu_size-1];
-    nalu->NumBytesInNALunit = nalu_size; 
-
-    return 1;
-}
-
-int va_FoolGetFrame(FILE *input_fp, char *frame_buf) 
-{
-    int i = 0, frame_pos = 0;
-    static slice_header sh; 
-    static nal_unit nalu;
-    char nal_head[4] = {0x00,0x00,0x00,0x01};
-
-    /* read the clip , here is the first frame,
-     * &let the clip go on frame by frame
-     */
-    if (!frame_no)
-        input_read(input_fp, ring_buf,RING_BUF_SIZE);
-
-    while (get_next_nal_unit(input_fp, &nalu)) {
-        if (nalu.nal_unit_type == 7 || nalu.nal_unit_type == 8) {
-            memcpy(frame_buf+frame_pos, nal_head, sizeof(char)*4);
-		    frame_pos = frame_pos + 4;
-            memcpy(frame_buf+frame_pos, nal_buf, sizeof(char)*(nalu.NumBytesInNALunit));
-            frame_pos += nalu.NumBytesInNALunit;
-        }
-        else if (nalu.nal_unit_type == 1 || nalu.nal_unit_type == 5) {
-            decode_slice_header(&sh);
-            if (0 == sh.first_mb_in_slice) {
-                ++frame_no;
-            }
-            memcpy(frame_buf+frame_pos, nal_head, sizeof(char)*4);
-		    frame_pos = frame_pos + 4;
-            memcpy(frame_buf+frame_pos, nal_buf, sizeof(char)*(nalu.NumBytesInNALunit));
-            frame_pos += nalu.NumBytesInNALunit;
-		    break;
-        }
-    }
-    
-    return frame_pos; 
-}
diff --git a/va/va_trace.c b/va/va_trace.c
index b954cd7..b89cfce 100644
--- a/va/va_trace.c
+++ b/va/va_trace.c
@@ -40,29 +40,25 @@
 #include <time.h>
 #include <errno.h>
 
-
 /*
  * Env. to debug some issue, e.g. the decode/encode issue in a video conference scenerio:
  * .LIBVA_TRACE=log_file: general VA parameters saved into log_file
  * .LIBVA_TRACE_BUFDATA: dump VA buffer data into log_file (if not set, just calculate a checksum)
  * .LIBVA_TRACE_CODEDBUF=coded_clip_file: save the coded clip into file coded_clip_file
- * .LIBVA_TRACE_SURFACE=decoded_yuv_file: save the decoded YUV file decoded_yuv_file
+ * .LIBVA_TRACE_SURFACE=yuv_file: save surface YUV into file yuv_file. Use file name to determine
+ *                                decode/encode or jpeg surfaces
  * .LIBVA_TRACE_LOGSIZE=numeric number: truncate the log_file or coded_clip_file, or decoded_yuv_file
  *                                      when the size is bigger than the number
  */
 
-
 /* global settings */
 
 /* LIBVA_TRACE */
-unsigned int trace_flag = 0;
+int trace_flag = 0;
 
 /* LIBVA_TRACE_LOGSIZE */
 static unsigned int trace_logsize = 0xffffffff; /* truncate the log when the size is bigger than it */
 
-/* LIBVA_TRACE_BUFDATA */
-static unsigned int trace_buffer_data; /* dump buffer data or not */
-
 #define TRACE_CONTEXT_MAX 4
 /* per context settings */
 static struct _trace_context {
@@ -70,15 +66,15 @@ static struct _trace_context {
     
     /* LIBVA_TRACE */
     FILE *trace_fp_log; /* save the log into a file */
-    char trace_log_fn[1024]; /* file name */
+    char *trace_log_fn; /* file name */
     
     /* LIBVA_TRACE_CODEDBUF */
     FILE *trace_fp_codedbuf; /* save the encode result into a file */
-    char trace_codedbuf_fn[1024]; /* file name */
+    char *trace_codedbuf_fn; /* file name */
     
     /* LIBVA_TRACE_SURFACE */
     FILE *trace_fp_surface; /* save the surface YUV into a file */
-    char trace_surface_fn[1024]; /* file name */
+    char *trace_surface_fn; /* file name */
 
     VAContextID  trace_context; /* current context */
     
@@ -94,7 +90,7 @@ static struct _trace_context {
     unsigned int trace_frame_width; /* current frame width */
     unsigned int trace_frame_height; /* current frame height */
     unsigned int trace_sequence_start; /* get a new sequence for encoding or not */
-} trace_context[TRACE_CONTEXT_MAX] = { {0} }; /* trace five context at the same time */
+} trace_context[TRACE_CONTEXT_MAX]; /* trace five context at the same time */
 
 #define DPY2INDEX(dpy)                                  \
     int idx;                                            \
@@ -141,11 +137,22 @@ VAStatus vaUnlockSurface(VADisplay dpy,
                          VASurfaceID surface
                          );
 
+#define FILE_NAME_SUFFIX(env_value)                      \
+do {                                                    \
+    int tmp = strnlen(env_value, sizeof(env_value));    \
+    int left = sizeof(env_value) - tmp;                 \
+                                                        \
+    snprintf(env_value+tmp,                             \
+             left,                                      \
+             ".%04d.%05d",                              \
+             trace_index,                               \
+             suffix);                                   \
+} while (0)
 
 void va_TraceInit(VADisplay dpy)
 {
     char env_value[1024];
-    unsigned int suffix = 0xffff & ((unsigned int)time(NULL));
+    unsigned short suffix = 0xffff & ((unsigned int)time(NULL));
     int trace_index = 0;
     FILE *tmp;    
     
@@ -156,79 +163,59 @@ void va_TraceInit(VADisplay dpy)
     if (trace_index == TRACE_CONTEXT_MAX)
         return;
 
+    memset(&trace_context[trace_index], 0, sizeof(struct _trace_context));
     if (va_parseConfig("LIBVA_TRACE", &env_value[0]) == 0) {
-        trace_flag = 1;
-
-        /*Check if there is still room for suffix .%d.%d*/
-    if (strnlen(env_value, 1024) < (1024 - 8))
-        snprintf(env_value+strnlen(env_value, 1024),
-                 (1025 - 8 - strnlen(env_value, 1024)),
-                 ".%d.%d", trace_index, suffix);
-
+        FILE_NAME_SUFFIX(env_value);
+        trace_context[trace_index].trace_log_fn = strdup(env_value);
+        
         tmp = fopen(env_value, "w");
         if (tmp) {
             trace_context[trace_index].trace_fp_log = tmp;
-            strcpy(trace_context[trace_index].trace_log_fn, env_value);
-        } else {
+            va_infoMessage("LIBVA_TRACE is on, save log into %s\n", trace_context[trace_index].trace_log_fn);
+            trace_flag = VA_TRACE_FLAG_LOG;
+        } else
             va_errorMessage("Open file %s failed (%s)\n", env_value, strerror(errno));
-            trace_context[trace_index].trace_fp_log = stderr;
-            strcpy(trace_context[trace_index].trace_codedbuf_fn, "/dev/stderr");
-        }
-        va_infoMessage("LIBVA_TRACE is on, save log into %s\n", trace_context[trace_index].trace_log_fn);
     }
 
-    if (trace_flag == 0)
-        return;
-
     /* may re-get the global settings for multiple context */
     if (va_parseConfig("LIBVA_TRACE_LOGSIZE", &env_value[0]) == 0) {
         trace_logsize = atoi(env_value);
         va_infoMessage("LIBVA_TRACE_LOGSIZE is on, size is %d\n", trace_logsize);
     }
-    
 
-    if (va_parseConfig("LIBVA_TRACE_BUFDATA", NULL) == 0) {
-        trace_buffer_data = 1; /* dump buffer data */
+    if ((trace_flag & VA_TRACE_FLAG_LOG) && (va_parseConfig("LIBVA_TRACE_BUFDATA", NULL) == 0)) {
+        trace_flag |= VA_TRACE_FLAG_BUFDATA;
         va_infoMessage("LIBVA_TRACE_BUFDATA is on, dump buffer into log file\n");
     }
-    
 
     /* per-context setting */
     if (va_parseConfig("LIBVA_TRACE_CODEDBUF", &env_value[0]) == 0) {
-        if (strnlen(env_value, 1024) < (1024 - 8))
-            snprintf(env_value+strnlen(env_value, 1024),
-                     (1025 - 8 - strnlen(env_value, 1024)),
-                     ".%d.%d", trace_index, suffix);
-
-        tmp = fopen(env_value, "w");
-        
-        if (tmp) {
-            trace_context[trace_index].trace_fp_codedbuf = tmp;
-            strcpy(trace_context[trace_index].trace_codedbuf_fn, env_value);
-        } else {
-            va_errorMessage("Open file %s failed (%s)\n", env_value, strerror(errno));
-            trace_context[trace_index].trace_fp_codedbuf = stderr;
-            strcpy(trace_context[trace_index].trace_codedbuf_fn, "/dev/stderr");
-        }
-
-        va_infoMessage("LIBVA_TRACE_CODEDBUF is on, save coded clip into %s\n", trace_context[trace_index].trace_codedbuf_fn);
+        FILE_NAME_SUFFIX(env_value);
+        trace_context[trace_index].trace_codedbuf_fn = strdup(env_value);
+        va_infoMessage("LIBVA_TRACE_CODEDBUF is on, save codedbuf into log file %s\n",
+                       trace_context[trace_index].trace_codedbuf_fn);
+        trace_flag |= VA_TRACE_FLAG_CODEDBUF;
     }
 
     if (va_parseConfig("LIBVA_TRACE_SURFACE", &env_value[0]) == 0) {
-        sprintf(env_value+strnlen(env_value, 1024), ".%d.%d", trace_index, suffix);
-
-        tmp = fopen(env_value, "w");
+        FILE_NAME_SUFFIX(env_value);
+        trace_context[trace_index].trace_surface_fn = strdup(env_value);
         
-        if (tmp) {
-            trace_context[trace_index].trace_fp_surface = tmp;
-            strcpy(trace_context[trace_index].trace_surface_fn, env_value);
-        } else {
-            va_errorMessage("Open file %s failed (%s)\n", env_value, strerror(errno));
-            trace_context[trace_index].trace_fp_surface = stderr;
-            strcpy(trace_context[trace_index].trace_surface_fn, "/dev/stderr");
-        }
-
-        va_infoMessage("LIBVA_TRACE_SURFACE is on, save coded clip into %s\n", trace_context[trace_index].trace_surface_fn);
+        va_infoMessage("LIBVA_TRACE_SURFACE is on, save surface into %s\n",
+                       trace_context[trace_index].trace_surface_fn);
+
+        /* for surface data dump, it is time-consume, and may
+         * cause some side-effect, so only trace the needed surfaces
+         * to trace encode surface, set the trace file name to sth like *enc*
+         * to trace decode surface, set the trace file name to sth like *dec*
+         * if no dec/enc in file name, set both
+         */
+        if (strstr(env_value, "dec"))
+            trace_flag |= VA_TRACE_FLAG_SURFACE_DECODE;
+        if (strstr(env_value, "enc"))
+            trace_flag |= VA_TRACE_FLAG_SURFACE_ENCODE;
+        if (strstr(env_value, "jpeg") || strstr(env_value, "jpg"))
+            trace_flag |= VA_TRACE_FLAG_SURFACE_JPEG;
     }
 
     trace_context[trace_index].dpy = dpy;
@@ -239,15 +226,24 @@ void va_TraceEnd(VADisplay dpy)
 {
     DPY2INDEX(dpy);
     
-    if (trace_context[idx].trace_fp_log && (trace_context[idx].trace_fp_log != stderr))
+    if (trace_context[idx].trace_fp_log)
         fclose(trace_context[idx].trace_fp_log);
     
-    if (trace_context[idx].trace_fp_codedbuf && (trace_context[idx].trace_fp_codedbuf != stderr))
+    if (trace_context[idx].trace_fp_codedbuf)
         fclose(trace_context[idx].trace_fp_codedbuf);
     
-    if (trace_context[idx].trace_fp_surface && (trace_context[idx].trace_fp_surface != stderr))
+    if (trace_context[idx].trace_fp_surface)
         fclose(trace_context[idx].trace_fp_surface);
 
+    if (trace_context[idx].trace_log_fn)
+        free(trace_context[idx].trace_log_fn);
+    
+    if (trace_context[idx].trace_codedbuf_fn)
+        free(trace_context[idx].trace_codedbuf_fn);
+    
+    if (trace_context[idx].trace_surface_fn)
+        free(trace_context[idx].trace_surface_fn);
+    
     memset(&trace_context[idx], 0, sizeof(struct _trace_context));
 }
 
@@ -272,9 +268,11 @@ void va_TraceMsg(int idx, const char *msg, ...)
 {
     va_list args;
 
+    if (!(trace_flag & VA_TRACE_FLAG_LOG))
+        return;
+
     if (file_size(trace_context[idx].trace_fp_log) >= trace_logsize)
         truncate_file(trace_context[idx].trace_fp_log);
-
     if (msg)  {
         va_start(args, msg);
         vfprintf(trace_context[idx].trace_fp_log, msg, args);
@@ -341,7 +339,7 @@ void va_TraceSurface(VADisplay dpy)
     VAStatus va_status;
     unsigned char check_sum = 0;
     DPY2INDEX(dpy);
-    
+
     va_TraceMsg(idx, "==========dump surface data in file %s\n", trace_context[idx].trace_surface_fn);
 
     if ((file_size(trace_context[idx].trace_fp_surface) >= trace_logsize)) {
@@ -388,21 +386,14 @@ void va_TraceSurface(VADisplay dpy)
 
     tmp = Y_data;
     for (i=0; i<trace_context[idx].trace_frame_height; i++) {
-        for (j=0; j<trace_context[idx].trace_frame_width; j++)
-            check_sum ^= tmp[j];
-
         if (trace_context[idx].trace_fp_surface)
             fwrite(tmp, trace_context[idx].trace_frame_width, 1, trace_context[idx].trace_fp_surface);
         
         tmp = Y_data + i * luma_stride;
     }
-
     tmp = UV_data;
     if (fourcc == VA_FOURCC_NV12) {
         for (i=0; i<trace_context[idx].trace_frame_height/2; i++) {
-            for (j=0; j<trace_context[idx].trace_frame_width; j++)
-                check_sum ^= tmp[j];
-            
             if (trace_context[idx].trace_fp_surface)
                 fwrite(tmp, trace_context[idx].trace_frame_width, 1, trace_context[idx].trace_fp_surface);
             
@@ -412,7 +403,6 @@ void va_TraceSurface(VADisplay dpy)
 
     vaUnlockSurface(dpy, trace_context[idx].trace_rendertarget);
 
-    va_TraceMsg(idx, "\tchecksum = 0x%02x\n", check_sum & 0xff);
     va_TraceMsg(idx, NULL);
 }
 
@@ -446,6 +436,7 @@ void va_TraceCreateConfig(
 )
 {
     int i;
+    int encode, decode, jpeg;
     DPY2INDEX(dpy);
 
     TRACE_FUNCNAME(idx);
@@ -461,6 +452,40 @@ void va_TraceCreateConfig(
 
     trace_context[idx].trace_profile = profile;
     trace_context[idx].trace_entrypoint = entrypoint;
+
+    /* avoid to create so many empty files */
+    encode = (trace_context[idx].trace_entrypoint == VAEntrypointEncSlice);
+    decode = (trace_context[idx].trace_entrypoint == VAEntrypointVLD);
+    jpeg = (trace_context[idx].trace_entrypoint == VAEntrypointEncPicture);
+    if ((encode && (trace_flag & VA_TRACE_FLAG_SURFACE_ENCODE)) ||
+        (decode && (trace_flag & VA_TRACE_FLAG_SURFACE_DECODE)) ||
+        (jpeg && (trace_flag & VA_TRACE_FLAG_SURFACE_JPEG))) {
+        FILE *tmp = fopen(trace_context[idx].trace_surface_fn, "w");
+        
+        if (tmp)
+            trace_context[idx].trace_fp_surface = tmp;
+        else {
+            va_errorMessage("Open file %s failed (%s)\n",
+                            trace_context[idx].trace_surface_fn,
+                            strerror(errno));
+            trace_context[idx].trace_fp_surface = NULL;
+            trace_flag &= ~(VA_TRACE_FLAG_SURFACE);
+        }
+    }
+
+    if (encode && (trace_flag & VA_TRACE_FLAG_CODEDBUF)) {
+        FILE *tmp = fopen(trace_context[idx].trace_codedbuf_fn, "w");
+        
+        if (tmp)
+            trace_context[idx].trace_fp_codedbuf = tmp;
+        else {
+            va_errorMessage("Open file %s failed (%s)\n",
+                            trace_context[idx].trace_codedbuf_fn,
+                            strerror(errno));
+            trace_context[idx].trace_fp_codedbuf = NULL;
+            trace_flag &= ~VA_TRACE_FLAG_CODEDBUF;
+        }
+    }
 }
 
 
@@ -610,10 +635,10 @@ static void va_TraceVABuffers(
     for (i=0; i<size; i++) {
         unsigned char value =  p[i];
             
-        if ((trace_buffer_data) && ((i%16) == 0))
+        if ((trace_flag & VA_TRACE_FLAG_BUFDATA) && ((i%16) == 0))
             va_TraceMsg(idx, "\n0x%08x:", i);
 
-        if (trace_buffer_data)
+        if (trace_flag & VA_TRACE_FLAG_BUFDATA)
             va_TraceMsg(idx, " %02x", value);
 
         check_sum ^= value;
@@ -1901,7 +1926,7 @@ void va_TraceRenderPicture(
     va_TraceMsg(idx, "\tcontext = 0x%08x\n", context);
     va_TraceMsg(idx, "\tnum_buffers = %d\n", num_buffers);
     for (i = 0; i < num_buffers; i++) {
-        void *pbuf;
+        unsigned char *pbuf;
         unsigned int j;
         
         /* get buffer type information */
@@ -1913,7 +1938,7 @@ void va_TraceRenderPicture(
         va_TraceMsg(idx, "\t  size = %d\n", size);
         va_TraceMsg(idx, "\t  num_elements = %d\n", num_elements);
 
-        vaMapBuffer(dpy, buffers[i], &pbuf);
+        vaMapBuffer(dpy, buffers[i], (void **)&pbuf);
 
         switch (trace_context[idx].trace_profile) {
         case VAProfileMPEG2Simple:
@@ -1980,36 +2005,46 @@ void va_TraceRenderPicture(
     va_TraceMsg(idx, NULL);
 }
 
-
 void va_TraceEndPicture(
     VADisplay dpy,
-    VAContextID context
+    VAContextID context,
+    int endpic_done
 )
 {
+    int encode, decode, jpeg;
     DPY2INDEX(dpy);
 
     TRACE_FUNCNAME(idx);
-    
-    va_TraceMsg(idx, "\tcontext = 0x%08x\n", context);
-    va_TraceMsg(idx, "\trender_targets = 0x%08x\n", trace_context[idx].trace_rendertarget);
 
-    /* want to trace codedbuf, and it is encode */
-    if (trace_context[idx].trace_fp_codedbuf &&
-        ((trace_context[idx].trace_entrypoint == VAEntrypointEncSlice) ||
-         (trace_context[idx].trace_entrypoint == VAEntrypointEncPicture))) {
+    if (endpic_done == 0) {
+        va_TraceMsg(idx, "\tcontext = 0x%08x\n", context);
+        va_TraceMsg(idx, "\trender_targets = 0x%08x\n", trace_context[idx].trace_rendertarget);
+    }
+
+    encode = (trace_context[idx].trace_entrypoint == VAEntrypointEncSlice) &&
+        (trace_flag & VA_TRACE_FLAG_SURFACE_ENCODE);
+    decode = (trace_context[idx].trace_entrypoint == VAEntrypointVLD) &&
+        (trace_flag & VA_TRACE_FLAG_SURFACE_DECODE);
+    jpeg = (trace_context[idx].trace_entrypoint == VAEntrypointEncPicture) &&
+        (trace_flag & VA_TRACE_FLAG_SURFACE_JPEG);
+    
+    /* want to trace encode source surface, do it before vaEndPicture */
+    if ((encode || jpeg) && (endpic_done == 0))
+        va_TraceSurface(dpy);
+    
+    /* want to trace encoode codedbuf, do it after vaEndPicture */
+    if ((encode || jpeg) && (endpic_done == 1)) {
         /* force the pipleline finish rendering */
         vaSyncSurface(dpy, trace_context[idx].trace_rendertarget);
         va_TraceCodedBuf(dpy);
     }
 
-    /* trace decoded surface for decoding, or the source sourface for encoding */
-    if (trace_context[idx].trace_fp_surface) {
+    /* want to trace decode dest surface, do it after vaEndPicture */
+    if (decode && (endpic_done == 1)) {
         /* force the pipleline finish rendering */
         vaSyncSurface(dpy, trace_context[idx].trace_rendertarget);
-        
         va_TraceSurface(dpy);
     }
-
     va_TraceMsg(idx, NULL);
 }
 
diff --git a/va/va_trace.h b/va/va_trace.h
index e8ae4f7..43b9fb8 100644
--- a/va/va_trace.h
+++ b/va/va_trace.h
@@ -25,6 +25,35 @@
 #ifndef VA_TRACE_H
 #define VA_TRACE_H
 
+#ifdef __cplusplus
+extern "C" {
+#endif
+
+extern int trace_flag;
+
+#define VA_TRACE_FLAG_LOG             0x1
+#define VA_TRACE_FLAG_BUFDATA         0x2
+#define VA_TRACE_FLAG_CODEDBUF        0x4
+#define VA_TRACE_FLAG_SURFACE_DECODE  0x8
+#define VA_TRACE_FLAG_SURFACE_ENCODE  0x10
+#define VA_TRACE_FLAG_SURFACE_JPEG    0x20
+#define VA_TRACE_FLAG_SURFACE         (VA_TRACE_FLAG_SURFACE_DECODE | \
+                                       VA_TRACE_FLAG_SURFACE_ENCODE | \
+                                       VA_TRACE_FLAG_SURFACE_JPEG)
+
+#define VA_TRACE_FUNC(trace_func,...)           \
+    if (trace_flag) {                           \
+        trace_func(__VA_ARGS__);                \
+    }
+#define VA_TRACE_LOG(trace_func,...)            \
+    if (trace_flag & VA_TRACE_FLAG_LOG) {            \
+        trace_func(__VA_ARGS__);                \
+    }
+#define VA_TRACE_SURFACE(trace_func,...)        \
+    if (trace_flag & (VA_TRACE_FLAG_SURFACE | VA_TRACE_FLAG_CODEDBUF)) {  \
+        trace_func(__VA_ARGS__);                \
+    }
+
 void va_TraceInit(VADisplay dpy);
 void va_TraceEnd(VADisplay dpy);
 
@@ -92,7 +121,8 @@ void va_TraceRenderPicture(
 
 void va_TraceEndPicture(
     VADisplay dpy,
-    VAContextID context
+    VAContextID context,
+    int endpic_done
 );
 
 void va_TraceSyncSurface(
@@ -155,6 +185,9 @@ void va_TracePutSurface (
     unsigned int flags /* de-interlacing flags */
 );
 
-
+#ifdef __cplusplus
+}
+#endif
+    
 
 #endif /* VA_TRACE_H */
diff --git a/va/x11/va_x11.c b/va/x11/va_x11.c
index e44dfa7..86b040a 100644
--- a/va/x11/va_x11.c
+++ b/va/x11/va_x11.c
@@ -26,6 +26,8 @@
 #include "sysdeps.h"
 #include "va.h"
 #include "va_backend.h"
+#include "va_trace.h"
+#include "va_fool.h"
 #include "va_x11.h"
 #include "va_dri.h"
 #include "va_dri2.h"
@@ -208,13 +210,6 @@ VADisplay vaGetDisplay (
 #define CTX(dpy) (((VADisplayContextP)dpy)->pDriverContext)
 #define CHECK_DISPLAY(dpy) if( !vaDisplayIsValid(dpy) ) { return VA_STATUS_ERROR_INVALID_DISPLAY; }
 
-extern int fool_postp; /* do nothing for vaPutSurface if set */
-extern int trace_flag; /* trace vaPutSurface parameters */
-#define VA_TRACE(trace_func,...)                \
-    if (trace_flag) {                           \
-        trace_func(__VA_ARGS__);                \
-    }
-
 void va_TracePutSurface (
     VADisplay dpy,
     VASurfaceID surface,
@@ -258,8 +253,8 @@ VAStatus vaPutSurface (
   CHECK_DISPLAY(dpy);
   ctx = CTX(dpy);
   
-  VA_TRACE(va_TracePutSurface, dpy, surface, (void *)draw, srcx, srcy, srcw, srch,
-           destx, desty, destw, desth,
+  VA_TRACE_FUNC(va_TracePutSurface, dpy, surface, (void *)draw, srcx, srcy, srcw, srch,
+                destx, desty, destw, desth,
            cliprects, number_cliprects, flags );
   
   return ctx->vtable->vaPutSurface( ctx, surface, (void *)draw, srcx, srcy, srcw, srch,