[Libva] [PATCH Intel-FEI 2/2] Add two FEI samples to test/encode/...

Hu, Kelvin kelvin.hu at intel.com
Mon May 5 00:10:19 PDT 2014


>From c20c45b5a7a798d5727323a0b87c3f9b4e453394 Mon Sep 17 00:00:00 2001
From: Kelvin Hu <kelvin.hu at intel.com>
Date: Mon, 5 May 2014 22:35:32 +0800
Subject: [PATCH 2/2] Add two FEI samples to test/encode/...

The new preenc.c sample for statistics case and feiavcenc.c sample
for fei case
---
 test/encode/Makefile.am |   16 +-
 test/encode/feiavcenc.c | 2209 +++++++++++++++++++++++++++++++++++++++++++++++
 test/encode/preenc.c    |  895 +++++++++++++++++++
 3 files changed, 3119 insertions(+), 1 deletion(-)
 create mode 100644 test/encode/feiavcenc.c
 create mode 100644 test/encode/preenc.c

diff --git a/test/encode/Makefile.am b/test/encode/Makefile.am
index 5ddabd1..a5c05d8 100644
--- a/test/encode/Makefile.am
+++ b/test/encode/Makefile.am
@@ -20,7 +20,7 @@
 # TORT OR OTHERWISE, ARISING FROM, OUT OF OR IN CONNECTION WITH THE
 # SOFTWARE OR THE USE OR OTHER DEALINGS IN THE SOFTWARE.
 
-bin_PROGRAMS = avcenc mpeg2vaenc h264encode
+bin_PROGRAMS = avcenc feiavcenc preenc mpeg2vaenc h264encode
 
 INCLUDES = \
        -Wall                           \
@@ -42,6 +42,20 @@ avcenc_LDADD		= \
 	$(top_builddir)/test/common/libva-display.la \
 	-lpthread
 
+feiavcenc_SOURCES	= feiavcenc.c
+feiavcenc_CFLAGS	= -I$(top_srcdir)/test/common -g
+feiavcenc_LDADD		= \
+	$(top_builddir)/va/libva.la \
+	$(top_builddir)/test/common/libva-display.la \
+	-lpthread
+
+preenc_SOURCES		= preenc.c
+preenc_CFLAGS		= -I$(top_srcdir)/test/common -g
+preenc_LDADD		= \
+	$(top_builddir)/va/libva.la \
+	$(top_builddir)/test/common/libva-display.la \
+	-lpthread
+
 mpeg2vaenc_SOURCES	= mpeg2vaenc.c
 mpeg2vaenc_CFLAGS	= -I$(top_srcdir)/test/common
 mpeg2vaenc_LDADD	= \
diff --git a/test/encode/feiavcenc.c b/test/encode/feiavcenc.c
new file mode 100644
index 0000000..3edb0bd
--- /dev/null
+++ b/test/encode/feiavcenc.c
@@ -0,0 +1,2209 @@
+/*
+ * Simple AVC encoder based on libVA.
+ *
+ * Usage:
+ * ./avcenc <width> <height> <input file> <output file> [qp]
+ */  
+
+#include <stdio.h>
+#include <string.h>
+#include <stdlib.h>
+#include <getopt.h>
+#include <X11/Xlib.h>
+
+#include <unistd.h>
+
+#include <sys/time.h>
+#include <sys/types.h>
+#include <sys/stat.h>
+#include <fcntl.h>
+#include <assert.h>
+#include <time.h>
+
+#include <pthread.h>
+
+#include <va/va.h>
+#include <va/va_enc_h264.h>
+#include <va/vendor/intel/va_intel_fei.h>
+#include <va/vendor/intel/va_intel_statistics.h>
+#include <va/va_x11.h>
+# include "va/drm/va_drm.h"
+
+#define NAL_REF_IDC_NONE        0
+#define NAL_REF_IDC_LOW         1
+#define NAL_REF_IDC_MEDIUM      2
+#define NAL_REF_IDC_HIGH        3
+
+#define NAL_NON_IDR             1
+#define NAL_IDR                 5
+#define NAL_SPS                 7
+#define NAL_PPS                 8
+
+
+#define SLICE_TYPE_P            0
+#define SLICE_TYPE_B            1
+#define SLICE_TYPE_I            2
+
+#define ENTROPY_MODE_CAVLC      0
+#define ENTROPY_MODE_CABAC      1
+
+#define PROFILE_IDC_BASELINE    66
+#define PROFILE_IDC_MAIN        77
+#define PROFILE_IDC_HIGH        100
+
+#define CHECK_VASTATUS(va_status,func)                                  \
+    if (va_status != VA_STATUS_SUCCESS) {                               \
+        fprintf(stderr,"%s:%s (%d) failed,exit\n", __func__, func, __LINE__); \
+        exit(1);                                                        \
+    }
+
+static Display *x11_display;
+static VADisplay va_dpy;
+
+static int picture_width, picture_width_in_mbs;
+static int picture_height, picture_height_in_mbs;
+static int frame_size;
+static unsigned char *newImageBuffer = 0;
+static VAEncFEIMBControlH264Intel *feiMbCtrlBuffer = 0;
+static VAEncMVPredictorH264Intel  *feiMvPredBuffer = 0;
+
+static int qp_value = 26;
+static Bool stats_en = 0;
+
+static int intra_period = 30;
+static int ip_period = 1;
+static int pb_period = 5;
+static int frame_bit_rate = -1;
+
+static int next_input_id = -1;
+#define MAX_SLICES      32
+
+signed int Cur_POC[2];
+signed int ref0_POC[2];
+signed int ref1_POC[2];
+static int
+build_packed_pic_buffer(unsigned char **header_buffer);
+
+static int
+build_packed_seq_buffer(unsigned char **header_buffer);
+
+int 
+build_packed_slc_buffer(unsigned char **header_buffer, int frame_num, int display_frame, int slice_type, int is_idr);
+
+
+static int avcenc_destroy_buffers(VABufferID *va_buffers, unsigned int num_va_buffers);
+
+struct upload_thread_param
+{
+    FILE *yuv_fp;
+    VASurfaceID surface_id;
+};
+
+static void 
+upload_yuv_to_surface(FILE *yuv_fp, VASurfaceID surface_id);
+
+static struct {
+    VAProfile profile;
+    VAEncSequenceParameterBufferH264 seq_param;
+    VAEncPictureParameterBufferH264 pic_param;
+    VAEncSliceParameterBufferH264 slice_param[MAX_SLICES];
+    VAContextID context_id;
+    VAConfigID config_id;
+    VABufferID seq_param_buf_id;                /* Sequence level parameter */
+    VABufferID pic_param_buf_id;                /* Picture level parameter */
+    VABufferID slice_param_buf_id[MAX_SLICES];  /* Slice level parameter, multil slices */
+    VABufferID codedbuf_buf_id;                 /* Output buffer, compressed data */
+    VABufferID fei_distortion_buf_id;           /* Output buffer, distortion data */
+    VABufferID fei_mv_buf_id;                   /* Output buffer, mv data */
+    VABufferID fei_mbcode_buf_id;               /* Output buffer, mb code data */
+    VABufferID packed_seq_header_param_buf_id;
+    VABufferID packed_seq_buf_id;
+    VABufferID packed_pic_header_param_buf_id;
+    VABufferID packed_pic_buf_id;
+    VABufferID packed_slc_header_param_buf_id;
+    VABufferID packed_slc_buf_id;
+    
+    VABufferID misc_parameter_hrd_buf_id;
+    VABufferID misc_parameter_rc_buf_id;
+    VABufferID misc_parameter_fr_buf_id;
+    // FEI buffer
+    VAEncMiscParameterFEIFrameControlH264Intel fei_pic_param;
+    VABufferID misc_parameter_fei_pic_control_buf_id;
+    VABufferID fei_mbctrl_buf_id;
+    VABufferID fei_mvpred_buf_id;
+    int fei_mbctrl_buf_size;
+    int fei_mvpred_buf_size;
+
+    int num_slices;
+    int codedbuf_i_size;
+    int codedbuf_pb_size;
+    int current_input_surface;
+    int rate_control_method;
+    struct upload_thread_param upload_thread_param;
+    pthread_t upload_thread_id;
+    int upload_thread_value;
+} avcenc_context;
+
+static int drm_fd = -1;
+
+static void create_encode_pipe()
+{
+    VAEntrypoint entrypoints[5];
+    int num_entrypoints,slice_entrypoint;
+    VAConfigAttrib attrib[2];
+    int major_ver, minor_ver;
+    VAStatus va_status;
+    
+    drm_fd = open("/dev/dri/card0", O_RDWR);    
+    if (drm_fd < 0) 
+    {        
+        fprintf(stderr, "error: can't open DRM connection!\n");        
+        return NULL;    
+    }    
+    va_dpy = vaGetDisplayDRM(drm_fd);
+    //x11_display = XOpenDisplay(":0.0");
+    //assert(x11_display);
+    
+    //va_dpy = vaGetDisplay(x11_display);
+    va_status = vaInitialize(va_dpy, &major_ver, &minor_ver);
+    CHECK_VASTATUS(va_status, "vaInitialize");
+
+    vaQueryConfigEntrypoints(va_dpy, avcenc_context.profile, entrypoints, 
+                             &num_entrypoints);
+
+    for	(slice_entrypoint = 0; slice_entrypoint < num_entrypoints; slice_entrypoint++) {
+        if (entrypoints[slice_entrypoint] == VAEntrypointEncFEIIntel)
+            break;
+    }
+
+    if (slice_entrypoint == num_entrypoints) {
+        /* not find Slice entry point */
+        assert(0);
+    }
+
+    /* find out the format for the render target, and rate control mode */
+    attrib[0].type = VAConfigAttribRTFormat;
+    attrib[1].type = VAConfigAttribRateControl;
+    attrib[2].type = VAConfigAttribEncFunctionTypeIntel;
+    attrib[3].type = VAConfigAttribEncMVPredictorsIntel;
+    vaGetConfigAttributes(va_dpy, avcenc_context.profile, VAEntrypointEncFEIIntel,
+                          &attrib[0], 4);
+
+    if ((attrib[0].value & VA_RT_FORMAT_YUV420) == 0) {
+        /* not find desired YUV420 RT format */
+        assert(0);
+    }
+
+    if ((attrib[1].value & avcenc_context.rate_control_method) == 0) {
+        /* Can't find matched RC mode */
+        printf("Can't find the desired RC mode, exit\n");
+        assert(0);
+    }
+
+    attrib[0].value = VA_RT_FORMAT_YUV420; /* set to desired RT format */
+    attrib[1].value = avcenc_context.rate_control_method; /* set to desired RC mode */
+    attrib[2].value = VA_ENC_FUNCTION_ENC_PAK_INTEL; /* set FEI function to ENC_PAK */
+    attrib[3].value = 1; /* set 0 MV Predictor */
+
+    va_status = vaCreateConfig(va_dpy, avcenc_context.profile, VAEntrypointEncFEIIntel,
+                               &attrib[0], 4, &avcenc_context.config_id);
+    CHECK_VASTATUS(va_status, "vaCreateConfig");
+
+    
+}
+
+static void destory_encode_pipe()
+{
+    vaDestroyContext(va_dpy,avcenc_context.context_id);
+    vaDestroyConfig(va_dpy,avcenc_context.config_id);
+    vaTerminate(va_dpy);
+    if (drm_fd < 0)        
+        return;    
+    close(drm_fd);    
+    drm_fd = -1;
+    //XCloseDisplay(x11_display);
+}
+
+/***************************************************
+ *
+ *  The encode pipe resource define 
+ *
+ ***************************************************/
+#define SID_INPUT_PICTURE_0                     0
+#define SID_INPUT_PICTURE_1                     1
+#define SID_INPUT_PICTURE_2                     5   //Bo
+#define SID_INPUT_PICTURE_3                     6
+#define SID_INPUT_PICTURE_4                     7
+
+#define SID_REFERENCE_PICTURE_L0                2
+#define SID_REFERENCE_PICTURE_L1                3
+#define SID_RECON_PICTURE                       4
+#define SID_NUMBER                              SID_RECON_PICTURE + 1 + 3
+static  VASurfaceID surface_ids[SID_NUMBER];
+
+static int frame_number;
+static int enc_frame_number;
+
+/***************************************************/
+
+static void *
+upload_thread_function(void *data)
+{
+    struct upload_thread_param *param = data;
+
+    upload_yuv_to_surface(param->yuv_fp, param->surface_id);
+
+    return NULL;
+}
+
+static void alloc_encode_resource(FILE *yuv_fp)
+{
+    VAStatus va_status;
+
+    // Create surface
+    va_status = vaCreateSurfaces(
+        va_dpy,
+        VA_RT_FORMAT_YUV420, picture_width, picture_height,
+        &surface_ids[0], SID_NUMBER,
+        NULL, 0
+    );
+
+    CHECK_VASTATUS(va_status, "vaCreateSurfaces");
+
+    /* Create a context for this decode pipe */
+    va_status = vaCreateContext(va_dpy, avcenc_context.config_id,
+                                picture_width, picture_height,
+                                VA_PROGRESSIVE, 
+                                &surface_ids[0], 
+                                SID_NUMBER,
+                                &avcenc_context.context_id);
+    CHECK_VASTATUS(va_status, "vaCreateContext");
+    newImageBuffer  = (unsigned char *)malloc(frame_size);
+    feiMbCtrlBuffer = (VAEncFEIMBControlH264Intel *)malloc(avcenc_context.fei_mbctrl_buf_size);
+    feiMvPredBuffer = (VAEncMVPredictorH264Intel  *)malloc(avcenc_context.fei_mvpred_buf_size);
+
+    /* firstly upload YUV data to SID_INPUT_PICTURE_1 */
+    avcenc_context.upload_thread_param.yuv_fp = yuv_fp;
+    avcenc_context.upload_thread_param.surface_id = surface_ids[SID_INPUT_PICTURE_1];
+    next_input_id = SID_INPUT_PICTURE_1;
+#if 0
+    avcenc_context.upload_thread_value = pthread_create(&avcenc_context.upload_thread_id,
+                                                        NULL,
+                                                        upload_thread_function, 
+                                                        (void*)&avcenc_context.upload_thread_param);
+//#else
+    upload_thread_function((void *)(&avcenc_context.upload_thread_param));
+#endif
+}
+
+static void release_encode_resource()
+{
+#if 0
+    pthread_join(avcenc_context.upload_thread_id, NULL);
+#endif
+    avcenc_destroy_buffers(&avcenc_context.packed_seq_header_param_buf_id, 1);
+    avcenc_destroy_buffers(&avcenc_context.packed_seq_buf_id, 1);
+    //avcenc_destroy_buffers(&avcenc_context.packed_pic_header_param_buf_id, 1);
+    //avcenc_destroy_buffers(&avcenc_context.packed_pic_buf_id, 1);
+    //avcenc_destroy_buffers(&avcenc_context.packed_slc_header_param_buf_id, 1);
+    //avcenc_destroy_buffers(&avcenc_context.packed_slc_buf_id, 1);
+    
+    free(newImageBuffer);
+    free(feiMbCtrlBuffer);
+    free(feiMvPredBuffer);
+
+    // Release all the surfaces resource
+    vaDestroySurfaces(va_dpy, &surface_ids[0], SID_NUMBER);	
+}
+
+static void avcenc_update_picture_parameter(int slice_type, int frame_num, int display_num, int is_idr)
+{
+    VAEncPictureParameterBufferH264 *pic_param;
+    VAStatus va_status;
+    int i;
+    // Picture level
+    unsigned int length_in_bits, offset_in_bytes;
+    VAEncPackedHeaderParameterBuffer packed_header_param_buffer;
+    unsigned char *packed_pic_buffer = NULL;
+    pic_param = &avcenc_context.pic_param;
+    pic_param->CurrPic.picture_id = surface_ids[SID_RECON_PICTURE];
+    pic_param->CurrPic.flags = VA_PICTURE_H264_SHORT_TERM_REFERENCE;
+    pic_param->CurrPic.TopFieldOrderCnt = display_num * 2;
+    pic_param->CurrPic.BottomFieldOrderCnt = display_num * 2+1;
+    Cur_POC[0] = display_num * 2;
+    Cur_POC[1] = display_num * 2 + 1;
+    if (slice_type == SLICE_TYPE_I)
+    {
+        for (i=0; i<16; i++)
+        {
+            pic_param->ReferenceFrames[i].picture_id = VA_INVALID_ID;
+            pic_param->ReferenceFrames[i].flags = VA_PICTURE_H264_INVALID;
+        }
+    }
+    else
+    {
+        pic_param->ReferenceFrames[0].picture_id = surface_ids[SID_REFERENCE_PICTURE_L0];
+        pic_param->ReferenceFrames[0].flags = VA_PICTURE_H264_SHORT_TERM_REFERENCE;
+        pic_param->ReferenceFrames[0].TopFieldOrderCnt = ref0_POC[0];
+        pic_param->ReferenceFrames[0].BottomFieldOrderCnt = ref0_POC[1];
+        if (slice_type != SLICE_TYPE_B)
+        {
+            pic_param->ReferenceFrames[1].picture_id = VA_INVALID_ID;//surface_ids[SID_REFERENCE_PICTURE_L0]; 
+            pic_param->ReferenceFrames[1].flags = VA_PICTURE_H264_INVALID;
+        }
+        else
+        {
+            pic_param->ReferenceFrames[1].picture_id = surface_ids[SID_REFERENCE_PICTURE_L1];
+            pic_param->ReferenceFrames[1].flags = VA_PICTURE_H264_SHORT_TERM_REFERENCE;
+            pic_param->ReferenceFrames[1].TopFieldOrderCnt = ref1_POC[0];
+            pic_param->ReferenceFrames[1].BottomFieldOrderCnt = ref1_POC[1];
+        }
+        for (i=2; i<16; i++)
+        {
+            pic_param->ReferenceFrames[i].picture_id = VA_INVALID_ID;
+            pic_param->ReferenceFrames[i].flags = VA_PICTURE_H264_INVALID;
+        }
+    }
+    assert(avcenc_context.codedbuf_buf_id != VA_INVALID_ID);
+    pic_param->coded_buf = avcenc_context.codedbuf_buf_id;
+    pic_param->frame_num = frame_num;
+    pic_param->pic_fields.bits.idr_pic_flag = !!is_idr;
+    pic_param->pic_fields.bits.reference_pic_flag = (slice_type != SLICE_TYPE_B);
+
+    va_status = vaCreateBuffer(va_dpy,
+                               avcenc_context.context_id,
+                               VAEncPictureParameterBufferType,
+                               sizeof(*pic_param), 1, pic_param,
+                               &avcenc_context.pic_param_buf_id);
+    CHECK_VASTATUS(va_status,"vaCreateBuffer");
+      
+        length_in_bits = build_packed_pic_buffer(&packed_pic_buffer);
+        offset_in_bytes = 0;
+        packed_header_param_buffer.type = VAEncPackedHeaderPicture;
+        packed_header_param_buffer.bit_length = length_in_bits;
+        packed_header_param_buffer.has_emulation_bytes = 0;
+
+        va_status = vaCreateBuffer(va_dpy,
+                                   avcenc_context.context_id,
+                                   VAEncPackedHeaderParameterBufferType,
+                                   sizeof(packed_header_param_buffer), 1, &packed_header_param_buffer,
+                                   &avcenc_context.packed_pic_header_param_buf_id);
+        CHECK_VASTATUS(va_status,"vaCreateBuffer");
+
+        va_status = vaCreateBuffer(va_dpy,
+                                   avcenc_context.context_id,
+                                   VAEncPackedHeaderDataBufferType,
+                                   (length_in_bits + 7) / 8, 1, packed_pic_buffer,
+                                   &avcenc_context.packed_pic_buf_id);
+        CHECK_VASTATUS(va_status,"vaCreateBuffer");
+
+
+}
+
+#ifndef VA_FOURCC_I420
+#define VA_FOURCC_I420          0x30323449
+#endif
+
+static void upload_yuv_to_surface(FILE *yuv_fp, VASurfaceID surface_id)
+{
+    VAImage surface_image;
+    VAStatus va_status;
+    void *surface_p = NULL;
+    unsigned char *y_src, *u_src, *v_src;
+    unsigned char *y_dst, *u_dst, *v_dst;
+    int y_size = picture_width * picture_height;
+    int u_size = (picture_width >> 1) * (picture_height >> 1);
+    int row, col;
+    size_t n_items;
+
+    do {
+        n_items = fread(newImageBuffer, frame_size, 1, yuv_fp);
+    } while (n_items != 1);
+
+    va_status = vaDeriveImage(va_dpy, surface_id, &surface_image);
+    CHECK_VASTATUS(va_status,"vaDeriveImage");
+
+    vaMapBuffer(va_dpy, surface_image.buf, &surface_p);
+    assert(VA_STATUS_SUCCESS == va_status);
+        
+    y_src = newImageBuffer;
+    u_src = newImageBuffer + y_size; /* UV offset for NV12 */
+    v_src = newImageBuffer + y_size + u_size;
+
+    y_dst = surface_p + surface_image.offsets[0];
+    u_dst = surface_p + surface_image.offsets[1]; /* UV offset for NV12 */
+    v_dst = surface_p + surface_image.offsets[2];
+
+    /* Y plane */
+    for (row = 0; row < surface_image.height; row++) {
+        memcpy(y_dst, y_src, surface_image.width);
+        y_dst += surface_image.pitches[0];
+        y_src += picture_width;
+    }
+
+    if (surface_image.format.fourcc == VA_FOURCC_NV12) { /* UV plane */
+        for (row = 0; row < surface_image.height / 2; row++) {
+            for (col = 0; col < surface_image.width / 2; col++) {
+                u_dst[col * 2] = u_src[col];
+                u_dst[col * 2 + 1] = v_src[col];
+            }
+
+            u_dst += surface_image.pitches[1];
+            u_src += (picture_width / 2);
+            v_src += (picture_width / 2);
+        }
+    } else if (surface_image.format.fourcc == VA_FOURCC_YV12 ||
+               surface_image.format.fourcc == VA_FOURCC_I420) {
+        const int U = surface_image.format.fourcc == VA_FOURCC_I420 ? 1 : 2;
+        const int V = surface_image.format.fourcc == VA_FOURCC_I420 ? 2 : 1;
+
+        u_dst = surface_p + surface_image.offsets[U];
+        v_dst = surface_p + surface_image.offsets[V];
+
+        for (row = 0; row < surface_image.height / 2; row++) {
+            memcpy(u_dst, u_src, surface_image.width / 2);
+            memcpy(v_dst, v_src, surface_image.width / 2);
+            u_dst += surface_image.pitches[U];
+            v_dst += surface_image.pitches[V];
+            u_src += (picture_width / 2);
+            v_src += (picture_width / 2);
+        }
+    }
+
+    vaUnmapBuffer(va_dpy, surface_image.buf);
+    vaDestroyImage(va_dpy, surface_image.image_id);
+}
+
+static void avcenc_update_slice_parameter(int slice_type, int display_num)
+{
+    VAEncSliceParameterBufferH264 *slice_param;
+    VAStatus va_status;
+    int i,j;
+
+    // Slice level
+    i = 0;
+    slice_param = &avcenc_context.slice_param[i];
+    slice_param->macroblock_address = 0;
+    slice_param->num_macroblocks = picture_height_in_mbs * picture_width_in_mbs; 
+    slice_param->pic_parameter_set_id = 0;
+    slice_param->slice_type = slice_type;
+    slice_param->pic_order_cnt_lsb = display_num * 2;
+    slice_param->direct_spatial_mv_pred_flag = 1;  // to match header NALU
+
+    slice_param->num_ref_idx_l0_active_minus1 = 0;      
+    slice_param->num_ref_idx_l1_active_minus1 = 0;
+
+    if ( slice_type == SLICE_TYPE_I ) {
+        slice_param->num_ref_idx_l0_active_minus1 = -1;      
+        slice_param->num_ref_idx_l1_active_minus1 = -1;
+    } else if ( slice_type == SLICE_TYPE_P ) {
+        slice_param->num_ref_idx_l0_active_minus1 = 0;      
+        slice_param->num_ref_idx_l1_active_minus1 = -1;
+    } else if ( slice_type == SLICE_TYPE_B ) {
+        slice_param->num_ref_idx_l0_active_minus1 = 0;      
+        slice_param->num_ref_idx_l1_active_minus1 = 0;
+    }
+
+    slice_param->cabac_init_idc = 0;
+    slice_param->slice_qp_delta = 0;
+    slice_param->disable_deblocking_filter_idc = 0;
+    slice_param->slice_alpha_c0_offset_div2 = 2;
+    slice_param->slice_beta_offset_div2 = 2;
+    slice_param->idr_pic_id = 0;
+    /*    */
+    if ( slice_type == SLICE_TYPE_I )
+    {
+        slice_param->RefPicList0[0].picture_id = VA_INVALID_ID;
+        slice_param->RefPicList0[0].flags = VA_PICTURE_H264_INVALID;
+    }
+    else
+    {
+        slice_param->RefPicList0[0].picture_id = surface_ids[SID_REFERENCE_PICTURE_L0];
+        slice_param->RefPicList0[0].flags = VA_PICTURE_H264_SHORT_TERM_REFERENCE;
+    }
+    for (j=1; j<32; j++)
+    {
+        slice_param->RefPicList0[j].picture_id = VA_INVALID_ID;
+        slice_param->RefPicList0[j].flags = VA_PICTURE_H264_INVALID;
+    }
+    for (j=0; j<32; j++)
+    {
+        slice_param->RefPicList1[j].picture_id = VA_INVALID_ID;
+        slice_param->RefPicList1[j].flags = VA_PICTURE_H264_INVALID;
+    }
+    if ( slice_type == SLICE_TYPE_B )
+    {
+        slice_param->RefPicList1[0].picture_id = surface_ids[SID_REFERENCE_PICTURE_L1];
+        slice_param->RefPicList1[0].flags = VA_PICTURE_H264_SHORT_TERM_REFERENCE;
+
+    }
+    //slice_param->RefPicList1[0].picture_id = surface_ids[SID_REFERENCE_PICTURE_L1];
+    /* FIXME: fill other fields */
+
+    va_status = vaCreateBuffer(va_dpy,
+                               avcenc_context.context_id,
+                               VAEncSliceParameterBufferType,
+                               sizeof(*slice_param), 1, slice_param,
+                               &avcenc_context.slice_param_buf_id[i]);
+    CHECK_VASTATUS(va_status,"vaCreateBuffer");;
+    i++;
+
+#if 0
+    slice_param = &avcenc_context.slice_param[i];
+    slice_param->macroblock_address = picture_height_in_mbs * picture_width_in_mbs / 2;
+    slice_param->num_macroblocks = picture_height_in_mbs * picture_width_in_mbs / 2;
+    slice_param->pic_parameter_set_id = 0;
+    slice_param->slice_type = slice_type;
+    slice_param->direct_spatial_mv_pred_flag = 0;
+    slice_param->num_ref_idx_l0_active_minus1 = 0;      /* FIXME: ??? */
+    slice_param->num_ref_idx_l1_active_minus1 = 0;
+    slice_param->cabac_init_idc = 0;
+    slice_param->slice_qp_delta = 0;
+    slice_param->disable_deblocking_filter_idc = 0;
+    slice_param->slice_alpha_c0_offset_div2 = 2;
+    slice_param->slice_beta_offset_div2 = 2;
+    slice_param->idr_pic_id = 0;
+
+    /* FIXME: fill other fields */
+
+    va_status = vaCreateBuffer(va_dpy,
+                               avcenc_context.context_id,
+                               VAEncSliceParameterBufferType,
+                               sizeof(*slice_param), 1, slice_param,
+                               &avcenc_context.slice_param_buf_id[i]);
+    CHECK_VASTATUS(va_status,"vaCreateBuffer");;
+    i++;
+#endif
+
+    avcenc_context.num_slices = i;
+}
+
+static int begin_picture(FILE *yuv_fp, int frame_num, int display_num, int slice_type, int is_idr)
+{
+    VAStatus va_status;
+    VAEncPackedHeaderParameterBuffer packed_header_param_buffer;
+    unsigned int length_in_bits, offset_in_bytes;
+    unsigned char *packed_seq_buffer = NULL, *packed_slc_buffer=NULL;
+    
+#if 0
+    if (avcenc_context.upload_thread_value != 0) {
+        fprintf(stderr, "FATAL error!!!\n");
+        exit(1);
+    }
+
+    pthread_join(avcenc_context.upload_thread_id, NULL);
+#endif
+    avcenc_context.upload_thread_value = -1;
+/*
+    if (avcenc_context.current_input_surface == SID_INPUT_PICTURE_0)
+        avcenc_context.current_input_surface = SID_INPUT_PICTURE_1;
+    else if (avcenc_context.current_input_surface == SID_INPUT_PICTURE_1)
+        avcenc_context.current_input_surface = SID_INPUT_PICTURE_2;
+    else if (avcenc_context.current_input_surface == SID_INPUT_PICTURE_2)
+        avcenc_context.current_input_surface = SID_INPUT_PICTURE_3;
+    else if (avcenc_context.current_input_surface == SID_INPUT_PICTURE_3)
+        avcenc_context.current_input_surface = SID_INPUT_PICTURE_4;
+    else
+        avcenc_context.current_input_surface = SID_INPUT_PICTURE_0;
+*/  
+    avcenc_context.current_input_surface = next_input_id;
+    if (frame_num == 0) {
+
+        assert(slice_type == SLICE_TYPE_I);
+        length_in_bits = build_packed_seq_buffer(&packed_seq_buffer);
+        offset_in_bytes = 0;
+        packed_header_param_buffer.type = VAEncPackedHeaderSequence;
+        packed_header_param_buffer.bit_length = length_in_bits;
+        packed_header_param_buffer.has_emulation_bytes = 0;
+        va_status = vaCreateBuffer(va_dpy,
+                                   avcenc_context.context_id,
+                                   VAEncPackedHeaderParameterBufferType,
+                                   sizeof(packed_header_param_buffer), 1, &packed_header_param_buffer,
+                                   &avcenc_context.packed_seq_header_param_buf_id);
+        CHECK_VASTATUS(va_status,"vaCreateBuffer");
+
+        va_status = vaCreateBuffer(va_dpy,
+                                   avcenc_context.context_id,
+                                   VAEncPackedHeaderDataBufferType,
+                                   (length_in_bits + 7) / 8, 1, packed_seq_buffer,
+                                   &avcenc_context.packed_seq_buf_id);
+        CHECK_VASTATUS(va_status,"vaCreateBuffer");
+
+
+        free(packed_seq_buffer);
+    }
+
+    /* sequence parameter set */
+    VAEncSequenceParameterBufferH264 *seq_param = &avcenc_context.seq_param;
+    va_status = vaCreateBuffer(va_dpy,
+                               avcenc_context.context_id,
+                               VAEncSequenceParameterBufferType,
+                               sizeof(*seq_param), 1, seq_param,
+                               &avcenc_context.seq_param_buf_id);
+    CHECK_VASTATUS(va_status,"vaCreateBuffer");
+    
+    /* hrd parameter */
+    VAEncMiscParameterBuffer *misc_param;
+    VAEncMiscParameterHRD *misc_hrd_param;
+    vaCreateBuffer(va_dpy,
+                   avcenc_context.context_id,
+                   VAEncMiscParameterBufferType,
+                   sizeof(VAEncMiscParameterBuffer) + sizeof(VAEncMiscParameterHRD),
+                   1,
+                   NULL, 
+                   &avcenc_context.misc_parameter_hrd_buf_id);
+    CHECK_VASTATUS(va_status, "vaCreateBuffer");
+
+    vaMapBuffer(va_dpy,
+                avcenc_context.misc_parameter_hrd_buf_id,
+                (void **)&misc_param);
+    misc_param->type = VAEncMiscParameterTypeHRD;
+    misc_hrd_param = (VAEncMiscParameterHRD *)misc_param->data;
+
+    if (frame_bit_rate > 0) {
+        misc_hrd_param->initial_buffer_fullness = frame_bit_rate * 1000 * 4;
+        misc_hrd_param->buffer_size = frame_bit_rate * 1000 * 8;
+    } else {
+        misc_hrd_param->initial_buffer_fullness = 0;
+        misc_hrd_param->buffer_size = 0;
+    }
+
+    vaUnmapBuffer(va_dpy, avcenc_context.misc_parameter_hrd_buf_id);
+#if 0    
+    /* rate control parameter  */
+    VAEncMiscParameterRateControl  *misc_rc_param;
+    vaCreateBuffer(va_dpy,
+                   avcenc_context.context_id,
+                   VAEncMiscParameterBufferType,
+                   sizeof(VAEncMiscParameterBuffer) + sizeof(VAEncMiscParameterRateControl),
+                   1,
+                   NULL, 
+                   &avcenc_context.misc_parameter_rc_buf_id);
+    CHECK_VASTATUS(va_status, "vaCreateBuffer");
+
+    vaMapBuffer(va_dpy,
+                avcenc_context.misc_parameter_rc_buf_id,
+                (void **)&misc_param);
+    misc_param->type = VAEncMiscParameterTypeRateControl;
+    misc_rc_param = (VAEncMiscParameterRateControl  *)misc_param->data;
+
+    if (frame_bit_rate > 0) {
+        misc_rc_param->bits_per_second   = frame_bit_rate * 1000;
+        misc_rc_param->target_percentage = 95;
+        misc_rc_param->window_size       = 4000;
+        misc_rc_param->initial_qp        = 26;
+        misc_rc_param->min_qp            = 4;
+    } else {
+        misc_rc_param->bits_per_second   = 0;
+        misc_rc_param->target_percentage = 0;
+        misc_rc_param->window_size       = 0;
+        misc_rc_param->initial_qp        = 0;
+        misc_rc_param->min_qp            = 0;
+    }
+
+    vaUnmapBuffer(va_dpy, avcenc_context.misc_parameter_rc_buf_id);
+#endif
+    /* frame rate parameter  */
+    VAEncMiscParameterFrameRate    *misc_fr_param;
+    vaCreateBuffer(va_dpy,
+                   avcenc_context.context_id,
+                   VAEncMiscParameterBufferType,
+                   sizeof(VAEncMiscParameterBuffer) + sizeof(VAEncMiscParameterFrameRate),
+                   1,
+                   NULL, 
+                   &avcenc_context.misc_parameter_fr_buf_id);
+    CHECK_VASTATUS(va_status, "vaCreateBuffer");
+
+    vaMapBuffer(va_dpy,
+                avcenc_context.misc_parameter_fr_buf_id,
+                (void **)&misc_param);
+    misc_param->type = VAEncMiscParameterTypeFrameRate;
+    misc_fr_param = (VAEncMiscParameterRateControl*)misc_param->data;
+    misc_fr_param->framerate = 30*100;
+
+    vaUnmapBuffer(va_dpy, avcenc_context.misc_parameter_fr_buf_id);
+
+    /* FEI picture control parameter  */
+    VAEncMiscParameterFEIFrameControlH264Intel    *misc_fei_pic_control_param;
+    vaCreateBuffer(va_dpy,
+                   avcenc_context.context_id,
+                   VAEncMiscParameterBufferType,
+                   sizeof(VAEncMiscParameterBuffer) + sizeof(VAEncMiscParameterFEIFrameControlH264Intel),
+                   1,
+                   NULL, 
+                   &avcenc_context.misc_parameter_fei_pic_control_buf_id);
+    CHECK_VASTATUS(va_status, "vaCreateBuffer");
+
+    vaMapBuffer(va_dpy,
+                avcenc_context.misc_parameter_fei_pic_control_buf_id,
+                (void **)&misc_param);
+    misc_param->type = VAEncMiscParameterTypeFEIFrameControlIntel;
+    misc_fei_pic_control_param  = (VAEncMiscParameterFEIFrameControlH264Intel*)misc_param->data;
+    if(avcenc_context.fei_pic_param.mb_input)
+    {
+#if 0        
+        FILE *mbctrl_fp;
+        int  data_amount;
+        int  read_amount;
+        mbctrl_fp = fopen( "mbctrlin.bin", "rb" );
+        if ( ! mbctrl_fp ) {
+            printf( "Error: No input MB control file was found\n" );
+            exit(-1);
+        }
+        data_amount = avcenc_context.fei_mbctrl_buf_size;
+        read_amount = fread( feiMbCtrlBuffer, 1, data_amount, mbctrl_fp );
+        if( data_amount != read_amount )
+        {
+            printf( "Error: read input MVs file failed\n" );
+            exit(-1);
+        }
+#else
+        VAEncFEIMBControlH264Intel *pFeiMbCtrl = feiMbCtrlBuffer;
+        int i = 0;
+        memset(pFeiMbCtrl, 0, avcenc_context.fei_mbctrl_buf_size);
+
+        for(i = 0; i<(picture_width_in_mbs*picture_height_in_mbs); i++)
+        {
+            pFeiMbCtrl->force_to_intra      = 0;
+            pFeiMbCtrl->force_to_skip       = 0;
+
+            pFeiMbCtrl->target_size_in_word = 0xff;
+            pFeiMbCtrl->max_size_in_word    = 80;
+            pFeiMbCtrl++;
+        }
+        if(0)
+        {
+            FILE *out;
+            int  ret = 0;
+            out = fopen("mbctrlin.bin", "wb");
+            ret = fwrite(feiMbCtrlBuffer, 1, avcenc_context.fei_mbctrl_buf_size, out);
+            fclose(out);
+        }
+#endif
+        va_status = vaCreateBuffer(va_dpy,
+                                   avcenc_context.context_id,
+                                   VAEncFEIMBControlBufferTypeIntel,
+                                   sizeof(VAEncFEIMBControlH264Intel)*picture_width_in_mbs*picture_height_in_mbs, 1, feiMbCtrlBuffer,
+                                   &avcenc_context.fei_mbctrl_buf_id);
+        CHECK_VASTATUS(va_status,"vaCreateBuffer");
+        avcenc_context.fei_pic_param.mb_ctrl     = avcenc_context.fei_mbctrl_buf_id;
+    }
+    if(avcenc_context.fei_pic_param.mv_predictor_enable)
+    {
+#if 0
+        // FIXME get FEI MV Predictor from file input to feiMvPredBuffer first
+        FILE *mvpred_fp;
+        int  data_amount;
+        int  read_amount;
+        mvpred_fp = fopen( "mvpindata0.bin", "rb" );
+        if ( ! mvpred_fp ) {
+            printf( "Error: No input MVs file was found\n" );
+            exit(-1);
+        }
+        data_amount = avcenc_context.fei_mvpred_buf_size;
+        read_amount = fread( feiMvPredBuffer, 1, data_amount, mvpred_fp );
+        if( data_amount != read_amount )
+        {
+            printf( "Error: read input MVs file failed\n" );
+            exit(-1);
+        }
+#else
+        VAEncMVPredictorH264Intel *pMvPred = feiMvPredBuffer;
+        int i = 0;
+        memset(pMvPred, 0, avcenc_context.fei_mvpred_buf_size);
+
+        for(i = 0; i<(picture_width_in_mbs*picture_height_in_mbs); i++)
+        {
+            pMvPred->mv[0].mv0[0] = 1;
+            pMvPred->mv[1].mv0[0] = 1;
+            pMvPred++;
+        }
+        if(0)
+        {
+            FILE *out;
+            int  ret = 0;
+            out = fopen("mvpindata.bin", "wb");
+            ret = fwrite(feiMvPredBuffer, 1, avcenc_context.fei_mvpred_buf_size, out);
+            fclose(out);
+        }
+#endif
+        va_status = vaCreateBuffer(va_dpy,
+                                   avcenc_context.context_id,
+                                   VAEncFEIMVPredictorBufferTypeIntel,
+                                   sizeof(VAEncMVPredictorH264Intel)*picture_width_in_mbs*picture_height_in_mbs, 1, feiMvPredBuffer,
+                                   &avcenc_context.fei_mvpred_buf_id);
+        CHECK_VASTATUS(va_status,"vaCreateBuffer");
+        avcenc_context.fei_pic_param.mv_predictor = avcenc_context.fei_mvpred_buf_id;
+    }
+    /* distortion buffer */
+    if(stats_en)
+    {
+        va_status = vaCreateBuffer(va_dpy,
+                                   avcenc_context.context_id,
+                                   VAEncFEIDistortionBufferTypeIntel,
+                                   sizeof(VAEncFEIDistortionBufferH264Intel)*picture_width_in_mbs*picture_height_in_mbs, 1, NULL,
+                                   &avcenc_context.fei_distortion_buf_id);
+        CHECK_VASTATUS(va_status,"vaCreateBuffer");
+        avcenc_context.fei_pic_param.distortion = avcenc_context.fei_distortion_buf_id;
+       
+        va_status = vaCreateBuffer(va_dpy,
+                                   avcenc_context.context_id,
+                                   VAEncFEIMVBufferTypeIntel,
+                                   sizeof(VAMotionVectorIntel)*16*picture_width_in_mbs*picture_height_in_mbs, 1, NULL,
+                                   &avcenc_context.fei_mv_buf_id);
+        CHECK_VASTATUS(va_status,"vaCreateBuffer");
+        avcenc_context.fei_pic_param.mv_data = avcenc_context.fei_mv_buf_id;
+        
+        va_status = vaCreateBuffer(va_dpy,
+                                   avcenc_context.context_id,
+                                   VAEncFEIModeBufferTypeIntel,
+                                   sizeof(VAEncFEIModeBufferH264Intel)*picture_width_in_mbs*picture_height_in_mbs, 1, NULL,
+                                   &avcenc_context.fei_mbcode_buf_id);
+        CHECK_VASTATUS(va_status,"vaCreateBuffer");
+        avcenc_context.fei_pic_param.mb_code_data = avcenc_context.fei_mbcode_buf_id;
+
+    }
+    *misc_fei_pic_control_param = avcenc_context.fei_pic_param;
+
+    vaUnmapBuffer(va_dpy, avcenc_context.misc_parameter_fei_pic_control_buf_id);
+
+    /* slice parameter */
+    avcenc_update_slice_parameter(slice_type,display_num);
+
+    length_in_bits = build_packed_slc_buffer(&packed_slc_buffer,frame_num, display_num, slice_type, is_idr); 
+    offset_in_bytes = 0;
+    packed_header_param_buffer.type = VAEncPackedHeaderSlice;
+    packed_header_param_buffer.bit_length = length_in_bits;
+    packed_header_param_buffer.has_emulation_bytes = 0;
+    
+    va_status = vaCreateBuffer(va_dpy,
+                               avcenc_context.context_id,
+                               VAEncPackedHeaderParameterBufferType,
+                               sizeof(packed_header_param_buffer), 1, &packed_header_param_buffer,
+                               &avcenc_context.packed_slc_header_param_buf_id);
+    CHECK_VASTATUS(va_status,"vaCreateBuffer");
+    
+    va_status = vaCreateBuffer(va_dpy,
+                               avcenc_context.context_id,
+                               VAEncPackedHeaderDataBufferType,
+                               (length_in_bits + 7) / 8, 1, packed_slc_buffer,
+                               &avcenc_context.packed_slc_buf_id);
+    CHECK_VASTATUS(va_status,"vaCreateBuffer");
+
+    free(packed_slc_buffer);
+    
+    return 0;
+}
+
+int avcenc_render_picture()
+{
+    VAStatus va_status;
+    VABufferID va_buffers[10];
+    unsigned int num_va_buffers = 0;
+    int i;
+
+    va_buffers[num_va_buffers++] = avcenc_context.seq_param_buf_id;
+    va_buffers[num_va_buffers++] = avcenc_context.pic_param_buf_id;
+
+    if (avcenc_context.packed_seq_header_param_buf_id != VA_INVALID_ID)
+        va_buffers[num_va_buffers++] = avcenc_context.packed_seq_header_param_buf_id;
+
+    if (avcenc_context.packed_seq_buf_id != VA_INVALID_ID)
+        va_buffers[num_va_buffers++] = avcenc_context.packed_seq_buf_id;
+
+    if (avcenc_context.packed_pic_header_param_buf_id != VA_INVALID_ID)
+        va_buffers[num_va_buffers++] = avcenc_context.packed_pic_header_param_buf_id;
+
+    if (avcenc_context.packed_pic_buf_id != VA_INVALID_ID)
+        va_buffers[num_va_buffers++] = avcenc_context.packed_pic_buf_id;
+
+    if (avcenc_context.packed_slc_header_param_buf_id != VA_INVALID_ID)
+        va_buffers[num_va_buffers++] = avcenc_context.packed_slc_header_param_buf_id;
+
+    if (avcenc_context.packed_slc_buf_id != VA_INVALID_ID)
+        va_buffers[num_va_buffers++] = avcenc_context.packed_slc_buf_id;
+
+    if (avcenc_context.misc_parameter_hrd_buf_id != VA_INVALID_ID)
+        va_buffers[num_va_buffers++] =  avcenc_context.misc_parameter_hrd_buf_id;
+    
+    if (avcenc_context.misc_parameter_rc_buf_id != VA_INVALID_ID)
+        va_buffers[num_va_buffers++] =  avcenc_context.misc_parameter_rc_buf_id;
+
+    if (avcenc_context.misc_parameter_fr_buf_id != VA_INVALID_ID)
+        va_buffers[num_va_buffers++] =  avcenc_context.misc_parameter_fr_buf_id;
+    
+    if (avcenc_context.misc_parameter_fei_pic_control_buf_id != VA_INVALID_ID)
+        va_buffers[num_va_buffers++] =  avcenc_context.misc_parameter_fei_pic_control_buf_id;
+#if 0    
+    if (stats_en)
+    {
+        va_buffers[num_va_buffers++] =  avcenc_context.fei_distortion_buf_id;
+        va_buffers[num_va_buffers++] =  avcenc_context.fei_mv_buf_id;
+        va_buffers[num_va_buffers++] =  avcenc_context.fei_mbcode_buf_id;
+    }
+#endif
+#if 1
+    // there have been sent by avcenc_context.misc_parameter_fei_pic_control_buf_id, so no needed to render again. 
+    if(avcenc_context.fei_pic_param.mb_input)
+       va_buffers[num_va_buffers++] =  avcenc_context.fei_mbctrl_buf_id;
+
+    if(avcenc_context.fei_pic_param.mv_predictor_enable)
+        va_buffers[num_va_buffers++] =  avcenc_context.fei_mvpred_buf_id;
+#endif
+    va_status = vaBeginPicture(va_dpy,
+                               avcenc_context.context_id,
+                               surface_ids[avcenc_context.current_input_surface]);
+    CHECK_VASTATUS(va_status,"vaBeginPicture");
+    
+    va_status = vaRenderPicture(va_dpy,
+                                avcenc_context.context_id,
+                                va_buffers,
+                                num_va_buffers);
+    CHECK_VASTATUS(va_status,"vaRenderPicture");
+    
+    for(i = 0; i < avcenc_context.num_slices; i++) {
+        va_status = vaRenderPicture(va_dpy,
+                                avcenc_context.context_id,
+                                &avcenc_context.slice_param_buf_id[i],
+                                1);
+        CHECK_VASTATUS(va_status,"vaRenderPicture");
+    }
+
+    va_status = vaEndPicture(va_dpy, avcenc_context.context_id);
+    CHECK_VASTATUS(va_status,"vaEndPicture");
+
+    return 0;
+}
+
+static int avcenc_destroy_buffers(VABufferID *va_buffers, unsigned int num_va_buffers)
+{
+    VAStatus va_status;
+    unsigned int i;
+
+    for (i = 0; i < num_va_buffers; i++) {
+        if (va_buffers[i] != VA_INVALID_ID) {
+            va_status = vaDestroyBuffer(va_dpy, va_buffers[i]);
+            CHECK_VASTATUS(va_status,"vaDestroyBuffer");
+            va_buffers[i] = VA_INVALID_ID;
+        }
+    }
+
+    return 0;
+}
+
+static GetNextInput (int current_input)
+{
+    if (current_input == SID_INPUT_PICTURE_0)
+        return SID_INPUT_PICTURE_1;
+    else if (current_input == SID_INPUT_PICTURE_1)
+        return SID_INPUT_PICTURE_2;
+    else if (current_input == SID_INPUT_PICTURE_2)
+        return SID_INPUT_PICTURE_3;
+    else if (current_input == SID_INPUT_PICTURE_3)
+        return SID_INPUT_PICTURE_4;
+    else
+        return SID_INPUT_PICTURE_0; 
+}
+static void end_picture(int slice_type, int next_is_bpic)
+{
+    VABufferID tempID;
+
+    /* Prepare for next picture */
+    tempID = surface_ids[SID_RECON_PICTURE];  
+
+    if (slice_type != SLICE_TYPE_B) {
+        if (next_is_bpic) {
+            surface_ids[SID_RECON_PICTURE] = surface_ids[SID_REFERENCE_PICTURE_L1]; 
+            surface_ids[SID_REFERENCE_PICTURE_L1] = tempID;	
+            ref1_POC[0] = Cur_POC[0];
+            ref1_POC[1] = Cur_POC[1];
+            next_input_id = GetNextInput(next_input_id );
+        } else {
+            surface_ids[SID_RECON_PICTURE] = surface_ids[SID_REFERENCE_PICTURE_L0]; 
+            surface_ids[SID_REFERENCE_PICTURE_L0] = tempID;
+            ref0_POC[0] = Cur_POC[0];
+            ref0_POC[1] = Cur_POC[1];
+            next_input_id = GetNextInput(next_input_id );
+        }
+    } else {
+        if (!next_is_bpic) {
+            surface_ids[SID_RECON_PICTURE] = surface_ids[SID_REFERENCE_PICTURE_L0]; 
+            surface_ids[SID_REFERENCE_PICTURE_L0] = surface_ids[SID_REFERENCE_PICTURE_L1];
+            surface_ids[SID_REFERENCE_PICTURE_L1] = tempID;
+            ref0_POC[0] = ref1_POC[0];
+            ref0_POC[1] = ref1_POC[1];
+            ref1_POC[0] = Cur_POC[0];
+            ref1_POC[1] = Cur_POC[1];
+            next_input_id = GetNextInput(next_input_id );
+        }
+    }
+
+    avcenc_destroy_buffers(&avcenc_context.seq_param_buf_id, 1);
+    avcenc_destroy_buffers(&avcenc_context.pic_param_buf_id, 1);
+    /*
+    avcenc_destroy_buffers(&avcenc_context.packed_seq_header_param_buf_id, 1);
+    avcenc_destroy_buffers(&avcenc_context.packed_seq_buf_id, 1);
+    */
+    avcenc_destroy_buffers(&avcenc_context.packed_pic_header_param_buf_id, 1);
+    avcenc_destroy_buffers(&avcenc_context.packed_pic_buf_id, 1);
+    
+    avcenc_destroy_buffers(&avcenc_context.packed_slc_header_param_buf_id, 1);
+    avcenc_destroy_buffers(&avcenc_context.packed_slc_buf_id, 1);
+
+    
+    avcenc_destroy_buffers(&avcenc_context.slice_param_buf_id[0], avcenc_context.num_slices);
+    avcenc_destroy_buffers(&avcenc_context.codedbuf_buf_id, 1);
+    avcenc_destroy_buffers(&avcenc_context.misc_parameter_hrd_buf_id, 1);
+    avcenc_destroy_buffers(&avcenc_context.misc_parameter_rc_buf_id, 1);
+    avcenc_destroy_buffers(&avcenc_context.misc_parameter_fr_buf_id, 1);
+    if(avcenc_context.fei_pic_param.mb_input)
+        avcenc_destroy_buffers(&avcenc_context.fei_mbctrl_buf_id, 1);
+    if(avcenc_context.fei_pic_param.mv_predictor_enable)
+        avcenc_destroy_buffers(&avcenc_context.fei_mvpred_buf_id, 1);
+    if (stats_en)
+    {
+        avcenc_destroy_buffers(&avcenc_context.fei_distortion_buf_id, 1);
+        avcenc_destroy_buffers(&avcenc_context.fei_mv_buf_id,         1);
+        avcenc_destroy_buffers(&avcenc_context.fei_mbcode_buf_id,     1);
+    }
+
+    memset(avcenc_context.slice_param, 0, sizeof(avcenc_context.slice_param));
+    avcenc_context.num_slices = 0;
+}
+
+#define BITSTREAM_ALLOCATE_STEPPING     4096
+
+struct __bitstream {
+    unsigned int *buffer;
+    int bit_offset;
+    int max_size_in_dword;
+};
+
+typedef struct __bitstream bitstream;
+
+#if 0
+static int 
+get_coded_bitsteam_length(unsigned char *buffer, int buffer_length)
+{
+    int i;
+
+    for (i = 0; i < buffer_length - 3; i++) {
+        if (!buffer[i] &&
+            !buffer[i + 1] &&
+            !buffer[i + 2] &&
+            !buffer[i + 3])
+            break;
+    }
+
+    return i;
+}
+#endif
+
+static unsigned int 
+swap32(unsigned int val)
+{
+    unsigned char *pval = (unsigned char *)&val;
+
+    return ((pval[0] << 24)     |
+            (pval[1] << 16)     |
+            (pval[2] << 8)      |
+            (pval[3] << 0));
+}
+
+static void
+bitstream_start(bitstream *bs)
+{
+    bs->max_size_in_dword = BITSTREAM_ALLOCATE_STEPPING;
+    bs->buffer = calloc(bs->max_size_in_dword * sizeof(int), 1);
+    bs->bit_offset = 0;
+}
+
+static void
+bitstream_end(bitstream *bs)
+{
+    int pos = (bs->bit_offset >> 5);
+    int bit_offset = (bs->bit_offset & 0x1f);
+    int bit_left = 32 - bit_offset;
+
+    if (bit_offset) {
+        bs->buffer[pos] = swap32((bs->buffer[pos] << bit_left));
+    }
+}
+ 
+static void
+bitstream_put_ui(bitstream *bs, unsigned int val, int size_in_bits)
+{
+    int pos = (bs->bit_offset >> 5);
+    int bit_offset = (bs->bit_offset & 0x1f);
+    int bit_left = 32 - bit_offset;
+
+    if (!size_in_bits)
+        return;
+
+    bs->bit_offset += size_in_bits;
+
+    if (bit_left > size_in_bits) {
+        bs->buffer[pos] = (bs->buffer[pos] << size_in_bits | val);
+    } else {
+        size_in_bits -= bit_left;
+        bs->buffer[pos] = (bs->buffer[pos] << bit_left) | (val >> size_in_bits);
+        bs->buffer[pos] = swap32(bs->buffer[pos]);
+
+        if (pos + 1 == bs->max_size_in_dword) {
+            bs->max_size_in_dword += BITSTREAM_ALLOCATE_STEPPING;
+            bs->buffer = realloc(bs->buffer, bs->max_size_in_dword * sizeof(unsigned int));
+        }
+
+        bs->buffer[pos + 1] = val;
+    }
+}
+
+static void
+bitstream_put_ue(bitstream *bs, unsigned int val)
+{
+    int size_in_bits = 0;
+    int tmp_val = ++val;
+
+    while (tmp_val) {
+        tmp_val >>= 1;
+        size_in_bits++;
+    }
+
+    bitstream_put_ui(bs, 0, size_in_bits - 1); // leading zero
+    bitstream_put_ui(bs, val, size_in_bits);
+}
+
+static void
+bitstream_put_se(bitstream *bs, int val)
+{
+    unsigned int new_val;
+
+    if (val <= 0)
+        new_val = -2 * val;
+    else
+        new_val = 2 * val - 1;
+
+    bitstream_put_ue(bs, new_val);
+}
+
+static void
+bitstream_byte_aligning(bitstream *bs, int bit)
+{
+    int bit_offset = (bs->bit_offset & 0x7);
+    int bit_left = 8 - bit_offset;
+    int new_val;
+
+    if (!bit_offset)
+        return;
+
+    assert(bit == 0 || bit == 1);
+
+    if (bit)
+        new_val = (1 << bit_left) - 1;
+    else
+        new_val = 0;
+
+    bitstream_put_ui(bs, new_val, bit_left);
+}
+
+static void 
+rbsp_trailing_bits(bitstream *bs)
+{
+    bitstream_put_ui(bs, 1, 1);
+    bitstream_byte_aligning(bs, 0);
+}
+
+static void nal_start_code_prefix(bitstream *bs)
+{
+    bitstream_put_ui(bs, 0x00000001, 32);
+}
+
+static void nal_header(bitstream *bs, int nal_ref_idc, int nal_unit_type)
+{
+    bitstream_put_ui(bs, 0, 1);                /* forbidden_zero_bit: 0 */
+    bitstream_put_ui(bs, nal_ref_idc, 2);
+    bitstream_put_ui(bs, nal_unit_type, 5);
+}
+
+static void sps_rbsp(bitstream *bs)
+{
+    VAEncSequenceParameterBufferH264 *seq_param = &avcenc_context.seq_param;
+    int profile_idc = PROFILE_IDC_BASELINE;
+
+    if (avcenc_context.profile == VAProfileH264High)
+        profile_idc = PROFILE_IDC_HIGH;
+    else if (avcenc_context.profile == VAProfileH264Main)
+        profile_idc = PROFILE_IDC_MAIN;
+
+    bitstream_put_ui(bs, profile_idc, 8);               /* profile_idc */
+    bitstream_put_ui(bs, 0, 1);                         /* constraint_set0_flag */
+    bitstream_put_ui(bs, 1, 1);                         /* constraint_set1_flag */
+    bitstream_put_ui(bs, 0, 1);                         /* constraint_set2_flag */
+    bitstream_put_ui(bs, 0, 1);                         /* constraint_set3_flag */
+    bitstream_put_ui(bs, 0, 4);                         /* reserved_zero_4bits */
+    bitstream_put_ui(bs, seq_param->level_idc, 8);      /* level_idc */
+    bitstream_put_ue(bs, seq_param->seq_parameter_set_id);      /* seq_parameter_set_id */
+
+    if (avcenc_context.profile == VAProfileH264High)
+    {
+        bitstream_put_ue(bs, 1);  /* chroma_format_idc */
+        bitstream_put_ue(bs, 0);  /* bit_depth_luma_minus8 */
+        bitstream_put_ue(bs, 0);  /* bit_depth_chroma_minus8 */
+        bitstream_put_ui(bs, 0, 1); /* qpprime_y_zero_transform_bypass_flag */
+        bitstream_put_ui(bs, 0, 1); /* seq_scaling_matrix_present_flag */
+    }
+    
+    bitstream_put_ue(bs, seq_param->seq_fields.bits.log2_max_frame_num_minus4); /* log2_max_frame_num_minus4 */
+
+    bitstream_put_ue(bs, seq_param->seq_fields.bits.pic_order_cnt_type);        /* pic_order_cnt_type */
+
+    if (seq_param->seq_fields.bits.pic_order_cnt_type == 0)
+        bitstream_put_ue(bs, seq_param->seq_fields.bits.log2_max_pic_order_cnt_lsb_minus4);     /* log2_max_pic_order_cnt_lsb_minus4 */
+    else {
+        assert(0);
+    }
+
+    bitstream_put_ue(bs, seq_param->max_num_ref_frames);        /* num_ref_frames */
+    bitstream_put_ui(bs, 0, 1);                                 /* gaps_in_frame_num_value_allowed_flag */
+
+    bitstream_put_ue(bs, seq_param->picture_width_in_mbs - 1);  /* pic_width_in_mbs_minus1 */
+    bitstream_put_ue(bs, seq_param->picture_height_in_mbs - 1); /* pic_height_in_map_units_minus1 */
+    bitstream_put_ui(bs, seq_param->seq_fields.bits.frame_mbs_only_flag, 1);    /* frame_mbs_only_flag */
+
+    if (!seq_param->seq_fields.bits.frame_mbs_only_flag) {
+        assert(0);
+    }
+
+    bitstream_put_ui(bs, seq_param->seq_fields.bits.direct_8x8_inference_flag, 1);      /* direct_8x8_inference_flag */
+    bitstream_put_ui(bs, seq_param->frame_cropping_flag, 1);            /* frame_cropping_flag */
+
+    if (seq_param->frame_cropping_flag) {
+        bitstream_put_ue(bs, seq_param->frame_crop_left_offset);        /* frame_crop_left_offset */
+        bitstream_put_ue(bs, seq_param->frame_crop_right_offset);       /* frame_crop_right_offset */
+        bitstream_put_ue(bs, seq_param->frame_crop_top_offset);         /* frame_crop_top_offset */
+        bitstream_put_ue(bs, seq_param->frame_crop_bottom_offset);      /* frame_crop_bottom_offset */
+    }
+    
+    if ( frame_bit_rate < 0 ) {
+        bitstream_put_ui(bs, 0, 1); /* vui_parameters_present_flag */
+    } else {
+        bitstream_put_ui(bs, 1, 1); /* vui_parameters_present_flag */
+        bitstream_put_ui(bs, 0, 1); /* aspect_ratio_info_present_flag */
+        bitstream_put_ui(bs, 0, 1); /* overscan_info_present_flag */
+        bitstream_put_ui(bs, 0, 1); /* video_signal_type_present_flag */
+        bitstream_put_ui(bs, 0, 1); /* chroma_loc_info_present_flag */
+        bitstream_put_ui(bs, 1, 1); /* timing_info_present_flag */
+        {
+            bitstream_put_ui(bs, 15, 32);
+            bitstream_put_ui(bs, 900, 32);
+            bitstream_put_ui(bs, 1, 1);
+        }
+        bitstream_put_ui(bs, 1, 1); /* nal_hrd_parameters_present_flag */
+        {
+            // hrd_parameters 
+            bitstream_put_ue(bs, 0);    /* cpb_cnt_minus1 */
+            bitstream_put_ui(bs, 4, 4); /* bit_rate_scale */
+            bitstream_put_ui(bs, 6, 4); /* cpb_size_scale */
+           
+            bitstream_put_ue(bs, frame_bit_rate - 1); /* bit_rate_value_minus1[0] */
+            bitstream_put_ue(bs, frame_bit_rate*8 - 1); /* cpb_size_value_minus1[0] */
+            bitstream_put_ui(bs, 1, 1);  /* cbr_flag[0] */
+
+            bitstream_put_ui(bs, 23, 5);   /* initial_cpb_removal_delay_length_minus1 */
+            bitstream_put_ui(bs, 23, 5);   /* cpb_removal_delay_length_minus1 */
+            bitstream_put_ui(bs, 23, 5);   /* dpb_output_delay_length_minus1 */
+            bitstream_put_ui(bs, 23, 5);   /* time_offset_length  */
+        }
+        bitstream_put_ui(bs, 0, 1);   /* vcl_hrd_parameters_present_flag */
+        bitstream_put_ui(bs, 0, 1);   /* low_delay_hrd_flag */ 
+
+        bitstream_put_ui(bs, 0, 1); /* pic_struct_present_flag */
+        bitstream_put_ui(bs, 0, 1); /* bitstream_restriction_flag */
+    }
+
+    rbsp_trailing_bits(bs);     /* rbsp_trailing_bits */
+}
+
+#if 0
+static void build_nal_sps(FILE *avc_fp)
+{
+    bitstream bs;
+
+    bitstream_start(&bs);
+    nal_start_code_prefix(&bs);
+    nal_header(&bs, NAL_REF_IDC_HIGH, NAL_SPS);
+    sps_rbsp(&bs);
+    bitstream_end(&bs, avc_fp);
+}
+#endif
+
+static void pps_rbsp(bitstream *bs)
+{
+    VAEncPictureParameterBufferH264 *pic_param = &avcenc_context.pic_param;
+
+    bitstream_put_ue(bs, pic_param->pic_parameter_set_id);      /* pic_parameter_set_id */
+    bitstream_put_ue(bs, pic_param->seq_parameter_set_id);      /* seq_parameter_set_id */
+
+    bitstream_put_ui(bs, pic_param->pic_fields.bits.entropy_coding_mode_flag, 1);  /* entropy_coding_mode_flag */
+
+    bitstream_put_ui(bs, 0, 1);                         /* pic_order_present_flag: 0 */
+
+    bitstream_put_ue(bs, 0);                            /* num_slice_groups_minus1 */
+
+    bitstream_put_ue(bs, pic_param->num_ref_idx_l0_active_minus1);      /* num_ref_idx_l0_active_minus1 */
+    bitstream_put_ue(bs, pic_param->num_ref_idx_l1_active_minus1);      /* num_ref_idx_l1_active_minus1 1 */
+
+    bitstream_put_ui(bs, pic_param->pic_fields.bits.weighted_pred_flag, 1);     /* weighted_pred_flag: 0 */
+    bitstream_put_ui(bs, pic_param->pic_fields.bits.weighted_bipred_idc, 2);	/* weighted_bipred_idc: 0 */
+
+    bitstream_put_se(bs, pic_param->pic_init_qp - 26);  /* pic_init_qp_minus26 */
+    bitstream_put_se(bs, 0);                            /* pic_init_qs_minus26 */
+    bitstream_put_se(bs, 0);                            /* chroma_qp_index_offset */
+
+    bitstream_put_ui(bs, pic_param->pic_fields.bits.deblocking_filter_control_present_flag, 1); /* deblocking_filter_control_present_flag */
+    bitstream_put_ui(bs, 0, 1);                         /* constrained_intra_pred_flag */
+    bitstream_put_ui(bs, 0, 1);                         /* redundant_pic_cnt_present_flag */
+     
+    /* more_rbsp_data */
+    if (avcenc_context.profile == VAProfileH264High)
+    {
+        bitstream_put_ui(bs, pic_param->pic_fields.bits.transform_8x8_mode_flag, 1);    /*transform_8x8_mode_flag */
+        bitstream_put_ui(bs, 0, 1);                         /* pic_scaling_matrix_present_flag */
+        bitstream_put_se(bs, pic_param->second_chroma_qp_index_offset );    /*second_chroma_qp_index_offset */
+    }
+    rbsp_trailing_bits(bs);
+}
+
+#if 0
+static void build_nal_pps(FILE *avc_fp)
+{
+    bitstream bs;
+
+    bitstream_start(&bs);
+    nal_start_code_prefix(&bs);
+    nal_header(&bs, NAL_REF_IDC_HIGH, NAL_PPS);
+    pps_rbsp(&bs);
+    bitstream_end(&bs, avc_fp);
+}
+
+static void 
+build_header(FILE *avc_fp)
+{
+    build_nal_sps(avc_fp);
+    build_nal_pps(avc_fp);
+}
+#endif
+
+static int
+build_packed_pic_buffer(unsigned char **header_buffer)
+{
+    bitstream bs;
+
+    bitstream_start(&bs);
+    nal_start_code_prefix(&bs);
+    nal_header(&bs, NAL_REF_IDC_HIGH, NAL_PPS);
+    pps_rbsp(&bs);
+    bitstream_end(&bs);
+
+    *header_buffer = (unsigned char *)bs.buffer;
+    return bs.bit_offset;
+}
+
+static int
+build_packed_seq_buffer(unsigned char **header_buffer)
+{
+    bitstream bs;
+
+    bitstream_start(&bs);
+    nal_start_code_prefix(&bs);
+    nal_header(&bs, NAL_REF_IDC_HIGH, NAL_SPS);
+    sps_rbsp(&bs);
+    bitstream_end(&bs);
+
+    *header_buffer = (unsigned char *)bs.buffer;
+    return bs.bit_offset;
+}
+
+
+#if 1
+static void 
+slice_header(bitstream *bs, int frame_num, int display_frame, int slice_type, int nal_ref_idc, int is_idr)
+{
+    VAEncSequenceParameterBufferH264 *seq_param = &avcenc_context.seq_param;
+    VAEncPictureParameterBufferH264 *pic_param = &avcenc_context.pic_param;
+    int is_cabac = (pic_param->pic_fields.bits.entropy_coding_mode_flag == ENTROPY_MODE_CABAC);
+
+    bitstream_put_ue(bs, 0);                   /* first_mb_in_slice: 0 */
+    bitstream_put_ue(bs, slice_type);          /* slice_type */
+    bitstream_put_ue(bs, 0);                   /* pic_parameter_set_id: 0 */
+
+    if(is_idr) {
+        frame_num = 0; // !! IDR pic need to reset frame_num
+    }
+
+    bitstream_put_ui(bs, frame_num & 0x0F, seq_param->seq_fields.bits.log2_max_frame_num_minus4 + 4);    /* frame_num */
+
+    /* frame_mbs_only_flag == 1 */
+    if (!seq_param->seq_fields.bits.frame_mbs_only_flag) {
+        /* FIXME: */
+        assert(0);
+    }
+
+    if (is_idr)
+        bitstream_put_ue(bs, 0);		/* idr_pic_id: 0 */
+
+    if (seq_param->seq_fields.bits.pic_order_cnt_type == 0) {
+        bitstream_put_ui(bs, (display_frame*2) & 0x3F, seq_param->seq_fields.bits.log2_max_pic_order_cnt_lsb_minus4 + 4);
+        /* only support frame */
+    } else {
+        /* FIXME: */
+        assert(0);
+    }
+
+    /* redundant_pic_cnt_present_flag == 0 */
+    
+    /* slice type */
+    if (slice_type == SLICE_TYPE_P) {
+        bitstream_put_ui(bs, 0, 1);            /* num_ref_idx_active_override_flag: 0 */
+        /* ref_pic_list_reordering */
+        bitstream_put_ui(bs, 0, 1);            /* ref_pic_list_reordering_flag_l0: 0 */
+    } else if (slice_type == SLICE_TYPE_B) {
+        bitstream_put_ui(bs, 1, 1);            /* direct_spatial_mv_pred: 1 */
+        bitstream_put_ui(bs, 0, 1);            /* num_ref_idx_active_override_flag: 0 */
+        /* ref_pic_list_reordering */
+        bitstream_put_ui(bs, 0, 1);            /* ref_pic_list_reordering_flag_l0: 0 */
+        bitstream_put_ui(bs, 0, 1);            /* ref_pic_list_reordering_flag_l1: 0 */
+    } 
+
+    /* weighted_pred_flag == 0 */
+
+    /* dec_ref_pic_marking */
+    if (nal_ref_idc != 0) {
+        if ( is_idr) {
+            bitstream_put_ui(bs, 0, 1);            /* no_output_of_prior_pics_flag: 0 */
+            bitstream_put_ui(bs, 0, 1);            /* long_term_reference_flag: 0 */
+        } else {
+            bitstream_put_ui(bs, 0, 1);            /* adaptive_ref_pic_marking_mode_flag: 0 */
+        }
+    }
+
+    if (is_cabac && (slice_type != SLICE_TYPE_I))
+        bitstream_put_ue(bs, 0);               /* cabac_init_idc: 0 */
+
+    bitstream_put_se(bs, 0);                   /* slice_qp_delta: 0 */
+
+    if (pic_param->pic_fields.bits.deblocking_filter_control_present_flag == 1) {
+        bitstream_put_ue(bs, 0);               /* disable_deblocking_filter_idc: 0 */
+        bitstream_put_se(bs, 2);               /* slice_alpha_c0_offset_div2: 2 */
+        bitstream_put_se(bs, 2);               /* slice_beta_offset_div2: 2 */
+    }
+
+    if(is_cabac) {
+        while(bs->bit_offset % 8) {
+           bitstream_put_ui(bs, 1, 1); // cabac_alignment_one_bit
+        }
+    }
+
+}
+#endif
+
+
+int 
+build_packed_slc_buffer(unsigned char **header_buffer, int frame_num, int display_frame, int slice_type, int is_idr)
+{
+    int nal_ref_idc;    // NAL_REF_IDC_HIGH or zero
+    int nal_unit_type;  // NAL_IDR or NAL_NON_IDR
+    bitstream bs;
+    bitstream_start(&bs);
+    nal_start_code_prefix(&bs);
+
+    if(is_idr) {
+        nal_ref_idc = NAL_REF_IDC_HIGH;
+        nal_unit_type = NAL_IDR;
+    } else if(slice_type != SLICE_TYPE_B) {
+        nal_ref_idc = NAL_REF_IDC_HIGH;
+        nal_unit_type = NAL_NON_IDR;
+    } else {
+        nal_ref_idc = 0;
+        nal_unit_type = NAL_NON_IDR;
+    }
+   
+    nal_header(&bs, nal_ref_idc, nal_unit_type);
+
+    slice_header(&bs, frame_num, display_frame, slice_type, nal_ref_idc, is_idr);
+    bitstream_end(&bs);
+    *header_buffer = (unsigned char *)bs.buffer;
+    return bs.bit_offset;
+}
+
+
+#if 0
+static void 
+slice_data(bitstream *bs)
+{
+    VACodedBufferSegment *coded_buffer_segment;
+    unsigned char *coded_mem;
+    int i, slice_data_length;
+    VAStatus va_status;
+    VASurfaceStatus surface_status;
+
+    va_status = vaSyncSurface(va_dpy, surface_ids[avcenc_context.current_input_surface]);
+    CHECK_VASTATUS(va_status,"vaSyncSurface");
+
+    surface_status = 0;
+    va_status = vaQuerySurfaceStatus(va_dpy, surface_ids[avcenc_context.current_input_surface], &surface_status);
+    CHECK_VASTATUS(va_status,"vaQuerySurfaceStatus");
+
+    va_status = vaMapBuffer(va_dpy, avcenc_context.codedbuf_buf_id, (void **)(&coded_buffer_segment));
+    CHECK_VASTATUS(va_status,"vaMapBuffer");
+    coded_mem = coded_buffer_segment->buf;
+
+    slice_data_length = get_coded_bitsteam_length(coded_mem, codedbuf_size);
+
+    for (i = 0; i < slice_data_length; i++) {
+        bitstream_put_ui(bs, *coded_mem, 8);
+        coded_mem++;
+    }
+
+    vaUnmapBuffer(va_dpy, avcenc_context.codedbuf_buf_id);
+}
+
+static void 
+build_nal_slice(FILE *avc_fp, int frame_num, int display_frame, int slice_type, int is_idr)
+{
+    bitstream bs;
+
+    bitstream_start(&bs);
+    slice_data(&bs);
+    bitstream_end(&bs, avc_fp);
+}
+
+#endif
+
+void
+store_distortion_buffer()
+{
+    FILE *dist_fp;
+    unsigned char   *pbuf;
+    VAStatus        va_status;
+    size_t          n_items;
+
+    // distoriton output
+    va_status = vaMapBuffer(va_dpy, avcenc_context.fei_distortion_buf_id, (void **)(&pbuf));
+    CHECK_VASTATUS(va_status,"vaMapBuffer");
+
+    dist_fp = fopen("feidistortionout.bin", "wb");
+
+    do {
+        n_items = fwrite(pbuf, sizeof(VAEncFEIDistortionBufferH264Intel)*picture_width_in_mbs*picture_height_in_mbs, 1, dist_fp);
+    }
+    while (n_items != 1);
+
+    fclose(dist_fp);
+
+    va_status = vaUnmapBuffer(va_dpy, avcenc_context.fei_distortion_buf_id);
+    CHECK_VASTATUS(va_status,"vaUnmapBuffer");
+
+    // mv output
+    va_status = vaMapBuffer(va_dpy, avcenc_context.fei_mv_buf_id, (void **)(&pbuf));
+    CHECK_VASTATUS(va_status,"vaMapBuffer");
+
+    dist_fp = fopen("feimvout.bin", "wb");
+
+    do {
+        n_items = fwrite(pbuf, sizeof(VAMotionVectorIntel)*16*picture_width_in_mbs*picture_height_in_mbs, 1, dist_fp);
+    }
+    while (n_items != 1);
+
+    fclose(dist_fp);
+
+    va_status = vaUnmapBuffer(va_dpy, avcenc_context.fei_mv_buf_id);
+    CHECK_VASTATUS(va_status,"vaUnmapBuffer");
+
+    // mv output
+    va_status = vaMapBuffer(va_dpy, avcenc_context.fei_mbcode_buf_id, (void **)(&pbuf));
+    CHECK_VASTATUS(va_status,"vaMapBuffer");
+
+    dist_fp = fopen("feimbcodeout.bin", "wb");
+
+    do {
+        n_items = fwrite(pbuf, sizeof(VAEncFEIModeBufferH264Intel)*picture_width_in_mbs*picture_height_in_mbs, 1, dist_fp);
+    }
+    while (n_items != 1);
+
+    fclose(dist_fp);
+
+    va_status = vaUnmapBuffer(va_dpy, avcenc_context.fei_mbcode_buf_id);
+    CHECK_VASTATUS(va_status,"vaUnmapBuffer");
+
+    return;
+}
+
+static int
+store_coded_buffer(FILE *avc_fp, int slice_type)
+{
+    VACodedBufferSegment *coded_buffer_segment;
+    unsigned char *coded_mem;
+    int slice_data_length;
+    VAStatus va_status;
+    VASurfaceStatus surface_status;
+    size_t w_items;
+
+    va_status = vaSyncSurface(va_dpy, surface_ids[avcenc_context.current_input_surface]);
+    CHECK_VASTATUS(va_status,"vaSyncSurface");
+
+    surface_status = 0;
+    va_status = vaQuerySurfaceStatus(va_dpy, surface_ids[avcenc_context.current_input_surface], &surface_status);
+    CHECK_VASTATUS(va_status,"vaQuerySurfaceStatus");
+
+    va_status = vaMapBuffer(va_dpy, avcenc_context.codedbuf_buf_id, (void **)(&coded_buffer_segment));
+    CHECK_VASTATUS(va_status,"vaMapBuffer");
+    coded_mem = coded_buffer_segment->buf;
+   
+    
+
+    if (coded_buffer_segment->status & VA_CODED_BUF_STATUS_SLICE_OVERFLOW_MASK) {
+        if (slice_type == SLICE_TYPE_I)
+            avcenc_context.codedbuf_i_size *= 2;
+        else
+            avcenc_context.codedbuf_pb_size *= 2;
+
+        vaUnmapBuffer(va_dpy, avcenc_context.codedbuf_buf_id);
+        return -1;
+    }
+
+    slice_data_length = coded_buffer_segment->size;
+    printf("SIZE=%d\n",slice_data_length);
+    do {
+        w_items = fwrite(coded_mem, slice_data_length, 1, avc_fp);
+    } while (w_items != 1);
+#if 0    
+    if (slice_type == SLICE_TYPE_I) {
+        if (avcenc_context.codedbuf_i_size > slice_data_length * 3 / 2) {
+            avcenc_context.codedbuf_i_size = slice_data_length * 3 / 2;
+        }
+        
+        if (avcenc_context.codedbuf_pb_size < slice_data_length) {
+            avcenc_context.codedbuf_pb_size = slice_data_length;
+        }
+    } else {
+        if (avcenc_context.codedbuf_pb_size > slice_data_length * 3 / 2) {
+            avcenc_context.codedbuf_pb_size = slice_data_length * 3 / 2;
+        }
+    }
+#endif
+    vaUnmapBuffer(va_dpy, avcenc_context.codedbuf_buf_id);
+
+    return 0;
+}
+
+static void
+encode_picture(FILE *yuv_fp, FILE *avc_fp,
+               int frame_num, int display_num,
+               int is_idr,
+               int slice_type, int next_is_bpic,
+               int next_display_num)
+{
+    VAStatus va_status;
+    int ret = 0, codedbuf_size;
+    
+    printf("\r ---> encoded frame %d, type=%d ...\n", frame_num, slice_type);
+    begin_picture(yuv_fp, frame_num, display_num, slice_type, is_idr);
+
+    //if (next_display_num < frame_number) {
+    if (1) {
+        int index;
+
+        /* prepare for next frame */
+        /*
+        if (avcenc_context.current_input_surface == SID_INPUT_PICTURE_0)
+            index = SID_INPUT_PICTURE_1;
+        else if (avcenc_context.current_input_surface == SID_INPUT_PICTURE_1)
+            index = SID_INPUT_PICTURE_2;
+        else if (avcenc_context.current_input_surface == SID_INPUT_PICTURE_2)
+            index = SID_INPUT_PICTURE_3;
+        else if (avcenc_context.current_input_surface == SID_INPUT_PICTURE_3)
+            index = SID_INPUT_PICTURE_4;
+        else
+            index = SID_INPUT_PICTURE_0;
+        */
+        index=next_input_id;
+        if ( next_display_num >= frame_number )
+            next_display_num = frame_number - 1;
+        fseek(yuv_fp, frame_size * display_num, SEEK_SET);
+
+        avcenc_context.upload_thread_param.yuv_fp = yuv_fp;
+        avcenc_context.upload_thread_param.surface_id = surface_ids[index];
+#if 0
+        avcenc_context.upload_thread_value = pthread_create(&avcenc_context.upload_thread_id,
+                                                            NULL,
+                                                            upload_thread_function, 
+                                                            (void*)&avcenc_context.upload_thread_param);
+#else
+        upload_thread_function((void *)(&avcenc_context.upload_thread_param));
+#endif
+    }
+    usleep(10000);
+    do {
+        avcenc_destroy_buffers(&avcenc_context.pic_param_buf_id, 1);
+        avcenc_destroy_buffers(&avcenc_context.packed_pic_buf_id, 1);
+
+        if (SLICE_TYPE_I == slice_type) {
+            codedbuf_size = avcenc_context.codedbuf_i_size;
+        } else {
+            codedbuf_size = avcenc_context.codedbuf_pb_size;
+        }
+
+        /* coded buffer */
+        va_status = vaCreateBuffer(va_dpy,
+                                   avcenc_context.context_id,
+                                   VAEncCodedBufferType,
+                                   codedbuf_size, 1, NULL,
+                                   &avcenc_context.codedbuf_buf_id);
+        CHECK_VASTATUS(va_status,"vaCreateBuffer");
+
+        /* picture parameter set */
+        avcenc_update_picture_parameter(slice_type, frame_num, display_num, is_idr);
+
+        avcenc_render_picture();
+        usleep(10000);
+
+        if(stats_en)
+            store_distortion_buffer();
+        ret = store_coded_buffer(avc_fp, slice_type);
+    } while (ret);
+
+    end_picture(slice_type, next_is_bpic);
+}
+
+static void encode_pb_pictures(FILE *yuv_fp, FILE *avc_fp, int f, int nbframes, int next_f)
+{
+    int i;
+    encode_picture(yuv_fp, avc_fp,
+                   enc_frame_number, f + nbframes,
+                   0,
+                   SLICE_TYPE_P, 1, f);
+
+    for( i = 0; i < nbframes - 1; i++) {
+        encode_picture(yuv_fp, avc_fp,
+                       enc_frame_number + 1, f + i,
+                       0,
+                       SLICE_TYPE_B, 1, f + i + 1);
+    }
+    
+    encode_picture(yuv_fp, avc_fp,
+                   enc_frame_number + 1, f + nbframes - 1,
+                   0,
+                   SLICE_TYPE_B, 0, next_f);
+}
+
+static void show_help()
+{
+    //printf("Usage: avnenc <width> <height> <input_yuvfile> <output_avcfile> [qp=qp_value|fb=framebitrate] [mode=0(I frames only)/1(I and P frames)/2(I, P and B frames)\n");
+    printf("Usage: feiavnenc -size <width> <height> -i <input_yuvfile> -o <output_avcfile> -qp [QP value] -fb [frame_bit_rate] -mode [mode=0(I frames only)/1(I and P frames)/2(I, P and B frames) -frames [0 means encoded all frames, or encoded real frames] -stats <enable output statistics or not>\n");
+}
+
+static void avcenc_context_seq_param_init(VAEncSequenceParameterBufferH264 *seq_param,
+                                          int width, int height)
+
+{
+    int width_in_mbs = (width + 15) / 16;
+    int height_in_mbs = (height + 15) / 16;
+    int frame_cropping_flag = 0;
+    int frame_crop_bottom_offset = 0;
+
+    seq_param->seq_parameter_set_id = 0;
+    seq_param->level_idc = 41;
+    seq_param->intra_period = intra_period;
+    seq_param->ip_period = ip_period;   /* FIXME: ??? */
+    seq_param->max_num_ref_frames = 4;
+    seq_param->picture_width_in_mbs = width_in_mbs;
+    seq_param->picture_height_in_mbs = height_in_mbs;
+    seq_param->seq_fields.bits.frame_mbs_only_flag = 1;
+    
+    if (frame_bit_rate > 0)
+        seq_param->bits_per_second = 1024 * frame_bit_rate; /* use kbps as input */
+    else
+        seq_param->bits_per_second = 0;
+    
+    seq_param->time_scale = 900;
+    seq_param->num_units_in_tick = 15;			/* Tc = num_units_in_tick / time_sacle */
+
+    if (height_in_mbs * 16 - height) {
+        frame_cropping_flag = 1;
+        frame_crop_bottom_offset = 
+            (height_in_mbs * 16 - height) / (2 * (!seq_param->seq_fields.bits.frame_mbs_only_flag + 1));
+    }
+
+    seq_param->frame_cropping_flag = frame_cropping_flag;
+    seq_param->frame_crop_left_offset = 0;
+    seq_param->frame_crop_right_offset = 0;
+    seq_param->frame_crop_top_offset = 0;
+    seq_param->frame_crop_bottom_offset = frame_crop_bottom_offset;
+
+    seq_param->seq_fields.bits.pic_order_cnt_type = 0;
+    seq_param->seq_fields.bits.direct_8x8_inference_flag = 1;
+    
+    seq_param->seq_fields.bits.log2_max_frame_num_minus4 = 0;
+
+    seq_param->seq_fields.bits.log2_max_pic_order_cnt_lsb_minus4 = 2;
+	
+    if (frame_bit_rate > 0)
+        seq_param->vui_parameters_present_flag = 1;	//HRD info located in vui
+    else
+        seq_param->vui_parameters_present_flag = 0;
+}
+
+static void avcenc_context_pic_param_init(VAEncPictureParameterBufferH264 *pic_param)
+{
+    pic_param->seq_parameter_set_id = 0;
+    pic_param->pic_parameter_set_id = 0;
+
+    pic_param->last_picture = 0;
+    pic_param->frame_num = 0;
+    
+    pic_param->pic_init_qp = (qp_value >= 0 ?  qp_value : 26);
+    pic_param->num_ref_idx_l0_active_minus1 = 0;
+    pic_param->num_ref_idx_l1_active_minus1 = 0;
+
+    pic_param->pic_fields.bits.idr_pic_flag = 0;
+    pic_param->pic_fields.bits.reference_pic_flag = 0;
+    pic_param->pic_fields.bits.entropy_coding_mode_flag = ENTROPY_MODE_CABAC;
+    pic_param->pic_fields.bits.weighted_pred_flag = 0;
+    pic_param->pic_fields.bits.weighted_bipred_idc = 0;
+    pic_param->pic_fields.bits.transform_8x8_mode_flag = 1;
+    pic_param->pic_fields.bits.deblocking_filter_control_present_flag = 1;
+}
+
+static void avcenc_context_init(int width, int height)
+{
+    int i;
+    memset(&avcenc_context, 0, sizeof(avcenc_context));
+    avcenc_context.profile = VAProfileH264High;
+    avcenc_context.seq_param_buf_id = VA_INVALID_ID;
+    avcenc_context.pic_param_buf_id = VA_INVALID_ID;
+    avcenc_context.packed_seq_header_param_buf_id = VA_INVALID_ID;
+    avcenc_context.packed_seq_buf_id = VA_INVALID_ID;
+    avcenc_context.packed_pic_header_param_buf_id = VA_INVALID_ID;
+    avcenc_context.packed_pic_buf_id = VA_INVALID_ID;
+    avcenc_context.packed_slc_header_param_buf_id = VA_INVALID_ID;
+    avcenc_context.packed_slc_buf_id = VA_INVALID_ID;
+    avcenc_context.codedbuf_buf_id = VA_INVALID_ID;
+    avcenc_context.fei_distortion_buf_id = VA_INVALID_ID;
+    avcenc_context.fei_mv_buf_id         = VA_INVALID_ID;
+    avcenc_context.fei_mbcode_buf_id     = VA_INVALID_ID;
+    avcenc_context.misc_parameter_fei_pic_control_buf_id = VA_INVALID_ID;
+    avcenc_context.fei_mbctrl_buf_id = VA_INVALID_ID;
+    avcenc_context.fei_mvpred_buf_id = VA_INVALID_ID;
+    avcenc_context.misc_parameter_hrd_buf_id = VA_INVALID_ID;
+    avcenc_context.misc_parameter_rc_buf_id = VA_INVALID_ID;
+    avcenc_context.misc_parameter_fr_buf_id = VA_INVALID_ID;
+    avcenc_context.codedbuf_i_size = width * height * 3 / 2;
+    avcenc_context.codedbuf_pb_size = width * height * 3 / 2;
+    avcenc_context.fei_mbctrl_buf_size = 0;
+    avcenc_context.fei_mvpred_buf_size = 0;
+    avcenc_context.current_input_surface = SID_INPUT_PICTURE_0;
+    avcenc_context.upload_thread_value = -1;
+
+    if (qp_value == -1)
+        avcenc_context.rate_control_method = VA_RC_CBR;
+    else if (qp_value == -2)
+        avcenc_context.rate_control_method = VA_RC_VBR;
+    else {
+        assert(qp_value >= 0 && qp_value <= 51);
+        avcenc_context.rate_control_method = VA_RC_CQP;
+    }
+
+    for (i = 0; i < MAX_SLICES; i++) {
+        avcenc_context.slice_param_buf_id[i] = VA_INVALID_ID;
+    }
+
+    avcenc_context_seq_param_init(&avcenc_context.seq_param, width, height);
+    avcenc_context_pic_param_init(&avcenc_context.pic_param);
+
+    // init FEI pic level parameters
+    avcenc_context.fei_pic_param.function            = VA_ENC_FUNCTION_ENC_PAK_INTEL;
+    avcenc_context.fei_pic_param.num_mv_predictors   = 1;
+    avcenc_context.fei_pic_param.max_len_sp          = 57;
+    avcenc_context.fei_pic_param.len_sp              = 57;
+    avcenc_context.fei_pic_param.sub_mb_part_mask    = 0x77;  //1110111b
+    avcenc_context.fei_pic_param.intra_part_mask     = 0;
+    avcenc_context.fei_pic_param.multi_pred_l0       = 0;
+    avcenc_context.fei_pic_param.multi_pred_l1       = 0;
+    avcenc_context.fei_pic_param.sub_pel_mode        = 3;
+    avcenc_context.fei_pic_param.inter_sad           = 2;
+    avcenc_context.fei_pic_param.intra_sad           = 2;
+    avcenc_context.fei_pic_param.distortion_type     = 0;
+    avcenc_context.fei_pic_param.repartition_check_enable = 0;
+    avcenc_context.fei_pic_param.adaptive_search     = 1;
+    avcenc_context.fei_pic_param.mv_predictor_enable = 1;
+    avcenc_context.fei_pic_param.mb_qp               = 0;
+    avcenc_context.fei_pic_param.mb_input            = 1;
+    avcenc_context.fei_pic_param.mb_size_ctrl        = 0;
+    avcenc_context.fei_pic_param.ref_width           = 48;
+    avcenc_context.fei_pic_param.ref_height          = 40;
+    avcenc_context.fei_pic_param.search_window       = 0;
+
+    avcenc_context.fei_pic_param.distortion          = avcenc_context.fei_distortion_buf_id;
+    avcenc_context.fei_pic_param.mv_data             = avcenc_context.fei_mv_buf_id;
+    avcenc_context.fei_pic_param.mb_code_data        = avcenc_context.fei_mbcode_buf_id;
+
+    if(avcenc_context.fei_pic_param.mb_input)
+        avcenc_context.fei_mbctrl_buf_size   = picture_width_in_mbs * picture_height_in_mbs * sizeof(VAEncFEIMBControlH264Intel);
+    if(avcenc_context.fei_pic_param.mv_predictor_enable)
+        avcenc_context.fei_mvpred_buf_size   = picture_width_in_mbs * picture_height_in_mbs * sizeof(VAEncMVPredictorH264Intel);
+}
+
+int main(int argc, char *argv[])
+{
+    unsigned char inputyuv[256];
+    unsigned char outputstream[256];
+    int i = 0;
+    int f;
+    FILE *yuv_fp;
+    FILE *avc_fp;
+    long file_size;
+    int enc_frames = 0;
+    int i_frame_only=0,i_p_frame_only=1;
+    int mode_value;
+    struct timeval tpstart,tpend; 
+    float  timeuse;
+    
+    picture_width    = 352;
+    picture_height   = 288;
+    qp_value         = 28;      //default const QP mode
+    for (i=1; i<argc; i++)
+    {
+        //printf("---> i=%d, argv[i]=%s \n", i, argv[i]);
+        if(strcmp(argv[i],"-size")==0)
+        {
+            if(i+2<argc)
+            {
+                picture_width  = atoi(argv[i+1]);
+                picture_height = atoi(argv[i+2]);
+                i += 2;
+            } else
+            {
+                show_help();
+                return -1;
+            }
+        } else if(strcmp(argv[i],"-i")==0)
+        {
+            if(i+1<argc)
+            {
+                strcpy(inputyuv, argv[i+1]);
+                i++;
+                //printf("---> argv[i]=%s, inputyuv=%s \n", argv[i+1], inputyuv);
+            } else
+            {
+                show_help();
+                return -1;
+            }
+        } else if(strcmp(argv[i],"-o")==0)
+        {
+            if(i+1<argc)
+            {
+                strcpy(outputstream, argv[i+1]);
+                i++;
+                //printf("---> argv[i]=%s, outputstream=%s \n", argv[i+1], outputstream);
+            } else
+            {
+                show_help();
+                return -1;
+            }
+
+        } else if(strcmp(argv[i],"-qp")==0)
+        {
+            if(i+1<argc)
+            {
+                qp_value  = atoi(argv[i+1]);
+                i++;
+            } else
+            {
+                show_help();
+                return -1;
+            }
+        } else if(strcmp(argv[i],"-fb")==0)
+        {
+            if(i+1<argc)
+            {
+                frame_bit_rate  = atoi(argv[i+1]);
+                printf("frame_bit_rate: disregarded, FEI should have no notion for bitrate!\n");
+                frame_bit_rate = -1;  // force to -1 to match correct logic
+                i++;
+            } else
+            {
+                show_help();
+                return -1;
+            }
+        } else if(strcmp(argv[i],"-mode")==0)
+        {
+            if(i+1<argc)
+            {
+                mode_value  = atoi(argv[i+1]);
+                i++;
+                if ( mode_value == 0 ) 
+                {
+                    i_frame_only   = 1;
+                    i_p_frame_only = 0;
+                    intra_period   = 1;
+                    ip_period      = 1;
+                } else if ( mode_value == 1 )
+                {
+                    i_frame_only   = 0;
+                    i_p_frame_only = 1;
+                    intra_period   = 30;
+                    ip_period      = 1;
+                }
+                else if ( mode_value == 2 ) 
+                {
+                    i_frame_only   = 0;
+                    i_p_frame_only = 0;
+                    intra_period   = 30;
+                    ip_period      = 3;
+                }
+                else {
+                    printf("mode_value=%d\n",mode_value);
+                    show_help();
+                    return -1;
+                }
+            }
+        } else 
+        if(strcmp(argv[i],"-frames")==0)
+        {
+            if(i+1<argc)
+            {
+                enc_frames = atoi(argv[i+1]);
+                i++;
+            } else
+            {
+                show_help();
+                return -1;
+            }
+        } else
+        if(strcmp(argv[i],"-stats")==0)
+        {
+            if(i+1<argc)
+            {
+                stats_en = (atoi(argv[i+1]) != 0);
+                i++;
+            } else
+            {
+                show_help();
+                return -1;
+            }
+        } else
+        {
+            show_help();
+            return -1;
+        }
+    }
+
+    picture_width_in_mbs  = (picture_width  + 15) / 16;
+    picture_height_in_mbs = (picture_height + 15) / 16;
+
+    if((qp_value == -1) && (frame_bit_rate = -1))
+    {
+        show_help();
+        return -1;
+    } else if (qp_value > 51)
+    {
+        qp_value = 51;
+    } else if (qp_value < 0) {
+        qp_value = 0;
+    }
+
+    yuv_fp = fopen(inputyuv,"rb");
+    if ( yuv_fp == NULL){
+        printf("Can't open input YUV file\n");
+        show_help();
+        return -1;
+    }
+    fseek(yuv_fp,0l, SEEK_END);
+    file_size = ftell(yuv_fp);
+    frame_size = picture_width * picture_height +  ((picture_width * picture_height) >> 1) ;
+
+    if ( (file_size < frame_size) || (file_size % frame_size) ) {
+        fclose(yuv_fp);
+        printf("The YUV file's size is not correct\n");
+        show_help();
+        return -1;
+    }
+    if(enc_frames > 0 )
+        file_size = enc_frames * frame_size;   //default 0 run all frames, or run enc_frames frames. 
+    frame_number = file_size / frame_size;
+    fseek(yuv_fp, 0l, SEEK_SET);
+
+    avc_fp = fopen(outputstream, "wb");	
+    if ( avc_fp == NULL) {
+        fclose(yuv_fp);
+        printf("Can't open output avc file\n");
+        show_help();
+        return -1;
+    }	
+    gettimeofday(&tpstart,NULL);	
+    avcenc_context_init(picture_width, picture_height);
+    create_encode_pipe();
+    alloc_encode_resource(yuv_fp);
+
+    enc_frame_number = 0;
+    for ( f = 0; f < frame_number; ) {		//picture level loop
+        static int const frame_type_pattern[][2] = { {SLICE_TYPE_I,1}, 
+                                                     {SLICE_TYPE_P,3}, {SLICE_TYPE_P,3},{SLICE_TYPE_P,3},
+                                                     {SLICE_TYPE_P,3}, {SLICE_TYPE_P,3},{SLICE_TYPE_P,3},
+                                                     {SLICE_TYPE_P,3}, {SLICE_TYPE_P,3},{SLICE_TYPE_P,3},
+                                                     {SLICE_TYPE_P,2} };
+
+        if ( i_frame_only ) {
+            encode_picture(yuv_fp, avc_fp,enc_frame_number, f, f==0, SLICE_TYPE_I, 0, f+1);
+            f++;
+            enc_frame_number++;
+        } else if ( i_p_frame_only ) {
+            if ( (f % intra_period) == 0 ) {
+                encode_picture(yuv_fp, avc_fp,enc_frame_number, f, 1, SLICE_TYPE_I, 0, f+1); 
+                f++;
+                enc_frame_number++;
+            } else {
+                encode_picture(yuv_fp, avc_fp,enc_frame_number, f, 0, SLICE_TYPE_P, 0, f+1); 
+                f++;
+                enc_frame_number++;
+            }
+        } else { // follow the i,p,b pattern
+            static int fcurrent = 0;
+            int fnext;
+            
+            fcurrent = fcurrent % (sizeof(frame_type_pattern)/sizeof(int[2]));
+            fnext = (fcurrent+1) % (sizeof(frame_type_pattern)/sizeof(int[2]));
+            
+            if ( frame_type_pattern[fcurrent][0] == SLICE_TYPE_I ) {
+                encode_picture(yuv_fp, avc_fp,enc_frame_number, f, 1, SLICE_TYPE_I, 0, 
+                        f+frame_type_pattern[fnext][1]);  
+
+                f++;
+                enc_frame_number++;
+            } else {
+                encode_pb_pictures(yuv_fp, avc_fp, f, frame_type_pattern[fcurrent][1]-1, 
+                        f + frame_type_pattern[fcurrent][1] + frame_type_pattern[fnext][1] -1 );
+                f += frame_type_pattern[fcurrent][1];
+                enc_frame_number++;
+            }
+ 
+            fcurrent++;
+        }
+        //printf("\r ---> encoded %d/%d ...", f+1, frame_number);
+        fflush(stdout);
+    }
+
+    gettimeofday(&tpend,NULL);
+    timeuse=1000000*(tpend.tv_sec-tpstart.tv_sec)+ tpend.tv_usec-tpstart.tv_usec;
+    timeuse/=1000000;
+    printf("\ndone!\n");
+    printf("encode %d frames in %f secondes, FPS is %.1f\n",frame_number, timeuse, frame_number/timeuse);
+    release_encode_resource();
+    destory_encode_pipe();
+
+    fclose(yuv_fp);
+    fclose(avc_fp);
+
+    return 0;
+}
diff --git a/test/encode/preenc.c b/test/encode/preenc.c
new file mode 100644
index 0000000..a605e49
--- /dev/null
+++ b/test/encode/preenc.c
@@ -0,0 +1,895 @@
+/*
+ * Simple AVC encoder based on libVA.
+ *
+ * Usage:
+ * ./preenc -size <width> <height> -i <input_yuvfile> -mode [mode=0(P frames only)/mode=1(B frames)]
+ */  
+
+#include <stdio.h>
+#include <string.h>
+#include <stdlib.h>
+#include <getopt.h>
+#include <X11/Xlib.h>
+
+#include <unistd.h>
+
+#include <sys/time.h>
+#include <sys/types.h>
+#include <sys/stat.h>
+#include <fcntl.h>
+#include <assert.h>
+#include <time.h>
+
+#include <pthread.h>
+
+#include <va/va.h>
+#include <va/va_enc_h264.h>
+#include <va/vendor/intel/va_intel_statistics.h>
+#include <va/vendor/intel/va_intel_fei.h>
+#include <va/va_x11.h>
+#include "va/drm/va_drm.h"
+
+#define SLICE_TYPE_P            0
+#define SLICE_TYPE_B            1
+#define SLICE_TYPE_I            2
+
+#define CHECK_VASTATUS(va_status,func)                                  \
+    if (va_status != VA_STATUS_SUCCESS) {                               \
+        fprintf(stderr,"%s:%s (%d) failed,exit\n", __func__, func, __LINE__); \
+        exit(1);                                                        \
+    }
+
+#define MIN(a, b) ((a)>(b)?(b):(a))
+
+static VADisplay va_dpy;
+
+static int picture_width,  picture_width_in_mbs;
+static int picture_height, picture_height_in_mbs;
+static int frame_size;
+static unsigned char *newImageBuffer  = 0;
+static unsigned char *pMvPredInBuffer = 0;
+static unsigned char *pMbQpInBuffer   = 0;
+
+static VAStatsStatisticsParameter16x16Intel pic_param_back;
+
+struct upload_thread_param
+{
+    FILE *yuv_fp;
+    VASurfaceID surface_id;
+};
+
+// mv_predictor or qp
+struct upload_input_thread_param
+{
+    FILE *input_fp;
+    int  size;
+    VASurfaceID surface_id;
+};
+
+static struct {
+    VAProfile   profile;
+    VAStatsStatisticsParameter16x16Intel pic_param;
+    VAContextID context_id;
+    VAConfigID  config_id;
+    VABufferID  pic_param_buf_id;
+    VABufferID  mvpred_buf_id;
+    VABufferID  mbqp_buf_id;
+    VABufferID  out_buf_id[2];     // 0 for mv out and 1 for stats out buffer if both mv and stats out enabled
+
+    int         mvpred_buf_size;
+    int         mbqp_buf_size;
+    int         mvout_buf_size;
+    int         statsout_buf_size;
+    int         current_input_surface;
+    struct upload_thread_param       upload_thread_param;
+    struct upload_input_thread_param upload_input_thread_param;
+    pthread_t upload_thread_id;
+    int upload_thread_value;
+} preenc_context;
+
+
+static FILE *yuv_fp = 0;
+static FILE *mvpredin_fp = 0;
+static FILE *qpin_fp = 0;
+
+static FILE *mvout_fp = 0;
+static FILE *statsout_fp = 0;
+
+static void show_help()
+{
+    printf("Usage: preenc -size <width> <height> -i <input_yuvfile> -frames <#frames to process, >=2, default runs all frames> -bdistance <distance between I/P and P, 1 means no B> -gop_lenght <I period>\n");
+}
+
+static void preenc_context_init()
+{
+    memset(&preenc_context, 0, sizeof(preenc_context));
+    preenc_context.profile               = VAProfileNone;
+    preenc_context.pic_param_buf_id      = VA_INVALID_ID;
+    preenc_context.mvpred_buf_id         = VA_INVALID_ID;
+    preenc_context.mbqp_buf_id           = VA_INVALID_ID;
+    preenc_context.out_buf_id[0]         = VA_INVALID_ID;
+    preenc_context.out_buf_id[1]         = VA_INVALID_ID;
+    preenc_context.mvpred_buf_size       = 0;
+    preenc_context.mbqp_buf_size         = 0;
+    preenc_context.mvout_buf_size        = picture_width_in_mbs * picture_height_in_mbs * 16 * sizeof(VAMotionVectorIntel);
+    preenc_context.statsout_buf_size     = picture_width_in_mbs * picture_height_in_mbs * sizeof(VAStatsStatistics16x16Intel);
+
+    preenc_context.current_input_surface = 2;//SID_REFERENCE_PICTURE_L0;
+    preenc_context.upload_thread_value   = -1;
+    
+
+    // init picture level params
+    preenc_context.pic_param.num_past_references       = 1;
+    preenc_context.pic_param.num_future_references     = 0;
+    preenc_context.pic_param.frame_qp                  = 27;
+    preenc_context.pic_param.len_sp                    = 57;
+    preenc_context.pic_param.max_len_sp                = 57;
+    preenc_context.pic_param.sub_mb_part_mask          = 0x77; //1110111b
+    preenc_context.pic_param.sub_pel_mode              = 3;
+    preenc_context.pic_param.inter_sad                 = 2;
+    preenc_context.pic_param.intra_sad                 = 2;
+    preenc_context.pic_param.adaptive_search           = 1;
+    preenc_context.pic_param.mv_predictor_ctrl         = 1;
+    preenc_context.pic_param.mb_qp                     = 1;
+    preenc_context.pic_param.ft_enable                 = 0;
+    preenc_context.pic_param.ref_width                 = 48;
+    preenc_context.pic_param.ref_height                = 40;
+    preenc_context.pic_param.search_window             = 0;
+    preenc_context.pic_param.disable_mv_output         = 0;
+    preenc_context.pic_param.disable_statistics_output = 0;
+    preenc_context.pic_param.interlaced                = 0;
+    
+    if(preenc_context.pic_param.mv_predictor_ctrl != 0)
+        preenc_context.mvpred_buf_size   = picture_width_in_mbs * picture_height_in_mbs * sizeof(VAMotionVectorIntel);
+    if(preenc_context.pic_param.mb_qp != 0)
+        preenc_context.mbqp_buf_size     = picture_width_in_mbs * picture_height_in_mbs * sizeof(VAEncQpBufferH264);
+}
+
+static int drm_fd = -1;
+static void create_preenc_pipe()
+{
+    VAEntrypoint entrypoints[5];
+    int num_entrypoints = 1;
+    VAConfigAttrib attrib[2];
+    int major_ver, minor_ver;
+    VAStatus va_status;
+    
+    drm_fd = open("/dev/dri/card0", O_RDWR);    
+    if (drm_fd < 0) 
+    {        
+        fprintf(stderr, "error: can't open DRM connection!\n");        
+        return;    
+    }    
+    va_dpy = vaGetDisplayDRM(drm_fd);
+    //x11_display = XOpenDisplay(":0.0");
+    //assert(x11_display);
+    
+    //va_dpy = vaGetDisplay(x11_display);
+    va_status = vaInitialize(va_dpy, &major_ver, &minor_ver);
+    CHECK_VASTATUS(va_status, "vaInitialize");
+
+    vaQueryConfigEntrypoints(va_dpy, preenc_context.profile, entrypoints, 
+                             &num_entrypoints);
+
+    /* find out the format for the render target, and rate control mode */
+    attrib[0].type = VAConfigAttribStatisticsIntel;
+    vaGetConfigAttributes(va_dpy, preenc_context.profile, VAEntrypointStatisticsIntel,
+                          &attrib[0], 1);
+
+    if ((attrib[0].value & 0xff) == 0) {
+        /* not support past/future reference */
+        assert(0);
+    }
+
+    /* set to default mode */
+    attrib[0].value =  preenc_context.pic_param.num_past_references | 
+                      (preenc_context.pic_param.num_future_references << 4) |
+                      ((2 - preenc_context.pic_param.disable_mv_output - preenc_context.pic_param.disable_statistics_output) << 8) |
+                      (preenc_context.pic_param.interlaced << 11); 
+
+    va_status = vaCreateConfig(va_dpy, preenc_context.profile, VAEntrypointStatisticsIntel,
+                               &attrib[0], 1, &preenc_context.config_id);
+    CHECK_VASTATUS(va_status, "vaCreateConfig");
+    
+}
+
+static void destory_preenc_pipe()
+{
+    vaDestroyContext(va_dpy,preenc_context.context_id);
+    vaDestroyConfig(va_dpy,preenc_context.config_id);
+    vaTerminate(va_dpy);
+    if (drm_fd < 0)        
+        return;    
+    close(drm_fd);    
+    drm_fd = -1;
+    //XCloseDisplay(x11_display);
+}
+
+/***************************************************
+ *
+ *  The preenc pipe resource define 
+ *
+ ***************************************************/
+#define SID_INPUT_PICTURE_0                     0
+#define SID_INPUT_PICTURE_1                     1
+#define SID_REFERENCE_PICTURE_L0                2
+#define SID_REFERENCE_PICTURE_L1                3
+
+
+#define SID_NUMBER                              4
+static  VASurfaceID surface_ids[SID_NUMBER];
+
+static int frame_number;
+
+/***************************************************/
+#ifndef VA_FOURCC_I420
+#define VA_FOURCC_I420          0x30323449
+#endif
+
+static void upload_yuv_to_surface(FILE *yuv_fp, VASurfaceID surface_id)
+{
+    VAImage surface_image;
+    VAStatus va_status;
+    void *surface_p = NULL;
+    unsigned char *y_src, *u_src, *v_src;
+    unsigned char *y_dst, *u_dst, *v_dst;
+    int y_size = picture_width * picture_height;
+    int u_size = (picture_width >> 1) * (picture_height >> 1);
+    int row, col;
+    size_t n_items;
+
+    do {
+        n_items = fread(newImageBuffer, frame_size, 1, yuv_fp);
+    } while (n_items != 1);
+
+    va_status = vaDeriveImage(va_dpy, surface_id, &surface_image);
+    CHECK_VASTATUS(va_status,"vaDeriveImage");
+
+    vaMapBuffer(va_dpy, surface_image.buf, &surface_p);
+    assert(VA_STATUS_SUCCESS == va_status);
+        
+    y_src = newImageBuffer;
+    u_src = newImageBuffer + y_size; /* UV offset for NV12 */
+    v_src = newImageBuffer + y_size + u_size;
+
+    y_dst = surface_p + surface_image.offsets[0];
+    u_dst = surface_p + surface_image.offsets[1]; /* UV offset for NV12 */
+    v_dst = surface_p + surface_image.offsets[2];
+
+    /* Y plane */
+    memset(y_dst, 0, surface_image.pitches[0]*surface_image.height);
+    for (row = 0; row < picture_height; row++) {
+        memcpy(y_dst, y_src, picture_width);
+        y_dst += surface_image.pitches[0];
+        y_src += picture_width;
+    }
+
+    if (surface_image.format.fourcc == VA_FOURCC_NV12) { /* UV plane */
+        memset(u_dst, 0, surface_image.pitches[1]*(surface_image.height/2));
+        for (row = 0; row < picture_height / 2; row++) {
+            for (col = 0; col < picture_width / 2; col++) {
+                u_dst[col * 2]     = u_src[col];
+                u_dst[col * 2 + 1] = v_src[col];
+            }
+
+            u_dst += surface_image.pitches[1];
+            u_src += (picture_width / 2);
+            v_src += (picture_width / 2);
+        }
+    } else if (surface_image.format.fourcc == VA_FOURCC_YV12 ||
+               surface_image.format.fourcc == VA_FOURCC_I420) {
+        const int U = surface_image.format.fourcc == VA_FOURCC_I420 ? 1 : 2;
+        const int V = surface_image.format.fourcc == VA_FOURCC_I420 ? 2 : 1;
+
+        u_dst = surface_p + surface_image.offsets[U];
+        v_dst = surface_p + surface_image.offsets[V];
+
+        for (row = 0; row < picture_height / 2; row++) {
+            memcpy(u_dst, u_src, picture_width / 2);
+            memcpy(v_dst, v_src, picture_width / 2);
+            u_dst += surface_image.pitches[U];
+            v_dst += surface_image.pitches[V];
+            u_src += (picture_width / 2);
+            v_src += (picture_width / 2);
+        }
+    }
+
+    vaUnmapBuffer (va_dpy, surface_image.buf);
+    vaDestroyImage(va_dpy, surface_image.image_id);
+}
+
+static void *
+upload_thread_function(void *data)
+{
+    struct upload_thread_param *param = data;
+
+    upload_yuv_to_surface(param->yuv_fp, param->surface_id);
+
+    return NULL;
+}
+
+static void *
+upload_input_thread_function(void *data)
+{
+//    struct upload_inpyut_thread_param *param = data;
+
+//    upload_input_to_surface(param->input_fp, param->surface_id, param->size);
+
+    return NULL;
+}
+
+static void alloc_preenc_resource()
+{
+    VAStatus va_status;
+
+    // Create surface
+    va_status = vaCreateSurfaces(
+        va_dpy,
+        VA_RT_FORMAT_YUV420, picture_width_in_mbs*16, picture_height_in_mbs*16,
+        &surface_ids[0], SID_NUMBER,
+        NULL, 0
+    );
+
+    CHECK_VASTATUS(va_status, "vaCreateSurfaces");
+
+    /* Create a context for this decode pipe */
+    va_status = vaCreateContext(va_dpy, preenc_context.config_id,
+                                picture_width_in_mbs*16, picture_height_in_mbs*16,
+                                VA_PROGRESSIVE, 
+                                &surface_ids[0], 
+                                SID_NUMBER,
+                                &preenc_context.context_id);
+    CHECK_VASTATUS(va_status, "vaCreateContext");
+    newImageBuffer  = (unsigned char *)malloc(frame_size);
+    
+    if(preenc_context.pic_param.mv_predictor_ctrl != 0)
+    {
+        pMvPredInBuffer = (unsigned char *)malloc(preenc_context.mvpred_buf_size);
+        memset(pMvPredInBuffer, 0, preenc_context.mvpred_buf_size);
+    }
+
+    if(preenc_context.pic_param.mb_qp != 0)
+    {
+        pMbQpInBuffer   = (unsigned char *)malloc(preenc_context.mbqp_buf_size);
+        memset(pMbQpInBuffer,   0, preenc_context.mbqp_buf_size);
+    }
+
+}
+
+static int preenc_destroy_buffers(VABufferID *va_buffers, unsigned int num_va_buffers)
+{
+    VAStatus va_status;
+    unsigned int i;
+
+    for (i = 0; i < num_va_buffers; i++) {
+        if (va_buffers[i] != VA_INVALID_ID) {
+            va_status = vaDestroyBuffer(va_dpy, va_buffers[i]);
+            CHECK_VASTATUS(va_status,"vaDestroyBuffer");
+            va_buffers[i] = VA_INVALID_ID;
+        }
+    }
+
+    return 0;
+}
+
+static void release_preenc_resource()
+{
+    //preenc_destroy_buffers(&preenc_context.pic_param_buf_id, 1);
+    free(newImageBuffer);
+    if(pMvPredInBuffer)
+        free(pMvPredInBuffer);
+    if(pMbQpInBuffer)
+        free(pMbQpInBuffer);
+
+    // Release all the surfaces resource
+    vaDestroySurfaces(va_dpy, &surface_ids[0], SID_NUMBER);	
+}
+
+static int begin_picture(FILE *yuv_fp, int frame_num, int slice_type)
+{
+    VAStatus va_status;
+
+    VAStatsStatisticsParameter16x16Intel *pic_param = &preenc_context.pic_param;
+
+    preenc_context.current_input_surface = SID_INPUT_PICTURE_0;
+    if(frame_num == 0)
+    {
+        assert(slice_type == SLICE_TYPE_I);
+        return 0;
+    }
+
+    if(pic_param->mv_predictor_ctrl != 0)
+    {
+        // init input MV prediction buffer, FIXME reading from files
+        VAMotionVectorIntel *pMvPred = pMvPredInBuffer;
+        int i = 0;
+        memset(pMvPred, 0, preenc_context.mvpred_buf_size);
+
+        for(i = 0; i<(picture_width_in_mbs*picture_height_in_mbs); i++)
+        {
+            pMvPred->mv0[0] = 1;
+            pMvPred->mv0[1] = 1;
+            pMvPred++;
+        }
+#if 1
+        {
+            FILE* out = NULL;
+
+            out = fopen("mvpindata.bin", "wb");
+            fwrite(pMvPred, 1, preenc_context.mvpred_buf_size, out);
+            fclose(out);
+        }
+#endif
+    }
+
+    if(pic_param->mb_qp != 0)
+    {
+        // init input MB QP buffer, FIXME reading from files
+        VAEncQpBufferH264 *pQpPred = pMbQpInBuffer;
+        memset(pQpPred, 20, preenc_context.mbqp_buf_size);
+#if 1
+        {
+            FILE* out = NULL;
+
+            out = fopen("qpindata.bin", "wb");
+            fwrite(pQpPred, 1, preenc_context.mbqp_buf_size, out);
+            fclose(out);
+        }
+#endif  
+    }
+
+    return 0;
+}
+
+static void preenc_update_picture_parameter(int slice_type, int frame_num)
+{
+    VAStatsStatisticsParameter16x16Intel *pic_param;
+    VAStatus va_status;
+ 
+    // Picture level
+    pic_param        = &preenc_context.pic_param;
+    pic_param->input = surface_ids[SID_INPUT_PICTURE_0];
+    if(slice_type == SLICE_TYPE_I)
+    {
+        pic_param->num_past_references   = 0;
+        pic_param->num_future_references = 0;
+    }
+    else if(slice_type == SLICE_TYPE_P)
+    {
+        pic_param->num_past_references   = 1;
+        pic_param->num_future_references = 0;
+        pic_param->past_references       = &surface_ids[SID_REFERENCE_PICTURE_L0];
+    } 
+    else if(slice_type == SLICE_TYPE_B)
+    {
+        pic_param->num_past_references   = 1;
+        pic_param->num_future_references = 1;
+        pic_param->past_references       = &surface_ids[SID_REFERENCE_PICTURE_L0];
+        pic_param->future_references     = &surface_ids[SID_REFERENCE_PICTURE_L1];
+    } 
+
+
+    if(!pic_param->disable_mv_output || !pic_param->disable_statistics_output)
+        pic_param->outputs      = preenc_context.out_buf_id;
+    else
+        pic_param->outputs = 0;
+
+
+    if(pic_param->mv_predictor_ctrl != 0)
+    {
+        va_status = vaCreateBuffer(va_dpy,
+                                   preenc_context.context_id,
+                                   VAStatsMVPredictorBufferTypeIntel,
+                                   sizeof(VAMotionVectorIntel), picture_width_in_mbs*picture_height_in_mbs, pMvPredInBuffer,
+                                   &preenc_context.mvpred_buf_id);
+        CHECK_VASTATUS(va_status,"vaCreateBuffer");
+        pic_param->mv_predictor = preenc_context.mvpred_buf_id;
+    }
+
+    if(pic_param->mb_qp != 0)
+    {
+        va_status = vaCreateBuffer(va_dpy,
+                                   preenc_context.context_id,
+                                   VAEncQpBufferType,
+                                   sizeof(VAEncQpBufferH264), picture_width_in_mbs*picture_height_in_mbs, pMbQpInBuffer,
+                                   &preenc_context.mbqp_buf_id);
+        CHECK_VASTATUS(va_status,"vaCreateBuffer");
+        pic_param->qp           = preenc_context.mbqp_buf_id;
+    }
+
+    va_status = vaCreateBuffer(va_dpy,
+                               preenc_context.context_id,
+                               VAStatsStatisticsParameterBufferTypeIntel,
+                               sizeof(*pic_param), 1, pic_param,
+                               &preenc_context.pic_param_buf_id);
+    CHECK_VASTATUS(va_status,"vaCreateBuffer");
+
+}
+
+static int preenc_render_picture()
+{
+    VAStatus va_status;
+    VABufferID va_buffers[10];
+    unsigned int num_va_buffers = 0;
+
+
+    va_buffers[num_va_buffers++] = preenc_context.pic_param_buf_id;
+    if(preenc_context.pic_param.mv_predictor_ctrl != 0)
+        va_buffers[num_va_buffers++] = preenc_context.mvpred_buf_id;
+    if(preenc_context.pic_param.mb_qp != 0)
+        va_buffers[num_va_buffers++] = preenc_context.mbqp_buf_id;
+    
+    va_status = vaBeginPicture(va_dpy,
+                               preenc_context.context_id,
+                               surface_ids[preenc_context.current_input_surface]);
+    CHECK_VASTATUS(va_status,"vaBeginPicture");
+
+    va_status = vaRenderPicture(va_dpy,
+                                preenc_context.context_id,
+                                va_buffers,
+                                num_va_buffers);
+    CHECK_VASTATUS(va_status,"vaRenderPicture");
+ 
+    va_status = vaEndPicture(va_dpy, preenc_context.context_id);
+    CHECK_VASTATUS(va_status,"vaEndPicture");
+
+    return 0;
+}
+
+
+static void download_buffer_to_file(FILE *out_fp, VABufferID buf_id, int size)
+{
+    unsigned char   *pbuf;
+    VAStatus        va_status;
+    size_t          n_items;
+
+
+    va_status = vaMapBuffer(va_dpy, buf_id, (void **)(&pbuf));
+    CHECK_VASTATUS(va_status,"vaMapBuffer");
+
+    do {
+        n_items = fwrite(pbuf, size, 1, out_fp);
+        //printf("--> n_items = %d\n", n_items);
+    } 
+    while (n_items != 1);
+
+    va_status = vaUnmapBuffer(va_dpy, buf_id);
+    CHECK_VASTATUS(va_status,"vaUnmapBuffer");
+}
+
+static int store_output_buffer()
+{
+    VAStatus va_status;
+    VASurfaceStatus surface_status;
+    int i = 0;
+
+    va_status = vaSyncSurface(va_dpy, surface_ids[preenc_context.current_input_surface]);
+    CHECK_VASTATUS(va_status,"vaSyncSurface");
+    surface_status = 0;
+    va_status = vaQuerySurfaceStatus(va_dpy, surface_ids[preenc_context.current_input_surface], &surface_status);
+    CHECK_VASTATUS(va_status,"vaQuerySurfaceStatus");
+
+    // get MV output surface data
+    if(!preenc_context.pic_param.disable_mv_output)
+    {
+        download_buffer_to_file(mvout_fp, preenc_context.out_buf_id[i++], preenc_context.mvout_buf_size);
+    }
+
+    // get Statistics output surface data
+    if(!preenc_context.pic_param.disable_statistics_output)
+    {
+        download_buffer_to_file(statsout_fp, preenc_context.out_buf_id[i++], preenc_context.statsout_buf_size);
+    }
+
+    return 0;
+}
+
+static int GetNextInput(int current_input)
+{
+//    if (current_input == SID_REFERENCE_PICTURE_L0)
+//        return SID_INPUT_PICTURE_0;
+//    if (current_input == SID_INPUT_PICTURE_0)
+        return SID_INPUT_PICTURE_0;
+}
+
+static void end_picture(int slice_type, int next_bpic_num)
+{
+    VABufferID tempID;
+    int i = 0;
+
+    /* Prepare for next picture */
+    tempID = surface_ids[SID_INPUT_PICTURE_0];  
+
+    if (slice_type == SLICE_TYPE_I) {
+        surface_ids[SID_INPUT_PICTURE_0] = surface_ids[SID_REFERENCE_PICTURE_L0];
+        surface_ids[SID_REFERENCE_PICTURE_L0] = tempID;	
+    } else
+    if (slice_type != SLICE_TYPE_B) {
+        if (next_bpic_num) {
+            surface_ids[SID_INPUT_PICTURE_0]      = surface_ids[SID_REFERENCE_PICTURE_L1]; 
+            surface_ids[SID_REFERENCE_PICTURE_L1] = tempID;	
+        } else {
+            surface_ids[SID_INPUT_PICTURE_0] = surface_ids[SID_REFERENCE_PICTURE_L0]; 
+            surface_ids[SID_REFERENCE_PICTURE_L0] = tempID;
+        }
+    } else {
+        if (next_bpic_num == 0) {
+            surface_ids[SID_INPUT_PICTURE_0]      = surface_ids[SID_REFERENCE_PICTURE_L0]; 
+            surface_ids[SID_REFERENCE_PICTURE_L0] = surface_ids[SID_REFERENCE_PICTURE_L1];
+            surface_ids[SID_REFERENCE_PICTURE_L1] = tempID;
+        }
+    }
+
+    preenc_destroy_buffers(&preenc_context.pic_param_buf_id, 1);
+    if(preenc_context.pic_param.mv_predictor_ctrl != 0)
+        preenc_destroy_buffers(&preenc_context.mvpred_buf_id, 1);
+    if(preenc_context.pic_param.mb_qp != 0)
+        preenc_destroy_buffers(&preenc_context.mbqp_buf_id, 1);
+    if(!preenc_context.pic_param.disable_mv_output)
+        preenc_destroy_buffers(&preenc_context.out_buf_id[i++], 1);
+    if(!preenc_context.pic_param.disable_statistics_output)
+        preenc_destroy_buffers(&preenc_context.out_buf_id[i++], 1);
+}
+
+static void
+preenc_picture(int frame_num, int slice_type, int next_bpic_num)
+{
+    VAStatus va_status;
+    int ret = 0;
+    int index;
+    
+    printf("---> preenc picture %d, type=%d, next_bpic_num=%d \n", frame_num, slice_type, next_bpic_num);
+
+    pic_param_back = preenc_context.pic_param;
+    if(slice_type == SLICE_TYPE_I)
+    {
+        preenc_context.pic_param.disable_mv_output = 1;
+    }
+    begin_picture(yuv_fp, frame_num, slice_type);
+
+    index = SID_INPUT_PICTURE_0;
+    if(slice_type == SLICE_TYPE_P)
+        fseek(yuv_fp, frame_size * (frame_num + next_bpic_num), SEEK_SET);  // skip B frame to P frame
+    else if(slice_type == SLICE_TYPE_B)
+        fseek(yuv_fp, frame_size * (frame_num - 1), SEEK_SET);              // inserted one P frame before B frames, so minus 1.
+    else 
+        fseek(yuv_fp, frame_size * frame_num, SEEK_SET);
+    preenc_context.upload_thread_param.yuv_fp     = yuv_fp;
+    preenc_context.upload_thread_param.surface_id = surface_ids[index];
+    upload_thread_function((void *)(&preenc_context.upload_thread_param));
+    usleep(10000);
+
+    do {
+        int i = 0;
+        preenc_destroy_buffers(&preenc_context.pic_param_buf_id, 1);
+        if(preenc_context.pic_param.mv_predictor_ctrl != 0)
+            preenc_destroy_buffers(&preenc_context.mvpred_buf_id, 1);
+        if(preenc_context.pic_param.mb_qp != 0)
+            preenc_destroy_buffers(&preenc_context.mbqp_buf_id, 1);
+
+        if(!preenc_context.pic_param.disable_mv_output)
+        {
+            preenc_destroy_buffers(&preenc_context.out_buf_id[i], 1);
+            va_status = vaCreateBuffer(va_dpy,
+                                       preenc_context.context_id,
+                                       VAStatsMotionVectorBufferTypeIntel,
+                                       sizeof(VAMotionVectorIntel), picture_width_in_mbs*picture_height_in_mbs*16, NULL,
+                                       &preenc_context.out_buf_id[i++]);
+            CHECK_VASTATUS(va_status,"vaCreateBuffer");
+        }
+        if(!preenc_context.pic_param.disable_statistics_output)
+        {
+            preenc_destroy_buffers(&preenc_context.out_buf_id[i], 1);
+            va_status = vaCreateBuffer(va_dpy,
+                                       preenc_context.context_id,
+                                       VAStatsStatisticsBufferTypeIntel,
+                                       sizeof(VAStatsStatistics16x16Intel), picture_width_in_mbs*picture_height_in_mbs, NULL,
+                                       &preenc_context.out_buf_id[i++]);
+            CHECK_VASTATUS(va_status,"vaCreateBuffer");
+        }
+
+        /* picture parameter set */
+        preenc_update_picture_parameter(slice_type, frame_num);
+
+        preenc_render_picture();
+        usleep(10000);
+        ret = store_output_buffer();
+    } while (ret);
+
+    end_picture(slice_type, next_bpic_num);
+    preenc_context.pic_param = pic_param_back;
+}
+
+int main(int argc, char *argv[])
+{
+    unsigned char inputyuv[256];
+    int i = 0;
+    int f;
+    long file_size;
+    int mode_value;
+    int enc_frames = 0;
+    int bdistance  = 1; //Distance between I- or P- key frames (1 means no B-frames) (0 - by default(I frames) not allowed in this test)
+    int gop_length = -1;
+    struct timeval tpstart,tpend; 
+    float  timeuse;
+
+    picture_width  = 352;
+    picture_height = 288;
+    for (i=1; i<argc; i++)
+    {
+        //printf("---> i=%d, argv[i]=%s \n", i, argv[i]);
+        if(strcmp(argv[i],"-size")==0)
+        {
+            if(i+2<argc)
+            {
+                picture_width  = atoi(argv[i+1]);
+                picture_height = atoi(argv[i+2]);
+                i += 2;
+            } else
+            {
+                show_help();
+                return -1;
+            }
+        } else
+        if(strcmp(argv[i],"-i")==0)
+        {
+            if(i+1<argc) 
+            {
+                //strcat(inputyuv, argv[i+1]);
+                strcpy(inputyuv, argv[i+1]);
+                //printf("---> argv[i+1]=%s, inputyuv=%s \n", argv[i+1], inputyuv);
+                i++;
+            } else
+            {
+                show_help();
+                return -1;
+            }
+        } else
+        if(strcmp(argv[i],"-frames")==0)
+        {
+            if(i+1<argc)
+            {
+                enc_frames = atoi(argv[i+1]);
+                i++;
+            } else
+            {
+                show_help();
+                return -1;
+            }
+        } else
+        if(strcmp(argv[i],"-bdistance")==0)
+        {
+            if(i+1<argc)
+            {
+                bdistance = atoi(argv[i+1]);
+                i++;
+            } else
+            {
+                show_help();
+                return -1;
+            }
+        } else
+        if(strcmp(argv[i],"-gop_length")==0)
+        {
+            if(i+1<argc)
+            {
+                gop_length = atoi(argv[i+1]);
+                i++;
+            } else
+            {
+                show_help();
+                return -1;
+            }
+        } else
+        {
+            show_help();
+            return -1;
+        }
+    }
+
+    picture_width_in_mbs  = (picture_width  + 15) / 16;
+    picture_height_in_mbs = (picture_height + 15) / 16;
+
+    yuv_fp = fopen(inputyuv,"rb");
+    if ( yuv_fp == NULL){
+        printf("Can't open input YUV file\n");
+        show_help();
+        return -1;
+    }
+    fseek(yuv_fp,0l, SEEK_END);
+    file_size = ftell(yuv_fp);
+    frame_size = picture_width * picture_height +  ((picture_width * picture_height) >> 1) ;
+
+    if(enc_frames == 1 )
+        file_size = 2 * frame_size;          //FIXME default 0 run all frames, 1 only test first P frame first, first is I frame
+    else if(enc_frames > 1)
+        file_size = enc_frames * frame_size; //FIXME only test (enc_frames - 1) P frames
+
+    if ( (file_size < frame_size) || (file_size % frame_size) ) {
+        fclose(yuv_fp);
+        printf("The YUV file's size is not correct\n");
+        show_help();
+        return -1;
+    }
+
+    frame_number = file_size / frame_size;
+    fseek(yuv_fp, 0l, SEEK_SET);
+
+    if(bdistance < 1)
+        bdistance = 1;
+    if(gop_length < 0)
+        gop_length = frame_number;
+ 
+    if(preenc_context.pic_param.disable_mv_output)
+    {
+        mvout_fp = 0;
+    } else 
+    {
+        mvout_fp = fopen("mvoutdata.bin", "wb");	
+        if ( mvout_fp == NULL) {
+            fclose(mvout_fp);
+            printf("Can't open output MV data file\n");
+            show_help();
+            return -1;
+        }
+    }
+    if(preenc_context.pic_param.disable_statistics_output)
+    {
+        statsout_fp = 0;
+    } else 
+    {
+        statsout_fp = fopen("statsoutdata.bin", "wb");	
+        if ( statsout_fp == NULL) {
+            fclose(statsout_fp);
+            printf("Can't open output Statistics data file\n");
+            show_help();
+            return -1;
+        }
+    }
+
+    gettimeofday(&tpstart,NULL);
+    preenc_context_init();
+    create_preenc_pipe();
+    alloc_preenc_resource();
+
+    for ( f = 0; f < frame_number; f++) {		//picture level loop, start from second picture
+        if(bdistance==1)
+        {
+            if((f % gop_length) == 0)
+                preenc_picture(f, SLICE_TYPE_I, 0);
+            else
+                preenc_picture(f, SLICE_TYPE_P, 0);
+        } else 
+        {
+            if((f % gop_length) == 0)
+                preenc_picture(f, SLICE_TYPE_I, 0);
+            else if (((f-1)%bdistance) == 0)
+                preenc_picture(f, SLICE_TYPE_P, MIN((bdistance-1), (gop_length-1-(f%gop_length))));
+            else
+                preenc_picture(f, SLICE_TYPE_B, MIN((((f+bdistance-1)/bdistance)*bdistance-f), (gop_length-1-(f%gop_length))));
+
+        }
+    }
+
+    gettimeofday(&tpend,NULL);
+    timeuse=1000000*(tpend.tv_sec-tpstart.tv_sec)+ tpend.tv_usec-tpstart.tv_usec;
+    timeuse/=1000000;
+    printf("\ndone!\n");
+    printf("encode %d frames in %f secondes, FPS is %.1f\n",frame_number, timeuse, frame_number/timeuse);
+    release_preenc_resource();
+    destory_preenc_pipe();
+
+    if(yuv_fp)
+        fclose(yuv_fp);
+    if(mvpredin_fp)
+        fclose(mvpredin_fp);
+    if(qpin_fp)
+        fclose(qpin_fp);
+    if(mvout_fp)
+        fclose(mvout_fp);
+    if(statsout_fp)
+        fclose(statsout_fp);
+
+    return 0;
+}
+
-- 
1.7.9.5


Hi,
We add two new FEI samples into test/encode/ folder for test sample fei/statistics cases, feiavcenc.c is for fei case and preenc.c is for statistics case.


Thanks,
Kelvin




More information about the Libva mailing list