static void put_updated_rectangle(rfbClient *client, int x, int y, int width, int height, int f_width, int f_height, int first_for_frame)
{
if (curr_surface == VA_INVALID_ID) {
rfbClientErr("%s: called, but current surface is invalid\n", __FUNCTION__);
return;
}
VAStatus va_status;
if (client->outputWindow) {
/* use efficient vaPutSurface() method of putting the framebuffer on the screen */
if (first_for_frame) {
/* vaPutSurface() clears window contents outside the given destination rectangle => always update full screen. */
va_status = vaPutSurface(va_dpy, curr_surface, client->outputWindow, 0, 0, f_width, f_height, 0, 0, f_width, f_height, NULL, 0, VA_FRAME_PICTURE);
CHECK_VASTATUS(va_status, "vaPutSurface");
}
}
else if (client->frameBuffer) {
/* ... or copy the changed framebuffer region manually as a fallback */
VAImage decoded_image;
decoded_image.image_id = VA_INVALID_ID;
decoded_image.buf = VA_INVALID_ID;
va_status = vaDeriveImage(va_dpy, curr_surface, &decoded_image);
CHECK_VASTATUS(va_status, "vaDeriveImage");
if ((decoded_image.image_id == VA_INVALID_ID) || (decoded_image.buf == VA_INVALID_ID)) {
rfbClientErr("%s: vaDeriveImage() returned success but VA image is invalid (id: %d, buf: %d)\n", __FUNCTION__, decoded_image.image_id, decoded_image.buf);
}
nv12_to_rgba(decoded_image, client, x, y, width, height);
va_status = vaDestroyImage(va_dpy, decoded_image.image_id);
CHECK_VASTATUS(va_status, "vaDestroyImage");
}
}
int csc_preparation ()
{
VAStatus va_status;
// 1. make sure dst fourcc is supported for vaImage
if (!lookup_image_format(csc_dst_fourcc)) {
test_color_conversion = 0;
printf("VA driver doesn't support %s image, skip additional color conversion\n", map_vafourcc_to_str(csc_dst_fourcc));
goto cleanup;
}
// 2. make sure src_fourcc is supported for vaSurface
VASurfaceAttrib surface_attribs[1], * const s_attrib = &surface_attribs[0];
s_attrib->type = VASurfaceAttribPixelFormat;
s_attrib->flags = VA_SURFACE_ATTRIB_SETTABLE;
s_attrib->value.type = VAGenericValueTypeInteger;
s_attrib->value.value.i = csc_src_fourcc;
if (!lookup_surface_attrib(VASurfaceAttribPixelFormat, &s_attrib->value)) {
printf("VA driver doesn't support %s surface, skip additional color conversion\n", map_vafourcc_to_str(csc_src_fourcc));
test_color_conversion = 0;
goto cleanup;
}
// 3 create all objs required by csc
// 3.1 vaSurface with src fourcc
va_status = vaCreateSurfaces(
va_dpy,
VA_RT_FORMAT_YUV420, surface_width, surface_height,
&surface_id[0], SURFACE_NUM,
surface_attribs, 1
);
CHECK_VASTATUS(va_status,"vaCreateSurfaces");
// 3.2 vaImage with dst fourcc
VAImageFormat image_format;
image_format.fourcc = csc_dst_fourcc;
image_format.byte_order = VA_LSB_FIRST;
image_format.bits_per_pixel = 16;
va_status = vaCreateImage(va_dpy, &image_format,
surface_width, surface_height,
&csc_dst_fourcc_image);
CHECK_VASTATUS(va_status,"vaCreateImage");
// 3.3 create a temp VASurface for final rendering(vaPutSurface)
s_attrib->value.value.i = VA_FOURCC_NV12;
va_status = vaCreateSurfaces(va_dpy, VA_RT_FORMAT_YUV420,
surface_width, surface_height,
&csc_render_surface, 1,
surface_attribs, 1);
CHECK_VASTATUS(va_status,"vaCreateSurfaces");
cleanup:
return test_color_conversion;
}
static float run(const char* infile, const char* outfile1, const char* outfile2, bool doInterop)
{
VASurfaceID surface;
VAStatus status;
Timer t;
// initialize CL context for CL/VA interop
cv::va_intel::ocl::initializeContextFromVA(va::display, doInterop);
// load input image
cv::UMat u1 = readImage(infile);
cv::Size size2 = u1.size();
status = vaCreateSurfaces(va::display, VA_RT_FORMAT_YUV420, size2.width, size2.height, &surface, 1, NULL, 0);
CHECK_VASTATUS(status, "vaCreateSurfaces");
// transfer image into VA surface, make sure all CL initialization is done (kernels etc)
cv::va_intel::convertToVASurface(va::display, u1, surface, size2);
cv::va_intel::convertFromVASurface(va::display, surface, size2, u1);
cv::UMat u2;
cv::blur(u1, u2, cv::Size(7, 7), cv::Point(-3, -3));
// measure performance on some image processing
writeImage(u1, outfile1, doInterop);
t.start();
cv::va_intel::convertFromVASurface(va::display, surface, size2, u1);
cv::blur(u1, u2, cv::Size(7, 7), cv::Point(-3, -3));
cv::va_intel::convertToVASurface(va::display, u2, surface, size2);
t.stop();
writeImage(u2, outfile2, doInterop);
vaDestroySurfaces(va::display, &surface,1);
return t.time(Timer::MSEC);
}
static int
ensure_image_formats(void)
{
VAStatus va_status;
VAImageFormat *image_formats;
int num_image_formats;
if (va_num_image_formats >= 0)
return va_num_image_formats;
num_image_formats = vaMaxNumImageFormats(va_dpy);
if (num_image_formats == 0)
return 0;
image_formats = malloc(num_image_formats * sizeof(*image_formats));
if (!image_formats)
return 0;
va_status = vaQueryImageFormats(va_dpy, image_formats, &num_image_formats);
CHECK_VASTATUS(va_status, "vaQuerySurfaceAttributes()");
va_image_formats = image_formats;
va_num_image_formats = num_image_formats;
return num_image_formats;
}
static int
ensure_surface_attribs(void)
{
VAStatus va_status;
VASurfaceAttrib *surface_attribs;
unsigned int num_image_formats, num_surface_attribs;
if (va_num_surface_attribs >= 0)
return va_num_surface_attribs;
num_image_formats = vaMaxNumImageFormats(va_dpy);
if (num_image_formats == 0)
return 0;
va_status = vaCreateConfig(va_dpy, VAProfileNone, VAEntrypointVideoProc,
NULL, 0, &vpp_config_id);
CHECK_VASTATUS(va_status, "vaCreateConfig()");
/* Guess the number of surface attributes, thus including any
pixel-format supported by the VA driver */
num_surface_attribs = VASurfaceAttribCount + num_image_formats;
surface_attribs = malloc(num_surface_attribs * sizeof(*surface_attribs));
if (!surface_attribs)
return 0;
va_status = vaQuerySurfaceAttributes(va_dpy, vpp_config_id,
surface_attribs, &num_surface_attribs);
if (va_status == VA_STATUS_SUCCESS)
va_surface_attribs = surface_attribs;
else if (va_status == VA_STATUS_ERROR_MAX_NUM_EXCEEDED) {
va_surface_attribs = realloc(surface_attribs,
num_surface_attribs * sizeof(*va_surface_attribs));
if (!va_surface_attribs) {
free(surface_attribs);
return 0;
}
va_status = vaQuerySurfaceAttributes(va_dpy, vpp_config_id,
va_surface_attribs, &num_surface_attribs);
}
CHECK_VASTATUS(va_status, "vaQuerySurfaceAttributes()");
va_num_surface_attribs = num_surface_attribs;
return num_surface_attribs;
}
static void h264_cleanup_decoder()
{
VAStatus va_status;
rfbClientLog("%s()\n", __FUNCTION__);
if (va_surface_id[0] != VA_INVALID_ID) {
va_status = vaDestroySurfaces(va_dpy, &va_surface_id[0], SURFACE_NUM);
CHECK_VASTATUS(va_status, "vaDestroySurfaces");
}
if (va_context_id) {
va_status = vaDestroyContext(va_dpy, va_context_id);
CHECK_VASTATUS(va_status, "vaDestroyContext");
va_context_id = 0;
}
num_frames = 0;
sid = 0;
frame_id = 0;
field_order_count = 0;
}
static void checkIfAvailableYUV420()
{
VAEntrypoint entrypoints[5];
int num_entrypoints,vld_entrypoint;
VAConfigAttrib attrib;
VAStatus status;
status = vaQueryConfigEntrypoints(va::display, VAProfileMPEG2Main, entrypoints, &num_entrypoints);
CHECK_VASTATUS(status, "vaQueryConfigEntrypoints");
for (vld_entrypoint = 0; vld_entrypoint < num_entrypoints; ++vld_entrypoint)
{
if (entrypoints[vld_entrypoint] == VAEntrypointVLD)
break;
}
if (vld_entrypoint == num_entrypoints)
throw std::runtime_error("Failed to find VLD entry point");
attrib.type = VAConfigAttribRTFormat;
vaGetConfigAttributes(va::display, VAProfileMPEG2Main, VAEntrypointVLD, &attrib, 1);
if ((attrib.value & VA_RT_FORMAT_YUV420) == 0)
throw std::runtime_error("Desired YUV420 RT format not found");
}
int main(int argc,char **argv)
{
int major_ver, minor_ver;
VAStatus va_status;
pthread_t thread1;
int ret;
char c;
int i;
char str_src_fmt[5], str_dst_fmt[5];
static struct option long_options[] =
{
{"fmt1", required_argument, NULL, '1'},
{"fmt2", required_argument, NULL, '2'},
{0, 0, 0, 0}
};
while ((c =getopt_long(argc,argv,"w:h:g:r:d:f:tcep?n:1:2:v", long_options, NULL)) != EOF) {
switch (c) {
case '?':
printf("putsurface <options>\n");
printf(" -g <widthxheight+x_location+y_location> window geometry\n");
printf(" -w/-h resolution of surface\n");
printf(" -r <framerate>\n");
printf(" -d the dimension of black/write square box, default is 32\n");
printf(" -t multi-threads\n");
printf(" -c test clipbox\n");
printf(" -f <1/2> top field, or bottom field\n");
printf(" -1 source format (fourcc) for color conversion test\n");
printf(" -2 dest format (fourcc) for color conversion test\n");
printf(" --fmt1 same to -1\n");
printf(" --fmt2 same to -2\n");
printf(" -v verbose output\n");
exit(0);
break;
case 'g':
ret = sscanf(optarg, "%dx%d+%d+%d", &win_width, &win_height, &win_x, &win_y);
if (ret != 4) {
printf("invalid window geometry, must be widthxheight+x_location+y_location\n");
exit(0);
} else
printf("Create window at (%d, %d), width = %d, height = %d\n",
win_x, win_y, win_width, win_height);
break;
case 'r':
frame_rate = atoi(optarg);
break;
case 'w':
surface_width = atoi(optarg);
break;
case 'h':
surface_height = atoi(optarg);
break;
case 'n':
frame_num_total = atoi(optarg);
break;
case 'd':
box_width = atoi(optarg);
break;
case 't':
multi_thread = 1;
printf("Two threads to do vaPutSurface\n");
break;
case 'e':
check_event = 0;
break;
case 'p':
put_pixmap = 1;
break;
case 'c':
test_clip = 1;
break;
case 'f':
if (atoi(optarg) == 1) {
printf("Display TOP field\n");
display_field = VA_TOP_FIELD;
} else if (atoi(optarg) == 2) {
printf("Display BOTTOM field\n");
display_field = VA_BOTTOM_FIELD;
} else
printf("The validate input for -f is: 1(top field)/2(bottom field)\n");
break;
case '1':
sscanf(optarg, "%s", str_src_fmt);
csc_src_fourcc = map_str_to_vafourcc (str_src_fmt);
if (!csc_src_fourcc) {
printf("invalid fmt1: %s\n", str_src_fmt );
exit(0);
}
break;
case '2':
sscanf(optarg, "%s", str_dst_fmt);
csc_dst_fourcc = map_str_to_vafourcc (str_dst_fmt);
if (!csc_dst_fourcc) {
printf("invalid fmt1: %s\n", str_dst_fmt );
exit(0);
}
break;
case 'v':
verbose = 1;
printf("Enable verbose output\n");
break;
}
}
if (csc_src_fourcc && csc_dst_fourcc) {
test_color_conversion = 1;
}
win_display = (void *)open_display();
if (win_display == NULL) {
fprintf(stderr, "Can't open the connection of display!\n");
exit(-1);
}
create_window(win_display, win_x, win_y, win_width, win_height);
va_dpy = vaGetDisplay(win_display);
va_status = vaInitialize(va_dpy, &major_ver, &minor_ver);
CHECK_VASTATUS(va_status, "vaInitialize");
if (test_color_conversion) {
ret = csc_preparation();
}
if (!test_color_conversion || !ret ) {
va_status = vaCreateSurfaces(
va_dpy,
VA_RT_FORMAT_YUV420, surface_width, surface_height,
&surface_id[0], SURFACE_NUM,
NULL, 0
);
}
CHECK_VASTATUS(va_status, "vaCreateSurfaces");
if (multi_thread == 0) /* upload the content for all surfaces */
upload_source_YUV_once_for_all();
if (check_event)
pthread_mutex_init(&gmutex, NULL);
for(i = 0; i< SURFACE_NUM; i++)
pthread_mutex_init(&surface_mutex[i], NULL);
if (multi_thread == 1)
ret = pthread_create(&thread1, NULL, putsurface_thread, (void*)drawable_thread1);
putsurface_thread((void *)drawable_thread0);
if (multi_thread == 1)
pthread_join(thread1, (void **)&ret);
printf("thread1 is free\n");
if (test_color_conversion) {
// destroy temp surface/image
va_status = vaDestroySurfaces(va_dpy, &csc_render_surface, 1);
CHECK_VASTATUS(va_status,"vaDestroySurfaces");
va_status = vaDestroyImage(va_dpy, csc_dst_fourcc_image.image_id);
CHECK_VASTATUS(va_status,"vaDestroyImage");
}
if (vpp_config_id != VA_INVALID_ID) {
vaDestroyConfig (va_dpy, vpp_config_id);
vpp_config_id = VA_INVALID_ID;
}
vaDestroySurfaces(va_dpy,&surface_id[0],SURFACE_NUM);
vaTerminate(va_dpy);
free(va_image_formats);
free(va_surface_attribs);
close_display(win_display);
return 0;
}
static void* putsurface_thread(void *data)
{
int width=win_width, height=win_height;
void *drawable = data;
int quit = 0;
VAStatus vaStatus;
int row_shift = 0;
int index = 0;
unsigned int frame_num=0, start_time, putsurface_time;
VARectangle cliprects[2]; /* client supplied clip list */
int continue_display = 0;
if (drawable == drawable_thread0)
printf("Enter into thread0\n\n");
if (drawable == drawable_thread1)
printf("Enter into thread1\n\n");
putsurface_time = 0;
while (!quit) {
VASurfaceID surface_id = VA_INVALID_SURFACE;
while (surface_id == VA_INVALID_SURFACE)
surface_id = get_next_free_surface(&index);
if (verbose) printf("Thread: %p Display surface 0x%x,\n", drawable, surface_id);
if (multi_thread)
upload_surface(va_dpy, surface_id, box_width, row_shift, display_field);
if (check_event)
pthread_mutex_lock(&gmutex);
start_time = get_tick_count();
if ((continue_display == 0) && getenv("FRAME_STOP")) {
char c;
printf("Press any key to display frame %d...(c/C to continue)\n", frame_num);
c = getchar();
if (c == 'c' || c == 'C')
continue_display = 1;
}
if (test_color_conversion) {
static int _put_surface_count = 0;
if (_put_surface_count++ %50 == 0) {
printf("do additional colorcoversion from %s to %s\n", map_vafourcc_to_str(csc_src_fourcc), map_vafourcc_to_str(csc_dst_fourcc));
}
// get image from surface, csc_src_fourcc to csc_dst_fourcc conversion happens
vaStatus = vaGetImage(va_dpy, surface_id, 0, 0,
surface_width, surface_height, csc_dst_fourcc_image.image_id);
CHECK_VASTATUS(vaStatus,"vaGetImage");
// render csc_dst_fourcc image to temp surface
vaStatus = vaPutImage(va_dpy, csc_render_surface, csc_dst_fourcc_image.image_id,
0, 0, surface_width, surface_height,
0, 0, surface_width, surface_height);
CHECK_VASTATUS(vaStatus,"vaPutImage");
// render the temp surface, it should be same with original surface without color conversion test
vaStatus = vaPutSurface(va_dpy, csc_render_surface, CAST_DRAWABLE(drawable),
0,0,surface_width,surface_height,
0,0,width,height,
(test_clip==0)?NULL:&cliprects[0],
(test_clip==0)?0:2,
display_field);
CHECK_VASTATUS(vaStatus,"vaPutSurface");
}
else {
vaStatus = vaPutSurface(va_dpy, surface_id, CAST_DRAWABLE(drawable),
0,0,surface_width,surface_height,
0,0,width,height,
(test_clip==0)?NULL:&cliprects[0],
(test_clip==0)?0:2,
display_field);
CHECK_VASTATUS(vaStatus,"vaPutSurface");
}
putsurface_time += (get_tick_count() - start_time);
if (check_event)
pthread_mutex_unlock(&gmutex);
pthread_mutex_unlock(&surface_mutex[index]); /* locked in get_next_free_surface */
if ((frame_num % 0xff) == 0) {
fprintf(stderr, "%.2f FPS \r", 256000.0 / (float)putsurface_time);
putsurface_time = 0;
update_clipbox(cliprects, width, height);
}
if (check_event)
check_window_event(win_display, drawable, &width, &height, &quit);
if (multi_thread) { /* reload surface content */
row_shift++;
if (row_shift==(2*box_width)) row_shift= 0;
}
if (frame_rate != 0) /* rough framerate control */
usleep(1000/frame_rate*1000);
frame_num++;
if (frame_num >= frame_num_total)
quit = 1;
}
if (drawable == drawable_thread1)
pthread_exit(NULL);
return 0;
}
static void nv12_to_rgba(const VAImage vaImage, rfbClient *client, int ch_x, int ch_y, int ch_w, int ch_h)
{
DebugLog(("%s: converting region (%d, %d)-(%d, %d) from NV12->RGBA\n", __FUNCTION__, ch_x, ch_y, ch_w, ch_h));
VAStatus va_status;
uint8_t *nv12_buf;
va_status = vaMapBuffer(va_dpy, vaImage.buf, (void **)&nv12_buf);
CHECK_VASTATUS(va_status, "vaMapBuffer(DecodedData)");
/* adjust x, y, width, height of the affected area so
* x, y, width and height are always even.
*/
if (ch_x % 2) { --ch_x; ++ch_w; }
if (ch_y % 2) { --ch_y; ++ch_h; }
if ((ch_x + ch_w) % 2) { ++ch_w; }
if ((ch_y + ch_h) % 2) { ++ch_h; }
/* point nv12_buf and dst to upper left corner of changed area */
uint8_t *nv12_y = &nv12_buf[vaImage.offsets[0] + vaImage.pitches[0] * ch_y + ch_x];
uint8_t *nv12_uv = &nv12_buf[vaImage.offsets[1] + vaImage.pitches[1] * (ch_y / 2) + ch_x];
uint32_t *dst = &((uint32_t*)client->frameBuffer)[client->width * ch_y + ch_x];
/* TODO: optimize R, G, B calculation. Possible ways to do this:
* - use lookup tables
* - convert from floating point to integer arithmetic
* - use MMX/SSE to vectorize calculations
* - use GPU (VA VPP, shader...)
*/
int src_x, src_y;
for (src_y = 0; src_y < ch_h; src_y += 2) {
for (src_x = 0; src_x < ch_w; src_x += 2) {
uint8_t nv_u = nv12_uv[src_x];
uint8_t nv_v = nv12_uv[src_x + 1];
uint8_t nv_y[4] = { nv12_y[ src_x], nv12_y[ src_x + 1],
nv12_y[vaImage.pitches[0] + src_x], nv12_y[vaImage.pitches[0] + src_x + 1] };
int i;
for (i = 0; i < 4; ++i) {
double R = 1.164 * (nv_y[i] - 16) + 1.596 * (nv_v - 128);
double G = 1.164 * (nv_y[i] - 16) - 0.391 * (nv_u - 128) - 0.813 * (nv_v - 128);
double B = 1.164 * (nv_y[i] - 16) + 2.018 * (nv_u - 128);
/* clamp R, G, B values. For some Y, U, V combinations,
* the results of the above calculations fall outside of
* the range 0-255.
*/
if (R < 0.0) R = 0.0;
if (G < 0.0) G = 0.0;
if (B < 0.0) B = 0.0;
if (R > 255.0) R = 255.0;
if (G > 255.0) G = 255.0;
if (B > 255.0) B = 255.0;
dst[client->width * (i / 2) + src_x + (i % 2)] = 0
| ((unsigned int)(R + 0.5) << client->format.redShift)
| ((unsigned int)(G + 0.5) << client->format.greenShift)
| ((unsigned int)(B + 0.5) << client->format.blueShift);
}
}
nv12_y += 2 * vaImage.pitches[0];
nv12_uv += vaImage.pitches[1];
dst += 2 * client->width;
}
CHECK_SURF(va_surface_id[sid]);
va_status = vaUnmapBuffer(va_dpy, vaImage.buf);
CHECK_VASTATUS(va_status, "vaUnmapBuffer(DecodedData)");
}
static void h264_decode_frame(int f_width, int f_height, char *framedata, int framesize, int slice_type)
{
VAStatus va_status;
DebugLog(("%s: called for frame of %d bytes (%dx%d) slice_type=%d\n", __FUNCTION__, framesize, width, height, slice_type));
/* Initialize decode pipeline if necessary */
if ( (f_width > cur_width) || (f_height > cur_height) ) {
if (va_dpy != NULL)
h264_cleanup_decoder();
cur_width = f_width;
cur_height = f_height;
h264_init_decoder(f_width, f_height);
rfbClientLog("%s: decoder initialized\n", __FUNCTION__);
}
/* Decode frame */
static VAPictureH264 va_picture_h264, va_old_picture_h264;
/* The server should always send an I-frame when a new client connects
* or when the resolution of the framebuffer changes, but we check
* just in case.
*/
if ( (slice_type != SLICE_TYPE_I) && (num_frames == 0) ) {
rfbClientLog("First frame is not an I frame !!! Skipping!!!\n");
return;
}
DebugLog(("%s: frame_id=%d va_surface_id[%d]=0x%x field_order_count=%d\n", __FUNCTION__, frame_id, sid, va_surface_id[sid], field_order_count));
va_picture_h264.picture_id = va_surface_id[sid];
va_picture_h264.frame_idx = frame_id;
va_picture_h264.flags = 0;
va_picture_h264.BottomFieldOrderCnt = field_order_count;
va_picture_h264.TopFieldOrderCnt = field_order_count;
/* Set up picture parameter buffer */
if (va_pic_param_buf_id[sid] == VA_INVALID_ID) {
va_status = vaCreateBuffer(va_dpy, va_context_id, VAPictureParameterBufferType, sizeof(VAPictureParameterBufferH264), 1, NULL, &va_pic_param_buf_id[sid]);
CHECK_VASTATUS(va_status, "vaCreateBuffer(PicParam)");
}
CHECK_SURF(va_surface_id[sid]);
VAPictureParameterBufferH264 *pic_param_buf = NULL;
va_status = vaMapBuffer(va_dpy, va_pic_param_buf_id[sid], (void **)&pic_param_buf);
CHECK_VASTATUS(va_status, "vaMapBuffer(PicParam)");
SetVAPictureParameterBufferH264(pic_param_buf, f_width, f_height);
memcpy(&pic_param_buf->CurrPic, &va_picture_h264, sizeof(VAPictureH264));
if (slice_type == SLICE_TYPE_P) {
memcpy(&pic_param_buf->ReferenceFrames[0], &va_old_picture_h264, sizeof(VAPictureH264));
pic_param_buf->ReferenceFrames[0].flags = 0;
}
else if (slice_type != SLICE_TYPE_I) {
rfbClientLog("Frame type %d not supported!!!\n");
return;
}
pic_param_buf->frame_num = frame_id;
va_status = vaUnmapBuffer(va_dpy, va_pic_param_buf_id[sid]);
CHECK_VASTATUS(va_status, "vaUnmapBuffer(PicParam)");
/* Set up IQ matrix buffer */
if (va_mat_param_buf_id[sid] == VA_INVALID_ID) {
va_status = vaCreateBuffer(va_dpy, va_context_id, VAIQMatrixBufferType, sizeof(VAIQMatrixBufferH264), 1, NULL, &va_mat_param_buf_id[sid]);
CHECK_VASTATUS(va_status, "vaCreateBuffer(IQMatrix)");
}
CHECK_SURF(va_surface_id[sid]);
VAIQMatrixBufferH264 *iq_matrix_buf = NULL;
va_status = vaMapBuffer(va_dpy, va_mat_param_buf_id[sid], (void **)&iq_matrix_buf);
CHECK_VASTATUS(va_status, "vaMapBuffer(IQMatrix)");
static const unsigned char m_MatrixBufferH264[]= {
/* ScalingList4x4[6][16] */
0x10,0x10,0x10,0x10,0x10,0x10,0x10,0x10,0x10,0x10,0x10,0x10,0x10,0x10,0x10,0x10,
0x10,0x10,0x10,0x10,0x10,0x10,0x10,0x10,0x10,0x10,0x10,0x10,0x10,0x10,0x10,0x10,
0x10,0x10,0x10,0x10,0x10,0x10,0x10,0x10,0x10,0x10,0x10,0x10,0x10,0x10,0x10,0x10,
0x10,0x10,0x10,0x10,0x10,0x10,0x10,0x10,0x10,0x10,0x10,0x10,0x10,0x10,0x10,0x10,
0x10,0x10,0x10,0x10,0x10,0x10,0x10,0x10,0x10,0x10,0x10,0x10,0x10,0x10,0x10,0x10,
0x10,0x10,0x10,0x10,0x10,0x10,0x10,0x10,0x10,0x10,0x10,0x10,0x10,0x10,0x10,0x10,
/* ScalingList8x8[2][64] */
0x00,0x00,0x00,0x00,0x00,0x00,0x00,0x00,
0x00,0x00,0x00,0x00,0x00,0x00,0x00,0x00,
0x00,0x00,0x00,0x00,0x00,0x00,0x00,0x00,
0x00,0x00,0x00,0x00,0x00,0x00,0x00,0x00,
0x00,0x00,0x00,0x00,0x00,0x00,0x00,0x00,
0x00,0x00,0x00,0x00,0x00,0x00,0x00,0x00,
0x00,0x00,0x00,0x00,0x00,0x00,0x00,0x00,
0x00,0x00,0x00,0x00,0x00,0x00,0x00,0x00,
0x00,0x00,0x00,0x00,0x00,0x00,0x00,0x00,
0x00,0x00,0x00,0x00,0x00,0x00,0x00,0x00,
0x00,0x00,0x00,0x00,0x00,0x00,0x00,0x00,
0x00,0x00,0x00,0x00,0x00,0x00,0x00,0x00,
0x00,0x00,0x00,0x00,0x00,0x00,0x00,0x00,
0x00,0x00,0x00,0x00,0x00,0x00,0x00,0x00,
0x00,0x00,0x00,0x00,0x00,0x00,0x00,0x00,
0x00,0x00,0x00,0x00,0x00,0x00,0x00,0x00
};
memcpy(iq_matrix_buf, m_MatrixBufferH264, 224);
va_status = vaUnmapBuffer(va_dpy, va_mat_param_buf_id[sid]);
CHECK_VASTATUS(va_status, "vaUnmapBuffer(IQMatrix)");
VABufferID buffer_ids[2];
buffer_ids[0] = va_pic_param_buf_id[sid];
buffer_ids[1] = va_mat_param_buf_id[sid];
CHECK_SURF(va_surface_id[sid]);
va_status = vaRenderPicture(va_dpy, va_context_id, buffer_ids, 2);
CHECK_VASTATUS(va_status, "vaRenderPicture");
/* Set up slice parameter buffer */
if (va_sp_param_buf_id[sid] == VA_INVALID_ID) {
va_status = vaCreateBuffer(va_dpy, va_context_id, VASliceParameterBufferType, sizeof(VASliceParameterBufferH264), 1, NULL, &va_sp_param_buf_id[sid]);
CHECK_VASTATUS(va_status, "vaCreateBuffer(SliceParam)");
}
CHECK_SURF(va_surface_id[sid]);
VASliceParameterBufferH264 *slice_param_buf = NULL;
va_status = vaMapBuffer(va_dpy, va_sp_param_buf_id[sid], (void **)&slice_param_buf);
CHECK_VASTATUS(va_status, "vaMapBuffer(SliceParam)");
static int t2_first = 1;
if (slice_type == SLICE_TYPE_I) {
SetVASliceParameterBufferH264_Intra(slice_param_buf, t2_first);
t2_first = 0;
} else {
SetVASliceParameterBufferH264(slice_param_buf);
memcpy(&slice_param_buf->RefPicList0[0], &va_old_picture_h264, sizeof(VAPictureH264));
slice_param_buf->RefPicList0[0].flags = 0;
}
slice_param_buf->slice_data_bit_offset = 0;
slice_param_buf->slice_data_size = framesize;
va_status = vaUnmapBuffer(va_dpy, va_sp_param_buf_id[sid]);
CHECK_VASTATUS(va_status, "vaUnmapBuffer(SliceParam)");
CHECK_SURF(va_surface_id[sid]);
/* Set up slice data buffer and copy H.264 encoded data */
if (va_d_param_buf_id[sid] == VA_INVALID_ID) {
/* TODO use estimation matching framebuffer dimensions instead of this large value */
va_status = vaCreateBuffer(va_dpy, va_context_id, VASliceDataBufferType, 4177920, 1, NULL, &va_d_param_buf_id[sid]); /* 1080p size */
CHECK_VASTATUS(va_status, "vaCreateBuffer(SliceData)");
}
char *slice_data_buf;
va_status = vaMapBuffer(va_dpy, va_d_param_buf_id[sid], (void **)&slice_data_buf);
CHECK_VASTATUS(va_status, "vaMapBuffer(SliceData)");
memcpy(slice_data_buf, framedata, framesize);
CHECK_SURF(va_surface_id[sid]);
va_status = vaUnmapBuffer(va_dpy, va_d_param_buf_id[sid]);
CHECK_VASTATUS(va_status, "vaUnmapBuffer(SliceData)");
buffer_ids[0] = va_sp_param_buf_id[sid];
buffer_ids[1] = va_d_param_buf_id[sid];
CHECK_SURF(va_surface_id[sid]);
va_status = vaRenderPicture(va_dpy, va_context_id, buffer_ids, 2);
CHECK_VASTATUS(va_status, "vaRenderPicture");
va_status = vaEndPicture(va_dpy, va_context_id);
CHECK_VASTATUS(va_status, "vaEndPicture");
/* Prepare next one... */
int sid_new = (sid + 1) % SURFACE_NUM;
DebugLog(("%s: new Surface ID = %d\n", __FUNCTION__, sid_new));
va_status = vaBeginPicture(va_dpy, va_context_id, va_surface_id[sid_new]);
CHECK_VASTATUS(va_status, "vaBeginPicture");
/* Get decoded data */
va_status = vaSyncSurface(va_dpy, va_surface_id[sid]);
CHECK_VASTATUS(va_status, "vaSyncSurface");
CHECK_SURF(va_surface_id[sid]);
curr_surface = va_surface_id[sid];
sid = sid_new;
field_order_count += 2;
++frame_id;
if (frame_id > 15) {
frame_id = 0;
}
++num_frames;
memcpy(&va_old_picture_h264, &va_picture_h264, sizeof(VAPictureH264));
}
static void h264_init_decoder(int width, int height)
{
VAStatus va_status;
if (va_context_id) {
rfbClientLog("%s: va_dpy already initialized\n", __FUNCTION__);
}
if (va_dpy != NULL) {
rfbClientLog("%s: Re-initializing H.264 decoder\n", __FUNCTION__);
}
else {
rfbClientLog("%s: initializing H.264 decoder\n", __FUNCTION__);
/* Attach VA display to local X display */
Display *win_display = (Display *)XOpenDisplay(":0.0");
if (win_display == NULL) {
rfbClientErr("Can't connect to local display\n");
exit(-1);
}
int major_ver, minor_ver;
va_dpy = vaGetDisplay(win_display);
va_status = vaInitialize(va_dpy, &major_ver, &minor_ver);
CHECK_VASTATUS(va_status, "vaInitialize");
rfbClientLog("%s: libva version %d.%d found\n", __FUNCTION__, major_ver, minor_ver);
}
/* Check for VLD entrypoint */
int num_entrypoints;
VAEntrypoint entrypoints[5];
int vld_entrypoint_found = 0;
/* Change VAProfileH264High if needed */
VAProfile profile = VAProfileH264High;
va_status = vaQueryConfigEntrypoints(va_dpy, profile, entrypoints, &num_entrypoints);
CHECK_VASTATUS(va_status, "vaQueryConfigEntrypoints");
int i;
for (i = 0; i < num_entrypoints; ++i) {
if (entrypoints[i] == VAEntrypointVLD) {
vld_entrypoint_found = 1;
break;
}
}
if (vld_entrypoint_found == 0) {
rfbClientErr("VLD entrypoint not found\n");
exit(1);
}
/* Create configuration for the decode pipeline */
VAConfigAttrib attrib;
attrib.type = VAConfigAttribRTFormat;
va_status = vaCreateConfig(va_dpy, profile, VAEntrypointVLD, &attrib, 1, &va_config_id);
CHECK_VASTATUS(va_status, "vaCreateConfig");
/* Create VA surfaces */
for (i = 0; i < SURFACE_NUM; ++i) {
va_surface_id[i] = VA_INVALID_ID;
va_pic_param_buf_id[i] = VA_INVALID_ID;
va_mat_param_buf_id[i] = VA_INVALID_ID;
va_sp_param_buf_id[i] = VA_INVALID_ID;
va_d_param_buf_id[i] = VA_INVALID_ID;
}
va_status = vaCreateSurfaces(va_dpy, width, height, VA_RT_FORMAT_YUV420, SURFACE_NUM, &va_surface_id[0]);
CHECK_VASTATUS(va_status, "vaCreateSurfaces");
for (i = 0; i < SURFACE_NUM; ++i) {
DebugLog(("%s: va_surface_id[%d] = %p\n", __FUNCTION__, i, va_surface_id[i]));
}
/* Create VA context */
va_status = vaCreateContext(va_dpy, va_config_id, width, height, 0/*VA_PROGRESSIVE*/, &va_surface_id[0], SURFACE_NUM, &va_context_id);
CHECK_VASTATUS(va_status, "vaCreateContext");
DebugLog(("%s: VA context created (id: %d)\n", __FUNCTION__, va_context_id));
/* Instantiate decode pipeline */
va_status = vaBeginPicture(va_dpy, va_context_id, va_surface_id[0]);
CHECK_VASTATUS(va_status, "vaBeginPicture");
rfbClientLog("%s: H.264 decoder initialized\n", __FUNCTION__);
}
int main(int argc,char **argv)
{
int major_ver, minor_ver;
VAStatus va_status;
pthread_t thread1;
int ret;
char c;
int i;
while ((c =getopt(argc,argv,"w:h:g:r:d:f:tcep?n:v") ) != EOF) {
switch (c) {
case '?':
printf("putsurface <options>\n");
printf(" -g <widthxheight+x_location+y_location> window geometry\n");
printf(" -w/-h resolution of surface\n");
printf(" -r <framerate>\n");
printf(" -d the dimension of black/write square box, default is 32\n");
printf(" -t multi-threads\n");
printf(" -c test clipbox\n");
printf(" -f <1/2> top field, or bottom field\n");
printf(" -v verbose output\n");
exit(0);
break;
case 'g':
ret = sscanf(optarg, "%dx%d+%d+%d", &win_width, &win_height, &win_x, &win_y);
if (ret != 4) {
printf("invalid window geometry, must be widthxheight+x_location+y_location\n");
exit(0);
} else
printf("Create window at (%d, %d), width = %d, height = %d\n",
win_x, win_y, win_width, win_height);
break;
case 'r':
frame_rate = atoi(optarg);
break;
case 'w':
surface_width = atoi(optarg);
break;
case 'h':
surface_height = atoi(optarg);
break;
case 'n':
frame_num_total = atoi(optarg);
break;
case 'd':
box_width = atoi(optarg);
break;
case 't':
multi_thread = 1;
printf("Two threads to do vaPutSurface\n");
break;
case 'e':
check_event = 0;
break;
case 'p':
put_pixmap = 1;
break;
case 'c':
test_clip = 1;
break;
case 'f':
if (atoi(optarg) == 1) {
printf("Display TOP field\n");
display_field = VA_TOP_FIELD;
} else if (atoi(optarg) == 2) {
printf("Display BOTTOM field\n");
display_field = VA_BOTTOM_FIELD;
} else
printf("The validate input for -f is: 1(top field)/2(bottom field)\n");
break;
case 'v':
verbose = 1;
printf("Enable verbose output\n");
break;
}
}
win_display = (void *)open_display();
if (win_display == NULL) {
fprintf(stderr, "Can't open the connection of display!\n");
exit(-1);
}
create_window(win_display, win_x, win_y, win_width, win_height);
va_dpy = vaGetDisplay(win_display);
va_status = vaInitialize(va_dpy, &major_ver, &minor_ver);
CHECK_VASTATUS(va_status, "vaInitialize");
va_status = vaCreateSurfaces(va_dpy,surface_width, surface_height,
VA_RT_FORMAT_YUV420, SURFACE_NUM, &surface_id[0]);
CHECK_VASTATUS(va_status, "vaCreateSurfaces");
if (multi_thread == 0) /* upload the content for all surfaces */
upload_source_YUV_once_for_all();
if (check_event)
pthread_mutex_init(&gmutex, NULL);
for(i = 0; i< SURFACE_NUM; i++)
pthread_mutex_init(&surface_mutex[i], NULL);
if (multi_thread == 1)
ret = pthread_create(&thread1, NULL, putsurface_thread, (void*)drawable_thread1);
putsurface_thread((void *)drawable_thread0);
if (multi_thread == 1)
pthread_join(thread1, (void **)&ret);
printf("thread1 is free\n");
vaDestroySurfaces(va_dpy,&surface_id[0],SURFACE_NUM);
vaTerminate(va_dpy);
close_display(win_display);
return 0;
}
static void* putsurface_thread(void *data)
{
int width=win_width, height=win_height;
void *drawable = data;
int quit = 0;
VAStatus vaStatus;
int row_shift = 0;
int index = 0;
unsigned int frame_num=0, start_time, putsurface_time;
VARectangle cliprects[2]; /* client supplied clip list */
int continue_display = 0;
if (drawable == drawable_thread0)
printf("Enter into thread0\n\n");
if (drawable == drawable_thread1)
printf("Enter into thread1\n\n");
putsurface_time = 0;
while (!quit) {
VASurfaceID surface_id = VA_INVALID_SURFACE;
while (surface_id == VA_INVALID_SURFACE)
surface_id = get_next_free_surface(&index);
if (verbose) printf("Thread %x Display surface 0x%p,\n", (unsigned int)drawable, (void *)surface_id);
if (multi_thread)
upload_surface(va_dpy, surface_id, box_width, row_shift, display_field);
if (check_event)
pthread_mutex_lock(&gmutex);
start_time = get_tick_count();
if ((continue_display == 0) && getenv("FRAME_STOP")) {
char c;
printf("Press any key to display frame %d...(c/C to continue)\n", frame_num);
c = getchar();
if (c == 'c' || c == 'C')
continue_display = 1;
}
vaStatus = vaPutSurface(va_dpy, surface_id, CAST_DRAWABLE(drawable),
0,0,surface_width,surface_height,
0,0,width,height,
(test_clip==0)?NULL:&cliprects[0],
(test_clip==0)?0:2,
display_field);
CHECK_VASTATUS(vaStatus,"vaPutSurface");
putsurface_time += (get_tick_count() - start_time);
if (check_event)
pthread_mutex_unlock(&gmutex);
pthread_mutex_unlock(&surface_mutex[index]); /* locked in get_next_free_surface */
if ((frame_num % 0xff) == 0) {
fprintf(stderr, "%.2f FPS \r", 256000.0 / (float)putsurface_time);
putsurface_time = 0;
update_clipbox(cliprects, width, height);
}
if (check_event)
check_window_event(win_display, drawable, &width, &height, &quit);
if (multi_thread) { /* reload surface content */
row_shift++;
if (row_shift==(2*box_width)) row_shift= 0;
}
if (frame_rate != 0) /* rough framerate control */
usleep(1000/frame_rate*1000);
frame_num++;
if (frame_num >= frame_num_total)
quit = 1;
}
if (drawable == drawable_thread1)
pthread_exit(NULL);
return 0;
}
