//#include "sysdeps.h"
#include "h264encode.h"
+#include <movit/util.h>
#include <EGL/eglplatform.h>
#include <X11/X.h>
#include <X11/Xlib.h>
#include <stdio.h>
#include <stdlib.h>
#include <string.h>
+#include <fcntl.h>
#include <va/va.h>
+#include <va/va_drm.h>
#include <va/va_drmcommon.h>
#include <va/va_enc_h264.h>
#include <va/va_x11.h>
+#include <algorithm>
#include <condition_variable>
#include <cstdint>
#include <map>
#include "context.h"
#include "defs.h"
+#include "flags.h"
#include "httpd.h"
#include "timebase.h"
exit(1); \
}
+#define BUFFER_OFFSET(i) ((char *)NULL + (i))
+
//#include "loadsurface.h"
#define NAL_REF_IDC_NONE 0
#define BITSTREAM_ALLOCATE_STEPPING 4096
#define SURFACE_NUM 16 /* 16 surfaces for source YUV */
+#define MAX_NUM_REF1 16 // Seemingly a hardware-fixed value, not related to SURFACE_NUM
+#define MAX_NUM_REF2 32 // Seemingly a hardware-fixed value, not related to SURFACE_NUM
static constexpr unsigned int MaxFrameNum = (2<<16);
static constexpr unsigned int MaxPicOrderCntLsb = (2<<8);
using namespace std;
+// H.264 video comes out in encoding order (e.g. with two B-frames:
+// 0, 3, 1, 2, 6, 4, 5, etc.), but uncompressed video needs to
+// come in the right order. Since we do everything, including waiting
+// for the frames to come out of OpenGL, in encoding order, we need
+// a reordering buffer for uncompressed frames so that they come out
+// correctly. We go the super-lazy way of not making it understand
+// anything about the true order (which introduces some extra latency,
+// though); we know that for N B-frames we need at most (N-1) frames
+// in the reorder buffer, and can just sort on that.
+//
+// The class also deals with keeping a freelist as needed.
+class FrameReorderer {
+public:
+ FrameReorderer(unsigned queue_length, int width, int height);
+
+ // Returns the next frame to insert with its pts, if any. Otherwise -1 and nullptr.
+ // Does _not_ take ownership of data; a copy is taken if needed.
+ // The returned pointer is valid until the next call to reorder_frame, or destruction.
+ // As a special case, if queue_length == 0, will just return pts and data (no reordering needed).
+ pair<int64_t, const uint8_t *> reorder_frame(int64_t pts, const uint8_t *data);
+
+ // The same as reorder_frame, but without inserting anything. Used to empty the queue.
+ pair<int64_t, const uint8_t *> get_first_frame();
+
+ bool empty() const { return frames.empty(); }
+
+private:
+ unsigned queue_length;
+ int width, height;
+
+ priority_queue<pair<int64_t, uint8_t *>> frames;
+ stack<uint8_t *> freelist; // Includes the last value returned from reorder_frame.
+
+ // Owns all the pointers. Normally, freelist and frames could do this themselves,
+ // except priority_queue doesn't work well with movable-only types.
+ vector<unique_ptr<uint8_t[]>> owner;
+};
+
+FrameReorderer::FrameReorderer(unsigned queue_length, int width, int height)
+ : queue_length(queue_length), width(width), height(height)
+{
+ for (unsigned i = 0; i < queue_length; ++i) {
+ owner.emplace_back(new uint8_t[width * height * 2]);
+ freelist.push(owner.back().get());
+ }
+}
+
+pair<int64_t, const uint8_t *> FrameReorderer::reorder_frame(int64_t pts, const uint8_t *data)
+{
+ if (queue_length == 0) {
+ return make_pair(pts, data);
+ }
+
+ assert(!freelist.empty());
+ uint8_t *storage = freelist.top();
+ freelist.pop();
+ memcpy(storage, data, width * height * 2);
+ frames.emplace(-pts, storage); // Invert pts to get smallest first.
+
+ if (frames.size() >= queue_length) {
+ return get_first_frame();
+ } else {
+ return make_pair(-1, nullptr);
+ }
+}
+
+pair<int64_t, const uint8_t *> FrameReorderer::get_first_frame()
+{
+ assert(!frames.empty());
+ pair<int64_t, uint8_t *> storage = frames.top();
+ frames.pop();
+ int64_t pts = storage.first;
+ freelist.push(storage.second);
+ return make_pair(-pts, storage.second); // Re-invert pts (see reorder_frame()).
+}
+
class H264EncoderImpl {
public:
- H264EncoderImpl(QSurface *surface, int width, int height, HTTPD *httpd);
+ H264EncoderImpl(QSurface *surface, const string &va_display, int width, int height, HTTPD *httpd);
~H264EncoderImpl();
- void add_audio(int64_t pts, std::vector<float> audio); // Needs to come before end_frame() of same pts.
+ void add_audio(int64_t pts, vector<float> audio);
bool begin_frame(GLuint *y_tex, GLuint *cbcr_tex);
- void end_frame(RefCountedGLsync fence, int64_t pts, const std::vector<RefCountedFrame> &input_frames);
+ void end_frame(RefCountedGLsync fence, int64_t pts, const vector<RefCountedFrame> &input_frames);
+ void shutdown();
private:
struct storage_task {
unsigned long long display_order;
int frame_type;
- std::vector<float> audio;
+ vector<float> audio;
int64_t pts, dts;
};
struct PendingFrame {
RefCountedGLsync fence;
- std::vector<RefCountedFrame> input_frames;
+ vector<RefCountedFrame> input_frames;
int64_t pts;
};
void encode_thread_func();
void encode_remaining_frames_as_p(int encoding_frame_num, int gop_start_display_frame_num, int64_t last_dts);
+ void add_packet_for_uncompressed_frame(int64_t pts, const uint8_t *data);
void encode_frame(PendingFrame frame, int encoding_frame_num, int display_frame_num, int gop_start_display_frame_num,
int frame_type, int64_t pts, int64_t dts);
void storage_task_thread();
void slice_header(bitstream *bs);
int build_packed_seq_buffer(unsigned char **header_buffer);
int build_packed_slice_buffer(unsigned char **header_buffer);
- int init_va();
+ int init_va(const string &va_display);
int deinit_va();
- VADisplay va_open_display(void);
+ VADisplay va_open_display(const string &va_display);
void va_close_display(VADisplay va_dpy);
int setup_encode();
int release_encode();
void update_ReferenceFrames(int frame_type);
int update_RefPicList(int frame_type);
- std::thread encode_thread, storage_thread;
+ bool is_shutdown = false;
+ bool use_zerocopy;
+ int drm_fd = -1;
- std::mutex storage_task_queue_mutex;
- std::condition_variable storage_task_queue_changed;
+ thread encode_thread, storage_thread;
+
+ mutex storage_task_queue_mutex;
+ condition_variable storage_task_queue_changed;
int srcsurface_status[SURFACE_NUM]; // protected by storage_task_queue_mutex
- std::queue<storage_task> storage_task_queue; // protected by storage_task_queue_mutex
+ queue<storage_task> storage_task_queue; // protected by storage_task_queue_mutex
bool storage_thread_should_quit = false; // protected by storage_task_queue_mutex
- std::mutex frame_queue_mutex;
- std::condition_variable frame_queue_nonempty;
+ mutex frame_queue_mutex;
+ condition_variable frame_queue_nonempty;
bool encode_thread_should_quit = false; // under frame_queue_mutex
int current_storage_frame;
- std::map<int, PendingFrame> pending_video_frames; // under frame_queue_mutex
- std::map<int64_t, std::vector<float>> pending_audio_frames; // under frame_queue_mutex
+ map<int, PendingFrame> pending_video_frames; // under frame_queue_mutex
+ map<int64_t, vector<float>> pending_audio_frames; // under frame_queue_mutex
QSurface *surface;
AVCodecContext *context_audio;
+ AVFrame *audio_frame = nullptr;
HTTPD *httpd;
+ unique_ptr<FrameReorderer> reorderer;
- Display *x11_display;
- Window x11_window;
+ Display *x11_display = nullptr;
// Encoder parameters
VADisplay va_dpy;
VAImage surface_image;
GLuint y_tex, cbcr_tex;
+
+ // Only if use_zerocopy == true.
EGLImage y_egl_image, cbcr_egl_image;
+
+ // Only if use_zerocopy == false.
+ GLuint pbo;
+ uint8_t *y_ptr, *cbcr_ptr;
+ size_t y_offset, cbcr_offset;
};
GLSurface gl_surfaces[SURFACE_NUM];
VAEncPictureParameterBufferH264 pic_param;
VAEncSliceParameterBufferH264 slice_param;
VAPictureH264 CurrentCurrPic;
- VAPictureH264 ReferenceFrames[16], RefPicList0_P[32], RefPicList0_B[32], RefPicList1_B[32];
+ VAPictureH264 ReferenceFrames[MAX_NUM_REF1], RefPicList0_P[MAX_NUM_REF2], RefPicList0_B[MAX_NUM_REF2], RefPicList1_B[MAX_NUM_REF2];
// Static quality settings.
static constexpr unsigned int frame_bitrate = 15000000 / 60; // Doesn't really matter; only initial_qp does.
int frame_height_mbaligned;
};
-
// Supposedly vaRenderPicture() is supposed to destroy the buffer implicitly,
// but if we don't delete it here, we get leaks. The GStreamer implementation
// does the same.
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);
+ if (bit_left >= 32) {
+ bs->buffer[pos] = (val >> size_in_bits);
+ } else {
+ bs->buffer[pos] = (bs->buffer[pos] << bit_left) | (val >> size_in_bits);
+ }
bs->buffer[pos] = va_swap32(bs->buffer[pos]);
if (pos + 1 == bs->max_size_in_dword) {
}
}
-VADisplay H264EncoderImpl::va_open_display(void)
+VADisplay H264EncoderImpl::va_open_display(const string &va_display)
{
- x11_display = XOpenDisplay(NULL);
- if (!x11_display) {
- fprintf(stderr, "error: can't connect to X server!\n");
- return NULL;
- }
- return vaGetDisplay(x11_display);
+ if (va_display.empty()) {
+ x11_display = XOpenDisplay(NULL);
+ if (!x11_display) {
+ fprintf(stderr, "error: can't connect to X server!\n");
+ return NULL;
+ }
+ use_zerocopy = true;
+ if (global_flags.uncompressed_video_to_http) {
+ fprintf(stderr, "Disabling zerocopy H.264 encoding due to --uncompressed_video_to_http.\n");
+ use_zerocopy = false;
+ }
+ return vaGetDisplay(x11_display);
+ } else if (va_display[0] != '/') {
+ x11_display = XOpenDisplay(va_display.c_str());
+ if (!x11_display) {
+ fprintf(stderr, "error: can't connect to X server!\n");
+ return NULL;
+ }
+ use_zerocopy = true;
+ if (global_flags.uncompressed_video_to_http) {
+ fprintf(stderr, "Disabling zerocopy H.264 encoding due to --uncompressed_video_to_http.\n");
+ use_zerocopy = false;
+ }
+ return vaGetDisplay(x11_display);
+ } else {
+ drm_fd = open(va_display.c_str(), O_RDWR);
+ if (drm_fd == -1) {
+ perror(va_display.c_str());
+ return NULL;
+ }
+ use_zerocopy = false;
+ return vaGetDisplayDRM(drm_fd);
+ }
}
void H264EncoderImpl::va_close_display(VADisplay va_dpy)
{
- if (!x11_display)
- return;
-
- if (x11_window) {
- XUnmapWindow(x11_display, x11_window);
- XDestroyWindow(x11_display, x11_window);
- x11_window = None;
- }
- XCloseDisplay(x11_display);
- x11_display = NULL;
+ if (x11_display) {
+ XCloseDisplay(x11_display);
+ x11_display = nullptr;
+ }
+ if (drm_fd != -1) {
+ close(drm_fd);
+ }
}
-int H264EncoderImpl::init_va()
+int H264EncoderImpl::init_va(const string &va_display)
{
VAProfile profile_list[]={VAProfileH264High, VAProfileH264Main, VAProfileH264Baseline, VAProfileH264ConstrainedBaseline};
VAEntrypoint *entrypoints;
VAStatus va_status;
unsigned int i;
- va_dpy = va_open_display();
+ va_dpy = va_open_display(va_display);
va_status = vaInitialize(va_dpy, &major_ver, &minor_ver);
CHECK_VASTATUS(va_status, "vaInitialize");
}
if (support_encode == 0) {
- printf("Can't find VAEntrypointEncSlice for H264 profiles\n");
+ printf("Can't find VAEntrypointEncSlice for H264 profiles. If you are using a non-Intel GPU\n");
+ printf("but have one in your system, try launching Nageru with --va-display /dev/dri/renderD128\n");
+ printf("to use VA-API against DRM instead of X11.\n");
exit(1);
} else {
switch (h264_profile) {
for (i = 0; i < SURFACE_NUM; i++) {
glGenTextures(1, &gl_surfaces[i].y_tex);
glGenTextures(1, &gl_surfaces[i].cbcr_tex);
+
+ if (!use_zerocopy) {
+ // Create Y image.
+ glBindTexture(GL_TEXTURE_2D, gl_surfaces[i].y_tex);
+ glTexStorage2D(GL_TEXTURE_2D, 1, GL_R8, frame_width, frame_height);
+
+ // Create CbCr image.
+ glBindTexture(GL_TEXTURE_2D, gl_surfaces[i].cbcr_tex);
+ glTexStorage2D(GL_TEXTURE_2D, 1, GL_RG8, frame_width / 2, frame_height / 2);
+
+ // Generate a PBO to read into. It doesn't necessarily fit 1:1 with the VA-API
+ // buffers, due to potentially differing pitch.
+ glGenBuffers(1, &gl_surfaces[i].pbo);
+ glBindBuffer(GL_PIXEL_PACK_BUFFER, gl_surfaces[i].pbo);
+ glBufferStorage(GL_PIXEL_PACK_BUFFER, frame_width * frame_height * 2, nullptr, GL_MAP_READ_BIT | GL_MAP_WRITE_BIT | GL_MAP_PERSISTENT_BIT);
+ uint8_t *ptr = (uint8_t *)glMapBufferRange(GL_PIXEL_PACK_BUFFER, 0, frame_width * frame_height * 2, GL_MAP_READ_BIT | GL_MAP_PERSISTENT_BIT);
+ gl_surfaces[i].y_offset = 0;
+ gl_surfaces[i].cbcr_offset = frame_width * frame_height;
+ gl_surfaces[i].y_ptr = ptr + gl_surfaces[i].y_offset;
+ gl_surfaces[i].cbcr_ptr = ptr + gl_surfaces[i].cbcr_offset;
+ glBindBuffer(GL_PIXEL_PACK_BUFFER, 0);
+ }
}
for (i = 0; i < SURFACE_NUM; i++) {
return 0;
}
-
-
-#define partition(ref, field, key, ascending) \
- while (i <= j) { \
- if (ascending) { \
- while (ref[i].field < key) \
- i++; \
- while (ref[j].field > key) \
- j--; \
- } else { \
- while (ref[i].field > key) \
- i++; \
- while (ref[j].field < key) \
- j--; \
- } \
- if (i <= j) { \
- tmp = ref[i]; \
- ref[i] = ref[j]; \
- ref[j] = tmp; \
- i++; \
- j--; \
- } \
- } \
-
-static void sort_one(VAPictureH264 ref[], int left, int right,
- int ascending, int frame_idx)
+// Given a list like 1 9 3 0 2 8 4 and a pivot element 3, will produce
+//
+// 2 1 0 [3] 4 8 9
+template<class T, class C>
+static void sort_two(T *begin, T *end, const T &pivot, const C &less_than)
{
- int i = left, j = right;
- unsigned int key;
- VAPictureH264 tmp;
-
- if (frame_idx) {
- key = ref[(left + right) / 2].frame_idx;
- partition(ref, frame_idx, key, ascending);
- } else {
- key = ref[(left + right) / 2].TopFieldOrderCnt;
- partition(ref, TopFieldOrderCnt, (signed int)key, ascending);
- }
-
- /* recursion */
- if (left < j)
- sort_one(ref, left, j, ascending, frame_idx);
-
- if (i < right)
- sort_one(ref, i, right, ascending, frame_idx);
-}
-
-static void sort_two(VAPictureH264 ref[], int left, int right, unsigned int key, unsigned int frame_idx,
- int partition_ascending, int list0_ascending, int list1_ascending)
-{
- int i = left, j = right;
- VAPictureH264 tmp;
-
- if (frame_idx) {
- partition(ref, frame_idx, key, partition_ascending);
- } else {
- partition(ref, TopFieldOrderCnt, (signed int)key, partition_ascending);
- }
-
-
- sort_one(ref, left, i-1, list0_ascending, frame_idx);
- sort_one(ref, j+1, right, list1_ascending, frame_idx);
+ T *middle = partition(begin, end, [&](const T &elem) { return less_than(elem, pivot); });
+ sort(begin, middle, [&](const T &a, const T &b) { return less_than(b, a); });
+ sort(middle, end, less_than);
}
void H264EncoderImpl::update_ReferenceFrames(int frame_type)
int H264EncoderImpl::update_RefPicList(int frame_type)
{
- unsigned int current_poc = CurrentCurrPic.TopFieldOrderCnt;
+ const auto descending_by_frame_idx = [](const VAPictureH264 &a, const VAPictureH264 &b) {
+ return a.frame_idx > b.frame_idx;
+ };
+ const auto ascending_by_top_field_order_cnt = [](const VAPictureH264 &a, const VAPictureH264 &b) {
+ return a.TopFieldOrderCnt < b.TopFieldOrderCnt;
+ };
+ const auto descending_by_top_field_order_cnt = [](const VAPictureH264 &a, const VAPictureH264 &b) {
+ return a.TopFieldOrderCnt > b.TopFieldOrderCnt;
+ };
if (frame_type == FRAME_P) {
memcpy(RefPicList0_P, ReferenceFrames, numShortTerm * sizeof(VAPictureH264));
- sort_one(RefPicList0_P, 0, numShortTerm-1, 0, 1);
- }
-
- if (frame_type == FRAME_B) {
+ sort(&RefPicList0_P[0], &RefPicList0_P[numShortTerm], descending_by_frame_idx);
+ } else if (frame_type == FRAME_B) {
memcpy(RefPicList0_B, ReferenceFrames, numShortTerm * sizeof(VAPictureH264));
- sort_two(RefPicList0_B, 0, numShortTerm-1, current_poc, 0,
- 1, 0, 1);
+ sort_two(&RefPicList0_B[0], &RefPicList0_B[numShortTerm], CurrentCurrPic, ascending_by_top_field_order_cnt);
memcpy(RefPicList1_B, ReferenceFrames, numShortTerm * sizeof(VAPictureH264));
- sort_two(RefPicList1_B, 0, numShortTerm-1, current_poc, 0,
- 0, 1, 0);
+ sort_two(&RefPicList1_B[0], &RefPicList1_B[numShortTerm], CurrentCurrPic, descending_by_top_field_order_cnt);
}
return 0;
CurrentCurrPic = pic_param.CurrPic;
memcpy(pic_param.ReferenceFrames, ReferenceFrames, numShortTerm*sizeof(VAPictureH264));
- for (i = numShortTerm; i < SURFACE_NUM; i++) {
+ for (i = numShortTerm; i < MAX_NUM_REF1; i++) {
pic_param.ReferenceFrames[i].picture_id = VA_INVALID_SURFACE;
pic_param.ReferenceFrames[i].flags = VA_PICTURE_H264_INVALID;
}
int refpiclist0_max = h264_maxref & 0xffff;
memcpy(slice_param.RefPicList0, RefPicList0_P, refpiclist0_max*sizeof(VAPictureH264));
- for (i = refpiclist0_max; i < 32; i++) {
+ for (i = refpiclist0_max; i < MAX_NUM_REF2; i++) {
slice_param.RefPicList0[i].picture_id = VA_INVALID_SURFACE;
slice_param.RefPicList0[i].flags = VA_PICTURE_H264_INVALID;
}
int refpiclist1_max = (h264_maxref >> 16) & 0xffff;
memcpy(slice_param.RefPicList0, RefPicList0_B, refpiclist0_max*sizeof(VAPictureH264));
- for (i = refpiclist0_max; i < 32; i++) {
+ for (i = refpiclist0_max; i < MAX_NUM_REF2; i++) {
slice_param.RefPicList0[i].picture_id = VA_INVALID_SURFACE;
slice_param.RefPicList0[i].flags = VA_PICTURE_H264_INVALID;
}
memcpy(slice_param.RefPicList1, RefPicList1_B, refpiclist1_max*sizeof(VAPictureH264));
- for (i = refpiclist1_max; i < 32; i++) {
+ for (i = refpiclist1_max; i < MAX_NUM_REF2; i++) {
slice_param.RefPicList1[i].picture_id = VA_INVALID_SURFACE;
slice_param.RefPicList1[i].flags = VA_PICTURE_H264_INVALID;
}
string data;
- const int64_t global_delay = (ip_period - 1) * (TIMEBASE / MAX_FPS); // So we never get negative dts.
+ const int64_t global_delay = int64_t(ip_period - 1) * (TIMEBASE / MAX_FPS); // So we never get negative dts.
va_status = vaMapBuffer(va_dpy, gl_surfaces[task.display_order % SURFACE_NUM].coded_buf, (void **)(&buf_list));
CHECK_VASTATUS(va_status, "vaMapBuffer");
pkt.flags = 0;
}
//pkt.duration = 1;
- httpd->add_packet(pkt, task.pts + global_delay, task.dts + global_delay);
+ httpd->add_packet(pkt, task.pts + global_delay, task.dts + global_delay,
+ global_flags.uncompressed_video_to_http ? HTTPD::DESTINATION_FILE_ONLY : HTTPD::DESTINATION_FILE_AND_HTTP);
}
// Encode and add all audio frames up to and including the pts of this video frame.
for ( ;; ) {
pending_audio_frames.erase(it);
}
- AVFrame *frame = avcodec_alloc_frame();
- frame->nb_samples = audio.size() / 2;
- frame->format = AV_SAMPLE_FMT_S32;
- frame->channel_layout = AV_CH_LAYOUT_STEREO;
+ audio_frame->nb_samples = audio.size() / 2;
+ audio_frame->format = AV_SAMPLE_FMT_S32;
+ audio_frame->channel_layout = AV_CH_LAYOUT_STEREO;
unique_ptr<int32_t[]> int_samples(new int32_t[audio.size()]);
- int ret = avcodec_fill_audio_frame(frame, 2, AV_SAMPLE_FMT_S32, (const uint8_t*)int_samples.get(), audio.size() * sizeof(int32_t), 1);
+ int ret = avcodec_fill_audio_frame(audio_frame, 2, AV_SAMPLE_FMT_S32, (const uint8_t*)int_samples.get(), audio.size() * sizeof(int32_t), 1);
if (ret < 0) {
fprintf(stderr, "avcodec_fill_audio_frame() failed with %d\n", ret);
exit(1);
}
- for (int i = 0; i < frame->nb_samples * 2; ++i) {
+ for (int i = 0; i < audio_frame->nb_samples * 2; ++i) {
if (audio[i] >= 1.0f) {
int_samples[i] = 2147483647;
} else if (audio[i] <= -1.0f) {
pkt.data = nullptr;
pkt.size = 0;
int got_output;
- avcodec_encode_audio2(context_audio, &pkt, frame, &got_output);
+ avcodec_encode_audio2(context_audio, &pkt, audio_frame, &got_output);
if (got_output) {
pkt.stream_index = 1;
- httpd->add_packet(pkt, audio_pts + global_delay, audio_pts + global_delay);
+ httpd->add_packet(pkt, audio_pts + global_delay, audio_pts + global_delay, HTTPD::DESTINATION_FILE_AND_HTTP);
}
// TODO: Delayed frames.
- avcodec_free_frame(&frame);
+ av_frame_unref(audio_frame);
av_free_packet(&pkt);
if (audio_pts == task.pts) break;
}
-
-#if 0
- printf("\r "); /* return back to startpoint */
- switch (encode_order % 4) {
- case 0:
- printf("|");
- break;
- case 1:
- printf("/");
- break;
- case 2:
- printf("-");
- break;
- case 3:
- printf("\\");
- break;
- }
- printf("%08lld", encode_order);
-#endif
}
{
unique_lock<mutex> lock(storage_task_queue_mutex);
storage_task_queue.push(move(task));
- srcsurface_status[task.display_order % SURFACE_NUM] = SRC_SURFACE_IN_ENCODING;
storage_task_queue_changed.notify_all();
}
int H264EncoderImpl::release_encode()
{
- int i;
-
- for (i = 0; i < SURFACE_NUM; i++) {
- vaDestroyBuffer(va_dpy, gl_surfaces[i].coded_buf);
- vaDestroySurfaces(va_dpy, &gl_surfaces[i].src_surface, 1);
- vaDestroySurfaces(va_dpy, &gl_surfaces[i].ref_surface, 1);
- }
-
- vaDestroyContext(va_dpy, context_id);
- vaDestroyConfig(va_dpy, config_id);
+ for (unsigned i = 0; i < SURFACE_NUM; i++) {
+ vaDestroyBuffer(va_dpy, gl_surfaces[i].coded_buf);
+ vaDestroySurfaces(va_dpy, &gl_surfaces[i].src_surface, 1);
+ vaDestroySurfaces(va_dpy, &gl_surfaces[i].ref_surface, 1);
+
+ if (!use_zerocopy) {
+ glBindBuffer(GL_PIXEL_PACK_BUFFER, gl_surfaces[i].pbo);
+ glUnmapBuffer(GL_PIXEL_PACK_BUFFER);
+ glBindBuffer(GL_PIXEL_PACK_BUFFER, 0);
+ glDeleteBuffers(1, &gl_surfaces[i].pbo);
+ }
+ glDeleteTextures(1, &gl_surfaces[i].y_tex);
+ glDeleteTextures(1, &gl_surfaces[i].cbcr_tex);
+ }
- return 0;
+ vaDestroyContext(va_dpy, context_id);
+ vaDestroyConfig(va_dpy, config_id);
+
+ return 0;
}
int H264EncoderImpl::deinit_va()
}
-H264EncoderImpl::H264EncoderImpl(QSurface *surface, int width, int height, HTTPD *httpd)
+H264EncoderImpl::H264EncoderImpl(QSurface *surface, const string &va_display, int width, int height, HTTPD *httpd)
: current_storage_frame(0), surface(surface), httpd(httpd)
{
AVCodec *codec_audio = avcodec_find_encoder(AUDIO_OUTPUT_CODEC);
fprintf(stderr, "Could not open codec\n");
exit(1);
}
+ audio_frame = av_frame_alloc();
frame_width = width;
frame_height = height;
//print_input();
- init_va();
+ if (global_flags.uncompressed_video_to_http) {
+ reorderer.reset(new FrameReorderer(ip_period - 1, frame_width, frame_height));
+ }
+
+ init_va(va_display);
setup_encode();
// No frames are ready yet.
H264EncoderImpl::~H264EncoderImpl()
{
- {
- unique_lock<mutex> lock(frame_queue_mutex);
- encode_thread_should_quit = true;
- frame_queue_nonempty.notify_all();
- }
- encode_thread.join();
- {
- unique_lock<mutex> lock(storage_task_queue_mutex);
- storage_thread_should_quit = true;
- frame_queue_nonempty.notify_all();
- storage_task_queue_changed.notify_all();
- }
- storage_thread.join();
+ shutdown();
+ av_frame_free(&audio_frame);
- release_encode();
- deinit_va();
+ // TODO: Destroy context.
}
bool H264EncoderImpl::begin_frame(GLuint *y_tex, GLuint *cbcr_tex)
{
+ assert(!is_shutdown);
{
// Wait until this frame slot is done encoding.
unique_lock<mutex> lock(storage_task_queue_mutex);
+ if (srcsurface_status[current_storage_frame % SURFACE_NUM] != SRC_SURFACE_FREE) {
+ fprintf(stderr, "Warning: Slot %d (for frame %d) is still encoding, rendering has to wait for H.264 encoder\n",
+ current_storage_frame % SURFACE_NUM, current_storage_frame);
+ }
storage_task_queue_changed.wait(lock, [this]{ return storage_thread_should_quit || (srcsurface_status[current_storage_frame % SURFACE_NUM] == SRC_SURFACE_FREE); });
+ srcsurface_status[current_storage_frame % SURFACE_NUM] = SRC_SURFACE_IN_ENCODING;
if (storage_thread_should_quit) return false;
}
*y_tex = surf->y_tex;
*cbcr_tex = surf->cbcr_tex;
- VASurfaceID surface = surf->src_surface;
- VAStatus va_status = vaDeriveImage(va_dpy, surface, &surf->surface_image);
- CHECK_VASTATUS(va_status, "vaDeriveImage");
-
- VABufferInfo buf_info;
- buf_info.mem_type = VA_SURFACE_ATTRIB_MEM_TYPE_DRM_PRIME; // or VA_SURFACE_ATTRIB_MEM_TYPE_KERNEL_DRM?
- va_status = vaAcquireBufferHandle(va_dpy, surf->surface_image.buf, &buf_info);
- CHECK_VASTATUS(va_status, "vaAcquireBufferHandle");
-
- // Create Y image.
- surf->y_egl_image = EGL_NO_IMAGE_KHR;
- EGLint y_attribs[] = {
- EGL_WIDTH, frame_width,
- EGL_HEIGHT, frame_height,
- EGL_LINUX_DRM_FOURCC_EXT, fourcc_code('R', '8', ' ', ' '),
- EGL_DMA_BUF_PLANE0_FD_EXT, EGLint(buf_info.handle),
- EGL_DMA_BUF_PLANE0_OFFSET_EXT, EGLint(surf->surface_image.offsets[0]),
- EGL_DMA_BUF_PLANE0_PITCH_EXT, EGLint(surf->surface_image.pitches[0]),
- EGL_NONE
- };
-
- surf->y_egl_image = eglCreateImageKHR(eglGetCurrentDisplay(), EGL_NO_CONTEXT, EGL_LINUX_DMA_BUF_EXT, NULL, y_attribs);
- assert(surf->y_egl_image != EGL_NO_IMAGE_KHR);
-
- // Associate Y image to a texture.
- glBindTexture(GL_TEXTURE_2D, *y_tex);
- glEGLImageTargetTexture2DOES(GL_TEXTURE_2D, surf->y_egl_image);
-
- // Create CbCr image.
- surf->cbcr_egl_image = EGL_NO_IMAGE_KHR;
- EGLint cbcr_attribs[] = {
- EGL_WIDTH, frame_width,
- EGL_HEIGHT, frame_height,
- EGL_LINUX_DRM_FOURCC_EXT, fourcc_code('G', 'R', '8', '8'),
- EGL_DMA_BUF_PLANE0_FD_EXT, EGLint(buf_info.handle),
- EGL_DMA_BUF_PLANE0_OFFSET_EXT, EGLint(surf->surface_image.offsets[1]),
- EGL_DMA_BUF_PLANE0_PITCH_EXT, EGLint(surf->surface_image.pitches[1]),
- EGL_NONE
- };
-
- surf->cbcr_egl_image = eglCreateImageKHR(eglGetCurrentDisplay(), EGL_NO_CONTEXT, EGL_LINUX_DMA_BUF_EXT, NULL, cbcr_attribs);
- assert(surf->cbcr_egl_image != EGL_NO_IMAGE_KHR);
-
- // Associate CbCr image to a texture.
- glBindTexture(GL_TEXTURE_2D, *cbcr_tex);
- glEGLImageTargetTexture2DOES(GL_TEXTURE_2D, surf->cbcr_egl_image);
+ VAStatus va_status = vaDeriveImage(va_dpy, surf->src_surface, &surf->surface_image);
+ CHECK_VASTATUS(va_status, "vaDeriveImage");
+
+ if (use_zerocopy) {
+ VABufferInfo buf_info;
+ buf_info.mem_type = VA_SURFACE_ATTRIB_MEM_TYPE_DRM_PRIME; // or VA_SURFACE_ATTRIB_MEM_TYPE_KERNEL_DRM?
+ va_status = vaAcquireBufferHandle(va_dpy, surf->surface_image.buf, &buf_info);
+ CHECK_VASTATUS(va_status, "vaAcquireBufferHandle");
+
+ // Create Y image.
+ surf->y_egl_image = EGL_NO_IMAGE_KHR;
+ EGLint y_attribs[] = {
+ EGL_WIDTH, frame_width,
+ EGL_HEIGHT, frame_height,
+ EGL_LINUX_DRM_FOURCC_EXT, fourcc_code('R', '8', ' ', ' '),
+ EGL_DMA_BUF_PLANE0_FD_EXT, EGLint(buf_info.handle),
+ EGL_DMA_BUF_PLANE0_OFFSET_EXT, EGLint(surf->surface_image.offsets[0]),
+ EGL_DMA_BUF_PLANE0_PITCH_EXT, EGLint(surf->surface_image.pitches[0]),
+ EGL_NONE
+ };
+
+ surf->y_egl_image = eglCreateImageKHR(eglGetCurrentDisplay(), EGL_NO_CONTEXT, EGL_LINUX_DMA_BUF_EXT, NULL, y_attribs);
+ assert(surf->y_egl_image != EGL_NO_IMAGE_KHR);
+
+ // Associate Y image to a texture.
+ glBindTexture(GL_TEXTURE_2D, *y_tex);
+ glEGLImageTargetTexture2DOES(GL_TEXTURE_2D, surf->y_egl_image);
+
+ // Create CbCr image.
+ surf->cbcr_egl_image = EGL_NO_IMAGE_KHR;
+ EGLint cbcr_attribs[] = {
+ EGL_WIDTH, frame_width,
+ EGL_HEIGHT, frame_height,
+ EGL_LINUX_DRM_FOURCC_EXT, fourcc_code('G', 'R', '8', '8'),
+ EGL_DMA_BUF_PLANE0_FD_EXT, EGLint(buf_info.handle),
+ EGL_DMA_BUF_PLANE0_OFFSET_EXT, EGLint(surf->surface_image.offsets[1]),
+ EGL_DMA_BUF_PLANE0_PITCH_EXT, EGLint(surf->surface_image.pitches[1]),
+ EGL_NONE
+ };
+
+ surf->cbcr_egl_image = eglCreateImageKHR(eglGetCurrentDisplay(), EGL_NO_CONTEXT, EGL_LINUX_DMA_BUF_EXT, NULL, cbcr_attribs);
+ assert(surf->cbcr_egl_image != EGL_NO_IMAGE_KHR);
+
+ // Associate CbCr image to a texture.
+ glBindTexture(GL_TEXTURE_2D, *cbcr_tex);
+ glEGLImageTargetTexture2DOES(GL_TEXTURE_2D, surf->cbcr_egl_image);
+ }
return true;
}
void H264EncoderImpl::add_audio(int64_t pts, vector<float> audio)
{
+ assert(!is_shutdown);
{
unique_lock<mutex> lock(frame_queue_mutex);
pending_audio_frames[pts] = move(audio);
void H264EncoderImpl::end_frame(RefCountedGLsync fence, int64_t pts, const vector<RefCountedFrame> &input_frames)
{
+ assert(!is_shutdown);
+
+ if (!use_zerocopy) {
+ GLSurface *surf = &gl_surfaces[current_storage_frame % SURFACE_NUM];
+
+ glPixelStorei(GL_PACK_ROW_LENGTH, 0);
+ check_error();
+
+ glBindBuffer(GL_PIXEL_PACK_BUFFER, surf->pbo);
+ check_error();
+
+ glBindTexture(GL_TEXTURE_2D, surf->y_tex);
+ check_error();
+ glGetTexImage(GL_TEXTURE_2D, 0, GL_RED, GL_UNSIGNED_BYTE, BUFFER_OFFSET(surf->y_offset));
+ check_error();
+
+ glBindTexture(GL_TEXTURE_2D, surf->cbcr_tex);
+ check_error();
+ glGetTexImage(GL_TEXTURE_2D, 0, GL_RG, GL_UNSIGNED_BYTE, BUFFER_OFFSET(surf->cbcr_offset));
+ check_error();
+
+ glBindTexture(GL_TEXTURE_2D, 0);
+ check_error();
+ glBindBuffer(GL_PIXEL_PACK_BUFFER, 0);
+ check_error();
+
+ glMemoryBarrier(GL_TEXTURE_UPDATE_BARRIER_BIT | GL_CLIENT_MAPPED_BUFFER_BARRIER_BIT);
+ check_error();
+ fence = RefCountedGLsync(GL_SYNC_GPU_COMMANDS_COMPLETE, /*flags=*/0);
+ check_error();
+ }
+
{
unique_lock<mutex> lock(frame_queue_mutex);
pending_video_frames[current_storage_frame] = PendingFrame{ fence, input_frames, pts };
frame_queue_nonempty.notify_all();
}
+void H264EncoderImpl::shutdown()
+{
+ if (is_shutdown) {
+ return;
+ }
+
+ {
+ unique_lock<mutex> lock(frame_queue_mutex);
+ encode_thread_should_quit = true;
+ frame_queue_nonempty.notify_all();
+ }
+ encode_thread.join();
+ {
+ unique_lock<mutex> lock(storage_task_queue_mutex);
+ storage_thread_should_quit = true;
+ frame_queue_nonempty.notify_all();
+ storage_task_queue_changed.notify_all();
+ }
+ storage_thread.join();
+
+ release_encode();
+ deinit_va();
+ is_shutdown = true;
+}
+
void H264EncoderImpl::encode_thread_func()
{
int64_t last_dts = -1;
encode_frame(frame, encoding_frame_num++, display_frame_num, gop_start_display_frame_num, FRAME_P, frame.pts, dts);
last_dts = dts;
}
+
+ if (global_flags.uncompressed_video_to_http) {
+ // Add frames left in reorderer.
+ while (!reorderer->empty()) {
+ pair<int64_t, const uint8_t *> output_frame = reorderer->get_first_frame();
+ add_packet_for_uncompressed_frame(output_frame.first, output_frame.second);
+ }
+ }
}
+void H264EncoderImpl::add_packet_for_uncompressed_frame(int64_t pts, const uint8_t *data)
+{
+ AVPacket pkt;
+ memset(&pkt, 0, sizeof(pkt));
+ pkt.buf = nullptr;
+ pkt.data = const_cast<uint8_t *>(data);
+ pkt.size = frame_width * frame_height * 2;
+ pkt.stream_index = 0;
+ pkt.flags = AV_PKT_FLAG_KEY;
+ httpd->add_packet(pkt, pts, pts, HTTPD::DESTINATION_HTTP_ONLY);
+}
+
+namespace {
+
+void memcpy_with_pitch(uint8_t *dst, const uint8_t *src, size_t src_width, size_t dst_pitch, size_t height)
+{
+ if (src_width == dst_pitch) {
+ memcpy(dst, src, src_width * height);
+ } else {
+ for (size_t y = 0; y < height; ++y) {
+ const uint8_t *sptr = src + y * src_width;
+ uint8_t *dptr = dst + y * dst_pitch;
+ memcpy(dptr, sptr, src_width);
+ }
+ }
+}
+
+} // namespace
+
void H264EncoderImpl::encode_frame(H264EncoderImpl::PendingFrame frame, int encoding_frame_num, int display_frame_num, int gop_start_display_frame_num,
int frame_type, int64_t pts, int64_t dts)
{
// Wait for the GPU to be done with the frame.
- glClientWaitSync(frame.fence.get(), 0, 0);
+ GLenum sync_status;
+ do {
+ sync_status = glClientWaitSync(frame.fence.get(), 0, 1000000000);
+ check_error();
+ } while (sync_status == GL_TIMEOUT_EXPIRED);
+ assert(sync_status != GL_WAIT_FAILED);
// Release back any input frames we needed to render this frame.
frame.input_frames.clear();
- // Unmap the image.
GLSurface *surf = &gl_surfaces[display_frame_num % SURFACE_NUM];
- eglDestroyImageKHR(eglGetCurrentDisplay(), surf->y_egl_image);
- eglDestroyImageKHR(eglGetCurrentDisplay(), surf->cbcr_egl_image);
- VAStatus va_status = vaReleaseBufferHandle(va_dpy, surf->surface_image.buf);
- CHECK_VASTATUS(va_status, "vaReleaseBufferHandle");
+ VAStatus va_status;
+
+ if (use_zerocopy) {
+ eglDestroyImageKHR(eglGetCurrentDisplay(), surf->y_egl_image);
+ eglDestroyImageKHR(eglGetCurrentDisplay(), surf->cbcr_egl_image);
+ va_status = vaReleaseBufferHandle(va_dpy, surf->surface_image.buf);
+ CHECK_VASTATUS(va_status, "vaReleaseBufferHandle");
+ } else {
+ unsigned char *surface_p = nullptr;
+ vaMapBuffer(va_dpy, surf->surface_image.buf, (void **)&surface_p);
+
+ unsigned char *va_y_ptr = (unsigned char *)surface_p + surf->surface_image.offsets[0];
+ memcpy_with_pitch(va_y_ptr, surf->y_ptr, frame_width, surf->surface_image.pitches[0], frame_height);
+
+ unsigned char *va_cbcr_ptr = (unsigned char *)surface_p + surf->surface_image.offsets[1];
+ memcpy_with_pitch(va_cbcr_ptr, surf->cbcr_ptr, (frame_width / 2) * sizeof(uint16_t), surf->surface_image.pitches[1], frame_height / 2);
+
+ va_status = vaUnmapBuffer(va_dpy, surf->surface_image.buf);
+ CHECK_VASTATUS(va_status, "vaUnmapBuffer");
+
+ if (global_flags.uncompressed_video_to_http) {
+ // Add uncompressed video. (Note that pts == dts here.)
+ const int64_t global_delay = int64_t(ip_period - 1) * (TIMEBASE / MAX_FPS); // Needs to match audio.
+ pair<int64_t, const uint8_t *> output_frame = reorderer->reorder_frame(pts + global_delay, reinterpret_cast<uint8_t *>(surf->y_ptr));
+ if (output_frame.second != nullptr) {
+ add_packet_for_uncompressed_frame(output_frame.first, output_frame.second);
+ }
+ }
+ }
+
va_status = vaDestroyImage(va_dpy, surf->surface_image.image_id);
CHECK_VASTATUS(va_status, "vaDestroyImage");
- VASurfaceID surface = surf->src_surface;
-
// Schedule the frame for encoding.
- va_status = vaBeginPicture(va_dpy, context_id, surface);
+ VASurfaceID va_surface = surf->src_surface;
+ va_status = vaBeginPicture(va_dpy, context_id, va_surface);
CHECK_VASTATUS(va_status, "vaBeginPicture");
if (frame_type == FRAME_IDR) {
}
// Proxy object.
-H264Encoder::H264Encoder(QSurface *surface, int width, int height, HTTPD *httpd)
- : impl(new H264EncoderImpl(surface, width, height, httpd)) {}
+H264Encoder::H264Encoder(QSurface *surface, const string &va_display, int width, int height, HTTPD *httpd)
+ : impl(new H264EncoderImpl(surface, va_display, width, height, httpd)) {}
// Must be defined here because unique_ptr<> destructor needs to know the impl.
H264Encoder::~H264Encoder() {}
-void H264Encoder::add_audio(int64_t pts, std::vector<float> audio)
+void H264Encoder::add_audio(int64_t pts, vector<float> audio)
{
impl->add_audio(pts, audio);
}
return impl->begin_frame(y_tex, cbcr_tex);
}
-void H264Encoder::end_frame(RefCountedGLsync fence, int64_t pts, const std::vector<RefCountedFrame> &input_frames)
+void H264Encoder::end_frame(RefCountedGLsync fence, int64_t pts, const vector<RefCountedFrame> &input_frames)
{
impl->end_frame(fence, pts, input_frames);
}
+void H264Encoder::shutdown()
+{
+ impl->shutdown();
+}
+
// Real class.
projects / nageru / blobdiff