From e9e9555cac3c11c70277824d82d5b3fbd5e4afd2 Mon Sep 17 00:00:00 2001 From: "zejian.su" Date: Fri, 1 Dec 2023 16:30:27 +0800 Subject: [PATCH 2/2] Fix the sync issue between the Preview.Show() and the QWidget.update() 1. Add a mutex between the Preview.Show() and the QWidget.update(). 2. Speed up the Preview.Show() -- change the floating point to integer, merge some plus and multiplying operation and remove some data copy. Signed-off-by: zejian.su --- preview/qt_preview.cpp | 183 +++++++++++++++++++++++++---------------- 1 file changed, 113 insertions(+), 70 deletions(-) diff --git a/preview/qt_preview.cpp b/preview/qt_preview.cpp index 8196a8e..ef3c8a4 100644 --- a/preview/qt_preview.cpp +++ b/preview/qt_preview.cpp @@ -41,16 +41,47 @@ class MyWidget : public QWidget public: MyWidget(QWidget *parent, int w, int h) : QWidget(parent), size(w, h) { - image = QImage(size, QImage::Format_RGB888); - image.fill(0); + for(int i = 0; i < 2; i++) { + buffers_[i].image = QImage(size, QImage::Format_RGB888); + buffers_[i].image.fill(0); + availableBuffers_.push_back(&buffers_[i]); + } } QSize size; - QImage image; + + struct ImageBuffer { + QImage image; + uint8_t frameCounter; + + ImageBuffer() : frameCounter(0) {}; + } buffers_[2]; + + std::list freeBuffers_; + std::list availableBuffers_; + + std::mutex buffers_available_mutex_; + std::mutex buffers_free_mutex_; protected: void paintEvent(QPaintEvent *) override { QPainter painter(this); - painter.drawImage(rect(), image, image.rect()); + { + ImageBuffer *buffer = nullptr; + { + std::lock_guard lock(buffers_available_mutex_); + if(!availableBuffers_.size()) + return; + buffer = availableBuffers_.back(); + availableBuffers_.pop_back(); + } + + painter.drawImage(rect(), buffer->image, buffer->image.rect()); + + { + std::lock_guard lock(buffers_free_mutex_); + freeBuffers_.push_back(buffer); + } + } } QSize sizeHint() const override { return size; } }; @@ -58,7 +89,7 @@ protected: class QtPreview : public Preview { public: - QtPreview(Options const *options) : Preview(options), frame_counter_(0) + QtPreview(Options const *options) : Preview(options) { window_width_ = options->preview_width; window_height_ = options->preview_height; @@ -67,6 +98,8 @@ public: // This preview window is expensive, so make it small by default. if (window_width_ == 0 || window_height_ == 0) window_width_ = 512, window_height_ = 384; + + frameCounter_ = 0; // As a hint, reserve twice the binned width for our widest current camera (V3) tmp_stripe_.reserve(4608); thread_ = std::thread(&QtPreview::threadFunc, this, options); @@ -83,100 +116,105 @@ public: void SetInfoText(const std::string &text) override { main_window_->setWindowTitle(QString::fromStdString(text)); } virtual void Show(int fd, libcamera::Span span, StreamInfo const &info) override { - if((frame_counter_++) & 1) { - // Return the buffer to the camera system. - done_callback_(fd); - return; + MyWidget::ImageBuffer *buffer = nullptr; + { + std::lock_guard lock(pane_->buffers_free_mutex_); + if(pane_->freeBuffers_.size() > 0) { + buffer = pane_->freeBuffers_.front(); + pane_->freeBuffers_.pop_front(); + } + } + + if(!buffer) { + std::lock_guard lock(pane_->buffers_available_mutex_); + if(!pane_->availableBuffers_.size()) { + done_callback_(fd); + return; + } + buffer = pane_->availableBuffers_.front(); + pane_->availableBuffers_.pop_front(); } - // Quick and simple nearest-neighbour-ish resampling is used here. // We further share U,V samples between adjacent output pixel pairs // (even when downscaling) to speed up the conversion. unsigned x_step = (info.width << 16) / window_width_; unsigned y_step = (info.height << 16) / window_height_; - // Choose the right matrix to convert YUV back to RGB. - static const float YUV2RGB[3][9] = { - { 1.0, 0.0, 1.402, 1.0, -0.344, -0.714, 1.0, 1.772, 0.0 }, // JPEG - { 1.164, 0.0, 1.596, 1.164, -0.392, -0.813, 1.164, 2.017, 0.0 }, // SMPTE170M - { 1.164, 0.0, 1.793, 1.164, -0.213, -0.533, 1.164, 2.112, 0.0 }, // Rec709 + static const uint32_t YUV2RGB[3][9] = { + { 128, 0, 179, 128, 44, 91, 128, 227, 0 }, // JPEG + { 149, 0, 204, 149, 50, 104, 149, 258, 0 }, // SMPTE170M + { 149, 0, 230, 149, 27, 68, 149, 270, 0 }, // Rec709 }; - int offsetY; - float coeffY, coeffVR, coeffUG, coeffVG, coeffUB; - if (info.colour_space == libcamera::ColorSpace::Smpte170m) + static const int RGBOFFSET[3][3] = { + {24960, -15232, 31104}, {28496, -17328, 35408}, {31824, -9776, 36944} + }; + + uint32_t coeffY, coeffVR, coeffUG, coeffVG, coeffUB; + const int * rgbOffset = nullptr; + if(info.colour_space == libcamera::ColorSpace::Smpte170m) { - offsetY = 16; coeffY = YUV2RGB[1][0]; coeffVR = YUV2RGB[1][2]; coeffUG = YUV2RGB[1][4]; coeffVG = YUV2RGB[1][5]; coeffUB = YUV2RGB[1][7]; - } - else if (info.colour_space == libcamera::ColorSpace::Rec709) + rgbOffset = RGBOFFSET[1]; + } else if(info.colour_space == libcamera::ColorSpace::Rec709) { - offsetY = 16; coeffY = YUV2RGB[2][0]; coeffVR = YUV2RGB[2][2]; coeffUG = YUV2RGB[2][4]; coeffVG = YUV2RGB[2][5]; coeffUB = YUV2RGB[2][7]; - } - else + rgbOffset = RGBOFFSET[2]; + } else { - offsetY = 0; coeffY = YUV2RGB[0][0]; coeffVR = YUV2RGB[0][2]; coeffUG = YUV2RGB[0][4]; coeffVG = YUV2RGB[0][5]; coeffUB = YUV2RGB[0][7]; - if (info.colour_space != libcamera::ColorSpace::Sycc) - LOG(1, "QtPreview: unexpected colour space " << libcamera::ColorSpace::toString(info.colour_space)); + rgbOffset = RGBOFFSET[0]; } - // Because the source buffer is uncached, and we want to read it a byte at a time, - // take a copy of each row used. This is a speedup provided memcpy() is vectorized. - tmp_stripe_.resize(2 * info.stride); - uint8_t const *Y_start = span.data(); - uint8_t const *UV_start = Y_start + info.height * info.stride; - uint8_t *Y_row = &tmp_stripe_[0]; - uint8_t *U_row = Y_row + info.stride; - uint8_t *V_row = U_row + (info.stride >> 1); - - // Possibly this should be locked in case a repaint is happening? In practice the risk - // is only that there might be some tearing, so I don't think we worry. We could speed - // it up by getting the ISP to supply RGB, but I'm not sure I want to handle that extra - // possibility in our main application code, so we'll put up with the slow conversion. - for (unsigned int y = 0; y < window_height_; y++) + uint8_t const * Y_start = span.data(); + uint8_t const * UV_start = Y_start + info.height * info.stride; + int src_ypos = y_step >> 1; + uint32_t Y2 = 0; + uint32_t U2 = coeffUG | (coeffUB << 16); + uint32_t V2 = coeffVR | (coeffVG << 16); + uint16_t * y2 = (uint16_t *)&Y2; + uint32_t U; + uint32_t V; + uint16_t * u2 = (uint16_t *)&U; + uint16_t * v2 = (uint16_t *)&V; + + for(unsigned int y = 0; y < window_height_; y++, src_ypos += y_step) { - unsigned row = (y * y_step) >> 16; - uint8_t *dest = pane_->image.scanLine(y); - unsigned x_pos = x_step >> 1; - - memcpy(Y_row, Y_start + row * info.stride, info.stride); - //memcpy(U_row, Y_start + ((4 * info.height + row) >> 1) * (info.stride >> 1), info.stride >> 1); - //memcpy(V_row, Y_start + ((5 * info.height + row) >> 1) * (info.stride >> 1), info.stride >> 1); - uint8_t const *cur_uv = UV_start + (row >> 1) * info.width; - for(unsigned int uv_idx = 0; uv_idx < info.width >> 1; uv_idx++, cur_uv += 2) - U_row[uv_idx] = cur_uv[0], V_row[uv_idx] = cur_uv[1]; - - for (unsigned int x = 0; x < window_width_; x += 2) + const uint8_t * src_y = Y_start + (src_ypos >> 16) * info.width; + const uint8_t * src_uv = UV_start + (src_ypos >> 17) * info.width; + uint8_t * dest = buffer->image.scanLine(y); + uint32_t x_pos = x_step >> 1; + + for(unsigned int x = 0; x < window_width_; x += 2) { - int Y0 = Y_row[x_pos >> 16]; + y2[0] = src_y[x_pos >> 16]; x_pos += x_step; - int Y1 = Y_row[x_pos >> 16]; - int U = U_row[x_pos >> 17]; - int V = V_row[x_pos >> 17]; + y2[1] = src_y[x_pos >> 16]; + U = src_uv[(x_pos >> 16) & 0xfffffffe]; + V = src_uv[(x_pos >> 16) | 1]; x_pos += x_step; - Y0 -= offsetY; - Y1 -= offsetY; - U -= 128; - V -= 128; - int R0 = coeffY * Y0 + coeffVR * V; - int G0 = coeffY * Y0 + coeffUG * U + coeffVG * V; - int B0 = coeffY * Y0 + coeffUB * U; - int R1 = coeffY * Y1 + coeffVR * V; - int G1 = coeffY * Y1 + coeffUG * U + coeffVG * V; - int B1 = coeffY * Y1 + coeffUB * U; + + Y2 *= coeffY; + U *= U2; + V *= V2; + + int R0 = ((int)y2[0] + (int)v2[0] - rgbOffset[0]) >> 7; + int G0 = ((int)y2[0] - (int)u2[0] - (int)v2[1] - rgbOffset[1]) >> 7; + int B0 = ((int)y2[0] + (int)u2[1] - rgbOffset[2]) >> 7; + int R1 = ((int)y2[1] + (int)v2[0] - rgbOffset[0]) >> 7; + int G1 = ((int)y2[1] - (int)u2[0] - (int)v2[1] - rgbOffset[1]) >> 7; + int B1 = ((int)y2[1] + (int)u2[1] - rgbOffset[2]) >> 7; *(dest++) = std::clamp(R0, 0, 255); *(dest++) = std::clamp(G0, 0, 255); *(dest++) = std::clamp(B0, 0, 255); @@ -186,6 +224,12 @@ public: } } + { + std::lock_guard lock(pane_->buffers_available_mutex_); + buffer->frameCounter = ++frameCounter_; + pane_->availableBuffers_.push_back(buffer); + } + pane_->update(); // Return the buffer to the camera system. @@ -193,7 +237,7 @@ public: } // Reset the preview window, clearing the current buffers and being ready to // show new ones. - void Reset() override {frame_counter_ = 0;} + void Reset() override {} // Check if preview window has been shut down. bool Quit() override { return main_window_->quit; } // There is no particular limit to image sizes, though large images will be very slow. @@ -228,8 +272,7 @@ private: std::mutex mutex_; std::condition_variable cond_var_; std::vector tmp_stripe_; - - unsigned int frame_counter_; + uint8_t frameCounter_; }; Preview *make_qt_preview(Options const *options) -- 2.34.1