Change Futatabi frames to be cached as textures instead of in system memory.

[nageru] / futatabi / vaapi_jpeg_decoder.cpp

diff --git a/futatabi/vaapi_jpeg_decoder.cpp b/futatabi/vaapi_jpeg_decoder.cpp
index f34654d508504434e81fbcc816e12b3a8aba95a5..758d974bfeb1fc4e0e6e9b595bef779f5f12bfa3 100644 (file)
--- a/futatabi/vaapi_jpeg_decoder.cpp
+++ b/futatabi/vaapi_jpeg_decoder.cpp
@@ -3,6 +3,7 @@
 #include "jpeg_destroyer.h"
 #include "jpeg_frame.h"
 #include "jpeglib_error_wrapper.h"
+#include "pbo_pool.h"
 #include "shared/memcpy_interleaved.h"
 
 #include <X11/Xlib.h>
@@ -22,6 +23,8 @@
 #include <va/va_drm.h>
 #include <va/va_x11.h>
 
+#define BUFFER_OFFSET(i) ((char *)nullptr + (i))
+
 using namespace std;
 
 static unique_ptr<VADisplayWithCleanup> va_dpy;
@@ -549,24 +552,38 @@ shared_ptr<Frame> decode_jpeg_vaapi(const string &jpeg)
 #else
        // Convert Y'CbCr to separate Y' and CbCr.
        frame->is_semiplanar = true;
-       frame->y.reset(new uint8_t[dinfo.image_width * dinfo.image_height]);
-       frame->cbcr.reset(new uint8_t[dinfo.image_width * dinfo.image_height]);
+
+       PBO pbo = global_pbo_pool->alloc_pbo();
+       size_t cbcr_offset = dinfo.image_width * dinfo.image_height;
+       uint8_t *y_pix = pbo.ptr;
+       uint8_t *cbcr_pix = pbo.ptr + cbcr_offset;
+
        const uint8_t *src = (const uint8_t *)mapped + resources.image.offsets[0];
        if (resources.image.pitches[0] == dinfo.image_width * 2) {
-               memcpy_interleaved(frame->cbcr.get(), frame->y.get(), src, dinfo.image_width * dinfo.image_height * 2);
+               memcpy_interleaved(cbcr_pix, y_pix, src, dinfo.image_width * dinfo.image_height * 2);
        } else {
                for (unsigned y = 0; y < dinfo.image_height; ++y) {
-                       memcpy_interleaved(frame->cbcr.get() + y * dinfo.image_width, frame->y.get() + y * dinfo.image_width,
+                       memcpy_interleaved(cbcr_pix + y * dinfo.image_width, y_pix + y * dinfo.image_width,
                                           src + y * resources.image.pitches[0], dinfo.image_width * 2);
                }
        }
+
+       glBindBuffer(GL_PIXEL_UNPACK_BUFFER, pbo.pbo);
+       frame->y = create_texture_2d(dinfo.image_width, dinfo.image_height, GL_R8, GL_RED, GL_UNSIGNED_BYTE, BUFFER_OFFSET(0));
+       frame->cbcr = create_texture_2d(dinfo.image_width / 2, dinfo.image_height, GL_RG8, GL_RG, GL_UNSIGNED_BYTE, BUFFER_OFFSET(cbcr_offset));
+       glBindBuffer(GL_PIXEL_UNPACK_BUFFER, 0);
+
+       glFlushMappedNamedBufferRange(pbo.pbo, 0, dinfo.image_width * dinfo.image_height * 2);
+       glMemoryBarrier(GL_PIXEL_BUFFER_BARRIER_BIT);
+       pbo.upload_done = RefCountedGLsync(GL_SYNC_GPU_COMMANDS_COMPLETE, /*flags=*/0);
+       frame->uploaded_ui_thread = pbo.upload_done;
+       frame->uploaded_interpolation = pbo.upload_done;
+       global_pbo_pool->release_pbo(move(pbo));
 #endif
        frame->width = dinfo.image_width;
        frame->height = dinfo.image_height;
        frame->chroma_subsampling_x = 2;
        frame->chroma_subsampling_y = 1;
-       frame->pitch_y = dinfo.image_width;
-       frame->pitch_chroma = dinfo.image_width / 2;
 
        if (dinfo.marker_list != nullptr &&
            dinfo.marker_list->marker == JPEG_APP0 + 1 &&