X-Git-Url: https://git.sesse.net/?a=blobdiff_plain;f=futatabi%2Fvaapi_jpeg_decoder.cpp;h=0540db7da82061ff314ad26bdba2c76cb2cdca5a;hb=0c7201c2d136870ea8c5fe205bee21207369312c;hp=f34654d508504434e81fbcc816e12b3a8aba95a5;hpb=32b87c91cf51d730ff5abc8347884219918fad66;p=nageru diff --git a/futatabi/vaapi_jpeg_decoder.cpp b/futatabi/vaapi_jpeg_decoder.cpp index f34654d..0540db7 100644 --- a/futatabi/vaapi_jpeg_decoder.cpp +++ b/futatabi/vaapi_jpeg_decoder.cpp @@ -3,7 +3,9 @@ #include "jpeg_destroyer.h" #include "jpeg_frame.h" #include "jpeglib_error_wrapper.h" +#include "pbo_pool.h" #include "shared/memcpy_interleaved.h" +#include "shared/va_display.h" #include #include @@ -22,6 +24,8 @@ #include #include +#define BUFFER_OFFSET(i) ((char *)nullptr + (i)) + using namespace std; static unique_ptr va_dpy; @@ -135,90 +139,12 @@ private: bool committed = false; }; -VADisplayWithCleanup::~VADisplayWithCleanup() -{ - if (va_dpy != nullptr) { - vaTerminate(va_dpy); - } - if (x11_display != nullptr) { - XCloseDisplay(x11_display); - } - if (drm_fd != -1) { - close(drm_fd); - } -} - -unique_ptr va_open_display(const string &va_display) +static unique_ptr try_open_va_mjpeg(const string &va_display) { - if (va_display.empty() || va_display[0] != '/') { // An X display. - Display *x11_display = XOpenDisplay(va_display.empty() ? nullptr : va_display.c_str()); - if (x11_display == nullptr) { - fprintf(stderr, "error: can't connect to X server!\n"); - return nullptr; - } - - unique_ptr ret(new VADisplayWithCleanup); - ret->x11_display = x11_display; - ret->va_dpy = vaGetDisplay(x11_display); - if (ret->va_dpy == nullptr) { - return nullptr; - } - return ret; - } else { // A DRM node on the filesystem (e.g. /dev/dri/renderD128). - int drm_fd = open(va_display.c_str(), O_RDWR); - if (drm_fd == -1) { - perror(va_display.c_str()); - return nullptr; - } - unique_ptr ret(new VADisplayWithCleanup); - ret->drm_fd = drm_fd; - ret->va_dpy = vaGetDisplayDRM(drm_fd); - if (ret->va_dpy == nullptr) { - return nullptr; - } - return ret; - } -} - -unique_ptr try_open_va(const string &va_display, string *error) -{ - unique_ptr va_dpy = va_open_display(va_display); - if (va_dpy == nullptr) { - if (error) - *error = "Opening VA display failed"; - return nullptr; - } - int major_ver, minor_ver; - VAStatus va_status = vaInitialize(va_dpy->va_dpy, &major_ver, &minor_ver); - if (va_status != VA_STATUS_SUCCESS) { - char buf[256]; - snprintf(buf, sizeof(buf), "vaInitialize() failed with status %d\n", va_status); - if (error != nullptr) - *error = buf; - return nullptr; - } - - int num_entrypoints = vaMaxNumEntrypoints(va_dpy->va_dpy); - unique_ptr entrypoints(new VAEntrypoint[num_entrypoints]); - if (entrypoints == nullptr) { - if (error != nullptr) - *error = "Failed to allocate memory for VA entry points"; - return nullptr; - } - - vaQueryConfigEntrypoints(va_dpy->va_dpy, VAProfileJPEGBaseline, entrypoints.get(), &num_entrypoints); - for (int slice_entrypoint = 0; slice_entrypoint < num_entrypoints; slice_entrypoint++) { - if (entrypoints[slice_entrypoint] != VAEntrypointVLD) { - continue; - } - - // We found a usable decode, so return it. - return va_dpy; - } - - if (error != nullptr) - *error = "Can't find VAEntrypointVLD for the JPEG profile"; - return nullptr; + // Seemingly VA_FOURCC_422H is no good for vaGetImage(). :-/ + return try_open_va(va_display, { VAProfileJPEGBaseline }, VAEntrypointVLD, + { { "4:2:2", VA_RT_FORMAT_YUV422, VA_FOURCC_UYVY, &config_id, &uyvy_format } }, + /*chosen_profile=*/nullptr, /*error=*/nullptr); } string get_usable_va_display() @@ -232,7 +158,7 @@ string get_usable_va_display() } // First try the default (ie., whatever $DISPLAY is set to). - unique_ptr va_dpy = try_open_va("", nullptr); + unique_ptr va_dpy = try_open_va_mjpeg(""); if (va_dpy != nullptr) { if (need_env_reset) { unsetenv("LIBVA_MESSAGING_LEVEL"); @@ -250,7 +176,7 @@ string get_usable_va_display() } else { for (size_t i = 0; i < g.gl_pathc; ++i) { string path = g.gl_pathv[i]; - va_dpy = try_open_va(path, nullptr); + va_dpy = try_open_va_mjpeg(path); if (va_dpy != nullptr) { fprintf(stderr, "Autodetected %s as a suitable replacement; using it.\n", path.c_str()); @@ -277,37 +203,11 @@ void init_jpeg_vaapi() return; } - va_dpy = try_open_va(dpy, nullptr); + va_dpy = try_open_va_mjpeg(dpy); if (va_dpy == nullptr) { return; } - VAConfigAttrib attr = { VAConfigAttribRTFormat, VA_RT_FORMAT_YUV422 }; - - VAStatus va_status = vaCreateConfig(va_dpy->va_dpy, VAProfileJPEGBaseline, VAEntrypointVLD, - &attr, 1, &config_id); - CHECK_VASTATUS(va_status, "vaCreateConfig"); - - int num_formats = vaMaxNumImageFormats(va_dpy->va_dpy); - assert(num_formats > 0); - - unique_ptr formats(new VAImageFormat[num_formats]); - va_status = vaQueryImageFormats(va_dpy->va_dpy, formats.get(), &num_formats); - CHECK_VASTATUS(va_status, "vaQueryImageFormats"); - - bool found = false; - for (int i = 0; i < num_formats; ++i) { - // Seemingly VA_FOURCC_422H is no good for vaGetImage(). :-/ - if (formats[i].fourcc == VA_FOURCC_UYVY) { - memcpy(&uyvy_format, &formats[i], sizeof(VAImageFormat)); - found = true; - break; - } - } - if (!found) { - return; - } - fprintf(stderr, "VA-API JPEG decoding initialized.\n"); vaapi_jpeg_decoding_usable = true; } @@ -380,8 +280,11 @@ shared_ptr decode_jpeg_vaapi(const string &jpeg) pic_param.color_space = 0; // YUV. pic_param.rotation = VA_ROTATION_NONE; + VAResources resources = get_va_resources(dinfo.image_width, dinfo.image_height); + ReleaseVAResources release(resources); + VABufferID pic_param_buffer; - VAStatus va_status = vaCreateBuffer(va_dpy->va_dpy, config_id, VAPictureParameterBufferType, sizeof(pic_param), 1, &pic_param, &pic_param_buffer); + VAStatus va_status = vaCreateBuffer(va_dpy->va_dpy, resources.context, VAPictureParameterBufferType, sizeof(pic_param), 1, &pic_param, &pic_param_buffer); CHECK_VASTATUS_RET(va_status, "vaCreateBuffer"); VABufferDestroyer destroy_pic_param(va_dpy->va_dpy, pic_param_buffer); @@ -406,7 +309,7 @@ shared_ptr decode_jpeg_vaapi(const string &jpeg) } VABufferID iq_buffer; - va_status = vaCreateBuffer(va_dpy->va_dpy, config_id, VAIQMatrixBufferType, sizeof(iq), 1, &iq, &iq_buffer); + va_status = vaCreateBuffer(va_dpy->va_dpy, resources.context, VAIQMatrixBufferType, sizeof(iq), 1, &iq, &iq_buffer); CHECK_VASTATUS_RET(va_status, "vaCreateBuffer"); VABufferDestroyer destroy_iq(va_dpy->va_dpy, iq_buffer); @@ -440,7 +343,7 @@ shared_ptr decode_jpeg_vaapi(const string &jpeg) } VABufferID huff_buffer; - va_status = vaCreateBuffer(va_dpy->va_dpy, config_id, VAHuffmanTableBufferType, sizeof(huff), 1, &huff, &huff_buffer); + va_status = vaCreateBuffer(va_dpy->va_dpy, resources.context, VAHuffmanTableBufferType, sizeof(huff), 1, &huff, &huff_buffer); CHECK_VASTATUS_RET(va_status, "vaCreateBuffer"); VABufferDestroyer destroy_huff(va_dpy->va_dpy, huff_buffer); @@ -470,19 +373,16 @@ shared_ptr decode_jpeg_vaapi(const string &jpeg) parms.num_mcus = horiz_mcus * vert_mcus; VABufferID slice_param_buffer; - va_status = vaCreateBuffer(va_dpy->va_dpy, config_id, VASliceParameterBufferType, sizeof(parms), 1, &parms, &slice_param_buffer); + va_status = vaCreateBuffer(va_dpy->va_dpy, resources.context, VASliceParameterBufferType, sizeof(parms), 1, &parms, &slice_param_buffer); CHECK_VASTATUS_RET(va_status, "vaCreateBuffer"); VABufferDestroyer destroy_slice_param(va_dpy->va_dpy, slice_param_buffer); // The actual data. VA-API will destuff and all for us. VABufferID data_buffer; - va_status = vaCreateBuffer(va_dpy->va_dpy, config_id, VASliceDataBufferType, dinfo.src->bytes_in_buffer, 1, const_cast(dinfo.src->next_input_byte), &data_buffer); + va_status = vaCreateBuffer(va_dpy->va_dpy, resources.context, VASliceDataBufferType, dinfo.src->bytes_in_buffer, 1, const_cast(dinfo.src->next_input_byte), &data_buffer); CHECK_VASTATUS_RET(va_status, "vaCreateBuffer"); VABufferDestroyer destroy_data(va_dpy->va_dpy, data_buffer); - VAResources resources = get_va_resources(dinfo.image_width, dinfo.image_height); - ReleaseVAResources release(resources); - va_status = vaBeginPicture(va_dpy->va_dpy, resources.context, resources.surface); CHECK_VASTATUS_RET(va_status, "vaBeginPicture"); va_status = vaRenderPicture(va_dpy->va_dpy, resources.context, &pic_param_buffer, 1); @@ -549,24 +449,38 @@ shared_ptr 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 &&