diff options
Diffstat (limited to 'indra/llimage')
| -rw-r--r-- | indra/llimage/llimage.cpp | 51 | ||||
| -rw-r--r-- | indra/llimage/llimage.h | 6 | ||||
| -rw-r--r-- | indra/llimage/llimagej2c.cpp | 2 | ||||
| -rw-r--r-- | indra/llimage/llimageworker.cpp | 122 | ||||
| -rw-r--r-- | indra/llimage/llimageworker.h | 65 | ||||
| -rw-r--r-- | indra/llimage/tests/llimageworker_test.cpp | 99 |
6 files changed, 136 insertions, 209 deletions
diff --git a/indra/llimage/llimage.cpp b/indra/llimage/llimage.cpp index 0fa027c9c3..031471d1fe 100644 --- a/indra/llimage/llimage.cpp +++ b/indra/llimage/llimage.cpp @@ -798,7 +798,6 @@ U8* LLImageBase::allocateDataSize(S32 width, S32 height, S32 ncomponents, S32 si // LLImageRaw //--------------------------------------------------------------------------- -S32 LLImageRaw::sGlobalRawMemory = 0; S32 LLImageRaw::sRawImageCount = 0; LLImageRaw::LLImageRaw() @@ -815,6 +814,15 @@ LLImageRaw::LLImageRaw(U16 width, U16 height, S8 components) ++sRawImageCount; } +LLImageRaw::LLImageRaw(const U8* data, U16 width, U16 height, S8 components) + : LLImageBase() +{ + if (allocateDataSize(width, height, components)) + { + memcpy(getData(), data, width * height * components); + } +} + LLImageRaw::LLImageRaw(U8 *data, U16 width, U16 height, S8 components, bool no_copy) : LLImageBase() { @@ -847,16 +855,13 @@ LLImageRaw::~LLImageRaw() U8* LLImageRaw::allocateData(S32 size) { U8* res = LLImageBase::allocateData(size); - sGlobalRawMemory += getDataSize(); return res; } // virtual U8* LLImageRaw::reallocateData(S32 size) { - sGlobalRawMemory -= getDataSize(); U8* res = LLImageBase::reallocateData(size); - sGlobalRawMemory += getDataSize(); return res; } @@ -869,7 +874,6 @@ void LLImageRaw::releaseData() // virtual void LLImageRaw::deleteData() { - sGlobalRawMemory -= getDataSize(); LLImageBase::deleteData(); } @@ -985,6 +989,43 @@ void LLImageRaw::verticalFlip() } +bool LLImageRaw::optimizeAwayAlpha() +{ + if (getComponents() == 4) + { + U8* data = getData(); + U32 pixels = getWidth() * getHeight(); + + // check alpha channel for all 255 + for (U32 i = 0; i < pixels; ++i) + { + if (data[i * 4 + 3] != 255) + { + return false; + } + } + + // alpha channel is all 255, make a new copy of data without alpha channel + U8* new_data = (U8*) ll_aligned_malloc_16(getWidth() * getHeight() * 3); + + for (U32 i = 0; i < pixels; ++i) + { + U32 di = i * 3; + U32 si = i * 4; + for (U32 j = 0; j < 3; ++j) + { + new_data[di+j] = data[si+j]; + } + } + + setDataAndSize(new_data, getWidth(), getHeight(), 3); + + return true; + } + + return false; +} + void LLImageRaw::expandToPowerOfTwo(S32 max_dim, bool scale_image) { // Find new sizes diff --git a/indra/llimage/llimage.h b/indra/llimage/llimage.h index 7a588cfb03..8f9e1b3c54 100644 --- a/indra/llimage/llimage.h +++ b/indra/llimage/llimage.h @@ -184,6 +184,7 @@ protected: public: LLImageRaw(); LLImageRaw(U16 width, U16 height, S8 components); + LLImageRaw(const U8* data, U16 width, U16 height, S8 components); LLImageRaw(U8 *data, U16 width, U16 height, S8 components, bool no_copy = false); // Construct using createFromFile (used by tools) //LLImageRaw(const std::string& filename, bool j2c_lowest_mip_only = false); @@ -207,6 +208,10 @@ public: void clear(U8 r=0, U8 g=0, U8 b=0, U8 a=255); void verticalFlip(); + + // if the alpha channel is all 100% opaque, delete it + // returns true if alpha channel was deleted + bool optimizeAwayAlpha(); static S32 biasedDimToPowerOfTwo(S32 curr_dim, S32 max_dim = MAX_IMAGE_SIZE); static S32 expandDimToPowerOfTwo(S32 curr_dim, S32 max_dim = MAX_IMAGE_SIZE); @@ -275,7 +280,6 @@ protected: void setDataAndSize(U8 *data, S32 width, S32 height, S8 components) ; public: - static S32 sGlobalRawMemory; static S32 sRawImageCount; private: diff --git a/indra/llimage/llimagej2c.cpp b/indra/llimage/llimagej2c.cpp index e1809dbe59..8dba1641a6 100644 --- a/indra/llimage/llimagej2c.cpp +++ b/indra/llimage/llimagej2c.cpp @@ -146,6 +146,7 @@ bool LLImageJ2C::initEncode(LLImageRaw &raw_image, int blocks_size, int precinct bool LLImageJ2C::decode(LLImageRaw *raw_imagep, F32 decode_time) { + LL_PROFILE_ZONE_SCOPED_CATEGORY_TEXTURE; return decodeChannels(raw_imagep, decode_time, 0, 4); } @@ -153,6 +154,7 @@ bool LLImageJ2C::decode(LLImageRaw *raw_imagep, F32 decode_time) // Returns true to mean done, whether successful or not. bool LLImageJ2C::decodeChannels(LLImageRaw *raw_imagep, F32 decode_time, S32 first_channel, S32 max_channel_count ) { + LL_PROFILE_ZONE_SCOPED_CATEGORY_TEXTURE; LLTimer elapsed; bool res = true; diff --git a/indra/llimage/llimageworker.cpp b/indra/llimage/llimageworker.cpp index 0dbb744bcf..0093958e6d 100644 --- a/indra/llimage/llimageworker.cpp +++ b/indra/llimage/llimageworker.cpp @@ -28,64 +28,88 @@ #include "llimageworker.h" #include "llimagedxt.h" +#include "threadpool.h" + +/*--------------------------------------------------------------------------*/ +class ImageRequest +{ +public: + ImageRequest(const LLPointer<LLImageFormatted>& image, + S32 discard, BOOL needs_aux, + const LLPointer<LLImageDecodeThread::Responder>& responder); + virtual ~ImageRequest(); + + /*virtual*/ bool processRequest(); + /*virtual*/ void finishRequest(bool completed); + +private: + // LLPointers stored in ImageRequest MUST be LLPointer instances rather + // than references: we need to increment the refcount when storing these. + // input + LLPointer<LLImageFormatted> mFormattedImage; + S32 mDiscardLevel; + BOOL mNeedsAux; + // output + LLPointer<LLImageRaw> mDecodedImageRaw; + LLPointer<LLImageRaw> mDecodedImageAux; + BOOL mDecodedRaw; + BOOL mDecodedAux; + LLPointer<LLImageDecodeThread::Responder> mResponder; +}; + //---------------------------------------------------------------------------- // MAIN THREAD -LLImageDecodeThread::LLImageDecodeThread(bool threaded) - : LLQueuedThread("imagedecode", threaded) +LLImageDecodeThread::LLImageDecodeThread(bool /*threaded*/) { - mCreationMutex = new LLMutex(); + mThreadPool.reset(new LL::ThreadPool("ImageDecode", 8)); + mThreadPool->start(); } //virtual LLImageDecodeThread::~LLImageDecodeThread() -{ - delete mCreationMutex ; -} +{} // MAIN THREAD // virtual S32 LLImageDecodeThread::update(F32 max_time_ms) { LL_PROFILE_ZONE_SCOPED_CATEGORY_TEXTURE; - LLMutexLock lock(mCreationMutex); - for (creation_list_t::iterator iter = mCreationList.begin(); - iter != mCreationList.end(); ++iter) - { - creation_info& info = *iter; - ImageRequest* req = new ImageRequest(info.handle, info.image, - info.priority, info.discard, info.needs_aux, - info.responder); + return getPending(); +} - bool res = addRequest(req); - if (!res) - { - LL_ERRS() << "request added after LLLFSThread::cleanupClass()" << LL_ENDL; - } - } - mCreationList.clear(); - S32 res = LLQueuedThread::update(max_time_ms); - return res; +S32 LLImageDecodeThread::getPending() +{ + return mThreadPool->getQueue().size(); } -LLImageDecodeThread::handle_t LLImageDecodeThread::decodeImage(LLImageFormatted* image, - U32 priority, S32 discard, BOOL needs_aux, Responder* responder) +LLImageDecodeThread::handle_t LLImageDecodeThread::decodeImage( + const LLPointer<LLImageFormatted>& image, + S32 discard, + BOOL needs_aux, + const LLPointer<LLImageDecodeThread::Responder>& responder) { LL_PROFILE_ZONE_SCOPED_CATEGORY_TEXTURE; - LLMutexLock lock(mCreationMutex); - handle_t handle = generateHandle(); - mCreationList.push_back(creation_info(handle, image, priority, discard, needs_aux, responder)); - return handle; + + // Instantiate the ImageRequest right in the lambda, why not? + mThreadPool->getQueue().post( + [req = ImageRequest(image, discard, needs_aux, responder)] + () mutable + { + auto done = req.processRequest(); + req.finishRequest(done); + }); + + // It's important to our consumer (LLTextureFetchWorker) that we return a + // nonzero handle. It is NOT important that the nonzero handle be unique: + // nothing is ever done with it except to compare it to zero, or zero it. + return 17; } -// Used by unit test only -// Returns the size of the mutex guarded list as an indication of sanity -S32 LLImageDecodeThread::tut_size() +void LLImageDecodeThread::shutdown() { - LLMutexLock lock(mCreationMutex); - S32 res = mCreationList.size(); - return res; + mThreadPool->close(); } LLImageDecodeThread::Responder::~Responder() @@ -94,11 +118,10 @@ LLImageDecodeThread::Responder::~Responder() //---------------------------------------------------------------------------- -LLImageDecodeThread::ImageRequest::ImageRequest(handle_t handle, LLImageFormatted* image, - U32 priority, S32 discard, BOOL needs_aux, - LLImageDecodeThread::Responder* responder) - : LLQueuedThread::QueuedRequest(handle, priority, FLAG_AUTO_COMPLETE), - mFormattedImage(image), +ImageRequest::ImageRequest(const LLPointer<LLImageFormatted>& image, + S32 discard, BOOL needs_aux, + const LLPointer<LLImageDecodeThread::Responder>& responder) + : mFormattedImage(image), mDiscardLevel(discard), mNeedsAux(needs_aux), mDecodedRaw(FALSE), @@ -107,7 +130,7 @@ LLImageDecodeThread::ImageRequest::ImageRequest(handle_t handle, LLImageFormatte { } -LLImageDecodeThread::ImageRequest::~ImageRequest() +ImageRequest::~ImageRequest() { mDecodedImageRaw = NULL; mDecodedImageAux = NULL; @@ -118,10 +141,10 @@ LLImageDecodeThread::ImageRequest::~ImageRequest() // Returns true when done, whether or not decode was successful. -bool LLImageDecodeThread::ImageRequest::processRequest() +bool ImageRequest::processRequest() { LL_PROFILE_ZONE_SCOPED_CATEGORY_TEXTURE; - const F32 decode_time_slice = .1f; + const F32 decode_time_slice = 0.f; //disable time slicing bool done = true; if (!mDecodedRaw && mFormattedImage.notNull()) { @@ -145,7 +168,7 @@ bool LLImageDecodeThread::ImageRequest::processRequest() mFormattedImage->getHeight(), mFormattedImage->getComponents()); } - done = mFormattedImage->decode(mDecodedImageRaw, decode_time_slice); // 1ms + done = mFormattedImage->decode(mDecodedImageRaw, decode_time_slice); // some decoders are removing data when task is complete and there were errors mDecodedRaw = done && mDecodedImageRaw->getData(); } @@ -158,14 +181,14 @@ bool LLImageDecodeThread::ImageRequest::processRequest() mFormattedImage->getHeight(), 1); } - done = mFormattedImage->decodeChannels(mDecodedImageAux, decode_time_slice, 4, 4); // 1ms + done = mFormattedImage->decodeChannels(mDecodedImageAux, decode_time_slice, 4, 4); mDecodedAux = done && mDecodedImageAux->getData(); } return done; } -void LLImageDecodeThread::ImageRequest::finishRequest(bool completed) +void ImageRequest::finishRequest(bool completed) { LL_PROFILE_ZONE_SCOPED_CATEGORY_TEXTURE; if (mResponder.notNull()) @@ -175,10 +198,3 @@ void LLImageDecodeThread::ImageRequest::finishRequest(bool completed) } // Will automatically be deleted } - -// Used by unit test only -// Checks that a responder exists for this instance so that something can happen when completion is reached -bool LLImageDecodeThread::ImageRequest::tut_isOK() -{ - return mResponder.notNull(); -} diff --git a/indra/llimage/llimageworker.h b/indra/llimage/llimageworker.h index 1bfb0ddfd3..fb35108a7a 100644 --- a/indra/llimage/llimageworker.h +++ b/indra/llimage/llimageworker.h @@ -29,9 +29,9 @@ #include "llimage.h" #include "llpointer.h" -#include "llworkerthread.h" +#include "threadpool_fwd.h" -class LLImageDecodeThread : public LLQueuedThread +class LLImageDecodeThread { public: class Responder : public LLThreadSafeRefCount @@ -42,63 +42,24 @@ public: virtual void completed(bool success, LLImageRaw* raw, LLImageRaw* aux) = 0; }; - class ImageRequest : public LLQueuedThread::QueuedRequest - { - protected: - virtual ~ImageRequest(); // use deleteRequest() - - public: - ImageRequest(handle_t handle, LLImageFormatted* image, - U32 priority, S32 discard, BOOL needs_aux, - LLImageDecodeThread::Responder* responder); - - /*virtual*/ bool processRequest(); - /*virtual*/ void finishRequest(bool completed); - - // Used by unit tests to check the consitency of the request instance - bool tut_isOK(); - - private: - // input - LLPointer<LLImageFormatted> mFormattedImage; - S32 mDiscardLevel; - BOOL mNeedsAux; - // output - LLPointer<LLImageRaw> mDecodedImageRaw; - LLPointer<LLImageRaw> mDecodedImageAux; - BOOL mDecodedRaw; - BOOL mDecodedAux; - LLPointer<LLImageDecodeThread::Responder> mResponder; - }; - public: LLImageDecodeThread(bool threaded = true); virtual ~LLImageDecodeThread(); - handle_t decodeImage(LLImageFormatted* image, - U32 priority, S32 discard, BOOL needs_aux, - Responder* responder); + // meant to resemble LLQueuedThread::handle_t + typedef U32 handle_t; + handle_t decodeImage(const LLPointer<LLImageFormatted>& image, + S32 discard, BOOL needs_aux, + const LLPointer<Responder>& responder); + S32 getPending(); S32 update(F32 max_time_ms); + void shutdown(); - // Used by unit tests to check the consistency of the thread instance - S32 tut_size(); - private: - struct creation_info - { - handle_t handle; - LLPointer<LLImageFormatted> image; - U32 priority; - S32 discard; - BOOL needs_aux; - LLPointer<Responder> responder; - creation_info(handle_t h, LLImageFormatted* i, U32 p, S32 d, BOOL aux, Responder* r) - : handle(h), image(i), priority(p), discard(d), needs_aux(aux), responder(r) - {} - }; - typedef std::list<creation_info> creation_list_t; - creation_list_t mCreationList; - LLMutex* mCreationMutex; + // As of SL-17483, LLImageDecodeThread is no longer itself an + // LLQueuedThread - instead this is the API by which we submit work to the + // "ImageDecode" ThreadPool. + std::unique_ptr<LL::ThreadPool> mThreadPool; }; #endif diff --git a/indra/llimage/tests/llimageworker_test.cpp b/indra/llimage/tests/llimageworker_test.cpp index 9011ac615c..0a97b739b0 100644 --- a/indra/llimage/tests/llimageworker_test.cpp +++ b/indra/llimage/tests/llimageworker_test.cpp @@ -125,42 +125,11 @@ namespace tut } }; - // Test wrapper declaration : image worker - // Note: this class is not meant to be instantiated outside an LLImageDecodeThread instance - // but it's not a bad idea to get its public API a good shake as part of a thorough unit test set. - // Some gotcha with the destructor though (see below). - struct imagerequest_test - { - // Instance to be tested - LLImageDecodeThread::ImageRequest* mRequest; - bool done; - - // Constructor and destructor of the test wrapper - imagerequest_test() - { - done = false; - - mRequest = new LLImageDecodeThread::ImageRequest(0, 0, - LLQueuedThread::PRIORITY_NORMAL, 0, FALSE, - new responder_test(&done)); - } - ~imagerequest_test() - { - // We should delete the object *but*, because its destructor is protected, that cannot be - // done from outside an LLImageDecodeThread instance... So we leak memory here... It's fine... - //delete mRequest; - } - }; - // Tut templating thingamagic: test group, object and test instance typedef test_group<imagedecodethread_test> imagedecodethread_t; typedef imagedecodethread_t::object imagedecodethread_object_t; tut::imagedecodethread_t tut_imagedecodethread("LLImageDecodeThread"); - typedef test_group<imagerequest_test> imagerequest_t; - typedef imagerequest_t::object imagerequest_object_t; - tut::imagerequest_t tut_imagerequest("LLImageRequest"); - // --------------------------------------------------------------------------------------- // Test functions // Notes: @@ -172,64 +141,18 @@ namespace tut // --------------------------------------------------------------------------------------- // Test the LLImageDecodeThread interface // --------------------------------------------------------------------------------------- - // - // Note on Unit Testing Queued Thread Classes - // - // Since methods on such a class are called on a separate loop and that we can't insert tut - // ensure() calls in there, we exercise the class with 2 sets of tests: - // - 1: Test as a single threaded instance: We declare the class but ask for no thread - // to be spawned (easy with LLThreads since there's a boolean argument on the constructor - // just for that). We can then unit test each public method like we do on a normal class. - // - 2: Test as a threaded instance: We let the thread launch and check that its external - // behavior is as expected (i.e. it runs, can accept a work order and processes - // it). Typically though there's no guarantee that this exercises all the methods of the - // class which is why we also need the previous "non threaded" set of unit tests for - // complete coverage. - // - // --------------------------------------------------------------------------------------- template<> template<> void imagedecodethread_object_t::test<1>() { - // Test a *non threaded* instance of the class - mThread = new LLImageDecodeThread(false); - ensure("LLImageDecodeThread: non threaded constructor failed", mThread != NULL); - // Test that we start with an empty list right at creation - ensure("LLImageDecodeThread: non threaded init state incorrect", mThread->tut_size() == 0); - // Insert something in the queue - bool done = false; - LLImageDecodeThread::handle_t decodeHandle = mThread->decodeImage(NULL, LLQueuedThread::PRIORITY_NORMAL, 0, FALSE, new responder_test(&done)); - // Verifies we got a valid handle - ensure("LLImageDecodeThread: non threaded decodeImage(), returned handle is null", decodeHandle != 0); - // Verifies that we do now have something in the queued list - ensure("LLImageDecodeThread: non threaded decodeImage() insertion in threaded list failed", mThread->tut_size() == 1); - // Trigger queue handling "manually" (on a threaded instance, this is done on the thread loop) - S32 res = mThread->update(0); - // Verifies that we successfully handled the list - ensure("LLImageDecodeThread: non threaded update() list handling test failed", res == 0); - // Verifies that the list is now empty - ensure("LLImageDecodeThread: non threaded update() list emptying test failed", mThread->tut_size() == 0); - } - - template<> template<> - void imagedecodethread_object_t::test<2>() - { // Test a *threaded* instance of the class mThread = new LLImageDecodeThread(true); ensure("LLImageDecodeThread: threaded constructor failed", mThread != NULL); - // Test that we start with an empty list right at creation - ensure("LLImageDecodeThread: threaded init state incorrect", mThread->tut_size() == 0); // Insert something in the queue bool done = false; - LLImageDecodeThread::handle_t decodeHandle = mThread->decodeImage(NULL, LLQueuedThread::PRIORITY_NORMAL, 0, FALSE, new responder_test(&done)); + LLImageDecodeThread::handle_t decodeHandle = mThread->decodeImage(NULL, 0, FALSE, new responder_test(&done)); // Verifies we get back a valid handle ensure("LLImageDecodeThread: threaded decodeImage(), returned handle is null", decodeHandle != 0); - // Wait a little so to simulate the main thread doing something on its main loop... - ms_sleep(500); // 500 milliseconds - // Verifies that the responder has *not* been called yet in the meantime - ensure("LLImageDecodeThread: responder creation failed", done == false); - // Ask the thread to update: that means tells the queue to check itself and creates work requests - mThread->update(1); // Wait till the thread has time to handle the work order (though it doesn't do much per work order...) const U32 INCREMENT_TIME = 500; // 500 milliseconds const U32 MAX_TIME = 20 * INCREMENT_TIME; // Do the loop 20 times max, i.e. wait 10 seconds but no more @@ -242,24 +165,4 @@ namespace tut // Verifies that the responder has now been called ensure("LLImageDecodeThread: threaded work unit not processed", done == true); } - - // --------------------------------------------------------------------------------------- - // Test the LLImageDecodeThread::ImageRequest interface - // --------------------------------------------------------------------------------------- - - template<> template<> - void imagerequest_object_t::test<1>() - { - // Test that we start with a correct request at creation - ensure("LLImageDecodeThread::ImageRequest::ImageRequest() constructor test failed", mRequest->tut_isOK()); - bool res = mRequest->processRequest(); - // Verifies that we processed the request successfully - ensure("LLImageDecodeThread::ImageRequest::processRequest() processing request test failed", res == true); - // Check that we can call the finishing call safely - try { - mRequest->finishRequest(false); - } catch (...) { - fail("LLImageDecodeThread::ImageRequest::finishRequest() test failed"); - } - } } |