From 7c9aeed97d4ba3641971b9a1a92d334ec0adbb09 Mon Sep 17 00:00:00 2001 From: Nat Goodspeed Date: Fri, 1 Oct 2021 16:05:23 -0400 Subject: SL-16024: Enhance LLThreadSafeQueue for use with WorkQueue. First, parameterize LLThreadSafeQueue's queue type. This allows us to substitute (e.g.) a std::priority_queue for a particular instance. Use std::queue for the default queue type, changing the operations invoked on the queue type from std::deque methods to std::queue methods. Rename published methods from (e.g.) pushFront() and popBack() to simple push() and pop(), retaining legacy names as aliases. Not only are the overt Front and Back unnecessary; they're the opposite of how std::queue uses std::deque or std::list, so they only confuse the reader. Break out tryPushUntil() method. We already use that logic internally to tryPushFor(), so it's just as easy to publish it as its own entry point. Add tryPopFor() and tryPopUntil() to allow limiting the time we'll wait for a queue item to become available. --- indra/llcommon/llthreadsafequeue.h | 229 ++++++++++++++++++++++++++----------- 1 file changed, 163 insertions(+), 66 deletions(-) (limited to 'indra/llcommon/llthreadsafequeue.h') diff --git a/indra/llcommon/llthreadsafequeue.h b/indra/llcommon/llthreadsafequeue.h index 26e0d71d31..04f51816d7 100644 --- a/indra/llcommon/llthreadsafequeue.h +++ b/indra/llcommon/llthreadsafequeue.h @@ -1,6 +1,6 @@ /** * @file llthreadsafequeue.h - * @brief Base classes for thread, mutex and condition handling. + * @brief Queue protected with mutexes for cross-thread use * * $LicenseInfo:firstyear=2004&license=viewerlgpl$ * Second Life Viewer Source Code @@ -27,15 +27,15 @@ #ifndef LL_LLTHREADSAFEQUEUE_H #define LL_LLTHREADSAFEQUEUE_H -#include "llexception.h" -#include -#include -#include -#include "mutex.h" #include "llcoros.h" #include LLCOROS_MUTEX_HEADER #include #include LLCOROS_CONDVAR_HEADER +#include "llexception.h" +#include "mutex.h" +#include +#include +#include // // A general queue exception. @@ -66,61 +66,95 @@ public: } }; -// -// Implements a thread safe FIFO. -// -template +/** + * Implements a thread safe FIFO. + */ +// Let the default std::queue default to underlying std::deque. Override if +// desired. +template> class LLThreadSafeQueue { public: typedef ElementT value_type; - + // If the pool is set to NULL one will be allocated and managed by this // queue. LLThreadSafeQueue(U32 capacity = 1024); - - // Add an element to the front of queue (will block if the queue has + + // Add an element to the queue (will block if the queue has // reached capacity). // // This call will raise an interrupt error if the queue is closed while // the caller is blocked. - void pushFront(ElementT const & element); - - // Try to add an element to the front of queue without blocking. Returns + void push(ElementT const& element); + // legacy name + void pushFront(ElementT const & element) { return push(element); } + + // Try to add an element to the queue without blocking. Returns // true only if the element was actually added. - bool tryPushFront(ElementT const & element); + bool tryPush(ElementT const& element); + // legacy name + bool tryPushFront(ElementT const & element) { return tryPush(element); } - // Try to add an element to the front of queue, blocking if full but with - // timeout. Returns true if the element was added. + // Try to add an element to the queue, blocking if full but with timeout + // after specified duration. Returns true if the element was added. // There are potentially two different timeouts involved: how long to try // to lock the mutex, versus how long to wait for the queue to stop being // full. Careful settings for each timeout might be orders of magnitude // apart. However, this method conflates them. template + bool tryPushFor(const std::chrono::duration& timeout, + ElementT const & element); + // legacy name + template bool tryPushFrontFor(const std::chrono::duration& timeout, - ElementT const & element); + ElementT const & element) { return tryPushFor(timeout, element); } + + // Try to add an element to the queue, blocking if full but with + // timeout at specified time_point. Returns true if the element was added. + template + bool tryPushUntil(const std::chrono::time_point& timeout, + ElementT const& element); + // no legacy name because this is a newer method - // Pop the element at the end of the queue (will block if the queue is + // Pop the element at the head of the queue (will block if the queue is // empty). // // This call will raise an interrupt error if the queue is closed while // the caller is blocked. - ElementT popBack(void); - - // Pop an element from the end of the queue if there is one available. + ElementT pop(void); + // legacy name + ElementT popBack(void) { return pop(); } + + // Pop an element from the head of the queue if there is one available. // Returns true only if an element was popped. - bool tryPopBack(ElementT & element); - + bool tryPop(ElementT & element); + // legacy name + bool tryPopBack(ElementT & element) { return tryPop(element); } + + // Pop the element at the head of the queue, blocking if empty, with + // timeout after specified duration. Returns true if an element was popped. + template + bool tryPopFor(const std::chrono::duration& timeout, ElementT& element); + // no legacy name because this is a newer method + + // Pop the element at the head of the queue, blocking if empty, with + // timeout at specified time_point. Returns true if an element was popped. + template + bool tryPopUntil(const std::chrono::time_point& timeout, + ElementT& element); + // no legacy name because this is a newer method + // Returns the size of the queue. size_t size(); // closes the queue: - // - every subsequent pushFront() call will throw LLThreadSafeQueueInterrupt - // - every subsequent tryPushFront() call will return false - // - popBack() calls will return normally until the queue is drained, then - // every subsequent popBack() will throw LLThreadSafeQueueInterrupt - // - tryPopBack() calls will return normally until the queue is drained, - // then every subsequent tryPopBack() call will return false + // - every subsequent push() call will throw LLThreadSafeQueueInterrupt + // - every subsequent tryPush() call will return false + // - pop() calls will return normally until the queue is drained, then + // every subsequent pop() will throw LLThreadSafeQueueInterrupt + // - tryPop() calls will return normally until the queue is drained, + // then every subsequent tryPop() call will return false void close(); // detect closed state @@ -128,8 +162,9 @@ public: // inverse of isClosed() explicit operator bool(); -private: - std::deque< ElementT > mStorage; +protected: + typedef QueueT queue_type; + QueueT mStorage; U32 mCapacity; bool mClosed; @@ -142,16 +177,16 @@ private: // LLThreadSafeQueue //----------------------------------------------------------------------------- -template -LLThreadSafeQueue::LLThreadSafeQueue(U32 capacity) : +template +LLThreadSafeQueue::LLThreadSafeQueue(U32 capacity) : mCapacity(capacity), mClosed(false) { } -template -void LLThreadSafeQueue::pushFront(ElementT const & element) +template +void LLThreadSafeQueue::push(ElementT const & element) { lock_t lock1(mLock); while (true) @@ -163,7 +198,7 @@ void LLThreadSafeQueue::pushFront(ElementT const & element) if (mStorage.size() < mCapacity) { - mStorage.push_front(element); + mStorage.push(element); lock1.unlock(); mEmptyCond.notify_one(); return; @@ -175,15 +210,24 @@ void LLThreadSafeQueue::pushFront(ElementT const & element) } -template +template template -bool LLThreadSafeQueue::tryPushFrontFor(const std::chrono::duration& timeout, - ElementT const & element) +bool LLThreadSafeQueue::tryPushFor( + const std::chrono::duration& timeout, + ElementT const & element) { // Convert duration to time_point: passing the same timeout duration to // each of multiple calls is wrong. - auto endpoint = std::chrono::steady_clock::now() + timeout; + return tryPushUntil(std::chrono::steady_clock::now() + timeout, element); +} + +template +template +bool LLThreadSafeQueue::tryPushUntil( + const std::chrono::time_point& endpoint, + ElementT const& element) +{ lock_t lock1(mLock, std::defer_lock); if (!lock1.try_lock_until(endpoint)) return false; @@ -197,7 +241,7 @@ bool LLThreadSafeQueue::tryPushFrontFor(const std::chrono::duration::tryPushFrontFor(const std::chrono::duration -bool LLThreadSafeQueue::tryPushFront(ElementT const & element) +template +bool LLThreadSafeQueue::tryPush(ElementT const & element) { lock_t lock1(mLock, std::defer_lock); if (!lock1.try_lock()) @@ -228,23 +272,24 @@ bool LLThreadSafeQueue::tryPushFront(ElementT const & element) if (mStorage.size() >= mCapacity) return false; - mStorage.push_front(element); + mStorage.push(element); lock1.unlock(); mEmptyCond.notify_one(); return true; } -template -ElementT LLThreadSafeQueue::popBack(void) +template +ElementT LLThreadSafeQueue::pop(void) { lock_t lock1(mLock); while (true) { if (!mStorage.empty()) { - ElementT value = mStorage.back(); - mStorage.pop_back(); + // std::queue::front() is the element about to pop() + ElementT value = mStorage.front(); + mStorage.pop(); lock1.unlock(); mCapacityCond.notify_one(); return value; @@ -261,54 +306,106 @@ ElementT LLThreadSafeQueue::popBack(void) } -template -bool LLThreadSafeQueue::tryPopBack(ElementT & element) +template +bool LLThreadSafeQueue::tryPop(ElementT & element) { lock_t lock1(mLock, std::defer_lock); if (!lock1.try_lock()) return false; - // no need to check mClosed: tryPopBack() behavior when the queue is + // no need to check mClosed: tryPop() behavior when the queue is // closed is implemented by simple inability to push any new elements if (mStorage.empty()) return false; - element = mStorage.back(); - mStorage.pop_back(); + // std::queue::front() is the element about to pop() + element = mStorage.front(); + mStorage.pop(); lock1.unlock(); mCapacityCond.notify_one(); return true; } -template -size_t LLThreadSafeQueue::size(void) +template +template +bool LLThreadSafeQueue::tryPopFor( + const std::chrono::duration& timeout, + ElementT& element) +{ + // Convert duration to time_point: passing the same timeout duration to + // each of multiple calls is wrong. + return tryPopUntil(std::chrono::steady_clock::now() + timeout, element); +} + + +template +template +bool LLThreadSafeQueue::tryPopUntil( + const std::chrono::time_point& endpoint, + ElementT& element) +{ + lock_t lock1(mLock, std::defer_lock); + if (!lock1.try_lock_until(endpoint)) + return false; + + while (true) + { + if (!mStorage.empty()) + { + // std::queue::front() is the element about to pop() + element = mStorage.front(); + mStorage.pop(); + lock1.unlock(); + mCapacityCond.notify_one(); + return true; + } + + if (mClosed) + { + return false; + } + + // Storage empty. Wait for signal. + if (LLCoros::cv_status::timeout == mEmptyCond.wait_until(lock1, endpoint)) + { + // timed out -- formally we might recheck both conditions above + return false; + } + // If we didn't time out, we were notified for some reason. Loop back + // to check. + } +} + + +template +size_t LLThreadSafeQueue::size(void) { lock_t lock(mLock); return mStorage.size(); } -template -void LLThreadSafeQueue::close() +template +void LLThreadSafeQueue::close() { lock_t lock(mLock); mClosed = true; lock.unlock(); - // wake up any blocked popBack() calls + // wake up any blocked pop() calls mEmptyCond.notify_all(); - // wake up any blocked pushFront() calls + // wake up any blocked push() calls mCapacityCond.notify_all(); } -template -bool LLThreadSafeQueue::isClosed() +template +bool LLThreadSafeQueue::isClosed() { lock_t lock(mLock); return mClosed && mStorage.size() == 0; } -template -LLThreadSafeQueue::operator bool() +template +LLThreadSafeQueue::operator bool() { return ! isClosed(); } -- cgit v1.3 From ca60fbe72ce086fbdf0821043ad3be6aad06857c Mon Sep 17 00:00:00 2001 From: Nat Goodspeed Date: Mon, 4 Oct 2021 16:19:59 -0400 Subject: SL-16024: LLThreadSafeQueue enhancements Add LL::PriorityQueueAdapter, a wrapper for std::priority_queue to make its API more closely resemble std::queue for drop-in use as LLThreadSafeQueue's underlying QueueT container. Support move-only element types. Factor out some implementation redundancy: wrap actual push semantics as push_(), actual pop semantics as pop_(). push(), tryPush() and tryPushUntil() now call push_(); pop(), tryPop() and tryPopUntil() now call pop_(). Break out tryLock() and tryLockUntil() methods that, if they can lock, run the passed callable. Then tryPush(), tryPushUntil(), tryPop() and tryPopUntil() pass lambdas containing the meat of the original method body to tryLock() or tryLockUntil(), as appropriate. --- indra/llcommon/llthreadsafequeue.h | 358 ++++++++++++++++++++++++------------- 1 file changed, 235 insertions(+), 123 deletions(-) (limited to 'indra/llcommon/llthreadsafequeue.h') diff --git a/indra/llcommon/llthreadsafequeue.h b/indra/llcommon/llthreadsafequeue.h index 04f51816d7..c57520c01f 100644 --- a/indra/llcommon/llthreadsafequeue.h +++ b/indra/llcommon/llthreadsafequeue.h @@ -37,6 +37,9 @@ #include #include +/***************************************************************************** +* LLThreadSafeQueue +*****************************************************************************/ // // A general queue exception. // @@ -77,8 +80,8 @@ class LLThreadSafeQueue public: typedef ElementT value_type; - // If the pool is set to NULL one will be allocated and managed by this - // queue. + // Limiting the number of pending items prevents unbounded growth of the + // underlying queue. LLThreadSafeQueue(U32 capacity = 1024); // Add an element to the queue (will block if the queue has @@ -86,13 +89,15 @@ public: // // This call will raise an interrupt error if the queue is closed while // the caller is blocked. - void push(ElementT const& element); + template + void push(T&& element); // legacy name void pushFront(ElementT const & element) { return push(element); } // Try to add an element to the queue without blocking. Returns // true only if the element was actually added. - bool tryPush(ElementT const& element); + template + bool tryPush(T&& element); // legacy name bool tryPushFront(ElementT const & element) { return tryPush(element); } @@ -102,9 +107,9 @@ public: // to lock the mutex, versus how long to wait for the queue to stop being // full. Careful settings for each timeout might be orders of magnitude // apart. However, this method conflates them. - template + template bool tryPushFor(const std::chrono::duration& timeout, - ElementT const & element); + T&& element); // legacy name template bool tryPushFrontFor(const std::chrono::duration& timeout, @@ -112,9 +117,9 @@ public: // Try to add an element to the queue, blocking if full but with // timeout at specified time_point. Returns true if the element was added. - template - bool tryPushUntil(const std::chrono::time_point& timeout, - ElementT const& element); + template + bool tryPushUntil(const std::chrono::time_point& until, + T&& element); // no legacy name because this is a newer method // Pop the element at the head of the queue (will block if the queue is @@ -141,7 +146,7 @@ public: // Pop the element at the head of the queue, blocking if empty, with // timeout at specified time_point. Returns true if an element was popped. template - bool tryPopUntil(const std::chrono::time_point& timeout, + bool tryPopUntil(const std::chrono::time_point& until, ElementT& element); // no legacy name because this is a newer method @@ -172,11 +177,74 @@ protected: typedef std::unique_lock lock_t; boost::fibers::condition_variable_any mCapacityCond; boost::fibers::condition_variable_any mEmptyCond; -}; -// LLThreadSafeQueue -//----------------------------------------------------------------------------- + // if we're able to lock immediately, do so and run the passed callable, + // which must accept lock_t& and return bool + template + bool tryLock(CALLABLE&& callable); + // if we're able to lock before the passed time_point, do so and run the + // passed callable, which must accept lock_t& and return bool + template + bool tryLockUntil(const std::chrono::time_point& until, + CALLABLE&& callable); + // while lock is locked, really push the passed element, if we can + template + bool push_(lock_t& lock, T&& element); + // while lock is locked, really pop the head element, if we can + template + bool pop_(lock_t& lock, ElementT& element, + PRED&& pred=[](const ElementT&){ return true; }); +}; +/***************************************************************************** +* PriorityQueueAdapter +*****************************************************************************/ +namespace LL +{ + /** + * std::priority_queue's API is almost like std::queue, intentionally of + * course, but you must access the element about to pop() as top() rather + * than as front(). Make an adapter for use with LLThreadSafeQueue. + */ + template , + typename Compare=std::less> + class PriorityQueueAdapter + { + public: + // publish all the same types + typedef std::priority_queue queue_type; + typedef typename queue_type::container_type container_type; + typedef typename queue_type::value_compare value_compare; + typedef typename queue_type::value_type value_type; + typedef typename queue_type::size_type size_type; + typedef typename queue_type::reference reference; + typedef typename queue_type::const_reference const_reference; + + // Although std::queue defines both const and non-const front() + // methods, std::priority_queue defines only const top(). + const_reference front() const { return mQ.top(); } + // std::priority_queue has no equivalent to back(), so it's good that + // LLThreadSafeQueue doesn't use it. + + // All the rest of these merely forward to the corresponding + // queue_type methods. + bool empty() const { return mQ.empty(); } + size_type size() const { return mQ.size(); } + void push(const value_type& value) { mQ.push(value); } + void push(value_type&& value) { mQ.push(std::move(value)); } + template + void emplace(Args&&... args) { mQ.emplace(std::forward(args)...); } + void pop() { mQ.pop(); } + + private: + queue_type mQ; + }; +} // namespace LL + + +/***************************************************************************** +* LLThreadSafeQueue implementation +*****************************************************************************/ template LLThreadSafeQueue::LLThreadSafeQueue(U32 capacity) : mCapacity(capacity), @@ -185,24 +253,69 @@ LLThreadSafeQueue::LLThreadSafeQueue(U32 capacity) : } -template -void LLThreadSafeQueue::push(ElementT const & element) +// if we're able to lock immediately, do so and run the passed callable, which +// must accept lock_t& and return bool +template +template +bool LLThreadSafeQueue::tryLock(CALLABLE&& callable) +{ + lock_t lock1(mLock, std::defer_lock); + if (!lock1.try_lock()) + return false; + + return std::forward(callable)(lock1); +} + + +// if we're able to lock before the passed time_point, do so and run the +// passed callable, which must accept lock_t& and return bool +template +template +bool LLThreadSafeQueue::tryLockUntil( + const std::chrono::time_point& until, + CALLABLE&& callable) +{ + lock_t lock1(mLock, std::defer_lock); + if (!lock1.try_lock_until(until)) + return false; + + return std::forward(callable)(lock1); +} + + +// while lock is locked, really push the passed element, if we can +template +template +bool LLThreadSafeQueue::push_(lock_t& lock, T&& element) +{ + if (mStorage.size() >= mCapacity) + return false; + + mStorage.push(std::forward(element)); + lock.unlock(); + // now that we've pushed, if somebody's been waiting to pop, signal them + mEmptyCond.notify_one(); + return true; +} + + +template +template +void LLThreadSafeQueue::push(T&& element) { lock_t lock1(mLock); while (true) { + // On the producer side, it doesn't matter whether the queue has been + // drained or not: the moment either end calls close(), further push() + // operations will fail. if (mClosed) { LLTHROW(LLThreadSafeQueueInterrupt()); } - if (mStorage.size() < mCapacity) - { - mStorage.push(element); - lock1.unlock(); - mEmptyCond.notify_one(); + if (push_(lock1, std::forward(element))) return; - } // Storage Full. Wait for signal. mCapacityCond.wait(lock1); @@ -210,71 +323,85 @@ void LLThreadSafeQueue::push(ElementT const & element) } +template +template +bool LLThreadSafeQueue::tryPush(T&& element) +{ + return tryLock( + [this, element=std::move(element)](lock_t& lock) + { + if (mClosed) + return false; + return push_(lock, std::move(element)); + }); +} + + template -template +template bool LLThreadSafeQueue::tryPushFor( const std::chrono::duration& timeout, - ElementT const & element) + T&& element) { // Convert duration to time_point: passing the same timeout duration to // each of multiple calls is wrong. - return tryPushUntil(std::chrono::steady_clock::now() + timeout, element); + return tryPushUntil(std::chrono::steady_clock::now() + timeout, + std::forward(element)); } template -template +template bool LLThreadSafeQueue::tryPushUntil( - const std::chrono::time_point& endpoint, - ElementT const& element) + const std::chrono::time_point& until, + T&& element) { - lock_t lock1(mLock, std::defer_lock); - if (!lock1.try_lock_until(endpoint)) - return false; - - while (true) - { - if (mClosed) + return tryLockUntil( + until, + [this, until, element=std::move(element)](lock_t& lock) { - return false; - } - - if (mStorage.size() < mCapacity) - { - mStorage.push(element); - lock1.unlock(); - mEmptyCond.notify_one(); - return true; - } - - // Storage Full. Wait for signal. - if (LLCoros::cv_status::timeout == mCapacityCond.wait_until(lock1, endpoint)) - { - // timed out -- formally we might recheck both conditions above - return false; - } - // If we didn't time out, we were notified for some reason. Loop back - // to check. - } + while (true) + { + if (mClosed) + { + return false; + } + + if (push_(lock, std::move(element))) + return true; + + // Storage Full. Wait for signal. + if (LLCoros::cv_status::timeout == mCapacityCond.wait_until(lock, until)) + { + // timed out -- formally we might recheck both conditions above + return false; + } + // If we didn't time out, we were notified for some reason. Loop back + // to check. + } + }); } -template -bool LLThreadSafeQueue::tryPush(ElementT const & element) +// while lock is locked, really pop the head element, if we can +template +template +bool LLThreadSafeQueue::pop_( + lock_t& lock, ElementT& element, PRED&& pred) { - lock_t lock1(mLock, std::defer_lock); - if (!lock1.try_lock()) - return false; - - if (mClosed) - return false; - - if (mStorage.size() >= mCapacity) + // If mStorage is empty, there's no head element. + // If there's a head element, pass it to the predicate to see if caller + // considers it ready to pop. + // Unless both are satisfied, no point in continuing. + if (mStorage.empty() || ! std::forward(pred)(mStorage.front())) return false; - mStorage.push(element); - lock1.unlock(); - mEmptyCond.notify_one(); + // std::queue::front() is the element about to pop() + element = mStorage.front(); + mStorage.pop(); + lock.unlock(); + // now that we've popped, if somebody's been waiting to push, signal them + mCapacityCond.notify_one(); return true; } @@ -285,22 +412,20 @@ ElementT LLThreadSafeQueue::pop(void) lock_t lock1(mLock); while (true) { - if (!mStorage.empty()) - { - // std::queue::front() is the element about to pop() - ElementT value = mStorage.front(); - mStorage.pop(); - lock1.unlock(); - mCapacityCond.notify_one(); - return value; - } - + // On the consumer side, we always try to pop before checking mClosed + // so we can finish draining the queue. + ElementT value; + if (pop_(lock1, value)) + return std::move(value); + + // Once the queue is empty, mClosed lets us know if there will ever be + // any more coming. if (mClosed) { LLTHROW(LLThreadSafeQueueInterrupt()); } - // Storage empty. Wait for signal. + // Storage empty, queue still open. Wait for signal. mEmptyCond.wait(lock1); } } @@ -309,21 +434,14 @@ ElementT LLThreadSafeQueue::pop(void) template bool LLThreadSafeQueue::tryPop(ElementT & element) { - lock_t lock1(mLock, std::defer_lock); - if (!lock1.try_lock()) - return false; - - // no need to check mClosed: tryPop() behavior when the queue is - // closed is implemented by simple inability to push any new elements - if (mStorage.empty()) - return false; - - // std::queue::front() is the element about to pop() - element = mStorage.front(); - mStorage.pop(); - lock1.unlock(); - mCapacityCond.notify_one(); - return true; + return tryLock( + [this, &element](lock_t& lock) + { + // no need to check mClosed: tryPop() behavior when the queue is + // closed is implemented by simple inability to push any new + // elements + return pop_(lock, element); + }); } @@ -342,39 +460,33 @@ bool LLThreadSafeQueue::tryPopFor( template template bool LLThreadSafeQueue::tryPopUntil( - const std::chrono::time_point& endpoint, + const std::chrono::time_point& until, ElementT& element) { - lock_t lock1(mLock, std::defer_lock); - if (!lock1.try_lock_until(endpoint)) - return false; - - while (true) - { - if (!mStorage.empty()) + return tryLockUntil( + until, + [this, until, &element](lock_t& lock) { - // std::queue::front() is the element about to pop() - element = mStorage.front(); - mStorage.pop(); - lock1.unlock(); - mCapacityCond.notify_one(); - return true; - } - - if (mClosed) - { - return false; - } - - // Storage empty. Wait for signal. - if (LLCoros::cv_status::timeout == mEmptyCond.wait_until(lock1, endpoint)) - { - // timed out -- formally we might recheck both conditions above - return false; - } - // If we didn't time out, we were notified for some reason. Loop back - // to check. - } + while (true) + { + if (pop_(lock, element)) + return true; + + if (mClosed) + { + return false; + } + + // Storage empty. Wait for signal. + if (LLCoros::cv_status::timeout == mEmptyCond.wait_until(lock, until)) + { + // timed out -- formally we might recheck both conditions above + return false; + } + // If we didn't time out, we were notified for some reason. Loop back + // to check. + } + }); } -- cgit v1.3 From a35e266547e4d2c8dbd6b003c64b719d91eaaf87 Mon Sep 17 00:00:00 2001 From: Nat Goodspeed Date: Mon, 4 Oct 2021 17:21:39 -0400 Subject: SL-16024: Don't use a lambda as default arg for universal reference. Instead, break out a separate pop_() method that explicitly provides the lambda to the real pop_() implementation. --- indra/llcommon/llthreadsafequeue.h | 16 ++++++++++++++-- 1 file changed, 14 insertions(+), 2 deletions(-) (limited to 'indra/llcommon/llthreadsafequeue.h') diff --git a/indra/llcommon/llthreadsafequeue.h b/indra/llcommon/llthreadsafequeue.h index c57520c01f..1dffad6b89 100644 --- a/indra/llcommon/llthreadsafequeue.h +++ b/indra/llcommon/llthreadsafequeue.h @@ -191,9 +191,11 @@ protected: template bool push_(lock_t& lock, T&& element); // while lock is locked, really pop the head element, if we can + bool pop_(lock_t& lock, ElementT& element); + // pop_() with an explicit predicate indicating whether the head element + // is ready to be popped template - bool pop_(lock_t& lock, ElementT& element, - PRED&& pred=[](const ElementT&){ return true; }); + bool pop_(lock_t& lock, ElementT& element, PRED&& pred); }; /***************************************************************************** @@ -385,6 +387,16 @@ bool LLThreadSafeQueue::tryPushUntil( // while lock is locked, really pop the head element, if we can template +bool LLThreadSafeQueue::pop_(lock_t& lock, ElementT& element) +{ + // default predicate: head element, if present, is always ready to pop + return pop_(lock, element, [](const ElementT&){ return true; }); +} + + +// pop_() with an explicit predicate indicating whether the head element +// is ready to be popped +template template bool LLThreadSafeQueue::pop_( lock_t& lock, ElementT& element, PRED&& pred) -- cgit v1.3 From 955b967623983cb50ba09f7b82e5f01f2c6bcebb Mon Sep 17 00:00:00 2001 From: Nat Goodspeed Date: Tue, 5 Oct 2021 17:31:53 -0400 Subject: SL-16024: Add ThreadSafeSchedule, a timestamped LLThreadSafeQueue. ThreadSafeSchedule orders its items by timestamp, which can be passed either implicitly or explicitly. The timestamp specifies earliest delivery time: an item cannot be popped until that time. Add initial tests. Tweak the LLThreadSafeQueue base class to support ThreadSafeSchedule: introduce virtual canPop() method to report whether the current head item is available to pop. The base class unconditionally says yes, ThreadSafeSchedule says it depends on whether its timestamp is still in the future. This replaces the protected pop_() overload accepting a predicate. Rather than explicitly passing a predicate through a couple levels of function call, use canPop() at the level it matters. Runtime behavior that varies depending on an object's leaf class is what virtual functions were invented for. Give pop_() a three-state enum return so pop() can distinguish between "closed and empty" (throws exception) versus "closed, not yet drained because we're not yet ready to pop the head item" (waits). Also break out protected tryPopUntil_() method, the body logic of tryPopUntil(). The public method locks the data structure, the protected method requires that its caller has already done so. Add chrono.h with a more full-featured LL::time_point_cast() function than the one found in , which only converts between time_point durations, not between time_points based on different clocks. --- indra/llcommon/CMakeLists.txt | 3 + indra/llcommon/chrono.h | 65 +++++ indra/llcommon/llthreadsafequeue.h | 121 ++++---- indra/llcommon/tests/threadsafeschedule_test.cpp | 65 +++++ indra/llcommon/threadsafeschedule.h | 334 +++++++++++++++++++++++ 5 files changed, 535 insertions(+), 53 deletions(-) create mode 100644 indra/llcommon/chrono.h create mode 100644 indra/llcommon/tests/threadsafeschedule_test.cpp create mode 100644 indra/llcommon/threadsafeschedule.h (limited to 'indra/llcommon/llthreadsafequeue.h') diff --git a/indra/llcommon/CMakeLists.txt b/indra/llcommon/CMakeLists.txt index 6558219462..5efcfabf24 100644 --- a/indra/llcommon/CMakeLists.txt +++ b/indra/llcommon/CMakeLists.txt @@ -127,6 +127,7 @@ set(llcommon_SOURCE_FILES set(llcommon_HEADER_FILES CMakeLists.txt + chrono.h ctype_workaround.h fix_macros.h indra_constants.h @@ -253,6 +254,7 @@ set(llcommon_HEADER_FILES lockstatic.h stdtypes.h stringize.h + threadsafeschedule.h timer.h tuple.h u64.h @@ -359,6 +361,7 @@ if (LL_TESTS) LL_ADD_INTEGRATION_TEST(lluri "" "${test_libs}") LL_ADD_INTEGRATION_TEST(llunits "" "${test_libs}") LL_ADD_INTEGRATION_TEST(stringize "" "${test_libs}") + LL_ADD_INTEGRATION_TEST(threadsafeschedule "" "${test_libs}") LL_ADD_INTEGRATION_TEST(tuple "" "${test_libs}") ## llexception_test.cpp isn't a regression test, and doesn't need to be run diff --git a/indra/llcommon/chrono.h b/indra/llcommon/chrono.h new file mode 100644 index 0000000000..806e871892 --- /dev/null +++ b/indra/llcommon/chrono.h @@ -0,0 +1,65 @@ +/** + * @file chrono.h + * @author Nat Goodspeed + * @date 2021-10-05 + * @brief supplement with utility functions + * + * $LicenseInfo:firstyear=2021&license=viewerlgpl$ + * Copyright (c) 2021, Linden Research, Inc. + * $/LicenseInfo$ + */ + +#if ! defined(LL_CHRONO_H) +#define LL_CHRONO_H + +#include +#include // std::enable_if + +namespace LL +{ + +// time_point_cast() is derived from https://stackoverflow.com/a/35293183 +// without the iteration: we think errors in the ~1 microsecond range are +// probably acceptable. + +// This variant is for the optimal case when the source and dest use the same +// clock: that case is handled by std::chrono. +template ::value, + bool>::type = true> +DestTimePoint time_point_cast(const SrcTimePoint& time) +{ + return std::chrono::time_point_cast(time); +} + +// This variant is for when the source and dest use different clocks -- see +// the linked StackOverflow answer, also Howard Hinnant's, for more context. +template ::value, + bool>::type = true> +DestTimePoint time_point_cast(const SrcTimePoint& time) +{ + // The basic idea is that we must adjust the passed time_point by the + // difference between the clocks' epochs. But since time_point doesn't + // expose its epoch, we fall back on what each of them thinks is now(). + // However, since we necessarily make sequential calls to those now() + // functions, the answers differ not only by the cycles spent executing + // those calls, but by potential OS interruptions between them. Try to + // reduce that error by capturing the source clock time both before and + // after the dest clock, and splitting the difference. Of course an + // interruption between two of these now() calls without a comparable + // interruption between the other two will skew the result, but better is + // more expensive. + const auto src_before = typename SrcTimePoint::clock::now(); + const auto dest_now = typename DestTimePoint::clock::now(); + const auto src_after = typename SrcTimePoint::clock::now(); + const auto src_diff = src_after - src_before; + const auto src_now = src_before + src_diff / 2; + return dest_now + (time - src_now); +} + +} // namespace LL + +#endif /* ! defined(LL_CHRONO_H) */ diff --git a/indra/llcommon/llthreadsafequeue.h b/indra/llcommon/llthreadsafequeue.h index 1dffad6b89..bd2d82d4c3 100644 --- a/indra/llcommon/llthreadsafequeue.h +++ b/indra/llcommon/llthreadsafequeue.h @@ -83,6 +83,7 @@ public: // Limiting the number of pending items prevents unbounded growth of the // underlying queue. LLThreadSafeQueue(U32 capacity = 1024); + virtual ~LLThreadSafeQueue() {} // Add an element to the queue (will block if the queue has // reached capacity). @@ -162,10 +163,10 @@ public: // then every subsequent tryPop() call will return false void close(); - // detect closed state + // producer end: are we prevented from pushing any additional items? bool isClosed(); - // inverse of isClosed() - explicit operator bool(); + // consumer end: are we done, is the queue entirely drained? + bool done(); protected: typedef QueueT queue_type; @@ -178,6 +179,11 @@ protected: boost::fibers::condition_variable_any mCapacityCond; boost::fibers::condition_variable_any mEmptyCond; + // implementation logic, suitable for passing to tryLockUntil() + template + bool tryPopUntil_(lock_t& lock, + const std::chrono::time_point& until, + ElementT& element); // if we're able to lock immediately, do so and run the passed callable, // which must accept lock_t& and return bool template @@ -191,11 +197,11 @@ protected: template bool push_(lock_t& lock, T&& element); // while lock is locked, really pop the head element, if we can - bool pop_(lock_t& lock, ElementT& element); - // pop_() with an explicit predicate indicating whether the head element - // is ready to be popped - template - bool pop_(lock_t& lock, ElementT& element, PRED&& pred); + enum pop_result { EMPTY, WAITING, POPPED }; + pop_result pop_(lock_t& lock, ElementT& element); + // Is the current head element ready to pop? We say yes; subclass can + // override as needed. + virtual bool canPop(const ElementT& head) const { return true; } }; /***************************************************************************** @@ -387,26 +393,16 @@ bool LLThreadSafeQueue::tryPushUntil( // while lock is locked, really pop the head element, if we can template -bool LLThreadSafeQueue::pop_(lock_t& lock, ElementT& element) -{ - // default predicate: head element, if present, is always ready to pop - return pop_(lock, element, [](const ElementT&){ return true; }); -} - - -// pop_() with an explicit predicate indicating whether the head element -// is ready to be popped -template -template -bool LLThreadSafeQueue::pop_( - lock_t& lock, ElementT& element, PRED&& pred) +typename LLThreadSafeQueue::pop_result +LLThreadSafeQueue::pop_(lock_t& lock, ElementT& element) { // If mStorage is empty, there's no head element. - // If there's a head element, pass it to the predicate to see if caller - // considers it ready to pop. - // Unless both are satisfied, no point in continuing. - if (mStorage.empty() || ! std::forward(pred)(mStorage.front())) - return false; + if (mStorage.empty()) + return EMPTY; + + // If there's a head element, pass it to canPop() to see if it's ready to pop. + if (! canPop(mStorage.front())) + return WAITING; // std::queue::front() is the element about to pop() element = mStorage.front(); @@ -414,7 +410,7 @@ bool LLThreadSafeQueue::pop_( lock.unlock(); // now that we've popped, if somebody's been waiting to push, signal them mCapacityCond.notify_one(); - return true; + return POPPED; } @@ -422,17 +418,20 @@ template ElementT LLThreadSafeQueue::pop(void) { lock_t lock1(mLock); + ElementT value; while (true) { // On the consumer side, we always try to pop before checking mClosed // so we can finish draining the queue. - ElementT value; - if (pop_(lock1, value)) + pop_result popped = pop_(lock1, value); + if (popped == POPPED) return std::move(value); // Once the queue is empty, mClosed lets us know if there will ever be - // any more coming. - if (mClosed) + // any more coming. If we didn't pop because WAITING, i.e. canPop() + // returned false, then even if the producer end has been closed, + // there's still at least one item to drain: wait for it. + if (popped == EMPTY && mClosed) { LLTHROW(LLThreadSafeQueueInterrupt()); } @@ -452,7 +451,7 @@ bool LLThreadSafeQueue::tryPop(ElementT & element) // no need to check mClosed: tryPop() behavior when the queue is // closed is implemented by simple inability to push any new // elements - return pop_(lock, element); + return pop_(lock, element) == POPPED; }); } @@ -479,26 +478,38 @@ bool LLThreadSafeQueue::tryPopUntil( until, [this, until, &element](lock_t& lock) { - while (true) - { - if (pop_(lock, element)) - return true; + return tryPopUntil_(lock, until, element); + }); +} - if (mClosed) - { - return false; - } - // Storage empty. Wait for signal. - if (LLCoros::cv_status::timeout == mEmptyCond.wait_until(lock, until)) - { - // timed out -- formally we might recheck both conditions above - return false; - } - // If we didn't time out, we were notified for some reason. Loop back - // to check. - } - }); +// body of tryPopUntil(), called once we have the lock +template +template +bool LLThreadSafeQueue::tryPopUntil_( + lock_t& lock, + const std::chrono::time_point& until, + ElementT& element) +{ + while (true) + { + if (pop_(lock, element) == POPPED) + return true; + + if (mClosed) + { + return false; + } + + // Storage empty. Wait for signal. + if (LLCoros::cv_status::timeout == mEmptyCond.wait_until(lock, until)) + { + // timed out -- formally we might recheck both conditions above + return false; + } + // If we didn't time out, we were notified for some reason. Loop back + // to check. + } } @@ -509,6 +520,7 @@ size_t LLThreadSafeQueue::size(void) return mStorage.size(); } + template void LLThreadSafeQueue::close() { @@ -521,17 +533,20 @@ void LLThreadSafeQueue::close() mCapacityCond.notify_all(); } + template bool LLThreadSafeQueue::isClosed() { lock_t lock(mLock); - return mClosed && mStorage.size() == 0; + return mClosed; } + template -LLThreadSafeQueue::operator bool() +bool LLThreadSafeQueue::done() { - return ! isClosed(); + lock_t lock(mLock); + return mClosed && mStorage.size() == 0; } #endif diff --git a/indra/llcommon/tests/threadsafeschedule_test.cpp b/indra/llcommon/tests/threadsafeschedule_test.cpp new file mode 100644 index 0000000000..ec0fa0c928 --- /dev/null +++ b/indra/llcommon/tests/threadsafeschedule_test.cpp @@ -0,0 +1,65 @@ +/** + * @file threadsafeschedule_test.cpp + * @author Nat Goodspeed + * @date 2021-10-04 + * @brief Test for threadsafeschedule. + * + * $LicenseInfo:firstyear=2021&license=viewerlgpl$ + * Copyright (c) 2021, Linden Research, Inc. + * $/LicenseInfo$ + */ + +// Precompiled header +#include "linden_common.h" +// associated header +#include "threadsafeschedule.h" +// STL headers +// std headers +#include +// external library headers +// other Linden headers +#include "../test/lltut.h" + +using namespace std::literals::chrono_literals; // ms suffix +using namespace std::literals::string_literals; // s suffix +using Queue = LL::ThreadSafeSchedule; + +/***************************************************************************** +* TUT +*****************************************************************************/ +namespace tut +{ + struct threadsafeschedule_data + { + Queue queue; + }; + typedef test_group threadsafeschedule_group; + typedef threadsafeschedule_group::object object; + threadsafeschedule_group threadsafeschedulegrp("threadsafeschedule"); + + template<> template<> + void object::test<1>() + { + set_test_name("push"); + // Simply calling push() a few times might result in indeterminate + // delivery order if the resolution of steady_clock is coarser than + // the real time required for each push() call. Explicitly increment + // the timestamp for each one -- but since we're passing explicit + // timestamps, make the queue reorder them. + queue.push(Queue::TimeTuple(Queue::Clock::now() + 20ms, "ghi")); + queue.push("abc"s); + queue.push(Queue::Clock::now() + 10ms, "def"); + queue.close(); + auto entry = queue.pop(); + ensure_equals("failed to pop first", std::get<0>(entry), "abc"s); + entry = queue.pop(); + ensure_equals("failed to pop second", std::get<0>(entry), "def"s); + ensure("queue not closed", queue.isClosed()); + ensure("queue prematurely done", ! queue.done()); + entry = queue.pop(); + ensure_equals("failed to pop third", std::get<0>(entry), "ghi"s); + bool popped = queue.tryPop(entry); + ensure("queue not empty", ! popped); + ensure("queue not done", queue.done()); + } +} // namespace tut diff --git a/indra/llcommon/threadsafeschedule.h b/indra/llcommon/threadsafeschedule.h new file mode 100644 index 0000000000..545c820f53 --- /dev/null +++ b/indra/llcommon/threadsafeschedule.h @@ -0,0 +1,334 @@ +/** + * @file threadsafeschedule.h + * @author Nat Goodspeed + * @date 2021-10-02 + * @brief ThreadSafeSchedule is an ordered queue in which every item has an + * associated timestamp. + * + * $LicenseInfo:firstyear=2021&license=viewerlgpl$ + * Copyright (c) 2021, Linden Research, Inc. + * $/LicenseInfo$ + */ + +#if ! defined(LL_THREADSAFESCHEDULE_H) +#define LL_THREADSAFESCHEDULE_H + +#include "chrono.h" +#include "llexception.h" +#include "llthreadsafequeue.h" +#include "tuple.h" +#include +#include + +namespace LL +{ + namespace ThreadSafeSchedulePrivate + { + using TimePoint = std::chrono::steady_clock::time_point; + // Bundle consumer's data with a TimePoint to order items by timestamp. + template + using TimestampedTuple = std::tuple; + + // comparison functor for TimedTuples -- see TimedQueue comments + struct ReverseTupleOrder + { + template + bool operator()(const Tuple& left, const Tuple& right) const + { + return std::get<0>(left) > std::get<0>(right); + } + }; + + template + using TimedQueue = PriorityQueueAdapter< + TimestampedTuple, + // std::vector is the default storage for std::priority_queue, + // have to restate to specify comparison template parameter + std::vector>, + // std::priority_queue uses a counterintuitive comparison + // behavior: the default std::less comparator is used to present + // the *highest* value as top(). So to sort by earliest timestamp, + // we must invert by using >. + ReverseTupleOrder>; + } // namespace ThreadSafeSchedulePrivate + + /** + * ThreadSafeSchedule is an ordered LLThreadSafeQueue in which every item + * is given an associated timestamp. That is, TimePoint is implicitly + * prepended to the std::tuple with the specified types. + * + * Items are popped in increasing chronological order. Moreover, any item + * with a timestamp in the future is held back until + * std::chrono::steady_clock reaches that timestamp. + */ + template + class ThreadSafeSchedule: + public LLThreadSafeQueue, + ThreadSafeSchedulePrivate::TimedQueue> + { + public: + using DataTuple = std::tuple; + using TimeTuple = ThreadSafeSchedulePrivate::TimestampedTuple; + + private: + using super = LLThreadSafeQueue>; + using lock_t = typename super::lock_t; + using super::pop_; + using super::push_; + using super::mClosed; + using super::mEmptyCond; + using super::mCapacityCond; + + public: + using TimePoint = ThreadSafeSchedulePrivate::TimePoint; + using Clock = TimePoint::clock; + + ThreadSafeSchedule(U32 capacity=1024): + super(capacity) + {} + + /*----------------------------- push() -----------------------------*/ + /// explicitly pass TimeTuple + using super::push; + + /// pass DataTuple with implicit now + void push(const DataTuple& tuple) + { + push(tuple_cons(Clock::now(), tuple)); + } + + /// individually pass each component of the TimeTuple + void push(const TimePoint& time, Args&&... args) + { + push(TimeTuple(time, std::forward(args)...)); + } + + /// individually pass every component except the TimePoint (implies + /// now) -- could be ambiguous if the first specified template + /// parameter type is also TimePoint -- we could try to disambiguate, + /// but a simpler approach would be for the caller to explicitly + /// construct DataTuple and call that overload + void push(Args&&... args) + { + push(Clock::now(), std::forward(args)...); + } + + /*--------------------------- tryPush() ----------------------------*/ + /// explicit TimeTuple + using super::tryPush; + + /// DataTuple with implicit now + bool tryPush(const DataTuple& tuple) + { + return tryPush(tuple_cons(Clock::now(), tuple)); + } + + /// individually pass components + bool tryPush(const TimePoint& time, Args&&... args) + { + return tryPush(TimeTuple(time, std::forward(args)...)); + } + + /// individually pass components with implicit now + bool tryPush(Args&&... args) + { + return tryPush(Clock::now(), std::forward(args)...); + } + + /*-------------------------- tryPushFor() --------------------------*/ + /// explicit TimeTuple + using super::tryPushFor; + + /// DataTuple with implicit now + template + bool tryPushFor(const std::chrono::duration& timeout, + const DataTuple& tuple) + { + return tryPushFor(timeout, tuple_cons(Clock::now(), tuple)); + } + + /// individually pass components + template + bool tryPushFor(const std::chrono::duration& timeout, + const TimePoint& time, Args&&... args) + { + return tryPushFor(TimeTuple(time, std::forward(args)...)); + } + + /// individually pass components with implicit now + template + bool tryPushFor(const std::chrono::duration& timeout, + Args&&... args) + { + return tryPushFor(Clock::now(), std::forward(args)...); + } + + /*------------------------- tryPushUntil() -------------------------*/ + /// explicit TimeTuple + using super::tryPushUntil; + + /// DataTuple with implicit now + template + bool tryPushUntil(const std::chrono::time_point& until, + const DataTuple& tuple) + { + return tryPushUntil(until, tuple_cons(Clock::now(), tuple)); + } + + /// individually pass components + template + bool tryPushUntil(const std::chrono::time_point& until, + const TimePoint& time, Args&&... args) + { + return tryPushUntil(until, TimeTuple(time, std::forward(args)...)); + } + + /// individually pass components with implicit now + template + bool tryPushUntil(const std::chrono::time_point& until, + Args&&... args) + { + return tryPushUntil(until, Clock::now(), std::forward(args)...); + } + + /*----------------------------- pop() ------------------------------*/ + // Our consumer may or may not care about the timestamp associated + // with each popped item, so we allow retrieving either DataTuple or + // TimeTuple. One potential use would be to observe, and possibly + // adjust for, the time lag between the item time and the actual + // current time. + + /// pop DataTuple by value + DataTuple pop() + { + return tuple_cdr(popWithTime()); + } + + /// pop TimeTuple by value + TimeTuple popWithTime() + { + lock_t lock(super::mLock); + // We can't just sit around waiting forever, given that there may + // be items in the queue that are not yet ready but will *become* + // ready in the near future. So in fact, with this class, every + // pop() becomes a tryPopUntil(), constrained to the timestamp of + // the head item. It almost doesn't matter what we specify for the + // caller's time constraint -- all we really care about is the + // head item's timestamp. Since pop() and popWithTime() are + // defined to wait until either an item becomes available or the + // queue is closed, loop until one of those things happens. The + // constraint we pass just determines how often we'll loop while + // waiting. + TimeTuple tt; + while (true) + { + // Pick a point suitably far into the future. + TimePoint until = TimePoint::clock::now() + std::chrono::hours(24); + if (tryPopUntil_(lock, until, tt)) + return std::move(tt); + + // empty and closed: throw, just as super::pop() does + if (super::mStorage.empty() && super::mClosed) + { + LLTHROW(LLThreadSafeQueueInterrupt()); + } + // If not empty, we've still got items to drain. + // If not closed, it's worth waiting for more items. + // Either way, loop back to wait. + } + } + + // We can use tryPop(TimeTuple&) just as it stands; the only behavior + // difference is in our canPop() override method. + using super::tryPop; + + /// tryPop(DataTuple&) + bool tryPop(DataTuple& tuple) + { + TimeTuple tt; + if (! super::tryPop(tt)) + return false; + tuple = tuple_cdr(std::move(tt)); + return true; + } + + /// tryPopFor() + template + bool tryPopFor(const std::chrono::duration& timeout, Tuple& tuple) + { + // It's important to use OUR tryPopUntil() implementation, rather + // than delegating immediately to our base class. + return tryPopUntil(Clock::now() + timeout, tuple); + } + + /// tryPopUntil(TimeTuple&) + template + bool tryPopUntil(const std::chrono::time_point& until, + TimeTuple& tuple) + { + // super::tryPopUntil() wakes up when an item becomes available or + // we hit 'until', whichever comes first. Thing is, the current + // head of the queue could become ready sooner than either of + // those events, and we need to deliver it as soon as it does. + // Don't wait past the TimePoint of the head item. + // Naturally, lock the queue before peeking at mStorage. + return super::tryLockUntil( + until, + [this, until, &tuple](lock_t& lock) + { + // Use our time_point_cast to allow for 'until' that's a + // time_point type other than TimePoint. + return tryPopUntil_(lock, time_point_cast(until), tuple); + }); + } + + bool tryPopUntil_(lock_t& lock, const TimePoint& until, TimeTuple& tuple) + { + TimePoint adjusted = until; + if (! super::mStorage.empty()) + { + // use whichever is earlier: the head item's timestamp, or + // the caller's limit + adjusted = min(std::get<0>(super::mStorage.front()), adjusted); + } + // now delegate to base-class tryPopUntil_() + return super::tryPopUntil_(lock, adjusted, tuple); + } + + /// tryPopUntil(DataTuple&) + template + bool tryPopUntil(const std::chrono::time_point& until, + DataTuple& tuple) + { + TimeTuple tt; + if (! tryPopUntil(until, tt)) + return false; + tuple = tuple_cdr(std::move(tt)); + return true; + } + + /*------------------------------ etc. ------------------------------*/ + // We can't hide items that aren't yet ready because we can't traverse + // the underlying priority_queue: it has no iterators, only top(). So + // a consumer could observe size() > 0 and yet tryPop() returns false. + // Shrug, in a multi-consumer scenario that would be expected behavior. + using super::size; + // open/closed state + using super::close; + using super::isClosed; + using super::done; + + private: + // this method is called by base class pop_() every time we're + // considering whether to deliver the current head element + bool canPop(const TimeTuple& head) const override + { + // an item with a future timestamp isn't yet ready to pop + // (should we add some slop for overhead?) + return std::get<0>(head) <= Clock::now(); + } + }; + +} // namespace LL + +#endif /* ! defined(LL_THREADSAFESCHEDULE_H) */ -- cgit v1.3 From cf70766b4504f7ee745822926c526ed9c86c9339 Mon Sep 17 00:00:00 2001 From: Nat Goodspeed Date: Wed, 6 Oct 2021 12:54:29 -0400 Subject: SL-16024: Fix ThreadSafeSchedule::tryPopFor(), tryPopUntil(). ThreadSafeSchedule::tryPopUntil() (and therefore tryPopFor()) was simply delegating to LLThreadSafeQueue::tryPopUntil(), with an adjusted timeout since we want to wake up as soon as the head item, if any, becomes ready. But then we have to loop back to retry the pop to actually deal with that head item. In addition, ThreadSafeSchedule::popWithTime() was spinning rather than properly blocking on a timed condition variable. Fixed. --- indra/llcommon/llthreadsafequeue.h | 51 +++++++++-------- indra/llcommon/tests/threadsafeschedule_test.cpp | 10 +++- indra/llcommon/threadsafeschedule.h | 72 ++++++++++++++++++------ 3 files changed, 88 insertions(+), 45 deletions(-) (limited to 'indra/llcommon/llthreadsafequeue.h') diff --git a/indra/llcommon/llthreadsafequeue.h b/indra/llcommon/llthreadsafequeue.h index bd2d82d4c3..719edcd579 100644 --- a/indra/llcommon/llthreadsafequeue.h +++ b/indra/llcommon/llthreadsafequeue.h @@ -179,11 +179,12 @@ protected: boost::fibers::condition_variable_any mCapacityCond; boost::fibers::condition_variable_any mEmptyCond; + enum pop_result { EMPTY, DONE, WAITING, POPPED }; // implementation logic, suitable for passing to tryLockUntil() template - bool tryPopUntil_(lock_t& lock, - const std::chrono::time_point& until, - ElementT& element); + pop_result tryPopUntil_(lock_t& lock, + const std::chrono::time_point& until, + ElementT& element); // if we're able to lock immediately, do so and run the passed callable, // which must accept lock_t& and return bool template @@ -197,7 +198,6 @@ protected: template bool push_(lock_t& lock, T&& element); // while lock is locked, really pop the head element, if we can - enum pop_result { EMPTY, WAITING, POPPED }; pop_result pop_(lock_t& lock, ElementT& element); // Is the current head element ready to pop? We say yes; subclass can // override as needed. @@ -398,7 +398,7 @@ LLThreadSafeQueue::pop_(lock_t& lock, ElementT& element) { // If mStorage is empty, there's no head element. if (mStorage.empty()) - return EMPTY; + return mClosed? DONE : EMPTY; // If there's a head element, pass it to canPop() to see if it's ready to pop. if (! canPop(mStorage.front())) @@ -427,16 +427,16 @@ ElementT LLThreadSafeQueue::pop(void) if (popped == POPPED) return std::move(value); - // Once the queue is empty, mClosed lets us know if there will ever be - // any more coming. If we didn't pop because WAITING, i.e. canPop() - // returned false, then even if the producer end has been closed, - // there's still at least one item to drain: wait for it. - if (popped == EMPTY && mClosed) + // Once the queue is DONE, there will never be any more coming. + if (popped == DONE) { LLTHROW(LLThreadSafeQueueInterrupt()); } - // Storage empty, queue still open. Wait for signal. + // If we didn't pop because WAITING, i.e. canPop() returned false, + // then even if the producer end has been closed, there's still at + // least one item to drain: wait for it. Or we might be EMPTY, with + // the queue still open. Either way, wait for signal. mEmptyCond.wait(lock1); } } @@ -448,8 +448,8 @@ bool LLThreadSafeQueue::tryPop(ElementT & element) return tryLock( [this, &element](lock_t& lock) { - // no need to check mClosed: tryPop() behavior when the queue is - // closed is implemented by simple inability to push any new + // conflate EMPTY, DONE, WAITING: tryPop() behavior when the queue + // is closed is implemented by simple inability to push any new // elements return pop_(lock, element) == POPPED; }); @@ -478,7 +478,8 @@ bool LLThreadSafeQueue::tryPopUntil( until, [this, until, &element](lock_t& lock) { - return tryPopUntil_(lock, until, element); + // conflate EMPTY, DONE, WAITING + return tryPopUntil_(lock, until, element) == POPPED; }); } @@ -486,26 +487,28 @@ bool LLThreadSafeQueue::tryPopUntil( // body of tryPopUntil(), called once we have the lock template template -bool LLThreadSafeQueue::tryPopUntil_( +typename LLThreadSafeQueue::pop_result +LLThreadSafeQueue::tryPopUntil_( lock_t& lock, const std::chrono::time_point& until, ElementT& element) { while (true) { - if (pop_(lock, element) == POPPED) - return true; - - if (mClosed) + pop_result popped = pop_(lock, element); + if (popped == POPPED || popped == DONE) { - return false; + // If we succeeded, great! If we've drained the last item, so be + // it. Either way, break the loop and tell caller. + return popped; } - // Storage empty. Wait for signal. + // EMPTY or WAITING: wait for signal. if (LLCoros::cv_status::timeout == mEmptyCond.wait_until(lock, until)) { - // timed out -- formally we might recheck both conditions above - return false; + // timed out -- formally we might recheck + // as it is, break loop + return popped; } // If we didn't time out, we were notified for some reason. Loop back // to check. @@ -546,7 +549,7 @@ template bool LLThreadSafeQueue::done() { lock_t lock(mLock); - return mClosed && mStorage.size() == 0; + return mClosed && mStorage.empty(); } #endif diff --git a/indra/llcommon/tests/threadsafeschedule_test.cpp b/indra/llcommon/tests/threadsafeschedule_test.cpp index ec0fa0c928..af67b9f492 100644 --- a/indra/llcommon/tests/threadsafeschedule_test.cpp +++ b/indra/llcommon/tests/threadsafeschedule_test.cpp @@ -47,6 +47,8 @@ namespace tut // the timestamp for each one -- but since we're passing explicit // timestamps, make the queue reorder them. queue.push(Queue::TimeTuple(Queue::Clock::now() + 20ms, "ghi")); + // Given the various push() overloads, you have to match the type + // exactly: conversions are ambiguous. queue.push("abc"s); queue.push(Queue::Clock::now() + 10ms, "def"); queue.close(); @@ -56,9 +58,11 @@ namespace tut ensure_equals("failed to pop second", std::get<0>(entry), "def"s); ensure("queue not closed", queue.isClosed()); ensure("queue prematurely done", ! queue.done()); - entry = queue.pop(); - ensure_equals("failed to pop third", std::get<0>(entry), "ghi"s); - bool popped = queue.tryPop(entry); + std::string s; + bool popped = queue.tryPopFor(1s, s); + ensure("failed to pop third", popped); + ensure_equals("third is wrong", s, "ghi"s); + popped = queue.tryPop(s); ensure("queue not empty", ! popped); ensure("queue not done", queue.done()); } diff --git a/indra/llcommon/threadsafeschedule.h b/indra/llcommon/threadsafeschedule.h index 545c820f53..8ab4311ca1 100644 --- a/indra/llcommon/threadsafeschedule.h +++ b/indra/llcommon/threadsafeschedule.h @@ -73,11 +73,7 @@ namespace LL private: using super = LLThreadSafeQueue>; using lock_t = typename super::lock_t; - using super::pop_; - using super::push_; - using super::mClosed; - using super::mEmptyCond; - using super::mCapacityCond; + using pop_result = typename super::pop_result; public: using TimePoint = ThreadSafeSchedulePrivate::TimePoint; @@ -92,6 +88,11 @@ namespace LL using super::push; /// pass DataTuple with implicit now + // This could be ambiguous for Args with a single type. Unfortunately + // we can't enable_if an individual method with a condition based on + // the *class* template arguments, only on that method's template + // arguments. We could specialize this class for the single-Args case; + // we could minimize redundancy by breaking out a common base class... void push(const DataTuple& tuple) { push(tuple_cons(Clock::now(), tuple)); @@ -103,11 +104,11 @@ namespace LL push(TimeTuple(time, std::forward(args)...)); } - /// individually pass every component except the TimePoint (implies - /// now) -- could be ambiguous if the first specified template - /// parameter type is also TimePoint -- we could try to disambiguate, - /// but a simpler approach would be for the caller to explicitly - /// construct DataTuple and call that overload + /// individually pass every component except the TimePoint (implies now) + // This could be ambiguous if the first specified template parameter + // type is also TimePoint. We could try to disambiguate, but a simpler + // approach would be for the caller to explicitly construct DataTuple + // and call that overload. void push(Args&&... args) { push(Clock::now(), std::forward(args)...); @@ -199,6 +200,10 @@ namespace LL // current time. /// pop DataTuple by value + // It would be great to notice when sizeof...(Args) == 1 and directly + // return the first (only) value, instead of making pop()'s caller + // call std::get<0>(value). See push(DataTuple) remarks for why we + // haven't yet jumped through those hoops. DataTuple pop() { return tuple_cdr(popWithTime()); @@ -224,16 +229,17 @@ namespace LL { // Pick a point suitably far into the future. TimePoint until = TimePoint::clock::now() + std::chrono::hours(24); - if (tryPopUntil_(lock, until, tt)) + pop_result popped = tryPopUntil_(lock, until, tt); + if (popped == super::POPPED) return std::move(tt); - // empty and closed: throw, just as super::pop() does - if (super::mStorage.empty() && super::mClosed) + // DONE: throw, just as super::pop() does + if (popped == super::DONE) { LLTHROW(LLThreadSafeQueueInterrupt()); } - // If not empty, we've still got items to drain. - // If not closed, it's worth waiting for more items. + // WAITING: we've still got items to drain. + // EMPTY: not closed, so it's worth waiting for more items. // Either way, loop back to wait. } } @@ -252,6 +258,16 @@ namespace LL return true; } + /// for when Args has exactly one type + bool tryPop(typename std::tuple_element<1, TimeTuple>::type& value) + { + TimeTuple tt; + if (! super::tryPop(tt)) + return false; + value = std::get<1>(std::move(tt)); + return true; + } + /// tryPopFor() template bool tryPopFor(const std::chrono::duration& timeout, Tuple& tuple) @@ -278,11 +294,12 @@ namespace LL { // Use our time_point_cast to allow for 'until' that's a // time_point type other than TimePoint. - return tryPopUntil_(lock, time_point_cast(until), tuple); + return super::POPPED == + tryPopUntil_(lock, LL::time_point_cast(until), tuple); }); } - bool tryPopUntil_(lock_t& lock, const TimePoint& until, TimeTuple& tuple) + pop_result tryPopUntil_(lock_t& lock, const TimePoint& until, TimeTuple& tuple) { TimePoint adjusted = until; if (! super::mStorage.empty()) @@ -292,7 +309,14 @@ namespace LL adjusted = min(std::get<0>(super::mStorage.front()), adjusted); } // now delegate to base-class tryPopUntil_() - return super::tryPopUntil_(lock, adjusted, tuple); + pop_result popped; + while ((popped = super::tryPopUntil_(lock, adjusted, tuple)) == super::WAITING) + { + // If super::tryPopUntil_() returns WAITING, it means there's + // a head item, but it's not yet time. But it's worth looping + // back to recheck. + } + return popped; } /// tryPopUntil(DataTuple&) @@ -307,6 +331,18 @@ namespace LL return true; } + /// for when Args has exactly one type + template + bool tryPopUntil(const std::chrono::time_point& until, + typename std::tuple_element<1, TimeTuple>::type& value) + { + TimeTuple tt; + if (! tryPopUntil(until, tt)) + return false; + value = std::get<1>(std::move(tt)); + return true; + } + /*------------------------------ etc. ------------------------------*/ // We can't hide items that aren't yet ready because we can't traverse // the underlying priority_queue: it has no iterators, only top(). So -- cgit v1.3 From d00272e0cc9974f35a46f0c313ee2c0e11cddbda Mon Sep 17 00:00:00 2001 From: Dave Parks Date: Mon, 11 Oct 2021 16:03:40 +0000 Subject: SL-16099 Multi-threaded OpenGL usage on Windows, enable Core Profile and VAOs by default. --- indra/llcommon/llthreadsafequeue.h | 3 + indra/llimage/llimageworker.cpp | 4 + indra/llrender/llcubemap.cpp | 2 + indra/llrender/llfontfreetype.cpp | 1 + indra/llrender/llglslshader.cpp | 2 + indra/llrender/llgltexture.cpp | 2 + indra/llrender/llimagegl.cpp | 330 +++++++++++++++++++++----------- indra/llrender/llimagegl.h | 35 +++- indra/llrender/llrender.cpp | 2 +- indra/llrender/llrender.h | 2 +- indra/llwindow/llopenglview-objc.mm | 3 +- indra/llwindow/llwindow.h | 12 +- indra/llwindow/llwindowheadless.h | 3 + indra/llwindow/llwindowmacosx.cpp | 28 +++ indra/llwindow/llwindowmacosx.h | 144 +++++++------- indra/llwindow/llwindowwin32.cpp | 129 ++++++++----- indra/llwindow/llwindowwin32.h | 3 + indra/newview/app_settings/settings.xml | 4 +- indra/newview/llappviewer.cpp | 3 +- indra/newview/llnetmap.cpp | 1 + indra/newview/lltexturefetch.cpp | 4 + indra/newview/llviewertexture.cpp | 311 +++++++++++++++++++----------- indra/newview/llviewertexture.h | 8 +- indra/newview/llviewertexturelist.cpp | 36 ++++ indra/newview/llviewerwindow.cpp | 2 +- 25 files changed, 718 insertions(+), 356 deletions(-) (limited to 'indra/llcommon/llthreadsafequeue.h') diff --git a/indra/llcommon/llthreadsafequeue.h b/indra/llcommon/llthreadsafequeue.h index 719edcd579..06e8d8f609 100644 --- a/indra/llcommon/llthreadsafequeue.h +++ b/indra/llcommon/llthreadsafequeue.h @@ -154,6 +154,9 @@ public: // Returns the size of the queue. size_t size(); + //Returns the capacity of the queue. + U32 capacity() { return mCapacity; } + // closes the queue: // - every subsequent push() call will throw LLThreadSafeQueueInterrupt // - every subsequent tryPush() call will return false diff --git a/indra/llimage/llimageworker.cpp b/indra/llimage/llimageworker.cpp index 5f42fba866..33f8dce6ee 100644 --- a/indra/llimage/llimageworker.cpp +++ b/indra/llimage/llimageworker.cpp @@ -48,6 +48,7 @@ LLImageDecodeThread::~LLImageDecodeThread() // virtual S32 LLImageDecodeThread::update(F32 max_time_ms) { + LL_PROFILE_ZONE_SCOPED; LLMutexLock lock(mCreationMutex); for (creation_list_t::iterator iter = mCreationList.begin(); iter != mCreationList.end(); ++iter) @@ -71,6 +72,7 @@ S32 LLImageDecodeThread::update(F32 max_time_ms) LLImageDecodeThread::handle_t LLImageDecodeThread::decodeImage(LLImageFormatted* image, U32 priority, S32 discard, BOOL needs_aux, Responder* responder) { + LL_PROFILE_ZONE_SCOPED; LLMutexLock lock(mCreationMutex); handle_t handle = generateHandle(); mCreationList.push_back(creation_info(handle, image, priority, discard, needs_aux, responder)); @@ -118,6 +120,7 @@ LLImageDecodeThread::ImageRequest::~ImageRequest() // Returns true when done, whether or not decode was successful. bool LLImageDecodeThread::ImageRequest::processRequest() { + LL_PROFILE_ZONE_SCOPED; const F32 decode_time_slice = .1f; bool done = true; if (!mDecodedRaw && mFormattedImage.notNull()) @@ -164,6 +167,7 @@ bool LLImageDecodeThread::ImageRequest::processRequest() void LLImageDecodeThread::ImageRequest::finishRequest(bool completed) { + LL_PROFILE_ZONE_SCOPED; if (mResponder.notNull()) { bool success = completed && mDecodedRaw && (!mNeedsAux || mDecodedAux); diff --git a/indra/llrender/llcubemap.cpp b/indra/llrender/llcubemap.cpp index 5947bca670..d7f7b2f58e 100644 --- a/indra/llrender/llcubemap.cpp +++ b/indra/llrender/llcubemap.cpp @@ -150,6 +150,7 @@ void LLCubeMap::initRawData(const std::vector >& rawimages void LLCubeMap::initGLData() { + LL_PROFILE_ZONE_SCOPED; for (int i = 0; i < 6; i++) { mImages[i]->setSubImage(mRawImages[i], 0, 0, RESOLUTION, RESOLUTION); @@ -453,6 +454,7 @@ BOOL LLCubeMap::project(F32& v_min, F32& v_max, F32& h_min, F32& h_max, void LLCubeMap::paintIn(LLVector3 dir[4], const LLColor4U& col) { + LL_PROFILE_ZONE_SCOPED; F32 v_min, v_max, h_min, h_max; LLVector3 center = dir[0] + dir[1] + dir[2] + dir[3]; center.normVec(); diff --git a/indra/llrender/llfontfreetype.cpp b/indra/llrender/llfontfreetype.cpp index c41730ebaa..e18161e53c 100644 --- a/indra/llrender/llfontfreetype.cpp +++ b/indra/llrender/llfontfreetype.cpp @@ -460,6 +460,7 @@ LLFontGlyphInfo* LLFontFreetype::addGlyph(llwchar wch) const LLFontGlyphInfo* LLFontFreetype::addGlyphFromFont(const LLFontFreetype *fontp, llwchar wch, U32 glyph_index) const { + LL_PROFILE_ZONE_SCOPED; if (mFTFace == NULL) return NULL; diff --git a/indra/llrender/llglslshader.cpp b/indra/llrender/llglslshader.cpp index 2fb3b8257d..394fcd2b2f 100644 --- a/indra/llrender/llglslshader.cpp +++ b/indra/llrender/llglslshader.cpp @@ -208,6 +208,7 @@ void LLGLSLShader::dumpStats() //static void LLGLSLShader::startProfile() { + LL_PROFILE_ZONE_SCOPED; if (sProfileEnabled && sCurBoundShaderPtr) { sCurBoundShaderPtr->placeProfileQuery(); @@ -218,6 +219,7 @@ void LLGLSLShader::startProfile() //static void LLGLSLShader::stopProfile(U32 count, U32 mode) { + LL_PROFILE_ZONE_SCOPED; if (sProfileEnabled && sCurBoundShaderPtr) { sCurBoundShaderPtr->readProfileQuery(count, mode); diff --git a/indra/llrender/llgltexture.cpp b/indra/llrender/llgltexture.cpp index ad501687ed..a279e85bae 100644 --- a/indra/llrender/llgltexture.cpp +++ b/indra/llrender/llgltexture.cpp @@ -262,6 +262,7 @@ LLTexUnit::eTextureType LLGLTexture::getTarget(void) const BOOL LLGLTexture::setSubImage(const LLImageRaw* imageraw, S32 x_pos, S32 y_pos, S32 width, S32 height) { + LL_PROFILE_ZONE_SCOPED; llassert(mGLTexturep.notNull()) ; return mGLTexturep->setSubImage(imageraw, x_pos, y_pos, width, height) ; @@ -269,6 +270,7 @@ BOOL LLGLTexture::setSubImage(const LLImageRaw* imageraw, S32 x_pos, S32 y_pos, BOOL LLGLTexture::setSubImage(const U8* datap, S32 data_width, S32 data_height, S32 x_pos, S32 y_pos, S32 width, S32 height) { + LL_PROFILE_ZONE_SCOPED; llassert(mGLTexturep.notNull()) ; return mGLTexturep->setSubImage(datap, data_width, data_height, x_pos, y_pos, width, height) ; diff --git a/indra/llrender/llimagegl.cpp b/indra/llrender/llimagegl.cpp index 276fa55e15..aff29bd857 100644 --- a/indra/llrender/llimagegl.cpp +++ b/indra/llrender/llimagegl.cpp @@ -39,6 +39,7 @@ #include "llgl.h" #include "llglslshader.h" #include "llrender.h" +#include "llwindow.h" //---------------------------------------------------------------------------- const F32 MIN_TEXTURE_LIFETIME = 10.f; @@ -170,15 +171,32 @@ BOOL is_little_endian() return (*c == 0x78) ; } + +LLImageGLThread* LLImageGLThread::sInstance = nullptr; + //static -void LLImageGL::initClass(S32 num_catagories, BOOL skip_analyze_alpha /* = false */) +void LLImageGL::initClass(LLWindow* window, S32 num_catagories, BOOL skip_analyze_alpha /* = false */) { + LL_PROFILE_ZONE_SCOPED; sSkipAnalyzeAlpha = skip_analyze_alpha; + LLImageGLThread::sInstance = new LLImageGLThread(window); + LLImageGLThread::sInstance->start(); +} + +//static +void LLImageGL::updateClass() +{ + LL_PROFILE_ZONE_SCOPED; + LLImageGLThread::sInstance->executeCallbacks(); } //static void LLImageGL::cleanupClass() -{ +{ + LL_PROFILE_ZONE_SCOPED; + LLImageGLThread::sInstance->mFunctionQueue.close(); + delete LLImageGLThread::sInstance; + LLImageGLThread::sInstance = nullptr; } //static @@ -656,6 +674,7 @@ void LLImageGL::setExplicitFormat( LLGLint internal_format, LLGLenum primary_for void LLImageGL::setImage(const LLImageRaw* imageraw) { + LL_PROFILE_ZONE_SCOPED; llassert((imageraw->getWidth() == getWidth(mCurrentDiscardLevel)) && (imageraw->getHeight() == getHeight(mCurrentDiscardLevel)) && (imageraw->getComponents() == getComponents())); @@ -699,9 +718,8 @@ BOOL LLImageGL::setImage(const U8* data_in, BOOL data_hasmips) } llverify(gGL.getTexUnit(0)->bind(this)); - - - if (mUseMipMaps) + + if (mUseMipMaps) { if (data_hasmips) { @@ -781,7 +799,7 @@ BOOL LLImageGL::setImage(const U8* data_in, BOOL data_hasmips) glTexParameteri(mTarget, GL_GENERATE_MIPMAP, GL_TRUE); } - LLImageGL::setManualImage(mTarget, 0, mFormatInternal, + LLImageGL::setManualImage(mTarget, 0, mFormatInternal, w, h, mFormatPrimary, mFormatType, data_in, mAllowCompression); @@ -878,7 +896,7 @@ BOOL LLImageGL::setImage(const U8* data_in, BOOL data_hasmips) stop_glerror(); } - LLImageGL::setManualImage(mTarget, m, mFormatInternal, w, h, mFormatPrimary, mFormatType, cur_mip_data, mAllowCompression); + LLImageGL::setManualImage(mTarget, m, mFormatInternal, w, h, mFormatPrimary, mFormatType, cur_mip_data, mAllowCompression); if (m == 0) { analyzeAlpha(data_in, w, h); @@ -1067,6 +1085,7 @@ void LLImageGL::postAddToAtlas() BOOL LLImageGL::setSubImage(const U8* datap, S32 data_width, S32 data_height, S32 x_pos, S32 y_pos, S32 width, S32 height, BOOL force_fast_update) { + LL_PROFILE_ZONE_SCOPED; if (!width || !height) { return TRUE; @@ -1163,6 +1182,7 @@ BOOL LLImageGL::setSubImage(const U8* datap, S32 data_width, S32 data_height, S3 BOOL LLImageGL::setSubImage(const LLImageRaw* imageraw, S32 x_pos, S32 y_pos, S32 width, S32 height, BOOL force_fast_update) { + LL_PROFILE_ZONE_SCOPED; return setSubImage(imageraw->getData(), imageraw->getWidth(), imageraw->getHeight(), x_pos, y_pos, width, height, force_fast_update); } @@ -1201,119 +1221,119 @@ void LLImageGL::deleteTextures(S32 numTextures, U32 *textures) // static static LLTrace::BlockTimerStatHandle FTM_SET_MANUAL_IMAGE("setManualImage"); -void LLImageGL::setManualImage(U32 target, S32 miplevel, S32 intformat, S32 width, S32 height, U32 pixformat, U32 pixtype, const void *pixels, bool allow_compression) +void LLImageGL::setManualImage(U32 target, S32 miplevel, S32 intformat, S32 width, S32 height, U32 pixformat, U32 pixtype, const void* pixels, bool allow_compression) { - LL_RECORD_BLOCK_TIME(FTM_SET_MANUAL_IMAGE); - bool use_scratch = false; - U32* scratch = NULL; - if (LLRender::sGLCoreProfile) - { - if (pixformat == GL_ALPHA && pixtype == GL_UNSIGNED_BYTE) - { //GL_ALPHA is deprecated, convert to RGBA - use_scratch = true; - scratch = new U32[width*height]; + LL_RECORD_BLOCK_TIME(FTM_SET_MANUAL_IMAGE); + bool use_scratch = false; + U32* scratch = NULL; + if (LLRender::sGLCoreProfile) + { + if (pixformat == GL_ALPHA && pixtype == GL_UNSIGNED_BYTE) + { //GL_ALPHA is deprecated, convert to RGBA + use_scratch = true; + scratch = new U32[width * height]; - U32 pixel_count = (U32) (width*height); - for (U32 i = 0; i < pixel_count; i++) - { - U8* pix = (U8*) &scratch[i]; - pix[0] = pix[1] = pix[2] = 0; - pix[3] = ((U8*) pixels)[i]; - } - - pixformat = GL_RGBA; - intformat = GL_RGBA8; - } + U32 pixel_count = (U32)(width * height); + for (U32 i = 0; i < pixel_count; i++) + { + U8* pix = (U8*)&scratch[i]; + pix[0] = pix[1] = pix[2] = 0; + pix[3] = ((U8*)pixels)[i]; + } - if (pixformat == GL_LUMINANCE_ALPHA && pixtype == GL_UNSIGNED_BYTE) - { //GL_LUMINANCE_ALPHA is deprecated, convert to RGBA - use_scratch = true; - scratch = new U32[width*height]; + pixformat = GL_RGBA; + intformat = GL_RGBA8; + } - U32 pixel_count = (U32) (width*height); - for (U32 i = 0; i < pixel_count; i++) - { - U8 lum = ((U8*) pixels)[i*2+0]; - U8 alpha = ((U8*) pixels)[i*2+1]; + if (pixformat == GL_LUMINANCE_ALPHA && pixtype == GL_UNSIGNED_BYTE) + { //GL_LUMINANCE_ALPHA is deprecated, convert to RGBA + use_scratch = true; + scratch = new U32[width * height]; - U8* pix = (U8*) &scratch[i]; - pix[0] = pix[1] = pix[2] = lum; - pix[3] = alpha; - } - - pixformat = GL_RGBA; - intformat = GL_RGBA8; - } + U32 pixel_count = (U32)(width * height); + for (U32 i = 0; i < pixel_count; i++) + { + U8 lum = ((U8*)pixels)[i * 2 + 0]; + U8 alpha = ((U8*)pixels)[i * 2 + 1]; - if (pixformat == GL_LUMINANCE && pixtype == GL_UNSIGNED_BYTE) - { //GL_LUMINANCE_ALPHA is deprecated, convert to RGB - use_scratch = true; - scratch = new U32[width*height]; + U8* pix = (U8*)&scratch[i]; + pix[0] = pix[1] = pix[2] = lum; + pix[3] = alpha; + } - U32 pixel_count = (U32) (width*height); - for (U32 i = 0; i < pixel_count; i++) - { - U8 lum = ((U8*) pixels)[i]; - - U8* pix = (U8*) &scratch[i]; - pix[0] = pix[1] = pix[2] = lum; - pix[3] = 255; - } - - pixformat = GL_RGBA; - intformat = GL_RGB8; - } - } + pixformat = GL_RGBA; + intformat = GL_RGBA8; + } - if (LLImageGL::sCompressTextures && allow_compression) - { - switch (intformat) - { - case GL_RGB: - case GL_RGB8: - intformat = GL_COMPRESSED_RGB; - break; - case GL_SRGB: - case GL_SRGB8: - intformat = GL_COMPRESSED_SRGB; - break; - case GL_RGBA: - case GL_RGBA8: - intformat = GL_COMPRESSED_RGBA; - break; - case GL_SRGB_ALPHA: - case GL_SRGB8_ALPHA8: - intformat = GL_COMPRESSED_SRGB_ALPHA; - break; - case GL_LUMINANCE: - case GL_LUMINANCE8: - intformat = GL_COMPRESSED_LUMINANCE; - break; - case GL_LUMINANCE_ALPHA: - case GL_LUMINANCE8_ALPHA8: - intformat = GL_COMPRESSED_LUMINANCE_ALPHA; - break; - case GL_ALPHA: - case GL_ALPHA8: - intformat = GL_COMPRESSED_ALPHA; - break; - default: - LL_WARNS() << "Could not compress format: " << std::hex << intformat << LL_ENDL; - break; - } - } + if (pixformat == GL_LUMINANCE && pixtype == GL_UNSIGNED_BYTE) + { //GL_LUMINANCE_ALPHA is deprecated, convert to RGB + use_scratch = true; + scratch = new U32[width * height]; - stop_glerror(); - { - LL_PROFILE_ZONE_NAMED("glTexImage2D"); - glTexImage2D(target, miplevel, intformat, width, height, 0, pixformat, pixtype, use_scratch ? scratch : pixels); - } - stop_glerror(); + U32 pixel_count = (U32)(width * height); + for (U32 i = 0; i < pixel_count; i++) + { + U8 lum = ((U8*)pixels)[i]; - if (use_scratch) - { - delete [] scratch; - } + U8* pix = (U8*)&scratch[i]; + pix[0] = pix[1] = pix[2] = lum; + pix[3] = 255; + } + + pixformat = GL_RGBA; + intformat = GL_RGB8; + } + } + + if (LLImageGL::sCompressTextures && allow_compression) + { + switch (intformat) + { + case GL_RGB: + case GL_RGB8: + intformat = GL_COMPRESSED_RGB; + break; + case GL_SRGB: + case GL_SRGB8: + intformat = GL_COMPRESSED_SRGB; + break; + case GL_RGBA: + case GL_RGBA8: + intformat = GL_COMPRESSED_RGBA; + break; + case GL_SRGB_ALPHA: + case GL_SRGB8_ALPHA8: + intformat = GL_COMPRESSED_SRGB_ALPHA; + break; + case GL_LUMINANCE: + case GL_LUMINANCE8: + intformat = GL_COMPRESSED_LUMINANCE; + break; + case GL_LUMINANCE_ALPHA: + case GL_LUMINANCE8_ALPHA8: + intformat = GL_COMPRESSED_LUMINANCE_ALPHA; + break; + case GL_ALPHA: + case GL_ALPHA8: + intformat = GL_COMPRESSED_ALPHA; + break; + default: + LL_WARNS() << "Could not compress format: " << std::hex << intformat << LL_ENDL; + break; + } + } + + stop_glerror(); + { + LL_PROFILE_ZONE_NAMED("glTexImage2D"); + glTexImage2D(target, miplevel, intformat, width, height, 0, pixformat, pixtype, use_scratch ? scratch : pixels); + } + stop_glerror(); + + if (use_scratch) + { + delete[] scratch; + } } //create an empty GL texture: just create a texture name @@ -1336,6 +1356,7 @@ BOOL LLImageGL::createGLTexture() if(mTexName) { LLImageGL::deleteTextures(1, (reinterpret_cast(&mTexName))) ; + mTexName = 0; } @@ -1697,7 +1718,7 @@ void LLImageGL::destroyGLTexture() mTextureMemory = (S32Bytes)0; } - LLImageGL::deleteTextures(1, &mTexName); + LLImageGL::deleteTextures(1, &mTexName); mCurrentDiscardLevel = -1 ; //invalidate mCurrentDiscardLevel. mTexName = 0; mGLTextureCreated = FALSE ; @@ -2238,3 +2259,90 @@ void LLImageGL::resetCurTexSizebar() glTexParameteri(GL_TEXTURE_2D, GL_TEXTURE_BASE_LEVEL, 0); glTexParameteri(GL_TEXTURE_2D, GL_TEXTURE_MAX_LEVEL, nummips); */ + +LLImageGLThread::LLImageGLThread(LLWindow* window) + : LLThread("LLImageGL"), mWindow(window) +{ + mFinished = false; + + mContext = mWindow->createSharedContext(); +} + +// post a function to be executed on the LLImageGL background thread + +bool LLImageGLThread::post(const std::function& func) +{ + try + { + if (mFunctionQueue.size() < mFunctionQueue.capacity()) + { + //NOTE: tryPushFront will return immediately if the lock is held + // desired behavior here is to push and return true unless the + // queue is full or closed + mFunctionQueue.pushFront(func); + } + else + { + return false; + } + } + catch (LLThreadSafeQueueInterrupt e) + { + return false; + } + + return true; +} + +//post a callback to be executed on the main thread + +bool LLImageGLThread::postCallback(const std::function& callback) +{ + try + { + mCallbackQueue.pushFront(callback); + } + catch (LLThreadSafeQueueInterrupt e) + { + //thread is closing, drop request + return false; + } + + return true; +} + +void LLImageGLThread::executeCallbacks() +{ + LL_PROFILE_ZONE_SCOPED; + //executed from main thread + std::function callback; + while (mCallbackQueue.tryPopBack(callback)) + { + LL_PROFILE_ZONE_NAMED("iglt - callback"); + callback(); + } +} + +void LLImageGLThread::run() +{ + mWindow->makeContextCurrent(mContext); + gGL.init(); + try + { + while (true) + { + LL_PROFILE_ZONE_SCOPED; + std::function curFunc = mFunctionQueue.popBack(); + { + LL_PROFILE_ZONE_NAMED("iglt - function") + curFunc(); + } + } + } + catch (LLThreadSafeQueueInterrupt e) + { + //queue is closed, fall out of run loop + } + gGL.shutdown(); + mWindow->destroySharedContext(mContext); +} diff --git a/indra/llrender/llimagegl.h b/indra/llrender/llimagegl.h index 61ddc8d59b..8e9b483c2d 100644 --- a/indra/llrender/llimagegl.h +++ b/indra/llrender/llimagegl.h @@ -35,9 +35,11 @@ #include "llrefcount.h" #include "v2math.h" #include "llunits.h" - +#include "llthreadsafequeue.h" #include "llrender.h" class LLTextureAtlas ; +class LLWindow; + #define BYTES_TO_MEGA_BYTES(x) ((x) >> 20) #define MEGA_BYTES_TO_BYTES(x) ((x) << 20) @@ -102,7 +104,7 @@ public: void setAllowCompression(bool allow) { mAllowCompression = allow; } static void setManualImage(U32 target, S32 miplevel, S32 intformat, S32 width, S32 height, U32 pixformat, U32 pixtype, const void *pixels, bool allow_compression = true); - + BOOL createGLTexture() ; BOOL createGLTexture(S32 discard_level, const LLImageRaw* imageraw, S32 usename = 0, BOOL to_create = TRUE, S32 category = sMaxCategories-1); @@ -265,7 +267,8 @@ public: #endif public: - static void initClass(S32 num_catagories, BOOL skip_analyze_alpha = false); + static void initClass(LLWindow* window, S32 num_catagories, BOOL skip_analyze_alpha = false); + static void updateClass(); static void cleanupClass() ; private: @@ -301,4 +304,30 @@ public: }; +class LLImageGLThread : public LLThread +{ +public: + LLImageGLThread(LLWindow* window); + + // post a function to be executed on the LLImageGL background thread + bool post(const std::function& func); + + //post a callback to be executed on the main thread + bool postCallback(const std::function& callback); + + void executeCallbacks(); + + void run() override; + + LLThreadSafeQueue> mFunctionQueue; + LLThreadSafeQueue> mCallbackQueue; + + LLWindow* mWindow; + void* mContext; + LLAtomicBool mFinished; + + static LLImageGLThread* sInstance; +}; + + #endif // LL_LLIMAGEGL_H diff --git a/indra/llrender/llrender.cpp b/indra/llrender/llrender.cpp index 43b4441ea8..b6711e44e3 100644 --- a/indra/llrender/llrender.cpp +++ b/indra/llrender/llrender.cpp @@ -36,7 +36,7 @@ #include "lltexture.h" #include "llshadermgr.h" -LLRender gGL; +thread_local LLRender gGL; // Handy copies of last good GL matrices F32 gGLModelView[16]; diff --git a/indra/llrender/llrender.h b/indra/llrender/llrender.h index af8568f8a3..c08c2d6881 100644 --- a/indra/llrender/llrender.h +++ b/indra/llrender/llrender.h @@ -511,7 +511,7 @@ extern F32 gGLLastProjection[16]; extern F32 gGLProjection[16]; extern S32 gGLViewport[4]; -extern LLRender gGL; +extern thread_local LLRender gGL; // This rotation matrix moves the default OpenGL reference frame // (-Z at, Y up) to Cory's favorite reference frame (X at, Z up) diff --git a/indra/llwindow/llopenglview-objc.mm b/indra/llwindow/llopenglview-objc.mm index d2c5b11c3d..b647085b7e 100644 --- a/indra/llwindow/llopenglview-objc.mm +++ b/indra/llwindow/llopenglview-objc.mm @@ -288,7 +288,8 @@ attributedStringInfo getSegments(NSAttributedString *str) if (vsync) { - [glContext setValues:(const GLint*)1 forParameter:NSOpenGLCPSwapInterval]; + GLint value = 1; + [glContext setValues:&value forParameter:NSOpenGLCPSwapInterval]; } else { // supress this error after move to Xcode 7: // error: null passed to a callee that requires a non-null argument [-Werror,-Wnonnull] diff --git a/indra/llwindow/llwindow.h b/indra/llwindow/llwindow.h index d4d5b76937..10c0b6a424 100644 --- a/indra/llwindow/llwindow.h +++ b/indra/llwindow/llwindow.h @@ -78,7 +78,17 @@ public: BOOL setSize(LLCoordWindow size); virtual void setMinSize(U32 min_width, U32 min_height, bool enforce_immediately = true); virtual BOOL switchContext(BOOL fullscreen, const LLCoordScreen &size, BOOL disable_vsync, const LLCoordScreen * const posp = NULL) = 0; - virtual BOOL setCursorPosition(LLCoordWindow position) = 0; + + //create a new GL context that shares a namespace with this Window's main GL context and make it current on the current thread + // returns a pointer to be handed back to destroySharedConext/makeContextCurrent + virtual void* createSharedContext() = 0; + //make the given context current on the current thread + virtual void makeContextCurrent(void* context) = 0; + //destroy the given context that was retrieved by createSharedContext() + //Must be called on the same thread that called createSharedContext() + virtual void destroySharedContext(void* context) = 0; + + virtual BOOL setCursorPosition(LLCoordWindow position) = 0; virtual BOOL getCursorPosition(LLCoordWindow *position) = 0; virtual void showCursor() = 0; virtual void hideCursor() = 0; diff --git a/indra/llwindow/llwindowheadless.h b/indra/llwindow/llwindowheadless.h index c692666df1..a7ae28aa24 100644 --- a/indra/llwindow/llwindowheadless.h +++ b/indra/llwindow/llwindowheadless.h @@ -49,6 +49,9 @@ public: /*virtual*/ BOOL setSizeImpl(LLCoordScreen size) {return FALSE;}; /*virtual*/ BOOL setSizeImpl(LLCoordWindow size) {return FALSE;}; /*virtual*/ BOOL switchContext(BOOL fullscreen, const LLCoordScreen &size, BOOL disable_vsync, const LLCoordScreen * const posp = NULL) {return FALSE;}; + void* createSharedContext() { return nullptr; } + void makeContextCurrent(void*) {} + void destroySharedContext(void*) {} /*virtual*/ BOOL setCursorPosition(LLCoordWindow position) {return FALSE;}; /*virtual*/ BOOL getCursorPosition(LLCoordWindow *position) {return FALSE;}; /*virtual*/ void showCursor() {}; diff --git a/indra/llwindow/llwindowmacosx.cpp b/indra/llwindow/llwindowmacosx.cpp index 0d0607a0bb..23830dd24e 100644 --- a/indra/llwindow/llwindowmacosx.cpp +++ b/indra/llwindow/llwindowmacosx.cpp @@ -1907,6 +1907,34 @@ void LLWindowMacOSX::allowLanguageTextInput(LLPreeditor *preeditor, BOOL b) allowDirectMarkedTextInput(b, mGLView); // mLanguageTextInputAllowed and mMarkedTextAllowed should be updated at once (by Pell Smit } +class sharedContext +{ +public: + CGLContextObj mContext; +}; + +void* LLWindowMacOSX::createSharedContext() +{ + sharedContext* sc = new sharedContext(); + CGLCreateContext(mPixelFormat, mContext, &(sc->mContext)); + + return (void *)sc; +} + +void LLWindowMacOSX::makeContextCurrent(void* context) +{ + CGLSetCurrentContext(((sharedContext*)context)->mContext); +} + +void LLWindowMacOSX::destroySharedContext(void* context) +{ + sharedContext* sc = (sharedContext*)context; + + CGLDestroyContext(sc->mContext); + + delete sc; +} + void LLWindowMacOSX::interruptLanguageTextInput() { commitCurrentPreedit(mGLView); diff --git a/indra/llwindow/llwindowmacosx.h b/indra/llwindow/llwindowmacosx.h index bf45238c8d..ede2b453d5 100644 --- a/indra/llwindow/llwindowmacosx.h +++ b/indra/llwindow/llwindowmacosx.h @@ -41,85 +41,84 @@ #undef verify #undef require - class LLWindowMacOSX : public LLWindow { public: - /*virtual*/ void show(); - /*virtual*/ void hide(); - /*virtual*/ void close(); - /*virtual*/ BOOL getVisible(); - /*virtual*/ BOOL getMinimized(); - /*virtual*/ BOOL getMaximized(); - /*virtual*/ BOOL maximize(); - /*virtual*/ void minimize(); - /*virtual*/ void restore(); - /*virtual*/ BOOL getFullscreen(); - /*virtual*/ BOOL getPosition(LLCoordScreen *position); - /*virtual*/ BOOL getSize(LLCoordScreen *size); - /*virtual*/ BOOL getSize(LLCoordWindow *size); - /*virtual*/ BOOL setPosition(LLCoordScreen position); - /*virtual*/ BOOL setSizeImpl(LLCoordScreen size); - /*virtual*/ BOOL setSizeImpl(LLCoordWindow size); - /*virtual*/ BOOL switchContext(BOOL fullscreen, const LLCoordScreen &size, BOOL disable_vsync, const LLCoordScreen * const posp = NULL); - /*virtual*/ BOOL setCursorPosition(LLCoordWindow position); - /*virtual*/ BOOL getCursorPosition(LLCoordWindow *position); - /*virtual*/ void showCursor(); - /*virtual*/ void hideCursor(); - /*virtual*/ void showCursorFromMouseMove(); - /*virtual*/ void hideCursorUntilMouseMove(); - /*virtual*/ BOOL isCursorHidden(); - /*virtual*/ void updateCursor(); - /*virtual*/ ECursorType getCursor() const; - /*virtual*/ void captureMouse(); - /*virtual*/ void releaseMouse(); - /*virtual*/ void setMouseClipping( BOOL b ); - /*virtual*/ BOOL isClipboardTextAvailable(); - /*virtual*/ BOOL pasteTextFromClipboard(LLWString &dst); - /*virtual*/ BOOL copyTextToClipboard(const LLWString & src); - /*virtual*/ void flashIcon(F32 seconds); - /*virtual*/ F32 getGamma(); - /*virtual*/ BOOL setGamma(const F32 gamma); // Set the gamma - /*virtual*/ U32 getFSAASamples(); - /*virtual*/ void setFSAASamples(const U32 fsaa_samples); - /*virtual*/ BOOL restoreGamma(); // Restore original gamma table (before updating gamma) - /*virtual*/ ESwapMethod getSwapMethod() { return mSwapMethod; } - /*virtual*/ void gatherInput(); - /*virtual*/ void delayInputProcessing() {}; - /*virtual*/ void swapBuffers(); + void show() override; + void hide() override; + void close() override; + BOOL getVisible() override; + BOOL getMinimized() override; + BOOL getMaximized() override; + BOOL maximize() override; + void minimize() override; + void restore() override; + BOOL getFullscreen(); + BOOL getPosition(LLCoordScreen *position) override; + BOOL getSize(LLCoordScreen *size) override; + BOOL getSize(LLCoordWindow *size) override; + BOOL setPosition(LLCoordScreen position) override; + BOOL setSizeImpl(LLCoordScreen size) override; + BOOL setSizeImpl(LLCoordWindow size) override; + BOOL switchContext(BOOL fullscreen, const LLCoordScreen &size, BOOL disable_vsync, const LLCoordScreen * const posp = NULL) override; + BOOL setCursorPosition(LLCoordWindow position) override; + BOOL getCursorPosition(LLCoordWindow *position) override; + void showCursor() override; + void hideCursor() override; + void showCursorFromMouseMove() override; + void hideCursorUntilMouseMove() override; + BOOL isCursorHidden() override; + void updateCursor() override; + ECursorType getCursor() const override; + void captureMouse() override; + void releaseMouse() override; + void setMouseClipping( BOOL b ) override; + BOOL isClipboardTextAvailable() override; + BOOL pasteTextFromClipboard(LLWString &dst) override; + BOOL copyTextToClipboard(const LLWString & src) override; + void flashIcon(F32 seconds) override; + F32 getGamma() override; + BOOL setGamma(const F32 gamma) override; // Set the gamma + U32 getFSAASamples() override; + void setFSAASamples(const U32 fsaa_samples) override; + BOOL restoreGamma() override; // Restore original gamma table (before updating gamma) + ESwapMethod getSwapMethod() override { return mSwapMethod; } + void gatherInput() override; + void delayInputProcessing() override {}; + void swapBuffers() override; // handy coordinate space conversion routines - /*virtual*/ BOOL convertCoords(LLCoordScreen from, LLCoordWindow *to); - /*virtual*/ BOOL convertCoords(LLCoordWindow from, LLCoordScreen *to); - /*virtual*/ BOOL convertCoords(LLCoordWindow from, LLCoordGL *to); - /*virtual*/ BOOL convertCoords(LLCoordGL from, LLCoordWindow *to); - /*virtual*/ BOOL convertCoords(LLCoordScreen from, LLCoordGL *to); - /*virtual*/ BOOL convertCoords(LLCoordGL from, LLCoordScreen *to); + BOOL convertCoords(LLCoordScreen from, LLCoordWindow *to) override; + BOOL convertCoords(LLCoordWindow from, LLCoordScreen *to) override; + BOOL convertCoords(LLCoordWindow from, LLCoordGL *to) override; + BOOL convertCoords(LLCoordGL from, LLCoordWindow *to) override; + BOOL convertCoords(LLCoordScreen from, LLCoordGL *to) override; + BOOL convertCoords(LLCoordGL from, LLCoordScreen *to) override; - /*virtual*/ LLWindowResolution* getSupportedResolutions(S32 &num_resolutions); - /*virtual*/ F32 getNativeAspectRatio(); - /*virtual*/ F32 getPixelAspectRatio(); - /*virtual*/ void setNativeAspectRatio(F32 ratio) { mOverrideAspectRatio = ratio; } + LLWindowResolution* getSupportedResolutions(S32 &num_resolutions) override; + F32 getNativeAspectRatio() override; + F32 getPixelAspectRatio() override; + void setNativeAspectRatio(F32 ratio) override { mOverrideAspectRatio = ratio; } - /*virtual*/ void beforeDialog(); - /*virtual*/ void afterDialog(); + void beforeDialog() override; + void afterDialog() override; - /*virtual*/ BOOL dialogColorPicker(F32 *r, F32 *g, F32 *b); + BOOL dialogColorPicker(F32 *r, F32 *g, F32 *b) override; - /*virtual*/ void *getPlatformWindow(); - /*virtual*/ void bringToFront() {}; + void *getPlatformWindow() override; + void bringToFront() override {}; - /*virtual*/ void allowLanguageTextInput(LLPreeditor *preeditor, BOOL b); - /*virtual*/ void interruptLanguageTextInput(); - /*virtual*/ void spawnWebBrowser(const std::string& escaped_url, bool async); - /*virtual*/ F32 getSystemUISize(); + void allowLanguageTextInput(LLPreeditor *preeditor, BOOL b) override; + void interruptLanguageTextInput() override; + void spawnWebBrowser(const std::string& escaped_url, bool async) override; + F32 getSystemUISize() override; static std::vector getDisplaysResolutionList(); static std::vector getDynamicFallbackFontList(); // Provide native key event data - /*virtual*/ LLSD getNativeKeyData(); + LLSD getNativeKeyData() override; void* getWindow() { return mWindow; } LLWindowCallbacks* getCallbacks() { return mCallbacks; } @@ -132,6 +131,15 @@ public: bool allowsLanguageInput() { return mLanguageTextInputAllowed; } + //create a new GL context that shares a namespace with this Window's main GL context and make it current on the current thread + // returns a pointer to be handed back to destroySharedConext/makeContextCurrent + void* createSharedContext() override; + //make the given context current on the current thread + void makeContextCurrent(void* context) override; + //destroy the given context that was retrieved by createSharedContext() + //Must be called on the same thread that called createSharedContext() + void destroySharedContext(void* context) override; + protected: LLWindowMacOSX(LLWindowCallbacks* callbacks, const std::string& title, const std::string& name, int x, int y, int width, int height, U32 flags, @@ -141,7 +149,7 @@ protected: ~LLWindowMacOSX(); void initCursors(); - BOOL isValid(); + BOOL isValid() override; void moveWindow(const LLCoordScreen& position,const LLCoordScreen& size); @@ -157,7 +165,7 @@ protected: BOOL shouldPostQuit() { return mPostQuit; } //Satisfy MAINT-3135 and MAINT-3288 with a flag. - /*virtual */ void setOldResize(bool oldresize) {setResizeMode(oldresize, mGLView); } + /*virtual */ void setOldResize(bool oldresize) override {setResizeMode(oldresize, mGLView); } private: void restoreGLContext(); @@ -231,9 +239,9 @@ public: LLSplashScreenMacOSX(); virtual ~LLSplashScreenMacOSX(); - /*virtual*/ void showImpl(); - /*virtual*/ void updateImpl(const std::string& mesg); - /*virtual*/ void hideImpl(); + void showImpl(); + void updateImpl(const std::string& mesg); + void hideImpl(); private: WindowRef mWindow; diff --git a/indra/llwindow/llwindowwin32.cpp b/indra/llwindow/llwindowwin32.cpp index 4ce7c30bef..12d4c6c30e 100644 --- a/indra/llwindow/llwindowwin32.cpp +++ b/indra/llwindow/llwindowwin32.cpp @@ -1774,58 +1774,11 @@ const S32 max_format = (S32)num_formats - 1; mhRC = 0; if (wglCreateContextAttribsARB) { //attempt to create a specific versioned context - S32 attribs[] = - { //start at 4.2 - WGL_CONTEXT_MAJOR_VERSION_ARB, 4, - WGL_CONTEXT_MINOR_VERSION_ARB, 2, - WGL_CONTEXT_PROFILE_MASK_ARB, LLRender::sGLCoreProfile ? WGL_CONTEXT_CORE_PROFILE_BIT_ARB : WGL_CONTEXT_COMPATIBILITY_PROFILE_BIT_ARB, - WGL_CONTEXT_FLAGS_ARB, gDebugGL ? WGL_CONTEXT_DEBUG_BIT_ARB : 0, - 0 - }; - - bool done = false; - while (!done) - { - mhRC = wglCreateContextAttribsARB(mhDC, mhRC, attribs); - - if (!mhRC) - { - if (attribs[3] > 0) - { //decrement minor version - attribs[3]--; - } - else if (attribs[1] > 3) - { //decrement major version and start minor version over at 3 - attribs[1]--; - attribs[3] = 3; - } - else - { //we reached 3.0 and still failed, bail out - done = true; - } - } - else - { - LL_INFOS() << "Created OpenGL " << llformat("%d.%d", attribs[1], attribs[3]) << - (LLRender::sGLCoreProfile ? " core" : " compatibility") << " context." << LL_ENDL; - done = true; - - // force sNoFixedFunction iff we're trying to use nsight debugging which does not support many legacy API uses - - // nSight doesn't support use of legacy API funcs in the fixed function pipe - if (LLRender::sGLCoreProfile || LLRender::sNsightDebugSupport) - { - LLGLSLShader::sNoFixedFunction = true; - } - } - } - } - - if (!mhRC && !(mhRC = wglCreateContext(mhDC))) - { - close(); - OSMessageBox(mCallbacks->translateString("MBGLContextErr"), mCallbacks->translateString("MBError"), OSMB_OK); - return FALSE; + mhRC = (HGLRC) createSharedContext(); + if (!mhRC) + { + return FALSE; + } } if (!wglMakeCurrent(mhDC, mhRC)) @@ -1880,6 +1833,75 @@ const S32 max_format = (S32)num_formats - 1; return TRUE; } +void* LLWindowWin32::createSharedContext() +{ + S32 attribs[] = + { + WGL_CONTEXT_MAJOR_VERSION_ARB, 4, + WGL_CONTEXT_MINOR_VERSION_ARB, 2, + WGL_CONTEXT_PROFILE_MASK_ARB, LLRender::sGLCoreProfile ? WGL_CONTEXT_CORE_PROFILE_BIT_ARB : WGL_CONTEXT_COMPATIBILITY_PROFILE_BIT_ARB, + WGL_CONTEXT_FLAGS_ARB, gDebugGL ? WGL_CONTEXT_DEBUG_BIT_ARB : 0, + 0 + }; + + HGLRC rc = wglCreateContextAttribsARB(mhDC, mhRC, attribs); + + bool done = false; + while (!done) + { + rc = wglCreateContextAttribsARB(mhDC, mhRC, attribs); + + if (!rc) + { + if (attribs[3] > 0) + { //decrement minor version + attribs[3]--; + } + else if (attribs[1] > 3) + { //decrement major version and start minor version over at 3 + attribs[1]--; + attribs[3] = 3; + } + else + { //we reached 3.0 and still failed, bail out + done = true; + } + } + else + { + LL_INFOS() << "Created OpenGL " << llformat("%d.%d", attribs[1], attribs[3]) << + (LLRender::sGLCoreProfile ? " core" : " compatibility") << " context." << LL_ENDL; + done = true; + + // force sNoFixedFunction iff we're trying to use nsight debugging which does not support many legacy API uses + + // nSight doesn't support use of legacy API funcs in the fixed function pipe + if (LLRender::sGLCoreProfile || LLRender::sNsightDebugSupport) + { + LLGLSLShader::sNoFixedFunction = true; + } + } + } + + if (!rc && !(rc = wglCreateContext(mhDC))) + { + close(); + OSMessageBox(mCallbacks->translateString("MBGLContextErr"), mCallbacks->translateString("MBError"), OSMB_OK); + } + + return rc; +} + +void LLWindowWin32::makeContextCurrent(void* contextPtr) +{ + wglMakeCurrent(mhDC, (HGLRC) contextPtr); +} + +void LLWindowWin32::destroySharedContext(void* contextPtr) +{ + wglDeleteContext((HGLRC)contextPtr); +} + void LLWindowWin32::moveWindow( const LLCoordScreen& position, const LLCoordScreen& size ) { if( mIsMouseClipping ) @@ -3720,6 +3742,7 @@ BOOL LLWindowWin32::resetDisplayResolution() void LLWindowWin32::swapBuffers() { + LL_PROFILE_ZONE_SCOPED; ASSERT_MAIN_THREAD(); SwapBuffers(mhDC); @@ -4731,4 +4754,4 @@ void LLWindowWin32::post(const std::function& func) #else mFunctionQueue.pushFront(func); #endif -} \ No newline at end of file +} diff --git a/indra/llwindow/llwindowwin32.h b/indra/llwindow/llwindowwin32.h index 66647459b2..5f253b5df3 100644 --- a/indra/llwindow/llwindowwin32.h +++ b/indra/llwindow/llwindowwin32.h @@ -93,6 +93,9 @@ public: /*virtual*/ BOOL setSizeImpl(LLCoordScreen size); /*virtual*/ BOOL setSizeImpl(LLCoordWindow size); /*virtual*/ BOOL switchContext(BOOL fullscreen, const LLCoordScreen &size, BOOL disable_vsync, const LLCoordScreen * const posp = NULL); + void* createSharedContext() override; + void makeContextCurrent(void* context) override; + void destroySharedContext(void* context) override; /*virtual*/ BOOL setCursorPosition(LLCoordWindow position); /*virtual*/ BOOL getCursorPosition(LLCoordWindow *position); /*virtual*/ void showCursor(); diff --git a/indra/newview/app_settings/settings.xml b/indra/newview/app_settings/settings.xml index 05c3fc3bfe..d05d100ddf 100644 --- a/indra/newview/app_settings/settings.xml +++ b/indra/newview/app_settings/settings.xml @@ -9814,7 +9814,7 @@ Type Boolean Value - 0 + 1 RenderGlow @@ -10599,7 +10599,7 @@ Type Boolean Value - 0 + 1 RenderUseTransformFeedback diff --git a/indra/newview/llappviewer.cpp b/indra/newview/llappviewer.cpp index 722a6caa65..e46e8feb14 100644 --- a/indra/newview/llappviewer.cpp +++ b/indra/newview/llappviewer.cpp @@ -613,7 +613,7 @@ static void settings_modify() LLPipeline::sRenderDeferred = LLPipeline::sRenderTransparentWater && LLPipeline::sRenderBump && gSavedSettings.getBOOL("RenderDeferred"); LLVOSurfacePatch::sLODFactor = gSavedSettings.getF32("RenderTerrainLODFactor"); LLVOSurfacePatch::sLODFactor *= LLVOSurfacePatch::sLODFactor; //square lod factor to get exponential range of [1,4] - gDebugGL = gSavedSettings.getBOOL("RenderDebugGL") || gDebugSession; + gDebugGL = gSavedSettings.getBOOL("RenderDebugGL") || gDebugSession; gDebugPipeline = gSavedSettings.getBOOL("RenderDebugPipeline"); } @@ -4831,6 +4831,7 @@ void LLAppViewer::idle() LLNotificationsUI::LLToast::updateClass(); LLSmoothInterpolation::updateInterpolants(); LLMortician::updateClass(); + LLImageGL::updateClass(); LLFilePickerThread::clearDead(); //calls LLFilePickerThread::notify() LLDirPickerThread::clearDead(); F32 dt_raw = idle_timer.getElapsedTimeAndResetF32(); diff --git a/indra/newview/llnetmap.cpp b/indra/newview/llnetmap.cpp index 112da55682..111b45612e 100644 --- a/indra/newview/llnetmap.cpp +++ b/indra/newview/llnetmap.cpp @@ -147,6 +147,7 @@ void LLNetMap::setScale( F32 scale ) void LLNetMap::draw() { + LL_PROFILE_ZONE_SCOPED; static LLFrameTimer map_timer; static LLUIColor map_avatar_color = LLUIColorTable::instance().getColor("MapAvatarColor", LLColor4::white); static LLUIColor map_avatar_friend_color = LLUIColorTable::instance().getColor("MapAvatarFriendColor", LLColor4::white); diff --git a/indra/newview/lltexturefetch.cpp b/indra/newview/lltexturefetch.cpp index f64db7beb5..63e561147d 100644 --- a/indra/newview/lltexturefetch.cpp +++ b/indra/newview/lltexturefetch.cpp @@ -314,6 +314,7 @@ private: // Threads: Ttc virtual void completed(bool success) { + LL_PROFILE_ZONE_SCOPED; LLTextureFetchWorker* worker = mFetcher->getWorker(mID); if (worker) { @@ -338,6 +339,7 @@ private: // Threads: Ttc virtual void completed(bool success) { + LL_PROFILE_ZONE_SCOPED; LLTextureFetchWorker* worker = mFetcher->getWorker(mID); if (worker) { @@ -362,6 +364,7 @@ private: // Threads: Tid virtual void completed(bool success, LLImageRaw* raw, LLImageRaw* aux) { + LL_PROFILE_ZONE_SCOPED; LLTextureFetchWorker* worker = mFetcher->getWorker(mID); if (worker) { @@ -1139,6 +1142,7 @@ void LLTextureFetchWorker::startWork(S32 param) // Threads: Ttf bool LLTextureFetchWorker::doWork(S32 param) { + LL_PROFILE_ZONE_SCOPED; static const LLCore::HttpStatus http_not_found(HTTP_NOT_FOUND); // 404 static const LLCore::HttpStatus http_service_unavail(HTTP_SERVICE_UNAVAILABLE); // 503 static const LLCore::HttpStatus http_not_sat(HTTP_REQUESTED_RANGE_NOT_SATISFIABLE); // 416; diff --git a/indra/newview/llviewertexture.cpp b/indra/newview/llviewertexture.cpp index af55f68cd2..274f53a160 100644 --- a/indra/newview/llviewertexture.cpp +++ b/indra/newview/llviewertexture.cpp @@ -184,6 +184,7 @@ void LLViewerTextureManager::findFetchedTextures(const LLUUID& id, std::vector &output) { + LL_PROFILE_ZONE_SCOPED; std::vector fetched_output; gTextureList.findTexturesByID(id, fetched_output); std::vector::iterator iter = fetched_output.begin(); @@ -485,6 +486,7 @@ static LLTrace::BlockTimerStatHandle FTM_TEXTURE_MEMORY_CHECK("Memory Check"); //static bool LLViewerTexture::isMemoryForTextureLow() { + LL_PROFILE_ZONE_SCOPED; // Note: we need to figure out a better source for 'min' values, // what is free for low end at minimal settings is 'nothing left' // for higher end gpus at high settings. @@ -501,6 +503,7 @@ bool LLViewerTexture::isMemoryForTextureLow() //static bool LLViewerTexture::isMemoryForTextureSuficientlyFree() { + LL_PROFILE_ZONE_SCOPED; const S32Megabytes DESIRED_FREE_TEXTURE_MEMORY(50); const S32Megabytes DESIRED_FREE_MAIN_MEMORY(200); @@ -514,6 +517,7 @@ bool LLViewerTexture::isMemoryForTextureSuficientlyFree() //static void LLViewerTexture::getGPUMemoryForTextures(S32Megabytes &gpu, S32Megabytes &physical) { + LL_PROFILE_ZONE_SCOPED; static LLFrameTimer timer; static S32Megabytes gpu_res = S32Megabytes(S32_MAX); static S32Megabytes physical_res = S32Megabytes(S32_MAX); @@ -526,27 +530,29 @@ void LLViewerTexture::getGPUMemoryForTextures(S32Megabytes &gpu, S32Megabytes &p } timer.reset(); - LL_RECORD_BLOCK_TIME(FTM_TEXTURE_MEMORY_CHECK); - - if (gGLManager.mHasATIMemInfo) { - S32 meminfo[4]; - glGetIntegerv(GL_TEXTURE_FREE_MEMORY_ATI, meminfo); - gpu_res = (S32Megabytes)meminfo[0]; + LL_RECORD_BLOCK_TIME(FTM_TEXTURE_MEMORY_CHECK); - //check main memory, only works for windows. - LLMemory::updateMemoryInfo(); - physical_res = LLMemory::getAvailableMemKB(); - } - else if (gGLManager.mHasNVXMemInfo) - { - S32 free_memory; - glGetIntegerv(GL_GPU_MEMORY_INFO_CURRENT_AVAILABLE_VIDMEM_NVX, &free_memory); - gpu_res = (S32Megabytes)(free_memory / 1024); - } + if (gGLManager.mHasATIMemInfo) + { + S32 meminfo[4]; + glGetIntegerv(GL_TEXTURE_FREE_MEMORY_ATI, meminfo); + gpu_res = (S32Megabytes)meminfo[0]; - gpu = gpu_res; - physical = physical_res; + //check main memory, only works for windows. + LLMemory::updateMemoryInfo(); + physical_res = LLMemory::getAvailableMemKB(); + } + else if (gGLManager.mHasNVXMemInfo) + { + S32 free_memory; + glGetIntegerv(GL_GPU_MEMORY_INFO_CURRENT_AVAILABLE_VIDMEM_NVX, &free_memory); + gpu_res = (S32Megabytes)(free_memory / 1024); + } + + gpu = gpu_res; + physical = physical_res; + } } static LLTrace::BlockTimerStatHandle FTM_TEXTURE_UPDATE_MEDIA("Media"); @@ -555,6 +561,7 @@ static LLTrace::BlockTimerStatHandle FTM_TEXTURE_UPDATE_TEST("Test"); //static void LLViewerTexture::updateClass(const F32 velocity, const F32 angular_velocity) { + LL_PROFILE_ZONE_SCOPED; sCurrentTime = gFrameTimeSeconds; LLTexturePipelineTester* tester = (LLTexturePipelineTester*)LLMetricPerformanceTesterBasic::getTester(sTesterName); @@ -704,6 +711,7 @@ void LLViewerTexture::cleanup() void LLViewerTexture::notifyAboutCreatingTexture() { + LL_PROFILE_ZONE_SCOPED; for(U32 ch = 0; ch < LLRender::NUM_TEXTURE_CHANNELS; ++ch) { for(U32 f = 0; f < mNumFaces[ch]; f++) @@ -715,6 +723,7 @@ void LLViewerTexture::notifyAboutCreatingTexture() void LLViewerTexture::notifyAboutMissingAsset() { + LL_PROFILE_ZONE_SCOPED; for(U32 ch = 0; ch < LLRender::NUM_TEXTURE_CHANNELS; ++ch) { for(U32 f = 0; f < mNumFaces[ch]; f++) @@ -727,6 +736,7 @@ void LLViewerTexture::notifyAboutMissingAsset() // virtual void LLViewerTexture::dump() { + LL_PROFILE_ZONE_SCOPED; LLGLTexture::dump(); LL_INFOS() << "LLViewerTexture" @@ -762,6 +772,7 @@ bool LLViewerTexture::isActiveFetching() bool LLViewerTexture::bindDebugImage(const S32 stage) { + LL_PROFILE_ZONE_SCOPED; if (stage < 0) return false; bool res = true; @@ -780,6 +791,7 @@ bool LLViewerTexture::bindDebugImage(const S32 stage) bool LLViewerTexture::bindDefaultImage(S32 stage) { + LL_PROFILE_ZONE_SCOPED; if (stage < 0) return false; bool res = true; @@ -822,6 +834,7 @@ void LLViewerTexture::forceImmediateUpdate() void LLViewerTexture::addTextureStats(F32 virtual_size, BOOL needs_gltexture) const { + LL_PROFILE_ZONE_SCOPED; if(needs_gltexture) { mNeedsGLTexture = TRUE; @@ -864,6 +877,7 @@ void LLViewerTexture::setKnownDrawSize(S32 width, S32 height) //virtual void LLViewerTexture::addFace(U32 ch, LLFace* facep) { + LL_PROFILE_ZONE_SCOPED; llassert(ch < LLRender::NUM_TEXTURE_CHANNELS); if(mNumFaces[ch] >= mFaceList[ch].size()) @@ -879,6 +893,7 @@ void LLViewerTexture::addFace(U32 ch, LLFace* facep) //virtual void LLViewerTexture::removeFace(U32 ch, LLFace* facep) { + LL_PROFILE_ZONE_SCOPED; llassert(ch < LLRender::NUM_TEXTURE_CHANNELS); if(mNumFaces[ch] > 1) @@ -919,6 +934,7 @@ S32 LLViewerTexture::getNumFaces(U32 ch) const //virtual void LLViewerTexture::addVolume(U32 ch, LLVOVolume* volumep) { + LL_PROFILE_ZONE_SCOPED; if (mNumVolumes[ch] >= mVolumeList[ch].size()) { mVolumeList[ch].resize(2 * mNumVolumes[ch] + 1); @@ -932,6 +948,7 @@ void LLViewerTexture::addVolume(U32 ch, LLVOVolume* volumep) //virtual void LLViewerTexture::removeVolume(U32 ch, LLVOVolume* volumep) { + LL_PROFILE_ZONE_SCOPED; if (mNumVolumes[ch] > 1) { S32 index = volumep->getIndexInTex(ch); @@ -955,6 +972,7 @@ S32 LLViewerTexture::getNumVolumes(U32 ch) const void LLViewerTexture::reorganizeFaceList() { + LL_PROFILE_ZONE_SCOPED; static const F32 MAX_WAIT_TIME = 20.f; // seconds static const U32 MAX_EXTRA_BUFFER_SIZE = 4; @@ -978,6 +996,7 @@ void LLViewerTexture::reorganizeFaceList() void LLViewerTexture::reorganizeVolumeList() { + LL_PROFILE_ZONE_SCOPED; static const F32 MAX_WAIT_TIME = 20.f; // seconds static const U32 MAX_EXTRA_BUFFER_SIZE = 4; @@ -1180,6 +1199,7 @@ FTType LLViewerFetchedTexture::getFTType() const void LLViewerFetchedTexture::cleanup() { + LL_PROFILE_ZONE_SCOPED; for(callback_list_t::iterator iter = mLoadedCallbackList.begin(); iter != mLoadedCallbackList.end(); ) { @@ -1205,6 +1225,7 @@ void LLViewerFetchedTexture::cleanup() //access the fast cache void LLViewerFetchedTexture::loadFromFastCache() { + LL_PROFILE_ZONE_SCOPED; if(!mInFastCacheList) { return; //no need to access the fast cache. @@ -1350,6 +1371,7 @@ void LLViewerFetchedTexture::dump() // ONLY called from LLViewerFetchedTextureList void LLViewerFetchedTexture::destroyTexture() { + LL_PROFILE_ZONE_SCOPED; if(LLImageGL::sGlobalTextureMemory < sMaxDesiredTextureMem * 0.95f)//not ready to release unused memory. { return ; @@ -1366,6 +1388,7 @@ void LLViewerFetchedTexture::destroyTexture() void LLViewerFetchedTexture::addToCreateTexture() { + LL_PROFILE_ZONE_SCOPED; bool force_update = false; if (getComponents() != mRawImage->getComponents()) { @@ -1407,6 +1430,7 @@ void LLViewerFetchedTexture::addToCreateTexture() } else { + LL_PROFILE_ZONE_SCOPED; #if 1 // //if mRequestedDiscardLevel > mDesiredDiscardLevel, we assume the required image res keep going up, @@ -1451,99 +1475,100 @@ void LLViewerFetchedTexture::addToCreateTexture() } } #endif - mNeedsCreateTexture = TRUE; - gTextureList.mCreateTextureList.insert(this); - } + scheduleCreateTexture(); + } return; } // ONLY called from LLViewerTextureList -BOOL LLViewerFetchedTexture::createTexture(S32 usename/*= 0*/) +BOOL LLViewerFetchedTexture::preCreateTexture(S32 usename/*= 0*/) { - if (!mNeedsCreateTexture) - { - destroyRawImage(); - return FALSE; - } - mNeedsCreateTexture = FALSE; - if (mRawImage.isNull()) - { - LL_ERRS() << "LLViewerTexture trying to create texture with no Raw Image" << LL_ENDL; - } - if (mRawImage->isBufferInvalid()) - { - LL_WARNS() << "Can't create a texture: invalid image data" << LL_ENDL; - destroyRawImage(); - return FALSE; - } -// LL_INFOS() << llformat("IMAGE Creating (%d) [%d x %d] Bytes: %d ", -// mRawDiscardLevel, -// mRawImage->getWidth(), mRawImage->getHeight(),mRawImage->getDataSize()) -// << mID.getString() << LL_ENDL; - BOOL res = TRUE; + LL_PROFILE_ZONE_SCOPED; + if (!mNeedsCreateTexture) + { + destroyRawImage(); + return FALSE; + } + mNeedsCreateTexture = FALSE; - // store original size only for locally-sourced images - if (mUrl.compare(0, 7, "file://") == 0) - { - mOrigWidth = mRawImage->getWidth(); - mOrigHeight = mRawImage->getHeight(); + if (mRawImage.isNull()) + { + LL_ERRS() << "LLViewerTexture trying to create texture with no Raw Image" << LL_ENDL; + } + if (mRawImage->isBufferInvalid()) + { + LL_WARNS() << "Can't create a texture: invalid image data" << LL_ENDL; + destroyRawImage(); + return FALSE; + } + // LL_INFOS() << llformat("IMAGE Creating (%d) [%d x %d] Bytes: %d ", + // mRawDiscardLevel, + // mRawImage->getWidth(), mRawImage->getHeight(),mRawImage->getDataSize()) + // << mID.getString() << LL_ENDL; + BOOL res = TRUE; + + // store original size only for locally-sourced images + if (mUrl.compare(0, 7, "file://") == 0) + { + mOrigWidth = mRawImage->getWidth(); + mOrigHeight = mRawImage->getHeight(); // This is only safe because it's a local image and fetcher doesn't use raw data // from local images, but this might become unsafe in case of changes to fetcher - if (mBoostLevel == BOOST_PREVIEW) - { - mRawImage->biasedScaleToPowerOfTwo(1024); - } - else - { // leave black border, do not scale image content - mRawImage->expandToPowerOfTwo(MAX_IMAGE_SIZE, FALSE); - } - - mFullWidth = mRawImage->getWidth(); - mFullHeight = mRawImage->getHeight(); - setTexelsPerImage(); - } - else - { - mOrigWidth = mFullWidth; - mOrigHeight = mFullHeight; - } + if (mBoostLevel == BOOST_PREVIEW) + { + mRawImage->biasedScaleToPowerOfTwo(1024); + } + else + { // leave black border, do not scale image content + mRawImage->expandToPowerOfTwo(MAX_IMAGE_SIZE, FALSE); + } - bool size_okay = true; + mFullWidth = mRawImage->getWidth(); + mFullHeight = mRawImage->getHeight(); + setTexelsPerImage(); + } + else + { + mOrigWidth = mFullWidth; + mOrigHeight = mFullHeight; + } - S32 discard_level = mRawDiscardLevel; - if (mRawDiscardLevel < 0) - { - LL_DEBUGS() << "Negative raw discard level when creating image: " << mRawDiscardLevel << LL_ENDL; - discard_level = 0; - } + bool size_okay = true; - U32 raw_width = mRawImage->getWidth() << discard_level; - U32 raw_height = mRawImage->getHeight() << discard_level; + S32 discard_level = mRawDiscardLevel; + if (mRawDiscardLevel < 0) + { + LL_DEBUGS() << "Negative raw discard level when creating image: " << mRawDiscardLevel << LL_ENDL; + discard_level = 0; + } - if( raw_width > MAX_IMAGE_SIZE || raw_height > MAX_IMAGE_SIZE ) - { - LL_INFOS() << "Width or height is greater than " << MAX_IMAGE_SIZE << ": (" << raw_width << "," << raw_height << ")" << LL_ENDL; - size_okay = false; - } - - if (!LLImageGL::checkSize(mRawImage->getWidth(), mRawImage->getHeight())) - { - // A non power-of-two image was uploaded (through a non standard client) - LL_INFOS() << "Non power of two width or height: (" << mRawImage->getWidth() << "," << mRawImage->getHeight() << ")" << LL_ENDL; - size_okay = false; - } - - if( !size_okay ) - { - // An inappropriately-sized image was uploaded (through a non standard client) - // We treat these images as missing assets which causes them to - // be renderd as 'missing image' and to stop requesting data - LL_WARNS() << "!size_ok, setting as missing" << LL_ENDL; - setIsMissingAsset(); - destroyRawImage(); - return FALSE; - } + U32 raw_width = mRawImage->getWidth() << discard_level; + U32 raw_height = mRawImage->getHeight() << discard_level; + + if (raw_width > MAX_IMAGE_SIZE || raw_height > MAX_IMAGE_SIZE) + { + LL_INFOS() << "Width or height is greater than " << MAX_IMAGE_SIZE << ": (" << raw_width << "," << raw_height << ")" << LL_ENDL; + size_okay = false; + } + + if (!LLImageGL::checkSize(mRawImage->getWidth(), mRawImage->getHeight())) + { + // A non power-of-two image was uploaded (through a non standard client) + LL_INFOS() << "Non power of two width or height: (" << mRawImage->getWidth() << "," << mRawImage->getHeight() << ")" << LL_ENDL; + size_okay = false; + } + + if (!size_okay) + { + // An inappropriately-sized image was uploaded (through a non standard client) + // We treat these images as missing assets which causes them to + // be renderd as 'missing image' and to stop requesting data + LL_WARNS() << "!size_ok, setting as missing" << LL_ENDL; + setIsMissingAsset(); + destroyRawImage(); + return FALSE; + } if (mGLTexturep->getHasExplicitFormat()) { @@ -1565,19 +1590,79 @@ BOOL LLViewerFetchedTexture::createTexture(S32 usename/*= 0*/) } } - res = mGLTexturep->createGLTexture(mRawDiscardLevel, mRawImage, usename, TRUE, mBoostLevel); + return res; +} - notifyAboutCreatingTexture(); +BOOL LLViewerFetchedTexture::createTexture(S32 usename/*= 0*/) +{ + if (!mNeedsCreateTexture) + { + return FALSE; + } - setActive(); + BOOL res = mGLTexturep->createGLTexture(mRawDiscardLevel, mRawImage, usename, TRUE, mBoostLevel); + + return res; +} - if (!needsToSaveRawImage()) - { - mNeedsAux = FALSE; - destroyRawImage(); - } +void LLViewerFetchedTexture::postCreateTexture() +{ + if (!mNeedsCreateTexture) + { + return; + } - return res; + notifyAboutCreatingTexture(); + + setActive(); + + if (!needsToSaveRawImage()) + { + mNeedsAux = FALSE; + destroyRawImage(); + } + + mNeedsCreateTexture = FALSE; +} + +void LLViewerFetchedTexture::scheduleCreateTexture() +{ + ref(); + mNeedsCreateTexture = TRUE; + if (preCreateTexture()) + { + mNeedsCreateTexture = TRUE; +#if LL_WINDOWS //flip to 0 to revert to single-threaded OpenGL texture uploads + if (!LLImageGLThread::sInstance->post([this]() + { + //actually create the texture on a background thread + createTexture(); + { + LL_PROFILE_ZONE_NAMED("iglt - sync"); + if (gGLManager.mHasSync) + { + auto sync = glFenceSync(GL_SYNC_GPU_COMMANDS_COMPLETE, 0); + glClientWaitSync(sync, 0, 0); + glDeleteSync(sync); + } + else + { + glFinish(); + } + } + LLImageGLThread::sInstance->postCallback([this]() + { + //finalize on main thread + postCreateTexture(); + unref(); + }); + })) +#endif + { + gTextureList.mCreateTextureList.insert(this); + unref(); + } + } } // Call with 0,0 to turn this feature off. @@ -1869,6 +1954,7 @@ void LLViewerFetchedTexture::setAdditionalDecodePriority(F32 priority) void LLViewerFetchedTexture::updateVirtualSize() { + LL_PROFILE_ZONE_SCOPED; if(!mMaxVirtualSizeResetCounter) { addTextureStats(0.f, FALSE);//reset @@ -1960,6 +2046,7 @@ bool LLViewerFetchedTexture::isActiveFetching() bool LLViewerFetchedTexture::updateFetch() { + LL_PROFILE_ZONE_SCOPED; static LLCachedControl textures_decode_disabled(gSavedSettings,"TextureDecodeDisabled", false); static LLCachedControl sCameraMotionThreshold(gSavedSettings,"TextureCameraMotionThreshold", 0.2); static LLCachedControl sCameraMotionBoost(gSavedSettings,"TextureCameraMotionBoost", 3); @@ -2062,7 +2149,7 @@ bool LLViewerFetchedTexture::updateFetch() } else { - mIsRawImageValid = TRUE; + mIsRawImageValid = TRUE; addToCreateTexture(); } @@ -2890,6 +2977,7 @@ void LLViewerFetchedTexture::destroyRawImage() //virtual void LLViewerFetchedTexture::switchToCachedImage() { + LL_PROFILE_ZONE_SCOPED; if(mCachedRawImage.notNull()) { mRawImage = mCachedRawImage; @@ -2901,12 +2989,12 @@ void LLViewerFetchedTexture::switchToCachedImage() mComponents = mRawImage->getComponents(); mGLTexturep->setComponents(mComponents); gTextureList.dirtyImage(this); - } + } mIsRawImageValid = TRUE; mRawDiscardLevel = mCachedRawDiscardLevel; - gTextureList.mCreateTextureList.insert(this); - mNeedsCreateTexture = TRUE; + + scheduleCreateTexture(); } } @@ -3180,6 +3268,7 @@ bool LLViewerLODTexture::isUpdateFrozen() //virtual void LLViewerLODTexture::processTextureStats() { + LL_PROFILE_ZONE_SCOPED; updateVirtualSize(); static LLCachedControl textures_fullres(gSavedSettings,"TextureLoadFullRes", false); diff --git a/indra/newview/llviewertexture.h b/indra/newview/llviewertexture.h index 69568cc825..a5a1fb2c16 100644 --- a/indra/newview/llviewertexture.h +++ b/indra/newview/llviewertexture.h @@ -321,9 +321,13 @@ public: void addToCreateTexture(); - - // ONLY call from LLViewerTextureList + //call to determine if createTexture is necessary + BOOL preCreateTexture(S32 usename = 0); + // ONLY call from LLViewerTextureList or ImageGL background thread BOOL createTexture(S32 usename = 0); + void postCreateTexture(); + void scheduleCreateTexture(); + void destroyTexture() ; virtual void processTextureStats() ; diff --git a/indra/newview/llviewertexturelist.cpp b/indra/newview/llviewertexturelist.cpp index 12495078e9..db740b69e9 100644 --- a/indra/newview/llviewertexturelist.cpp +++ b/indra/newview/llviewertexturelist.cpp @@ -114,6 +114,7 @@ void LLViewerTextureList::init() void LLViewerTextureList::doPreloadImages() { + LL_PROFILE_ZONE_SCOPED; LL_DEBUGS("ViewerImages") << "Preloading images..." << LL_ENDL; llassert_always(mInitialized) ; @@ -205,6 +206,7 @@ static std::string get_texture_list_name() void LLViewerTextureList::doPrefetchImages() { + LL_PROFILE_ZONE_SCOPED; if (LLAppViewer::instance()->getPurgeCache()) { // cache was purged, no point @@ -258,6 +260,7 @@ LLViewerTextureList::~LLViewerTextureList() void LLViewerTextureList::shutdown() { + LL_PROFILE_ZONE_SCOPED; // clear out preloads mImagePreloads.clear(); @@ -333,6 +336,7 @@ void LLViewerTextureList::shutdown() void LLViewerTextureList::dump() { + LL_PROFILE_ZONE_SCOPED; LL_INFOS() << "LLViewerTextureList::dump()" << LL_ENDL; for (image_priority_list_t::iterator it = mImageList.begin(); it != mImageList.end(); ++it) { @@ -377,6 +381,7 @@ LLViewerFetchedTexture* LLViewerTextureList::getImageFromFile(const std::string& LLGLenum primary_format, const LLUUID& force_id) { + LL_PROFILE_ZONE_SCOPED; if(!mInitialized) { return NULL ; @@ -404,6 +409,7 @@ LLViewerFetchedTexture* LLViewerTextureList::getImageFromUrl(const std::string& LLGLenum primary_format, const LLUUID& force_id) { + LL_PROFILE_ZONE_SCOPED; if(!mInitialized) { return NULL ; @@ -492,6 +498,7 @@ LLViewerFetchedTexture* LLViewerTextureList::getImage(const LLUUID &image_id, LLGLenum primary_format, LLHost request_from_host) { + LL_PROFILE_ZONE_SCOPED; if(!mInitialized) { return NULL ; @@ -554,6 +561,7 @@ LLViewerFetchedTexture* LLViewerTextureList::createImage(const LLUUID &image_id, LLGLenum primary_format, LLHost request_from_host) { + LL_PROFILE_ZONE_SCOPED; static LLCachedControl fast_cache_fetching_enabled(gSavedSettings, "FastCacheFetchEnabled", true); LLPointer imagep ; @@ -609,6 +617,7 @@ LLViewerFetchedTexture* LLViewerTextureList::createImage(const LLUUID &image_id, void LLViewerTextureList::findTexturesByID(const LLUUID &image_id, std::vector &output) { + LL_PROFILE_ZONE_SCOPED; LLTextureKey search_key(image_id, TEX_LIST_STANDARD); uuid_map_t::iterator iter = mUUIDMap.lower_bound(search_key); while (iter != mUUIDMap.end() && iter->first.textureId == image_id) @@ -634,6 +643,7 @@ LLViewerFetchedTexture *LLViewerTextureList::findImage(const LLUUID &image_id, E void LLViewerTextureList::addImageToList(LLViewerFetchedTexture *image) { + LL_PROFILE_ZONE_SCOPED; assert_main_thread(); llassert_always(mInitialized) ; llassert(image); @@ -653,6 +663,7 @@ void LLViewerTextureList::addImageToList(LLViewerFetchedTexture *image) void LLViewerTextureList::removeImageFromList(LLViewerFetchedTexture *image) { + LL_PROFILE_ZONE_SCOPED; assert_main_thread(); llassert_always(mInitialized) ; llassert(image); @@ -701,6 +712,7 @@ void LLViewerTextureList::removeImageFromList(LLViewerFetchedTexture *image) void LLViewerTextureList::addImage(LLViewerFetchedTexture *new_image, ETexListType tex_type) { + LL_PROFILE_ZONE_SCOPED; if (!new_image) { return; @@ -724,6 +736,7 @@ void LLViewerTextureList::addImage(LLViewerFetchedTexture *new_image, ETexListTy void LLViewerTextureList::deleteImage(LLViewerFetchedTexture *image) { + LL_PROFILE_ZONE_SCOPED; if( image) { if (image->hasCallbacks()) @@ -845,6 +858,7 @@ void LLViewerTextureList::updateImages(F32 max_time) void LLViewerTextureList::clearFetchingRequests() { + LL_PROFILE_ZONE_SCOPED; if (LLAppViewer::getTextureFetch()->getNumRequests() == 0) { return; @@ -862,6 +876,7 @@ void LLViewerTextureList::clearFetchingRequests() void LLViewerTextureList::updateImagesDecodePriorities() { + LL_PROFILE_ZONE_SCOPED; // Update the decode priority for N images each frame { F32 lazy_flush_timeout = 30.f; // stop decoding @@ -977,6 +992,7 @@ void LLViewerTextureList::updateImagesDecodePriorities() void LLViewerTextureList::setDebugFetching(LLViewerFetchedTexture* tex, S32 debug_level) { + LL_PROFILE_ZONE_SCOPED; if(!tex->setDebugFetching(debug_level)) { return; @@ -1025,6 +1041,7 @@ void LLViewerTextureList::setDebugFetching(LLViewerFetchedTexture* tex, S32 debu F32 LLViewerTextureList::updateImagesCreateTextures(F32 max_time) { + LL_PROFILE_ZONE_SCOPED; if (gGLManager.mIsDisabled) return 0.0f; // @@ -1041,6 +1058,7 @@ F32 LLViewerTextureList::updateImagesCreateTextures(F32 max_time) enditer = iter; LLViewerFetchedTexture *imagep = *curiter; imagep->createTexture(); + imagep->postCreateTexture(); if (create_timer.getElapsedTimeF32() > max_time) { break; @@ -1052,6 +1070,7 @@ F32 LLViewerTextureList::updateImagesCreateTextures(F32 max_time) F32 LLViewerTextureList::updateImagesLoadingFastCache(F32 max_time) { + LL_PROFILE_ZONE_SCOPED; if (gGLManager.mIsDisabled) return 0.0f; if(mFastCacheList.empty()) { @@ -1082,6 +1101,7 @@ F32 LLViewerTextureList::updateImagesLoadingFastCache(F32 max_time) void LLViewerTextureList::forceImmediateUpdate(LLViewerFetchedTexture* imagep) { + LL_PROFILE_ZONE_SCOPED; if(!imagep) { return ; @@ -1101,6 +1121,7 @@ void LLViewerTextureList::forceImmediateUpdate(LLViewerFetchedTexture* imagep) F32 LLViewerTextureList::updateImagesFetchTextures(F32 max_time) { + LL_PROFILE_ZONE_SCOPED; LLTimer image_op_timer; // Update fetch for N images each frame @@ -1176,6 +1197,7 @@ F32 LLViewerTextureList::updateImagesFetchTextures(F32 max_time) void LLViewerTextureList::updateImagesUpdateStats() { + LL_PROFILE_ZONE_SCOPED; if (mForceResetTextureStats) { for (image_priority_list_t::iterator iter = mImageList.begin(); @@ -1190,6 +1212,7 @@ void LLViewerTextureList::updateImagesUpdateStats() void LLViewerTextureList::decodeAllImages(F32 max_time) { + LL_PROFILE_ZONE_SCOPED; LLTimer timer; //loading from fast cache @@ -1259,6 +1282,7 @@ BOOL LLViewerTextureList::createUploadFile(const std::string& filename, const std::string& out_filename, const U8 codec) { + LL_PROFILE_ZONE_SCOPED; // Load the image LLPointer image = LLImageFormatted::createFromType(codec); if (image.isNull()) @@ -1312,6 +1336,7 @@ BOOL LLViewerTextureList::createUploadFile(const std::string& filename, // note: modifies the argument raw_image!!!! LLPointer LLViewerTextureList::convertToUploadFile(LLPointer raw_image) { + LL_PROFILE_ZONE_SCOPED; raw_image->biasedScaleToPowerOfTwo(LLViewerFetchedTexture::MAX_IMAGE_SIZE_DEFAULT); LLPointer compressedImage = new LLImageJ2C(); @@ -1345,6 +1370,7 @@ LLPointer LLViewerTextureList::convertToUploadFile(LLPointer U32Megabytes(1500)) ? S32Megabytes(64) : gMinVideoRam ; @@ -1354,6 +1380,7 @@ S32Megabytes LLViewerTextureList::getMinVideoRamSetting() // Returns max setting for TextureMemory (in MB) S32Megabytes LLViewerTextureList::getMaxVideoRamSetting(bool get_recommended, float mem_multiplier) { + LL_PROFILE_ZONE_SCOPED; S32Megabytes max_texmem; if (gGLManager.mVRAM != 0) { @@ -1407,6 +1434,7 @@ const S32Megabytes VIDEO_CARD_FRAMEBUFFER_MEM(12); const S32Megabytes MIN_MEM_FOR_NON_TEXTURE(512); void LLViewerTextureList::updateMaxResidentTexMem(S32Megabytes mem) { + LL_PROFILE_ZONE_SCOPED; // Initialize the image pipeline VRAM settings S32Megabytes cur_mem(gSavedSettings.getS32("TextureMemory")); F32 mem_multiplier = gSavedSettings.getF32("RenderTextureMemoryMultiple"); @@ -1647,6 +1675,7 @@ void LLUIImageList::cleanUp() LLUIImagePtr LLUIImageList::getUIImageByID(const LLUUID& image_id, S32 priority) { + LL_PROFILE_ZONE_SCOPED; // use id as image name std::string image_name = image_id.asString(); @@ -1665,6 +1694,7 @@ LLUIImagePtr LLUIImageList::getUIImageByID(const LLUUID& image_id, S32 priority) LLUIImagePtr LLUIImageList::getUIImage(const std::string& image_name, S32 priority) { + LL_PROFILE_ZONE_SCOPED; // look for existing image uuid_ui_image_map_t::iterator found_it = mUIImages.find(image_name); if (found_it != mUIImages.end()) @@ -1682,6 +1712,7 @@ LLUIImagePtr LLUIImageList::loadUIImageByName(const std::string& name, const std BOOL use_mips, const LLRect& scale_rect, const LLRect& clip_rect, LLViewerTexture::EBoostLevel boost_priority, LLUIImage::EScaleStyle scale_style) { + LL_PROFILE_ZONE_SCOPED; if (boost_priority == LLGLTexture::BOOST_NONE) { boost_priority = LLGLTexture::BOOST_UI; @@ -1694,6 +1725,7 @@ LLUIImagePtr LLUIImageList::loadUIImageByID(const LLUUID& id, BOOL use_mips, const LLRect& scale_rect, const LLRect& clip_rect, LLViewerTexture::EBoostLevel boost_priority, LLUIImage::EScaleStyle scale_style) { + LL_PROFILE_ZONE_SCOPED; if (boost_priority == LLGLTexture::BOOST_NONE) { boost_priority = LLGLTexture::BOOST_UI; @@ -1705,6 +1737,7 @@ LLUIImagePtr LLUIImageList::loadUIImageByID(const LLUUID& id, LLUIImagePtr LLUIImageList::loadUIImage(LLViewerFetchedTexture* imagep, const std::string& name, BOOL use_mips, const LLRect& scale_rect, const LLRect& clip_rect, LLUIImage::EScaleStyle scale_style) { + LL_PROFILE_ZONE_SCOPED; if (!imagep) return NULL; imagep->setAddressMode(LLTexUnit::TAM_CLAMP); @@ -1742,6 +1775,7 @@ LLUIImagePtr LLUIImageList::loadUIImage(LLViewerFetchedTexture* imagep, const st LLUIImagePtr LLUIImageList::preloadUIImage(const std::string& name, const std::string& filename, BOOL use_mips, const LLRect& scale_rect, const LLRect& clip_rect, LLUIImage::EScaleStyle scale_style) { + LL_PROFILE_ZONE_SCOPED; // look for existing image uuid_ui_image_map_t::iterator found_it = mUIImages.find(name); if (found_it != mUIImages.end()) @@ -1756,6 +1790,7 @@ LLUIImagePtr LLUIImageList::preloadUIImage(const std::string& name, const std::s //static void LLUIImageList::onUIImageLoaded( BOOL success, LLViewerFetchedTexture *src_vi, LLImageRaw* src, LLImageRaw* src_aux, S32 discard_level, BOOL final, void* user_data ) { + LL_PROFILE_ZONE_SCOPED; if(!success || !user_data) { return; @@ -1857,6 +1892,7 @@ struct UIImageDeclarations : public LLInitParam::Block bool LLUIImageList::initFromFile() { + LL_PROFILE_ZONE_SCOPED; // Look for textures.xml in all the right places. Pass // constraint=LLDir::ALL_SKINS because we want to overlay textures.xml // from all the skins directories. diff --git a/indra/newview/llviewerwindow.cpp b/indra/newview/llviewerwindow.cpp index c52a4e98d3..82ece85c1b 100644 --- a/indra/newview/llviewerwindow.cpp +++ b/indra/newview/llviewerwindow.cpp @@ -2011,7 +2011,7 @@ LLViewerWindow::LLViewerWindow(const Params& p) // Init the image list. Must happen after GL is initialized and before the images that // LLViewerWindow needs are requested. - LLImageGL::initClass(LLViewerTexture::MAX_GL_IMAGE_CATEGORY) ; + LLImageGL::initClass(mWindow, LLViewerTexture::MAX_GL_IMAGE_CATEGORY) ; gTextureList.init(); LLViewerTextureManager::init() ; gBumpImageList.init(); -- cgit v1.3