store context in device context cache per thread

6 files changed, 99 insertions, 16 deletions

diff --git a/include/vkvg.h b/include/vkvg.h
index c38b59f..769a330 100644
--- a/include/vkvg.h
+++ b/include/vkvg.h
@@ -89,6 +89,7 @@ extern "C" {
 #define VKVG_LOG_INFO_VBO	0x00000020
 #define VKVG_LOG_INFO_IBO	0x00000040
 #define VKVG_LOG_INFO_VAO	(VKVG_LOG_INFO_VBO|VKVG_LOG_INFO_IBO)
+#define VKVG_LOG_THREAD		0x00000080
 #define VKVG_LOG_DBG_ARRAYS	0x00001000
 #define VKVG_LOG_FULL		0xffffffff
 
@@ -553,7 +554,17 @@ void vkvg_matrix_get_scale (const vkvg_matrix_t *matrix, float *sx, float *sy);
 
 vkvg_public
 void vkvg_device_set_thread_aware (VkvgDevice dev, uint32_t thread_awayre);
-
+vkvg_public
+/**
+ * @brief Set maximum cached context count.
+ *
+ * The context cache stored destroyed contexts per thread to speed-up new context creation.
+ * To disable context cache, call this method with maxCount=0.
+ *
+ * @param dev A valid vkvg device pointer.
+ * @param maxCount The maximum count of saved contexts for fast context instanciation.
+ */
+void vkvg_device_set_context_cache_size (VkvgDevice dev, uint32_t maxCount);
 /**
  * @brief Create a new vkvg device.
  *
diff --git a/src/vkvg_context.c b/src/vkvg_context.c
index 6e5d125..dfcd76e 100644
--- a/src/vkvg_context.c
+++ b/src/vkvg_context.c
@@ -62,12 +62,17 @@ void _init_ctx (VkvgContext ctx) {
 	ctx->renderPassBeginInfo.renderArea.extent.height = ctx->pSurf->height;
 	ctx->renderPassBeginInfo.pClearValues = clearValues;
 
+	LOCK_SURFACE (ctx->pSurf)
+
 	if (ctx->pSurf->new)
 		ctx->renderPassBeginInfo.renderPass = ctx->dev->renderPass_ClearAll;
 	else
 		ctx->renderPassBeginInfo.renderPass = ctx->dev->renderPass_ClearStencil;
 
 	ctx->pSurf->new = false;
+
+	UNLOCK_SURFACE (ctx->pSurf);
+
 	vkvg_surface_reference (ctx->pSurf);
 
 	if (ctx->dev->samples == VK_SAMPLE_COUNT_1_BIT)
diff --git a/src/vkvg_context_internal.c b/src/vkvg_context_internal.c
index 415037a..2eb20d3 100644
--- a/src/vkvg_context_internal.c
+++ b/src/vkvg_context_internal.c
@@ -873,7 +873,7 @@ void _release_context_ressources (VkvgContext ctx) {
 	VkDescriptorSet dss[] = {ctx->dsFont, ctx->dsSrc, ctx->dsGrad};
 	vkFreeDescriptorSets	(dev, ctx->descriptorPool, 3, dss);
 
-	vkDestroyDescriptorPool (dev, ctx->descriptorPool,NULL);
+	vkDestroyDescriptorPool (dev, ctx->descriptorPool, NULL);
 
 	vkvg_buffer_destroy (&ctx->uboGrad);
 	vkvg_buffer_destroy (&ctx->indices);
diff --git a/src/vkvg_device.c b/src/vkvg_device.c
index 4a266cc..eb50287 100644
--- a/src/vkvg_device.c
+++ b/src/vkvg_device.c
@@ -31,6 +31,22 @@
 if (vkh_phyinfo_try_get_extension_properties(pi, #ext, NULL))	\
 	enabledExts[enabledExtsCount++] = #ext;						\
 }
+void vkvg_device_set_context_cache_size (VkvgDevice dev, uint32_t maxCount) {
+	if (maxCount == dev->cachedContextMaxCount)
+		return;
+
+	dev->cachedContextMaxCount = maxCount;
+
+	_cached_ctx* cur = dev->cachedContextLast;
+	while (cur && dev->cachedContextCount > dev->cachedContextMaxCount) {
+		_release_context_ressources (cur->ctx);
+		_cached_ctx* prev = cur;
+		cur = cur->pNext;
+		free (prev);
+		dev->cachedContextCount--;
+	}
+	dev->cachedContextLast = cur;
+}
 void _device_init (VkvgDevice dev, VkInstance inst, VkPhysicalDevice phy, VkDevice vkdev, uint32_t qFamIdx, uint32_t qIndex, VkSampleCountFlags samples, bool deferredResolve) {
 	dev->instance = inst;
 	dev->hdpi	= 72;
@@ -40,6 +56,8 @@ void _device_init (VkvgDevice dev, VkInstance inst, VkPhysicalDevice phy, VkDevi
 	dev->vkDev	= vkdev;
 	dev->phy	= phy;
 
+	dev->cachedContextMaxCount = VKVG_MAX_CACHED_CONTEXT_COUNT;
+
 #if VKVG_DBG_STATS
 	dev->debug_stats = (vkvg_debug_stats_t) {0};
 #endif
@@ -205,7 +223,7 @@ const void* vkvg_get_device_requirements (VkPhysicalDeviceFeatures* pEnabledFeat
 }
 
 
-VkvgDevice vkvg_device_create(VkSampleCountFlags samples, bool deferredResolve) {
+VkvgDevice vkvg_device_create (VkSampleCountFlags samples, bool deferredResolve) {
 	LOG(VKVG_LOG_INFO, "CREATE Device\n");
 	VkvgDevice dev = (vkvg_device*)calloc(1,sizeof(vkvg_device));
 	if (!dev) {
@@ -337,11 +355,22 @@ void vkvg_device_destroy (VkvgDevice dev)
 	}
 	UNLOCK_DEVICE
 
-	while (dev->cachedContextCount > 0)
-		_release_context_ressources (dev->cachedContext[--dev->cachedContextCount]);
+
+	if (dev->cachedContextCount > 0) {
+		_cached_ctx* cur = dev->cachedContextLast;
+		while (cur) {
+			_release_context_ressources (cur->ctx);
+			_cached_ctx* prev = cur;
+			cur = cur->pNext;
+			free (prev);
+		}
+	}
+
 
 	LOG(VKVG_LOG_INFO, "DESTROY Device\n");
 
+	vkDeviceWaitIdle (dev->vkDev);
+
 	vkh_image_destroy				(dev->emptyImg);
 
 	vkDestroyDescriptorSetLayout	(dev->vkDev, dev->dslGrad,NULL);
diff --git a/src/vkvg_device_internal.c b/src/vkvg_device_internal.c
index cbec2c4..87784c2 100644
--- a/src/vkvg_device_internal.c
+++ b/src/vkvg_device_internal.c
@@ -414,8 +414,10 @@ void _device_wait_idle (VkvgDevice dev) {
 	vkDeviceWaitIdle (dev->vkDev);
 }
 void _device_wait_and_reset_device_fence (VkvgDevice dev) {
+	LOCK_DEVICE
 	vkWaitForFences (dev->vkDev, 1, &dev->fence, VK_TRUE, UINT64_MAX);
 	_device_reset_fence(dev, dev->fence);
+	UNLOCK_DEVICE
 }
 
 void _device_destroy_fence (VkvgDevice dev, VkFence fence) {
@@ -441,14 +443,33 @@ void _device_reset_fence (VkvgDevice dev, VkFence fence){
 bool _device_try_get_cached_context (VkvgDevice dev, VkvgContext* pCtx) {
 	LOCK_DEVICE
 
-	if (dev->cachedContextCount)
-		*pCtx = dev->cachedContext[--dev->cachedContextCount];
-	else
-		*pCtx = NULL;
+	if (dev->cachedContextCount) {
+		thrd_t curThread = thrd_current ();
+		_cached_ctx* prev = NULL;
+		_cached_ctx* cur = dev->cachedContextLast;
+		while (cur) {
+			if (thrd_equal (cur->thread, curThread)) {
+				if (prev)
+					prev->pNext = cur->pNext;
+				else
+					dev->cachedContextLast = cur->pNext;
 
-	UNLOCK_DEVICE
+				dev->cachedContextCount--;
+
+				LOG(VKVG_LOG_THREAD,"get cached context: %p, thd:%lu cached ctx: %d\n", cur->ctx, cur->thread, dev->cachedContextCount);
 
-	return *pCtx != NULL;
+				*pCtx = cur->ctx;
+				free (cur);
+				UNLOCK_DEVICE
+				return true;
+			}
+			prev = cur;
+			cur = cur->pNext;
+		}
+	}
+	*pCtx = NULL;
+	UNLOCK_DEVICE
+	return false;
 }
 void _device_store_context (VkvgContext ctx) {
 	VkvgDevice dev = ctx->dev;
@@ -457,7 +478,17 @@ void _device_store_context (VkvgContext ctx) {
 
 	if (dev->gQLastFence == ctx->flushFence)
 		dev->gQLastFence = VK_NULL_HANDLE;
-	dev->cachedContext[dev->cachedContextCount++] = ctx;
+
+	_cached_ctx* cur = (_cached_ctx*)calloc(1, sizeof(_cached_ctx));
+	cur->ctx	= ctx;
+	cur->thread	= thrd_current ();
+	cur->pNext	= dev->cachedContextLast;
+
+	dev->cachedContextLast = cur;
+	dev->cachedContextCount++;
+
+	LOG(VKVG_LOG_THREAD,"store context: %p, thd:%lu cached ctx: %d\n", cur->ctx, cur->thread, dev->cachedContextCount);
+
 	ctx->references++;
 
 	UNLOCK_DEVICE
diff --git a/src/vkvg_device_internal.h b/src/vkvg_device_internal.h
index f776880..eba61a1 100644
--- a/src/vkvg_device_internal.h
+++ b/src/vkvg_device_internal.h
@@ -30,7 +30,7 @@
 #define STENCIL_CLIP_BIT	0x2
 #define STENCIL_ALL_BIT		0x3
 
-#define VKVG_MAX_CACHED_CONTEXT_COUNT 0
+#define VKVG_MAX_CACHED_CONTEXT_COUNT 2
 
 extern PFN_vkCmdBindPipeline			CmdBindPipeline;
 extern PFN_vkCmdBindDescriptorSets		CmdBindDescriptorSets;
@@ -53,7 +53,13 @@ extern PFN_vkWaitForFences				WaitForFences;
 extern PFN_vkResetFences				ResetFences;
 extern PFN_vkResetCommandBuffer			ResetCommandBuffer;
 
-typedef struct _vkvg_device_t{
+typedef struct _cached_ctx{
+	thrd_t				thread;
+	VkvgContext			ctx;
+	struct _cached_ctx*	pNext;
+} _cached_ctx;
+
+typedef struct _vkvg_device_t {
 	VkDevice				vkDev;					/**< Vulkan Logical Device */
 	VkPhysicalDeviceMemoryProperties phyMemProps;	/**< Vulkan Physical device memory properties */
 	VkPhysicalDevice		phy;					/**< Vulkan Physical device */
@@ -106,8 +112,9 @@ typedef struct _vkvg_device_t{
 
 	VkvgContext				lastCtx;				/**< last element of double linked list of context, used to trigger font caching system update on all contexts*/
 
-	int32_t					cachedContextCount;
-	VkvgContext				cachedContext[VKVG_MAX_CACHED_CONTEXT_COUNT];
+	int32_t					cachedContextMaxCount;	/**< Maximum context cache element count.*/
+	int32_t					cachedContextCount;		/**< Current context cache element count.*/
+	_cached_ctx*			cachedContextLast;		/**< Last element of single linked list of saved context for fast reuse.*/
 
 #ifdef VKVG_WIRED_DEBUG
 	VkPipeline				pipelineWired;