Cycles: Pack kernel textures into buffers for OpenCL

Image textures were being packed into a single buffer for OpenCL, which
limited the amount of memory available for images to the size of one
buffer (usually 4gb on AMD hardware). By packing textures into multiple
buffers that limit is removed, while simultaneously reducing the number
of buffers that need to be passed to each kernel.

Benchmarks were within 2%.

Fixes T51554.

Differential Revision: https://developer.blender.org/D2745

5 files changed, 7 insertions, 196 deletions

diff --git a/intern/cycles/render/image.cpp b/intern/cycles/render/image.cpp
index a490f10aee4..80ec77f8b4a 100644
--- a/intern/cycles/render/image.cpp
+++ b/intern/cycles/render/image.cpp
@@ -43,7 +43,6 @@ static bool isfinite(half /*value*/)
 ImageManager::ImageManager(const DeviceInfo& info)
 {
 	need_update = true;
-	pack_images = false;
 	osl_texture_system = NULL;
 	animation_frame = 0;
 
@@ -87,11 +86,6 @@ ImageManager::~ImageManager()
 	}
 }
 
-void ImageManager::set_pack_images(bool pack_images_)
-{
-	pack_images = pack_images_;
-}
-
 void ImageManager::set_osl_texture_system(void *texture_system)
 {
 	osl_texture_system = texture_system;
@@ -742,7 +736,7 @@ void ImageManager::device_load_image(Device *device,
 			pixels[3] = TEX_IMAGE_MISSING_A;
 		}
 
-		if(!pack_images) {
+		{
 			thread_scoped_lock device_lock(device_mutex);
 			device->tex_alloc(name.c_str(),
 			                  tex_img,
@@ -771,7 +765,7 @@ void ImageManager::device_load_image(Device *device,
 			pixels[0] = TEX_IMAGE_MISSING_R;
 		}
 
-		if(!pack_images) {
+		{
 			thread_scoped_lock device_lock(device_mutex);
 			device->tex_alloc(name.c_str(),
 			                  tex_img,
@@ -803,7 +797,7 @@ void ImageManager::device_load_image(Device *device,
 			pixels[3] = (TEX_IMAGE_MISSING_A * 255);
 		}
 
-		if(!pack_images) {
+		{
 			thread_scoped_lock device_lock(device_mutex);
 			device->tex_alloc(name.c_str(),
 			                  tex_img,
@@ -831,7 +825,7 @@ void ImageManager::device_load_image(Device *device,
 			pixels[0] = (TEX_IMAGE_MISSING_R * 255);
 		}
 
-		if(!pack_images) {
+		{
 			thread_scoped_lock device_lock(device_mutex);
 			device->tex_alloc(name.c_str(),
 			                  tex_img,
@@ -862,7 +856,7 @@ void ImageManager::device_load_image(Device *device,
 			pixels[3] = TEX_IMAGE_MISSING_A;
 		}
 
-		if(!pack_images) {
+		{
 			thread_scoped_lock device_lock(device_mutex);
 			device->tex_alloc(name.c_str(),
 			                  tex_img,
@@ -890,7 +884,7 @@ void ImageManager::device_load_image(Device *device,
 			pixels[0] = TEX_IMAGE_MISSING_R;
 		}
 
-		if(!pack_images) {
+		{
 			thread_scoped_lock device_lock(device_mutex);
 			device->tex_alloc(name.c_str(),
 			                  tex_img,
@@ -1047,9 +1041,6 @@ void ImageManager::device_update(Device *device,
 
 	pool.wait_work();
 
-	if(pack_images)
-		device_pack_images(device, dscene, progress);
-
 	need_update = false;
 }
 
@@ -1079,141 +1070,6 @@ void ImageManager::device_update_slot(Device *device,
 	}
 }
 
-uint8_t ImageManager::pack_image_options(ImageDataType type, size_t slot)
-{
-	uint8_t options = 0;
-	/* Image Options are packed into one uint:
-	 * bit 0 -> Interpolation
-	 * bit 1 + 2 + 3 -> Extension
-	 */
-	if(images[type][slot]->interpolation == INTERPOLATION_CLOSEST) {
-		options |= (1 << 0);
-	}
-	if(images[type][slot]->extension == EXTENSION_REPEAT) {
-		options |= (1 << 1);
-	}
-	else if(images[type][slot]->extension == EXTENSION_EXTEND) {
-		options |= (1 << 2);
-	}
-	else /* EXTENSION_CLIP */ {
-		options |= (1 << 3);
-	}
-	return options;
-}
-
-template<typename T>
-void ImageManager::device_pack_images_type(
-        ImageDataType type,
-        const vector<device_vector<T>*>& cpu_textures,
-        device_vector<T> *device_image,
-        uint4 *info)
-{
-	size_t size = 0, offset = 0;
-	/* First step is to calculate size of the texture we need. */
-	for(size_t slot = 0; slot < images[type].size(); slot++) {
-		if(images[type][slot] == NULL) {
-			continue;
-		}
-		device_vector<T>& tex_img = *cpu_textures[slot];
-		size += tex_img.size();
-	}
-	/* Now we know how much memory we need, so we can allocate and fill. */
-	T *pixels = device_image->resize(size);
-	for(size_t slot = 0; slot < images[type].size(); slot++) {
-		if(images[type][slot] == NULL) {
-			continue;
-		}
-		device_vector<T>& tex_img = *cpu_textures[slot];
-		uint8_t options = pack_image_options(type, slot);
-		const int index = type_index_to_flattened_slot(slot, type) * 2;
-		info[index] = make_uint4(tex_img.data_width,
-		                         tex_img.data_height,
-		                         offset,
-		                         options);
-		info[index+1] = make_uint4(tex_img.data_depth, 0, 0, 0);
-		memcpy(pixels + offset,
-		       (void*)tex_img.data_pointer,
-		       tex_img.memory_size());
-		offset += tex_img.size();
-	}
-}
-
-void ImageManager::device_pack_images(Device *device,
-                                      DeviceScene *dscene,
-                                      Progress& /*progess*/)
-{
-	/* For OpenCL, we pack all image textures into a single large texture, and
-	 * do our own interpolation in the kernel.
-	 */
-
-	/* TODO(sergey): This will over-allocate a bit, but this is constant memory
-	 * so should be fine for a short term.
-	 */
-	const size_t info_size = max4(max_flattened_slot(IMAGE_DATA_TYPE_FLOAT4),
-	                              max_flattened_slot(IMAGE_DATA_TYPE_BYTE4),
-	                              max_flattened_slot(IMAGE_DATA_TYPE_FLOAT),
-	                              max_flattened_slot(IMAGE_DATA_TYPE_BYTE));
-	uint4 *info = dscene->tex_image_packed_info.resize(info_size*2);
-
-	/* Pack byte4 textures. */
-	device_pack_images_type(IMAGE_DATA_TYPE_BYTE4,
-	                        dscene->tex_byte4_image,
-	                        &dscene->tex_image_byte4_packed,
-	                        info);
-	/* Pack float4 textures. */
-	device_pack_images_type(IMAGE_DATA_TYPE_FLOAT4,
-	                        dscene->tex_float4_image,
-	                        &dscene->tex_image_float4_packed,
-	                        info);
-	/* Pack byte textures. */
-	device_pack_images_type(IMAGE_DATA_TYPE_BYTE,
-	                        dscene->tex_byte_image,
-	                        &dscene->tex_image_byte_packed,
-	                        info);
-	/* Pack float textures. */
-	device_pack_images_type(IMAGE_DATA_TYPE_FLOAT,
-	                        dscene->tex_float_image,
-	                        &dscene->tex_image_float_packed,
-	                        info);
-
-	/* Push textures to the device. */
-	if(dscene->tex_image_byte4_packed.size()) {
-		if(dscene->tex_image_byte4_packed.device_pointer) {
-			thread_scoped_lock device_lock(device_mutex);
-			device->tex_free(dscene->tex_image_byte4_packed);
-		}
-		device->tex_alloc("__tex_image_byte4_packed", dscene->tex_image_byte4_packed);
-	}
-	if(dscene->tex_image_float4_packed.size()) {
-		if(dscene->tex_image_float4_packed.device_pointer) {
-			thread_scoped_lock device_lock(device_mutex);
-			device->tex_free(dscene->tex_image_float4_packed);
-		}
-		device->tex_alloc("__tex_image_float4_packed", dscene->tex_image_float4_packed);
-	}
-	if(dscene->tex_image_byte_packed.size()) {
-		if(dscene->tex_image_byte_packed.device_pointer) {
-			thread_scoped_lock device_lock(device_mutex);
-			device->tex_free(dscene->tex_image_byte_packed);
-		}
-		device->tex_alloc("__tex_image_byte_packed", dscene->tex_image_byte_packed);
-	}
-	if(dscene->tex_image_float_packed.size()) {
-		if(dscene->tex_image_float_packed.device_pointer) {
-			thread_scoped_lock device_lock(device_mutex);
-			device->tex_free(dscene->tex_image_float_packed);
-		}
-		device->tex_alloc("__tex_image_float_packed", dscene->tex_image_float_packed);
-	}
-	if(dscene->tex_image_packed_info.size()) {
-		if(dscene->tex_image_packed_info.device_pointer) {
-			thread_scoped_lock device_lock(device_mutex);
-			device->tex_free(dscene->tex_image_packed_info);
-		}
-		device->tex_alloc("__tex_image_packed_info", dscene->tex_image_packed_info);
-	}
-}
-
 void ImageManager::device_free_builtin(Device *device, DeviceScene *dscene)
 {
 	for(int type = 0; type < IMAGE_DATA_NUM_TYPES; type++) {
@@ -1239,18 +1095,6 @@ void ImageManager::device_free(Device *device, DeviceScene *dscene)
 	dscene->tex_float_image.clear();
 	dscene->tex_byte_image.clear();
 	dscene->tex_half_image.clear();
-
-	device->tex_free(dscene->tex_image_float4_packed);
-	device->tex_free(dscene->tex_image_byte4_packed);
-	device->tex_free(dscene->tex_image_float_packed);
-	device->tex_free(dscene->tex_image_byte_packed);
-	device->tex_free(dscene->tex_image_packed_info);
-
-	dscene->tex_image_float4_packed.clear();
-	dscene->tex_image_byte4_packed.clear();
-	dscene->tex_image_float_packed.clear();
-	dscene->tex_image_byte_packed.clear();
-	dscene->tex_image_packed_info.clear();
 }
 
 CCL_NAMESPACE_END
diff --git a/intern/cycles/render/image.h b/intern/cycles/render/image.h
index db7e28a5e44..c86d1cbedbf 100644
--- a/intern/cycles/render/image.h
+++ b/intern/cycles/render/image.h
@@ -76,7 +76,6 @@ public:
 	void device_free_builtin(Device *device, DeviceScene *dscene);
 
 	void set_osl_texture_system(void *texture_system);
-	void set_pack_images(bool pack_images_);
 	bool set_animation_frame_update(int frame);
 
 	bool need_update;
@@ -130,7 +129,6 @@ private:
 
 	vector<Image*> images[IMAGE_DATA_NUM_TYPES];
 	void *osl_texture_system;
-	bool pack_images;
 
 	bool file_load_image_generic(Image *img,
 	                             ImageInput **in,
@@ -152,8 +150,6 @@ private:
 	int flattened_slot_to_type_index(int flat_slot, ImageDataType *type);
 	string name_from_type(int type);
 
-	uint8_t pack_image_options(ImageDataType type, size_t slot);
-
 	void device_load_image(Device *device,
 	                       DeviceScene *dscene,
 	                       Scene *scene,
@@ -164,17 +160,6 @@ private:
 	                       DeviceScene *dscene,
 	                       ImageDataType type,
 	                       int slot);
-
-	template<typename T>
-	void device_pack_images_type(
-	        ImageDataType type,
-	        const vector<device_vector<T>*>& cpu_textures,
-	        device_vector<T> *device_image,
-	        uint4 *info);
-
-	void device_pack_images(Device *device,
-	                        DeviceScene *dscene,
-	                        Progress& progess);
 };
 
 CCL_NAMESPACE_END
diff --git a/intern/cycles/render/mesh.cpp b/intern/cycles/render/mesh.cpp
index 03825f780e0..84537bf5993 100644
--- a/intern/cycles/render/mesh.cpp
+++ b/intern/cycles/render/mesh.cpp
@@ -1925,16 +1925,7 @@ void MeshManager::device_update_displacement_images(Device *device,
 					if(node->special_type != SHADER_SPECIAL_TYPE_IMAGE_SLOT) {
 						continue;
 					}
-					if(device->info.pack_images) {
-						/* If device requires packed images we need to update all
-						 * images now, even if they're not used for displacement.
-						 */
-						image_manager->device_update(device,
-						                             dscene,
-						                             scene,
-						                             progress);
-						return;
-					}
+
 					ImageSlotTextureNode *image_node = static_cast<ImageSlotTextureNode*>(node);
 					int slot = image_node->slot;
 					if(slot != -1) {
diff --git a/intern/cycles/render/scene.cpp b/intern/cycles/render/scene.cpp
index 4db20338744..c59a5d97df5 100644
--- a/intern/cycles/render/scene.cpp
+++ b/intern/cycles/render/scene.cpp
@@ -148,8 +148,6 @@ void Scene::device_update(Device *device_, Progress& progress)
 	 * - Film needs light manager to run for use_light_visibility
 	 * - Lookup tables are done a second time to handle film tables
 	 */
-	
-	image_manager->set_pack_images(device->info.pack_images);
 
 	progress.set_status("Updating Shaders");
 	shader_manager->device_update(device, &dscene, this, progress);
diff --git a/intern/cycles/render/scene.h b/intern/cycles/render/scene.h
index 4c2c4f5fcc3..0194327f567 100644
--- a/intern/cycles/render/scene.h
+++ b/intern/cycles/render/scene.h
@@ -121,13 +121,6 @@ public:
 	vector<device_vector<uchar>* > tex_byte_image;
 	vector<device_vector<half>* > tex_half_image;
 
-	/* opencl images */
-	device_vector<float4> tex_image_float4_packed;
-	device_vector<uchar4> tex_image_byte4_packed;
-	device_vector<float> tex_image_float_packed;
-	device_vector<uchar> tex_image_byte_packed;
-	device_vector<uint4> tex_image_packed_info;
-
 	KernelData data;
 };