Greg Daniel | 48cf268 | 2016-02-22 09:11:32 -0500 | [diff] [blame^] | 1 | /* |
| 2 | * Copyright 2015 Google Inc. |
| 3 | * |
| 4 | * Use of this source code is governed by a BSD-style license that can be |
| 5 | * found in the LICENSE file. |
| 6 | */ |
| 7 | |
| 8 | #include "GrVkGpu.h" |
| 9 | |
| 10 | #include "GrContextOptions.h" |
| 11 | #include "GrGeometryProcessor.h" |
| 12 | #include "GrGpuResourceCacheAccess.h" |
| 13 | #include "GrPipeline.h" |
| 14 | #include "GrRenderTargetPriv.h" |
| 15 | #include "GrSurfacePriv.h" |
| 16 | #include "GrTexturePriv.h" |
| 17 | #include "GrVertices.h" |
| 18 | |
| 19 | #include "GrVkCommandBuffer.h" |
| 20 | #include "GrVkImage.h" |
| 21 | #include "GrVkIndexBuffer.h" |
| 22 | #include "GrVkMemory.h" |
| 23 | #include "GrVkPipeline.h" |
| 24 | #include "GrVkProgram.h" |
| 25 | #include "GrVkProgramBuilder.h" |
| 26 | #include "GrVkProgramDesc.h" |
| 27 | #include "GrVkRenderPass.h" |
| 28 | #include "GrVkResourceProvider.h" |
| 29 | #include "GrVkTexture.h" |
| 30 | #include "GrVkTextureRenderTarget.h" |
| 31 | #include "GrVkTransferBuffer.h" |
| 32 | #include "GrVkVertexBuffer.h" |
| 33 | |
| 34 | #include "SkConfig8888.h" |
| 35 | |
| 36 | #include "vk/GrVkInterface.h" |
| 37 | |
| 38 | #define VK_CALL(X) GR_VK_CALL(this->vkInterface(), X) |
| 39 | #define VK_CALL_RET(RET, X) GR_VK_CALL_RET(this->vkInterface(), RET, X) |
| 40 | #define VK_CALL_ERRCHECK(X) GR_VK_CALL_ERRCHECK(this->vkInterface(), X) |
| 41 | |
| 42 | //////////////////////////////////////////////////////////////////////////////// |
| 43 | // Stuff used to set up a GrVkGpu secrectly for now. |
| 44 | |
| 45 | // For now the VkGpuCreate is using the same signature as GL. This is mostly for ease of |
| 46 | // hiding this code from offical skia. In the end the VkGpuCreate will not take a GrBackendContext |
| 47 | // and mostly likely would take an optional device and queues to use. |
| 48 | GrGpu* vk_gpu_create(GrBackendContext backendContext, const GrContextOptions& options, |
| 49 | GrContext* context) { |
| 50 | // Below is Vulkan setup code that normal would be done by a client, but will do here for now |
| 51 | // for testing purposes. |
| 52 | VkPhysicalDevice physDev; |
| 53 | VkDevice device; |
| 54 | VkInstance inst; |
| 55 | VkResult err; |
| 56 | |
| 57 | const VkApplicationInfo app_info = { |
| 58 | VK_STRUCTURE_TYPE_APPLICATION_INFO, // sType |
| 59 | nullptr, // pNext |
| 60 | "vktest", // pApplicationName |
| 61 | 0, // applicationVersion |
| 62 | "vktest", // pEngineName |
| 63 | 0, // engineVerison |
| 64 | VK_API_VERSION, // apiVersion |
| 65 | }; |
| 66 | const VkInstanceCreateInfo instance_create = { |
| 67 | VK_STRUCTURE_TYPE_INSTANCE_CREATE_INFO, // sType |
| 68 | nullptr, // pNext |
| 69 | 0, // flags |
| 70 | &app_info, // pApplicationInfo |
| 71 | 0, // enabledLayerNameCount |
| 72 | nullptr, // ppEnabledLayerNames |
| 73 | 0, // enabledExtensionNameCount |
| 74 | nullptr, // ppEnabledExtensionNames |
| 75 | }; |
| 76 | err = vkCreateInstance(&instance_create, nullptr, &inst); |
| 77 | if (err < 0) { |
| 78 | SkDebugf("vkCreateInstanced failed: %d\n", err); |
| 79 | SkFAIL("failing"); |
| 80 | } |
| 81 | |
| 82 | uint32_t gpuCount; |
| 83 | err = vkEnumeratePhysicalDevices(inst, &gpuCount, nullptr); |
| 84 | if (err) { |
| 85 | SkDebugf("vkEnumeratePhysicalDevices failed: %d\n", err); |
| 86 | SkFAIL("failing"); |
| 87 | } |
| 88 | SkASSERT(gpuCount > 0); |
| 89 | // Just returning the first physical device instead of getting the whole array. |
| 90 | gpuCount = 1; |
| 91 | err = vkEnumeratePhysicalDevices(inst, &gpuCount, &physDev); |
| 92 | if (err) { |
| 93 | SkDebugf("vkEnumeratePhysicalDevices failed: %d\n", err); |
| 94 | SkFAIL("failing"); |
| 95 | } |
| 96 | |
| 97 | // query to get the initial queue props size |
| 98 | uint32_t queueCount; |
| 99 | vkGetPhysicalDeviceQueueFamilyProperties(physDev, &queueCount, nullptr); |
| 100 | SkASSERT(queueCount >= 1); |
| 101 | |
| 102 | SkAutoMalloc queuePropsAlloc(queueCount * sizeof(VkQueueFamilyProperties)); |
| 103 | // now get the actual queue props |
| 104 | VkQueueFamilyProperties* queueProps = (VkQueueFamilyProperties*)queuePropsAlloc.get(); |
| 105 | |
| 106 | vkGetPhysicalDeviceQueueFamilyProperties(physDev, &queueCount, queueProps); |
| 107 | |
| 108 | // iterate to find the graphics queue |
| 109 | uint32_t graphicsQueueIndex = -1; |
| 110 | for (uint32_t i = 0; i < queueCount; i++) { |
| 111 | if (queueProps[i].queueFlags & VK_QUEUE_GRAPHICS_BIT) { |
| 112 | graphicsQueueIndex = i; |
| 113 | break; |
| 114 | } |
| 115 | } |
| 116 | SkASSERT(graphicsQueueIndex < queueCount); |
| 117 | |
| 118 | float queuePriorities[1] = { 0.0 }; |
| 119 | const VkDeviceQueueCreateInfo queueInfo = { |
| 120 | VK_STRUCTURE_TYPE_DEVICE_QUEUE_CREATE_INFO, // sType |
| 121 | nullptr, // pNext |
| 122 | 0, // VkDeviceQueueCreateFlags |
| 123 | 0, // queueFamilyIndex |
| 124 | 1, // queueCount |
| 125 | queuePriorities, // pQueuePriorities |
| 126 | }; |
| 127 | const VkDeviceCreateInfo deviceInfo = { |
| 128 | VK_STRUCTURE_TYPE_DEVICE_CREATE_INFO, // sType |
| 129 | nullptr, // pNext |
| 130 | 0, // VkDeviceCreateFlags |
| 131 | 1, // queueCreateInfoCount |
| 132 | &queueInfo, // pQueueCreateInfos |
| 133 | 0, // layerCount |
| 134 | nullptr, // ppEnabledLayerNames |
| 135 | 0, // extensionCount |
| 136 | nullptr, // ppEnabledExtensionNames |
| 137 | nullptr // ppEnabledFeatures |
| 138 | }; |
| 139 | |
| 140 | err = vkCreateDevice(physDev, &deviceInfo, nullptr, &device); |
| 141 | if (err) { |
| 142 | SkDebugf("CreateDevice failed: %d\n", err); |
| 143 | SkFAIL("failing"); |
| 144 | } |
| 145 | |
| 146 | VkQueue queue; |
| 147 | vkGetDeviceQueue(device, graphicsQueueIndex, 0, &queue); |
| 148 | |
| 149 | const VkCommandPoolCreateInfo cmdPoolInfo = { |
| 150 | VK_STRUCTURE_TYPE_COMMAND_POOL_CREATE_INFO, // sType |
| 151 | nullptr, // pNext |
| 152 | 0, // CmdPoolCreateFlags |
| 153 | graphicsQueueIndex, // queueFamilyIndex |
| 154 | }; |
| 155 | |
| 156 | VkCommandPool cmdPool; |
| 157 | err = vkCreateCommandPool(device, &cmdPoolInfo, nullptr, &cmdPool); |
| 158 | if (err) { |
| 159 | SkDebugf("CreateCommandPool failed: %d\n", err); |
| 160 | SkFAIL("failing"); |
| 161 | } |
| 162 | |
| 163 | return new GrVkGpu(context, options, physDev, device, queue, cmdPool, inst); |
| 164 | } |
| 165 | |
| 166 | //////////////////////////////////////////////////////////////////////////////// |
| 167 | |
| 168 | GrVkGpu::GrVkGpu(GrContext* context, const GrContextOptions& options, |
| 169 | VkPhysicalDevice physDev, VkDevice device, VkQueue queue, VkCommandPool cmdPool, |
| 170 | VkInstance inst) |
| 171 | : INHERITED(context) |
| 172 | , fDevice(device) |
| 173 | , fQueue(queue) |
| 174 | , fCmdPool(cmdPool) |
| 175 | , fResourceProvider(this) |
| 176 | , fVkInstance(inst) { |
| 177 | fInterface.reset(GrVkCreateInterface(fVkInstance)); |
| 178 | fCompiler = shaderc_compiler_initialize(); |
| 179 | |
| 180 | fVkCaps.reset(new GrVkCaps(options, fInterface, physDev)); |
| 181 | fCaps.reset(SkRef(fVkCaps.get())); |
| 182 | |
| 183 | fCurrentCmdBuffer = fResourceProvider.createCommandBuffer(); |
| 184 | SkASSERT(fCurrentCmdBuffer); |
| 185 | fCurrentCmdBuffer->begin(this); |
| 186 | VK_CALL(GetPhysicalDeviceMemoryProperties(physDev, &fPhysDevMemProps)); |
| 187 | |
| 188 | } |
| 189 | |
| 190 | GrVkGpu::~GrVkGpu() { |
| 191 | shaderc_compiler_release(fCompiler); |
| 192 | fCurrentCmdBuffer->end(this); |
| 193 | fCurrentCmdBuffer->unref(this); |
| 194 | |
| 195 | // wait for all commands to finish |
| 196 | VK_CALL(QueueWaitIdle(fQueue)); |
| 197 | |
| 198 | // must call this just before we destroy the VkDevice |
| 199 | fResourceProvider.destroyResources(); |
| 200 | |
| 201 | VK_CALL(DestroyCommandPool(fDevice, fCmdPool, nullptr)); |
| 202 | VK_CALL(DestroyDevice(fDevice, nullptr)); |
| 203 | VK_CALL(DestroyInstance(fVkInstance, nullptr)); |
| 204 | } |
| 205 | |
| 206 | /////////////////////////////////////////////////////////////////////////////// |
| 207 | |
| 208 | void GrVkGpu::submitCommandBuffer(SyncQueue sync) { |
| 209 | SkASSERT(fCurrentCmdBuffer); |
| 210 | fCurrentCmdBuffer->end(this); |
| 211 | |
| 212 | fCurrentCmdBuffer->submitToQueue(this, fQueue, sync); |
| 213 | fResourceProvider.checkCommandBuffers(); |
| 214 | |
| 215 | // Release old command buffer and create a new one |
| 216 | fCurrentCmdBuffer->unref(this); |
| 217 | fCurrentCmdBuffer = fResourceProvider.createCommandBuffer(); |
| 218 | SkASSERT(fCurrentCmdBuffer); |
| 219 | |
| 220 | fCurrentCmdBuffer->begin(this); |
| 221 | } |
| 222 | |
| 223 | /////////////////////////////////////////////////////////////////////////////// |
| 224 | GrVertexBuffer* GrVkGpu::onCreateVertexBuffer(size_t size, bool dynamic) { |
| 225 | return GrVkVertexBuffer::Create(this, size, dynamic); |
| 226 | } |
| 227 | |
| 228 | GrIndexBuffer* GrVkGpu::onCreateIndexBuffer(size_t size, bool dynamic) { |
| 229 | return GrVkIndexBuffer::Create(this, size, dynamic); |
| 230 | } |
| 231 | |
| 232 | GrTransferBuffer* GrVkGpu::onCreateTransferBuffer(size_t size, TransferType type) { |
| 233 | GrVkBuffer::Type bufferType = kCpuToGpu_TransferType ? GrVkBuffer::kCopyRead_Type |
| 234 | : GrVkBuffer::kCopyWrite_Type; |
| 235 | return GrVkTransferBuffer::Create(this, size, bufferType); |
| 236 | } |
| 237 | |
| 238 | //////////////////////////////////////////////////////////////////////////////// |
| 239 | bool GrVkGpu::onGetWritePixelsInfo(GrSurface* dstSurface, int width, int height, |
| 240 | GrPixelConfig srcConfig, DrawPreference* drawPreference, |
| 241 | WritePixelTempDrawInfo* tempDrawInfo) { |
| 242 | if (kIndex_8_GrPixelConfig == srcConfig || GrPixelConfigIsCompressed(dstSurface->config())) { |
| 243 | return false; |
| 244 | } |
| 245 | |
| 246 | // Currently we don't handle draws, so if the caller wants/needs to do a draw we need to fail |
| 247 | if (kNoDraw_DrawPreference != *drawPreference) { |
| 248 | return false; |
| 249 | } |
| 250 | |
| 251 | if (dstSurface->config() != srcConfig) { |
| 252 | // TODO: This should fall back to drawing or copying to change config of dstSurface to |
| 253 | // match that of srcConfig. |
| 254 | return false; |
| 255 | } |
| 256 | |
| 257 | return true; |
| 258 | } |
| 259 | |
| 260 | bool GrVkGpu::onWritePixels(GrSurface* surface, |
| 261 | int left, int top, int width, int height, |
| 262 | GrPixelConfig config, const void* buffer, |
| 263 | size_t rowBytes) { |
| 264 | GrVkTexture* vkTex = static_cast<GrVkTexture*>(surface->asTexture()); |
| 265 | if (!vkTex) { |
| 266 | return false; |
| 267 | } |
| 268 | |
| 269 | // We assume Vulkan doesn't do sRGB <-> linear conversions when reading and writing pixels. |
| 270 | if (GrPixelConfigIsSRGB(surface->config()) != GrPixelConfigIsSRGB(config)) { |
| 271 | return false; |
| 272 | } |
| 273 | |
| 274 | bool success = false; |
| 275 | if (GrPixelConfigIsCompressed(vkTex->desc().fConfig)) { |
| 276 | // We check that config == desc.fConfig in GrGpu::getWritePixelsInfo() |
| 277 | SkASSERT(config == vkTex->desc().fConfig); |
| 278 | // TODO: add compressed texture support |
| 279 | // delete the following two lines and uncomment the two after that when ready |
| 280 | vkTex->unref(); |
| 281 | return false; |
| 282 | //success = this->uploadCompressedTexData(vkTex->desc(), buffer, false, left, top, width, |
| 283 | // height); |
| 284 | } else { |
| 285 | bool linearTiling = vkTex->isLinearTiled(); |
| 286 | if (linearTiling && VK_IMAGE_LAYOUT_PREINITIALIZED != vkTex->currentLayout()) { |
| 287 | // Need to change the layout to general in order to perform a host write |
| 288 | VkImageLayout layout = vkTex->currentLayout(); |
| 289 | VkPipelineStageFlags srcStageMask = GrVkMemory::LayoutToPipelineStageFlags(layout); |
| 290 | VkPipelineStageFlags dstStageMask = VK_PIPELINE_STAGE_HOST_BIT; |
| 291 | VkAccessFlags srcAccessMask = GrVkMemory::LayoutToSrcAccessMask(layout); |
| 292 | VkAccessFlags dstAccessMask = VK_ACCESS_HOST_WRITE_BIT; |
| 293 | vkTex->setImageLayout(this, |
| 294 | VK_IMAGE_LAYOUT_GENERAL, |
| 295 | srcAccessMask, |
| 296 | dstAccessMask, |
| 297 | srcStageMask, |
| 298 | dstStageMask, |
| 299 | false); |
| 300 | } |
| 301 | success = this->uploadTexData(vkTex, left, top, width, height, config, |
| 302 | buffer, rowBytes); |
| 303 | } |
| 304 | |
| 305 | if (success) { |
| 306 | vkTex->texturePriv().dirtyMipMaps(true); |
| 307 | return true; |
| 308 | } |
| 309 | |
| 310 | return false; |
| 311 | } |
| 312 | |
| 313 | bool GrVkGpu::uploadTexData(GrVkTexture* tex, |
| 314 | int left, int top, int width, int height, |
| 315 | GrPixelConfig dataConfig, |
| 316 | const void* data, |
| 317 | size_t rowBytes) { |
| 318 | SkASSERT(data); |
| 319 | |
| 320 | // If we're uploading compressed data then we should be using uploadCompressedTexData |
| 321 | SkASSERT(!GrPixelConfigIsCompressed(dataConfig)); |
| 322 | |
| 323 | bool linearTiling = tex->isLinearTiled(); |
| 324 | |
| 325 | size_t bpp = GrBytesPerPixel(dataConfig); |
| 326 | |
| 327 | const GrSurfaceDesc& desc = tex->desc(); |
| 328 | |
| 329 | if (!GrSurfacePriv::AdjustWritePixelParams(desc.fWidth, desc.fHeight, bpp, &left, &top, |
| 330 | &width, &height, &data, &rowBytes)) { |
| 331 | return false; |
| 332 | } |
| 333 | size_t trimRowBytes = width * bpp; |
| 334 | |
| 335 | if (linearTiling) { |
| 336 | SkASSERT(VK_IMAGE_LAYOUT_PREINITIALIZED == tex->currentLayout() || |
| 337 | VK_IMAGE_LAYOUT_GENERAL == tex->currentLayout()); |
| 338 | const VkImageSubresource subres = { |
| 339 | VK_IMAGE_ASPECT_COLOR_BIT, |
| 340 | 0, // mipLevel |
| 341 | 0, // arraySlice |
| 342 | }; |
| 343 | VkSubresourceLayout layout; |
| 344 | VkResult err; |
| 345 | |
| 346 | const GrVkInterface* interface = this->vkInterface(); |
| 347 | |
| 348 | GR_VK_CALL(interface, GetImageSubresourceLayout(fDevice, |
| 349 | tex->textureImage(), |
| 350 | &subres, |
| 351 | &layout)); |
| 352 | |
| 353 | int texTop = kBottomLeft_GrSurfaceOrigin == desc.fOrigin ? tex->height() - top - height |
| 354 | : top; |
| 355 | VkDeviceSize offset = texTop*layout.rowPitch + left*bpp; |
| 356 | VkDeviceSize size = height*layout.rowPitch; |
| 357 | void* mapPtr; |
| 358 | err = GR_VK_CALL(interface, MapMemory(fDevice, tex->textureMemory(), offset, size, 0, |
| 359 | &mapPtr)); |
| 360 | if (err) { |
| 361 | return false; |
| 362 | } |
| 363 | |
| 364 | if (kBottomLeft_GrSurfaceOrigin == desc.fOrigin) { |
| 365 | // copy into buffer by rows |
| 366 | const char* srcRow = reinterpret_cast<const char*>(data); |
| 367 | char* dstRow = reinterpret_cast<char*>(mapPtr)+(height - 1)*layout.rowPitch; |
| 368 | for (int y = 0; y < height; y++) { |
| 369 | memcpy(dstRow, srcRow, trimRowBytes); |
| 370 | srcRow += rowBytes; |
| 371 | dstRow -= layout.rowPitch; |
| 372 | } |
| 373 | } else { |
| 374 | // If there is no padding on the src (rowBytes) or dst (layout.rowPitch) we can memcpy |
| 375 | if (trimRowBytes == rowBytes && trimRowBytes == layout.rowPitch) { |
| 376 | memcpy(mapPtr, data, trimRowBytes * height); |
| 377 | } else { |
| 378 | SkRectMemcpy(mapPtr, layout.rowPitch, data, rowBytes, trimRowBytes, height); |
| 379 | } |
| 380 | } |
| 381 | |
| 382 | GR_VK_CALL(interface, UnmapMemory(fDevice, tex->textureMemory())); |
| 383 | } else { |
| 384 | GrVkTransferBuffer* transferBuffer = |
| 385 | GrVkTransferBuffer::Create(this, trimRowBytes * height, GrVkBuffer::kCopyRead_Type); |
| 386 | |
| 387 | void* mapPtr = transferBuffer->map(); |
| 388 | |
| 389 | if (kBottomLeft_GrSurfaceOrigin == desc.fOrigin) { |
| 390 | // copy into buffer by rows |
| 391 | const char* srcRow = reinterpret_cast<const char*>(data); |
| 392 | char* dstRow = reinterpret_cast<char*>(mapPtr)+(height - 1)*trimRowBytes; |
| 393 | for (int y = 0; y < height; y++) { |
| 394 | memcpy(dstRow, srcRow, trimRowBytes); |
| 395 | srcRow += rowBytes; |
| 396 | dstRow -= trimRowBytes; |
| 397 | } |
| 398 | } else { |
| 399 | // If there is no padding on the src data rows, we can do a single memcpy |
| 400 | if (trimRowBytes == rowBytes) { |
| 401 | memcpy(mapPtr, data, trimRowBytes * height); |
| 402 | } else { |
| 403 | SkRectMemcpy(mapPtr, trimRowBytes, data, rowBytes, trimRowBytes, height); |
| 404 | } |
| 405 | } |
| 406 | |
| 407 | transferBuffer->unmap(); |
| 408 | |
| 409 | // make sure the unmap has finished |
| 410 | transferBuffer->addMemoryBarrier(this, |
| 411 | VK_ACCESS_HOST_WRITE_BIT, |
| 412 | VK_ACCESS_TRANSFER_READ_BIT, |
| 413 | VK_PIPELINE_STAGE_TOP_OF_PIPE_BIT, |
| 414 | VK_PIPELINE_STAGE_TRANSFER_BIT, |
| 415 | false); |
| 416 | |
| 417 | // Set up copy region |
| 418 | bool flipY = kBottomLeft_GrSurfaceOrigin == tex->origin(); |
| 419 | VkOffset3D offset = { |
| 420 | left, |
| 421 | flipY ? tex->height() - top - height : top, |
| 422 | 0 |
| 423 | }; |
| 424 | |
| 425 | VkBufferImageCopy region; |
| 426 | memset(®ion, 0, sizeof(VkBufferImageCopy)); |
| 427 | region.bufferOffset = 0; |
| 428 | region.bufferRowLength = width; |
| 429 | region.bufferImageHeight = height; |
| 430 | region.imageSubresource = { VK_IMAGE_ASPECT_COLOR_BIT, 0, 0, 1 }; |
| 431 | region.imageOffset = offset; |
| 432 | region.imageExtent = { (uint32_t)width, (uint32_t)height, 1 }; |
| 433 | |
| 434 | // Change layout of our target so it can be copied to |
| 435 | VkImageLayout layout = tex->currentLayout(); |
| 436 | VkPipelineStageFlags srcStageMask = GrVkMemory::LayoutToPipelineStageFlags(layout); |
| 437 | VkPipelineStageFlags dstStageMask = VK_PIPELINE_STAGE_TRANSFER_BIT; |
| 438 | VkAccessFlags srcAccessMask = GrVkMemory::LayoutToSrcAccessMask(layout); |
| 439 | VkAccessFlags dstAccessMask = VK_ACCESS_TRANSFER_WRITE_BIT; |
| 440 | tex->setImageLayout(this, |
| 441 | VK_IMAGE_LAYOUT_TRANSFER_DST_OPTIMAL, |
| 442 | srcAccessMask, |
| 443 | dstAccessMask, |
| 444 | srcStageMask, |
| 445 | dstStageMask, |
| 446 | false); |
| 447 | |
| 448 | // Copy the buffer to the image |
| 449 | fCurrentCmdBuffer->copyBufferToImage(this, |
| 450 | transferBuffer, |
| 451 | tex, |
| 452 | VK_IMAGE_LAYOUT_TRANSFER_DST_OPTIMAL, |
| 453 | 1, |
| 454 | ®ion); |
| 455 | |
| 456 | // Submit the current command buffer to the Queue |
| 457 | this->submitCommandBuffer(kSkip_SyncQueue); |
| 458 | |
| 459 | transferBuffer->unref(); |
| 460 | } |
| 461 | |
| 462 | return true; |
| 463 | } |
| 464 | |
| 465 | //////////////////////////////////////////////////////////////////////////////// |
| 466 | GrTexture* GrVkGpu::onCreateTexture(const GrSurfaceDesc& desc, GrGpuResource::LifeCycle lifeCycle, |
| 467 | const void* srcData, size_t rowBytes) { |
| 468 | bool renderTarget = SkToBool(desc.fFlags & kRenderTarget_GrSurfaceFlag); |
| 469 | |
| 470 | VkFormat pixelFormat; |
| 471 | if (!GrPixelConfigToVkFormat(desc.fConfig, &pixelFormat)) { |
| 472 | return nullptr; |
| 473 | } |
| 474 | |
| 475 | if (!fVkCaps->isConfigTexturable(desc.fConfig)) { |
| 476 | return nullptr; |
| 477 | } |
| 478 | |
| 479 | bool linearTiling = false; |
| 480 | if (SkToBool(desc.fFlags & kZeroCopy_GrSurfaceFlag)) { |
| 481 | if (fVkCaps->isConfigTexurableLinearly(desc.fConfig) && |
| 482 | (!renderTarget || fVkCaps->isConfigRenderableLinearly(desc.fConfig, false))) { |
| 483 | linearTiling = true; |
| 484 | } else { |
| 485 | return nullptr; |
| 486 | } |
| 487 | } |
| 488 | |
| 489 | VkImageUsageFlags usageFlags = VK_IMAGE_USAGE_SAMPLED_BIT; |
| 490 | if (renderTarget) { |
| 491 | usageFlags |= VK_IMAGE_USAGE_COLOR_ATTACHMENT_BIT; |
| 492 | } |
| 493 | |
| 494 | // For now we will set the VK_IMAGE_USAGE_TRANSFER_DESTINATION_BIT and |
| 495 | // VK_IMAGE_USAGE_TRANSFER_SOURCE_BIT on every texture since we do not know whether or not we |
| 496 | // will be using this texture in some copy or not. Also this assumes, as is the current case, |
| 497 | // that all render targets in vulkan are also texutres. If we change this practice of setting |
| 498 | // both bits, we must make sure to set the destination bit if we are uploading srcData to the |
| 499 | // texture. |
| 500 | usageFlags |= VK_IMAGE_USAGE_TRANSFER_SRC_BIT | VK_IMAGE_USAGE_TRANSFER_DST_BIT; |
| 501 | |
| 502 | VkFlags memProps = (srcData && linearTiling) ? VK_MEMORY_PROPERTY_HOST_VISIBLE_BIT : |
| 503 | VK_MEMORY_PROPERTY_DEVICE_LOCAL_BIT; |
| 504 | |
| 505 | // This ImageDesc refers to the texture that will be read by the client. Thus even if msaa is |
| 506 | // requested, this ImageDesc describes the resolved texutre. Therefore we always have samples set |
| 507 | // to 1. |
| 508 | GrVkImage::ImageDesc imageDesc; |
| 509 | imageDesc.fImageType = VK_IMAGE_TYPE_2D; |
| 510 | imageDesc.fFormat = pixelFormat; |
| 511 | imageDesc.fWidth = desc.fWidth; |
| 512 | imageDesc.fHeight = desc.fHeight; |
| 513 | imageDesc.fLevels = 1; |
| 514 | imageDesc.fSamples = 1; |
| 515 | imageDesc.fImageTiling = linearTiling ? VK_IMAGE_TILING_LINEAR : VK_IMAGE_TILING_OPTIMAL; |
| 516 | imageDesc.fUsageFlags = usageFlags; |
| 517 | imageDesc.fMemProps = memProps; |
| 518 | |
| 519 | GrVkTexture* tex; |
| 520 | if (renderTarget) { |
| 521 | tex = GrVkTextureRenderTarget::CreateNewTextureRenderTarget(this, desc, lifeCycle, |
| 522 | imageDesc); |
| 523 | } else { |
| 524 | tex = GrVkTexture::CreateNewTexture(this, desc, lifeCycle, imageDesc); |
| 525 | } |
| 526 | |
| 527 | if (!tex) { |
| 528 | return nullptr; |
| 529 | } |
| 530 | |
| 531 | if (srcData) { |
| 532 | if (!this->uploadTexData(tex, 0, 0, desc.fWidth, desc.fHeight, desc.fConfig, srcData, |
| 533 | rowBytes)) { |
| 534 | tex->unref(); |
| 535 | return nullptr; |
| 536 | } |
| 537 | } |
| 538 | |
| 539 | return tex; |
| 540 | } |
| 541 | |
| 542 | //////////////////////////////////////////////////////////////////////////////// |
| 543 | |
| 544 | static GrSurfaceOrigin resolve_origin(GrSurfaceOrigin origin) { |
| 545 | // By default, all textures in Vk use TopLeft |
| 546 | if (kDefault_GrSurfaceOrigin == origin) { |
| 547 | return kTopLeft_GrSurfaceOrigin; |
| 548 | } else { |
| 549 | return origin; |
| 550 | } |
| 551 | } |
| 552 | |
| 553 | GrTexture* GrVkGpu::onWrapBackendTexture(const GrBackendTextureDesc& desc, |
| 554 | GrWrapOwnership ownership) { |
| 555 | VkFormat format; |
| 556 | if (!GrPixelConfigToVkFormat(desc.fConfig, &format)) { |
| 557 | return nullptr; |
| 558 | } |
| 559 | |
| 560 | if (0 == desc.fTextureHandle) { |
| 561 | return nullptr; |
| 562 | } |
| 563 | |
| 564 | int maxSize = this->caps()->maxTextureSize(); |
| 565 | if (desc.fWidth > maxSize || desc.fHeight > maxSize) { |
| 566 | return nullptr; |
| 567 | } |
| 568 | |
| 569 | // TODO: determine what format Chrome will actually send us and turn it into a Resource |
| 570 | GrVkImage::Resource* imageRsrc = reinterpret_cast<GrVkImage::Resource*>(desc.fTextureHandle); |
| 571 | |
| 572 | GrGpuResource::LifeCycle lifeCycle; |
| 573 | switch (ownership) { |
| 574 | case kAdopt_GrWrapOwnership: |
| 575 | lifeCycle = GrGpuResource::kAdopted_LifeCycle; |
| 576 | break; |
| 577 | case kBorrow_GrWrapOwnership: |
| 578 | lifeCycle = GrGpuResource::kBorrowed_LifeCycle; |
| 579 | break; |
| 580 | } |
| 581 | |
| 582 | GrSurfaceDesc surfDesc; |
| 583 | // next line relies on GrBackendTextureDesc's flags matching GrTexture's |
| 584 | surfDesc.fFlags = (GrSurfaceFlags)desc.fFlags; |
| 585 | surfDesc.fWidth = desc.fWidth; |
| 586 | surfDesc.fHeight = desc.fHeight; |
| 587 | surfDesc.fConfig = desc.fConfig; |
| 588 | surfDesc.fSampleCnt = SkTMin(desc.fSampleCnt, this->caps()->maxSampleCount()); |
| 589 | bool renderTarget = SkToBool(desc.fFlags & kRenderTarget_GrBackendTextureFlag); |
| 590 | // In GL, Chrome assumes all textures are BottomLeft |
| 591 | // In VK, we don't have this restriction |
| 592 | surfDesc.fOrigin = resolve_origin(desc.fOrigin); |
| 593 | |
| 594 | GrVkTexture* texture = nullptr; |
| 595 | if (renderTarget) { |
| 596 | texture = GrVkTextureRenderTarget::CreateWrappedTextureRenderTarget(this, surfDesc, |
| 597 | lifeCycle, format, |
| 598 | imageRsrc); |
| 599 | } else { |
| 600 | texture = GrVkTexture::CreateWrappedTexture(this, surfDesc, lifeCycle, format, imageRsrc); |
| 601 | } |
| 602 | if (!texture) { |
| 603 | return nullptr; |
| 604 | } |
| 605 | |
| 606 | return texture; |
| 607 | } |
| 608 | |
| 609 | GrRenderTarget* GrVkGpu::onWrapBackendRenderTarget(const GrBackendRenderTargetDesc& wrapDesc, |
| 610 | GrWrapOwnership ownership) { |
| 611 | |
| 612 | // TODO: determine what format Chrome will actually send us and turn it into a Resource |
| 613 | GrVkImage::Resource* imageRsrc = |
| 614 | reinterpret_cast<GrVkImage::Resource*>(wrapDesc.fRenderTargetHandle); |
| 615 | |
| 616 | GrGpuResource::LifeCycle lifeCycle; |
| 617 | switch (ownership) { |
| 618 | case kAdopt_GrWrapOwnership: |
| 619 | lifeCycle = GrGpuResource::kAdopted_LifeCycle; |
| 620 | break; |
| 621 | case kBorrow_GrWrapOwnership: |
| 622 | lifeCycle = GrGpuResource::kBorrowed_LifeCycle; |
| 623 | break; |
| 624 | } |
| 625 | |
| 626 | GrSurfaceDesc desc; |
| 627 | desc.fConfig = wrapDesc.fConfig; |
| 628 | desc.fFlags = kCheckAllocation_GrSurfaceFlag; |
| 629 | desc.fWidth = wrapDesc.fWidth; |
| 630 | desc.fHeight = wrapDesc.fHeight; |
| 631 | desc.fSampleCnt = SkTMin(wrapDesc.fSampleCnt, this->caps()->maxSampleCount()); |
| 632 | |
| 633 | desc.fOrigin = resolve_origin(wrapDesc.fOrigin); |
| 634 | |
| 635 | GrVkRenderTarget* tgt = GrVkRenderTarget::CreateWrappedRenderTarget(this, desc, |
| 636 | lifeCycle, imageRsrc); |
| 637 | if (tgt && wrapDesc.fStencilBits) { |
| 638 | if (!createStencilAttachmentForRenderTarget(tgt, desc.fWidth, desc.fHeight)) { |
| 639 | tgt->unref(); |
| 640 | return nullptr; |
| 641 | } |
| 642 | } |
| 643 | return tgt; |
| 644 | } |
| 645 | |
| 646 | //////////////////////////////////////////////////////////////////////////////// |
| 647 | |
| 648 | void GrVkGpu::bindGeometry(const GrPrimitiveProcessor& primProc, |
| 649 | const GrNonInstancedVertices& vertices) { |
| 650 | GrVkVertexBuffer* vbuf; |
| 651 | vbuf = (GrVkVertexBuffer*)vertices.vertexBuffer(); |
| 652 | SkASSERT(vbuf); |
| 653 | SkASSERT(!vbuf->isMapped()); |
| 654 | |
| 655 | vbuf->addMemoryBarrier(this, |
| 656 | VK_ACCESS_HOST_WRITE_BIT, |
| 657 | VK_ACCESS_VERTEX_ATTRIBUTE_READ_BIT, |
| 658 | VK_PIPELINE_STAGE_HOST_BIT, |
| 659 | VK_PIPELINE_STAGE_VERTEX_INPUT_BIT, |
| 660 | false); |
| 661 | |
| 662 | fCurrentCmdBuffer->bindVertexBuffer(this, vbuf); |
| 663 | |
| 664 | if (vertices.isIndexed()) { |
| 665 | GrVkIndexBuffer* ibuf = (GrVkIndexBuffer*)vertices.indexBuffer(); |
| 666 | SkASSERT(ibuf); |
| 667 | SkASSERT(!ibuf->isMapped()); |
| 668 | |
| 669 | ibuf->addMemoryBarrier(this, |
| 670 | VK_ACCESS_HOST_WRITE_BIT, |
| 671 | VK_ACCESS_INDEX_READ_BIT, |
| 672 | VK_PIPELINE_STAGE_HOST_BIT, |
| 673 | VK_PIPELINE_STAGE_VERTEX_INPUT_BIT, |
| 674 | false); |
| 675 | |
| 676 | fCurrentCmdBuffer->bindIndexBuffer(this, ibuf); |
| 677 | } |
| 678 | } |
| 679 | |
| 680 | void GrVkGpu::buildProgramDesc(GrProgramDesc* desc, |
| 681 | const GrPrimitiveProcessor& primProc, |
| 682 | const GrPipeline& pipeline) const { |
| 683 | if (!GrVkProgramDescBuilder::Build(desc, primProc, pipeline, *this->vkCaps().glslCaps())) { |
| 684 | SkDEBUGFAIL("Failed to generate GL program descriptor"); |
| 685 | } |
| 686 | } |
| 687 | |
| 688 | //////////////////////////////////////////////////////////////////////////////// |
| 689 | |
| 690 | GrStencilAttachment* GrVkGpu::createStencilAttachmentForRenderTarget(const GrRenderTarget* rt, |
| 691 | int width, |
| 692 | int height) { |
| 693 | SkASSERT(rt->asTexture()); |
| 694 | SkASSERT(width >= rt->width()); |
| 695 | SkASSERT(height >= rt->height()); |
| 696 | |
| 697 | int samples = rt->numStencilSamples(); |
| 698 | |
| 699 | SkASSERT(this->vkCaps().stencilFormats().count()); |
| 700 | const GrVkCaps::StencilFormat& sFmt = this->vkCaps().stencilFormats()[0]; |
| 701 | |
| 702 | GrVkStencilAttachment* stencil(GrVkStencilAttachment::Create(this, |
| 703 | GrGpuResource::kCached_LifeCycle, |
| 704 | width, |
| 705 | height, |
| 706 | samples, |
| 707 | sFmt)); |
| 708 | fStats.incStencilAttachmentCreates(); |
| 709 | return stencil; |
| 710 | } |
| 711 | |
| 712 | //////////////////////////////////////////////////////////////////////////////// |
| 713 | |
| 714 | GrBackendObject GrVkGpu::createTestingOnlyBackendTexture(void* srcData, int w, int h, |
| 715 | GrPixelConfig config) { |
| 716 | |
| 717 | VkFormat pixelFormat; |
| 718 | if (!GrPixelConfigToVkFormat(config, &pixelFormat)) { |
| 719 | return 0; |
| 720 | } |
| 721 | |
| 722 | bool linearTiling = false; |
| 723 | if (!fVkCaps->isConfigTexturable(config)) { |
| 724 | return 0; |
| 725 | } |
| 726 | |
| 727 | if (fVkCaps->isConfigTexurableLinearly(config)) { |
| 728 | linearTiling = true; |
| 729 | } |
| 730 | |
| 731 | // Currently this is not supported since it requires a copy which has not yet been implemented. |
| 732 | if (srcData && !linearTiling) { |
| 733 | return 0; |
| 734 | } |
| 735 | |
| 736 | VkImageUsageFlags usageFlags = VK_IMAGE_USAGE_SAMPLED_BIT; |
| 737 | usageFlags |= VK_IMAGE_USAGE_TRANSFER_SRC_BIT; |
| 738 | usageFlags |= VK_IMAGE_USAGE_TRANSFER_DST_BIT; |
| 739 | |
| 740 | VkFlags memProps = (srcData && linearTiling) ? VK_MEMORY_PROPERTY_HOST_VISIBLE_BIT : |
| 741 | VK_MEMORY_PROPERTY_DEVICE_LOCAL_BIT; |
| 742 | |
| 743 | // This ImageDesc refers to the texture that will be read by the client. Thus even if msaa is |
| 744 | // requested, this ImageDesc describes the resolved texutre. Therefore we always have samples set |
| 745 | // to 1. |
| 746 | GrVkImage::ImageDesc imageDesc; |
| 747 | imageDesc.fImageType = VK_IMAGE_TYPE_2D; |
| 748 | imageDesc.fFormat = pixelFormat; |
| 749 | imageDesc.fWidth = w; |
| 750 | imageDesc.fHeight = h; |
| 751 | imageDesc.fLevels = 1; |
| 752 | imageDesc.fSamples = 1; |
| 753 | imageDesc.fImageTiling = linearTiling ? VK_IMAGE_TILING_LINEAR : VK_IMAGE_TILING_OPTIMAL; |
| 754 | imageDesc.fUsageFlags = usageFlags; |
| 755 | imageDesc.fMemProps = memProps; |
| 756 | |
| 757 | const GrVkImage::Resource* imageRsrc = GrVkImage::CreateResource(this, imageDesc); |
| 758 | if (!imageRsrc) { |
| 759 | return 0; |
| 760 | } |
| 761 | |
| 762 | if (srcData) { |
| 763 | if (linearTiling) { |
| 764 | const VkImageSubresource subres = { |
| 765 | VK_IMAGE_ASPECT_COLOR_BIT, |
| 766 | 0, // mipLevel |
| 767 | 0, // arraySlice |
| 768 | }; |
| 769 | VkSubresourceLayout layout; |
| 770 | VkResult err; |
| 771 | |
| 772 | const GrVkInterface* interface = this->vkInterface(); |
| 773 | |
| 774 | GR_VK_CALL(interface, GetImageSubresourceLayout(fDevice, |
| 775 | imageRsrc->fImage, |
| 776 | &subres, |
| 777 | &layout)); |
| 778 | |
| 779 | void* mapPtr; |
| 780 | err = GR_VK_CALL(interface, MapMemory(fDevice, |
| 781 | imageRsrc->fAlloc, |
| 782 | 0, |
| 783 | layout.rowPitch * h, |
| 784 | 0, |
| 785 | &mapPtr)); |
| 786 | if (err) { |
| 787 | imageRsrc->unref(this); |
| 788 | return 0; |
| 789 | } |
| 790 | |
| 791 | size_t bpp = GrBytesPerPixel(config); |
| 792 | size_t rowCopyBytes = bpp * w; |
| 793 | // If there is no padding on dst (layout.rowPitch) we can do a single memcopy. |
| 794 | // This assumes the srcData comes in with no padding. |
| 795 | if (rowCopyBytes == layout.rowPitch) { |
| 796 | memcpy(mapPtr, srcData, rowCopyBytes * h); |
| 797 | } else { |
| 798 | SkRectMemcpy(mapPtr, layout.rowPitch, srcData, w, rowCopyBytes, h); |
| 799 | } |
| 800 | GR_VK_CALL(interface, UnmapMemory(fDevice, imageRsrc->fAlloc)); |
| 801 | } else { |
| 802 | // TODO: Add support for copying to optimal tiling |
| 803 | SkASSERT(false); |
| 804 | } |
| 805 | } |
| 806 | |
| 807 | return (GrBackendObject)imageRsrc; |
| 808 | } |
| 809 | |
| 810 | bool GrVkGpu::isTestingOnlyBackendTexture(GrBackendObject id) const { |
| 811 | GrVkImage::Resource* backend = reinterpret_cast<GrVkImage::Resource*>(id); |
| 812 | |
| 813 | if (backend && backend->fImage && backend->fAlloc) { |
| 814 | VkMemoryRequirements req; |
| 815 | memset(&req, 0, sizeof(req)); |
| 816 | GR_VK_CALL(this->vkInterface(), GetImageMemoryRequirements(fDevice, |
| 817 | backend->fImage, |
| 818 | &req)); |
| 819 | // TODO: find a better check |
| 820 | // This will probably fail with a different driver |
| 821 | return (req.size > 0) && (req.size <= 8192 * 8192); |
| 822 | } |
| 823 | |
| 824 | return false; |
| 825 | } |
| 826 | |
| 827 | void GrVkGpu::deleteTestingOnlyBackendTexture(GrBackendObject id, bool abandon) { |
| 828 | GrVkImage::Resource* backend = reinterpret_cast<GrVkImage::Resource*>(id); |
| 829 | |
| 830 | if (backend) { |
| 831 | if (!abandon) { |
| 832 | backend->unref(this); |
| 833 | } else { |
| 834 | backend->unrefAndAbandon(); |
| 835 | } |
| 836 | } |
| 837 | } |
| 838 | |
| 839 | //////////////////////////////////////////////////////////////////////////////// |
| 840 | |
| 841 | void GrVkGpu::addMemoryBarrier(VkPipelineStageFlags srcStageMask, |
| 842 | VkPipelineStageFlags dstStageMask, |
| 843 | bool byRegion, |
| 844 | VkMemoryBarrier* barrier) const { |
| 845 | SkASSERT(fCurrentCmdBuffer); |
| 846 | fCurrentCmdBuffer->pipelineBarrier(this, |
| 847 | srcStageMask, |
| 848 | dstStageMask, |
| 849 | byRegion, |
| 850 | GrVkCommandBuffer::kMemory_BarrierType, |
| 851 | barrier); |
| 852 | } |
| 853 | |
| 854 | void GrVkGpu::addBufferMemoryBarrier(VkPipelineStageFlags srcStageMask, |
| 855 | VkPipelineStageFlags dstStageMask, |
| 856 | bool byRegion, |
| 857 | VkBufferMemoryBarrier* barrier) const { |
| 858 | SkASSERT(fCurrentCmdBuffer); |
| 859 | fCurrentCmdBuffer->pipelineBarrier(this, |
| 860 | srcStageMask, |
| 861 | dstStageMask, |
| 862 | byRegion, |
| 863 | GrVkCommandBuffer::kBufferMemory_BarrierType, |
| 864 | barrier); |
| 865 | } |
| 866 | |
| 867 | void GrVkGpu::addImageMemoryBarrier(VkPipelineStageFlags srcStageMask, |
| 868 | VkPipelineStageFlags dstStageMask, |
| 869 | bool byRegion, |
| 870 | VkImageMemoryBarrier* barrier) const { |
| 871 | SkASSERT(fCurrentCmdBuffer); |
| 872 | fCurrentCmdBuffer->pipelineBarrier(this, |
| 873 | srcStageMask, |
| 874 | dstStageMask, |
| 875 | byRegion, |
| 876 | GrVkCommandBuffer::kImageMemory_BarrierType, |
| 877 | barrier); |
| 878 | } |
| 879 | |
| 880 | void GrVkGpu::finishDrawTarget() { |
| 881 | // Submit the current command buffer to the Queue |
| 882 | this->submitCommandBuffer(kSkip_SyncQueue); |
| 883 | } |
| 884 | |
| 885 | void GrVkGpu::onClear(GrRenderTarget* target, const SkIRect& rect, GrColor color) { |
| 886 | // parent class should never let us get here with no RT |
| 887 | SkASSERT(target); |
| 888 | |
| 889 | VkClearColorValue vkColor; |
| 890 | GrColorToRGBAFloat(color, vkColor.float32); |
| 891 | |
| 892 | GrVkRenderTarget* vkRT = static_cast<GrVkRenderTarget*>(target); |
| 893 | VkImageLayout origDstLayout = vkRT->currentLayout(); |
| 894 | |
| 895 | if (rect.width() != target->width() || rect.height() != target->height()) { |
| 896 | VkAccessFlags srcAccessMask = GrVkMemory::LayoutToSrcAccessMask(origDstLayout); |
| 897 | VkAccessFlags dstAccessMask = VK_ACCESS_COLOR_ATTACHMENT_WRITE_BIT; |
| 898 | VkPipelineStageFlags srcStageMask = |
| 899 | GrVkMemory::LayoutToPipelineStageFlags(vkRT->currentLayout()); |
| 900 | VkPipelineStageFlags dstStageMask = VK_PIPELINE_STAGE_TOP_OF_PIPE_BIT; |
| 901 | vkRT->setImageLayout(this, |
| 902 | VK_IMAGE_LAYOUT_COLOR_ATTACHMENT_OPTIMAL, |
| 903 | srcAccessMask, |
| 904 | dstAccessMask, |
| 905 | srcStageMask, |
| 906 | dstStageMask, |
| 907 | false); |
| 908 | |
| 909 | VkClearRect clearRect; |
| 910 | clearRect.rect.offset = { rect.fLeft, rect.fTop }; |
| 911 | clearRect.rect.extent = { (uint32_t)rect.width(), (uint32_t)rect.height() }; |
| 912 | clearRect.baseArrayLayer = 0; |
| 913 | clearRect.layerCount = 1; |
| 914 | |
| 915 | |
| 916 | |
| 917 | const GrVkRenderPass* renderPass = vkRT->simpleRenderPass(); |
| 918 | SkASSERT(renderPass); |
| 919 | fCurrentCmdBuffer->beginRenderPass(this, renderPass, *vkRT); |
| 920 | |
| 921 | uint32_t colorIndex; |
| 922 | SkAssertResult(renderPass->colorAttachmentIndex(&colorIndex)); |
| 923 | |
| 924 | VkClearAttachment attachment; |
| 925 | attachment.aspectMask = VK_IMAGE_ASPECT_COLOR_BIT; |
| 926 | attachment.colorAttachment = colorIndex; |
| 927 | attachment.clearValue.color = vkColor; |
| 928 | |
| 929 | fCurrentCmdBuffer->clearAttachments(this, 1, &attachment, 1, &clearRect); |
| 930 | fCurrentCmdBuffer->endRenderPass(this); |
| 931 | return; |
| 932 | } |
| 933 | |
| 934 | VkPipelineStageFlags srcStageMask = GrVkMemory::LayoutToPipelineStageFlags(origDstLayout); |
| 935 | VkPipelineStageFlags dstStageMask = VK_PIPELINE_STAGE_TRANSFER_BIT; |
| 936 | |
| 937 | VkAccessFlags srcAccessMask = GrVkMemory::LayoutToSrcAccessMask(origDstLayout);; |
| 938 | VkAccessFlags dstAccessMask = VK_ACCESS_TRANSFER_WRITE_BIT; |
| 939 | |
| 940 | vkRT->setImageLayout(this, |
| 941 | VK_IMAGE_LAYOUT_TRANSFER_DST_OPTIMAL, |
| 942 | srcAccessMask, |
| 943 | dstAccessMask, |
| 944 | srcStageMask, |
| 945 | dstStageMask, |
| 946 | false); |
| 947 | |
| 948 | |
| 949 | VkImageSubresourceRange subRange; |
| 950 | memset(&subRange, 0, sizeof(VkImageSubresourceRange)); |
| 951 | subRange.aspectMask = VK_IMAGE_ASPECT_COLOR_BIT; |
| 952 | subRange.baseMipLevel = 0; |
| 953 | subRange.levelCount = 1; |
| 954 | subRange.baseArrayLayer = 0; |
| 955 | subRange.layerCount = 1; |
| 956 | |
| 957 | // In the future we may not actually be doing this type of clear at all. If we are inside a |
| 958 | // render pass or doing a non full clear then we will use CmdClearColorAttachment. The more |
| 959 | // common use case will be clearing an attachment at the start of a render pass, in which case |
| 960 | // we will use the clear load ops. |
| 961 | fCurrentCmdBuffer->clearColorImage(this, |
| 962 | vkRT, |
| 963 | &vkColor, |
| 964 | 1, &subRange); |
| 965 | } |
| 966 | |
| 967 | inline bool can_copy_image(const GrSurface* dst, |
| 968 | const GrSurface* src, |
| 969 | const GrVkGpu* gpu) { |
| 970 | if (src->asTexture() && |
| 971 | dst->asTexture() && |
| 972 | src->origin() == dst->origin() && |
| 973 | src->config() == dst->config()) { |
| 974 | return true; |
| 975 | } |
| 976 | |
| 977 | // How does msaa play into this? If a VkTexture is multisampled, are we copying the multisampled |
| 978 | // or the resolved image here? |
| 979 | |
| 980 | return false; |
| 981 | } |
| 982 | |
| 983 | void GrVkGpu::copySurfaceAsCopyImage(GrSurface* dst, |
| 984 | GrSurface* src, |
| 985 | const SkIRect& srcRect, |
| 986 | const SkIPoint& dstPoint) { |
| 987 | SkASSERT(can_copy_image(dst, src, this)); |
| 988 | |
| 989 | // Insert memory barriers to switch src and dst to transfer_source and transfer_dst layouts |
| 990 | GrVkTexture* dstTex = static_cast<GrVkTexture*>(dst->asTexture()); |
| 991 | GrVkTexture* srcTex = static_cast<GrVkTexture*>(src->asTexture()); |
| 992 | |
| 993 | VkImageLayout origDstLayout = dstTex->currentLayout(); |
| 994 | VkImageLayout origSrcLayout = srcTex->currentLayout(); |
| 995 | |
| 996 | VkPipelineStageFlags srcStageMask = GrVkMemory::LayoutToPipelineStageFlags(origDstLayout); |
| 997 | VkPipelineStageFlags dstStageMask = VK_PIPELINE_STAGE_TRANSFER_BIT; |
| 998 | |
| 999 | // These flags are for flushing/invalidating caches and for the dst image it doesn't matter if |
| 1000 | // the cache is flushed since it is only being written to. |
| 1001 | VkAccessFlags srcAccessMask = GrVkMemory::LayoutToSrcAccessMask(origDstLayout);; |
| 1002 | VkAccessFlags dstAccessMask = VK_ACCESS_TRANSFER_WRITE_BIT; |
| 1003 | |
| 1004 | dstTex->setImageLayout(this, |
| 1005 | VK_IMAGE_LAYOUT_TRANSFER_DST_OPTIMAL, |
| 1006 | srcAccessMask, |
| 1007 | dstAccessMask, |
| 1008 | srcStageMask, |
| 1009 | dstStageMask, |
| 1010 | false); |
| 1011 | |
| 1012 | srcStageMask = GrVkMemory::LayoutToPipelineStageFlags(origSrcLayout); |
| 1013 | dstStageMask = VK_PIPELINE_STAGE_TRANSFER_BIT; |
| 1014 | |
| 1015 | srcAccessMask = GrVkMemory::LayoutToSrcAccessMask(origSrcLayout); |
| 1016 | dstAccessMask = VK_ACCESS_TRANSFER_READ_BIT; |
| 1017 | |
| 1018 | srcTex->setImageLayout(this, |
| 1019 | VK_IMAGE_LAYOUT_TRANSFER_SRC_OPTIMAL, |
| 1020 | srcAccessMask, |
| 1021 | dstAccessMask, |
| 1022 | srcStageMask, |
| 1023 | dstStageMask, |
| 1024 | false); |
| 1025 | |
| 1026 | // Flip rect if necessary |
| 1027 | SkIRect srcVkRect = srcRect; |
| 1028 | int32_t dstY = dstPoint.fY; |
| 1029 | |
| 1030 | if (kBottomLeft_GrSurfaceOrigin == src->origin()) { |
| 1031 | SkASSERT(kBottomLeft_GrSurfaceOrigin == dst->origin()); |
| 1032 | srcVkRect.fTop = src->height() - srcRect.fBottom; |
| 1033 | srcVkRect.fBottom = src->height() - srcRect.fTop; |
| 1034 | dstY = dst->height() - dstPoint.fY - srcVkRect.height(); |
| 1035 | } |
| 1036 | |
| 1037 | VkImageCopy copyRegion; |
| 1038 | memset(©Region, 0, sizeof(VkImageCopy)); |
| 1039 | copyRegion.srcSubresource = { VK_IMAGE_ASPECT_COLOR_BIT, 0, 0, 1 }; |
| 1040 | copyRegion.srcOffset = { srcVkRect.fLeft, srcVkRect.fTop, 0 }; |
| 1041 | copyRegion.dstSubresource = { VK_IMAGE_ASPECT_COLOR_BIT, 0, 0, 1 }; |
| 1042 | copyRegion.dstOffset = { dstPoint.fX, dstY, 0 }; |
| 1043 | copyRegion.extent = { (uint32_t)srcVkRect.width(), (uint32_t)srcVkRect.height(), 0 }; |
| 1044 | |
| 1045 | fCurrentCmdBuffer->copyImage(this, |
| 1046 | srcTex, |
| 1047 | VK_IMAGE_LAYOUT_TRANSFER_SRC_OPTIMAL, |
| 1048 | dstTex, |
| 1049 | VK_IMAGE_LAYOUT_TRANSFER_DST_OPTIMAL, |
| 1050 | 1, |
| 1051 | ©Region); |
| 1052 | } |
| 1053 | |
| 1054 | inline bool can_copy_as_draw(const GrSurface* dst, |
| 1055 | const GrSurface* src, |
| 1056 | const GrVkGpu* gpu) { |
| 1057 | return false; |
| 1058 | } |
| 1059 | |
| 1060 | void GrVkGpu::copySurfaceAsDraw(GrSurface* dst, |
| 1061 | GrSurface* src, |
| 1062 | const SkIRect& srcRect, |
| 1063 | const SkIPoint& dstPoint) { |
| 1064 | SkASSERT(false); |
| 1065 | } |
| 1066 | |
| 1067 | bool GrVkGpu::onCopySurface(GrSurface* dst, |
| 1068 | GrSurface* src, |
| 1069 | const SkIRect& srcRect, |
| 1070 | const SkIPoint& dstPoint) { |
| 1071 | if (can_copy_image(dst, src, this)) { |
| 1072 | this->copySurfaceAsCopyImage(dst, src, srcRect, dstPoint); |
| 1073 | return true; |
| 1074 | } |
| 1075 | |
| 1076 | if (can_copy_as_draw(dst, src, this)) { |
| 1077 | this->copySurfaceAsDraw(dst, src, srcRect, dstPoint); |
| 1078 | return true; |
| 1079 | } |
| 1080 | |
| 1081 | return false; |
| 1082 | } |
| 1083 | |
| 1084 | bool GrVkGpu::onGetReadPixelsInfo(GrSurface* srcSurface, int width, int height, size_t rowBytes, |
| 1085 | GrPixelConfig readConfig, DrawPreference* drawPreference, |
| 1086 | ReadPixelTempDrawInfo* tempDrawInfo) { |
| 1087 | // Currently we don't handle draws, so if the caller wants/needs to do a draw we need to fail |
| 1088 | if (kNoDraw_DrawPreference != *drawPreference) { |
| 1089 | return false; |
| 1090 | } |
| 1091 | |
| 1092 | if (srcSurface->config() != readConfig) { |
| 1093 | // TODO: This should fall back to drawing or copying to change config of srcSurface to match |
| 1094 | // that of readConfig. |
| 1095 | return false; |
| 1096 | } |
| 1097 | |
| 1098 | return true; |
| 1099 | } |
| 1100 | |
| 1101 | bool GrVkGpu::onReadPixels(GrSurface* surface, |
| 1102 | int left, int top, int width, int height, |
| 1103 | GrPixelConfig config, |
| 1104 | void* buffer, |
| 1105 | size_t rowBytes) { |
| 1106 | VkFormat pixelFormat; |
| 1107 | if (!GrPixelConfigToVkFormat(config, &pixelFormat)) { |
| 1108 | return false; |
| 1109 | } |
| 1110 | |
| 1111 | GrVkTexture* tgt = static_cast<GrVkTexture*>(surface->asTexture()); |
| 1112 | if (!tgt) { |
| 1113 | return false; |
| 1114 | } |
| 1115 | |
| 1116 | // Change layout of our target so it can be used as copy |
| 1117 | VkImageLayout layout = tgt->currentLayout(); |
| 1118 | VkPipelineStageFlags srcStageMask = GrVkMemory::LayoutToPipelineStageFlags(layout); |
| 1119 | VkPipelineStageFlags dstStageMask = VK_PIPELINE_STAGE_TRANSFER_BIT; |
| 1120 | VkAccessFlags srcAccessMask = GrVkMemory::LayoutToSrcAccessMask(layout); |
| 1121 | VkAccessFlags dstAccessMask = VK_ACCESS_TRANSFER_READ_BIT; |
| 1122 | tgt->setImageLayout(this, |
| 1123 | VK_IMAGE_LAYOUT_TRANSFER_SRC_OPTIMAL, |
| 1124 | srcAccessMask, |
| 1125 | dstAccessMask, |
| 1126 | srcStageMask, |
| 1127 | dstStageMask, |
| 1128 | false); |
| 1129 | |
| 1130 | GrVkTransferBuffer* transferBuffer = |
| 1131 | reinterpret_cast<GrVkTransferBuffer*>(this->createTransferBuffer(rowBytes * height, |
| 1132 | kGpuToCpu_TransferType)); |
| 1133 | |
| 1134 | bool flipY = kBottomLeft_GrSurfaceOrigin == surface->origin(); |
| 1135 | VkOffset3D offset = { |
| 1136 | left, |
| 1137 | flipY ? surface->height() - top - height : top, |
| 1138 | 0 |
| 1139 | }; |
| 1140 | |
| 1141 | // Copy the image to a buffer so we can map it to cpu memory |
| 1142 | VkBufferImageCopy region; |
| 1143 | memset(®ion, 0, sizeof(VkBufferImageCopy)); |
| 1144 | region.bufferOffset = 0; |
| 1145 | region.bufferRowLength = 0; // Forces RowLength to be imageExtent.width |
| 1146 | region.bufferImageHeight = 0; // Forces height to be tightly packed. Only useful for 3d images. |
| 1147 | region.imageSubresource = { VK_IMAGE_ASPECT_COLOR_BIT, 0, 0, 1 }; |
| 1148 | region.imageOffset = offset; |
| 1149 | region.imageExtent = { (uint32_t)width, (uint32_t)height, 1 }; |
| 1150 | |
| 1151 | fCurrentCmdBuffer->copyImageToBuffer(this, |
| 1152 | tgt, |
| 1153 | VK_IMAGE_LAYOUT_TRANSFER_SRC_OPTIMAL, |
| 1154 | transferBuffer, |
| 1155 | 1, |
| 1156 | ®ion); |
| 1157 | |
| 1158 | // make sure the copy to buffer has finished |
| 1159 | transferBuffer->addMemoryBarrier(this, |
| 1160 | VK_ACCESS_TRANSFER_WRITE_BIT, |
| 1161 | VK_ACCESS_HOST_READ_BIT, |
| 1162 | VK_PIPELINE_STAGE_TRANSFER_BIT, |
| 1163 | VK_PIPELINE_STAGE_HOST_BIT, |
| 1164 | false); |
| 1165 | |
| 1166 | // We need to submit the current command buffer to the Queue and make sure it finishes before |
| 1167 | // we can copy the data out of the buffer. |
| 1168 | this->submitCommandBuffer(kForce_SyncQueue); |
| 1169 | |
| 1170 | void* mappedMemory = transferBuffer->map(); |
| 1171 | |
| 1172 | memcpy(buffer, mappedMemory, rowBytes*height); |
| 1173 | |
| 1174 | transferBuffer->unmap(); |
| 1175 | transferBuffer->unref(); |
| 1176 | |
| 1177 | if (flipY) { |
| 1178 | SkAutoSMalloc<32 * sizeof(GrColor)> scratch; |
| 1179 | size_t tightRowBytes = GrBytesPerPixel(config) * width; |
| 1180 | scratch.reset(tightRowBytes); |
| 1181 | void* tmpRow = scratch.get(); |
| 1182 | // flip y in-place by rows |
| 1183 | const int halfY = height >> 1; |
| 1184 | char* top = reinterpret_cast<char*>(buffer); |
| 1185 | char* bottom = top + (height - 1) * rowBytes; |
| 1186 | for (int y = 0; y < halfY; y++) { |
| 1187 | memcpy(tmpRow, top, tightRowBytes); |
| 1188 | memcpy(top, bottom, tightRowBytes); |
| 1189 | memcpy(bottom, tmpRow, tightRowBytes); |
| 1190 | top += rowBytes; |
| 1191 | bottom -= rowBytes; |
| 1192 | } |
| 1193 | } |
| 1194 | |
| 1195 | return true; |
| 1196 | } |
| 1197 | |
| 1198 | void GrVkGpu::onDraw(const DrawArgs& args, const GrNonInstancedVertices& vertices) { |
| 1199 | GrRenderTarget* rt = args.fPipeline->getRenderTarget(); |
| 1200 | GrVkRenderTarget* vkRT = static_cast<GrVkRenderTarget*>(rt); |
| 1201 | const GrVkRenderPass* renderPass = vkRT->simpleRenderPass(); |
| 1202 | SkASSERT(renderPass); |
| 1203 | |
| 1204 | |
| 1205 | GrVkProgram* program = GrVkProgramBuilder::CreateProgram(this, args, |
| 1206 | vertices.primitiveType(), |
| 1207 | *renderPass); |
| 1208 | |
| 1209 | if (!program) { |
| 1210 | return; |
| 1211 | } |
| 1212 | |
| 1213 | program->setData(this, *args.fPrimitiveProcessor, *args.fPipeline); |
| 1214 | |
| 1215 | fCurrentCmdBuffer->beginRenderPass(this, renderPass, *vkRT); |
| 1216 | |
| 1217 | program->bind(this, fCurrentCmdBuffer); |
| 1218 | |
| 1219 | this->bindGeometry(*args.fPrimitiveProcessor, vertices); |
| 1220 | |
| 1221 | // Change layout of our render target so it can be used as the color attachment |
| 1222 | VkImageLayout layout = vkRT->currentLayout(); |
| 1223 | // Our color attachment is purely a destination and won't be read so don't need to flush or |
| 1224 | // invalidate any caches |
| 1225 | VkPipelineStageFlags srcStageMask = GrVkMemory::LayoutToPipelineStageFlags(layout); |
| 1226 | VkPipelineStageFlags dstStageMask = VK_PIPELINE_STAGE_TOP_OF_PIPE_BIT; |
| 1227 | VkAccessFlags srcAccessMask = GrVkMemory::LayoutToSrcAccessMask(layout); |
| 1228 | VkAccessFlags dstAccessMask = VK_ACCESS_COLOR_ATTACHMENT_WRITE_BIT; |
| 1229 | vkRT->setImageLayout(this, |
| 1230 | VK_IMAGE_LAYOUT_COLOR_ATTACHMENT_OPTIMAL, |
| 1231 | srcAccessMask, |
| 1232 | dstAccessMask, |
| 1233 | srcStageMask, |
| 1234 | dstStageMask, |
| 1235 | false); |
| 1236 | |
| 1237 | if (vertices.isIndexed()) { |
| 1238 | fCurrentCmdBuffer->drawIndexed(this, |
| 1239 | vertices.indexCount(), |
| 1240 | 1, |
| 1241 | vertices.startIndex(), |
| 1242 | vertices.startVertex(), |
| 1243 | 0); |
| 1244 | } else { |
| 1245 | fCurrentCmdBuffer->draw(this, vertices.vertexCount(), 1, vertices.startVertex(), 0); |
| 1246 | } |
| 1247 | |
| 1248 | fCurrentCmdBuffer->endRenderPass(this); |
| 1249 | |
| 1250 | // Technically we don't have to call this here (since there is a safety check in program:setData |
| 1251 | // but this will allow for quicker freeing of resources if the program sits in a cache for a |
| 1252 | // while. |
| 1253 | program->freeTempResources(this); |
| 1254 | // This free will go away once we setup a program cache, and then the cache will be responsible |
| 1255 | // for call freeGpuResources. |
| 1256 | program->freeGPUResources(this); |
| 1257 | program->unref(); |
| 1258 | |
| 1259 | #if SWAP_PER_DRAW |
| 1260 | glFlush(); |
| 1261 | #if defined(SK_BUILD_FOR_MAC) |
| 1262 | aglSwapBuffers(aglGetCurrentContext()); |
| 1263 | int set_a_break_pt_here = 9; |
| 1264 | aglSwapBuffers(aglGetCurrentContext()); |
| 1265 | #elif defined(SK_BUILD_FOR_WIN32) |
| 1266 | SwapBuf(); |
| 1267 | int set_a_break_pt_here = 9; |
| 1268 | SwapBuf(); |
| 1269 | #endif |
| 1270 | #endif |
| 1271 | } |
| 1272 | |