void GrVkGpuRTCommandBuffer::executeDrawable(std::unique_ptr<SkDrawable::GpuDrawHandler> drawable) {
GrVkRenderTarget* target = static_cast<GrVkRenderTarget*>(fRenderTarget);
GrVkImage* targetImage = target->msaaImage() ? target->msaaImage() : target;
CommandBufferInfo& cbInfo = fCommandBufferInfos[fCurrentCmdInfo];
VkRect2D bounds;
bounds.offset = { 0, 0 };
bounds.extent = { 0, 0 };
GrVkDrawableInfo vkInfo;
vkInfo.fSecondaryCommandBuffer = cbInfo.currentCmdBuf()->vkCommandBuffer();
vkInfo.fCompatibleRenderPass = cbInfo.fRenderPass->vkRenderPass();
SkAssertResult(cbInfo.fRenderPass->colorAttachmentIndex(&vkInfo.fColorAttachmentIndex));
vkInfo.fFormat = targetImage->imageFormat();
vkInfo.fDrawBounds = &bounds;
#ifdef SK_BUILD_FOR_ANDROID_FRAMEWORK
vkInfo.fImage = targetImage->image();
#else
vkInfo.fImage = VK_NULL_HANDLE;
#endif //SK_BUILD_FOR_ANDROID_FRAMEWORK
GrBackendDrawableInfo info(vkInfo);
// After we draw into the command buffer via the drawable, cached state we have may be invalid.
cbInfo.currentCmdBuf()->invalidateState();
// Also assume that the drawable produced output.
cbInfo.fIsEmpty = false;
drawable->draw(info);
fGpu->addDrawable(std::move(drawable));
if (bounds.extent.width == 0 || bounds.extent.height == 0) {
cbInfo.fBounds.join(target->getBoundsRect());
} else {
cbInfo.fBounds.join(SkRect::MakeXYWH(bounds.offset.x, bounds.offset.y,
bounds.extent.width, bounds.extent.height));
}
}
void GrVkGpuRTCommandBuffer::submit() {
if (!fRenderTarget) {
return;
}
GrVkRenderTarget* vkRT = static_cast<GrVkRenderTarget*>(fRenderTarget);
GrVkImage* targetImage = vkRT->msaaImage() ? vkRT->msaaImage() : vkRT;
GrStencilAttachment* stencil = fRenderTarget->renderTargetPriv().getStencilAttachment();
auto currPreCmd = fPreCommandBufferTasks.begin();
GrVkPrimaryCommandBufferTask::Args taskArgs{fGpu, fRenderTarget, fOrigin};
for (int i = 0; i < fCommandBufferInfos.count(); ++i) {
CommandBufferInfo& cbInfo = fCommandBufferInfos[i];
for (int c = 0; c < cbInfo.fNumPreCmds; ++c, ++currPreCmd) {
currPreCmd->execute(taskArgs);
}
// TODO: Many things create a scratch texture which adds the discard immediately, but then
// don't draw to it right away. This causes the discard to be ignored and we get yelled at
// for loading uninitialized data. However, once MDB lands with reordering, the discard will
// get reordered with the rest of the draw commands and we can remove the discard check.
if (cbInfo.fIsEmpty &&
cbInfo.fLoadStoreState != LoadStoreState::kStartsWithClear &&
cbInfo.fLoadStoreState != LoadStoreState::kStartsWithDiscard) {
// We have sumbitted no actual draw commands to the command buffer and we are not using
// the render pass to do a clear so there is no need to submit anything.
continue;
}
// We don't want to actually submit the secondary command buffer if it is wrapped.
if (this->wrapsSecondaryCommandBuffer()) {
// If we have any sampled images set their layout now.
for (int j = 0; j < cbInfo.fSampledTextures.count(); ++j) {
cbInfo.fSampledTextures[j]->setImageLayout(
fGpu, VK_IMAGE_LAYOUT_SHADER_READ_ONLY_OPTIMAL, VK_ACCESS_SHADER_READ_BIT,
VK_PIPELINE_STAGE_FRAGMENT_SHADER_BIT, false);
}
// There should have only been one secondary command buffer in the wrapped case so it is
// safe to just return here.
SkASSERT(fCommandBufferInfos.count() == 1);
return;
}
// Make sure if we only have a discard load that we execute the discard on the whole image.
// TODO: Once we improve our tracking of discards so that we never end up flushing a discard
// call with no actually ops, remove this.
if (cbInfo.fIsEmpty && cbInfo.fLoadStoreState == LoadStoreState::kStartsWithDiscard) {
cbInfo.fBounds = SkRect::MakeWH(vkRT->width(), vkRT->height());
}
if (cbInfo.fBounds.intersect(0, 0,
SkIntToScalar(fRenderTarget->width()),
SkIntToScalar(fRenderTarget->height()))) {
// Make sure we do the following layout changes after all copies, uploads, or any other
// pre-work is done since we may change the layouts in the pre-work. Also since the
// draws will be submitted in different render passes, we need to guard againts write
// and write issues.
// Change layout of our render target so it can be used as the color attachment.
// TODO: If we know that we will never be blending or loading the attachment we could
// drop the VK_ACCESS_COLOR_ATTACHMENT_READ_BIT.
targetImage->setImageLayout(fGpu,
VK_IMAGE_LAYOUT_COLOR_ATTACHMENT_OPTIMAL,
VK_ACCESS_COLOR_ATTACHMENT_READ_BIT |
VK_ACCESS_COLOR_ATTACHMENT_WRITE_BIT,
VK_PIPELINE_STAGE_COLOR_ATTACHMENT_OUTPUT_BIT,
false);
// If we are using a stencil attachment we also need to update its layout
if (stencil) {
GrVkStencilAttachment* vkStencil = (GrVkStencilAttachment*)stencil;
// We need the write and read access bits since we may load and store the stencil.
// The initial load happens in the VK_PIPELINE_STAGE_EARLY_FRAGMENT_TESTS_BIT so we
// wait there.
vkStencil->setImageLayout(fGpu,
VK_IMAGE_LAYOUT_DEPTH_STENCIL_ATTACHMENT_OPTIMAL,
VK_ACCESS_DEPTH_STENCIL_ATTACHMENT_WRITE_BIT |
VK_ACCESS_DEPTH_STENCIL_ATTACHMENT_READ_BIT,
VK_PIPELINE_STAGE_EARLY_FRAGMENT_TESTS_BIT,
false);
}
// If we have any sampled images set their layout now.
for (int j = 0; j < cbInfo.fSampledTextures.count(); ++j) {
cbInfo.fSampledTextures[j]->setImageLayout(
fGpu, VK_IMAGE_LAYOUT_SHADER_READ_ONLY_OPTIMAL, VK_ACCESS_SHADER_READ_BIT,
VK_PIPELINE_STAGE_FRAGMENT_SHADER_BIT, false);
}
SkIRect iBounds;
cbInfo.fBounds.roundOut(&iBounds);
fGpu->submitSecondaryCommandBuffer(cbInfo.fCommandBuffers, cbInfo.fRenderPass,
&cbInfo.fColorClearValue, vkRT, fOrigin, iBounds);
}
}
SkASSERT(currPreCmd == fPreCommandBufferTasks.end());
}
bool GrVkCopyManager::copySurfaceAsDraw(GrVkGpu* gpu,
GrSurface* dst, GrSurfaceOrigin dstOrigin,
GrSurface* src, GrSurfaceOrigin srcOrigin,
const SkIRect& srcRect, const SkIPoint& dstPoint,
bool canDiscardOutsideDstRect) {
// None of our copy methods can handle a swizzle. TODO: Make copySurfaceAsDraw handle the
// swizzle.
if (gpu->caps()->shaderCaps()->configOutputSwizzle(src->config()) !=
gpu->caps()->shaderCaps()->configOutputSwizzle(dst->config())) {
return false;
}
GrVkRenderTarget* rt = static_cast<GrVkRenderTarget*>(dst->asRenderTarget());
if (!rt) {
return false;
}
GrVkTexture* srcTex = static_cast<GrVkTexture*>(src->asTexture());
if (!srcTex) {
return false;
}
if (VK_NULL_HANDLE == fVertShaderModule) {
SkASSERT(VK_NULL_HANDLE == fFragShaderModule &&
nullptr == fPipelineLayout &&
nullptr == fVertexBuffer.get() &&
nullptr == fUniformBuffer.get());
if (!this->createCopyProgram(gpu)) {
SkDebugf("Failed to create copy program.\n");
return false;
}
}
SkASSERT(fPipelineLayout);
GrVkResourceProvider& resourceProv = gpu->resourceProvider();
GrVkCopyPipeline* pipeline = resourceProv.findOrCreateCopyPipeline(rt,
fShaderStageInfo,
fPipelineLayout->layout());
if (!pipeline) {
return false;
}
// UPDATE UNIFORM DESCRIPTOR SET
int w = srcRect.width();
int h = srcRect.height();
// dst rect edges in NDC (-1 to 1)
int dw = dst->width();
int dh = dst->height();
float dx0 = 2.f * dstPoint.fX / dw - 1.f;
float dx1 = 2.f * (dstPoint.fX + w) / dw - 1.f;
float dy0 = 2.f * dstPoint.fY / dh - 1.f;
float dy1 = 2.f * (dstPoint.fY + h) / dh - 1.f;
if (kBottomLeft_GrSurfaceOrigin == dstOrigin) {
dy0 = -dy0;
dy1 = -dy1;
}
float sx0 = (float)srcRect.fLeft;
float sx1 = (float)(srcRect.fLeft + w);
float sy0 = (float)srcRect.fTop;
float sy1 = (float)(srcRect.fTop + h);
int sh = src->height();
if (kBottomLeft_GrSurfaceOrigin == srcOrigin) {
sy0 = sh - sy0;
sy1 = sh - sy1;
}
// src rect edges in normalized texture space (0 to 1).
int sw = src->width();
sx0 /= sw;
sx1 /= sw;
sy0 /= sh;
sy1 /= sh;
float uniData[] = { dx1 - dx0, dy1 - dy0, dx0, dy0, // posXform
sx1 - sx0, sy1 - sy0, sx0, sy0 }; // texCoordXform
fUniformBuffer->updateData(gpu, uniData, sizeof(uniData), nullptr);
const GrVkDescriptorSet* uniformDS = resourceProv.getUniformDescriptorSet();
SkASSERT(uniformDS);
VkDescriptorBufferInfo uniBufferInfo;
uniBufferInfo.buffer = fUniformBuffer->buffer();
uniBufferInfo.offset = fUniformBuffer->offset();
uniBufferInfo.range = fUniformBuffer->size();
VkWriteDescriptorSet descriptorWrites;
descriptorWrites.sType = VK_STRUCTURE_TYPE_WRITE_DESCRIPTOR_SET;
descriptorWrites.pNext = nullptr;
descriptorWrites.dstSet = uniformDS->descriptorSet();
descriptorWrites.dstBinding = GrVkUniformHandler::kGeometryBinding;
descriptorWrites.dstArrayElement = 0;
descriptorWrites.descriptorCount = 1;
descriptorWrites.descriptorType = VK_DESCRIPTOR_TYPE_UNIFORM_BUFFER;
descriptorWrites.pImageInfo = nullptr;
descriptorWrites.pBufferInfo = &uniBufferInfo;
descriptorWrites.pTexelBufferView = nullptr;
GR_VK_CALL(gpu->vkInterface(), UpdateDescriptorSets(gpu->device(),
1,
&descriptorWrites,
0, nullptr));
// UPDATE SAMPLER DESCRIPTOR SET
const GrVkDescriptorSet* samplerDS =
gpu->resourceProvider().getSamplerDescriptorSet(fSamplerDSHandle);
GrSamplerState samplerState = GrSamplerState::ClampNearest();
GrVkSampler* sampler = resourceProv.findOrCreateCompatibleSampler(
samplerState, GrVkYcbcrConversionInfo());
VkDescriptorImageInfo imageInfo;
memset(&imageInfo, 0, sizeof(VkDescriptorImageInfo));
imageInfo.sampler = sampler->sampler();
imageInfo.imageView = srcTex->textureView()->imageView();
imageInfo.imageLayout = VK_IMAGE_LAYOUT_SHADER_READ_ONLY_OPTIMAL;
VkWriteDescriptorSet writeInfo;
memset(&writeInfo, 0, sizeof(VkWriteDescriptorSet));
writeInfo.sType = VK_STRUCTURE_TYPE_WRITE_DESCRIPTOR_SET;
writeInfo.pNext = nullptr;
writeInfo.dstSet = samplerDS->descriptorSet();
writeInfo.dstBinding = 0;
writeInfo.dstArrayElement = 0;
writeInfo.descriptorCount = 1;
writeInfo.descriptorType = VK_DESCRIPTOR_TYPE_COMBINED_IMAGE_SAMPLER;
writeInfo.pImageInfo = &imageInfo;
writeInfo.pBufferInfo = nullptr;
writeInfo.pTexelBufferView = nullptr;
GR_VK_CALL(gpu->vkInterface(), UpdateDescriptorSets(gpu->device(),
1,
&writeInfo,
0, nullptr));
VkDescriptorSet vkDescSets[] = { uniformDS->descriptorSet(), samplerDS->descriptorSet() };
GrVkRenderTarget* texRT = static_cast<GrVkRenderTarget*>(srcTex->asRenderTarget());
if (texRT) {
gpu->resolveRenderTargetNoFlush(texRT);
}
// TODO: Make tighter bounds and then adjust bounds for origin and granularity if we see
// any perf issues with using the whole bounds
SkIRect bounds = SkIRect::MakeWH(rt->width(), rt->height());
// Change layouts of rt and texture. We aren't blending so we don't need color attachment read
// access for blending.
GrVkImage* targetImage = rt->msaaImage() ? rt->msaaImage() : rt;
VkAccessFlags dstAccessFlags = VK_ACCESS_COLOR_ATTACHMENT_WRITE_BIT;
if (!canDiscardOutsideDstRect) {
// We need to load the color attachment so need to be able to read it.
dstAccessFlags |= VK_ACCESS_COLOR_ATTACHMENT_READ_BIT;
}
targetImage->setImageLayout(gpu,
VK_IMAGE_LAYOUT_COLOR_ATTACHMENT_OPTIMAL,
dstAccessFlags,
VK_PIPELINE_STAGE_COLOR_ATTACHMENT_OUTPUT_BIT,
false);
srcTex->setImageLayout(gpu,
VK_IMAGE_LAYOUT_SHADER_READ_ONLY_OPTIMAL,
VK_ACCESS_SHADER_READ_BIT,
VK_PIPELINE_STAGE_FRAGMENT_SHADER_BIT,
false);
GrStencilAttachment* stencil = rt->renderTargetPriv().getStencilAttachment();
if (stencil) {
GrVkStencilAttachment* vkStencil = (GrVkStencilAttachment*)stencil;
// We aren't actually using the stencil but we still load and store it so we need
// appropriate barriers.
// TODO: Once we refactor surface and how we conntect stencil to RTs, we should not even
// have the stencil on this render pass if possible.
vkStencil->setImageLayout(gpu,
VK_IMAGE_LAYOUT_DEPTH_STENCIL_ATTACHMENT_OPTIMAL,
VK_ACCESS_DEPTH_STENCIL_ATTACHMENT_READ_BIT |
VK_ACCESS_DEPTH_STENCIL_ATTACHMENT_WRITE_BIT,
VK_PIPELINE_STAGE_EARLY_FRAGMENT_TESTS_BIT,
false);
}
VkAttachmentLoadOp loadOp = canDiscardOutsideDstRect ? VK_ATTACHMENT_LOAD_OP_DONT_CARE
: VK_ATTACHMENT_LOAD_OP_LOAD;
GrVkRenderPass::LoadStoreOps vkColorOps(loadOp, VK_ATTACHMENT_STORE_OP_STORE);
GrVkRenderPass::LoadStoreOps vkStencilOps(VK_ATTACHMENT_LOAD_OP_LOAD,
VK_ATTACHMENT_STORE_OP_STORE);
const GrVkRenderPass* renderPass;
const GrVkResourceProvider::CompatibleRPHandle& rpHandle = rt->compatibleRenderPassHandle();
if (rpHandle.isValid()) {
renderPass = gpu->resourceProvider().findRenderPass(rpHandle,
vkColorOps,
vkStencilOps);
} else {
renderPass = gpu->resourceProvider().findRenderPass(*rt,
vkColorOps,
vkStencilOps);
}
SkASSERT(renderPass->isCompatible(*rt->simpleRenderPass()));
GrVkPrimaryCommandBuffer* cmdBuffer = gpu->currentCommandBuffer();
cmdBuffer->beginRenderPass(gpu, renderPass, nullptr, *rt, bounds, true);
GrVkSecondaryCommandBuffer* secondary = gpu->cmdPool()->findOrCreateSecondaryCommandBuffer(gpu);
if (!secondary) {
return false;
}
secondary->begin(gpu, rt->framebuffer(), renderPass);
secondary->bindPipeline(gpu, pipeline);
// Uniform DescriptorSet, Sampler DescriptorSet, and vertex shader uniformBuffer
SkSTArray<3, const GrVkRecycledResource*> descriptorRecycledResources;
descriptorRecycledResources.push_back(uniformDS);
descriptorRecycledResources.push_back(samplerDS);
descriptorRecycledResources.push_back(fUniformBuffer->resource());
// One sampler, texture view, and texture
SkSTArray<3, const GrVkResource*> descriptorResources;
descriptorResources.push_back(sampler);
descriptorResources.push_back(srcTex->textureView());
descriptorResources.push_back(srcTex->resource());
secondary->bindDescriptorSets(gpu,
descriptorRecycledResources,
descriptorResources,
fPipelineLayout,
0,
2,
vkDescSets,
0,
nullptr);
// Set Dynamic viewport and stencil
// We always use one viewport the size of the RT
VkViewport viewport;
viewport.x = 0.0f;
viewport.y = 0.0f;
viewport.width = SkIntToScalar(rt->width());
viewport.height = SkIntToScalar(rt->height());
viewport.minDepth = 0.0f;
viewport.maxDepth = 1.0f;
secondary->setViewport(gpu, 0, 1, &viewport);
// We assume the scissor is not enabled so just set it to the whole RT
VkRect2D scissor;
scissor.extent.width = rt->width();
scissor.extent.height = rt->height();
scissor.offset.x = 0;
scissor.offset.y = 0;
secondary->setScissor(gpu, 0, 1, &scissor);
secondary->bindInputBuffer(gpu, 0, fVertexBuffer.get());
secondary->draw(gpu, 4, 1, 0, 0);
secondary->end(gpu);
cmdBuffer->executeCommands(gpu, secondary);
cmdBuffer->endRenderPass(gpu);
secondary->unref(gpu);
// Release all temp resources which should now be reffed by the cmd buffer
pipeline->unref(gpu);
uniformDS->unref(gpu);
samplerDS->unref(gpu);
sampler->unref(gpu);
renderPass->unref(gpu);
return true;
}