mirror of
https://github.com/hrydgard/ppsspp.git
synced 2026-07-11 01:25:07 +02:00
Vulkan: duplicate fields move to separate const reference var
This commit is contained in:
Herman Semenoff
committed by
Henrik Rydgård
Henrik Rydgård
parent
5f8e631167
commit
eaae4dca82
@@ -213,51 +213,54 @@ void VulkanQueueRunner::PreprocessSteps(std::vector<VKRStep *> &steps) {
|
||||
// substitute descriptors, alternatively using texture array layers creatively.
|
||||
|
||||
for (int j = 0; j < (int)steps.size(); j++) {
|
||||
decltype(steps[j]->render) &render = steps[j]->render;
|
||||
if (steps[j]->stepType == VKRStepType::RENDER &&
|
||||
steps[j]->render.framebuffer) {
|
||||
if (steps[j]->render.finalColorLayout == VK_IMAGE_LAYOUT_UNDEFINED) {
|
||||
steps[j]->render.finalColorLayout = VK_IMAGE_LAYOUT_COLOR_ATTACHMENT_OPTIMAL;
|
||||
render.framebuffer) {
|
||||
if (render.finalColorLayout == VK_IMAGE_LAYOUT_UNDEFINED) {
|
||||
render.finalColorLayout = VK_IMAGE_LAYOUT_COLOR_ATTACHMENT_OPTIMAL;
|
||||
}
|
||||
if (steps[j]->render.finalDepthStencilLayout == VK_IMAGE_LAYOUT_UNDEFINED) {
|
||||
steps[j]->render.finalDepthStencilLayout = VK_IMAGE_LAYOUT_DEPTH_STENCIL_ATTACHMENT_OPTIMAL;
|
||||
if (render.finalDepthStencilLayout == VK_IMAGE_LAYOUT_UNDEFINED) {
|
||||
render.finalDepthStencilLayout = VK_IMAGE_LAYOUT_DEPTH_STENCIL_ATTACHMENT_OPTIMAL;
|
||||
}
|
||||
}
|
||||
}
|
||||
|
||||
for (int j = 0; j < (int)steps.size() - 1; j++) {
|
||||
// Push down empty "Clear/Store" renderpasses, and merge them with the first "Load/Store" to the same framebuffer.
|
||||
decltype(steps[j]->render) &render_j = steps[j]->render;
|
||||
if (steps.size() > 1 && steps[j]->stepType == VKRStepType::RENDER &&
|
||||
steps[j]->render.numDraws == 0 &&
|
||||
steps[j]->render.numReads == 0 &&
|
||||
steps[j]->render.colorLoad == VKRRenderPassLoadAction::CLEAR &&
|
||||
steps[j]->render.stencilLoad == VKRRenderPassLoadAction::CLEAR &&
|
||||
steps[j]->render.depthLoad == VKRRenderPassLoadAction::CLEAR) {
|
||||
render_j.numDraws == 0 &&
|
||||
render_j.numReads == 0 &&
|
||||
render_j.colorLoad == VKRRenderPassLoadAction::CLEAR &&
|
||||
render_j.stencilLoad == VKRRenderPassLoadAction::CLEAR &&
|
||||
render_j.depthLoad == VKRRenderPassLoadAction::CLEAR) {
|
||||
|
||||
// Drop the clear step, and merge it into the next step that touches the same framebuffer.
|
||||
for (int i = j + 1; i < (int)steps.size(); i++) {
|
||||
decltype(steps[i]->render) &render_i = steps[i]->render;
|
||||
if (steps[i]->stepType == VKRStepType::RENDER &&
|
||||
steps[i]->render.framebuffer == steps[j]->render.framebuffer) {
|
||||
if (steps[i]->render.colorLoad != VKRRenderPassLoadAction::CLEAR) {
|
||||
steps[i]->render.colorLoad = VKRRenderPassLoadAction::CLEAR;
|
||||
steps[i]->render.clearColor = steps[j]->render.clearColor;
|
||||
render_i.framebuffer == render_j.framebuffer) {
|
||||
if (render_i.colorLoad != VKRRenderPassLoadAction::CLEAR) {
|
||||
render_i.colorLoad = VKRRenderPassLoadAction::CLEAR;
|
||||
render_i.clearColor = render_j.clearColor;
|
||||
}
|
||||
if (steps[i]->render.depthLoad != VKRRenderPassLoadAction::CLEAR) {
|
||||
steps[i]->render.depthLoad = VKRRenderPassLoadAction::CLEAR;
|
||||
steps[i]->render.clearDepth = steps[j]->render.clearDepth;
|
||||
if (render_i.depthLoad != VKRRenderPassLoadAction::CLEAR) {
|
||||
render_i.depthLoad = VKRRenderPassLoadAction::CLEAR;
|
||||
render_i.clearDepth = render_j.clearDepth;
|
||||
}
|
||||
if (steps[i]->render.stencilLoad != VKRRenderPassLoadAction::CLEAR) {
|
||||
steps[i]->render.stencilLoad = VKRRenderPassLoadAction::CLEAR;
|
||||
steps[i]->render.clearStencil = steps[j]->render.clearStencil;
|
||||
if (render_i.stencilLoad != VKRRenderPassLoadAction::CLEAR) {
|
||||
render_i.stencilLoad = VKRRenderPassLoadAction::CLEAR;
|
||||
render_i.clearStencil = render_j.clearStencil;
|
||||
}
|
||||
MergeRenderAreaRectInto(&steps[i]->render.renderArea, steps[j]->render.renderArea);
|
||||
steps[i]->render.renderPassType = MergeRPTypes(steps[i]->render.renderPassType, steps[j]->render.renderPassType);
|
||||
steps[i]->render.numDraws += steps[j]->render.numDraws;
|
||||
steps[i]->render.numReads += steps[j]->render.numReads;
|
||||
MergeRenderAreaRectInto(&render_i.renderArea, render_j.renderArea);
|
||||
render_i.renderPassType = MergeRPTypes(render_i.renderPassType, render_j.renderPassType);
|
||||
render_i.numDraws += render_j.numDraws;
|
||||
render_i.numReads += render_j.numReads;
|
||||
// Cheaply skip the first step.
|
||||
steps[j]->stepType = VKRStepType::RENDER_SKIP;
|
||||
break;
|
||||
} else if (steps[i]->stepType == VKRStepType::COPY &&
|
||||
steps[i]->copy.src == steps[j]->render.framebuffer) {
|
||||
steps[i]->copy.src == render_j.framebuffer) {
|
||||
// Can't eliminate the clear if a game copies from it before it's
|
||||
// rendered to. However this should be rare.
|
||||
// TODO: This should never happen when we check numReads now.
|
||||
@@ -404,17 +407,19 @@ void VulkanQueueRunner::ApplyMGSHack(std::vector<VKRStep *> &steps) {
|
||||
|
||||
for (int i = 0; i < (int)steps.size() - 3; i++) {
|
||||
int last = -1;
|
||||
const decltype(steps[i + 1]->render) &render_i_plus_1 = steps[i + 1]->render;
|
||||
if (!(steps[i]->stepType == VKRStepType::COPY &&
|
||||
steps[i + 1]->stepType == VKRStepType::RENDER &&
|
||||
steps[i + 2]->stepType == VKRStepType::COPY &&
|
||||
steps[i + 1]->render.numDraws == 1 &&
|
||||
render_i_plus_1.numDraws == 1 &&
|
||||
steps[i]->copy.dst == steps[i + 2]->copy.dst))
|
||||
continue;
|
||||
// Looks promising! Let's start by finding the last one.
|
||||
for (int j = i; j < (int)steps.size(); j++) {
|
||||
const decltype(steps[j]->render) &render_j = steps[j]->render;
|
||||
switch (steps[j]->stepType) {
|
||||
case VKRStepType::RENDER:
|
||||
if (steps[j]->render.numDraws > 1)
|
||||
if (render_j.numDraws > 1)
|
||||
last = j - 1;
|
||||
// should really also check descriptor sets...
|
||||
if (steps[j]->commands.size()) {
|
||||
@@ -480,41 +485,45 @@ void VulkanQueueRunner::ApplyMGSHack(std::vector<VKRStep *> &steps) {
|
||||
// of the game.
|
||||
for (int i = 0; i < (int)steps.size() - 3; i++) {
|
||||
int last = -1;
|
||||
const decltype(steps[i]->render) &render_i = steps[i]->render;
|
||||
const decltype(steps[i + 1]->render) &render_i_plus_1 = steps[i + 1]->render;
|
||||
const decltype(steps[i + 2]->render) &render_i_plus_2 = steps[i + 2]->render;
|
||||
if (!(steps[i]->stepType == VKRStepType::RENDER &&
|
||||
steps[i + 1]->stepType == VKRStepType::RENDER &&
|
||||
steps[i + 2]->stepType == VKRStepType::RENDER &&
|
||||
steps[i]->render.numDraws == 1 &&
|
||||
steps[i + 1]->render.numDraws == 1 &&
|
||||
steps[i + 2]->render.numDraws == 1 &&
|
||||
steps[i]->render.colorLoad == VKRRenderPassLoadAction::DONT_CARE &&
|
||||
steps[i + 1]->render.colorLoad == VKRRenderPassLoadAction::KEEP &&
|
||||
steps[i + 2]->render.colorLoad == VKRRenderPassLoadAction::DONT_CARE)) {
|
||||
render_i.numDraws == 1 &&
|
||||
render_i_plus_1.numDraws == 1 &&
|
||||
render_i_plus_2.numDraws == 1 &&
|
||||
render_i.colorLoad == VKRRenderPassLoadAction::DONT_CARE &&
|
||||
render_i_plus_1.colorLoad == VKRRenderPassLoadAction::KEEP &&
|
||||
render_i_plus_2.colorLoad == VKRRenderPassLoadAction::DONT_CARE)) {
|
||||
continue;
|
||||
}
|
||||
VKRFramebuffer *depalFramebuffer = steps[i]->render.framebuffer;
|
||||
VKRFramebuffer *targetFramebuffer = steps[i + 1]->render.framebuffer;
|
||||
VKRFramebuffer *depalFramebuffer = render_i.framebuffer;
|
||||
VKRFramebuffer *targetFramebuffer = render_i_plus_1.framebuffer;
|
||||
// OK, found the start of a post-process sequence. Let's scan until we find the end.
|
||||
for (int j = i; j < (int)steps.size() - 3; j++) {
|
||||
const decltype(steps[j]->render) &render_j = steps[j]->render;
|
||||
if (((j - i) & 1) == 0) {
|
||||
// This should be a depal draw.
|
||||
if (steps[j]->render.numDraws != 1)
|
||||
if (render_j.numDraws != 1)
|
||||
break;
|
||||
if (steps[j]->commands.size() > 5) // TODO: Not the greatest heuristic! This may change if we merge commands.
|
||||
if (steps[j]->commands.size() > 5)
|
||||
break;
|
||||
if (steps[j]->render.colorLoad != VKRRenderPassLoadAction::DONT_CARE)
|
||||
if (render_j.colorLoad != VKRRenderPassLoadAction::DONT_CARE)
|
||||
break;
|
||||
if (steps[j]->render.framebuffer != depalFramebuffer)
|
||||
if (render_j.framebuffer != depalFramebuffer)
|
||||
break;
|
||||
last = j;
|
||||
} else {
|
||||
// This should be a target draw.
|
||||
if (steps[j]->render.numDraws != 1)
|
||||
if (render_j.numDraws != 1)
|
||||
break;
|
||||
if (steps[j]->commands.size() > 5) // TODO: Not the greatest heuristic! This may change if we merge commands.
|
||||
if (steps[j]->commands.size() > 5)
|
||||
break;
|
||||
if (steps[j]->render.colorLoad != VKRRenderPassLoadAction::KEEP)
|
||||
if (render_j.colorLoad != VKRRenderPassLoadAction::KEEP)
|
||||
break;
|
||||
if (steps[j]->render.framebuffer != targetFramebuffer)
|
||||
if (render_j.framebuffer != targetFramebuffer)
|
||||
break;
|
||||
last = j;
|
||||
}
|
||||
@@ -536,6 +545,7 @@ void VulkanQueueRunner::ApplyMGSHack(std::vector<VKRStep *> &steps) {
|
||||
|
||||
// Combine the depal renders. Also record scissor bounds.
|
||||
for (int j = i + 2; j <= last + 1; j += 2) {
|
||||
const decltype(steps[j]->render) &render_j = steps[j]->render;
|
||||
for (int k = 0; k < (int)steps[j]->commands.size(); k++) {
|
||||
switch (steps[j]->commands[k].cmd) {
|
||||
case VKRRenderCommand::DRAW:
|
||||
@@ -564,7 +574,7 @@ void VulkanQueueRunner::ApplyMGSHack(std::vector<VKRStep *> &steps) {
|
||||
break;
|
||||
}
|
||||
}
|
||||
MergeRenderAreaRectInto(&steps[i]->render.renderArea, steps[j]->render.renderArea);
|
||||
MergeRenderAreaRectInto(&steps[i]->render.renderArea, render_j.renderArea);
|
||||
steps[j]->stepType = VKRStepType::RENDER_SKIP;
|
||||
}
|
||||
|
||||
@@ -586,6 +596,7 @@ void VulkanQueueRunner::ApplyMGSHack(std::vector<VKRStep *> &steps) {
|
||||
|
||||
// Combine the target renders.
|
||||
for (int j = i + 3; j <= last; j += 2) {
|
||||
const decltype(steps[j]->render) &render_j = steps[j]->render;
|
||||
for (int k = 0; k < (int)steps[j]->commands.size(); k++) {
|
||||
switch (steps[j]->commands[k].cmd) {
|
||||
case VKRRenderCommand::DRAW:
|
||||
@@ -596,7 +607,7 @@ void VulkanQueueRunner::ApplyMGSHack(std::vector<VKRStep *> &steps) {
|
||||
break;
|
||||
}
|
||||
}
|
||||
MergeRenderAreaRectInto(&steps[i + 1]->render.renderArea, steps[j]->render.renderArea);
|
||||
MergeRenderAreaRectInto(&steps[i + 1]->render.renderArea, render_j.renderArea);
|
||||
steps[j]->stepType = VKRStepType::RENDER_SKIP;
|
||||
}
|
||||
|
||||
@@ -613,35 +624,39 @@ void VulkanQueueRunner::ApplySonicHack(std::vector<VKRStep *> &steps) {
|
||||
|
||||
for (int i = 0; i < (int)steps.size() - 4; i++) {
|
||||
int last = -1;
|
||||
const decltype(steps[i]->render) &render_i = steps[i]->render;
|
||||
const decltype(steps[i + 1]->render) &render_i_plus_1 = steps[i + 1]->render;
|
||||
const decltype(steps[i + 2]->render) &render_i_plus_2 = steps[i + 2]->render;
|
||||
const decltype(steps[i + 3]->render) &render_i_plus_3 = steps[i + 3]->render;
|
||||
if (!(steps[i]->stepType == VKRStepType::RENDER &&
|
||||
steps[i + 1]->stepType == VKRStepType::RENDER &&
|
||||
steps[i + 2]->stepType == VKRStepType::RENDER &&
|
||||
steps[i + 3]->stepType == VKRStepType::RENDER &&
|
||||
steps[i]->render.numDraws == 3 &&
|
||||
steps[i + 1]->render.numDraws == 1 &&
|
||||
steps[i + 2]->render.numDraws == 6 &&
|
||||
steps[i + 3]->render.numDraws == 1 &&
|
||||
steps[i]->render.framebuffer == steps[i + 2]->render.framebuffer &&
|
||||
steps[i + 1]->render.framebuffer == steps[i + 3]->render.framebuffer))
|
||||
render_i.numDraws == 3 &&
|
||||
render_i_plus_1.numDraws == 1 &&
|
||||
render_i_plus_2.numDraws == 6 &&
|
||||
render_i_plus_3.numDraws == 1 &&
|
||||
render_i.framebuffer == render_i_plus_2.framebuffer &&
|
||||
render_i_plus_1.framebuffer == render_i_plus_3.framebuffer))
|
||||
continue;
|
||||
// Looks promising! Let's start by finding the last one.
|
||||
for (int j = i; j < (int)steps.size(); j++) {
|
||||
switch (steps[j]->stepType) {
|
||||
case VKRStepType::RENDER:
|
||||
const decltype(steps[j]->render) &render_j = steps[j]->render;
|
||||
if (steps[j]->stepType == VKRStepType::RENDER) {
|
||||
if ((j - i) & 1) {
|
||||
if (steps[j]->render.framebuffer != steps[i + 1]->render.framebuffer)
|
||||
if (render_j.framebuffer != render_i_plus_1.framebuffer)
|
||||
last = j - 1;
|
||||
if (steps[j]->render.numDraws != 1)
|
||||
if (render_j.numDraws != 1)
|
||||
last = j - 1;
|
||||
} else {
|
||||
if (steps[j]->render.framebuffer != steps[i]->render.framebuffer)
|
||||
if (render_j.framebuffer != render_i.framebuffer)
|
||||
last = j - 1;
|
||||
if (steps[j]->render.numDraws != 3 && steps[j]->render.numDraws != 6)
|
||||
if (render_j.numDraws != 3 && render_j.numDraws != 6)
|
||||
last = j - 1;
|
||||
}
|
||||
break;
|
||||
default:
|
||||
break;
|
||||
} else {
|
||||
// Not a render step, so break.
|
||||
last = j - 1;
|
||||
}
|
||||
if (last != -1)
|
||||
break;
|
||||
@@ -769,8 +784,9 @@ void VulkanQueueRunner::ApplyRenderPassMerge(std::vector<VKRStep *> &steps) {
|
||||
// Now, let's go through the steps. If we find one that is rendered to more than once,
|
||||
// we'll scan forward and slurp up any rendering that can be merged across.
|
||||
for (int i = 0; i < (int)steps.size(); i++) {
|
||||
if (steps[i]->stepType == VKRStepType::RENDER && counts[steps[i]->render.framebuffer] > 1) {
|
||||
auto fb = steps[i]->render.framebuffer;
|
||||
decltype(steps[i]->render) &render_i = steps[i]->render;
|
||||
if (steps[i]->stepType == VKRStepType::RENDER && counts[render_i.framebuffer] > 1) {
|
||||
auto fb = render_i.framebuffer;
|
||||
TinySet<VKRFramebuffer *, 8> touchedFramebuffers; // must be the same fast-size as the dependencies TinySet for annoying reasons.
|
||||
for (int j = i + 1; j < (int)steps.size(); j++) {
|
||||
// If any other passes are reading from this framebuffer as-is, we cancel the scan.
|
||||
@@ -890,7 +906,8 @@ void VulkanQueueRunner::LogRenderPass(const VKRStep &pass, bool verbose) {
|
||||
INFO_LOG(Log::G3D, "RENDER %s Begin(%s, draws: %d, %dx%d, %s, %s, %s)", pass.tag, framebuf, r.numDraws, w, h, RenderPassActionName(r.colorLoad), RenderPassActionName(r.depthLoad), RenderPassActionName(r.stencilLoad));
|
||||
// TODO: Log these in detail.
|
||||
for (int i = 0; i < (int)pass.preTransitions.size(); i++) {
|
||||
INFO_LOG(Log::G3D, " PRETRANSITION: %s %s -> %s", pass.preTransitions[i].fb->Tag(), AspectToString(pass.preTransitions[i].aspect), ImageLayoutToString(pass.preTransitions[i].targetLayout));
|
||||
const TransitionRequest &preTransition = pass.preTransitions[i];
|
||||
INFO_LOG(Log::G3D, " PRETRANSITION: %s %s -> %s", preTransition.fb->Tag(), AspectToString(preTransition.aspect), ImageLayoutToString(preTransition.targetLayout));
|
||||
}
|
||||
|
||||
if (verbose) {
|
||||
@@ -1871,4 +1888,4 @@ const char *VKRRenderCommandToString(VKRRenderCommand cmd) {
|
||||
} else {
|
||||
return "N/A";
|
||||
}
|
||||
}
|
||||
}
|
||||
Reference in New Issue
Block a user