ccpr: Prefer atlas sizes under 4k on ARM

Bug: skia:
Change-Id: Ib5afb84647efe2e64a3ec2f9da422b39228431e9
Reviewed-on: https://skia-review.googlesource.com/108871
Commit-Queue: Chris Dalton <csmartdalton@google.com>
Reviewed-by: Brian Salomon <bsalomon@google.com>

9 files changed, 49 insertions, 24 deletions

diff --git a/include/gpu/GrCaps.h b/include/gpu/GrCaps.h
index 9d6ef0deaf..e8449bbcb4 100644
--- a/include/gpu/GrCaps.h
+++ b/include/gpu/GrCaps.h
@@ -133,7 +133,13 @@ public:
     int maxVertexAttributes() const { return fMaxVertexAttributes; }
 
     int maxRenderTargetSize() const { return fMaxRenderTargetSize; }
+
+    /** This is the largest render target size that can be used without incurring extra perfomance
+        cost. It is usually the max RT size, unless larger render targets are known to be slower. */
+    int maxPreferredRenderTargetSize() const { return fMaxPreferredRenderTargetSize; }
+
     int maxTextureSize() const { return fMaxTextureSize; }
+
     /** This is the maximum tile size to use by GPU devices for rendering sw-backed images/bitmaps.
         It is usually the max texture size, unless we're overriding it for testing. */
     int maxTileSize() const { SkASSERT(fMaxTileSize <= fMaxTextureSize); return fMaxTileSize; }
@@ -276,6 +282,7 @@ protected:
     int fBufferMapThreshold;
 
     int fMaxRenderTargetSize;
+    int fMaxPreferredRenderTargetSize;
     int fMaxVertexAttributes;
     int fMaxTextureSize;
     int fMaxTileSize;
diff --git a/src/gpu/GrCaps.cpp b/src/gpu/GrCaps.cpp
index 08e26bd1b5..8cfbc5faea 100644
--- a/src/gpu/GrCaps.cpp
+++ b/src/gpu/GrCaps.cpp
@@ -68,6 +68,7 @@ GrCaps::GrCaps(const GrContextOptions& options) {
 
     fMaxVertexAttributes = 0;
     fMaxRenderTargetSize = 1;
+    fMaxPreferredRenderTargetSize = 1;
     fMaxTextureSize = 1;
     fMaxRasterSamples = 0;
     fMaxWindowRectangles = 0;
@@ -179,6 +180,7 @@ void GrCaps::dumpJSON(SkJSONWriter* writer) const {
     writer->appendS32("Max Vertex Attributes", fMaxVertexAttributes);
     writer->appendS32("Max Texture Size", fMaxTextureSize);
     writer->appendS32("Max Render Target Size", fMaxRenderTargetSize);
+    writer->appendS32("Max Preferred Render Target Size", fMaxPreferredRenderTargetSize);
     writer->appendS32("Max Raster Samples", fMaxRasterSamples);
     writer->appendS32("Max Window Rectangles", fMaxWindowRectangles);
     writer->appendS32("Max Clip Analytic Fragment Processors", fMaxClipAnalyticFPs);
diff --git a/src/gpu/ccpr/GrCCAtlas.cpp b/src/gpu/ccpr/GrCCAtlas.cpp
index 97b28fb480..ce5cddb6d4 100644
--- a/src/gpu/ccpr/GrCCAtlas.cpp
+++ b/src/gpu/ccpr/GrCCAtlas.cpp
@@ -18,6 +18,9 @@
 #include "ccpr/GrCCPathParser.h"
 #include "ops/GrDrawOp.h"
 
+static constexpr int kAtlasMinSize = 1024;
+static constexpr int kPadding = 1;
+
 class GrCCAtlas::Node {
 public:
     Node(std::unique_ptr<Node> previous, int l, int t, int r, int b)
@@ -25,10 +28,15 @@ public:
 
     Node* previous() const { return fPrevious.get(); }
 
-    bool addRect(int w, int h, SkIPoint16* loc) {
-        static constexpr int kPad = 1;
-
-        if (!fRectanizer.addRect(w + kPad, h + kPad, loc)) {
+    bool addRect(int w, int h, SkIPoint16* loc, int maxAtlasSize) {
+        // Pad all paths except those that are expected to take up an entire physical texture.
+        if (w < maxAtlasSize) {
+            w = SkTMin(w + kPadding, maxAtlasSize);
+        }
+        if (h < maxAtlasSize) {
+            h = SkTMin(h + kPadding, maxAtlasSize);
+        }
+        if (!fRectanizer.addRect(w, h, loc)) {
             return false;
         }
         loc->fX += fX;
@@ -76,18 +84,20 @@ private:
     typedef GrDrawOp INHERITED;
 };
 
-GrCCAtlas::GrCCAtlas(const GrCaps& caps, int minWidth, int minHeight)
-        : fMaxAtlasSize(caps.maxRenderTargetSize()), fDrawBounds{0, 0} {
-    SkASSERT(fMaxAtlasSize <= caps.maxTextureSize());
-    SkASSERT(SkTMax(minWidth, minHeight) <= fMaxAtlasSize);
-    int initialSize = GrNextPow2(SkTMax(minWidth, minHeight));
-    initialSize = SkTMax(int(kMinSize), initialSize);
+GrCCAtlas::GrCCAtlas(const GrCaps& caps, int minSize)
+        : fMaxAtlasSize(SkTMax(minSize, caps.maxPreferredRenderTargetSize())) {
+    // Caller should have cropped any paths to the destination render target instead of asking for
+    // an atlas larger than maxRenderTargetSize.
+    SkASSERT(fMaxAtlasSize <= caps.maxRenderTargetSize());
+    int initialSize = GrNextPow2(minSize + kPadding);
+    initialSize = SkTMax(kAtlasMinSize, initialSize);
     initialSize = SkTMin(initialSize, fMaxAtlasSize);
     fHeight = fWidth = initialSize;
-    fTopNode = skstd::make_unique<Node>(nullptr, 0, 0, initialSize, initialSize);
+    fTopNode = skstd::make_unique<Node>(nullptr, 0, 0, fWidth, fHeight);
 }
 
-GrCCAtlas::~GrCCAtlas() {}
+GrCCAtlas::~GrCCAtlas() {
+}
 
 bool GrCCAtlas::addRect(int w, int h, SkIPoint16* loc) {
     // This can't be called anymore once setCoverageCountBatchID() has been called.
@@ -104,17 +114,14 @@ bool GrCCAtlas::addRect(int w, int h, SkIPoint16* loc) {
 }
 
 bool GrCCAtlas::internalPlaceRect(int w, int h, SkIPoint16* loc) {
-    SkASSERT(SkTMax(w, h) < fMaxAtlasSize);
-
     for (Node* node = fTopNode.get(); node; node = node->previous()) {
-        if (node->addRect(w, h, loc)) {
+        if (node->addRect(w, h, loc, fMaxAtlasSize)) {
             return true;
         }
     }
 
     // The rect didn't fit. Grow the atlas and try again.
     do {
-        SkASSERT(SkTMax(fWidth, fHeight) <= fMaxAtlasSize);
         if (fWidth == fMaxAtlasSize && fHeight == fMaxAtlasSize) {
             return false;
         }
@@ -127,7 +134,7 @@ bool GrCCAtlas::internalPlaceRect(int w, int h, SkIPoint16* loc) {
             fWidth = SkTMin(fWidth * 2, fMaxAtlasSize);
             fTopNode = skstd::make_unique<Node>(std::move(fTopNode), left, 0, fWidth, fHeight);
         }
-    } while (!fTopNode->addRect(w, h, loc));
+    } while (!fTopNode->addRect(w, h, loc, fMaxAtlasSize));
 
     return true;
 }
diff --git a/src/gpu/ccpr/GrCCAtlas.h b/src/gpu/ccpr/GrCCAtlas.h
index cac82af894..184022ee96 100644
--- a/src/gpu/ccpr/GrCCAtlas.h
+++ b/src/gpu/ccpr/GrCCAtlas.h
@@ -26,11 +26,9 @@ struct SkIPoint16;
  */
 class GrCCAtlas {
 public:
-    static constexpr int kMinSize = 1024;
-
     using CoverageCountBatchID = int;
 
-    GrCCAtlas(const GrCaps&, int minWidth, int minHeight);
+    GrCCAtlas(const GrCaps&, int minSize);
     ~GrCCAtlas();
 
     bool addRect(int devWidth, int devHeight, SkIPoint16* loc);
@@ -55,10 +53,9 @@ private:
 
     const int fMaxAtlasSize;
 
-    int fWidth;
-    int fHeight;
-    SkISize fDrawBounds;
+    int fWidth, fHeight;
     std::unique_ptr<Node> fTopNode;
+    SkISize fDrawBounds = {0, 0};
 
     CoverageCountBatchID fCoverageCountBatchID SkDEBUGCODE(= 0);
     sk_sp<GrTextureProxy> fTextureProxy;
diff --git a/src/gpu/ccpr/GrCoverageCountingPathRenderer.cpp b/src/gpu/ccpr/GrCoverageCountingPathRenderer.cpp
index ec3681dfcf..90d89dec60 100644
--- a/src/gpu/ccpr/GrCoverageCountingPathRenderer.cpp
+++ b/src/gpu/ccpr/GrCoverageCountingPathRenderer.cpp
@@ -486,7 +486,8 @@ GrCCAtlas* GrCoverageCountingPathRenderer::placeParsedPathInAtlas(
             auto coverageCountBatchID = fPerFlushPathParser->closeCurrentBatch();
             fPerFlushAtlases.back().setCoverageCountBatchID(coverageCountBatchID);
         }
-        fPerFlushAtlases.emplace_back(*onFlushRP->caps(), w, h).addRect(w, h, &atlasLocation);
+        fPerFlushAtlases.emplace_back(*onFlushRP->caps(), SkTMax(w, h));
+        SkAssertResult(fPerFlushAtlases.back().addRect(w, h, &atlasLocation));
     }
 
     *atlasOffsetX = atlasLocation.x() - static_cast<int16_t>(clippedPathIBounds.left());
diff --git a/src/gpu/gl/GrGLCaps.cpp b/src/gpu/gl/GrGLCaps.cpp
index d71194ed6f..3bdf5d0e60 100644
--- a/src/gpu/gl/GrGLCaps.cpp
+++ b/src/gpu/gl/GrGLCaps.cpp
@@ -483,6 +483,12 @@ void GrGLCaps::init(const GrContextOptions& contextOptions,
     // Our render targets are always created with textures as the color
     // attachment, hence this min:
     fMaxRenderTargetSize = SkTMin(fMaxTextureSize, fMaxRenderTargetSize);
+    fMaxPreferredRenderTargetSize = fMaxRenderTargetSize;
+
+    if (kARM_GrGLVendor == ctxInfo.vendor()) {
+        // On Mali G71, RT's above 4k have been observed to incur a performance cost.
+        fMaxPreferredRenderTargetSize = SkTMin(4096, fMaxPreferredRenderTargetSize);
+    }
 
     fGpuTracingSupport = ctxInfo.hasExtension("GL_EXT_debug_marker");
 
diff --git a/src/gpu/mock/GrMockCaps.h b/src/gpu/mock/GrMockCaps.h
index d96a690b31..e75a2c63ff 100644
--- a/src/gpu/mock/GrMockCaps.h
+++ b/src/gpu/mock/GrMockCaps.h
@@ -20,6 +20,7 @@ public:
         fBufferMapThreshold = SK_MaxS32; // Overridable in GrContextOptions.
         fMaxTextureSize = options.fMaxTextureSize;
         fMaxRenderTargetSize = SkTMin(options.fMaxRenderTargetSize, fMaxTextureSize);
+        fMaxPreferredRenderTargetSize = fMaxRenderTargetSize;
         fMaxVertexAttributes = options.fMaxVertexAttributes;
 
         fShaderCaps.reset(new GrShaderCaps(contextOptions));
diff --git a/src/gpu/mtl/GrMtlCaps.mm b/src/gpu/mtl/GrMtlCaps.mm
index 865aa37fe6..93d801fd72 100644
--- a/src/gpu/mtl/GrMtlCaps.mm
+++ b/src/gpu/mtl/GrMtlCaps.mm
@@ -118,6 +118,7 @@ void GrMtlCaps::initGrCaps(const id<MTLDevice> device) {
             }
         }
     }
+    fMaxPreferredRenderTargetSize = fMaxRenderTargetSize;
     fMaxTextureSize = fMaxRenderTargetSize;
 
     // Init sample counts. All devices support 1 (i.e. 0 in skia).
diff --git a/src/gpu/vk/GrVkCaps.cpp b/src/gpu/vk/GrVkCaps.cpp
index 3fc5a08acb..88d820a62c 100644
--- a/src/gpu/vk/GrVkCaps.cpp
+++ b/src/gpu/vk/GrVkCaps.cpp
@@ -165,6 +165,9 @@ void GrVkCaps::initGrCaps(const VkPhysicalDeviceProperties& properties,
     fMaxRenderTargetSize = SkTMin(properties.limits.maxImageDimension2D, (uint32_t)INT_MAX);
     fMaxTextureSize = SkTMin(properties.limits.maxImageDimension2D, (uint32_t)INT_MAX);
 
+    // TODO: check if RT's larger than 4k incur a performance cost on ARM.
+    fMaxPreferredRenderTargetSize = fMaxRenderTargetSize;
+
     // Assuming since we will always map in the end to upload the data we might as well just map
     // from the get go. There is no hard data to suggest this is faster or slower.
     fBufferMapThreshold = 0;