SkRasterPipeline preliminaries

src/core/SkRasterPipeline.h

+/*
+ * Copyright 2016 Google Inc.
+ *
+ * Use of this source code is governed by a BSD-style license that can be
+ * found in the LICENSE file.
+ */
+
+#ifndef SkRasterPipeline_DEFINED
+#define SkRasterPipeline_DEFINED
+
+#include "SkNx.h"
+#include "SkTArray.h"
+#include "SkTypes.h"
+
+/**
+ * SkRasterPipeline provide a cheap way to chain together a pixel processing pipeline.
+ *
+ * It's particularly designed for situations where the potential pipeline is extremely
+ * combinatoric: {N dst formats} x {M source formats} x {K mask formats} x {C transfer modes} ...
+ * No one wants to write specialized routines for all those combinations, and if we did, we'd
+ * end up bloating our code size dramatically.  SkRasterPipeline stages can be chained together
+ * at runtime, so we can scale this problem linearly rather than combinatorically.
+ *
+ * Each stage is represented by a function conforming to a common interface, SkRasterPipeline::Fn,
+ * and by an arbitrary context pointer (TODO: make this tidier with lambdas?).  Fn's arguments,
+ * and sometimes custom calling convention, are designed to maximize the amount of data we can
+ * pass along the pipeline cheaply.  On many machines they all stay in registers the entire time.
+ *
+ * The meaning of the arguments to Fn are sometimes fixed...
+ *    - The Stage* always represents the current stage, mainly providing access to ctx().
+ *    - The size_t is always the destination x coordinate.  If you need y, put it in your context.
+ *    - By the time the shader's done, the first four vectors should hold source red,
+ *      green, blue, and alpha, up to 4 pixels' worth each.
+ *
+ * ...and sometimes flexible:
+ *    - In the shader, the first four vectors can be used for anything, e.g. sample coordinates.
+ *    - The last four vectors are scratch registers that can be used to communicate between
+ *      stages; transfer modes use these to hold the original destination pixel components.
+ *
+ * On some platforms the last four vectors are slower to work with than the other arguments.
+ *
+ * When done mutating its arguments and/or context, a stage can either:
+ *   1) call st->next() with its mutated arguments, chaining to the next stage of the pipeline; or
+ *   2) return, indicating the pipeline is complete for these pixels.
+ *
+ * Some obvious stages that typically return are those that write a color to a destination pointer,
+ * but any stage can short-circuit the rest of the pipeline by returning instead of calling next().
+ */
+
+class SkRasterPipeline {
+public:
+    struct Stage;
+    using Fn = SK_VECTORCALL void(*)(Stage*, size_t, Sk4f,Sk4f,Sk4f,Sk4f,
+                                                     Sk4f,Sk4f,Sk4f,Sk4f);
+    struct Stage {
+        template <typename T = void>
+        T ctx() { return static_cast<T>(fCtx); }
+
+        SK_ALWAYS_INLINE void next(size_t x, Sk4f v0, Sk4f v1, Sk4f v2, Sk4f v3,
+                                             Sk4f v4, Sk4f v5, Sk4f v6, Sk4f v7) {
+            // Stages are logically a pipeline, and physically are contiguous in an array.
+            // To get to the next stage, we just increment our pointer to the next array element.
+            fNext(this+1, x, v0,v1,v2,v3, v4,v5,v6,v7);
+        }
+
+        // It makes next() a good bit cheaper if we hold the next function to call here,
+        // rather than logically simpler choice of the function implementing this stage.
+        Fn fNext;
+        void* fCtx;
+    };
+
+
+    SkRasterPipeline();
+
+    // Run the pipeline constructed with append(), walking x through [0,n),
+    // generally in 4 pixel steps, but sometimes 1 pixel at a time.
+    void run(size_t n);
+
+    // Use this append() if your stage is sensitive to the number of pixels you're working with:
+    //   - body will always be called for a full 4 pixels
+    //   - tail will always be called for a single pixel
+    // Typically this is only an essential distintion for stages that read or write memory.
+    void append(Fn body, const void* body_ctx,
+                Fn tail, const void* tail_ctx);
+
+    // Most stages don't actually care if they're working on 4 or 1 pixel.
+    void append(Fn fn, const void* ctx = nullptr) {
+        this->append(fn, ctx, fn, ctx);
+    }
+
+    // Most 4 pixel or 1 pixel variants share the same context pointer.
+    void append(Fn body, Fn tail, const void* ctx = nullptr) {
+        this->append(body, ctx, tail, ctx);
+    }
+
+private:
+    using Stages = SkSTArray<10, Stage, /*MEM_COPY=*/true>;
+
+    static void Rewire(Stages*);
+
+    Stages fBody,
+           fTail;
+    bool   fReadyToRun = false;
+};
+
+#endif//SkRasterPipeline_DEFINED