Production quality fast image up/downsampler

src/core/SkConvolver.h

Index: src/core/SkConvolver.h
diff --git a/src/core/SkConvolver.h b/src/core/SkConvolver.h
new file mode 100644
index 0000000000000000000000000000000000000000..a6e98f8bc5ea6ed6914325794ee38bcb5842b218
--- /dev/null
+++ b/src/core/SkConvolver.h
@@ -0,0 +1,215 @@
+// Copyright (c) 2012 The Chromium Authors. All rights reserved.
+// Use of this source code is governed by a BSD-style license that can be
+// found in the LICENSE file.
+
+#ifndef SK_CONVOLVER_H
+#define SK_CONVOLVER_H
+
+#include <cmath>

[reed1, 2013/07/18 13:42:12]

must we bring in these (would be a first for skia)

[humper, 2013/07/18 17:11:04]

cmath can probably be replaced by math.h; the other C++ libraries will take some
rewriting.

+#include <vector>
+
+#include "SkSize.h"
+#include "SkTypes.h"
+
+// We can build SSE2 optimized versions for all x86 CPUs
+// except when building for the IOS emulator.
+#if defined(SKIA_SSE)
+#define SIMD_SSE2 1
+#define SIMD_PADDING 8  // 8 * int16
+#endif
+
+// avoid confusion with Mac OS X's math library (Carbon)
+#if defined(__APPLE__)
+#undef FloatToFixed
+#undef FixedToFloat
+#endif
+
+// Represents a filter in one dimension. Each output pixel has one entry in this
+// object for the filter values contributing to it. You build up the filter
+// list by calling AddFilter for each output pixel (in order).
+//
+// We do 2-dimensional convolution by first convolving each row by one
+// SkConvolutionFilter1D, then convolving each column by another one.
+//
+// Entries are stored in fixed point, shifted left by kShiftBits.
+class SkConvolutionFilter1D {
+public:
+    typedef short Fixed;
+
+    // The number of bits that fixed point values are shifted by.
+    enum { kShiftBits = 14 };
+
+    SK_API SkConvolutionFilter1D();
+    SK_API ~SkConvolutionFilter1D();
+
+    // Convert between floating point and our fixed point representation.
+    static Fixed FloatToFixed(float f) {
+        return static_cast<Fixed>(f * (1 << kShiftBits));
+    }
+    static unsigned char FixedToChar(Fixed x) {
+        return static_cast<unsigned char>(x >> kShiftBits);
+    }
+    static float FixedToFloat(Fixed x) {
+        // The cast relies on Fixed being a short, implying that on
+        // the platforms we care about all (16) bits will fit into
+        // the mantissa of a (32-bit) float.
+        SK_COMPILE_ASSERT(sizeof(Fixed) == 2, fixed_type_should_fit_in_float_mantissa);
+        float raw = static_cast<float>(x);
+        return ldexpf(raw, -kShiftBits);
+    }
+
+    // Returns the maximum pixel span of a filter.
+    int maxFilter() const { return fMaxFilter; }
+
+    // Returns the number of filters in this filter. This is the dimension of the
+    // output image.
+    int numValues() const { return static_cast<int>(fFilters.size()); }
+
+    // Appends the given list of scaling values for generating a given output
+    // pixel. |filterOffset| is the distance from the edge of the image to where
+    // the scaling factors start. The scaling factors apply to the source pixels
+    // starting from this position, and going for the next |filterLength| pixels.
+    //
+    // You will probably want to make sure your input is normalized (that is,
+    // all entries in |filterValuesg| sub to one) to prevent affecting the overall
+    // brighness of the image.
+    //
+    // The filterLength must be > 0.
+    //
+    // This version will automatically convert your input to fixed point.
+    SK_API void AddFilter(int filterOffset,
+                          const float* filterValues,
+                          int filterLength);
+
+    // Same as the above version, but the input is already fixed point.
+    void AddFilter(int filterOffset,
+                   const Fixed* filterValues,
+                   int filterLength);
+
+    // Retrieves a filter for the given |valueOffset|, a position in the output
+    // image in the direction we're convolving. The offset and length of the
+    // filter values are put into the corresponding out arguments (see AddFilter
+    // above for what these mean), and a pointer to the first scaling factor is
+    // returned. There will be |filterLength| values in this array.
+    inline const Fixed* FilterForValue(int valueOffset, 
+                                       int* filterOffset,
+                                       int* filterLength) const {
+        const FilterInstance& filter = fFilters[valueOffset];
+        *filterOffset = filter.fOffset;
+        *filterLength = filter.fTrimmedLength;
+        if (filter.fTrimmedLength == 0) {
+            return NULL;
+        }
+        return &fFilterValues[filter.fDataLocation];
+    }
+
+  // Retrieves the filter for the offset 0, presumed to be the one and only.
+  // The offset and length of the filter values are put into the corresponding
+  // out arguments (see AddFilter). Note that |filterLegth| and
+  // |specifiedFilterLength| may be different if leading/trailing zeros of the
+  // original floating point form were clipped.
+  // There will be |filterLength| values in the return array.
+  // Returns NULL if the filter is 0-length (for instance when all floating
+  // point values passed to AddFilter were clipped to 0).
+    SK_API const Fixed* GetSingleFilter(int* specifiedFilterLength,
+        int* filterOffset,
+        int* filterLength) const;
+
+    inline void PaddingForSIMD() {
+    // Padding |paddingCount| of more dummy coefficients after the coefficients
+    // of last filter to prevent SIMD instructions which load 8 or 16 bytes
+    // together to access invalid memory areas. We are not trying to align the
+    // coefficients right now due to the opaqueness of <vector> implementation.
+    // This has to be done after all |AddFilter| calls.
+#ifdef SIMD_PADDING
+        for (int i = 0; i < SIMD_PADDING; ++i)
+            fFilterValues.push_back(static_cast<Fixed>(0));
+#endif
+    }
+
+private:
+    struct FilterInstance {
+        // Offset within filterValues for this instance of the filter.
+        int fDataLocation;
+
+        // Distance from the left of the filter to the center. IN PIXELS
+        int fOffset;
+
+        // Number of values in this filter instance.
+        int fTrimmedLength;
+
+        // Filter length as specified. Note that this may be different from
+        // 'trimmed_length' if leading/trailing zeros of the original floating
+        // point form were clipped differently on each tail.
+        int fLength;
+    };
+
+    // Stores the information for each filter added to this class.
+    std::vector<FilterInstance> fFilters;
+
+    // We store all the filter values in this flat list, indexed by
+    // |FilterInstance.data_location| to avoid the mallocs required for storing
+    // each one separately.
+    std::vector<Fixed> fFilterValues;
+
+    // The maximum size of any filter we've added.
+    int fMaxFilter;
+};
+
+typedef void (*SkConvolveVertically_pointer)(
+    const SkConvolutionFilter1D::Fixed* filterValues,
+    int filterLength,
+    unsigned char* const* sourceDataRows,
+    int pixelWidth,
+    unsigned char* outRow,
+    bool hasAlpha);
+typedef void (*SkConvolve4RowsHorizontally_pointer)(
+    const unsigned char* srcData[4],
+    const SkConvolutionFilter1D& filter,
+    unsigned char* outRow[4]);
+typedef void (*SkConvolveHorizontally_pointer)(
+    const unsigned char* srcData,
+    const SkConvolutionFilter1D& filter,
+    unsigned char* outRow,
+    bool hasAlpha);
+
+struct SkConvolutionProcs {
+  // This is how many extra pixels may be read by the
+  // conolve*horizontally functions.
+    int fExtraHorizontalReads;
+    SkConvolveVertically_pointer fConvolveVertically;
+    SkConvolve4RowsHorizontally_pointer fConvolve4RowsHorizontally;
+    SkConvolveHorizontally_pointer fConvolveHorizontally;
+};
+
+
+
+// Does a two-dimensional convolution on the given source image.
+//
+// It is assumed the source pixel offsets referenced in the input filters
+// reference only valid pixels, so the source image size is not required. Each
+// row of the source image starts |sourceByteRowStride| after the previous
+// one (this allows you to have rows with some padding at the end).
+//
+// The result will be put into the given output buffer. The destination image
+// size will be xfilter.numValues() * yfilter.numValues() pixels. It will be
+// in rows of exactly xfilter.numValues() * 4 bytes.
+//
+// |sourceHasAlpha| is a hint that allows us to avoid doing computations on
+// the alpha channel if the image is opaque. If you don't know, set this to
+// true and it will work properly, but setting this to false will be a few
+// percent faster if you know the image is opaque.
+//
+// The layout in memory is assumed to be 4-bytes per pixel in B-G-R-A order
+// (this is ARGB when loaded into 32-bit words on a little-endian machine).
+SK_API void BGRAConvolve2D(const unsigned char* sourceData,
+    int sourceByteRowStride,
+    bool sourceHasAlpha,
+    const SkConvolutionFilter1D& xfilter,
+    const SkConvolutionFilter1D& yfilter,
+    int outputByteRowStride,
+    unsigned char* output,
+    SkConvolutionProcs *convolveProcs,
+    bool useSimdIfPossible);
+
+#endif  // SK_CONVOLVER_H