More general version of image filtering; reworked to be robust and easier to SSE

src/opts/SkBitmapFilter_opts_SSE2.cpp

Index: src/opts/SkBitmapFilter_opts_SSE2.cpp
diff --git a/src/opts/SkBitmapFilter_opts_SSE2.cpp b/src/opts/SkBitmapFilter_opts_SSE2.cpp
new file mode 100644
index 0000000000000000000000000000000000000000..eaba46b2348b9b15980cac38b44849ce679d1f2d
--- /dev/null
+++ b/src/opts/SkBitmapFilter_opts_SSE2.cpp
@@ -0,0 +1,236 @@
+/*
+ * Copyright 2013 Google Inc.
+ *
+ * Use of this source code is governed by a BSD-style license that can be
+ * found in the LICENSE file.
+ */
+
+#include "SkBitmapProcState.h"
+#include "SkBitmap.h"
+#include "SkColor.h"
+#include "SkColorPriv.h"
+#include "SkUnPreMultiply.h"
+#include "SkShader.h"
+
+#include <emmintrin.h>
+
+#define DS(x) SkDoubleToScalar(x)
+
+#define MUL(a, b)   ((a) * (b))
+
+static inline void print128i(__m128i value) {
+    int *v = (int*) &value;
+    printf("% .11d % .11d % .11d % .11d\n", v[0], v[1], v[2], v[3]);
+}
+
+static inline void print128i_16(__m128i value) {
+    short *v = (short*) &value;
+    printf("% .5d % .5d % .5d % .5d % .5d % .5d % .5d % .5d\n", v[0], v[1], v[2], v[3], v[4], v[5], v[6], v[7]);
+}
+
+static inline void print128i_8(__m128i value) {
+    unsigned char *v = (unsigned char*) &value;
+    printf("%.3u %.3u %.3u %.3u %.3u %.3u %.3u %.3u %.3u %.3u %.3u %.3u %.3u %.3u %.3u %.3u\n", 
+           v[0], v[1], v[2], v[3], v[4], v[5], v[6], v[7],
+           v[8], v[9], v[10], v[11], v[12], v[13], v[14], v[15]
+           );
+}
+
+static inline void print128f(__m128 value) {
+    float *f = (float*) &value;
+    printf("%3.4f %3.4f %3.4f %3.4f\n", f[0], f[1], f[2], f[3]);
+}
+
+#define BLAH( s ) if (debug) { s; }
+
+// some SSE blending code adapted from theowl84's bilinear blog:
+// http://fastcpp.blogspot.com/2011/06/bilinear-pixel-interpolation-using-sse.html
+
+#define WEIGHT_BITS 6
+
+static inline SkPMColor cubicBlendSSE(__m128 coeffs, SkPMColor *p, bool debug=false) {
+    
+    BLAH( printf( "*(p+0) = (%.3u %.3u %.3u %.3u)\n", SkGetPackedR32(*(p+0)), SkGetPackedG32(*(p+0)), SkGetPackedB32(*(p+0)), SkGetPackedA32(*(p+0)) ) )
+    BLAH( printf( "*(p+1) = (%.3u %.3u %.3u %.3u)\n", SkGetPackedR32(*(p+1)), SkGetPackedG32(*(p+1)), SkGetPackedB32(*(p+1)), SkGetPackedA32(*(p+1)) ) )
+    BLAH( printf( "*(p+2) = (%.3u %.3u %.3u %.3u)\n", SkGetPackedR32(*(p+2)), SkGetPackedG32(*(p+2)), SkGetPackedB32(*(p+2)), SkGetPackedA32(*(p+2)) ) )
+    BLAH( printf( "*(p+3) = (%.3u %.3u %.3u %.3u)\n", SkGetPackedR32(*(p+3)), SkGetPackedG32(*(p+3)), SkGetPackedB32(*(p+3)), SkGetPackedA32(*(p+3)) ) )
+    
+    __m128 p1 = _mm_cvtpu8_ps( _mm_set1_pi32( *(p+0) ) );
+    __m128 p2 = _mm_cvtpu8_ps( _mm_set1_pi32( *(p+1) ) );
+    __m128 p3 = _mm_cvtpu8_ps( _mm_set1_pi32( *(p+2) ) );
+    __m128 p4 = _mm_cvtpu8_ps( _mm_set1_pi32( *(p+3) ) );
+    
+    BLAH( printf( "p1 = " ) );
+    BLAH( print128f( p1 ));
+    BLAH( printf( "p2 = " ) );
+    BLAH( print128f( p1 ));
+    BLAH( printf( "p3 = " ) );
+    BLAH( print128f( p1 ));
+    BLAH( printf( "p4 = " ) );
+    BLAH( print128f( p1 ));
+    
+        
+    
+    return (SkPMColor) 0;
+
+}
+
+static __m128 build_coeff_sse(float t, bool debug=false) {
+    static __m128 coeffs[4];
+    static const __m128 CONST_WEIGHT_FIXED = _mm_set1_ps(1 << WEIGHT_BITS);
+    static bool once = true;
+    if (once) {
+        once = false;
+        // note these coefficients are stored as the transpose of the ones in the
+        // scalar CPU code so we can do all four dot products at once (SOA organization)
+        coeffs[0] = _mm_set_ps(   0.0f / 18.0f ,  1.0f / 18.0f,  16.0f / 18.0f , 1.0f / 18.0f);
+        coeffs[1] = _mm_set_ps(   0.0f / 18.0f ,  9.0f / 18.0f,   0.0f / 18.0f ,-9.0f / 18.0f);
+        coeffs[2] = _mm_set_ps(  -6.0f / 18.0f , 27.0f / 18.0f, -36.0f / 18.0f ,15.0f / 18.0f);
+        coeffs[3] = _mm_set_ps(   7.0f / 18.0f ,-21.0f / 18.0f,  21.0f / 18.0f ,-7.0f / 18.0f);
+    };
+    
+    // evaluate the cubic polynomial
+    
+    __m128 tvec = _mm_set1_ps( t );
+    __m128 t2vec = _mm_mul_ps( tvec, tvec );
+    __m128 t3vec = _mm_mul_ps( tvec, t2vec );
+    
+    __m128 weight = _mm_mul_ps( t3vec, coeffs[3] );
+    weight = _mm_add_ps( weight, _mm_mul_ps( t2vec, coeffs[2]));
+    weight = _mm_add_ps( weight, _mm_mul_ps( tvec, coeffs[1]));
+    weight = _mm_add_ps( weight, coeffs[0] );
+    
+    BLAH( printf( "t = %f\n", t ) )
+    BLAH( printf( "weight = " ) )
+    BLAH( print128f( weight ) )
+    
+    return weight;
+}
+
+// because the border is handled specially, this is guaranteed to have all 16 pixels
+// available to it without running off the bitmap's edge.
+
+// the coefficients are split across two registers as prepared by the build_coeff function above.
+static SkPMColor doBicubicFilterSSE(const SkBitmap *bm, 
+                                    __m128 coeffX, __m128 coeffY,
+                                    int sx, int sy, 
+                                    bool debug=false )
+{
+    SkPMColor x_blends[4];
+    
+    const int maxX = bm->width() - 1;
+    const int maxY = bm->height() - 1;
+    
+    SkPMColor temp[4][4]; // used for special-casing the border
+    SkPMColor *rows[4];
+    
+    BLAH( printf( "in doBicubicFilterSSE with sx=%d, sy=%d\n", sx, sy))
+    
+    if (sx <= 0 || sy <= 0 || sx >= maxX-2 || sy >= maxY-2) {
+        // handle border specially
+        int x0 = SkClampMax(sx - 1, maxX);
+        int x1 = SkClampMax(sx    , maxX);
+        int x2 = SkClampMax(sx + 1, maxX);
+        int x3 = SkClampMax(sx + 2, maxX);
+        int y0 = SkClampMax(sy - 1, maxY);
+        int y1 = SkClampMax(sy    , maxY);
+        int y2 = SkClampMax(sy + 1, maxY);
+        int y3 = SkClampMax(sy + 2, maxY);
+        temp[0][0] = *bm->getAddr32(x0, y0);
+        temp[0][1] = *bm->getAddr32(x1, y0);
+        temp[0][2] = *bm->getAddr32(x2, y0);
+        temp[0][3] = *bm->getAddr32(x3, y0);
+        temp[1][0] = *bm->getAddr32(x0, y1);
+        temp[1][1] = *bm->getAddr32(x1, y1);
+        temp[1][2] = *bm->getAddr32(x2, y1);
+        temp[1][3] = *bm->getAddr32(x3, y1);
+        temp[2][0] = *bm->getAddr32(x0, y2);
+        temp[2][1] = *bm->getAddr32(x1, y2);
+        temp[2][2] = *bm->getAddr32(x2, y2);
+        temp[2][3] = *bm->getAddr32(x3, y2);
+        temp[3][0] = *bm->getAddr32(x0, y3);
+        temp[3][1] = *bm->getAddr32(x1, y3);
+        temp[3][2] = *bm->getAddr32(x2, y3);
+        temp[3][3] = *bm->getAddr32(x3, y3);
+        
+        rows[0] = temp[0];
+        rows[1] = temp[1];
+        rows[2] = temp[2];
+        rows[3] = temp[3];
+    } else {
+        rows[0] = bm->getAddr32(sx-1,sy-1);
+        rows[1] = bm->getAddr32(sx-1,sy-0);
+        rows[2] = bm->getAddr32(sx-1,sy+1);
+        rows[3] = bm->getAddr32(sx-1,sy+2);
+    }
+
+    x_blends[0] = cubicBlendSSE(coeffX, rows[0], debug);
+    x_blends[1] = cubicBlendSSE(coeffX, rows[1], debug);
+    x_blends[2] = cubicBlendSSE(coeffX, rows[2], debug);    
+    x_blends[3] = cubicBlendSSE(coeffX, rows[3], debug);
+
+    return cubicBlendSSE(coeffY, x_blends, debug);
+}
+
+
+int debug_x = 20;
+int debug_y = 255;
+
+void highQualityFilter_SSE2(const SkBitmapProcState& s, int x, int y,
+                        SkPMColor* SK_RESTRICT colors, int count) {
+
+    SkPMColor *orig_colors = colors;
+
+    while (count-- > 0) {
+        SkPoint srcPt;
+        s.fInvProc(*s.fInvMatrix, SkIntToScalar(x),
+                    SkIntToScalar(y), &srcPt);
+                    
+        srcPt.fX -= SK_ScalarHalf;
+        srcPt.fY -= SK_ScalarHalf;
+        SkScalar fractx = srcPt.fX - SkScalarFloorToScalar(srcPt.fX);
+        SkScalar fracty = srcPt.fY - SkScalarFloorToScalar(srcPt.fY);
+
+        __m128 coeffX, coeffY;
+        int sx = SkScalarFloorToInt(srcPt.fX);
+        int sy = SkScalarFloorToInt(srcPt.fY);
+        
+        coeffX = build_coeff_sse(fractx);
+        coeffY = build_coeff_sse(fracty);
+
+        SkPMColor color = doBicubicFilterSSE( s.fBitmap, coeffX, coeffY, sx, sy, 1 );
+        SkPMColorAssert(color);
+        *colors++ = color;
+        x++;
+    }
+
+}
+
+void highQualityFilter_ScaleOnly_SSE2(const SkBitmapProcState &s, int x, int y,
+                             SkPMColor *SK_RESTRICT colors, int count) {
+    const int maxX = s.fBitmap->width() - 1;
+    const int maxY = s.fBitmap->height() - 1;
+
+    SkPoint srcPt;
+    s.fInvProc(*s.fInvMatrix, SkIntToScalar(x), SkIntToScalar(y), &srcPt);
+    srcPt.fY -= SK_ScalarHalf;
+    SkScalar fracty = srcPt.fY - SkScalarFloorToScalar(srcPt.fY);
+    __m128 coeffX, coeffY;
+    coeffY = build_coeff_sse(fracty);
+    int sy = SkScalarFloorToInt(srcPt.fY);
+
+    while (count-- > 0) {
+        s.fInvProc(*s.fInvMatrix, SkIntToScalar(x), SkIntToScalar(y), &srcPt);
+        srcPt.fX -= SK_ScalarHalf;
+        SkScalar fractx = srcPt.fX - SkScalarFloorToScalar(srcPt.fX);
+
+        int sx = SkScalarFloorToInt(srcPt.fX);
+
+        coeffX = build_coeff_sse(fractx);
+        SkPMColor c = doBicubicFilterSSE( s.fBitmap, coeffX, coeffY, sx, sy );
+        SkPMColorAssert(c);
+        *colors++ = c;
+
+        x++;
+    }
+}