michael@0: /* NEON optimized code (C) COPYRIGHT 2009 Motorola
michael@0:  *
michael@0:  * Use of this source code is governed by a BSD-style license that can be
michael@0:  * found in the LICENSE file.
michael@0:  */
michael@0: 
michael@0: #include "SkBitmapProcState.h"
michael@0: #include "SkPerspIter.h"
michael@0: #include "SkShader.h"
michael@0: #include "SkUtils.h"
michael@0: #include "SkUtilsArm.h"
michael@0: #include "SkBitmapProcState_utils.h"
michael@0: 
michael@0: /*  returns 0...(n-1) given any x (positive or negative).
michael@0: 
michael@0:     As an example, if n (which is always positive) is 5...
michael@0: 
michael@0:           x: -8 -7 -6 -5 -4 -3 -2 -1  0  1  2  3  4  5  6  7  8
michael@0:     returns:  2  3  4  0  1  2  3  4  0  1  2  3  4  0  1  2  3
michael@0:  */
michael@0: static inline int sk_int_mod(int x, int n) {
michael@0:     SkASSERT(n > 0);
michael@0:     if ((unsigned)x >= (unsigned)n) {
michael@0:         if (x < 0) {
michael@0:             x = n + ~(~x % n);
michael@0:         } else {
michael@0:             x = x % n;
michael@0:         }
michael@0:     }
michael@0:     return x;
michael@0: }
michael@0: 
michael@0: void decal_nofilter_scale(uint32_t dst[], SkFixed fx, SkFixed dx, int count);
michael@0: void decal_filter_scale(uint32_t dst[], SkFixed fx, SkFixed dx, int count);
michael@0: 
michael@0: // Compile neon code paths if needed
michael@0: #if !SK_ARM_NEON_IS_NONE
michael@0: 
michael@0: // These are defined in src/opts/SkBitmapProcState_matrixProcs_neon.cpp
michael@0: extern const SkBitmapProcState::MatrixProc ClampX_ClampY_Procs_neon[];
michael@0: extern const SkBitmapProcState::MatrixProc RepeatX_RepeatY_Procs_neon[];
michael@0: 
michael@0: #endif // !SK_ARM_NEON_IS_NONE
michael@0: 
michael@0: // Compile non-neon code path if needed
michael@0: #if !SK_ARM_NEON_IS_ALWAYS
michael@0: #define MAKENAME(suffix)        ClampX_ClampY ## suffix
michael@0: #define TILEX_PROCF(fx, max)    SkClampMax((fx) >> 16, max)
michael@0: #define TILEY_PROCF(fy, max)    SkClampMax((fy) >> 16, max)
michael@0: #define TILEX_LOW_BITS(fx, max) (((fx) >> 12) & 0xF)
michael@0: #define TILEY_LOW_BITS(fy, max) (((fy) >> 12) & 0xF)
michael@0: #define CHECK_FOR_DECAL
michael@0: #include "SkBitmapProcState_matrix.h"
michael@0: 
michael@0: #define MAKENAME(suffix)        RepeatX_RepeatY ## suffix
michael@0: #define TILEX_PROCF(fx, max)    SK_USHIFT16(((fx) & 0xFFFF) * ((max) + 1))
michael@0: #define TILEY_PROCF(fy, max)    SK_USHIFT16(((fy) & 0xFFFF) * ((max) + 1))
michael@0: #define TILEX_LOW_BITS(fx, max) ((((fx) & 0xFFFF) * ((max) + 1) >> 12) & 0xF)
michael@0: #define TILEY_LOW_BITS(fy, max) ((((fy) & 0xFFFF) * ((max) + 1) >> 12) & 0xF)
michael@0: #include "SkBitmapProcState_matrix.h"
michael@0: #endif
michael@0: 
michael@0: #define MAKENAME(suffix)        GeneralXY ## suffix
michael@0: #define PREAMBLE(state)         SkBitmapProcState::FixedTileProc tileProcX = (state).fTileProcX; (void) tileProcX; \
michael@0:                                 SkBitmapProcState::FixedTileProc tileProcY = (state).fTileProcY; (void) tileProcY; \
michael@0:                                 SkBitmapProcState::FixedTileLowBitsProc tileLowBitsProcX = (state).fTileLowBitsProcX; (void) tileLowBitsProcX; \
michael@0:                                 SkBitmapProcState::FixedTileLowBitsProc tileLowBitsProcY = (state).fTileLowBitsProcY; (void) tileLowBitsProcY
michael@0: #define PREAMBLE_PARAM_X        , SkBitmapProcState::FixedTileProc tileProcX, SkBitmapProcState::FixedTileLowBitsProc tileLowBitsProcX
michael@0: #define PREAMBLE_PARAM_Y        , SkBitmapProcState::FixedTileProc tileProcY, SkBitmapProcState::FixedTileLowBitsProc tileLowBitsProcY
michael@0: #define PREAMBLE_ARG_X          , tileProcX, tileLowBitsProcX
michael@0: #define PREAMBLE_ARG_Y          , tileProcY, tileLowBitsProcY
michael@0: #define TILEX_PROCF(fx, max)    SK_USHIFT16(tileProcX(fx) * ((max) + 1))
michael@0: #define TILEY_PROCF(fy, max)    SK_USHIFT16(tileProcY(fy) * ((max) + 1))
michael@0: #define TILEX_LOW_BITS(fx, max) tileLowBitsProcX(fx, (max) + 1)
michael@0: #define TILEY_LOW_BITS(fy, max) tileLowBitsProcY(fy, (max) + 1)
michael@0: #include "SkBitmapProcState_matrix.h"
michael@0: 
michael@0: static inline U16CPU fixed_clamp(SkFixed x)
michael@0: {
michael@0:     if (x < 0) {
michael@0:         x = 0;
michael@0:     }
michael@0:     if (x >> 16) {
michael@0:         x = 0xFFFF;
michael@0:     }
michael@0:     return x;
michael@0: }
michael@0: 
michael@0: static inline U16CPU fixed_repeat(SkFixed x)
michael@0: {
michael@0:     return x & 0xFFFF;
michael@0: }
michael@0: 
michael@0: // Visual Studio 2010 (MSC_VER=1600) optimizes bit-shift code incorrectly.
michael@0: // See http://code.google.com/p/skia/issues/detail?id=472
michael@0: #if defined(_MSC_VER) && (_MSC_VER >= 1600)
michael@0: #pragma optimize("", off)
michael@0: #endif
michael@0: 
michael@0: static inline U16CPU fixed_mirror(SkFixed x)
michael@0: {
michael@0:     SkFixed s = x << 15 >> 31;
michael@0:     // s is FFFFFFFF if we're on an odd interval, or 0 if an even interval
michael@0:     return (x ^ s) & 0xFFFF;
michael@0: }
michael@0: 
michael@0: #if defined(_MSC_VER) && (_MSC_VER >= 1600)
michael@0: #pragma optimize("", on)
michael@0: #endif
michael@0: 
michael@0: static SkBitmapProcState::FixedTileProc choose_tile_proc(unsigned m)
michael@0: {
michael@0:     if (SkShader::kClamp_TileMode == m)
michael@0:         return fixed_clamp;
michael@0:     if (SkShader::kRepeat_TileMode == m)
michael@0:         return fixed_repeat;
michael@0:     SkASSERT(SkShader::kMirror_TileMode == m);
michael@0:     return fixed_mirror;
michael@0: }
michael@0: 
michael@0: static inline U16CPU fixed_clamp_lowbits(SkFixed x, int) {
michael@0:     return (x >> 12) & 0xF;
michael@0: }
michael@0: 
michael@0: static inline U16CPU fixed_repeat_or_mirrow_lowbits(SkFixed x, int scale) {
michael@0:     return ((x * scale) >> 12) & 0xF;
michael@0: }
michael@0: 
michael@0: static SkBitmapProcState::FixedTileLowBitsProc choose_tile_lowbits_proc(unsigned m) {
michael@0:     if (SkShader::kClamp_TileMode == m) {
michael@0:         return fixed_clamp_lowbits;
michael@0:     } else {
michael@0:         SkASSERT(SkShader::kMirror_TileMode == m ||
michael@0:                  SkShader::kRepeat_TileMode == m);
michael@0:         // mirror and repeat have the same behavior for the low bits.
michael@0:         return fixed_repeat_or_mirrow_lowbits;
michael@0:     }
michael@0: }
michael@0: 
michael@0: static inline U16CPU int_clamp(int x, int n) {
michael@0:     if (x >= n) {
michael@0:         x = n - 1;
michael@0:     }
michael@0:     if (x < 0) {
michael@0:         x = 0;
michael@0:     }
michael@0:     return x;
michael@0: }
michael@0: 
michael@0: static inline U16CPU int_repeat(int x, int n) {
michael@0:     return sk_int_mod(x, n);
michael@0: }
michael@0: 
michael@0: static inline U16CPU int_mirror(int x, int n) {
michael@0:     x = sk_int_mod(x, 2 * n);
michael@0:     if (x >= n) {
michael@0:         x = n + ~(x - n);
michael@0:     }
michael@0:     return x;
michael@0: }
michael@0: 
michael@0: #if 0
michael@0: static void test_int_tileprocs() {
michael@0:     for (int i = -8; i <= 8; i++) {
michael@0:         SkDebugf(" int_mirror(%2d, 3) = %d\n", i, int_mirror(i, 3));
michael@0:     }
michael@0: }
michael@0: #endif
michael@0: 
michael@0: static SkBitmapProcState::IntTileProc choose_int_tile_proc(unsigned tm) {
michael@0:     if (SkShader::kClamp_TileMode == tm)
michael@0:         return int_clamp;
michael@0:     if (SkShader::kRepeat_TileMode == tm)
michael@0:         return int_repeat;
michael@0:     SkASSERT(SkShader::kMirror_TileMode == tm);
michael@0:     return int_mirror;
michael@0: }
michael@0: 
michael@0: //////////////////////////////////////////////////////////////////////////////
michael@0: 
michael@0: void decal_nofilter_scale(uint32_t dst[], SkFixed fx, SkFixed dx, int count)
michael@0: {
michael@0:     int i;
michael@0: 
michael@0:     for (i = (count >> 2); i > 0; --i)
michael@0:     {
michael@0:         *dst++ = pack_two_shorts(fx >> 16, (fx + dx) >> 16);
michael@0:         fx += dx+dx;
michael@0:         *dst++ = pack_two_shorts(fx >> 16, (fx + dx) >> 16);
michael@0:         fx += dx+dx;
michael@0:     }
michael@0:     count &= 3;
michael@0: 
michael@0:     uint16_t* xx = (uint16_t*)dst;
michael@0:     for (i = count; i > 0; --i) {
michael@0:         *xx++ = SkToU16(fx >> 16); fx += dx;
michael@0:     }
michael@0: }
michael@0: 
michael@0: void decal_filter_scale(uint32_t dst[], SkFixed fx, SkFixed dx, int count)
michael@0: {
michael@0: 
michael@0: 
michael@0:     if (count & 1)
michael@0:     {
michael@0:         SkASSERT((fx >> (16 + 14)) == 0);
michael@0:         *dst++ = (fx >> 12 << 14) | ((fx >> 16) + 1);
michael@0:         fx += dx;
michael@0:     }
michael@0:     while ((count -= 2) >= 0)
michael@0:     {
michael@0:         SkASSERT((fx >> (16 + 14)) == 0);
michael@0:         *dst++ = (fx >> 12 << 14) | ((fx >> 16) + 1);
michael@0:         fx += dx;
michael@0: 
michael@0:         *dst++ = (fx >> 12 << 14) | ((fx >> 16) + 1);
michael@0:         fx += dx;
michael@0:     }
michael@0: }
michael@0: 
michael@0: ///////////////////////////////////////////////////////////////////////////////
michael@0: // stores the same as SCALE, but is cheaper to compute. Also since there is no
michael@0: // scale, we don't need/have a FILTER version
michael@0: 
michael@0: static void fill_sequential(uint16_t xptr[], int start, int count) {
michael@0: #if 1
michael@0:     if (reinterpret_cast<intptr_t>(xptr) & 0x2) {
michael@0:         *xptr++ = start++;
michael@0:         count -= 1;
michael@0:     }
michael@0:     if (count > 3) {
michael@0:         uint32_t* xxptr = reinterpret_cast<uint32_t*>(xptr);
michael@0:         uint32_t pattern0 = PACK_TWO_SHORTS(start + 0, start + 1);
michael@0:         uint32_t pattern1 = PACK_TWO_SHORTS(start + 2, start + 3);
michael@0:         start += count & ~3;
michael@0:         int qcount = count >> 2;
michael@0:         do {
michael@0:             *xxptr++ = pattern0;
michael@0:             pattern0 += 0x40004;
michael@0:             *xxptr++ = pattern1;
michael@0:             pattern1 += 0x40004;
michael@0:         } while (--qcount != 0);
michael@0:         xptr = reinterpret_cast<uint16_t*>(xxptr);
michael@0:         count &= 3;
michael@0:     }
michael@0:     while (--count >= 0) {
michael@0:         *xptr++ = start++;
michael@0:     }
michael@0: #else
michael@0:     for (int i = 0; i < count; i++) {
michael@0:         *xptr++ = start++;
michael@0:     }
michael@0: #endif
michael@0: }
michael@0: 
michael@0: static int nofilter_trans_preamble(const SkBitmapProcState& s, uint32_t** xy,
michael@0:                                    int x, int y) {
michael@0:     SkPoint pt;
michael@0:     s.fInvProc(s.fInvMatrix, SkIntToScalar(x) + SK_ScalarHalf,
michael@0:                SkIntToScalar(y) + SK_ScalarHalf, &pt);
michael@0:     **xy = s.fIntTileProcY(SkScalarToFixed(pt.fY) >> 16,
michael@0:                            s.fBitmap->height());
michael@0:     *xy += 1;   // bump the ptr
michael@0:     // return our starting X position
michael@0:     return SkScalarToFixed(pt.fX) >> 16;
michael@0: }
michael@0: 
michael@0: static void clampx_nofilter_trans(const SkBitmapProcState& s,
michael@0:                                   uint32_t xy[], int count, int x, int y) {
michael@0:     SkASSERT((s.fInvType & ~SkMatrix::kTranslate_Mask) == 0);
michael@0: 
michael@0:     int xpos = nofilter_trans_preamble(s, &xy, x, y);
michael@0:     const int width = s.fBitmap->width();
michael@0:     if (1 == width) {
michael@0:         // all of the following X values must be 0
michael@0:         memset(xy, 0, count * sizeof(uint16_t));
michael@0:         return;
michael@0:     }
michael@0: 
michael@0:     uint16_t* xptr = reinterpret_cast<uint16_t*>(xy);
michael@0:     int n;
michael@0: 
michael@0:     // fill before 0 as needed
michael@0:     if (xpos < 0) {
michael@0:         n = -xpos;
michael@0:         if (n > count) {
michael@0:             n = count;
michael@0:         }
michael@0:         memset(xptr, 0, n * sizeof(uint16_t));
michael@0:         count -= n;
michael@0:         if (0 == count) {
michael@0:             return;
michael@0:         }
michael@0:         xptr += n;
michael@0:         xpos = 0;
michael@0:     }
michael@0: 
michael@0:     // fill in 0..width-1 if needed
michael@0:     if (xpos < width) {
michael@0:         n = width - xpos;
michael@0:         if (n > count) {
michael@0:             n = count;
michael@0:         }
michael@0:         fill_sequential(xptr, xpos, n);
michael@0:         count -= n;
michael@0:         if (0 == count) {
michael@0:             return;
michael@0:         }
michael@0:         xptr += n;
michael@0:     }
michael@0: 
michael@0:     // fill the remaining with the max value
michael@0:     sk_memset16(xptr, width - 1, count);
michael@0: }
michael@0: 
michael@0: static void repeatx_nofilter_trans(const SkBitmapProcState& s,
michael@0:                                    uint32_t xy[], int count, int x, int y) {
michael@0:     SkASSERT((s.fInvType & ~SkMatrix::kTranslate_Mask) == 0);
michael@0: 
michael@0:     int xpos = nofilter_trans_preamble(s, &xy, x, y);
michael@0:     const int width = s.fBitmap->width();
michael@0:     if (1 == width) {
michael@0:         // all of the following X values must be 0
michael@0:         memset(xy, 0, count * sizeof(uint16_t));
michael@0:         return;
michael@0:     }
michael@0: 
michael@0:     uint16_t* xptr = reinterpret_cast<uint16_t*>(xy);
michael@0:     int start = sk_int_mod(xpos, width);
michael@0:     int n = width - start;
michael@0:     if (n > count) {
michael@0:         n = count;
michael@0:     }
michael@0:     fill_sequential(xptr, start, n);
michael@0:     xptr += n;
michael@0:     count -= n;
michael@0: 
michael@0:     while (count >= width) {
michael@0:         fill_sequential(xptr, 0, width);
michael@0:         xptr += width;
michael@0:         count -= width;
michael@0:     }
michael@0: 
michael@0:     if (count > 0) {
michael@0:         fill_sequential(xptr, 0, count);
michael@0:     }
michael@0: }
michael@0: 
michael@0: static void fill_backwards(uint16_t xptr[], int pos, int count) {
michael@0:     for (int i = 0; i < count; i++) {
michael@0:         SkASSERT(pos >= 0);
michael@0:         xptr[i] = pos--;
michael@0:     }
michael@0: }
michael@0: 
michael@0: static void mirrorx_nofilter_trans(const SkBitmapProcState& s,
michael@0:                                    uint32_t xy[], int count, int x, int y) {
michael@0:     SkASSERT((s.fInvType & ~SkMatrix::kTranslate_Mask) == 0);
michael@0: 
michael@0:     int xpos = nofilter_trans_preamble(s, &xy, x, y);
michael@0:     const int width = s.fBitmap->width();
michael@0:     if (1 == width) {
michael@0:         // all of the following X values must be 0
michael@0:         memset(xy, 0, count * sizeof(uint16_t));
michael@0:         return;
michael@0:     }
michael@0: 
michael@0:     uint16_t* xptr = reinterpret_cast<uint16_t*>(xy);
michael@0:     // need to know our start, and our initial phase (forward or backward)
michael@0:     bool forward;
michael@0:     int n;
michael@0:     int start = sk_int_mod(xpos, 2 * width);
michael@0:     if (start >= width) {
michael@0:         start = width + ~(start - width);
michael@0:         forward = false;
michael@0:         n = start + 1;  // [start .. 0]
michael@0:     } else {
michael@0:         forward = true;
michael@0:         n = width - start;  // [start .. width)
michael@0:     }
michael@0:     if (n > count) {
michael@0:         n = count;
michael@0:     }
michael@0:     if (forward) {
michael@0:         fill_sequential(xptr, start, n);
michael@0:     } else {
michael@0:         fill_backwards(xptr, start, n);
michael@0:     }
michael@0:     forward = !forward;
michael@0:     xptr += n;
michael@0:     count -= n;
michael@0: 
michael@0:     while (count >= width) {
michael@0:         if (forward) {
michael@0:             fill_sequential(xptr, 0, width);
michael@0:         } else {
michael@0:             fill_backwards(xptr, width - 1, width);
michael@0:         }
michael@0:         forward = !forward;
michael@0:         xptr += width;
michael@0:         count -= width;
michael@0:     }
michael@0: 
michael@0:     if (count > 0) {
michael@0:         if (forward) {
michael@0:             fill_sequential(xptr, 0, count);
michael@0:         } else {
michael@0:             fill_backwards(xptr, width - 1, count);
michael@0:         }
michael@0:     }
michael@0: }
michael@0: 
michael@0: ///////////////////////////////////////////////////////////////////////////////
michael@0: 
michael@0: SkBitmapProcState::MatrixProc
michael@0: SkBitmapProcState::chooseMatrixProc(bool trivial_matrix) {
michael@0: //    test_int_tileprocs();
michael@0:     // check for our special case when there is no scale/affine/perspective
michael@0:     if (trivial_matrix) {
michael@0:         SkASSERT(SkPaint::kNone_FilterLevel == fFilterLevel);
michael@0:         fIntTileProcY = choose_int_tile_proc(fTileModeY);
michael@0:         switch (fTileModeX) {
michael@0:             case SkShader::kClamp_TileMode:
michael@0:                 return clampx_nofilter_trans;
michael@0:             case SkShader::kRepeat_TileMode:
michael@0:                 return repeatx_nofilter_trans;
michael@0:             case SkShader::kMirror_TileMode:
michael@0:                 return mirrorx_nofilter_trans;
michael@0:         }
michael@0:     }
michael@0: 
michael@0:     int index = 0;
michael@0:     if (fFilterLevel != SkPaint::kNone_FilterLevel) {
michael@0:         index = 1;
michael@0:     }
michael@0:     if (fInvType & SkMatrix::kPerspective_Mask) {
michael@0:         index += 4;
michael@0:     } else if (fInvType & SkMatrix::kAffine_Mask) {
michael@0:         index += 2;
michael@0:     }
michael@0: 
michael@0:     if (SkShader::kClamp_TileMode == fTileModeX &&
michael@0:         SkShader::kClamp_TileMode == fTileModeY)
michael@0:     {
michael@0:         // clamp gets special version of filterOne
michael@0:         fFilterOneX = SK_Fixed1;
michael@0:         fFilterOneY = SK_Fixed1;
michael@0:         return SK_ARM_NEON_WRAP(ClampX_ClampY_Procs)[index];
michael@0:     }
michael@0: 
michael@0:     // all remaining procs use this form for filterOne
michael@0:     fFilterOneX = SK_Fixed1 / fBitmap->width();
michael@0:     fFilterOneY = SK_Fixed1 / fBitmap->height();
michael@0: 
michael@0:     if (SkShader::kRepeat_TileMode == fTileModeX &&
michael@0:         SkShader::kRepeat_TileMode == fTileModeY)
michael@0:     {
michael@0:         return SK_ARM_NEON_WRAP(RepeatX_RepeatY_Procs)[index];
michael@0:     }
michael@0: 
michael@0:     fTileProcX = choose_tile_proc(fTileModeX);
michael@0:     fTileProcY = choose_tile_proc(fTileModeY);
michael@0:     fTileLowBitsProcX = choose_tile_lowbits_proc(fTileModeX);
michael@0:     fTileLowBitsProcY = choose_tile_lowbits_proc(fTileModeY);
michael@0:     return GeneralXY_Procs[index];
michael@0: }