1/*
2 * Copyright 2009 The Android Open Source Project
3 *
4 * Use of this source code is governed by a BSD-style license that can be
5 * found in the LICENSE file.
6 */
7
8
9#include "SkBitmapProcState.h"
10#include "SkColorPriv.h"
11#include "SkPaint.h"
12#include "SkTypes.h"
13#include "SkUtils.h"
14#include "SkUtilsArm.h"
15
16#include "SkConvolver.h"
17
18#if SK_ARM_ARCH >= 6 && !defined(SK_CPU_BENDIAN)
19void SI8_D16_nofilter_DX_arm(
20    const SkBitmapProcState& s,
21    const uint32_t* SK_RESTRICT xy,
22    int count,
23    uint16_t* SK_RESTRICT colors) SK_ATTRIBUTE_OPTIMIZE_O1;
24
25void SI8_D16_nofilter_DX_arm(const SkBitmapProcState& s,
26                             const uint32_t* SK_RESTRICT xy,
27                             int count, uint16_t* SK_RESTRICT colors) {
28    SkASSERT(count > 0 && colors != NULL);
29    SkASSERT(s.fInvType <= (SkMatrix::kTranslate_Mask | SkMatrix::kScale_Mask));
30    SkASSERT(SkPaint::kNone_FilterLevel == s.fFilterLevel);
31
32    const uint16_t* SK_RESTRICT table = s.fBitmap->getColorTable()->lock16BitCache();
33    const uint8_t* SK_RESTRICT srcAddr = (const uint8_t*)s.fBitmap->getPixels();
34
35    // buffer is y32, x16, x16, x16, x16, x16
36    // bump srcAddr to the proper row, since we're told Y never changes
37    SkASSERT((unsigned)xy[0] < (unsigned)s.fBitmap->height());
38    srcAddr = (const uint8_t*)((const char*)srcAddr +
39                               xy[0] * s.fBitmap->rowBytes());
40
41    uint8_t src;
42
43    if (1 == s.fBitmap->width()) {
44        src = srcAddr[0];
45        uint16_t dstValue = table[src];
46        sk_memset16(colors, dstValue, count);
47    } else {
48        int i;
49        int count8 = count >> 3;
50        const uint16_t* SK_RESTRICT xx = (const uint16_t*)(xy + 1);
51
52        asm volatile (
53                      "cmp        %[count8], #0                   \n\t"   // compare loop counter with 0
54                      "beq        2f                              \n\t"   // if loop counter == 0, exit
55                      "1:                                             \n\t"
56                      "ldmia      %[xx]!, {r5, r7, r9, r11}       \n\t"   // load ptrs to pixels 0-7
57                      "subs       %[count8], %[count8], #1        \n\t"   // decrement loop counter
58                      "uxth       r4, r5                          \n\t"   // extract ptr 0
59                      "mov        r5, r5, lsr #16                 \n\t"   // extract ptr 1
60                      "uxth       r6, r7                          \n\t"   // extract ptr 2
61                      "mov        r7, r7, lsr #16                 \n\t"   // extract ptr 3
62                      "ldrb       r4, [%[srcAddr], r4]            \n\t"   // load pixel 0 from image
63                      "uxth       r8, r9                          \n\t"   // extract ptr 4
64                      "ldrb       r5, [%[srcAddr], r5]            \n\t"   // load pixel 1 from image
65                      "mov        r9, r9, lsr #16                 \n\t"   // extract ptr 5
66                      "ldrb       r6, [%[srcAddr], r6]            \n\t"   // load pixel 2 from image
67                      "uxth       r10, r11                        \n\t"   // extract ptr 6
68                      "ldrb       r7, [%[srcAddr], r7]            \n\t"   // load pixel 3 from image
69                      "mov        r11, r11, lsr #16               \n\t"   // extract ptr 7
70                      "ldrb       r8, [%[srcAddr], r8]            \n\t"   // load pixel 4 from image
71                      "add        r4, r4, r4                      \n\t"   // double pixel 0 for RGB565 lookup
72                      "ldrb       r9, [%[srcAddr], r9]            \n\t"   // load pixel 5 from image
73                      "add        r5, r5, r5                      \n\t"   // double pixel 1 for RGB565 lookup
74                      "ldrb       r10, [%[srcAddr], r10]          \n\t"   // load pixel 6 from image
75                      "add        r6, r6, r6                      \n\t"   // double pixel 2 for RGB565 lookup
76                      "ldrb       r11, [%[srcAddr], r11]          \n\t"   // load pixel 7 from image
77                      "add        r7, r7, r7                      \n\t"   // double pixel 3 for RGB565 lookup
78                      "ldrh       r4, [%[table], r4]              \n\t"   // load pixel 0 RGB565 from colmap
79                      "add        r8, r8, r8                      \n\t"   // double pixel 4 for RGB565 lookup
80                      "ldrh       r5, [%[table], r5]              \n\t"   // load pixel 1 RGB565 from colmap
81                      "add        r9, r9, r9                      \n\t"   // double pixel 5 for RGB565 lookup
82                      "ldrh       r6, [%[table], r6]              \n\t"   // load pixel 2 RGB565 from colmap
83                      "add        r10, r10, r10                   \n\t"   // double pixel 6 for RGB565 lookup
84                      "ldrh       r7, [%[table], r7]              \n\t"   // load pixel 3 RGB565 from colmap
85                      "add        r11, r11, r11                   \n\t"   // double pixel 7 for RGB565 lookup
86                      "ldrh       r8, [%[table], r8]              \n\t"   // load pixel 4 RGB565 from colmap
87                      "ldrh       r9, [%[table], r9]              \n\t"   // load pixel 5 RGB565 from colmap
88                      "ldrh       r10, [%[table], r10]            \n\t"   // load pixel 6 RGB565 from colmap
89                      "ldrh       r11, [%[table], r11]            \n\t"   // load pixel 7 RGB565 from colmap
90                      "pkhbt      r5, r4, r5, lsl #16             \n\t"   // pack pixels 0 and 1
91                      "pkhbt      r6, r6, r7, lsl #16             \n\t"   // pack pixels 2 and 3
92                      "pkhbt      r8, r8, r9, lsl #16             \n\t"   // pack pixels 4 and 5
93                      "pkhbt      r10, r10, r11, lsl #16          \n\t"   // pack pixels 6 and 7
94                      "stmia      %[colors]!, {r5, r6, r8, r10}   \n\t"   // store last 8 pixels
95                      "bgt        1b                              \n\t"   // loop if counter > 0
96                      "2:                                             \n\t"
97                      : [xx] "+r" (xx), [count8] "+r" (count8), [colors] "+r" (colors)
98                      : [table] "r" (table), [srcAddr] "r" (srcAddr)
99                      : "memory", "cc", "r4", "r5", "r6", "r7", "r8", "r9", "r10", "r11"
100                      );
101
102        for (i = (count & 7); i > 0; --i) {
103            src = srcAddr[*xx++]; *colors++ = table[src];
104        }
105    }
106
107    s.fBitmap->getColorTable()->unlock16BitCache();
108}
109
110void SI8_opaque_D32_nofilter_DX_arm(
111    const SkBitmapProcState& s,
112    const uint32_t* SK_RESTRICT xy,
113    int count,
114    SkPMColor* SK_RESTRICT colors) SK_ATTRIBUTE_OPTIMIZE_O1;
115
116void SI8_opaque_D32_nofilter_DX_arm(const SkBitmapProcState& s,
117                                    const uint32_t* SK_RESTRICT xy,
118                                    int count, SkPMColor* SK_RESTRICT colors) {
119    SkASSERT(count > 0 && colors != NULL);
120    SkASSERT(s.fInvType <= (SkMatrix::kTranslate_Mask | SkMatrix::kScale_Mask));
121    SkASSERT(SkPaint::kNone_FilterLevel == s.fFilterLevel);
122
123    const SkPMColor* SK_RESTRICT table = s.fBitmap->getColorTable()->lockColors();
124    const uint8_t* SK_RESTRICT srcAddr = (const uint8_t*)s.fBitmap->getPixels();
125
126    // buffer is y32, x16, x16, x16, x16, x16
127    // bump srcAddr to the proper row, since we're told Y never changes
128    SkASSERT((unsigned)xy[0] < (unsigned)s.fBitmap->height());
129    srcAddr = (const uint8_t*)((const char*)srcAddr + xy[0] * s.fBitmap->rowBytes());
130
131    if (1 == s.fBitmap->width()) {
132        uint8_t src = srcAddr[0];
133        SkPMColor dstValue = table[src];
134        sk_memset32(colors, dstValue, count);
135    } else {
136        const uint16_t* xx = (const uint16_t*)(xy + 1);
137
138        asm volatile (
139                      "subs       %[count], %[count], #8          \n\t"   // decrement count by 8, set flags
140                      "blt        2f                              \n\t"   // if count < 0, branch to singles
141                      "1:                                             \n\t"   // eights loop
142                      "ldmia      %[xx]!, {r5, r7, r9, r11}       \n\t"   // load ptrs to pixels 0-7
143                      "uxth       r4, r5                          \n\t"   // extract ptr 0
144                      "mov        r5, r5, lsr #16                 \n\t"   // extract ptr 1
145                      "uxth       r6, r7                          \n\t"   // extract ptr 2
146                      "mov        r7, r7, lsr #16                 \n\t"   // extract ptr 3
147                      "ldrb       r4, [%[srcAddr], r4]            \n\t"   // load pixel 0 from image
148                      "uxth       r8, r9                          \n\t"   // extract ptr 4
149                      "ldrb       r5, [%[srcAddr], r5]            \n\t"   // load pixel 1 from image
150                      "mov        r9, r9, lsr #16                 \n\t"   // extract ptr 5
151                      "ldrb       r6, [%[srcAddr], r6]            \n\t"   // load pixel 2 from image
152                      "uxth       r10, r11                        \n\t"   // extract ptr 6
153                      "ldrb       r7, [%[srcAddr], r7]            \n\t"   // load pixel 3 from image
154                      "mov        r11, r11, lsr #16               \n\t"   // extract ptr 7
155                      "ldrb       r8, [%[srcAddr], r8]            \n\t"   // load pixel 4 from image
156                      "ldrb       r9, [%[srcAddr], r9]            \n\t"   // load pixel 5 from image
157                      "ldrb       r10, [%[srcAddr], r10]          \n\t"   // load pixel 6 from image
158                      "ldrb       r11, [%[srcAddr], r11]          \n\t"   // load pixel 7 from image
159                      "ldr        r4, [%[table], r4, lsl #2]      \n\t"   // load pixel 0 SkPMColor from colmap
160                      "ldr        r5, [%[table], r5, lsl #2]      \n\t"   // load pixel 1 SkPMColor from colmap
161                      "ldr        r6, [%[table], r6, lsl #2]      \n\t"   // load pixel 2 SkPMColor from colmap
162                      "ldr        r7, [%[table], r7, lsl #2]      \n\t"   // load pixel 3 SkPMColor from colmap
163                      "ldr        r8, [%[table], r8, lsl #2]      \n\t"   // load pixel 4 SkPMColor from colmap
164                      "ldr        r9, [%[table], r9, lsl #2]      \n\t"   // load pixel 5 SkPMColor from colmap
165                      "ldr        r10, [%[table], r10, lsl #2]    \n\t"   // load pixel 6 SkPMColor from colmap
166                      "ldr        r11, [%[table], r11, lsl #2]    \n\t"   // load pixel 7 SkPMColor from colmap
167                      "subs       %[count], %[count], #8          \n\t"   // decrement loop counter
168                      "stmia      %[colors]!, {r4-r11}            \n\t"   // store 8 pixels
169                      "bge        1b                              \n\t"   // loop if counter >= 0
170                      "2:                                             \n\t"
171                      "adds       %[count], %[count], #8          \n\t"   // fix up counter, set flags
172                      "beq        4f                              \n\t"   // if count == 0, branch to exit
173                      "3:                                             \n\t"   // singles loop
174                      "ldrh       r4, [%[xx]], #2                 \n\t"   // load pixel ptr
175                      "subs       %[count], %[count], #1          \n\t"   // decrement loop counter
176                      "ldrb       r5, [%[srcAddr], r4]            \n\t"   // load pixel from image
177                      "ldr        r6, [%[table], r5, lsl #2]      \n\t"   // load SkPMColor from colmap
178                      "str        r6, [%[colors]], #4             \n\t"   // store pixel, update ptr
179                      "bne        3b                              \n\t"   // loop if counter != 0
180                      "4:                                             \n\t"   // exit
181                      : [xx] "+r" (xx), [count] "+r" (count), [colors] "+r" (colors)
182                      : [table] "r" (table), [srcAddr] "r" (srcAddr)
183                      : "memory", "cc", "r4", "r5", "r6", "r7", "r8", "r9", "r10", "r11"
184                      );
185    }
186
187    s.fBitmap->getColorTable()->unlockColors(false);
188}
189#endif // SK_ARM_ARCH >= 6 && !defined(SK_CPU_BENDIAN)
190
191///////////////////////////////////////////////////////////////////////////////
192
193/*  If we replace a sampleproc, then we null-out the associated shaderproc,
194    otherwise the shader won't even look at the matrix/sampler
195 */
196void SkBitmapProcState::platformProcs() {
197    bool isOpaque = 256 == fAlphaScale;
198    bool justDx = false;
199
200    if (fInvType <= (SkMatrix::kTranslate_Mask | SkMatrix::kScale_Mask)) {
201        justDx = true;
202    }
203
204    switch (fBitmap->config()) {
205        case SkBitmap::kIndex8_Config:
206#if SK_ARM_ARCH >= 6 && !defined(SK_CPU_BENDIAN)
207            if (justDx && SkPaint::kNone_FilterLevel == fFilterLevel) {
208#if 0   /* crashing on android device */
209                fSampleProc16 = SI8_D16_nofilter_DX_arm;
210                fShaderProc16 = NULL;
211#endif
212                if (isOpaque) {
213                    // this one is only very slighty faster than the C version
214                    fSampleProc32 = SI8_opaque_D32_nofilter_DX_arm;
215                    fShaderProc32 = NULL;
216                }
217            }
218#endif
219            break;
220        default:
221            break;
222    }
223}
224
225/////////////////////////////////////
226
227/* FUNCTIONS BELOW ARE SCALAR STUBS INTENDED FOR ARM DEVELOPERS TO REPLACE */
228
229/////////////////////////////////////
230
231
232static inline unsigned char ClampTo8(int a) {
233    if (static_cast<unsigned>(a) < 256) {
234        return a;  // Avoid the extra check in the common case.
235    }
236    if (a < 0) {
237        return 0;
238    }
239    return 255;
240}
241
242// Convolves horizontally along a single row. The row data is given in
243// |srcData| and continues for the numValues() of the filter.
244void convolveHorizontally_arm(const unsigned char* srcData,
245                              const SkConvolutionFilter1D& filter,
246                              unsigned char* outRow,
247                              bool hasAlpha) {
248    // Loop over each pixel on this row in the output image.
249    int numValues = filter.numValues();
250    for (int outX = 0; outX < numValues; outX++) {
251        // Get the filter that determines the current output pixel.
252        int filterOffset, filterLength;
253        const SkConvolutionFilter1D::ConvolutionFixed* filterValues =
254            filter.FilterForValue(outX, &filterOffset, &filterLength);
255
256        // Compute the first pixel in this row that the filter affects. It will
257        // touch |filterLength| pixels (4 bytes each) after this.
258        const unsigned char* rowToFilter = &srcData[filterOffset * 4];
259
260        // Apply the filter to the row to get the destination pixel in |accum|.
261        int accum[4] = {0};
262        for (int filterX = 0; filterX < filterLength; filterX++) {
263            SkConvolutionFilter1D::ConvolutionFixed curFilter = filterValues[filterX];
264            accum[0] += curFilter * rowToFilter[filterX * 4 + 0];
265            accum[1] += curFilter * rowToFilter[filterX * 4 + 1];
266            accum[2] += curFilter * rowToFilter[filterX * 4 + 2];
267            if (hasAlpha) {
268                accum[3] += curFilter * rowToFilter[filterX * 4 + 3];
269            }
270        }
271
272        // Bring this value back in range. All of the filter scaling factors
273        // are in fixed point with kShiftBits bits of fractional part.
274        accum[0] >>= SkConvolutionFilter1D::kShiftBits;
275        accum[1] >>= SkConvolutionFilter1D::kShiftBits;
276        accum[2] >>= SkConvolutionFilter1D::kShiftBits;
277        if (hasAlpha) {
278            accum[3] >>= SkConvolutionFilter1D::kShiftBits;
279        }
280
281        // Store the new pixel.
282        outRow[outX * 4 + 0] = ClampTo8(accum[0]);
283        outRow[outX * 4 + 1] = ClampTo8(accum[1]);
284        outRow[outX * 4 + 2] = ClampTo8(accum[2]);
285        if (hasAlpha) {
286            outRow[outX * 4 + 3] = ClampTo8(accum[3]);
287        }
288    }
289}
290
291// Does vertical convolution to produce one output row. The filter values and
292// length are given in the first two parameters. These are applied to each
293// of the rows pointed to in the |sourceDataRows| array, with each row
294// being |pixelWidth| wide.
295//
296// The output must have room for |pixelWidth * 4| bytes.
297template<bool hasAlpha>
298    void convolveVertically_arm(const SkConvolutionFilter1D::ConvolutionFixed* filterValues,
299                            int filterLength,
300                            unsigned char* const* sourceDataRows,
301                            int pixelWidth,
302                            unsigned char* outRow) {
303        // We go through each column in the output and do a vertical convolution,
304        // generating one output pixel each time.
305        for (int outX = 0; outX < pixelWidth; outX++) {
306            // Compute the number of bytes over in each row that the current column
307            // we're convolving starts at. The pixel will cover the next 4 bytes.
308            int byteOffset = outX * 4;
309
310            // Apply the filter to one column of pixels.
311            int accum[4] = {0};
312            for (int filterY = 0; filterY < filterLength; filterY++) {
313                SkConvolutionFilter1D::ConvolutionFixed curFilter = filterValues[filterY];
314                accum[0] += curFilter * sourceDataRows[filterY][byteOffset + 0];
315                accum[1] += curFilter * sourceDataRows[filterY][byteOffset + 1];
316                accum[2] += curFilter * sourceDataRows[filterY][byteOffset + 2];
317                if (hasAlpha) {
318                    accum[3] += curFilter * sourceDataRows[filterY][byteOffset + 3];
319                }
320            }
321
322            // Bring this value back in range. All of the filter scaling factors
323            // are in fixed point with kShiftBits bits of precision.
324            accum[0] >>= SkConvolutionFilter1D::kShiftBits;
325            accum[1] >>= SkConvolutionFilter1D::kShiftBits;
326            accum[2] >>= SkConvolutionFilter1D::kShiftBits;
327            if (hasAlpha) {
328                accum[3] >>= SkConvolutionFilter1D::kShiftBits;
329            }
330
331            // Store the new pixel.
332            outRow[byteOffset + 0] = ClampTo8(accum[0]);
333            outRow[byteOffset + 1] = ClampTo8(accum[1]);
334            outRow[byteOffset + 2] = ClampTo8(accum[2]);
335            if (hasAlpha) {
336                unsigned char alpha = ClampTo8(accum[3]);
337
338                // Make sure the alpha channel doesn't come out smaller than any of the
339                // color channels. We use premultipled alpha channels, so this should
340                // never happen, but rounding errors will cause this from time to time.
341                // These "impossible" colors will cause overflows (and hence random pixel
342                // values) when the resulting bitmap is drawn to the screen.
343                //
344                // We only need to do this when generating the final output row (here).
345                int maxColorChannel = SkTMax(outRow[byteOffset + 0],
346                                               SkTMax(outRow[byteOffset + 1],
347                                                      outRow[byteOffset + 2]));
348                if (alpha < maxColorChannel) {
349                    outRow[byteOffset + 3] = maxColorChannel;
350                } else {
351                    outRow[byteOffset + 3] = alpha;
352                }
353            } else {
354                // No alpha channel, the image is opaque.
355                outRow[byteOffset + 3] = 0xff;
356            }
357        }
358    }
359
360void convolveVertically_arm(const SkConvolutionFilter1D::ConvolutionFixed* filterValues,
361                            int filterLength,
362                            unsigned char* const* sourceDataRows,
363                            int pixelWidth,
364                            unsigned char* outRow,
365                            bool sourceHasAlpha) {
366    if (sourceHasAlpha) {
367        convolveVertically_arm<true>(filterValues, filterLength,
368                                     sourceDataRows, pixelWidth,
369                                     outRow);
370    } else {
371        convolveVertically_arm<false>(filterValues, filterLength,
372                                      sourceDataRows, pixelWidth,
373                                      outRow);
374    }
375}
376
377// Convolves horizontally along four rows. The row data is given in
378// |src_data| and continues for the num_values() of the filter.
379// The algorithm is almost same as |ConvolveHorizontally_SSE2|. Please
380// refer to that function for detailed comments.
381void convolve4RowsHorizontally_arm(const unsigned char* src_data[4],
382                                   const SkConvolutionFilter1D& filter,
383                                   unsigned char* out_row[4]) {
384}
385
386///////////////////////////
387
388/* STOP REWRITING FUNCTIONS HERE, BUT DON'T FORGET TO EDIT THE
389   PLATFORM CONVOLUTION PROCS BELOW */
390
391///////////////////////////
392
393void applySIMDPadding_arm(SkConvolutionFilter1D *filter) {
394    // Padding |paddingCount| of more dummy coefficients after the coefficients
395    // of last filter to prevent SIMD instructions which load 8 or 16 bytes
396    // together to access invalid memory areas. We are not trying to align the
397    // coefficients right now due to the opaqueness of <vector> implementation.
398    // This has to be done after all |AddFilter| calls.
399    for (int i = 0; i < 8; ++i) {
400        filter->addFilterValue(static_cast<SkConvolutionFilter1D::ConvolutionFixed>(0));
401    }
402}
403
404void SkBitmapProcState::platformConvolutionProcs() {
405    if (sk_cpu_arm_has_neon()) {
406        fConvolutionProcs->fExtraHorizontalReads = 3;
407        fConvolutionProcs->fConvolveVertically = &convolveVertically_arm;
408
409        // next line is commented out because the four-row convolution function above is
410        // just a no-op.  Please see the comment above its definition, and the SSE implementation
411        // in SkBitmapProcState_opts_SSE2.cpp for guidance on its semantics.
412        // leaving it as NULL will just cause the convolution system to not attempt
413        // to operate on four rows at once, which is correct but not performance-optimal.
414
415        // fConvolutionProcs->fConvolve4RowsHorizontally = &convolve4RowsHorizontally_arm;
416
417        fConvolutionProcs->fConvolve4RowsHorizontally = NULL;
418
419        fConvolutionProcs->fConvolveHorizontally = &convolveHorizontally_arm;
420        fConvolutionProcs->fApplySIMDPadding = &applySIMDPadding_arm;
421    }
422}
423