michael@0: // Copyright (c) 2006-2011 The Chromium Authors. All rights reserved.
michael@0: //
michael@0: // Redistribution and use in source and binary forms, with or without
michael@0: // modification, are permitted provided that the following conditions
michael@0: // are met:
michael@0: //  * Redistributions of source code must retain the above copyright
michael@0: //    notice, this list of conditions and the following disclaimer.
michael@0: //  * Redistributions in binary form must reproduce the above copyright
michael@0: //    notice, this list of conditions and the following disclaimer in
michael@0: //    the documentation and/or other materials provided with the
michael@0: //    distribution.
michael@0: //  * Neither the name of Google, Inc. nor the names of its contributors
michael@0: //    may be used to endorse or promote products derived from this
michael@0: //    software without specific prior written permission.
michael@0: //
michael@0: // THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS
michael@0: // "AS IS" AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT
michael@0: // LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS
michael@0: // FOR A PARTICULAR PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE
michael@0: // COPYRIGHT OWNER OR CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT,
michael@0: // INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING,
michael@0: // BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS
michael@0: // OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED
michael@0: // AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY,
michael@0: // OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT
michael@0: // OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF
michael@0: // SUCH DAMAGE.
michael@0: 
michael@0: #include "convolver.h"
michael@0: 
michael@0: #include <algorithm>
michael@0: 
michael@0: #include "skia/SkTypes.h"
michael@0: 
michael@0: // note: SIMD_SSE2 is not enabled because of bugs, apparently
michael@0: 
michael@0: #if defined(SIMD_SSE2)
michael@0: #include <emmintrin.h>  // ARCH_CPU_X86_FAMILY was defined in build/config.h
michael@0: #endif
michael@0: 
michael@0: #if defined(SK_CPU_LENDIAN)
michael@0: #define R_OFFSET_IDX 0
michael@0: #define G_OFFSET_IDX 1
michael@0: #define B_OFFSET_IDX 2
michael@0: #define A_OFFSET_IDX 3
michael@0: #else
michael@0: #define R_OFFSET_IDX 3
michael@0: #define G_OFFSET_IDX 2
michael@0: #define B_OFFSET_IDX 1
michael@0: #define A_OFFSET_IDX 0
michael@0: #endif
michael@0: 
michael@0: namespace skia {
michael@0: 
michael@0: namespace {
michael@0: 
michael@0: // Converts the argument to an 8-bit unsigned value by clamping to the range
michael@0: // 0-255.
michael@0: inline unsigned char ClampTo8(int a) {
michael@0:   if (static_cast<unsigned>(a) < 256)
michael@0:     return a;  // Avoid the extra check in the common case.
michael@0:   if (a < 0)
michael@0:     return 0;
michael@0:   return 255;
michael@0: }
michael@0: 
michael@0: // Stores a list of rows in a circular buffer. The usage is you write into it
michael@0: // by calling AdvanceRow. It will keep track of which row in the buffer it
michael@0: // should use next, and the total number of rows added.
michael@0: class CircularRowBuffer {
michael@0:  public:
michael@0:   // The number of pixels in each row is given in |source_row_pixel_width|.
michael@0:   // The maximum number of rows needed in the buffer is |max_y_filter_size|
michael@0:   // (we only need to store enough rows for the biggest filter).
michael@0:   //
michael@0:   // We use the |first_input_row| to compute the coordinates of all of the
michael@0:   // following rows returned by Advance().
michael@0:   CircularRowBuffer(int dest_row_pixel_width, int max_y_filter_size,
michael@0:                     int first_input_row)
michael@0:       : row_byte_width_(dest_row_pixel_width * 4),
michael@0:         num_rows_(max_y_filter_size),
michael@0:         next_row_(0),
michael@0:         next_row_coordinate_(first_input_row) {
michael@0:     buffer_.resize(row_byte_width_ * max_y_filter_size);
michael@0:     row_addresses_.resize(num_rows_);
michael@0:   }
michael@0: 
michael@0:   // Moves to the next row in the buffer, returning a pointer to the beginning
michael@0:   // of it.
michael@0:   unsigned char* AdvanceRow() {
michael@0:     unsigned char* row = &buffer_[next_row_ * row_byte_width_];
michael@0:     next_row_coordinate_++;
michael@0: 
michael@0:     // Set the pointer to the next row to use, wrapping around if necessary.
michael@0:     next_row_++;
michael@0:     if (next_row_ == num_rows_)
michael@0:       next_row_ = 0;
michael@0:     return row;
michael@0:   }
michael@0: 
michael@0:   // Returns a pointer to an "unrolled" array of rows. These rows will start
michael@0:   // at the y coordinate placed into |*first_row_index| and will continue in
michael@0:   // order for the maximum number of rows in this circular buffer.
michael@0:   //
michael@0:   // The |first_row_index_| may be negative. This means the circular buffer
michael@0:   // starts before the top of the image (it hasn't been filled yet).
michael@0:   unsigned char* const* GetRowAddresses(int* first_row_index) {
michael@0:     // Example for a 4-element circular buffer holding coords 6-9.
michael@0:     //   Row 0   Coord 8
michael@0:     //   Row 1   Coord 9
michael@0:     //   Row 2   Coord 6  <- next_row_ = 2, next_row_coordinate_ = 10.
michael@0:     //   Row 3   Coord 7
michael@0:     //
michael@0:     // The "next" row is also the first (lowest) coordinate. This computation
michael@0:     // may yield a negative value, but that's OK, the math will work out
michael@0:     // since the user of this buffer will compute the offset relative
michael@0:     // to the first_row_index and the negative rows will never be used.
michael@0:     *first_row_index = next_row_coordinate_ - num_rows_;
michael@0: 
michael@0:     int cur_row = next_row_;
michael@0:     for (int i = 0; i < num_rows_; i++) {
michael@0:       row_addresses_[i] = &buffer_[cur_row * row_byte_width_];
michael@0: 
michael@0:       // Advance to the next row, wrapping if necessary.
michael@0:       cur_row++;
michael@0:       if (cur_row == num_rows_)
michael@0:         cur_row = 0;
michael@0:     }
michael@0:     return &row_addresses_[0];
michael@0:   }
michael@0: 
michael@0:  private:
michael@0:   // The buffer storing the rows. They are packed, each one row_byte_width_.
michael@0:   std::vector<unsigned char> buffer_;
michael@0: 
michael@0:   // Number of bytes per row in the |buffer_|.
michael@0:   int row_byte_width_;
michael@0: 
michael@0:   // The number of rows available in the buffer.
michael@0:   int num_rows_;
michael@0: 
michael@0:   // The next row index we should write into. This wraps around as the
michael@0:   // circular buffer is used.
michael@0:   int next_row_;
michael@0: 
michael@0:   // The y coordinate of the |next_row_|. This is incremented each time a
michael@0:   // new row is appended and does not wrap.
michael@0:   int next_row_coordinate_;
michael@0: 
michael@0:   // Buffer used by GetRowAddresses().
michael@0:   std::vector<unsigned char*> row_addresses_;
michael@0: };
michael@0: 
michael@0: // Convolves horizontally along a single row. The row data is given in
michael@0: // |src_data| and continues for the num_values() of the filter.
michael@0: template<bool has_alpha>
michael@0: // This function is miscompiled with gcc 4.5 with pgo. See bug 827946.
michael@0: #if defined(__GNUC__) && defined(MOZ_GCC_VERSION_AT_LEAST)
michael@0: #if MOZ_GCC_VERSION_AT_LEAST(4, 5, 0) && !MOZ_GCC_VERSION_AT_LEAST(4, 6, 0)
michael@0: __attribute__((optimize("-O1")))
michael@0: #endif
michael@0: #endif
michael@0: void ConvolveHorizontally(const unsigned char* src_data,
michael@0:                           const ConvolutionFilter1D& filter,
michael@0:                           unsigned char* out_row) {
michael@0:   // Loop over each pixel on this row in the output image.
michael@0:   int num_values = filter.num_values();
michael@0:   for (int out_x = 0; out_x < num_values; out_x++) {
michael@0:     // Get the filter that determines the current output pixel.
michael@0:     int filter_offset, filter_length;
michael@0:     const ConvolutionFilter1D::Fixed* filter_values =
michael@0:         filter.FilterForValue(out_x, &filter_offset, &filter_length);
michael@0: 
michael@0:     // Compute the first pixel in this row that the filter affects. It will
michael@0:     // touch |filter_length| pixels (4 bytes each) after this.
michael@0:     const unsigned char* row_to_filter = &src_data[filter_offset * 4];
michael@0: 
michael@0:     // Apply the filter to the row to get the destination pixel in |accum|.
michael@0:     int accum[4] = {0};
michael@0:     for (int filter_x = 0; filter_x < filter_length; filter_x++) {
michael@0:       ConvolutionFilter1D::Fixed cur_filter = filter_values[filter_x];
michael@0:       accum[0] += cur_filter * row_to_filter[filter_x * 4 + R_OFFSET_IDX];
michael@0:       accum[1] += cur_filter * row_to_filter[filter_x * 4 + G_OFFSET_IDX];
michael@0:       accum[2] += cur_filter * row_to_filter[filter_x * 4 + B_OFFSET_IDX];
michael@0:       if (has_alpha)
michael@0:         accum[3] += cur_filter * row_to_filter[filter_x * 4 + A_OFFSET_IDX];
michael@0:     }
michael@0: 
michael@0:     // Bring this value back in range. All of the filter scaling factors
michael@0:     // are in fixed point with kShiftBits bits of fractional part.
michael@0:     accum[0] >>= ConvolutionFilter1D::kShiftBits;
michael@0:     accum[1] >>= ConvolutionFilter1D::kShiftBits;
michael@0:     accum[2] >>= ConvolutionFilter1D::kShiftBits;
michael@0:     if (has_alpha)
michael@0:       accum[3] >>= ConvolutionFilter1D::kShiftBits;
michael@0: 
michael@0:     // Store the new pixel.
michael@0:     out_row[out_x * 4 + R_OFFSET_IDX] = ClampTo8(accum[0]);
michael@0:     out_row[out_x * 4 + G_OFFSET_IDX] = ClampTo8(accum[1]);
michael@0:     out_row[out_x * 4 + B_OFFSET_IDX] = ClampTo8(accum[2]);
michael@0:     if (has_alpha)
michael@0:       out_row[out_x * 4 + A_OFFSET_IDX] = ClampTo8(accum[3]);
michael@0:   }
michael@0: }
michael@0: 
michael@0: // Does vertical convolution to produce one output row. The filter values and
michael@0: // length are given in the first two parameters. These are applied to each
michael@0: // of the rows pointed to in the |source_data_rows| array, with each row
michael@0: // being |pixel_width| wide.
michael@0: //
michael@0: // The output must have room for |pixel_width * 4| bytes.
michael@0: template<bool has_alpha>
michael@0: void ConvolveVertically(const ConvolutionFilter1D::Fixed* filter_values,
michael@0:                         int filter_length,
michael@0:                         unsigned char* const* source_data_rows,
michael@0:                         int pixel_width,
michael@0:                         unsigned char* out_row) {
michael@0:   // We go through each column in the output and do a vertical convolution,
michael@0:   // generating one output pixel each time.
michael@0:   for (int out_x = 0; out_x < pixel_width; out_x++) {
michael@0:     // Compute the number of bytes over in each row that the current column
michael@0:     // we're convolving starts at. The pixel will cover the next 4 bytes.
michael@0:     int byte_offset = out_x * 4;
michael@0: 
michael@0:     // Apply the filter to one column of pixels.
michael@0:     int accum[4] = {0};
michael@0:     for (int filter_y = 0; filter_y < filter_length; filter_y++) {
michael@0:       ConvolutionFilter1D::Fixed cur_filter = filter_values[filter_y];
michael@0:       accum[0] += cur_filter 
michael@0: 	* source_data_rows[filter_y][byte_offset + R_OFFSET_IDX];
michael@0:       accum[1] += cur_filter 
michael@0: 	* source_data_rows[filter_y][byte_offset + G_OFFSET_IDX];
michael@0:       accum[2] += cur_filter 
michael@0: 	* source_data_rows[filter_y][byte_offset + B_OFFSET_IDX];
michael@0:       if (has_alpha)
michael@0:         accum[3] += cur_filter 
michael@0: 	  * source_data_rows[filter_y][byte_offset + A_OFFSET_IDX];
michael@0:     }
michael@0: 
michael@0:     // Bring this value back in range. All of the filter scaling factors
michael@0:     // are in fixed point with kShiftBits bits of precision.
michael@0:     accum[0] >>= ConvolutionFilter1D::kShiftBits;
michael@0:     accum[1] >>= ConvolutionFilter1D::kShiftBits;
michael@0:     accum[2] >>= ConvolutionFilter1D::kShiftBits;
michael@0:     if (has_alpha)
michael@0:       accum[3] >>= ConvolutionFilter1D::kShiftBits;
michael@0: 
michael@0:     // Store the new pixel.
michael@0:     out_row[byte_offset + R_OFFSET_IDX] = ClampTo8(accum[0]);
michael@0:     out_row[byte_offset + G_OFFSET_IDX] = ClampTo8(accum[1]);
michael@0:     out_row[byte_offset + B_OFFSET_IDX] = ClampTo8(accum[2]);
michael@0:     if (has_alpha) {
michael@0:       unsigned char alpha = ClampTo8(accum[3]);
michael@0: 
michael@0:       // Make sure the alpha channel doesn't come out smaller than any of the
michael@0:       // color channels. We use premultipled alpha channels, so this should
michael@0:       // never happen, but rounding errors will cause this from time to time.
michael@0:       // These "impossible" colors will cause overflows (and hence random pixel
michael@0:       // values) when the resulting bitmap is drawn to the screen.
michael@0:       //
michael@0:       // We only need to do this when generating the final output row (here).
michael@0:       int max_color_channel = std::max(out_row[byte_offset + R_OFFSET_IDX],
michael@0:           std::max(out_row[byte_offset + G_OFFSET_IDX], out_row[byte_offset + B_OFFSET_IDX]));
michael@0:       if (alpha < max_color_channel)
michael@0:         out_row[byte_offset + A_OFFSET_IDX] = max_color_channel;
michael@0:       else
michael@0:         out_row[byte_offset + A_OFFSET_IDX] = alpha;
michael@0:     } else {
michael@0:       // No alpha channel, the image is opaque.
michael@0:       out_row[byte_offset + A_OFFSET_IDX] = 0xff;
michael@0:     }
michael@0:   }
michael@0: }
michael@0: 
michael@0: 
michael@0: // Convolves horizontally along a single row. The row data is given in
michael@0: // |src_data| and continues for the num_values() of the filter.
michael@0: void ConvolveHorizontally_SSE2(const unsigned char* src_data,
michael@0:                                const ConvolutionFilter1D& filter,
michael@0:                                unsigned char* out_row) {
michael@0: #if defined(SIMD_SSE2)
michael@0:   int num_values = filter.num_values();
michael@0: 
michael@0:   int filter_offset, filter_length;
michael@0:   __m128i zero = _mm_setzero_si128();
michael@0:   __m128i mask[4];
michael@0:   // |mask| will be used to decimate all extra filter coefficients that are
michael@0:   // loaded by SIMD when |filter_length| is not divisible by 4.
michael@0:   // mask[0] is not used in following algorithm.
michael@0:   mask[1] = _mm_set_epi16(0, 0, 0, 0, 0, 0, 0, -1);
michael@0:   mask[2] = _mm_set_epi16(0, 0, 0, 0, 0, 0, -1, -1);
michael@0:   mask[3] = _mm_set_epi16(0, 0, 0, 0, 0, -1, -1, -1);
michael@0: 
michael@0:   // Output one pixel each iteration, calculating all channels (RGBA) together.
michael@0:   for (int out_x = 0; out_x < num_values; out_x++) {
michael@0:     const ConvolutionFilter1D::Fixed* filter_values =
michael@0:         filter.FilterForValue(out_x, &filter_offset, &filter_length);
michael@0: 
michael@0:     __m128i accum = _mm_setzero_si128();
michael@0: 
michael@0:     // Compute the first pixel in this row that the filter affects. It will
michael@0:     // touch |filter_length| pixels (4 bytes each) after this.
michael@0:     const __m128i* row_to_filter =
michael@0:         reinterpret_cast<const __m128i*>(&src_data[filter_offset << 2]);
michael@0: 
michael@0:     // We will load and accumulate with four coefficients per iteration.
michael@0:     for (int filter_x = 0; filter_x < filter_length >> 2; filter_x++) {
michael@0: 
michael@0:       // Load 4 coefficients => duplicate 1st and 2nd of them for all channels.
michael@0:       __m128i coeff, coeff16;
michael@0:       // [16] xx xx xx xx c3 c2 c1 c0
michael@0:       coeff = _mm_loadl_epi64(reinterpret_cast<const __m128i*>(filter_values));
michael@0:       // [16] xx xx xx xx c1 c1 c0 c0
michael@0:       coeff16 = _mm_shufflelo_epi16(coeff, _MM_SHUFFLE(1, 1, 0, 0));
michael@0:       // [16] c1 c1 c1 c1 c0 c0 c0 c0
michael@0:       coeff16 = _mm_unpacklo_epi16(coeff16, coeff16);
michael@0: 
michael@0:       // Load four pixels => unpack the first two pixels to 16 bits =>
michael@0:       // multiply with coefficients => accumulate the convolution result.
michael@0:       // [8] a3 b3 g3 r3 a2 b2 g2 r2 a1 b1 g1 r1 a0 b0 g0 r0
michael@0:       __m128i src8 = _mm_loadu_si128(row_to_filter);
michael@0:       // [16] a1 b1 g1 r1 a0 b0 g0 r0
michael@0:       __m128i src16 = _mm_unpacklo_epi8(src8, zero);
michael@0:       __m128i mul_hi = _mm_mulhi_epi16(src16, coeff16);
michael@0:       __m128i mul_lo = _mm_mullo_epi16(src16, coeff16);
michael@0:       // [32]  a0*c0 b0*c0 g0*c0 r0*c0
michael@0:       __m128i t = _mm_unpacklo_epi16(mul_lo, mul_hi);
michael@0:       accum = _mm_add_epi32(accum, t);
michael@0:       // [32]  a1*c1 b1*c1 g1*c1 r1*c1
michael@0:       t = _mm_unpackhi_epi16(mul_lo, mul_hi);
michael@0:       accum = _mm_add_epi32(accum, t);
michael@0: 
michael@0:       // Duplicate 3rd and 4th coefficients for all channels =>
michael@0:       // unpack the 3rd and 4th pixels to 16 bits => multiply with coefficients
michael@0:       // => accumulate the convolution results.
michael@0:       // [16] xx xx xx xx c3 c3 c2 c2
michael@0:       coeff16 = _mm_shufflelo_epi16(coeff, _MM_SHUFFLE(3, 3, 2, 2));
michael@0:       // [16] c3 c3 c3 c3 c2 c2 c2 c2
michael@0:       coeff16 = _mm_unpacklo_epi16(coeff16, coeff16);
michael@0:       // [16] a3 g3 b3 r3 a2 g2 b2 r2
michael@0:       src16 = _mm_unpackhi_epi8(src8, zero);
michael@0:       mul_hi = _mm_mulhi_epi16(src16, coeff16);
michael@0:       mul_lo = _mm_mullo_epi16(src16, coeff16);
michael@0:       // [32]  a2*c2 b2*c2 g2*c2 r2*c2
michael@0:       t = _mm_unpacklo_epi16(mul_lo, mul_hi);
michael@0:       accum = _mm_add_epi32(accum, t);
michael@0:       // [32]  a3*c3 b3*c3 g3*c3 r3*c3
michael@0:       t = _mm_unpackhi_epi16(mul_lo, mul_hi);
michael@0:       accum = _mm_add_epi32(accum, t);
michael@0: 
michael@0:       // Advance the pixel and coefficients pointers.
michael@0:       row_to_filter += 1;
michael@0:       filter_values += 4;
michael@0:     }
michael@0: 
michael@0:     // When |filter_length| is not divisible by 4, we need to decimate some of
michael@0:     // the filter coefficient that was loaded incorrectly to zero; Other than
michael@0:     // that the algorithm is same with above, exceot that the 4th pixel will be
michael@0:     // always absent.
michael@0:     int r = filter_length&3;
michael@0:     if (r) {
michael@0:       // Note: filter_values must be padded to align_up(filter_offset, 8).
michael@0:       __m128i coeff, coeff16;
michael@0:       coeff = _mm_loadl_epi64(reinterpret_cast<const __m128i*>(filter_values));
michael@0:       // Mask out extra filter taps.
michael@0:       coeff = _mm_and_si128(coeff, mask[r]);
michael@0:       coeff16 = _mm_shufflelo_epi16(coeff, _MM_SHUFFLE(1, 1, 0, 0));
michael@0:       coeff16 = _mm_unpacklo_epi16(coeff16, coeff16);
michael@0: 
michael@0:       // Note: line buffer must be padded to align_up(filter_offset, 16).
michael@0:       // We resolve this by use C-version for the last horizontal line.
michael@0:       __m128i src8 = _mm_loadu_si128(row_to_filter);
michael@0:       __m128i src16 = _mm_unpacklo_epi8(src8, zero);
michael@0:       __m128i mul_hi = _mm_mulhi_epi16(src16, coeff16);
michael@0:       __m128i mul_lo = _mm_mullo_epi16(src16, coeff16);
michael@0:       __m128i t = _mm_unpacklo_epi16(mul_lo, mul_hi);
michael@0:       accum = _mm_add_epi32(accum, t);
michael@0:       t = _mm_unpackhi_epi16(mul_lo, mul_hi);
michael@0:       accum = _mm_add_epi32(accum, t);
michael@0: 
michael@0:       src16 = _mm_unpackhi_epi8(src8, zero);
michael@0:       coeff16 = _mm_shufflelo_epi16(coeff, _MM_SHUFFLE(3, 3, 2, 2));
michael@0:       coeff16 = _mm_unpacklo_epi16(coeff16, coeff16);
michael@0:       mul_hi = _mm_mulhi_epi16(src16, coeff16);
michael@0:       mul_lo = _mm_mullo_epi16(src16, coeff16);
michael@0:       t = _mm_unpacklo_epi16(mul_lo, mul_hi);
michael@0:       accum = _mm_add_epi32(accum, t);
michael@0:     }
michael@0: 
michael@0:     // Shift right for fixed point implementation.
michael@0:     accum = _mm_srai_epi32(accum, ConvolutionFilter1D::kShiftBits);
michael@0: 
michael@0:     // Packing 32 bits |accum| to 16 bits per channel (signed saturation).
michael@0:     accum = _mm_packs_epi32(accum, zero);
michael@0:     // Packing 16 bits |accum| to 8 bits per channel (unsigned saturation).
michael@0:     accum = _mm_packus_epi16(accum, zero);
michael@0: 
michael@0:     // Store the pixel value of 32 bits.
michael@0:     *(reinterpret_cast<int*>(out_row)) = _mm_cvtsi128_si32(accum);
michael@0:     out_row += 4;
michael@0:   }
michael@0: #endif
michael@0: }
michael@0: 
michael@0: // Convolves horizontally along four rows. The row data is given in
michael@0: // |src_data| and continues for the num_values() of the filter.
michael@0: // The algorithm is almost same as |ConvolveHorizontally_SSE2|. Please
michael@0: // refer to that function for detailed comments.
michael@0: void ConvolveHorizontally4_SSE2(const unsigned char* src_data[4],
michael@0:                                 const ConvolutionFilter1D& filter,
michael@0:                                 unsigned char* out_row[4]) {
michael@0: #if defined(SIMD_SSE2)
michael@0:   int num_values = filter.num_values();
michael@0: 
michael@0:   int filter_offset, filter_length;
michael@0:   __m128i zero = _mm_setzero_si128();
michael@0:   __m128i mask[4];
michael@0:   // |mask| will be used to decimate all extra filter coefficients that are
michael@0:   // loaded by SIMD when |filter_length| is not divisible by 4.
michael@0:   // mask[0] is not used in following algorithm.
michael@0:   mask[1] = _mm_set_epi16(0, 0, 0, 0, 0, 0, 0, -1);
michael@0:   mask[2] = _mm_set_epi16(0, 0, 0, 0, 0, 0, -1, -1);
michael@0:   mask[3] = _mm_set_epi16(0, 0, 0, 0, 0, -1, -1, -1);
michael@0: 
michael@0:   // Output one pixel each iteration, calculating all channels (RGBA) together.
michael@0:   for (int out_x = 0; out_x < num_values; out_x++) {
michael@0:     const ConvolutionFilter1D::Fixed* filter_values =
michael@0:         filter.FilterForValue(out_x, &filter_offset, &filter_length);
michael@0: 
michael@0:     // four pixels in a column per iteration.
michael@0:     __m128i accum0 = _mm_setzero_si128();
michael@0:     __m128i accum1 = _mm_setzero_si128();
michael@0:     __m128i accum2 = _mm_setzero_si128();
michael@0:     __m128i accum3 = _mm_setzero_si128();
michael@0:     int start = (filter_offset<<2);
michael@0:     // We will load and accumulate with four coefficients per iteration.
michael@0:     for (int filter_x = 0; filter_x < (filter_length >> 2); filter_x++) {
michael@0:       __m128i coeff, coeff16lo, coeff16hi;
michael@0:       // [16] xx xx xx xx c3 c2 c1 c0
michael@0:       coeff = _mm_loadl_epi64(reinterpret_cast<const __m128i*>(filter_values));
michael@0:       // [16] xx xx xx xx c1 c1 c0 c0
michael@0:       coeff16lo = _mm_shufflelo_epi16(coeff, _MM_SHUFFLE(1, 1, 0, 0));
michael@0:       // [16] c1 c1 c1 c1 c0 c0 c0 c0
michael@0:       coeff16lo = _mm_unpacklo_epi16(coeff16lo, coeff16lo);
michael@0:       // [16] xx xx xx xx c3 c3 c2 c2
michael@0:       coeff16hi = _mm_shufflelo_epi16(coeff, _MM_SHUFFLE(3, 3, 2, 2));
michael@0:       // [16] c3 c3 c3 c3 c2 c2 c2 c2
michael@0:       coeff16hi = _mm_unpacklo_epi16(coeff16hi, coeff16hi);
michael@0: 
michael@0:       __m128i src8, src16, mul_hi, mul_lo, t;
michael@0: 
michael@0: #define ITERATION(src, accum)                                          \
michael@0:       src8 = _mm_loadu_si128(reinterpret_cast<const __m128i*>(src));   \
michael@0:       src16 = _mm_unpacklo_epi8(src8, zero);                           \
michael@0:       mul_hi = _mm_mulhi_epi16(src16, coeff16lo);                      \
michael@0:       mul_lo = _mm_mullo_epi16(src16, coeff16lo);                      \
michael@0:       t = _mm_unpacklo_epi16(mul_lo, mul_hi);                          \
michael@0:       accum = _mm_add_epi32(accum, t);                                 \
michael@0:       t = _mm_unpackhi_epi16(mul_lo, mul_hi);                          \
michael@0:       accum = _mm_add_epi32(accum, t);                                 \
michael@0:       src16 = _mm_unpackhi_epi8(src8, zero);                           \
michael@0:       mul_hi = _mm_mulhi_epi16(src16, coeff16hi);                      \
michael@0:       mul_lo = _mm_mullo_epi16(src16, coeff16hi);                      \
michael@0:       t = _mm_unpacklo_epi16(mul_lo, mul_hi);                          \
michael@0:       accum = _mm_add_epi32(accum, t);                                 \
michael@0:       t = _mm_unpackhi_epi16(mul_lo, mul_hi);                          \
michael@0:       accum = _mm_add_epi32(accum, t)
michael@0: 
michael@0:       ITERATION(src_data[0] + start, accum0);
michael@0:       ITERATION(src_data[1] + start, accum1);
michael@0:       ITERATION(src_data[2] + start, accum2);
michael@0:       ITERATION(src_data[3] + start, accum3);
michael@0: 
michael@0:       start += 16;
michael@0:       filter_values += 4;
michael@0:     }
michael@0: 
michael@0:     int r = filter_length & 3;
michael@0:     if (r) {
michael@0:       // Note: filter_values must be padded to align_up(filter_offset, 8);
michael@0:       __m128i coeff;
michael@0:       coeff = _mm_loadl_epi64(reinterpret_cast<const __m128i*>(filter_values));
michael@0:       // Mask out extra filter taps.
michael@0:       coeff = _mm_and_si128(coeff, mask[r]);
michael@0: 
michael@0:       __m128i coeff16lo = _mm_shufflelo_epi16(coeff, _MM_SHUFFLE(1, 1, 0, 0));
michael@0:       /* c1 c1 c1 c1 c0 c0 c0 c0 */
michael@0:       coeff16lo = _mm_unpacklo_epi16(coeff16lo, coeff16lo);
michael@0:       __m128i coeff16hi = _mm_shufflelo_epi16(coeff, _MM_SHUFFLE(3, 3, 2, 2));
michael@0:       coeff16hi = _mm_unpacklo_epi16(coeff16hi, coeff16hi);
michael@0: 
michael@0:       __m128i src8, src16, mul_hi, mul_lo, t;
michael@0: 
michael@0:       ITERATION(src_data[0] + start, accum0);
michael@0:       ITERATION(src_data[1] + start, accum1);
michael@0:       ITERATION(src_data[2] + start, accum2);
michael@0:       ITERATION(src_data[3] + start, accum3);
michael@0:     }
michael@0: 
michael@0:     accum0 = _mm_srai_epi32(accum0, ConvolutionFilter1D::kShiftBits);
michael@0:     accum0 = _mm_packs_epi32(accum0, zero);
michael@0:     accum0 = _mm_packus_epi16(accum0, zero);
michael@0:     accum1 = _mm_srai_epi32(accum1, ConvolutionFilter1D::kShiftBits);
michael@0:     accum1 = _mm_packs_epi32(accum1, zero);
michael@0:     accum1 = _mm_packus_epi16(accum1, zero);
michael@0:     accum2 = _mm_srai_epi32(accum2, ConvolutionFilter1D::kShiftBits);
michael@0:     accum2 = _mm_packs_epi32(accum2, zero);
michael@0:     accum2 = _mm_packus_epi16(accum2, zero);
michael@0:     accum3 = _mm_srai_epi32(accum3, ConvolutionFilter1D::kShiftBits);
michael@0:     accum3 = _mm_packs_epi32(accum3, zero);
michael@0:     accum3 = _mm_packus_epi16(accum3, zero);
michael@0: 
michael@0:     *(reinterpret_cast<int*>(out_row[0])) = _mm_cvtsi128_si32(accum0);
michael@0:     *(reinterpret_cast<int*>(out_row[1])) = _mm_cvtsi128_si32(accum1);
michael@0:     *(reinterpret_cast<int*>(out_row[2])) = _mm_cvtsi128_si32(accum2);
michael@0:     *(reinterpret_cast<int*>(out_row[3])) = _mm_cvtsi128_si32(accum3);
michael@0: 
michael@0:     out_row[0] += 4;
michael@0:     out_row[1] += 4;
michael@0:     out_row[2] += 4;
michael@0:     out_row[3] += 4;
michael@0:   }
michael@0: #endif
michael@0: }
michael@0: 
michael@0: // Does vertical convolution to produce one output row. The filter values and
michael@0: // length are given in the first two parameters. These are applied to each
michael@0: // of the rows pointed to in the |source_data_rows| array, with each row
michael@0: // being |pixel_width| wide.
michael@0: //
michael@0: // The output must have room for |pixel_width * 4| bytes.
michael@0: template<bool has_alpha>
michael@0: void ConvolveVertically_SSE2(const ConvolutionFilter1D::Fixed* filter_values,
michael@0:                              int filter_length,
michael@0:                              unsigned char* const* source_data_rows,
michael@0:                              int pixel_width,
michael@0:                              unsigned char* out_row) {
michael@0: #if defined(SIMD_SSE2)
michael@0:   int width = pixel_width & ~3;
michael@0: 
michael@0:   __m128i zero = _mm_setzero_si128();
michael@0:   __m128i accum0, accum1, accum2, accum3, coeff16;
michael@0:   const __m128i* src;
michael@0:   // Output four pixels per iteration (16 bytes).
michael@0:   for (int out_x = 0; out_x < width; out_x += 4) {
michael@0: 
michael@0:     // Accumulated result for each pixel. 32 bits per RGBA channel.
michael@0:     accum0 = _mm_setzero_si128();
michael@0:     accum1 = _mm_setzero_si128();
michael@0:     accum2 = _mm_setzero_si128();
michael@0:     accum3 = _mm_setzero_si128();
michael@0: 
michael@0:     // Convolve with one filter coefficient per iteration.
michael@0:     for (int filter_y = 0; filter_y < filter_length; filter_y++) {
michael@0: 
michael@0:       // Duplicate the filter coefficient 8 times.
michael@0:       // [16] cj cj cj cj cj cj cj cj
michael@0:       coeff16 = _mm_set1_epi16(filter_values[filter_y]);
michael@0: 
michael@0:       // Load four pixels (16 bytes) together.
michael@0:       // [8] a3 b3 g3 r3 a2 b2 g2 r2 a1 b1 g1 r1 a0 b0 g0 r0
michael@0:       src = reinterpret_cast<const __m128i*>(
michael@0:           &source_data_rows[filter_y][out_x << 2]);
michael@0:       __m128i src8 = _mm_loadu_si128(src);
michael@0: 
michael@0:       // Unpack 1st and 2nd pixels from 8 bits to 16 bits for each channels =>
michael@0:       // multiply with current coefficient => accumulate the result.
michael@0:       // [16] a1 b1 g1 r1 a0 b0 g0 r0
michael@0:       __m128i src16 = _mm_unpacklo_epi8(src8, zero);
michael@0:       __m128i mul_hi = _mm_mulhi_epi16(src16, coeff16);
michael@0:       __m128i mul_lo = _mm_mullo_epi16(src16, coeff16);
michael@0:       // [32] a0 b0 g0 r0
michael@0:       __m128i t = _mm_unpacklo_epi16(mul_lo, mul_hi);
michael@0:       accum0 = _mm_add_epi32(accum0, t);
michael@0:       // [32] a1 b1 g1 r1
michael@0:       t = _mm_unpackhi_epi16(mul_lo, mul_hi);
michael@0:       accum1 = _mm_add_epi32(accum1, t);
michael@0: 
michael@0:       // Unpack 3rd and 4th pixels from 8 bits to 16 bits for each channels =>
michael@0:       // multiply with current coefficient => accumulate the result.
michael@0:       // [16] a3 b3 g3 r3 a2 b2 g2 r2
michael@0:       src16 = _mm_unpackhi_epi8(src8, zero);
michael@0:       mul_hi = _mm_mulhi_epi16(src16, coeff16);
michael@0:       mul_lo = _mm_mullo_epi16(src16, coeff16);
michael@0:       // [32] a2 b2 g2 r2
michael@0:       t = _mm_unpacklo_epi16(mul_lo, mul_hi);
michael@0:       accum2 = _mm_add_epi32(accum2, t);
michael@0:       // [32] a3 b3 g3 r3
michael@0:       t = _mm_unpackhi_epi16(mul_lo, mul_hi);
michael@0:       accum3 = _mm_add_epi32(accum3, t);
michael@0:     }
michael@0: 
michael@0:     // Shift right for fixed point implementation.
michael@0:     accum0 = _mm_srai_epi32(accum0, ConvolutionFilter1D::kShiftBits);
michael@0:     accum1 = _mm_srai_epi32(accum1, ConvolutionFilter1D::kShiftBits);
michael@0:     accum2 = _mm_srai_epi32(accum2, ConvolutionFilter1D::kShiftBits);
michael@0:     accum3 = _mm_srai_epi32(accum3, ConvolutionFilter1D::kShiftBits);
michael@0: 
michael@0:     // Packing 32 bits |accum| to 16 bits per channel (signed saturation).
michael@0:     // [16] a1 b1 g1 r1 a0 b0 g0 r0
michael@0:     accum0 = _mm_packs_epi32(accum0, accum1);
michael@0:     // [16] a3 b3 g3 r3 a2 b2 g2 r2
michael@0:     accum2 = _mm_packs_epi32(accum2, accum3);
michael@0: 
michael@0:     // Packing 16 bits |accum| to 8 bits per channel (unsigned saturation).
michael@0:     // [8] a3 b3 g3 r3 a2 b2 g2 r2 a1 b1 g1 r1 a0 b0 g0 r0
michael@0:     accum0 = _mm_packus_epi16(accum0, accum2);
michael@0: 
michael@0:     if (has_alpha) {
michael@0:       // Compute the max(ri, gi, bi) for each pixel.
michael@0:       // [8] xx a3 b3 g3 xx a2 b2 g2 xx a1 b1 g1 xx a0 b0 g0
michael@0:       __m128i a = _mm_srli_epi32(accum0, 8);
michael@0:       // [8] xx xx xx max3 xx xx xx max2 xx xx xx max1 xx xx xx max0
michael@0:       __m128i b = _mm_max_epu8(a, accum0);  // Max of r and g.
michael@0:       // [8] xx xx a3 b3 xx xx a2 b2 xx xx a1 b1 xx xx a0 b0
michael@0:       a = _mm_srli_epi32(accum0, 16);
michael@0:       // [8] xx xx xx max3 xx xx xx max2 xx xx xx max1 xx xx xx max0
michael@0:       b = _mm_max_epu8(a, b);  // Max of r and g and b.
michael@0:       // [8] max3 00 00 00 max2 00 00 00 max1 00 00 00 max0 00 00 00
michael@0:       b = _mm_slli_epi32(b, 24);
michael@0: 
michael@0:       // Make sure the value of alpha channel is always larger than maximum
michael@0:       // value of color channels.
michael@0:       accum0 = _mm_max_epu8(b, accum0);
michael@0:     } else {
michael@0:       // Set value of alpha channels to 0xFF.
michael@0:       __m128i mask = _mm_set1_epi32(0xff000000);
michael@0:       accum0 = _mm_or_si128(accum0, mask);
michael@0:     }
michael@0: 
michael@0:     // Store the convolution result (16 bytes) and advance the pixel pointers.
michael@0:     _mm_storeu_si128(reinterpret_cast<__m128i*>(out_row), accum0);
michael@0:     out_row += 16;
michael@0:   }
michael@0: 
michael@0:   // When the width of the output is not divisible by 4, We need to save one
michael@0:   // pixel (4 bytes) each time. And also the fourth pixel is always absent.
michael@0:   if (pixel_width & 3) {
michael@0:     accum0 = _mm_setzero_si128();
michael@0:     accum1 = _mm_setzero_si128();
michael@0:     accum2 = _mm_setzero_si128();
michael@0:     for (int filter_y = 0; filter_y < filter_length; ++filter_y) {
michael@0:       coeff16 = _mm_set1_epi16(filter_values[filter_y]);
michael@0:       // [8] a3 b3 g3 r3 a2 b2 g2 r2 a1 b1 g1 r1 a0 b0 g0 r0
michael@0:       src = reinterpret_cast<const __m128i*>(
michael@0:           &source_data_rows[filter_y][width<<2]);
michael@0:       __m128i src8 = _mm_loadu_si128(src);
michael@0:       // [16] a1 b1 g1 r1 a0 b0 g0 r0
michael@0:       __m128i src16 = _mm_unpacklo_epi8(src8, zero);
michael@0:       __m128i mul_hi = _mm_mulhi_epi16(src16, coeff16);
michael@0:       __m128i mul_lo = _mm_mullo_epi16(src16, coeff16);
michael@0:       // [32] a0 b0 g0 r0
michael@0:       __m128i t = _mm_unpacklo_epi16(mul_lo, mul_hi);
michael@0:       accum0 = _mm_add_epi32(accum0, t);
michael@0:       // [32] a1 b1 g1 r1
michael@0:       t = _mm_unpackhi_epi16(mul_lo, mul_hi);
michael@0:       accum1 = _mm_add_epi32(accum1, t);
michael@0:       // [16] a3 b3 g3 r3 a2 b2 g2 r2
michael@0:       src16 = _mm_unpackhi_epi8(src8, zero);
michael@0:       mul_hi = _mm_mulhi_epi16(src16, coeff16);
michael@0:       mul_lo = _mm_mullo_epi16(src16, coeff16);
michael@0:       // [32] a2 b2 g2 r2
michael@0:       t = _mm_unpacklo_epi16(mul_lo, mul_hi);
michael@0:       accum2 = _mm_add_epi32(accum2, t);
michael@0:     }
michael@0: 
michael@0:     accum0 = _mm_srai_epi32(accum0, ConvolutionFilter1D::kShiftBits);
michael@0:     accum1 = _mm_srai_epi32(accum1, ConvolutionFilter1D::kShiftBits);
michael@0:     accum2 = _mm_srai_epi32(accum2, ConvolutionFilter1D::kShiftBits);
michael@0:     // [16] a1 b1 g1 r1 a0 b0 g0 r0
michael@0:     accum0 = _mm_packs_epi32(accum0, accum1);
michael@0:     // [16] a3 b3 g3 r3 a2 b2 g2 r2
michael@0:     accum2 = _mm_packs_epi32(accum2, zero);
michael@0:     // [8] a3 b3 g3 r3 a2 b2 g2 r2 a1 b1 g1 r1 a0 b0 g0 r0
michael@0:     accum0 = _mm_packus_epi16(accum0, accum2);
michael@0:     if (has_alpha) {
michael@0:       // [8] xx a3 b3 g3 xx a2 b2 g2 xx a1 b1 g1 xx a0 b0 g0
michael@0:       __m128i a = _mm_srli_epi32(accum0, 8);
michael@0:       // [8] xx xx xx max3 xx xx xx max2 xx xx xx max1 xx xx xx max0
michael@0:       __m128i b = _mm_max_epu8(a, accum0);  // Max of r and g.
michael@0:       // [8] xx xx a3 b3 xx xx a2 b2 xx xx a1 b1 xx xx a0 b0
michael@0:       a = _mm_srli_epi32(accum0, 16);
michael@0:       // [8] xx xx xx max3 xx xx xx max2 xx xx xx max1 xx xx xx max0
michael@0:       b = _mm_max_epu8(a, b);  // Max of r and g and b.
michael@0:       // [8] max3 00 00 00 max2 00 00 00 max1 00 00 00 max0 00 00 00
michael@0:       b = _mm_slli_epi32(b, 24);
michael@0:       accum0 = _mm_max_epu8(b, accum0);
michael@0:     } else {
michael@0:       __m128i mask = _mm_set1_epi32(0xff000000);
michael@0:       accum0 = _mm_or_si128(accum0, mask);
michael@0:     }
michael@0: 
michael@0:     for (int out_x = width; out_x < pixel_width; out_x++) {
michael@0:       *(reinterpret_cast<int*>(out_row)) = _mm_cvtsi128_si32(accum0);
michael@0:       accum0 = _mm_srli_si128(accum0, 4);
michael@0:       out_row += 4;
michael@0:     }
michael@0:   }
michael@0: #endif
michael@0: }
michael@0: 
michael@0: }  // namespace
michael@0: 
michael@0: // ConvolutionFilter1D ---------------------------------------------------------
michael@0: 
michael@0: ConvolutionFilter1D::ConvolutionFilter1D()
michael@0:     : max_filter_(0) {
michael@0: }
michael@0: 
michael@0: ConvolutionFilter1D::~ConvolutionFilter1D() {
michael@0: }
michael@0: 
michael@0: void ConvolutionFilter1D::AddFilter(int filter_offset,
michael@0:                                     const float* filter_values,
michael@0:                                     int filter_length) {
michael@0:   SkASSERT(filter_length > 0);
michael@0: 
michael@0:   std::vector<Fixed> fixed_values;
michael@0:   fixed_values.reserve(filter_length);
michael@0: 
michael@0:   for (int i = 0; i < filter_length; ++i)
michael@0:     fixed_values.push_back(FloatToFixed(filter_values[i]));
michael@0: 
michael@0:   AddFilter(filter_offset, &fixed_values[0], filter_length);
michael@0: }
michael@0: 
michael@0: void ConvolutionFilter1D::AddFilter(int filter_offset,
michael@0:                                     const Fixed* filter_values,
michael@0:                                     int filter_length) {
michael@0:   // It is common for leading/trailing filter values to be zeros. In such
michael@0:   // cases it is beneficial to only store the central factors.
michael@0:   // For a scaling to 1/4th in each dimension using a Lanczos-2 filter on
michael@0:   // a 1080p image this optimization gives a ~10% speed improvement.
michael@0:   int first_non_zero = 0;
michael@0:   while (first_non_zero < filter_length && filter_values[first_non_zero] == 0)
michael@0:     first_non_zero++;
michael@0: 
michael@0:   if (first_non_zero < filter_length) {
michael@0:     // Here we have at least one non-zero factor.
michael@0:     int last_non_zero = filter_length - 1;
michael@0:     while (last_non_zero >= 0 && filter_values[last_non_zero] == 0)
michael@0:       last_non_zero--;
michael@0: 
michael@0:     filter_offset += first_non_zero;
michael@0:     filter_length = last_non_zero + 1 - first_non_zero;
michael@0:     SkASSERT(filter_length > 0);
michael@0: 
michael@0:     for (int i = first_non_zero; i <= last_non_zero; i++)
michael@0:       filter_values_.push_back(filter_values[i]);
michael@0:   } else {
michael@0:     // Here all the factors were zeroes.
michael@0:     filter_length = 0;
michael@0:   }
michael@0: 
michael@0:   FilterInstance instance;
michael@0: 
michael@0:   // We pushed filter_length elements onto filter_values_
michael@0:   instance.data_location = (static_cast<int>(filter_values_.size()) -
michael@0:                             filter_length);
michael@0:   instance.offset = filter_offset;
michael@0:   instance.length = filter_length;
michael@0:   filters_.push_back(instance);
michael@0: 
michael@0:   max_filter_ = std::max(max_filter_, filter_length);
michael@0: }
michael@0: 
michael@0: void BGRAConvolve2D(const unsigned char* source_data,
michael@0:                     int source_byte_row_stride,
michael@0:                     bool source_has_alpha,
michael@0:                     const ConvolutionFilter1D& filter_x,
michael@0:                     const ConvolutionFilter1D& filter_y,
michael@0:                     int output_byte_row_stride,
michael@0:                     unsigned char* output,
michael@0:                     bool use_sse2) {
michael@0: #if !defined(SIMD_SSE2)
michael@0:   // Even we have runtime support for SSE2 instructions, since the binary
michael@0:   // was not built with SSE2 support, we had to fallback to C version.
michael@0:   use_sse2 = false;
michael@0: #endif
michael@0: 
michael@0:   int max_y_filter_size = filter_y.max_filter();
michael@0: 
michael@0:   // The next row in the input that we will generate a horizontally
michael@0:   // convolved row for. If the filter doesn't start at the beginning of the
michael@0:   // image (this is the case when we are only resizing a subset), then we
michael@0:   // don't want to generate any output rows before that. Compute the starting
michael@0:   // row for convolution as the first pixel for the first vertical filter.
michael@0:   int filter_offset, filter_length;
michael@0:   const ConvolutionFilter1D::Fixed* filter_values =
michael@0:       filter_y.FilterForValue(0, &filter_offset, &filter_length);
michael@0:   int next_x_row = filter_offset;
michael@0: 
michael@0:   // We loop over each row in the input doing a horizontal convolution. This
michael@0:   // will result in a horizontally convolved image. We write the results into
michael@0:   // a circular buffer of convolved rows and do vertical convolution as rows
michael@0:   // are available. This prevents us from having to store the entire
michael@0:   // intermediate image and helps cache coherency.
michael@0:   // We will need four extra rows to allow horizontal convolution could be done
michael@0:   // simultaneously. We also padding each row in row buffer to be aligned-up to
michael@0:   // 16 bytes.
michael@0:   // TODO(jiesun): We do not use aligned load from row buffer in vertical
michael@0:   // convolution pass yet. Somehow Windows does not like it.
michael@0:   int row_buffer_width = (filter_x.num_values() + 15) & ~0xF;
michael@0:   int row_buffer_height = max_y_filter_size + (use_sse2 ? 4 : 0);
michael@0:   CircularRowBuffer row_buffer(row_buffer_width,
michael@0:                                row_buffer_height,
michael@0:                                filter_offset);
michael@0: 
michael@0:   // Loop over every possible output row, processing just enough horizontal
michael@0:   // convolutions to run each subsequent vertical convolution.
michael@0:   SkASSERT(output_byte_row_stride >= filter_x.num_values() * 4);
michael@0:   int num_output_rows = filter_y.num_values();
michael@0: 
michael@0:   // We need to check which is the last line to convolve before we advance 4
michael@0:   // lines in one iteration.
michael@0:   int last_filter_offset, last_filter_length;
michael@0:   filter_y.FilterForValue(num_output_rows - 1, &last_filter_offset,
michael@0:                           &last_filter_length);
michael@0: 
michael@0:   for (int out_y = 0; out_y < num_output_rows; out_y++) {
michael@0:     filter_values = filter_y.FilterForValue(out_y,
michael@0:                                             &filter_offset, &filter_length);
michael@0: 
michael@0:     // Generate output rows until we have enough to run the current filter.
michael@0:     if (use_sse2) {
michael@0:       while (next_x_row < filter_offset + filter_length) {
michael@0:         if (next_x_row + 3 < last_filter_offset + last_filter_length - 1) {
michael@0:           const unsigned char* src[4];
michael@0:           unsigned char* out_row[4];
michael@0:           for (int i = 0; i < 4; ++i) {
michael@0:             src[i] = &source_data[(next_x_row + i) * source_byte_row_stride];
michael@0:             out_row[i] = row_buffer.AdvanceRow();
michael@0:           }
michael@0:           ConvolveHorizontally4_SSE2(src, filter_x, out_row);
michael@0:           next_x_row += 4;
michael@0:         } else {
michael@0:           // For the last row, SSE2 load possibly to access data beyond the
michael@0:           // image area. therefore we use C version here. 
michael@0:           if (next_x_row == last_filter_offset + last_filter_length - 1) {
michael@0:             if (source_has_alpha) {
michael@0:               ConvolveHorizontally<true>(
michael@0:                   &source_data[next_x_row * source_byte_row_stride],
michael@0:                   filter_x, row_buffer.AdvanceRow());
michael@0:             } else {
michael@0:               ConvolveHorizontally<false>(
michael@0:                   &source_data[next_x_row * source_byte_row_stride],
michael@0:                   filter_x, row_buffer.AdvanceRow());
michael@0:             }
michael@0:           } else {
michael@0:             ConvolveHorizontally_SSE2(
michael@0:                 &source_data[next_x_row * source_byte_row_stride],
michael@0:                 filter_x, row_buffer.AdvanceRow());
michael@0:           }
michael@0:           next_x_row++;
michael@0:         }
michael@0:       }
michael@0:     } else {
michael@0:       while (next_x_row < filter_offset + filter_length) {
michael@0:         if (source_has_alpha) {
michael@0:           ConvolveHorizontally<true>(
michael@0:               &source_data[next_x_row * source_byte_row_stride],
michael@0:               filter_x, row_buffer.AdvanceRow());
michael@0:         } else {
michael@0:           ConvolveHorizontally<false>(
michael@0:               &source_data[next_x_row * source_byte_row_stride],
michael@0:               filter_x, row_buffer.AdvanceRow());
michael@0:         }
michael@0:         next_x_row++;
michael@0:       }
michael@0:     }
michael@0: 
michael@0:     // Compute where in the output image this row of final data will go.
michael@0:     unsigned char* cur_output_row = &output[out_y * output_byte_row_stride];
michael@0: 
michael@0:     // Get the list of rows that the circular buffer has, in order.
michael@0:     int first_row_in_circular_buffer;
michael@0:     unsigned char* const* rows_to_convolve =
michael@0:         row_buffer.GetRowAddresses(&first_row_in_circular_buffer);
michael@0: 
michael@0:     // Now compute the start of the subset of those rows that the filter
michael@0:     // needs.
michael@0:     unsigned char* const* first_row_for_filter =
michael@0:         &rows_to_convolve[filter_offset - first_row_in_circular_buffer];
michael@0: 
michael@0:     if (source_has_alpha) {
michael@0:       if (use_sse2) {
michael@0:         ConvolveVertically_SSE2<true>(filter_values, filter_length,
michael@0:                                       first_row_for_filter,
michael@0:                                       filter_x.num_values(), cur_output_row);
michael@0:       } else {
michael@0:         ConvolveVertically<true>(filter_values, filter_length,
michael@0:                                  first_row_for_filter,
michael@0:                                  filter_x.num_values(), cur_output_row);
michael@0:       }
michael@0:     } else {
michael@0:       if (use_sse2) {
michael@0:         ConvolveVertically_SSE2<false>(filter_values, filter_length,
michael@0:                                        first_row_for_filter,
michael@0:                                        filter_x.num_values(), cur_output_row);
michael@0:       } else {
michael@0:         ConvolveVertically<false>(filter_values, filter_length,
michael@0:                                  first_row_for_filter,
michael@0:                                  filter_x.num_values(), cur_output_row);
michael@0:       }
michael@0:     }
michael@0:   }
michael@0: }
michael@0: 
michael@0: }  // namespace skia