michael@0: /*
michael@0:  *  Copyright 2011 The LibYuv Project Authors. All rights reserved.
michael@0:  *
michael@0:  *  Use of this source code is governed by a BSD-style license
michael@0:  *  that can be found in the LICENSE file in the root of the source
michael@0:  *  tree. An additional intellectual property rights grant can be found
michael@0:  *  in the file PATENTS. All contributing project authors may
michael@0:  *  be found in the AUTHORS file in the root of the source tree.
michael@0:  */
michael@0: 
michael@0: #include "libyuv/scale.h"
michael@0: 
michael@0: #include <assert.h>
michael@0: #include <string.h>
michael@0: 
michael@0: #include "libyuv/cpu_id.h"
michael@0: #include "libyuv/planar_functions.h"  // For CopyARGB
michael@0: #include "libyuv/row.h"
michael@0: #include "libyuv/scale_row.h"
michael@0: 
michael@0: #ifdef __cplusplus
michael@0: namespace libyuv {
michael@0: extern "C" {
michael@0: #endif
michael@0: 
michael@0: static __inline int Abs(int v) {
michael@0:   return v >= 0 ? v : -v;
michael@0: }
michael@0: 
michael@0: // ScaleARGB ARGB, 1/2
michael@0: // This is an optimized version for scaling down a ARGB to 1/2 of
michael@0: // its original size.
michael@0: static void ScaleARGBDown2(int src_width, int src_height,
michael@0:                            int dst_width, int dst_height,
michael@0:                            int src_stride, int dst_stride,
michael@0:                            const uint8* src_argb, uint8* dst_argb,
michael@0:                            int x, int dx, int y, int dy,
michael@0:                            enum FilterMode filtering) {
michael@0:   int j;
michael@0:   int row_stride = src_stride * (dy >> 16);
michael@0:   void (*ScaleARGBRowDown2)(const uint8* src_argb, ptrdiff_t src_stride,
michael@0:                             uint8* dst_argb, int dst_width) =
michael@0:     filtering == kFilterNone ? ScaleARGBRowDown2_C :
michael@0:         (filtering == kFilterLinear ? ScaleARGBRowDown2Linear_C :
michael@0:         ScaleARGBRowDown2Box_C);
michael@0:   assert(dx == 65536 * 2);  // Test scale factor of 2.
michael@0:   assert((dy & 0x1ffff) == 0);  // Test vertical scale is multiple of 2.
michael@0:   // Advance to odd row, even column.
michael@0:   if (filtering == kFilterBilinear) {
michael@0:     src_argb += (y >> 16) * src_stride + (x >> 16) * 4;
michael@0:   } else {
michael@0:     src_argb += (y >> 16) * src_stride + ((x >> 16) - 1) * 4;
michael@0:   }
michael@0: 
michael@0: #if defined(HAS_SCALEARGBROWDOWN2_SSE2)
michael@0:   if (TestCpuFlag(kCpuHasSSE2) && IS_ALIGNED(dst_width, 4) &&
michael@0:       IS_ALIGNED(src_argb, 16) && IS_ALIGNED(row_stride, 16) &&
michael@0:       IS_ALIGNED(dst_argb, 16) && IS_ALIGNED(dst_stride, 16)) {
michael@0:     ScaleARGBRowDown2 = filtering == kFilterNone ? ScaleARGBRowDown2_SSE2 :
michael@0:         (filtering == kFilterLinear ? ScaleARGBRowDown2Linear_SSE2 :
michael@0:         ScaleARGBRowDown2Box_SSE2);
michael@0:   }
michael@0: #elif defined(HAS_SCALEARGBROWDOWN2_NEON)
michael@0:   if (TestCpuFlag(kCpuHasNEON) && IS_ALIGNED(dst_width, 8) &&
michael@0:       IS_ALIGNED(src_argb, 4) && IS_ALIGNED(row_stride, 4)) {
michael@0:     ScaleARGBRowDown2 = filtering ? ScaleARGBRowDown2Box_NEON :
michael@0:         ScaleARGBRowDown2_NEON;
michael@0:   }
michael@0: #endif
michael@0: 
michael@0:   if (filtering == kFilterLinear) {
michael@0:     src_stride = 0;
michael@0:   }
michael@0:   for (j = 0; j < dst_height; ++j) {
michael@0:     ScaleARGBRowDown2(src_argb, src_stride, dst_argb, dst_width);
michael@0:     src_argb += row_stride;
michael@0:     dst_argb += dst_stride;
michael@0:   }
michael@0: }
michael@0: 
michael@0: // ScaleARGB ARGB, 1/4
michael@0: // This is an optimized version for scaling down a ARGB to 1/4 of
michael@0: // its original size.
michael@0: static void ScaleARGBDown4Box(int src_width, int src_height,
michael@0:                               int dst_width, int dst_height,
michael@0:                               int src_stride, int dst_stride,
michael@0:                               const uint8* src_argb, uint8* dst_argb,
michael@0:                               int x, int dx, int y, int dy) {
michael@0:   int j;
michael@0:   // Allocate 2 rows of ARGB.
michael@0:   const int kRowSize = (dst_width * 2 * 4 + 15) & ~15;
michael@0:   align_buffer_64(row, kRowSize * 2);
michael@0:   int row_stride = src_stride * (dy >> 16);
michael@0:   void (*ScaleARGBRowDown2)(const uint8* src_argb, ptrdiff_t src_stride,
michael@0:     uint8* dst_argb, int dst_width) = ScaleARGBRowDown2Box_C;
michael@0:   // Advance to odd row, even column.
michael@0:   src_argb += (y >> 16) * src_stride + (x >> 16) * 4;
michael@0:   assert(dx == 65536 * 4);  // Test scale factor of 4.
michael@0:   assert((dy & 0x3ffff) == 0);  // Test vertical scale is multiple of 4.
michael@0: #if defined(HAS_SCALEARGBROWDOWN2_SSE2)
michael@0:   if (TestCpuFlag(kCpuHasSSE2) && IS_ALIGNED(dst_width, 4) &&
michael@0:       IS_ALIGNED(src_argb, 16) && IS_ALIGNED(row_stride, 16) &&
michael@0:       IS_ALIGNED(dst_argb, 16) && IS_ALIGNED(dst_stride, 16)) {
michael@0:     ScaleARGBRowDown2 = ScaleARGBRowDown2Box_SSE2;
michael@0:   }
michael@0: #elif defined(HAS_SCALEARGBROWDOWN2_NEON)
michael@0:   if (TestCpuFlag(kCpuHasNEON) && IS_ALIGNED(dst_width, 8) &&
michael@0:       IS_ALIGNED(src_argb, 4) && IS_ALIGNED(row_stride, 4)) {
michael@0:     ScaleARGBRowDown2 = ScaleARGBRowDown2Box_NEON;
michael@0:   }
michael@0: #endif
michael@0:   for (j = 0; j < dst_height; ++j) {
michael@0:     ScaleARGBRowDown2(src_argb, src_stride, row, dst_width * 2);
michael@0:     ScaleARGBRowDown2(src_argb + src_stride * 2, src_stride,
michael@0:                       row + kRowSize, dst_width * 2);
michael@0:     ScaleARGBRowDown2(row, kRowSize, dst_argb, dst_width);
michael@0:     src_argb += row_stride;
michael@0:     dst_argb += dst_stride;
michael@0:   }
michael@0:   free_aligned_buffer_64(row);
michael@0: }
michael@0: 
michael@0: // ScaleARGB ARGB Even
michael@0: // This is an optimized version for scaling down a ARGB to even
michael@0: // multiple of its original size.
michael@0: static void ScaleARGBDownEven(int src_width, int src_height,
michael@0:                               int dst_width, int dst_height,
michael@0:                               int src_stride, int dst_stride,
michael@0:                               const uint8* src_argb, uint8* dst_argb,
michael@0:                               int x, int dx, int y, int dy,
michael@0:                               enum FilterMode filtering) {
michael@0:   int j;
michael@0:   int col_step = dx >> 16;
michael@0:   int row_stride = (dy >> 16) * src_stride;
michael@0:   void (*ScaleARGBRowDownEven)(const uint8* src_argb, ptrdiff_t src_stride,
michael@0:                                int src_step, uint8* dst_argb, int dst_width) =
michael@0:       filtering ? ScaleARGBRowDownEvenBox_C : ScaleARGBRowDownEven_C;
michael@0:   assert(IS_ALIGNED(src_width, 2));
michael@0:   assert(IS_ALIGNED(src_height, 2));
michael@0:   src_argb += (y >> 16) * src_stride + (x >> 16) * 4;
michael@0: #if defined(HAS_SCALEARGBROWDOWNEVEN_SSE2)
michael@0:   if (TestCpuFlag(kCpuHasSSE2) && IS_ALIGNED(dst_width, 4) &&
michael@0:       IS_ALIGNED(dst_argb, 16) && IS_ALIGNED(dst_stride, 16)) {
michael@0:     ScaleARGBRowDownEven = filtering ? ScaleARGBRowDownEvenBox_SSE2 :
michael@0:         ScaleARGBRowDownEven_SSE2;
michael@0:   }
michael@0: #elif defined(HAS_SCALEARGBROWDOWNEVEN_NEON)
michael@0:   if (TestCpuFlag(kCpuHasNEON) && IS_ALIGNED(dst_width, 4) &&
michael@0:       IS_ALIGNED(src_argb, 4)) {
michael@0:     ScaleARGBRowDownEven = filtering ? ScaleARGBRowDownEvenBox_NEON :
michael@0:         ScaleARGBRowDownEven_NEON;
michael@0:   }
michael@0: #endif
michael@0: 
michael@0:   if (filtering == kFilterLinear) {
michael@0:     src_stride = 0;
michael@0:   }
michael@0:   for (j = 0; j < dst_height; ++j) {
michael@0:     ScaleARGBRowDownEven(src_argb, src_stride, col_step, dst_argb, dst_width);
michael@0:     src_argb += row_stride;
michael@0:     dst_argb += dst_stride;
michael@0:   }
michael@0: }
michael@0: 
michael@0: // Scale ARGB down with bilinear interpolation.
michael@0: static void ScaleARGBBilinearDown(int src_width, int src_height,
michael@0:                                   int dst_width, int dst_height,
michael@0:                                   int src_stride, int dst_stride,
michael@0:                                   const uint8* src_argb, uint8* dst_argb,
michael@0:                                   int x, int dx, int y, int dy,
michael@0:                                   enum FilterMode filtering) {
michael@0:   int j;
michael@0:   int64 xlast = x + (int64)(dst_width - 1) * dx;
michael@0:   int64 xl = (dx >= 0) ? x : xlast;
michael@0:   int64 xr = (dx >= 0) ? xlast : x;
michael@0:   int clip_src_width;
michael@0:   xl = (xl >> 16) & ~3;  // Left edge aligned.
michael@0:   xr = (xr >> 16) + 1;  // Right most pixel used.
michael@0:   clip_src_width = (((xr - xl) + 1 + 3) & ~3) * 4;  // Width aligned to 4.
michael@0:   src_argb += xl * 4;
michael@0:   x -= (int)(xl << 16);
michael@0:   void (*InterpolateRow)(uint8* dst_argb, const uint8* src_argb,
michael@0:       ptrdiff_t src_stride, int dst_width, int source_y_fraction) =
michael@0:       InterpolateRow_C;
michael@0: #if defined(HAS_INTERPOLATEROW_SSE2)
michael@0:   if (TestCpuFlag(kCpuHasSSE2) && clip_src_width >= 16) {
michael@0:     InterpolateRow = InterpolateRow_Any_SSE2;
michael@0:     if (IS_ALIGNED(clip_src_width, 16)) {
michael@0:       InterpolateRow = InterpolateRow_Unaligned_SSE2;
michael@0:       if (IS_ALIGNED(src_argb, 16) && IS_ALIGNED(src_stride, 16)) {
michael@0:         InterpolateRow = InterpolateRow_SSE2;
michael@0:       }
michael@0:     }
michael@0:   }
michael@0: #endif
michael@0: #if defined(HAS_INTERPOLATEROW_SSSE3)
michael@0:   if (TestCpuFlag(kCpuHasSSSE3) && clip_src_width >= 16) {
michael@0:     InterpolateRow = InterpolateRow_Any_SSSE3;
michael@0:     if (IS_ALIGNED(clip_src_width, 16)) {
michael@0:       InterpolateRow = InterpolateRow_Unaligned_SSSE3;
michael@0:       if (IS_ALIGNED(src_argb, 16) && IS_ALIGNED(src_stride, 16)) {
michael@0:         InterpolateRow = InterpolateRow_SSSE3;
michael@0:       }
michael@0:     }
michael@0:   }
michael@0: #endif
michael@0: #if defined(HAS_INTERPOLATEROW_AVX2)
michael@0:   if (TestCpuFlag(kCpuHasAVX2) && clip_src_width >= 32) {
michael@0:     InterpolateRow = InterpolateRow_Any_AVX2;
michael@0:     if (IS_ALIGNED(clip_src_width, 32)) {
michael@0:       InterpolateRow = InterpolateRow_AVX2;
michael@0:     }
michael@0:   }
michael@0: #endif
michael@0: #if defined(HAS_INTERPOLATEROW_NEON)
michael@0:   if (TestCpuFlag(kCpuHasNEON) && clip_src_width >= 16) {
michael@0:     InterpolateRow = InterpolateRow_Any_NEON;
michael@0:     if (IS_ALIGNED(clip_src_width, 16)) {
michael@0:       InterpolateRow = InterpolateRow_NEON;
michael@0:     }
michael@0:   }
michael@0: #endif
michael@0: #if defined(HAS_INTERPOLATEROWS_MIPS_DSPR2)
michael@0:   if (TestCpuFlag(kCpuHasMIPS_DSPR2) && clip_src_width >= 4 &&
michael@0:       IS_ALIGNED(src_argb, 4) && IS_ALIGNED(src_stride, 4)) {
michael@0:     InterpolateRow = InterpolateRow_Any_MIPS_DSPR2;
michael@0:     if (IS_ALIGNED(clip_src_width, 4)) {
michael@0:       InterpolateRow = InterpolateRow_MIPS_DSPR2;
michael@0:     }
michael@0:   }
michael@0: #endif
michael@0:   void (*ScaleARGBFilterCols)(uint8* dst_argb, const uint8* src_argb,
michael@0:       int dst_width, int x, int dx) =
michael@0:       (src_width >= 32768) ? ScaleARGBFilterCols64_C : ScaleARGBFilterCols_C;
michael@0: #if defined(HAS_SCALEARGBFILTERCOLS_SSSE3)
michael@0:   if (TestCpuFlag(kCpuHasSSSE3) && src_width < 32768) {
michael@0:     ScaleARGBFilterCols = ScaleARGBFilterCols_SSSE3;
michael@0:   }
michael@0: #endif
michael@0:   // TODO(fbarchard): Consider not allocating row buffer for kFilterLinear.
michael@0:   // Allocate a row of ARGB.
michael@0:   align_buffer_64(row, clip_src_width * 4);
michael@0: 
michael@0:   const int max_y = (src_height - 1) << 16;
michael@0:   for (j = 0; j < dst_height; ++j) {
michael@0:     if (y > max_y) {
michael@0:       y = max_y;
michael@0:     }
michael@0:     int yi = y >> 16;
michael@0:     const uint8* src = src_argb + yi * src_stride;
michael@0:     if (filtering == kFilterLinear) {
michael@0:       ScaleARGBFilterCols(dst_argb, src, dst_width, x, dx);
michael@0:     } else {
michael@0:       int yf = (y >> 8) & 255;
michael@0:       InterpolateRow(row, src, src_stride, clip_src_width, yf);
michael@0:       ScaleARGBFilterCols(dst_argb, row, dst_width, x, dx);
michael@0:     }
michael@0:     dst_argb += dst_stride;
michael@0:     y += dy;
michael@0:   }
michael@0:   free_aligned_buffer_64(row);
michael@0: }
michael@0: 
michael@0: // Scale ARGB up with bilinear interpolation.
michael@0: static void ScaleARGBBilinearUp(int src_width, int src_height,
michael@0:                                 int dst_width, int dst_height,
michael@0:                                 int src_stride, int dst_stride,
michael@0:                                 const uint8* src_argb, uint8* dst_argb,
michael@0:                                 int x, int dx, int y, int dy,
michael@0:                                 enum FilterMode filtering) {
michael@0:   int j;
michael@0:   void (*InterpolateRow)(uint8* dst_argb, const uint8* src_argb,
michael@0:       ptrdiff_t src_stride, int dst_width, int source_y_fraction) =
michael@0:       InterpolateRow_C;
michael@0:   void (*ScaleARGBFilterCols)(uint8* dst_argb, const uint8* src_argb,
michael@0:       int dst_width, int x, int dx) =
michael@0:       filtering ? ScaleARGBFilterCols_C : ScaleARGBCols_C;
michael@0: #if defined(HAS_INTERPOLATEROW_SSE2)
michael@0:   if (TestCpuFlag(kCpuHasSSE2) && dst_width >= 4) {
michael@0:     InterpolateRow = InterpolateRow_Any_SSE2;
michael@0:     if (IS_ALIGNED(dst_width, 4)) {
michael@0:       InterpolateRow = InterpolateRow_Unaligned_SSE2;
michael@0:       if (IS_ALIGNED(dst_argb, 16) && IS_ALIGNED(dst_stride, 16)) {
michael@0:         InterpolateRow = InterpolateRow_SSE2;
michael@0:       }
michael@0:     }
michael@0:   }
michael@0: #endif
michael@0: #if defined(HAS_INTERPOLATEROW_SSSE3)
michael@0:   if (TestCpuFlag(kCpuHasSSSE3) && dst_width >= 4) {
michael@0:     InterpolateRow = InterpolateRow_Any_SSSE3;
michael@0:     if (IS_ALIGNED(dst_width, 4)) {
michael@0:       InterpolateRow = InterpolateRow_Unaligned_SSSE3;
michael@0:       if (IS_ALIGNED(dst_argb, 16) && IS_ALIGNED(dst_stride, 16)) {
michael@0:         InterpolateRow = InterpolateRow_SSSE3;
michael@0:       }
michael@0:     }
michael@0:   }
michael@0: #endif
michael@0: #if defined(HAS_INTERPOLATEROW_AVX2)
michael@0:   if (TestCpuFlag(kCpuHasAVX2) && dst_width >= 8) {
michael@0:     InterpolateRow = InterpolateRow_Any_AVX2;
michael@0:     if (IS_ALIGNED(dst_width, 8)) {
michael@0:       InterpolateRow = InterpolateRow_AVX2;
michael@0:     }
michael@0:   }
michael@0: #endif
michael@0: #if defined(HAS_INTERPOLATEROW_NEON)
michael@0:   if (TestCpuFlag(kCpuHasNEON) && dst_width >= 4) {
michael@0:     InterpolateRow = InterpolateRow_Any_NEON;
michael@0:     if (IS_ALIGNED(dst_width, 4)) {
michael@0:       InterpolateRow = InterpolateRow_NEON;
michael@0:     }
michael@0:   }
michael@0: #endif
michael@0: #if defined(HAS_INTERPOLATEROWS_MIPS_DSPR2)
michael@0:   if (TestCpuFlag(kCpuHasMIPS_DSPR2) && dst_width >= 1 &&
michael@0:       IS_ALIGNED(dst_argb, 4) && IS_ALIGNED(dst_stride, 4)) {
michael@0:     InterpolateRow = InterpolateRow_MIPS_DSPR2;
michael@0:   }
michael@0: #endif
michael@0:   if (src_width >= 32768) {
michael@0:     ScaleARGBFilterCols = filtering ?
michael@0:         ScaleARGBFilterCols64_C : ScaleARGBCols64_C;
michael@0:   }
michael@0: #if defined(HAS_SCALEARGBFILTERCOLS_SSSE3)
michael@0:   if (filtering && TestCpuFlag(kCpuHasSSSE3) && src_width < 32768) {
michael@0:     ScaleARGBFilterCols = ScaleARGBFilterCols_SSSE3;
michael@0:   }
michael@0: #endif
michael@0: #if defined(HAS_SCALEARGBCOLS_SSE2)
michael@0:   if (!filtering && TestCpuFlag(kCpuHasSSE2) && src_width < 32768) {
michael@0:     ScaleARGBFilterCols = ScaleARGBCols_SSE2;
michael@0:   }
michael@0: #endif
michael@0:   if (!filtering && src_width * 2 == dst_width && x < 0x8000) {
michael@0:     ScaleARGBFilterCols = ScaleARGBColsUp2_C;
michael@0: #if defined(HAS_SCALEARGBCOLSUP2_SSE2)
michael@0:     if (TestCpuFlag(kCpuHasSSE2) && IS_ALIGNED(dst_width, 8) &&
michael@0:         IS_ALIGNED(src_argb, 16) && IS_ALIGNED(src_stride, 16) &&
michael@0:         IS_ALIGNED(dst_argb, 16) && IS_ALIGNED(dst_stride, 16)) {
michael@0:       ScaleARGBFilterCols = ScaleARGBColsUp2_SSE2;
michael@0:     }
michael@0: #endif
michael@0:   }
michael@0: 
michael@0:   const int max_y = (src_height - 1) << 16;
michael@0:   if (y > max_y) {
michael@0:     y = max_y;
michael@0:   }
michael@0:   int yi = y >> 16;
michael@0:   const uint8* src = src_argb + yi * src_stride;
michael@0: 
michael@0:   // Allocate 2 rows of ARGB.
michael@0:   const int kRowSize = (dst_width * 4 + 15) & ~15;
michael@0:   align_buffer_64(row, kRowSize * 2);
michael@0: 
michael@0:   uint8* rowptr = row;
michael@0:   int rowstride = kRowSize;
michael@0:   int lasty = yi;
michael@0: 
michael@0:   ScaleARGBFilterCols(rowptr, src, dst_width, x, dx);
michael@0:   if (src_height > 1) {
michael@0:     src += src_stride;
michael@0:   }
michael@0:   ScaleARGBFilterCols(rowptr + rowstride, src, dst_width, x, dx);
michael@0:   src += src_stride;
michael@0: 
michael@0:   for (j = 0; j < dst_height; ++j) {
michael@0:     yi = y >> 16;
michael@0:     if (yi != lasty) {
michael@0:       if (y > max_y) {
michael@0:         y = max_y;
michael@0:         yi = y >> 16;
michael@0:         src = src_argb + yi * src_stride;
michael@0:       }
michael@0:       if (yi != lasty) {
michael@0:         ScaleARGBFilterCols(rowptr, src, dst_width, x, dx);
michael@0:         rowptr += rowstride;
michael@0:         rowstride = -rowstride;
michael@0:         lasty = yi;
michael@0:         src += src_stride;
michael@0:       }
michael@0:     }
michael@0:     if (filtering == kFilterLinear) {
michael@0:       InterpolateRow(dst_argb, rowptr, 0, dst_width * 4, 0);
michael@0:     } else {
michael@0:       int yf = (y >> 8) & 255;
michael@0:       InterpolateRow(dst_argb, rowptr, rowstride, dst_width * 4, yf);
michael@0:     }
michael@0:     dst_argb += dst_stride;
michael@0:     y += dy;
michael@0:   }
michael@0:   free_aligned_buffer_64(row);
michael@0: }
michael@0: 
michael@0: #ifdef YUVSCALEUP
michael@0: // Scale YUV to ARGB up with bilinear interpolation.
michael@0: static void ScaleYUVToARGBBilinearUp(int src_width, int src_height,
michael@0:                                      int dst_width, int dst_height,
michael@0:                                      int src_stride_y,
michael@0:                                      int src_stride_u,
michael@0:                                      int src_stride_v,
michael@0:                                      int dst_stride_argb,
michael@0:                                      const uint8* src_y,
michael@0:                                      const uint8* src_u,
michael@0:                                      const uint8* src_v,
michael@0:                                      uint8* dst_argb,
michael@0:                                      int x, int dx, int y, int dy,
michael@0:                                      enum FilterMode filtering) {
michael@0:   int j;
michael@0:   void (*I422ToARGBRow)(const uint8* y_buf,
michael@0:                         const uint8* u_buf,
michael@0:                         const uint8* v_buf,
michael@0:                         uint8* rgb_buf,
michael@0:                         int width) = I422ToARGBRow_C;
michael@0: #if defined(HAS_I422TOARGBROW_SSSE3)
michael@0:   if (TestCpuFlag(kCpuHasSSSE3) && src_width >= 8) {
michael@0:     I422ToARGBRow = I422ToARGBRow_Any_SSSE3;
michael@0:     if (IS_ALIGNED(src_width, 8)) {
michael@0:       I422ToARGBRow = I422ToARGBRow_Unaligned_SSSE3;
michael@0:       if (IS_ALIGNED(dst_argb, 16) && IS_ALIGNED(dst_stride_argb, 16)) {
michael@0:         I422ToARGBRow = I422ToARGBRow_SSSE3;
michael@0:       }
michael@0:     }
michael@0:   }
michael@0: #endif
michael@0: #if defined(HAS_I422TOARGBROW_AVX2)
michael@0:   if (TestCpuFlag(kCpuHasAVX2) && src_width >= 16) {
michael@0:     I422ToARGBRow = I422ToARGBRow_Any_AVX2;
michael@0:     if (IS_ALIGNED(src_width, 16)) {
michael@0:       I422ToARGBRow = I422ToARGBRow_AVX2;
michael@0:     }
michael@0:   }
michael@0: #endif
michael@0: #if defined(HAS_I422TOARGBROW_NEON)
michael@0:   if (TestCpuFlag(kCpuHasNEON) && src_width >= 8) {
michael@0:     I422ToARGBRow = I422ToARGBRow_Any_NEON;
michael@0:     if (IS_ALIGNED(src_width, 8)) {
michael@0:       I422ToARGBRow = I422ToARGBRow_NEON;
michael@0:     }
michael@0:   }
michael@0: #endif
michael@0: #if defined(HAS_I422TOARGBROW_MIPS_DSPR2)
michael@0:   if (TestCpuFlag(kCpuHasMIPS_DSPR2) && IS_ALIGNED(src_width, 4) &&
michael@0:       IS_ALIGNED(src_y, 4) && IS_ALIGNED(src_stride_y, 4) &&
michael@0:       IS_ALIGNED(src_u, 2) && IS_ALIGNED(src_stride_u, 2) &&
michael@0:       IS_ALIGNED(src_v, 2) && IS_ALIGNED(src_stride_v, 2) &&
michael@0:       IS_ALIGNED(dst_argb, 4) && IS_ALIGNED(dst_stride_argb, 4)) {
michael@0:     I422ToARGBRow = I422ToARGBRow_MIPS_DSPR2;
michael@0:   }
michael@0: #endif
michael@0: 
michael@0:   void (*InterpolateRow)(uint8* dst_argb, const uint8* src_argb,
michael@0:       ptrdiff_t src_stride, int dst_width, int source_y_fraction) =
michael@0:       InterpolateRow_C;
michael@0: #if defined(HAS_INTERPOLATEROW_SSE2)
michael@0:   if (TestCpuFlag(kCpuHasSSE2) && dst_width >= 4) {
michael@0:     InterpolateRow = InterpolateRow_Any_SSE2;
michael@0:     if (IS_ALIGNED(dst_width, 4)) {
michael@0:       InterpolateRow = InterpolateRow_Unaligned_SSE2;
michael@0:       if (IS_ALIGNED(dst_argb, 16) && IS_ALIGNED(dst_stride_argb, 16)) {
michael@0:         InterpolateRow = InterpolateRow_SSE2;
michael@0:       }
michael@0:     }
michael@0:   }
michael@0: #endif
michael@0: #if defined(HAS_INTERPOLATEROW_SSSE3)
michael@0:   if (TestCpuFlag(kCpuHasSSSE3) && dst_width >= 4) {
michael@0:     InterpolateRow = InterpolateRow_Any_SSSE3;
michael@0:     if (IS_ALIGNED(dst_width, 4)) {
michael@0:       InterpolateRow = InterpolateRow_Unaligned_SSSE3;
michael@0:       if (IS_ALIGNED(dst_argb, 16) && IS_ALIGNED(dst_stride_argb, 16)) {
michael@0:         InterpolateRow = InterpolateRow_SSSE3;
michael@0:       }
michael@0:     }
michael@0:   }
michael@0: #endif
michael@0: #if defined(HAS_INTERPOLATEROW_AVX2)
michael@0:   if (TestCpuFlag(kCpuHasAVX2) && dst_width >= 8) {
michael@0:     InterpolateRow = InterpolateRow_Any_AVX2;
michael@0:     if (IS_ALIGNED(dst_width, 8)) {
michael@0:       InterpolateRow = InterpolateRow_AVX2;
michael@0:     }
michael@0:   }
michael@0: #endif
michael@0: #if defined(HAS_INTERPOLATEROW_NEON)
michael@0:   if (TestCpuFlag(kCpuHasNEON) && dst_width >= 4) {
michael@0:     InterpolateRow = InterpolateRow_Any_NEON;
michael@0:     if (IS_ALIGNED(dst_width, 4)) {
michael@0:       InterpolateRow = InterpolateRow_NEON;
michael@0:     }
michael@0:   }
michael@0: #endif
michael@0: #if defined(HAS_INTERPOLATEROWS_MIPS_DSPR2)
michael@0:   if (TestCpuFlag(kCpuHasMIPS_DSPR2) && dst_width >= 1 &&
michael@0:       IS_ALIGNED(dst_argb, 4) && IS_ALIGNED(dst_stride_argb, 4)) {
michael@0:     InterpolateRow = InterpolateRow_MIPS_DSPR2;
michael@0:   }
michael@0: #endif
michael@0: 
michael@0:   void (*ScaleARGBFilterCols)(uint8* dst_argb, const uint8* src_argb,
michael@0:       int dst_width, int x, int dx) =
michael@0:       filtering ? ScaleARGBFilterCols_C : ScaleARGBCols_C;
michael@0:   if (src_width >= 32768) {
michael@0:     ScaleARGBFilterCols = filtering ?
michael@0:         ScaleARGBFilterCols64_C : ScaleARGBCols64_C;
michael@0:   }
michael@0: #if defined(HAS_SCALEARGBFILTERCOLS_SSSE3)
michael@0:   if (filtering && TestCpuFlag(kCpuHasSSSE3) && src_width < 32768) {
michael@0:     ScaleARGBFilterCols = ScaleARGBFilterCols_SSSE3;
michael@0:   }
michael@0: #endif
michael@0: #if defined(HAS_SCALEARGBCOLS_SSE2)
michael@0:   if (!filtering && TestCpuFlag(kCpuHasSSE2) && src_width < 32768) {
michael@0:     ScaleARGBFilterCols = ScaleARGBCols_SSE2;
michael@0:   }
michael@0: #endif
michael@0:   if (!filtering && src_width * 2 == dst_width && x < 0x8000) {
michael@0:     ScaleARGBFilterCols = ScaleARGBColsUp2_C;
michael@0: #if defined(HAS_SCALEARGBCOLSUP2_SSE2)
michael@0:     if (TestCpuFlag(kCpuHasSSE2) && IS_ALIGNED(dst_width, 8) &&
michael@0:         IS_ALIGNED(src_argb, 16) && IS_ALIGNED(src_stride, 16) &&
michael@0:         IS_ALIGNED(dst_argb, 16) && IS_ALIGNED(dst_stride, 16)) {
michael@0:       ScaleARGBFilterCols = ScaleARGBColsUp2_SSE2;
michael@0:     }
michael@0: #endif
michael@0:   }
michael@0: 
michael@0:   const int max_y = (src_height - 1) << 16;
michael@0:   if (y > max_y) {
michael@0:     y = max_y;
michael@0:   }
michael@0:   const int kYShift = 1;  // Shift Y by 1 to convert Y plane to UV coordinate.
michael@0:   int yi = y >> 16;
michael@0:   int uv_yi = yi >> kYShift;
michael@0:   const uint8* src_row_y = src_y + yi * src_stride_y;
michael@0:   const uint8* src_row_u = src_u + uv_yi * src_stride_u;
michael@0:   const uint8* src_row_v = src_v + uv_yi * src_stride_v;
michael@0: 
michael@0:   // Allocate 2 rows of ARGB.
michael@0:   const int kRowSize = (dst_width * 4 + 15) & ~15;
michael@0:   align_buffer_64(row, kRowSize * 2);
michael@0: 
michael@0:   // Allocate 1 row of ARGB for source conversion.
michael@0:   align_buffer_64(argb_row, src_width * 4);
michael@0: 
michael@0:   uint8* rowptr = row;
michael@0:   int rowstride = kRowSize;
michael@0:   int lasty = yi;
michael@0: 
michael@0:   // TODO(fbarchard): Convert first 2 rows of YUV to ARGB.
michael@0:   ScaleARGBFilterCols(rowptr, src_row_y, dst_width, x, dx);
michael@0:   if (src_height > 1) {
michael@0:     src_row_y += src_stride_y;
michael@0:     if (yi & 1) {
michael@0:       src_row_u += src_stride_u;
michael@0:       src_row_v += src_stride_v;
michael@0:     }
michael@0:   }
michael@0:   ScaleARGBFilterCols(rowptr + rowstride, src_row_y, dst_width, x, dx);
michael@0:   if (src_height > 2) {
michael@0:     src_row_y += src_stride_y;
michael@0:     if (!(yi & 1)) {
michael@0:       src_row_u += src_stride_u;
michael@0:       src_row_v += src_stride_v;
michael@0:     }
michael@0:   }
michael@0: 
michael@0:   for (j = 0; j < dst_height; ++j) {
michael@0:     yi = y >> 16;
michael@0:     if (yi != lasty) {
michael@0:       if (y > max_y) {
michael@0:         y = max_y;
michael@0:         yi = y >> 16;
michael@0:         uv_yi = yi >> kYShift;
michael@0:         src_row_y = src_y + yi * src_stride_y;
michael@0:         src_row_u = src_u + uv_yi * src_stride_u;
michael@0:         src_row_v = src_v + uv_yi * src_stride_v;
michael@0:       }
michael@0:       if (yi != lasty) {
michael@0:         // TODO(fbarchard): Convert the clipped region of row.
michael@0:         I422ToARGBRow(src_row_y, src_row_u, src_row_v, argb_row, src_width);
michael@0:         ScaleARGBFilterCols(rowptr, argb_row, dst_width, x, dx);
michael@0:         rowptr += rowstride;
michael@0:         rowstride = -rowstride;
michael@0:         lasty = yi;
michael@0:         src_row_y += src_stride_y;
michael@0:         if (yi & 1) {
michael@0:           src_row_u += src_stride_u;
michael@0:           src_row_v += src_stride_v;
michael@0:         }
michael@0:       }
michael@0:     }
michael@0:     if (filtering == kFilterLinear) {
michael@0:       InterpolateRow(dst_argb, rowptr, 0, dst_width * 4, 0);
michael@0:     } else {
michael@0:       int yf = (y >> 8) & 255;
michael@0:       InterpolateRow(dst_argb, rowptr, rowstride, dst_width * 4, yf);
michael@0:     }
michael@0:     dst_argb += dst_stride_argb;
michael@0:     y += dy;
michael@0:   }
michael@0:   free_aligned_buffer_64(row);
michael@0:   free_aligned_buffer_64(row_argb);
michael@0: }
michael@0: #endif
michael@0: 
michael@0: // Scale ARGB to/from any dimensions, without interpolation.
michael@0: // Fixed point math is used for performance: The upper 16 bits
michael@0: // of x and dx is the integer part of the source position and
michael@0: // the lower 16 bits are the fixed decimal part.
michael@0: 
michael@0: static void ScaleARGBSimple(int src_width, int src_height,
michael@0:                             int dst_width, int dst_height,
michael@0:                             int src_stride, int dst_stride,
michael@0:                             const uint8* src_argb, uint8* dst_argb,
michael@0:                             int x, int dx, int y, int dy) {
michael@0:   int j;
michael@0:   void (*ScaleARGBCols)(uint8* dst_argb, const uint8* src_argb,
michael@0:       int dst_width, int x, int dx) =
michael@0:       (src_width >= 32768) ? ScaleARGBCols64_C : ScaleARGBCols_C;
michael@0: #if defined(HAS_SCALEARGBCOLS_SSE2)
michael@0:   if (TestCpuFlag(kCpuHasSSE2) && src_width < 32768) {
michael@0:     ScaleARGBCols = ScaleARGBCols_SSE2;
michael@0:   }
michael@0: #endif
michael@0:   if (src_width * 2 == dst_width && x < 0x8000) {
michael@0:     ScaleARGBCols = ScaleARGBColsUp2_C;
michael@0: #if defined(HAS_SCALEARGBCOLSUP2_SSE2)
michael@0:     if (TestCpuFlag(kCpuHasSSE2) && IS_ALIGNED(dst_width, 8) &&
michael@0:         IS_ALIGNED(src_argb, 16) && IS_ALIGNED(src_stride, 16) &&
michael@0:         IS_ALIGNED(dst_argb, 16) && IS_ALIGNED(dst_stride, 16)) {
michael@0:       ScaleARGBCols = ScaleARGBColsUp2_SSE2;
michael@0:     }
michael@0: #endif
michael@0:   }
michael@0: 
michael@0:   for (j = 0; j < dst_height; ++j) {
michael@0:     ScaleARGBCols(dst_argb, src_argb + (y >> 16) * src_stride,
michael@0:                   dst_width, x, dx);
michael@0:     dst_argb += dst_stride;
michael@0:     y += dy;
michael@0:   }
michael@0: }
michael@0: 
michael@0: // ScaleARGB a ARGB.
michael@0: // This function in turn calls a scaling function
michael@0: // suitable for handling the desired resolutions.
michael@0: static void ScaleARGB(const uint8* src, int src_stride,
michael@0:                       int src_width, int src_height,
michael@0:                       uint8* dst, int dst_stride,
michael@0:                       int dst_width, int dst_height,
michael@0:                       int clip_x, int clip_y, int clip_width, int clip_height,
michael@0:                       enum FilterMode filtering) {
michael@0:   // Initial source x/y coordinate and step values as 16.16 fixed point.
michael@0:   int x = 0;
michael@0:   int y = 0;
michael@0:   int dx = 0;
michael@0:   int dy = 0;
michael@0:   // ARGB does not support box filter yet, but allow the user to pass it.
michael@0:   // Simplify filtering when possible.
michael@0:   filtering = ScaleFilterReduce(src_width, src_height,
michael@0:                                 dst_width, dst_height,
michael@0:                                 filtering);
michael@0: 
michael@0:   // Negative src_height means invert the image.
michael@0:   if (src_height < 0) {
michael@0:     src_height = -src_height;
michael@0:     src = src + (src_height - 1) * src_stride;
michael@0:     src_stride = -src_stride;
michael@0:   }
michael@0:   ScaleSlope(src_width, src_height, dst_width, dst_height, filtering,
michael@0:              &x, &y, &dx, &dy);
michael@0:   src_width = Abs(src_width);
michael@0:   if (clip_x) {
michael@0:     int64 clipf = (int64)(clip_x) * dx;
michael@0:     x += (clipf & 0xffff);
michael@0:     src += (clipf >> 16) * 4;
michael@0:     dst += clip_x * 4;
michael@0:   }
michael@0:   if (clip_y) {
michael@0:     int64 clipf = (int64)(clip_y) * dy;
michael@0:     y += (clipf & 0xffff);
michael@0:     src += (clipf >> 16) * src_stride;
michael@0:     dst += clip_y * dst_stride;
michael@0:   }
michael@0: 
michael@0:   // Special case for integer step values.
michael@0:   if (((dx | dy) & 0xffff) == 0) {
michael@0:     if (!dx || !dy) {  // 1 pixel wide and/or tall.
michael@0:       filtering = kFilterNone;
michael@0:     } else {
michael@0:       // Optimized even scale down. ie 2, 4, 6, 8, 10x.
michael@0:       if (!(dx & 0x10000) && !(dy & 0x10000)) {
michael@0:         if (dx == 0x20000) {
michael@0:           // Optimized 1/2 downsample.
michael@0:           ScaleARGBDown2(src_width, src_height,
michael@0:                          clip_width, clip_height,
michael@0:                          src_stride, dst_stride, src, dst,
michael@0:                          x, dx, y, dy, filtering);
michael@0:           return;
michael@0:         }
michael@0:         if (dx == 0x40000 && filtering == kFilterBox) {
michael@0:           // Optimized 1/4 box downsample.
michael@0:           ScaleARGBDown4Box(src_width, src_height,
michael@0:                             clip_width, clip_height,
michael@0:                             src_stride, dst_stride, src, dst,
michael@0:                             x, dx, y, dy);
michael@0:           return;
michael@0:         }
michael@0:         ScaleARGBDownEven(src_width, src_height,
michael@0:                           clip_width, clip_height,
michael@0:                           src_stride, dst_stride, src, dst,
michael@0:                           x, dx, y, dy, filtering);
michael@0:         return;
michael@0:       }
michael@0:       // Optimized odd scale down. ie 3, 5, 7, 9x.
michael@0:       if ((dx & 0x10000) && (dy & 0x10000)) {
michael@0:         filtering = kFilterNone;
michael@0:         if (dx == 0x10000 && dy == 0x10000) {
michael@0:           // Straight copy.
michael@0:           ARGBCopy(src + (y >> 16) * src_stride + (x >> 16) * 4, src_stride,
michael@0:                    dst, dst_stride, clip_width, clip_height);
michael@0:           return;
michael@0:         }
michael@0:       }
michael@0:     }
michael@0:   }
michael@0:   if (dx == 0x10000 && (x & 0xffff) == 0) {
michael@0:     // Arbitrary scale vertically, but unscaled vertically.
michael@0:     ScalePlaneVertical(src_height,
michael@0:                        clip_width, clip_height,
michael@0:                        src_stride, dst_stride, src, dst,
michael@0:                        x, y, dy, 4, filtering);
michael@0:     return;
michael@0:   }
michael@0:   if (filtering && dy < 65536) {
michael@0:     ScaleARGBBilinearUp(src_width, src_height,
michael@0:                         clip_width, clip_height,
michael@0:                         src_stride, dst_stride, src, dst,
michael@0:                         x, dx, y, dy, filtering);
michael@0:     return;
michael@0:   }
michael@0:   if (filtering) {
michael@0:     ScaleARGBBilinearDown(src_width, src_height,
michael@0:                           clip_width, clip_height,
michael@0:                           src_stride, dst_stride, src, dst,
michael@0:                           x, dx, y, dy, filtering);
michael@0:     return;
michael@0:   }
michael@0:   ScaleARGBSimple(src_width, src_height, clip_width, clip_height,
michael@0:                   src_stride, dst_stride, src, dst,
michael@0:                   x, dx, y, dy);
michael@0: }
michael@0: 
michael@0: LIBYUV_API
michael@0: int ARGBScaleClip(const uint8* src_argb, int src_stride_argb,
michael@0:                   int src_width, int src_height,
michael@0:                   uint8* dst_argb, int dst_stride_argb,
michael@0:                   int dst_width, int dst_height,
michael@0:                   int clip_x, int clip_y, int clip_width, int clip_height,
michael@0:                   enum FilterMode filtering) {
michael@0:   if (!src_argb || src_width == 0 || src_height == 0 ||
michael@0:       !dst_argb || dst_width <= 0 || dst_height <= 0 ||
michael@0:       clip_x < 0 || clip_y < 0 ||
michael@0:       (clip_x + clip_width) > dst_width ||
michael@0:       (clip_y + clip_height) > dst_height) {
michael@0:     return -1;
michael@0:   }
michael@0:   ScaleARGB(src_argb, src_stride_argb, src_width, src_height,
michael@0:             dst_argb, dst_stride_argb, dst_width, dst_height,
michael@0:             clip_x, clip_y, clip_width, clip_height, filtering);
michael@0:   return 0;
michael@0: }
michael@0: 
michael@0: // Scale an ARGB image.
michael@0: LIBYUV_API
michael@0: int ARGBScale(const uint8* src_argb, int src_stride_argb,
michael@0:               int src_width, int src_height,
michael@0:               uint8* dst_argb, int dst_stride_argb,
michael@0:               int dst_width, int dst_height,
michael@0:               enum FilterMode filtering) {
michael@0:   if (!src_argb || src_width == 0 || src_height == 0 ||
michael@0:       !dst_argb || dst_width <= 0 || dst_height <= 0) {
michael@0:     return -1;
michael@0:   }
michael@0:   ScaleARGB(src_argb, src_stride_argb, src_width, src_height,
michael@0:             dst_argb, dst_stride_argb, dst_width, dst_height,
michael@0:             0, 0, dst_width, dst_height, filtering);
michael@0:   return 0;
michael@0: }
michael@0: 
michael@0: #ifdef __cplusplus
michael@0: }  // extern "C"
michael@0: }  // namespace libyuv
michael@0: #endif