Import,RGBA: fix for BigEndian import

+ simplification of the logic

Change-Id: Ia20ce844793ed35ea03a17cef45838f3d0ae4afa
(cherry picked from commit 3b07d32712)

ChangeLog

@@ -1,3 +1,5 @@
+f66955de WEBP_REDUCE_CSP: restrict colorspace support
+a289d8e7 update ChangeLog (tag: v0.6.1-rc2)
 c10a493c vwebp: disable double buffering on windows & mac
 0d4466c2 webp_to_sdl.c: fix file mode
 1b27bf8b WEBP_REDUCE_SIZE: disable all rescaler code

extras/get_disto.c

@@ -322,6 +322,7 @@ int main(int argc, const char *argv[]) {
       fprintf(stderr, "Can only compute the difference map in ARGB format.\n");
       goto End;
     }
+#if !defined(WEBP_REDUCE_CSP)
     data_size = WebPEncodeLosslessBGRA((const uint8_t*)pic1.argb,
                                        pic1.width, pic1.height,
                                        pic1.argb_stride * 4,
@@ -333,6 +334,12 @@ int main(int argc, const char *argv[]) {
     ret = ImgIoUtilWriteFile(output, data, data_size) ? 0 : 1;
     WebPFree(data);
     if (ret) goto End;
+#else
+    (void)data;
+    (void)data_size;
+    fprintf(stderr, "Cannot save the difference map. Please recompile "
+                    "without the WEBP_REDUCE_CSP flag.\n");
+#endif  // WEBP_REDUCE_CSP
   }
   ret = 0;
 

imageio/webpdec.c

@@ -9,6 +9,10 @@
 //
 // WebP decode.
 
+#ifdef HAVE_CONFIG_H
+#include "webp/config.h"
+#endif
+
 #include "./webpdec.h"
 
 #include <stdio.h>
@@ -162,7 +166,11 @@ int ReadWebP(const uint8_t* const data, size_t data_size,
       break;
     }
     if (pic->use_argb) {
+#ifdef WORDS_BIGENDIAN
+      output_buffer->colorspace = MODE_ARGB;
+#else
       output_buffer->colorspace = MODE_BGRA;
+#endif
       output_buffer->u.RGBA.rgba = (uint8_t*)pic->argb;
       output_buffer->u.RGBA.stride = pic->argb_stride * sizeof(uint32_t);
       output_buffer->u.RGBA.size = output_buffer->u.RGBA.stride * pic->height;

src/dsp/alpha_processing.c

@@ -366,6 +366,16 @@ static WEBP_INLINE uint32_t MakeARGB32(int a, int r, int g, int b) {
   return (((uint32_t)a << 24) | (r << 16) | (g << 8) | b);
 }
 
+#ifdef WORDS_BIGENDIAN
+static void PackARGB_C(const uint8_t* a, const uint8_t* r, const uint8_t* g,
+                       const uint8_t* b, int len, uint32_t* out) {
+  int i;
+  for (i = 0; i < len; ++i) {
+    out[i] = MakeARGB32(a[4 * i], r[4 * i], g[4 * i], b[4 * i]);
+  }
+}
+#endif
+
 static void PackRGB_C(const uint8_t* r, const uint8_t* g, const uint8_t* b,
                       int len, int step, uint32_t* out) {
   int i, offset = 0;
@@ -381,6 +391,10 @@ int (*WebPDispatchAlpha)(const uint8_t*, int, int, int, uint8_t*, int);
 void (*WebPDispatchAlphaToGreen)(const uint8_t*, int, int, int, uint32_t*, int);
 int (*WebPExtractAlpha)(const uint8_t*, int, int, int, uint8_t*, int);
 void (*WebPExtractGreen)(const uint32_t* argb, uint8_t* alpha, int size);
+#ifdef WORDS_BIGENDIAN
+void (*WebPPackARGB)(const uint8_t* a, const uint8_t* r, const uint8_t* g,
+                     const uint8_t* b, int, uint32_t*);
+#endif
 void (*WebPPackRGB)(const uint8_t* r, const uint8_t* g, const uint8_t* b,
                     int len, int step, uint32_t* out);
 
@@ -405,6 +419,9 @@ WEBP_TSAN_IGNORE_FUNCTION void WebPInitAlphaProcessing(void) {
   WebPMultRow = WebPMultRow_C;
   WebPApplyAlphaMultiply4444 = ApplyAlphaMultiply_16b_C;
 
+#ifdef WORDS_BIGENDIAN
+  WebPPackARGB = PackARGB_C;
+#endif
   WebPPackRGB = PackRGB_C;
 #if !WEBP_NEON_OMIT_C_CODE
   WebPApplyAlphaMultiply = ApplyAlphaMultiply_C;
@@ -451,6 +468,9 @@ WEBP_TSAN_IGNORE_FUNCTION void WebPInitAlphaProcessing(void) {
   assert(WebPDispatchAlphaToGreen != NULL);
   assert(WebPExtractAlpha != NULL);
   assert(WebPExtractGreen != NULL);
+#ifdef WORDS_BIGENDIAN
+  assert(WebPPackARGB != NULL);
+#endif
   assert(WebPPackRGB != NULL);
   assert(WebPHasAlpha8b != NULL);
   assert(WebPHasAlpha32b != NULL);

src/dsp/alpha_processing_mips_dsp_r2.c

@@ -125,6 +125,49 @@ static void MultARGBRow_MIPSdspR2(uint32_t* const ptr, int width,
   }
 }
 
+#ifdef WORDS_BIGENDIAN
+static void PackARGB_MIPSdspR2(const uint8_t* a, const uint8_t* r,
+                               const uint8_t* g, const uint8_t* b, int len,
+                               uint32_t* out) {
+  int temp0, temp1, temp2, temp3, offset;
+  const int rest = len & 1;
+  const uint32_t* const loop_end = out + len - rest;
+  const int step = 4;
+  __asm__ volatile (
+    "xor          %[offset],   %[offset], %[offset]    \n\t"
+    "beq          %[loop_end], %[out],    0f           \n\t"
+  "2:                                                  \n\t"
+    "lbux         %[temp0],    %[offset](%[a])         \n\t"
+    "lbux         %[temp1],    %[offset](%[r])         \n\t"
+    "lbux         %[temp2],    %[offset](%[g])         \n\t"
+    "lbux         %[temp3],    %[offset](%[b])         \n\t"
+    "ins          %[temp1],    %[temp0],  16,     16   \n\t"
+    "ins          %[temp3],    %[temp2],  16,     16   \n\t"
+    "addiu        %[out],      %[out],    4            \n\t"
+    "precr.qb.ph  %[temp0],    %[temp1],  %[temp3]     \n\t"
+    "sw           %[temp0],    -4(%[out])              \n\t"
+    "addu         %[offset],   %[offset], %[step]      \n\t"
+    "bne          %[loop_end], %[out],    2b           \n\t"
+  "0:                                                  \n\t"
+    "beq          %[rest],     $zero,     1f           \n\t"
+    "lbux         %[temp0],    %[offset](%[a])         \n\t"
+    "lbux         %[temp1],    %[offset](%[r])         \n\t"
+    "lbux         %[temp2],    %[offset](%[g])         \n\t"
+    "lbux         %[temp3],    %[offset](%[b])         \n\t"
+    "ins          %[temp1],    %[temp0],  16,     16   \n\t"
+    "ins          %[temp3],    %[temp2],  16,     16   \n\t"
+    "precr.qb.ph  %[temp0],    %[temp1],  %[temp3]     \n\t"
+    "sw           %[temp0],    0(%[out])               \n\t"
+  "1:                                                  \n\t"
+    : [temp0]"=&r"(temp0), [temp1]"=&r"(temp1), [temp2]"=&r"(temp2),
+      [temp3]"=&r"(temp3), [offset]"=&r"(offset), [out]"+&r"(out)
+    : [a]"r"(a), [r]"r"(r), [g]"r"(g), [b]"r"(b), [step]"r"(step),
+      [loop_end]"r"(loop_end), [rest]"r"(rest)
+    : "memory"
+  );
+}
+#endif  // WORDS_BIGENDIAN
+
 static void PackRGB_MIPSdspR2(const uint8_t* r, const uint8_t* g,
                               const uint8_t* b, int len, int step,
                               uint32_t* out) {
@@ -172,6 +215,9 @@ extern void WebPInitAlphaProcessingMIPSdspR2(void);
 WEBP_TSAN_IGNORE_FUNCTION void WebPInitAlphaProcessingMIPSdspR2(void) {
   WebPDispatchAlpha = DispatchAlpha_MIPSdspR2;
   WebPMultARGBRow = MultARGBRow_MIPSdspR2;
+#ifdef WORDS_BIGENDIAN
+  WebPPackARGB = PackARGB_MIPSdspR2;
+#endif
   WebPPackRGB = PackRGB_MIPSdspR2;
 }
 

src/dsp/dsp.h

@@ -166,6 +166,13 @@ extern "C" {
 #define WEBP_SWAP_16BIT_CSP 0
 #endif
 
+// some endian fix (e.g.: mips-gcc doesn't define __BIG_ENDIAN__)
+#if !defined(WORDS_BIGENDIAN) && \
+    (defined(__BIG_ENDIAN__) || defined(_M_PPC) || \
+     (defined(__BYTE_ORDER__) && (__BYTE_ORDER__ == __ORDER_BIG_ENDIAN__)))
+#define WORDS_BIGENDIAN
+#endif
+
 typedef enum {
   kSSE2,
   kSSE3,
@@ -578,6 +585,13 @@ void WebPMultRow_C(uint8_t* const ptr, const uint8_t* const alpha,
                    int width, int inverse);
 void WebPMultARGBRow_C(uint32_t* const ptr, int width, int inverse);
 
+#ifdef WORDS_BIGENDIAN
+// ARGB packing function: a/r/g/b input is rgba or bgra order.
+extern void (*WebPPackARGB)(const uint8_t* a, const uint8_t* r,
+                            const uint8_t* g, const uint8_t* b, int len,
+                            uint32_t* out);
+#endif
+
 // RGB packing function. 'step' can be 3 or 4. r/g/b input is rgb or bgr order.
 extern void (*WebPPackRGB)(const uint8_t* r, const uint8_t* g, const uint8_t* b,
                            int len, int step, uint32_t* out);

src/dsp/upsampling.c

@@ -94,13 +94,37 @@ static void FUNC_NAME(const uint8_t* top_y, const uint8_t* bottom_y,           \
 
 // All variants implemented.
 #if !WEBP_NEON_OMIT_C_CODE
-UPSAMPLE_FUNC(UpsampleRgbLinePair_C,  VP8YuvToRgb,  3)
-UPSAMPLE_FUNC(UpsampleBgrLinePair_C,  VP8YuvToBgr,  3)
 UPSAMPLE_FUNC(UpsampleRgbaLinePair_C, VP8YuvToRgba, 4)
 UPSAMPLE_FUNC(UpsampleBgraLinePair_C, VP8YuvToBgra, 4)
+#if !defined(WEBP_REDUCE_CSP)
 UPSAMPLE_FUNC(UpsampleArgbLinePair_C, VP8YuvToArgb, 4)
+UPSAMPLE_FUNC(UpsampleRgbLinePair_C,  VP8YuvToRgb,  3)
+UPSAMPLE_FUNC(UpsampleBgrLinePair_C,  VP8YuvToBgr,  3)
 UPSAMPLE_FUNC(UpsampleRgba4444LinePair_C, VP8YuvToRgba4444, 2)
 UPSAMPLE_FUNC(UpsampleRgb565LinePair_C,  VP8YuvToRgb565,  2)
+#else
+static void EmptyUpsampleFunc(const uint8_t* top_y, const uint8_t* bottom_y,
+                              const uint8_t* top_u, const uint8_t* top_v,
+                              const uint8_t* cur_u, const uint8_t* cur_v,
+                              uint8_t* top_dst, uint8_t* bottom_dst, int len) {
+  (void)top_y;
+  (void)bottom_y;
+  (void)top_u;
+  (void)top_v;
+  (void)cur_u;
+  (void)cur_v;
+  (void)top_dst;
+  (void)bottom_dst;
+  (void)len;
+  assert(0);   // COLORSPACE SUPPORT NOT COMPILED
+}
+#define UpsampleArgbLinePair_C EmptyUpsampleFunc
+#define UpsampleRgbLinePair_C EmptyUpsampleFunc
+#define UpsampleBgrLinePair_C EmptyUpsampleFunc
+#define UpsampleRgba4444LinePair_C EmptyUpsampleFunc
+#define UpsampleRgb565LinePair_C EmptyUpsampleFunc
+#endif   // WEBP_REDUCE_CSP
+
 #endif
 
 #undef LOAD_UV
@@ -162,13 +186,30 @@ void FUNC_NAME(const uint8_t* y, const uint8_t* u, const uint8_t* v,           \
   for (i = 0; i < len; ++i) FUNC(y[i], u[i], v[i], &dst[i * (XSTEP)]);         \
 }
 
-YUV444_FUNC(WebPYuv444ToRgb_C,      VP8YuvToRgb,  3)
-YUV444_FUNC(WebPYuv444ToBgr_C,      VP8YuvToBgr,  3)
 YUV444_FUNC(WebPYuv444ToRgba_C,     VP8YuvToRgba, 4)
 YUV444_FUNC(WebPYuv444ToBgra_C,     VP8YuvToBgra, 4)
+#if !defined(WEBP_REDUCE_CSP)
+YUV444_FUNC(WebPYuv444ToRgb_C,      VP8YuvToRgb,  3)
+YUV444_FUNC(WebPYuv444ToBgr_C,      VP8YuvToBgr,  3)
 YUV444_FUNC(WebPYuv444ToArgb_C,     VP8YuvToArgb, 4)
 YUV444_FUNC(WebPYuv444ToRgba4444_C, VP8YuvToRgba4444, 2)
 YUV444_FUNC(WebPYuv444ToRgb565_C,   VP8YuvToRgb565, 2)
+#else
+static void EmptyYuv444Func(const uint8_t* y,
+                            const uint8_t* u, const uint8_t* v,
+                            uint8_t* dst, int len) {
+  (void)y;
+  (void)u;
+  (void)v;
+  (void)dst;
+  (void)len;
+}
+#define WebPYuv444ToRgb_C EmptyYuv444Func
+#define WebPYuv444ToBgr_C EmptyYuv444Func
+#define WebPYuv444ToArgb_C EmptyYuv444Func
+#define WebPYuv444ToRgba4444_C EmptyYuv444Func
+#define WebPYuv444ToRgb565_C EmptyYuv444Func
+#endif   // WEBP_REDUCE_CSP
 
 #undef YUV444_FUNC
 
@@ -183,10 +224,10 @@ static volatile VP8CPUInfo upsampling_last_cpuinfo_used1 =
 WEBP_TSAN_IGNORE_FUNCTION void WebPInitYUV444Converters(void) {
   if (upsampling_last_cpuinfo_used1 == VP8GetCPUInfo) return;
 
-  WebPYUV444Converters[MODE_RGB]       = WebPYuv444ToRgb_C;
   WebPYUV444Converters[MODE_RGBA]      = WebPYuv444ToRgba_C;
-  WebPYUV444Converters[MODE_BGR]       = WebPYuv444ToBgr_C;
   WebPYUV444Converters[MODE_BGRA]      = WebPYuv444ToBgra_C;
+  WebPYUV444Converters[MODE_RGB]       = WebPYuv444ToRgb_C;
+  WebPYUV444Converters[MODE_BGR]       = WebPYuv444ToBgr_C;
   WebPYUV444Converters[MODE_ARGB]      = WebPYuv444ToArgb_C;
   WebPYUV444Converters[MODE_RGBA_4444] = WebPYuv444ToRgba4444_C;
   WebPYUV444Converters[MODE_RGB_565]   = WebPYuv444ToRgb565_C;
@@ -226,15 +267,15 @@ WEBP_TSAN_IGNORE_FUNCTION void WebPInitUpsamplers(void) {
 
 #ifdef FANCY_UPSAMPLING
 #if !WEBP_NEON_OMIT_C_CODE
-  WebPUpsamplers[MODE_RGB]       = UpsampleRgbLinePair_C;
   WebPUpsamplers[MODE_RGBA]      = UpsampleRgbaLinePair_C;
-  WebPUpsamplers[MODE_BGR]       = UpsampleBgrLinePair_C;
   WebPUpsamplers[MODE_BGRA]      = UpsampleBgraLinePair_C;
+  WebPUpsamplers[MODE_rgbA]      = UpsampleRgbaLinePair_C;
+  WebPUpsamplers[MODE_bgrA]      = UpsampleBgraLinePair_C;
+  WebPUpsamplers[MODE_RGB]       = UpsampleRgbLinePair_C;
+  WebPUpsamplers[MODE_BGR]       = UpsampleBgrLinePair_C;
   WebPUpsamplers[MODE_ARGB]      = UpsampleArgbLinePair_C;
   WebPUpsamplers[MODE_RGBA_4444] = UpsampleRgba4444LinePair_C;
   WebPUpsamplers[MODE_RGB_565]   = UpsampleRgb565LinePair_C;
-  WebPUpsamplers[MODE_rgbA]      = UpsampleRgbaLinePair_C;
-  WebPUpsamplers[MODE_bgrA]      = UpsampleBgraLinePair_C;
   WebPUpsamplers[MODE_Argb]      = UpsampleArgbLinePair_C;
   WebPUpsamplers[MODE_rgbA_4444] = UpsampleRgba4444LinePair_C;
 #endif
@@ -265,15 +306,15 @@ WEBP_TSAN_IGNORE_FUNCTION void WebPInitUpsamplers(void) {
   }
 #endif
 
-  assert(WebPUpsamplers[MODE_RGB] != NULL);
   assert(WebPUpsamplers[MODE_RGBA] != NULL);
-  assert(WebPUpsamplers[MODE_BGR] != NULL);
   assert(WebPUpsamplers[MODE_BGRA] != NULL);
+  assert(WebPUpsamplers[MODE_rgbA] != NULL);
+  assert(WebPUpsamplers[MODE_bgrA] != NULL);
+  assert(WebPUpsamplers[MODE_RGB] != NULL);
+  assert(WebPUpsamplers[MODE_BGR] != NULL);
   assert(WebPUpsamplers[MODE_ARGB] != NULL);
   assert(WebPUpsamplers[MODE_RGBA_4444] != NULL);
   assert(WebPUpsamplers[MODE_RGB_565] != NULL);
-  assert(WebPUpsamplers[MODE_rgbA] != NULL);
-  assert(WebPUpsamplers[MODE_bgrA] != NULL);
   assert(WebPUpsamplers[MODE_Argb] != NULL);
   assert(WebPUpsamplers[MODE_rgbA_4444] != NULL);
 

src/dsp/upsampling_mips_dsp_r2.c

@@ -46,6 +46,7 @@
     );                                                                         \
   } while (0)
 
+#if !defined(WEBP_REDUCE_CSP)
 static WEBP_INLINE void YuvToRgb(int y, int u, int v, uint8_t* const rgb) {
   int r, g, b;
   YUV_TO_RGB(y, u, v, r, g, b);
@@ -91,10 +92,12 @@ static WEBP_INLINE void YuvToRgba4444(int y, int u, int v,
 #endif
    }
 }
+#endif   // WEBP_REDUCE_CSP
 
 //-----------------------------------------------------------------------------
 // Alpha handling variants
 
+#if !defined(WEBP_REDUCE_CSP)
 static WEBP_INLINE void YuvToArgb(uint8_t y, uint8_t u, uint8_t v,
                                   uint8_t* const argb) {
   int r, g, b;
@@ -104,6 +107,7 @@ static WEBP_INLINE void YuvToArgb(uint8_t y, uint8_t u, uint8_t v,
   argb[2] = g;
   argb[3] = b;
 }
+#endif   // WEBP_REDUCE_CSP
 static WEBP_INLINE void YuvToBgra(uint8_t y, uint8_t u, uint8_t v,
                                   uint8_t* const bgra) {
   int r, g, b;
@@ -197,13 +201,15 @@ static void FUNC_NAME(const uint8_t* top_y, const uint8_t* bottom_y,           \
 }
 
 // All variants implemented.
-UPSAMPLE_FUNC(UpsampleRgbLinePair,      YuvToRgb,      3)
-UPSAMPLE_FUNC(UpsampleBgrLinePair,      YuvToBgr,      3)
 UPSAMPLE_FUNC(UpsampleRgbaLinePair,     YuvToRgba,     4)
 UPSAMPLE_FUNC(UpsampleBgraLinePair,     YuvToBgra,     4)
+#if !defined(WEBP_REDUCE_CSP)
+UPSAMPLE_FUNC(UpsampleRgbLinePair,      YuvToRgb,      3)
+UPSAMPLE_FUNC(UpsampleBgrLinePair,      YuvToBgr,      3)
 UPSAMPLE_FUNC(UpsampleArgbLinePair,     YuvToArgb,     4)
 UPSAMPLE_FUNC(UpsampleRgba4444LinePair, YuvToRgba4444, 2)
 UPSAMPLE_FUNC(UpsampleRgb565LinePair,   YuvToRgb565,   2)
+#endif   // WEBP_REDUCE_CSP
 
 #undef LOAD_UV
 #undef UPSAMPLE_FUNC
@@ -214,17 +220,19 @@ UPSAMPLE_FUNC(UpsampleRgb565LinePair,   YuvToRgb565,   2)
 extern void WebPInitUpsamplersMIPSdspR2(void);
 
 WEBP_TSAN_IGNORE_FUNCTION void WebPInitUpsamplersMIPSdspR2(void) {
-  WebPUpsamplers[MODE_RGB]       = UpsampleRgbLinePair;
   WebPUpsamplers[MODE_RGBA]      = UpsampleRgbaLinePair;
-  WebPUpsamplers[MODE_BGR]       = UpsampleBgrLinePair;
   WebPUpsamplers[MODE_BGRA]      = UpsampleBgraLinePair;
+  WebPUpsamplers[MODE_rgbA]      = UpsampleRgbaLinePair;
+  WebPUpsamplers[MODE_bgrA]      = UpsampleBgraLinePair;
+#if !defined(WEBP_REDUCE_CSP)
+  WebPUpsamplers[MODE_RGB]       = UpsampleRgbLinePair;
+  WebPUpsamplers[MODE_BGR]       = UpsampleBgrLinePair;
   WebPUpsamplers[MODE_ARGB]      = UpsampleArgbLinePair;
   WebPUpsamplers[MODE_RGBA_4444] = UpsampleRgba4444LinePair;
   WebPUpsamplers[MODE_RGB_565]   = UpsampleRgb565LinePair;
-  WebPUpsamplers[MODE_rgbA]      = UpsampleRgbaLinePair;
-  WebPUpsamplers[MODE_bgrA]      = UpsampleBgraLinePair;
   WebPUpsamplers[MODE_Argb]      = UpsampleArgbLinePair;
   WebPUpsamplers[MODE_rgbA_4444] = UpsampleRgba4444LinePair;
+#endif   // WEBP_REDUCE_CSP
 }
 
 #endif  // FANCY_UPSAMPLING
@@ -239,13 +247,15 @@ static void FUNC_NAME(const uint8_t* y, const uint8_t* u, const uint8_t* v,    \
   for (i = 0; i < len; ++i) FUNC(y[i], u[i], v[i], &dst[i * XSTEP]);           \
 }
 
-YUV444_FUNC(Yuv444ToRgb,      YuvToRgb,      3)
-YUV444_FUNC(Yuv444ToBgr,      YuvToBgr,      3)
 YUV444_FUNC(Yuv444ToRgba,     YuvToRgba,     4)
 YUV444_FUNC(Yuv444ToBgra,     YuvToBgra,     4)
+#if !defined(WEBP_REDUCE_CSP)
+YUV444_FUNC(Yuv444ToRgb,      YuvToRgb,      3)
+YUV444_FUNC(Yuv444ToBgr,      YuvToBgr,      3)
 YUV444_FUNC(Yuv444ToArgb,     YuvToArgb,     4)
 YUV444_FUNC(Yuv444ToRgba4444, YuvToRgba4444, 2)
 YUV444_FUNC(Yuv444ToRgb565,   YuvToRgb565,   2)
+#endif   // WEBP_REDUCE_CSP
 
 #undef YUV444_FUNC
 
@@ -255,17 +265,19 @@ YUV444_FUNC(Yuv444ToRgb565,   YuvToRgb565,   2)
 extern void WebPInitYUV444ConvertersMIPSdspR2(void);
 
 WEBP_TSAN_IGNORE_FUNCTION void WebPInitYUV444ConvertersMIPSdspR2(void) {
-  WebPYUV444Converters[MODE_RGB]       = Yuv444ToRgb;
   WebPYUV444Converters[MODE_RGBA]      = Yuv444ToRgba;
-  WebPYUV444Converters[MODE_BGR]       = Yuv444ToBgr;
   WebPYUV444Converters[MODE_BGRA]      = Yuv444ToBgra;
+  WebPYUV444Converters[MODE_rgbA]      = Yuv444ToRgba;
+  WebPYUV444Converters[MODE_bgrA]      = Yuv444ToBgra;
+#if !defined(WEBP_REDUCE_CSP)
+  WebPYUV444Converters[MODE_RGB]       = Yuv444ToRgb;
+  WebPYUV444Converters[MODE_BGR]       = Yuv444ToBgr;
   WebPYUV444Converters[MODE_ARGB]      = Yuv444ToArgb;
   WebPYUV444Converters[MODE_RGBA_4444] = Yuv444ToRgba4444;
   WebPYUV444Converters[MODE_RGB_565]   = Yuv444ToRgb565;
-  WebPYUV444Converters[MODE_rgbA]      = Yuv444ToRgba;
-  WebPYUV444Converters[MODE_bgrA]      = Yuv444ToBgra;
   WebPYUV444Converters[MODE_Argb]      = Yuv444ToArgb;
   WebPYUV444Converters[MODE_rgbA_4444] = Yuv444ToRgba4444;
+#endif   // WEBP_REDUCE_CSP
 }
 
 #else  // !WEBP_USE_MIPS_DSP_R2

src/dsp/upsampling_msa.c

@@ -640,13 +640,15 @@ static void FUNC_NAME(const uint8_t* top_y, const uint8_t* bot_y,        \
   }                                                                      \
 }
 
-UPSAMPLE_FUNC(UpsampleRgbLinePair,      YuvToRgb,      3)
-UPSAMPLE_FUNC(UpsampleBgrLinePair,      YuvToBgr,      3)
 UPSAMPLE_FUNC(UpsampleRgbaLinePair,     YuvToRgba,     4)
 UPSAMPLE_FUNC(UpsampleBgraLinePair,     YuvToBgra,     4)
+#if !defined(WEBP_REDUCE_CSP)
+UPSAMPLE_FUNC(UpsampleRgbLinePair,      YuvToRgb,      3)
+UPSAMPLE_FUNC(UpsampleBgrLinePair,      YuvToBgr,      3)
 UPSAMPLE_FUNC(UpsampleArgbLinePair,     YuvToArgb,     4)
 UPSAMPLE_FUNC(UpsampleRgba4444LinePair, YuvToRgba4444, 2)
 UPSAMPLE_FUNC(UpsampleRgb565LinePair,   YuvToRgb565,   2)
+#endif   // WEBP_REDUCE_CSP
 
 //------------------------------------------------------------------------------
 // Entry point
@@ -656,17 +658,19 @@ extern WebPUpsampleLinePairFunc WebPUpsamplers[/* MODE_LAST */];
 extern void WebPInitUpsamplersMSA(void);
 
 WEBP_TSAN_IGNORE_FUNCTION void WebPInitUpsamplersMSA(void) {
-  WebPUpsamplers[MODE_RGB]       = UpsampleRgbLinePair;
   WebPUpsamplers[MODE_RGBA]      = UpsampleRgbaLinePair;
-  WebPUpsamplers[MODE_BGR]       = UpsampleBgrLinePair;
   WebPUpsamplers[MODE_BGRA]      = UpsampleBgraLinePair;
-  WebPUpsamplers[MODE_ARGB]      = UpsampleArgbLinePair;
   WebPUpsamplers[MODE_rgbA]      = UpsampleRgbaLinePair;
   WebPUpsamplers[MODE_bgrA]      = UpsampleBgraLinePair;
+#if !defined(WEBP_REDUCE_CSP)
+  WebPUpsamplers[MODE_RGB]       = UpsampleRgbLinePair;
+  WebPUpsamplers[MODE_BGR]       = UpsampleBgrLinePair;
+  WebPUpsamplers[MODE_ARGB]      = UpsampleArgbLinePair;
   WebPUpsamplers[MODE_Argb]      = UpsampleArgbLinePair;
   WebPUpsamplers[MODE_RGB_565]   = UpsampleRgb565LinePair;
   WebPUpsamplers[MODE_RGBA_4444] = UpsampleRgba4444LinePair;
   WebPUpsamplers[MODE_rgbA_4444] = UpsampleRgba4444LinePair;
+#endif   // WEBP_REDUCE_CSP
 }
 
 #endif  // FANCY_UPSAMPLING

src/dsp/upsampling_neon.c

@@ -243,13 +243,15 @@ static void FUNC_NAME(const uint8_t *top_y, const uint8_t *bottom_y,    \
 }
 
 // NEON variants of the fancy upsampler.
-NEON_UPSAMPLE_FUNC(UpsampleRgbLinePair_NEON,  Rgb,  3)
-NEON_UPSAMPLE_FUNC(UpsampleBgrLinePair_NEON,  Bgr,  3)
 NEON_UPSAMPLE_FUNC(UpsampleRgbaLinePair_NEON, Rgba, 4)
 NEON_UPSAMPLE_FUNC(UpsampleBgraLinePair_NEON, Bgra, 4)
+#if !defined(WEBP_REDUCE_CSP)
+NEON_UPSAMPLE_FUNC(UpsampleRgbLinePair_NEON,  Rgb,  3)
+NEON_UPSAMPLE_FUNC(UpsampleBgrLinePair_NEON,  Bgr,  3)
 NEON_UPSAMPLE_FUNC(UpsampleArgbLinePair_NEON, Argb, 4)
 NEON_UPSAMPLE_FUNC(UpsampleRgba4444LinePair_NEON, Rgba4444, 2)
 NEON_UPSAMPLE_FUNC(UpsampleRgb565LinePair_NEON, Rgb565, 2)
+#endif   // WEBP_REDUCE_CSP
 
 //------------------------------------------------------------------------------
 // Entry point
@@ -259,17 +261,19 @@ extern WebPUpsampleLinePairFunc WebPUpsamplers[/* MODE_LAST */];
 extern void WebPInitUpsamplersNEON(void);
 
 WEBP_TSAN_IGNORE_FUNCTION void WebPInitUpsamplersNEON(void) {
-  WebPUpsamplers[MODE_RGB]  = UpsampleRgbLinePair_NEON;
   WebPUpsamplers[MODE_RGBA] = UpsampleRgbaLinePair_NEON;
-  WebPUpsamplers[MODE_BGR]  = UpsampleBgrLinePair_NEON;
   WebPUpsamplers[MODE_BGRA] = UpsampleBgraLinePair_NEON;
-  WebPUpsamplers[MODE_ARGB] = UpsampleArgbLinePair_NEON;
   WebPUpsamplers[MODE_rgbA] = UpsampleRgbaLinePair_NEON;
   WebPUpsamplers[MODE_bgrA] = UpsampleBgraLinePair_NEON;
+#if !defined(WEBP_REDUCE_CSP)
+  WebPUpsamplers[MODE_RGB]  = UpsampleRgbLinePair_NEON;
+  WebPUpsamplers[MODE_BGR]  = UpsampleBgrLinePair_NEON;
+  WebPUpsamplers[MODE_ARGB] = UpsampleArgbLinePair_NEON;
   WebPUpsamplers[MODE_Argb] = UpsampleArgbLinePair_NEON;
   WebPUpsamplers[MODE_RGB_565] = UpsampleRgb565LinePair_NEON;
   WebPUpsamplers[MODE_RGBA_4444] = UpsampleRgba4444LinePair_NEON;
   WebPUpsamplers[MODE_rgbA_4444] = UpsampleRgba4444LinePair_NEON;
+#endif   // WEBP_REDUCE_CSP
 }
 
 #endif  // FANCY_UPSAMPLING

src/dsp/upsampling_sse2.c

@@ -169,13 +169,16 @@ static void FUNC_NAME(const uint8_t* top_y, const uint8_t* bottom_y,           \
 }
 
 // SSE2 variants of the fancy upsampler.
-SSE2_UPSAMPLE_FUNC(UpsampleRgbLinePair_SSE2,  VP8YuvToRgb,  3)
-SSE2_UPSAMPLE_FUNC(UpsampleBgrLinePair_SSE2,  VP8YuvToBgr,  3)
 SSE2_UPSAMPLE_FUNC(UpsampleRgbaLinePair_SSE2, VP8YuvToRgba, 4)
 SSE2_UPSAMPLE_FUNC(UpsampleBgraLinePair_SSE2, VP8YuvToBgra, 4)
+
+#if !defined(WEBP_REDUCE_CSP)
+SSE2_UPSAMPLE_FUNC(UpsampleRgbLinePair_SSE2,  VP8YuvToRgb,  3)
+SSE2_UPSAMPLE_FUNC(UpsampleBgrLinePair_SSE2,  VP8YuvToBgr,  3)
 SSE2_UPSAMPLE_FUNC(UpsampleArgbLinePair_SSE2, VP8YuvToArgb, 4)
 SSE2_UPSAMPLE_FUNC(UpsampleRgba4444LinePair_SSE2, VP8YuvToRgba4444, 2)
 SSE2_UPSAMPLE_FUNC(UpsampleRgb565LinePair_SSE2, VP8YuvToRgb565, 2)
+#endif   // WEBP_REDUCE_CSP
 
 #undef GET_M
 #undef PACK_AND_STORE
@@ -193,17 +196,19 @@ extern WebPUpsampleLinePairFunc WebPUpsamplers[/* MODE_LAST */];
 extern void WebPInitUpsamplersSSE2(void);
 
 WEBP_TSAN_IGNORE_FUNCTION void WebPInitUpsamplersSSE2(void) {
-  WebPUpsamplers[MODE_RGB]  = UpsampleRgbLinePair_SSE2;
   WebPUpsamplers[MODE_RGBA] = UpsampleRgbaLinePair_SSE2;
-  WebPUpsamplers[MODE_BGR]  = UpsampleBgrLinePair_SSE2;
   WebPUpsamplers[MODE_BGRA] = UpsampleBgraLinePair_SSE2;
-  WebPUpsamplers[MODE_ARGB] = UpsampleArgbLinePair_SSE2;
   WebPUpsamplers[MODE_rgbA] = UpsampleRgbaLinePair_SSE2;
   WebPUpsamplers[MODE_bgrA] = UpsampleBgraLinePair_SSE2;
+#if !defined(WEBP_REDUCE_CSP)
+  WebPUpsamplers[MODE_RGB]  = UpsampleRgbLinePair_SSE2;
+  WebPUpsamplers[MODE_BGR]  = UpsampleBgrLinePair_SSE2;
+  WebPUpsamplers[MODE_ARGB] = UpsampleArgbLinePair_SSE2;
   WebPUpsamplers[MODE_Argb] = UpsampleArgbLinePair_SSE2;
   WebPUpsamplers[MODE_RGB_565] = UpsampleRgb565LinePair_SSE2;
   WebPUpsamplers[MODE_RGBA_4444] = UpsampleRgba4444LinePair_SSE2;
   WebPUpsamplers[MODE_rgbA_4444] = UpsampleRgba4444LinePair_SSE2;
+#endif   // WEBP_REDUCE_CSP
 }
 
 #endif  // FANCY_UPSAMPLING
@@ -230,25 +235,29 @@ static void FUNC_NAME(const uint8_t* y, const uint8_t* u, const uint8_t* v,    \
 
 YUV444_FUNC(Yuv444ToRgba_SSE2, VP8YuvToRgba32_SSE2, WebPYuv444ToRgba_C, 4);
 YUV444_FUNC(Yuv444ToBgra_SSE2, VP8YuvToBgra32_SSE2, WebPYuv444ToBgra_C, 4);
+#if !defined(WEBP_REDUCE_CSP)
 YUV444_FUNC(Yuv444ToRgb_SSE2, VP8YuvToRgb32_SSE2, WebPYuv444ToRgb_C, 3);
 YUV444_FUNC(Yuv444ToBgr_SSE2, VP8YuvToBgr32_SSE2, WebPYuv444ToBgr_C, 3);
 YUV444_FUNC(Yuv444ToArgb_SSE2, VP8YuvToArgb32_SSE2, WebPYuv444ToArgb_C, 4)
 YUV444_FUNC(Yuv444ToRgba4444_SSE2, VP8YuvToRgba444432_SSE2, \
             WebPYuv444ToRgba4444_C, 2)
 YUV444_FUNC(Yuv444ToRgb565_SSE2, VP8YuvToRgb56532_SSE2, WebPYuv444ToRgb565_C, 2)
+#endif   // WEBP_REDUCE_CSP
 
 WEBP_TSAN_IGNORE_FUNCTION void WebPInitYUV444ConvertersSSE2(void) {
   WebPYUV444Converters[MODE_RGBA]      = Yuv444ToRgba_SSE2;
   WebPYUV444Converters[MODE_BGRA]      = Yuv444ToBgra_SSE2;
+  WebPYUV444Converters[MODE_rgbA]      = Yuv444ToRgba_SSE2;
+  WebPYUV444Converters[MODE_bgrA]      = Yuv444ToBgra_SSE2;
+#if !defined(WEBP_REDUCE_CSP)
   WebPYUV444Converters[MODE_RGB]       = Yuv444ToRgb_SSE2;
   WebPYUV444Converters[MODE_BGR]       = Yuv444ToBgr_SSE2;
   WebPYUV444Converters[MODE_ARGB]      = Yuv444ToArgb_SSE2;
   WebPYUV444Converters[MODE_RGBA_4444] = Yuv444ToRgba4444_SSE2;
   WebPYUV444Converters[MODE_RGB_565]   = Yuv444ToRgb565_SSE2;
-  WebPYUV444Converters[MODE_rgbA]      = Yuv444ToRgba_SSE2;
-  WebPYUV444Converters[MODE_bgrA]      = Yuv444ToBgra_SSE2;
   WebPYUV444Converters[MODE_Argb]      = Yuv444ToArgb_SSE2;
   WebPYUV444Converters[MODE_rgbA_4444] = Yuv444ToRgba4444_SSE2;
+#endif   // WEBP_REDUCE_CSP
 }
 
 #else

src/enc/picture_csp_enc.c

@@ -28,11 +28,11 @@
 // If defined, use table to compute x / alpha.
 #define USE_INVERSE_ALPHA_TABLE
 
-static const union {
-  uint32_t argb;
-  uint8_t  bytes[4];
-} test_endian = { 0xff000000u };
-#define ALPHA_IS_LAST (test_endian.bytes[3] == 0xff)
+#ifdef WORDS_BIGENDIAN
+#define ALPHA_OFFSET 0   // uint32_t 0xff000000 is 0xff,00,00,00 in memory
+#else
+#define ALPHA_OFFSET 3   // uint32_t 0xff000000 is 0x00,00,00,ff in memory
+#endif
 
 //------------------------------------------------------------------------------
 // Detection of non-trivial transparency
@@ -61,7 +61,7 @@ int WebPPictureHasTransparency(const WebPPicture* picture) {
     return CheckNonOpaque(picture->a, picture->width, picture->height,
                           1, picture->a_stride);
   } else {
-    const int alpha_offset = ALPHA_IS_LAST ? 3 : 0;
+    const int alpha_offset = ALPHA_OFFSET;
     return CheckNonOpaque((const uint8_t*)picture->argb + alpha_offset,
                           picture->width, picture->height,
                           4, picture->argb_stride * sizeof(*picture->argb));
@@ -990,10 +990,10 @@ static int PictureARGBToYUVA(WebPPicture* picture, WebPEncCSP colorspace,
     return WebPEncodingSetError(picture, VP8_ENC_ERROR_INVALID_CONFIGURATION);
   } else {
     const uint8_t* const argb = (const uint8_t*)picture->argb;
-    const uint8_t* const r = ALPHA_IS_LAST ? argb + 2 : argb + 1;
-    const uint8_t* const g = ALPHA_IS_LAST ? argb + 1 : argb + 2;
-    const uint8_t* const b = ALPHA_IS_LAST ? argb + 0 : argb + 3;
-    const uint8_t* const a = ALPHA_IS_LAST ? argb + 3 : argb + 0;
+    const uint8_t* const a = argb + (0 ^ ALPHA_OFFSET);
+    const uint8_t* const r = argb + (1 ^ ALPHA_OFFSET);
+    const uint8_t* const g = argb + (2 ^ ALPHA_OFFSET);
+    const uint8_t* const b = argb + (3 ^ ALPHA_OFFSET);
 
     picture->colorspace = WEBP_YUV420;
     return ImportYUVAFromRGBA(r, g, b, a, 4, 4 * picture->argb_stride,
@@ -1044,7 +1044,8 @@ int WebPPictureYUVAToARGB(WebPPicture* picture) {
     const int argb_stride = 4 * picture->argb_stride;
     uint8_t* dst = (uint8_t*)picture->argb;
     const uint8_t *cur_u = picture->u, *cur_v = picture->v, *cur_y = picture->y;
-    WebPUpsampleLinePairFunc upsample = WebPGetLinePairConverter(ALPHA_IS_LAST);
+    WebPUpsampleLinePairFunc upsample =
+        WebPGetLinePairConverter(ALPHA_OFFSET > 0);
 
     // First row, with replicated top samples.
     upsample(cur_y, NULL, cur_u, cur_v, cur_u, cur_v, dst, NULL, width);
@@ -1087,6 +1088,7 @@ static int Import(WebPPicture* const picture,
                   const uint8_t* rgb, int rgb_stride,
                   int step, int swap_rb, int import_alpha) {
   int y;
+  // swap_rb -> b,g,r,a , !swap_rb -> r,g,b,a
   const uint8_t* r_ptr = rgb + (swap_rb ? 2 : 0);
   const uint8_t* g_ptr = rgb + 1;
   const uint8_t* b_ptr = rgb + (swap_rb ? 0 : 2);
@@ -1104,16 +1106,25 @@ static int Import(WebPPicture* const picture,
   WebPInitAlphaProcessing();
 
   if (import_alpha) {
+    // dst[] byte order is {a,r,g,b} for big-endian, {b,g,r,a} for little endian
     uint32_t* dst = picture->argb;
-    const int do_copy =
-        (!swap_rb && !ALPHA_IS_LAST) || (swap_rb && ALPHA_IS_LAST);
+    const int do_copy = (ALPHA_OFFSET == 3) && swap_rb;
     assert(step == 4);
     for (y = 0; y < height; ++y) {
       if (do_copy) {
         memcpy(dst, rgb, width * 4);
       } else {
+#ifdef WORDS_BIGENDIAN
+        // BGRA or RGBA input order.
+        const uint8_t* a_ptr = rgb + 3;
+        WebPPackARGB(a_ptr, r_ptr, g_ptr, b_ptr, width, dst);
+        r_ptr += rgb_stride;
+        g_ptr += rgb_stride;
+        b_ptr += rgb_stride;
+#else
         // RGBA input order. Need to swap R and B.
         VP8LConvertBGRAToRGBA((const uint32_t*)rgb, width, (uint8_t*)dst);
+#endif
       }
       rgb += rgb_stride;
       dst += picture->argb_stride;
@@ -1134,12 +1145,7 @@ static int Import(WebPPicture* const picture,
 
 // Public API
 
-int WebPPictureImportRGB(WebPPicture* picture,
-                         const uint8_t* rgb, int rgb_stride) {
-  return (picture != NULL && rgb != NULL)
-             ? Import(picture, rgb, rgb_stride, 3, 0, 0)
-             : 0;
-}
+#if !defined(WEBP_REDUCE_CSP)
 
 int WebPPictureImportBGR(WebPPicture* picture,
                          const uint8_t* rgb, int rgb_stride) {
@@ -1148,13 +1154,6 @@ int WebPPictureImportBGR(WebPPicture* picture,
              : 0;
 }
 
-int WebPPictureImportRGBA(WebPPicture* picture,
-                          const uint8_t* rgba, int rgba_stride) {
-  return (picture != NULL && rgba != NULL)
-             ? Import(picture, rgba, rgba_stride, 4, 0, 1)
-             : 0;
-}
-
 int WebPPictureImportBGRA(WebPPicture* picture,
                           const uint8_t* rgba, int rgba_stride) {
   return (picture != NULL && rgba != NULL)
@@ -1162,12 +1161,6 @@ int WebPPictureImportBGRA(WebPPicture* picture,
              : 0;
 }
 
-int WebPPictureImportRGBX(WebPPicture* picture,
-                          const uint8_t* rgba, int rgba_stride) {
-  return (picture != NULL && rgba != NULL)
-             ? Import(picture, rgba, rgba_stride, 4, 0, 0)
-             : 0;
-}
 
 int WebPPictureImportBGRX(WebPPicture* picture,
                           const uint8_t* rgba, int rgba_stride) {
@@ -1176,4 +1169,27 @@ int WebPPictureImportBGRX(WebPPicture* picture,
              : 0;
 }
 
+#endif   // WEBP_REDUCE_CSP
+
+int WebPPictureImportRGB(WebPPicture* picture,
+                         const uint8_t* rgb, int rgb_stride) {
+  return (picture != NULL && rgb != NULL)
+             ? Import(picture, rgb, rgb_stride, 3, 0, 0)
+             : 0;
+}
+
+int WebPPictureImportRGBA(WebPPicture* picture,
+                          const uint8_t* rgba, int rgba_stride) {
+  return (picture != NULL && rgba != NULL)
+             ? Import(picture, rgba, rgba_stride, 4, 0, 1)
+             : 0;
+}
+
+int WebPPictureImportRGBX(WebPPicture* picture,
+                          const uint8_t* rgba, int rgba_stride) {
+  return (picture != NULL && rgba != NULL)
+             ? Import(picture, rgba, rgba_stride, 4, 0, 0)
+             : 0;
+}
+
 //------------------------------------------------------------------------------

src/enc/picture_enc.c

@@ -270,9 +270,11 @@ size_t NAME(const uint8_t* in, int w, int h, int bps, float q,          \
 }
 
 ENCODE_FUNC(WebPEncodeRGB, WebPPictureImportRGB)
-ENCODE_FUNC(WebPEncodeBGR, WebPPictureImportBGR)
 ENCODE_FUNC(WebPEncodeRGBA, WebPPictureImportRGBA)
+#if !defined(WEBP_REDUCE_CSP)
+ENCODE_FUNC(WebPEncodeBGR, WebPPictureImportBGR)
 ENCODE_FUNC(WebPEncodeBGRA, WebPPictureImportBGRA)
+#endif  // WEBP_REDUCE_CSP
 
 #undef ENCODE_FUNC
 
@@ -283,9 +285,11 @@ size_t NAME(const uint8_t* in, int w, int h, int bps, uint8_t** out) {       \
 }
 
 LOSSLESS_ENCODE_FUNC(WebPEncodeLosslessRGB, WebPPictureImportRGB)
-LOSSLESS_ENCODE_FUNC(WebPEncodeLosslessBGR, WebPPictureImportBGR)
 LOSSLESS_ENCODE_FUNC(WebPEncodeLosslessRGBA, WebPPictureImportRGBA)
+#if !defined(WEBP_REDUCE_CSP)
+LOSSLESS_ENCODE_FUNC(WebPEncodeLosslessBGR, WebPPictureImportBGR)
 LOSSLESS_ENCODE_FUNC(WebPEncodeLosslessBGRA, WebPPictureImportBGRA)
+#endif  // WEBP_REDUCE_CSP
 
 #undef LOSSLESS_ENCODE_FUNC
 

src/utils/endian_inl_utils.h

@@ -19,13 +19,6 @@
 #include "src/dsp/dsp.h"
 #include "src/webp/types.h"
 
-// some endian fix (e.g.: mips-gcc doesn't define __BIG_ENDIAN__)
-#if !defined(WORDS_BIGENDIAN) && \
-    (defined(__BIG_ENDIAN__) || defined(_M_PPC) || \
-     (defined(__BYTE_ORDER__) && (__BYTE_ORDER__ == __ORDER_BIG_ENDIAN__)))
-#define WORDS_BIGENDIAN
-#endif
-
 #if defined(WORDS_BIGENDIAN)
 #define HToLE32 BSwap32
 #define HToLE16 BSwap16