Merge "Make libsharpyuv self-contained by removing dependency on cpu.c" into main

sharpyuv/sharpyuv.c

@@ -62,9 +62,8 @@ static uint32_t kLinearToGammaTabS[kGammaTabSize + 2];
 #define GAMMA_TO_LINEAR_BITS 14
 static uint32_t kGammaToLinearTabS[MAX_Y_T + 1];   // size scales with Y_FIX
 static volatile int kGammaTablesSOk = 0;
-static void InitGammaTablesS(void);
 
-WEBP_DSP_INIT_FUNC(InitGammaTablesS) {
+static void InitGammaTablesS(void) {
   assert(2 * GAMMA_TO_LINEAR_BITS < 32);  // we use uint32_t intermediate values
   if (!kGammaTablesSOk) {
     int v;
@@ -180,6 +179,7 @@ static void UpdateChroma(const fixed_y_t* src1, const fixed_y_t* src2,
 
 static void StoreGray(const fixed_y_t* rgb, fixed_y_t* y, int w) {
   int i;
+  assert(w > 0);
   for (i = 0; i < w; ++i) {
     y[i] = RGBToGray(rgb[0 * w + i], rgb[1 * w + i], rgb[2 * w + i]);
   }
@@ -348,9 +348,6 @@ static int DoSharpArgbToYuv(const uint8_t* r_ptr, const uint8_t* g_ptr,
     goto End;
   }
 
-  InitGammaTablesS();
-  InitSharpYuv();
-
   // Import RGB samples to W/RGB representation.
   for (j = 0; j < height; j += 2) {
     const int is_last_row = (j == height - 1);
@@ -438,6 +435,26 @@ static int DoSharpArgbToYuv(const uint8_t* r_ptr, const uint8_t* g_ptr,
  }
 #undef SAFE_ALLOC
 
+// Hidden exported init function.
+// By default SharpYuvConvert calls it with NULL. If needed, users can declare
+// it as extern and call it with a VP8CPUInfo function.
+extern void SharpYuvInit(VP8CPUInfo cpu_info_func);
+void SharpYuvInit(VP8CPUInfo cpu_info_func) {
+  static volatile VP8CPUInfo sharpyuv_last_cpuinfo_used =
+      (VP8CPUInfo)&sharpyuv_last_cpuinfo_used;
+  const int initialized =
+      (sharpyuv_last_cpuinfo_used != (VP8CPUInfo)&sharpyuv_last_cpuinfo_used);
+  if (cpu_info_func == NULL && initialized) return;
+  if (sharpyuv_last_cpuinfo_used == cpu_info_func) return;
+
+  SharpYuvInitDsp(cpu_info_func);
+  if (!initialized) {
+    InitGammaTablesS();
+  }
+
+  sharpyuv_last_cpuinfo_used = cpu_info_func;
+}
+
 // In YUV_FIX fixed point precision.
 static const SharpYuvConversionMatrix kWebpYuvMatrix = {
   {16839,  33059,  6420,  16 << 16},
@@ -459,6 +476,7 @@ int SharpYuvConvert(const uint8_t* r_ptr, const uint8_t* g_ptr,
       height < kMinDimensionIterativeConversion) {
     return 0;
   }
+  SharpYuvInit(NULL);
   return DoSharpArgbToYuv(r_ptr, g_ptr, b_ptr, step, rgb_stride, dst_y,
                           dst_stride_y, dst_u, dst_stride_u, dst_v,
                           dst_stride_v, width, height, yuv_matrix);

sharpyuv/sharpyuv_dsp.c

@@ -74,7 +74,8 @@ void (*SharpYUVFilterRow)(const int16_t* A, const int16_t* B, int len,
 extern void InitSharpYUVSSE2(void);
 extern void InitSharpYUVNEON(void);
 
-WEBP_DSP_INIT_FUNC(InitSharpYuv) {
+void SharpYuvInitDsp(VP8CPUInfo cpu_info_func) {
+  (void)cpu_info_func;
 
 #if !WEBP_NEON_OMIT_C_CODE
   SharpYUVUpdateY = SharpYUVUpdateY_C;
@@ -83,14 +84,13 @@ WEBP_DSP_INIT_FUNC(InitSharpYuv) {
 #endif
 
 #if defined(WEBP_HAVE_SSE2)
-  if (VP8GetCPUInfo != NULL && VP8GetCPUInfo(kSSE2)) {
+  if (cpu_info_func == NULL || cpu_info_func(kSSE2)) {
     InitSharpYUVSSE2();
   }
 #endif  // WEBP_HAVE_SSE2
 
 #if defined(WEBP_HAVE_NEON)
-  if (WEBP_NEON_OMIT_C_CODE ||
+  if (WEBP_NEON_OMIT_C_CODE || cpu_info_func == NULL || cpu_info_func(kNEON)) {
-      (VP8GetCPUInfo != NULL && VP8GetCPUInfo(kNEON))) {
     InitSharpYUVNEON();
   }
 #endif  // WEBP_HAVE_NEON

sharpyuv/sharpyuv_dsp.h

@@ -23,6 +23,6 @@ extern void (*SharpYUVUpdateRGB)(const int16_t* src, const int16_t* ref,
 extern void (*SharpYUVFilterRow)(const int16_t* A, const int16_t* B, int len,
                                  const uint16_t* best_y, uint16_t* out);
 
-void InitSharpYuv(void);
+void SharpYuvInitDsp(VP8CPUInfo cpu_info_func);
 
 #endif  // WEBP_SHARPYUV_DSP_H_

src/dsp/cpu.c

@@ -11,7 +11,7 @@
 //
 // Author: Christian Duvivier (cduvivier@google.com)
 
-#include "src/dsp/dsp.h"
+#include "src/dsp/cpu.h"
 
 #if defined(WEBP_HAVE_NEON_RTCD)
 #include <stdio.h>

src/enc/picture_csp_enc.c

@@ -22,6 +22,11 @@
 #include "src/dsp/dsp.h"
 #include "src/dsp/lossless.h"
 #include "src/dsp/yuv.h"
+#include "src/dsp/cpu.h"
+
+#if defined(WEBP_USE_THREAD) && !defined(_WIN32)
+#include <pthread.h>
+#endif
 
 // Uncomment to disable gamma-compression during RGB->U/V averaging
 #define USE_GAMMA_COMPRESSION
@@ -174,6 +179,21 @@ typedef uint16_t fixed_y_t;   // unsigned type with extra SFIX precision for W
 //------------------------------------------------------------------------------
 // Main function
 
+extern void SharpYuvInit(VP8CPUInfo cpu_info_func);
+
+static void SafeInitSharpYuv(void) {
+#if defined(WEBP_USE_THREAD) && !defined(_WIN32)
+  static pthread_mutex_t initsharpyuv_lock = PTHREAD_MUTEX_INITIALIZER;
+  if (pthread_mutex_lock(&initsharpyuv_lock)) return;
+#endif
+
+  SharpYuvInit(VP8GetCPUInfo);
+
+#if defined(WEBP_USE_THREAD) && !defined(_WIN32)
+  (void)pthread_mutex_unlock(&initsharpyuv_lock);
+#endif
+}
+
 static int PreprocessARGB(const uint8_t* r_ptr,
                           const uint8_t* g_ptr,
                           const uint8_t* b_ptr,
@@ -495,6 +515,7 @@ static int ImportYUVAFromRGBA(const uint8_t* r_ptr,
   }
 
   if (use_iterative_conversion) {
+    SafeInitSharpYuv();
     if (!PreprocessARGB(r_ptr, g_ptr, b_ptr, step, rgb_stride, picture)) {
       return 0;
     }