Add ARGBExtractAlpha function

BUG=libyuv:572
R=fbarchard@google.com

Review URL: https://codereview.chromium.org/1995293002 .

include/libyuv/planar_functions.h

@@ -288,6 +288,12 @@ int ARGBCopyAlpha(const uint8* src_argb, int src_stride_argb,
                   uint8* dst_argb, int dst_stride_argb,
                   int width, int height);
 
+// Extract the alpha channel from ARGB.
+LIBYUV_API
+int ARGBExtractAlpha(const uint8* src_argb, int src_stride_argb,
+                     uint8* dst_a, int dst_stride_a,
+                     int width, int height);
+
 // Copy Y channel to Alpha of ARGB.
 LIBYUV_API
 int ARGBCopyYToAlpha(const uint8* src_y, int src_stride_y,

include/libyuv/row.h

@@ -104,6 +104,7 @@ extern "C" {
 #define HAS_ARGBTOUVROW_SSSE3
 #define HAS_ARGBTOYJROW_SSSE3
 #define HAS_ARGBTOYROW_SSSE3
+#define HAS_ARGBEXTRACTALPHAROW_SSE2
 #define HAS_BGRATOUVROW_SSSE3
 #define HAS_BGRATOYROW_SSSE3
 #define HAS_COPYROW_ERMS
@@ -291,6 +292,7 @@ extern "C" {
 #define HAS_ARGBTOUVROW_NEON
 #define HAS_ARGBTOYJROW_NEON
 #define HAS_ARGBTOYROW_NEON
+#define HAS_ARGBEXTRACTALPHAROW_NEON
 #define HAS_BGRATOUVROW_NEON
 #define HAS_BGRATOYROW_NEON
 #define HAS_COPYROW_NEON
@@ -877,6 +879,14 @@ void ARGBCopyAlphaRow_Any_SSE2(const uint8* src_argb, uint8* dst_argb,
 void ARGBCopyAlphaRow_Any_AVX2(const uint8* src_argb, uint8* dst_argb,
                                int width);
 
+void ARGBExtractAlphaRow_C(const uint8* src_argb, uint8* dst_a, int width);
+void ARGBExtractAlphaRow_SSE2(const uint8* src_argb, uint8* dst_a, int width);
+void ARGBExtractAlphaRow_NEON(const uint8* src_argb, uint8* dst_a, int width);
+void ARGBExtractAlphaRow_Any_SSE2(const uint8* src_argb, uint8* dst_a,
+                                  int width);
+void ARGBExtractAlphaRow_Any_NEON(const uint8* src_argb, uint8* dst_a,
+                                  int width);
+
 void ARGBCopyYToAlphaRow_C(const uint8* src_y, uint8* dst_argb, int width);
 void ARGBCopyYToAlphaRow_SSE2(const uint8* src_y, uint8* dst_argb, int width);
 void ARGBCopyYToAlphaRow_AVX2(const uint8* src_y, uint8* dst_argb, int width);

source/planar_functions.cc

@@ -2374,6 +2374,49 @@ int ARGBCopyAlpha(const uint8* src_argb, int src_stride_argb,
   return 0;
 }
 
+// Extract just the alpha channel from ARGB.
+LIBYUV_API
+int ARGBExtractAlpha(const uint8* src_argb, int src_stride,
+                     uint8* dst_a, int dst_stride,
+                     int width, int height) {
+  if (!src_argb || !dst_a || width <= 0 || height == 0) {
+    return -1;
+  }
+  // Negative height means invert the image.
+  if (height < 0) {
+    height = -height;
+    src_argb += (height - 1) * src_stride;
+    src_stride = -src_stride;
+  }
+  // Coalesce rows.
+  if (src_stride == width * 4 && dst_stride == width) {
+    width *= height;
+    height = 1;
+    src_stride = dst_stride = 0;
+  }
+  void (*ARGBExtractAlphaRow)(const uint8 *src_argb, uint8 *dst_a, int width) =
+      ARGBExtractAlphaRow_C;
+#if defined(HAS_ARGBEXTRACTALPHAROW_SSE2)
+  if (TestCpuFlag(kCpuHasSSE2)) {
+    ARGBExtractAlphaRow = IS_ALIGNED(width, 8) ? ARGBExtractAlphaRow_SSE2
+                                               : ARGBExtractAlphaRow_Any_SSE2;
+  }
+#endif
+#if defined(HAS_ARGBEXTRACTALPHAROW_NEON)
+  if (TestCpuFlag(kCpuHasNEON)) {
+    ARGBExtractAlphaRow = IS_ALIGNED(width, 8) ? ARGBExtractAlphaRow_NEON
+                                               : ARGBExtractAlphaRow_Any_NEON;
+  }
+#endif
+
+  for (int y = 0; y < height; ++y) {
+    ARGBExtractAlphaRow(src_argb, dst_a, width);
+    src_argb += src_stride;
+    dst_a += dst_stride;
+  }
+  return 0;
+}
+
 // Copy a planar Y channel to the alpha channel of a destination ARGB image.
 LIBYUV_API
 int ARGBCopyYToAlpha(const uint8* src_y, int src_stride_y,

source/row_any.cc

@@ -466,6 +466,12 @@ ANY11(ARGBUnattenuateRow_Any_AVX2, ARGBUnattenuateRow_AVX2, 0, 4, 4, 7)
 #ifdef HAS_ARGBATTENUATEROW_NEON
 ANY11(ARGBAttenuateRow_Any_NEON, ARGBAttenuateRow_NEON, 0, 4, 4, 7)
 #endif
+#ifdef HAS_ARGBEXTRACTALPHAROW_SSE2
+ANY11(ARGBExtractAlphaRow_Any_SSE2, ARGBExtractAlphaRow_SSE2, 0, 4, 1, 7)
+#endif
+#ifdef HAS_ARGBEXTRACTALPHAROW_NEON
+ANY11(ARGBExtractAlphaRow_Any_NEON, ARGBExtractAlphaRow_NEON, 0, 4, 1, 7)
+#endif
 #undef ANY11
 
 // Any 1 to 1 blended.  Destination is read, modify, write.

source/row_common.cc

@@ -2381,6 +2381,19 @@ void ARGBCopyAlphaRow_C(const uint8* src, uint8* dst, int width) {
   }
 }
 
+void ARGBExtractAlphaRow_C(const uint8* src_argb, uint8* dst_a, int width) {
+  int i;
+  for (i = 0; i < width - 1; i += 2) {
+    dst_a[0] = src_argb[3];
+    dst_a[1] = src_argb[7];
+    dst_a += 2;
+    src_argb += 8;
+  }
+  if (width & 1) {
+    dst_a[0] = src_argb[3];
+  }
+}
+
 void ARGBCopyYToAlphaRow_C(const uint8* src, uint8* dst, int width) {
   int i;
   for (i = 0; i < width - 1; i += 2) {

source/row_gcc.cc

@@ -2936,6 +2936,33 @@ void ARGBCopyAlphaRow_AVX2(const uint8* src, uint8* dst, int width) {
 }
 #endif  // HAS_ARGBCOPYALPHAROW_AVX2
 
+#ifdef HAS_ARGBEXTRACTALPHAROW_SSE2
+// width in pixels
+void ARGBExtractAlphaRow_SSE2(const uint8* src_argb, uint8* dst_a, int width) {
+ asm volatile (
+    LABELALIGN
+  "1:                                          \n"
+    "movdqu    " MEMACCESS(0) ", %%xmm0        \n"
+    "movdqu    " MEMACCESS2(0x10, 0) ", %%xmm1 \n"
+    "lea       " MEMLEA(0x20, 0) ", %0         \n"
+    "psrld     $0x18, %%xmm0                   \n"
+    "psrld     $0x18, %%xmm1                   \n"
+    "packssdw  %%xmm1, %%xmm0                  \n"
+    "packuswb  %%xmm0, %%xmm0                  \n"
+    "movq      %%xmm0," MEMACCESS(1) "         \n"
+    "lea       " MEMLEA(0x8, 1) ", %1          \n"
+    "sub       $0x8, %2                        \n"
+    "jg        1b                              \n"
+  : "+r"(src_argb),  // %0
+    "+r"(dst_a),     // %1
+    "+rm"(width)     // %2
+  :
+  : "memory", "cc"
+    , "xmm0", "xmm1"
+  );
+}
+#endif  // HAS_ARGBEXTRACTALPHAROW_SSE2
+
 #ifdef HAS_ARGBCOPYYTOALPHAROW_SSE2
 // width in pixels
 void ARGBCopyYToAlphaRow_SSE2(const uint8* src, uint8* dst, int width) {

source/row_neon.cc

@@ -1298,6 +1298,23 @@ void ARGBToYRow_NEON(const uint8* src_argb, uint8* dst_y, int width) {
   );
 }
 
+void ARGBExtractAlphaRow_NEON(const uint8* src_argb, uint8* dst_a, int width) {
+  asm volatile (
+  "1:                                          \n"
+    MEMACCESS(0)
+    "vld4.8     {d0, d1, d2, d3}, [%0]!        \n"  // load row 8 pixels
+    "subs       %2, %2, #8                     \n"  // 8 processed per loop
+    MEMACCESS(1)
+    "vst1.8     {d3}, [%1]!                    \n"  // store 8 A's.
+    "bgt       1b                              \n"
+  : "+r"(src_argb),   // %0
+    "+r"(dst_a),      // %1
+    "+r"(width)       // %2
+  :
+  : "cc", "memory", "d0", "d1", "d2", "d3"  // Clobber List
+  );
+}
+
 void ARGBToYJRow_NEON(const uint8* src_argb, uint8* dst_y, int width) {
   asm volatile (
     "vmov.u8    d24, #15                       \n"  // B * 0.11400 coefficient

source/row_neon64.cc

@@ -1399,6 +1399,25 @@ void ARGBToYRow_NEON(const uint8* src_argb, uint8* dst_y, int width) {
 }
 #endif  // HAS_ARGBTOYROW_NEON
 
+#ifdef HAS_ARGBEXTRACTALPHAROW_NEON
+void ARGBExtractAlphaRow_NEON(const uint8* src_argb, uint8* dst_a, int width) {
+  asm volatile (
+  "1:                                          \n"
+    MEMACCESS(0)
+    "ld4        {v0.8b,v1.8b,v2.8b,v3.8b}, [%0], #32 \n"  // load row 8 pixels
+    "subs       %w2, %w2, #8                   \n"  // 8 processed per loop
+    MEMACCESS(1)
+    "st1        {v3.8b}, [%1], #8              \n"  // store 8 A's.
+    "b.gt       1b                             \n"
+  : "+r"(src_argb),   // %0
+    "+r"(dst_a),      // %1
+    "+r"(width)       // %2
+  :
+  : "cc", "memory", "v0", "v1", "v2", "v3"  // Clobber List
+  );
+}
+#endif  // HAS_ARGBEXTRACTALPHAROW_NEON
+
 #ifdef HAS_ARGBTOYJROW_NEON
 void ARGBToYJRow_NEON(const uint8* src_argb, uint8* dst_y, int width) {
   asm volatile (

source/row_win.cc

@@ -3532,6 +3532,33 @@ void ARGBCopyAlphaRow_AVX2(const uint8* src, uint8* dst, int width) {
 }
 #endif  // HAS_ARGBCOPYALPHAROW_AVX2
 
+#ifdef HAS_ARGBEXTRACTALPHAROW_SSE2
+// width in pixels
+__declspec(naked)
+void ARGBExtractAlphaRow_SSE2(const uint8* src_argb, uint8* dst_a, int width) {
+  __asm {
+    mov        eax, [esp + 4]   // src_argb
+    mov        edx, [esp + 8]   // dst_a
+    mov        ecx, [esp + 12]  // width
+
+  extractloop:
+    movdqu     xmm0, [eax]
+    movdqu     xmm1, [eax + 16]
+    lea        eax, [eax + 32]
+    psrld      xmm0, 24
+    psrld      xmm1, 24
+    packssdw   xmm0, xmm1
+    packuswb   xmm0, xmm0
+    movq       qword ptr [edx], xmm0
+    lea        edx, [edx + 8]
+    sub        ecx, 8
+    jg         extractloop
+
+    ret
+  }
+}
+#endif  // HAS_ARGBEXTRACTALPHAROW_SSE2
+
 #ifdef HAS_ARGBCOPYYTOALPHAROW_SSE2
 // width in pixels
 __declspec(naked)

unit_test/planar_test.cc

@@ -2390,6 +2390,36 @@ TEST_F(LibYUVPlanarTest, TestARGBCopyAlpha) {
   free_aligned_buffer_64(orig_pixels);
 }
 
+TEST_F(LibYUVPlanarTest, TestARGBExtractAlpha) {
+  const int kPixels = benchmark_width_ * benchmark_height_;
+  align_buffer_64(src_pixels, kPixels * 4);
+  align_buffer_64(dst_pixels_opt, kPixels);
+  align_buffer_64(dst_pixels_c, kPixels);
+
+  MemRandomize(src_pixels, kPixels * 4);
+  MemRandomize(dst_pixels_opt, kPixels);
+  memcpy(dst_pixels_c, dst_pixels_opt, kPixels);
+
+  MaskCpuFlags(disable_cpu_flags_);
+  ARGBExtractAlpha(src_pixels, benchmark_width_ * 4,
+                   dst_pixels_c, benchmark_width_,
+                   benchmark_width_, benchmark_height_);
+  MaskCpuFlags(benchmark_cpu_info_);
+
+  for (int i = 0; i < benchmark_iterations_; ++i) {
+    ARGBExtractAlpha(src_pixels, benchmark_width_ * 4,
+                     dst_pixels_opt, benchmark_width_,
+                     benchmark_width_, benchmark_height_);
+  }
+  for (int i = 0; i < kPixels; ++i) {
+    EXPECT_EQ(dst_pixels_c[i], dst_pixels_opt[i]);
+  }
+
+  free_aligned_buffer_64(dst_pixels_c);
+  free_aligned_buffer_64(dst_pixels_opt);
+  free_aligned_buffer_64(src_pixels);
+}
+
 TEST_F(LibYUVPlanarTest, TestARGBCopyYToAlpha) {
   const int kPixels = benchmark_width_ * benchmark_height_;
   align_buffer_64(orig_pixels, kPixels);