Added some usage text to the driver.

1 files changed, 56 insertions, 3 deletions

diff --git a/unpremultiply.c b/unpremultiply.c
index 1919a15..9beebd0 100644
--- a/unpremultiply.c
+++ b/unpremultiply.c
@@ -185,6 +185,57 @@ main(int argc, char **argv)
     double elapsed_ms;
     int i;
 
+    if (1 == argc) {
+	printf(
+	    "usage: unpremultiply [aliased] <data> [solid] [method] [verify]\n"
+	    "\n"
+	    "Prints out the time in ms taken to unpremultiply a data buffer\n"
+	    "using the given method. Optionally verifies the result.\n"
+	    "If 'aliased' is given on the command line then the source\n"
+	    "and destination buffers are the same.\n"
+	    "\n"
+	    "Available methods:\n"
+	    "  div         The reference unpremultiplier using three\n"
+	    "              divisions per pixel.\n"
+	    "  lut         A division table based unpremultiplier.\n"
+	    "  inv         A reciprocal table based unpremultiplier.\n"
+	    "  lutb        'lut' with special casings for constant and\n"
+	    "              solid pixel runs.\n"
+	    "  invb        'inv' with special casings for constant and\n"
+	    "              solid pixel runs.\n"
+	    "  sse2        An AMD64/SSE2 unpremultiplier.\n"
+	    "\n"
+	    "Methods suitable as timing references:\n"
+	    "  copy        memcpy(dst, src, pixels);\n"
+	    "  write       memset(dst, 0, pixels);\n"
+	    "  read        Sums the source pixels.\n"
+	    "\n"
+	    "Available data generators:\n"
+	    "  random      Random pixel values.\n"
+	    "  gradient    A smoothly varying alpha gradient.\n"
+	    "              Usually gives the same results as\n"
+	    "              the random data.\n"
+	    "  clear       All zeros.\n"
+	    "\n"
+	    "If the 'solid' suffix is given after the data, then the\n"
+	    "data is forced to be completely opaque by setting the alpha\n"
+	    "of every pixel to 255.  If 'verify' is given on the command\n"
+	    "line then the final result is verified against a reference\n"
+	    "unpremultiplier. Verification using aliased data buffers\n"
+	    "works correctly when loops=1.\n"
+	    "\n"
+	    "Environment variables affecting the run:\n"
+	    "  loops   [500]     Number of times to do the\n"
+	    "                    unpremultiplication\n"
+	    "  pixels  [2097152] Size of the data buffers in pixels.\n"
+	    "  offset  [0]       Byte offset added to the data buffer\n"
+	    "                    pointers to misalign the pixel data.\n"
+	    "                    This option is liable to segfault the\n"
+	    "                    program.\n"
+	    );
+	return 1;
+    }
+
     udst.u8 = calloc(n*4+offset, 1) + offset;
     usrc.u8 = calloc(n*4+offset, 1) + offset;
 #define dst udst.u32
@@ -290,7 +341,7 @@ main(int argc, char **argv)
     printf("%f\n", elapsed_ms);
 
     if (verify) {
-	size_t i;
+	unsigned i;
 	int maxdiff = 0;
 	for (i=0; i<n; i++) {
 	    uint32_t x = dst[i];
@@ -300,8 +351,10 @@ main(int argc, char **argv)
 		int diff = (x & 255) - (y & 255);
 		if (diff < 0) diff = -diff;
 		if (diff > maxdiff) {
-		    printf("src:%08x dst:%08x ref:%08x diff:%d, i:%ld.%d\n",
-			   src[i], dst[i], ref[i], diff, i,j);
+		    printf("maxdiff now %d: "
+			   "src[%u]=%08x dst[%u]=%08x ref[%u]=%08x "
+			   "(component %d)\n",
+			   diff, i, src[i], i, dst[i], i, ref[i], j);
 		    maxdiff = diff;
 		}
 		x >>= 8;