[FFmpeg-devel] Small CRC32 standard tables

[James Almer]

Currently, standard tables like AV_CRC_32_IEEE and such are being generated (or 
provided in case the user compiles with hardcoded tables) with only 257 elements.
We're missing a considerable boost in performance by not making them with a size 
of 1024 elements instead.

By applying something like this

---
diff --git a/libavutil/crc.c b/libavutil/crc.c
index 9ee5efe..bec0d80 100644
--- a/libavutil/crc.c
+++ b/libavutil/crc.c
@@ -211,6 +211,11 @@ static const AVCRC av_crc_table[AV_CRC_MAX][257] = {
     },
 };
 #else
+#if CONFIG_SMALL
+#define AV_CRC_TABLE_SIZE 257
+#else
+#define AV_CRC_TABLE_SIZE 1024
+#endif
 static struct {
     uint8_t  le;
     uint8_t  bits;
@@ -222,7 +227,7 @@ static struct {
     [AV_CRC_32_IEEE]    = { 0, 32, 0x04C11DB7 },
     [AV_CRC_32_IEEE_LE] = { 1, 32, 0xEDB88320 },
 };
-static AVCRC av_crc_table[AV_CRC_MAX][257];
+static AVCRC av_crc_table[AV_CRC_MAX][AV_CRC_TABLE_SIZE];
 #endif
 
 int av_crc_init(AVCRC *ctx, int le, int bits, uint32_t poly, int ctx_size)

---

I'm getting 2x to 3x faster calculations when data.
A ~260mb file is processed in ~0.9 seconds using ffhash git head, but in ~0.4 
seconds if i apply the above change.
Similarly, using a modified version of Nicolas George's hash_bench tool to run
CRC32 instead of md5 or sha, i get ~2.5 seconds with the patch above, compared to 
~7 seconds without it.

Is there a reason why we're providing small tables with av_crc_get_table()?
As things are right now, using av_crc_init() with any of the standard polynomials 
and a context of size 1024 will pretty much always be the better option.

Regards.