Optimize the loop for writing varints in ProtoBuf #1294
Conversation
It has no performance impact on small (~1-2 bytes) varints, but increases throughput of large (7-9 bytes) varints by up to 25%
| } | ||
|
|
||
| private fun encodeVarint32SlowPath(value: Int) { | ||
| private inline fun encodeVarint(value: Long, length: Int) { |
There was a problem hiding this comment.
It is unlikely that someone now is using a 32-bit CPU but these calculations can be slow on them.
There was a problem hiding this comment.
Could you please elaborate on what you mean here?
There was a problem hiding this comment.
current now have long type, in old encodeVarint32SlowPath it was int. So perhaps for 32-bit CPU and 32-bit varints operations and, or, ushr will be slower and total speed will degrade.
There was a problem hiding this comment.
We generally do not care about 32-bit platforms
| // Fast-path: unrolled loop for single byte | ||
| ensureCapacity(5) |
There was a problem hiding this comment.
Why now for long we ensure dynamic length but here always 5?
There was a problem hiding this comment.
Because 32-bit varints are used on a regular basis for tags, it's always beneficial to allocate all the space.
Also, because tags are small, one-byte fast-path is used and we do not have to calculate varint length at all for such values
|
|
||
| fun encodeVarint64(value: Long) { | ||
| val length = varIntLength(value) | ||
| ensureCapacity(length + 1) |
There was a problem hiding this comment.
Suggest do +1 inside varIntLength or VAR_INT_LENGTHS init block to match the name.
There was a problem hiding this comment.
Then it will slightly complicate the loop (have to do decrement), let's leave it as is
| } | ||
|
|
||
| private fun encodeVarint32SlowPath(value: Int) { | ||
| private inline fun encodeVarint(value: Long, length: Int) { |
There was a problem hiding this comment.
current now have long type, in old encodeVarint32SlowPath it was int. So perhaps for 32-bit CPU and 32-bit varints operations and, or, ushr will be slower and total speed will degrade.
It has no performance impact on small (~1-2 bytes) varints, but increases throughput of large (7-9 bytes) varints by up to 25%