On Thursday 09 June 2016 05:45 PM, Richard Biener wrote:
On Thu, Jun 9, 2016 at 10:54 AM, Richard Biener
<richard.guent...@gmail.com> wrote:
On Tue, Jun 7, 2016 at 3:59 PM, Sameera Deshpande
<sameera.deshpa...@imgtec.com> wrote:
Hi Richard,
This is with reference to our discussion at GNU Tools Cauldron 2015 regarding my talk
titled "Improving the effectiveness and generality of GCC auto-vectorization."
Further to our prototype implementation of the concept, we have started implementing this
concept in GCC.
We are following incremental model to add language support in our front-end,
and corresponding back-end (for auto-vectorizer) will be added for feature
completion.
Looking at the complexity and scale of the project, we have divided this
project into subtasks listed below, for ease of implementation, testing and
review.
0. Add new pass to perform autovectorization using unified representation -
Current GCC framework does not give complete overview of the loop to be
vectorized : it either breaks the loop across body, or across iterations.
Because of which these data structures can not be reused for our approach which
gathers all the information of loop body at one place using primitive permute
operations. Hence, define new data structures and populate them.
1. Add support for vectorization of LOAD/STORE instructions
a. Create permute order tree for the loop with LOAD and STORE instructions
for single or multi-dimensional arrays, aggregates within nested loops.
b. Basic transformation phase to generate vectorized code for the
primitive reorder tree generated at stage 1a using tree tiling algorithm. This
phase handles code generation for SCATTER, GATHER, stridded memory accesses
etc. along with permute instruction generation.
2. Implementation of k-arity promotion/reduction : The permute nodes within
primitive reorder tree generated from input program can have any arity.
However, the target can support maximum of arity = 2 in most of the cases.
Hence, we need to promote or reduce the arity of permute order tree to enable
successful tree tiling.
3. Vector size reduction : Depending upon the vector size for target, reduce
vector size per statement and adjust the loop count for vectorized loop
accordingly.
4. Support simple arithmetic operations :
a. Add support for analyzing statements with simple arithmetic operations
like +, -, *, / for vectorization, and create primitive reorder tree with
compute_op.
b. Generate vector code for primitive reorder tree generated at stage 4a
using tree tiling algorithm - here support for complex patterns like
multiply-add should be checked and appropriate instruction to be generated.
5. Support reduction operation :
a. Add support for reduction operation analysis and primitive reorder tree
generation. The reduction operation needs special handling, as the finish
statement should COLLAPSE the temporary reduction vector TEMP_VAR into original
reduction variable.
b. The code generation for primitive reorder tree does not need any
handling - as reduction tree is same as tree generated in 4a, with only
difference that in 4a, the destination is MEMREF (because of STORE operation)
and for reduction it is TEMP_VAR. At this stage, generate code for COLLAPSE
node in finish statements.
6. Support other vectorizable statements like complex arithmetic operations,
bitwise operations, type conversions etc.
a. Add support for analysis and primitive reorder tree generation.
b. Vector code generation.
7. Cost effective tree tiling algorithm : Till now, the tree tiling is
happening without considering cost of computation. However, there can be
multiple target instructions covering the tree - hence, instead of picking
first matched largest instruction cover, select the instruction cover based on
cost of instruction given in .md for the target.
8. Optimizations on created primitive reorder tree : This stage is open ended,
and depending upon perf analysis, the scope of it can be defined.
The current patch I have attached herewith handles stage 0 and 1a : Adds new
pass to perform autovectorization using unified representation, defines new
data structures to cater to this requirement and creates primitive reorder tree
for LOAD/STORE instructions within the loop.
The whole loop is represented using the ITER_NODE, which have information about
- The preparatory statements for vectorization to be executed before entering
the loop (like initialization of vectors, prepping for reduction operations,
peeling etc.)
- Vectorizable loop body represented as PRIMOP_TREE (primitive reordering tree)
- Final statements (For peeling, variable loop bound, COLLAPSE operation for
reduction etc.)
- Other loop attributes (loop bound, peeling needed, dependences, etc.)
Memory accesses within a loop have definite repetitive pattern which can be
captured using primitive permute operators which can be used to determine
desired permute order for the vector computations. The PRIMOP_TREE is AST
which records all computations and permutations required to store destination
vector into continuous memory at the end of all iterations of the loop. It can
have INTERLEAVE, CONCAT, EXTRACT, SPLIT, ITER or any compute operation as
intermediate node. Leaf nodes can either be memory reference, constant or
vector of loop invariants. Depending upon the operation, PRIMOP_TREE holds
appropriate information about the statement within the loop which is necessary
for vectorization.
At this stage, these data structures are populated by gathering all the
information of the loop, statements within the loop and correlation of the
statements within the loop. Moreover the loop body is analyzed to check if
vectorization of each statement is possible. One has to note however that this
analysis phase will give worst-case estimate of instruction selection, as it
checks if specific named pattern is defined in .md for the target. It not
necessarily give optimal cover which is aim of the transformation phase using
tree tiling algorithm - and can be invoked only once the loop body is
represented using primitive reoder tree.
At this stage, the focus is to create permute order tree for the loop with LOAD
and STORE instructions only. The code we intend to compile is of the form
FOR(i = 0; i < N; i + +)
{
stmt 1 : D[k ∗ i + d 1 ] =S 1 [k ∗ i + c 11 ]
stmt 2 : D[k ∗ i + d 2 ] =S 1 [k ∗ i + c 21 ]
...
stmt k : D[k ∗ i + d k ] =S 1 [k ∗ i + c k 1 ]
}
Here we are assuming that any data reference can be represented using base + k
* index + offset (The data structure struct data_reference from GCC is used
currently for this purpose). If not, the address is normalized to convert to
such representation.
We are looking forward to your suggestions and insight in this regard for
better execution of this project.
I will look at the patch in detail this afternoon and will write up
some comments.
Ok, so here we go.
Hi Richard,
Thanks for your detailed review. Please find my comments inlined.
I see you copy quite some code from the existing vectorizer - rather than
doing this and adding a "new" pass I'd make the flag_tree_loop_vectorize_unified
flag guard code inside the existing vectorizer - thus share the pass.
I agree with you that I have copied lot of code from current implementation of GCC, and some data structures seem redundant as they create same
information as that is already available. However, as per our discussion at cauldron, I tried to generate ptrees after all the analysis phases were
done, using the information generated there. However, because of overwhelming and scattered information in various statements, loops and their
respective info nodes, it did not yield much. Moreover, it was seen that many functions or parts of the functions were not very useful, and some stmt
or loop related information needed different way of computation for our scheme than GCC's framework. So, instead of tweaking GCC's codebase, and
corrupting the information, thereby making it unusable if our optimization fails to vectorize by default vectorizer, we created new pass for the same.
I agree, that currently it looks like we are reusing most of the checks and conditions for vectorization as in GCC - as for first phase, our aim is to
meet performance of GCC before adding different optimizations. However, we plan to increase the scope further, thereby need to change the checks and
data generated accordingly.
e.g.: Peeling information won't be generated till transformation phase is started, where it will be added in the ITER_node depending upon alignment
requirements, ITER_count and VEC_SIZE of target.
Or, scatter/gather information is not needed as tree tiling for vector
load/store has different tiles for those instructions if target supports them.
Also, the way reduction operations are to be implemented in this scheme makes
all the categorization of reduction operation in current GCC redundant.
However, if you still think it is good idea to reuse same data structures and functions, I have older patch available, which I can clean up and update
to add updated ptree generation part in it, and share it.
Similarly I'd re-use loop_vinfo and simply add fields to it for the
new vectorizer
so you can dispatch to existing vectorizer functions "easily". Otherwise
it's going to be hard to keep things in-sync. Likewise for your stmt_attr
and the existing stmt_vec_info.
The main reason behind creating new ITER_node instead of reusing loop_vinfo is to capture whole loop as single entity, thereby allowing optimizations
on nested loops as well. The ITER_node, when implemented completely, can handle vectorization of multiple nested loops, as similar to all permute
nodes, even ITER_node can be distributed over each element in ITER_node.stmts - and can be propagated across compute_tree in permute order tree. So,
for future expansion, we are creating ITER_node with copy of some of loop_vinfo fields, and do not compute the fields which are not needed for this
scheme.
Similar is the case with stmt_vinfo - the information gathered in stmt_vinfo is mostly usable for SLP optimizations, which we cannot use as is. So,
instead of generating redundant information, or alter generated information, we chose to create new data structure.
Why isn't the existing data_reference structure good enough and you need
to invent a struct mem_ref? Even if that might be leaner adding new concepts
makes the code harder to understand. You can always use dr->aux to
add auxiliar data you need (like the existing vectorizer does).
We can use data_reference structure as is, however, each of it's component has very specific meaning associated with it - whereas the memref has only
3 components which are important - stride, offset - which is less than stride, and remaining base - for address of first element. So, again we thought
instead of overriding current semantics of components of data_ref, we created new data structure.
It helps if you follow the GCC coding-conventions from the start - a lot
of the new functions miss comments.
Sorry about that. I will tidy up the patch.
I realize the patch is probably still a "hack" and nowhere a final step 0/1a,
but for example you copied the iteration scheme over vector sizes. I hope
the new vectorizer can do without that and instead decide on the vector
size as part of the cost effective tiling algorithm.
I didn't get this point clearly. However, for my understanding, is the objection with use of ITER_COUNT instead of VEC_SIZE for each statement?
Because, if that is the case, there is basic difference between current GCC's implementation and this scheme - in case of GCC, we always look for the
standard pattern name to be matched for scalar instruction - for which VEC_TYPE becomes crucial. Whereas, this scheme represents each statement within
the loop as a vector with element of type SCALAR_TYPE and initial_vec_size = ITER_COUNT (as those many instances of the statement will be executed).
Then, depending upon the VEC_SIZE for target, VEC_SIZE reduction will be applied to it, on top of which tiling algorithm functions. So, for tiling, we
will have to use VEC_SIZE. However, for all other optimizations and transformations, we use each node of permute order tree to have vec_size =
ITER_COUNT to allow free movement of permute-nodes across compute-nodes, and various optimizations on them.
As the main feature of the patch is creation of the ptree (no code generation
seems to be done?) the biggest missing part is dumping the ptree in
some nice form (like in DOT format and/or to the dump file).
Yes, the ptree dump was put on back burner because I was focusing on the functionality. However, I will share the patch for dumping ptree in DOT
format shortly.
You do seem to check for target capabilities at ptree build time (looking at
vectorizable_store). I think that is a mistake as we'd ideally have the same
ptree when considering different vector sizes (or even generic vectors).
Target capabilities will differ between vector sizes.
That's right. However as long as ITER_COUNT > VEC_SIZE of target, there is at least a cover available for the ptree that we are creating. Hence, this
check is just to check if certain instruction can be represented anyhow on target architecture. This will actually be handled at transform phase after
vec_size reduction is performed. (I have kept the loop checking for different vector types in vect_analyze_loop_with_prim_tree() for now as I am still
not very clear at what point in transformation phase will it play any role... Once, the end-to-end solution for it is developed for load/store chains,
there will be some clarity - and the loop can be moved there or eliminated completely.)
That's all for now.
Last but not least - at the Cauldron I suggested to incrementally rewrite
the existing vectorizer by building the ptree at the point it is "ready"
for code generation and perform the permute optimizations on it and then
re-write the code generation routines to work off the ptree.
Thanks,
Richard.
- Thanks and regards,
Sameera D.
