Author: Richard Plangger <r...@pasra.at>
Branch: vecopt-merge
Changeset: r79117:3c01987639c7
Date: 2015-08-21 15:43 +0200
http://bitbucket.org/pypy/pypy/changeset/3c01987639c7/
Log: documentation additions (command line flags), added description of
the ABC optimization, note on limitations
diff --git a/rpython/doc/jit/index.rst b/rpython/doc/jit/index.rst
--- a/rpython/doc/jit/index.rst
+++ b/rpython/doc/jit/index.rst
@@ -25,6 +25,7 @@
pyjitpl5
optimizer
virtualizable
+ vectorization
- :doc:`Overview <overview>`: motivating our approach
diff --git a/rpython/doc/jit/optimizer.rst b/rpython/doc/jit/optimizer.rst
--- a/rpython/doc/jit/optimizer.rst
+++ b/rpython/doc/jit/optimizer.rst
@@ -178,6 +178,10 @@
It is prepended to all optimizations and thus extends the Optimizer class
and unrolls the loop once before it proceeds.
+Vectorization
+-------------
+
+- :doc:`Vectorization <vectorization>`
What is missing from this document
----------------------------------
diff --git a/rpython/doc/jit/vectorization.rst
b/rpython/doc/jit/vectorization.rst
--- a/rpython/doc/jit/vectorization.rst
+++ b/rpython/doc/jit/vectorization.rst
@@ -7,6 +7,18 @@
that is that they use the same index variable and offset can be expressed as a
a linear or affine combination.
+Command line flags:
+
+* --jit vec=1: turns on the vectorization for marked jitdrivers
+ (e.g. those in the NumPyPy module).
+* --jit vec_all=1: turns on the vectorization for any jit driver. See
parameters for
+ the filtering heuristics of traces.
+* --jit vec_ratio=2: A number from 0 to 10 that represents a real number
(vec_ratio / 10).
+ This filters traces if vec_all is enabled. N is the trace count then the
number of
+ vector transformable operations (add_int -> vec_add_int) M, the following
must hold:
+ M / N >= (vec_ratio / 10)
+* --jit vec_length=60: The maximum number of trace instructions the vectorizer
filters for.
+
Features
--------
@@ -38,6 +50,28 @@
load/store instructions) are not removed. The backend removes these
instructions
while assembling the trace.
+In addition a simple heuristic (enabled by --jit vec_all=1) tries to remove
+array bound checks for application level loops. It tries to identify the array
+bound checks and adds a transitive guard at the top of the loop::
+
+ label(...)
+ ...
+ guard(i < n) # index guard
+ ...
+ guard(i < len(a))
+ a = load(..., i, ...)
+ ...
+ jump(...)
+ # becomes
+ guard(n < len(a))
+ label(...)
+ guard(i < n) # index guard
+ ...
+ a = load(..., i, ...)
+ ...
+ jump(...)
+
+
Future Work and Limitations
---------------------------
@@ -54,5 +88,9 @@
to have 2 xmm registers (one filled with zero bits and the other with one
every bit).
This cuts down 2 instructions for guard checking, trading for higher
register pressure.
* prod, sum are only supported by 64 bit data types
+* isomorphic function prevents the following cases for combination into a pair:
+ 1) getarrayitem_gc, getarrayitem_gc_pure
+ 2) int_add(v,1), int_sub(v,-1)
.. _PMUL:
http://stackoverflow.com/questions/8866973/can-long-integer-routines-benefit-from-sse/8867025#8867025
+
_______________________________________________
pypy-commit mailing list
pypy-commit@python.org
https://mail.python.org/mailman/listinfo/pypy-commit
