Tom Stellard <thomas.stell...@amd.com> writes: > We were serializing the binaries once when clGetProgramInfo was called > with CL_PROGRAM_BINARY_SIZES and then again when it was called with > CL_PROGRAM_BINARIES. This was slowing down some OpenCV tests which were > building binary kernel caches. > > This improves the run-time of OpenCV's OCL_ImgProc/CvtColor8u.* > test from 7 minutes to 1 minute. > ---
Can you give the attached two patches a try? I'm curious to see if they have a comparable effect -- If they do I'd prefer to fix the underlying object rather than caching binaries in serialized form. Thanks. >[...]
From a500126213b073793184b0b6f170a58229340778 Mon Sep 17 00:00:00 2001
From: Francisco Jerez <curroje...@riseup.net>
Date: Sat, 14 Jun 2014 20:53:35 +0200
Subject: [PATCH 1/2] clover: Optimize module serialization for vectors of
fundamental types.
---
src/gallium/state_trackers/clover/core/module.cpp | 23 ++++++++++++++++++++++-
1 file changed, 22 insertions(+), 1 deletion(-)
diff --git a/src/gallium/state_trackers/clover/core/module.cpp b/src/gallium/state_trackers/clover/core/module.cpp
index 3e3ad99..41de734 100644
--- a/src/gallium/state_trackers/clover/core/module.cpp
+++ b/src/gallium/state_trackers/clover/core/module.cpp
@@ -69,7 +69,9 @@ namespace {
/// (De)serialize a vector.
template<typename T>
- struct _serializer<compat::vector<T>> {
+ struct _serializer<compat::vector<T>,
+ typename std::enable_if<
+ !std::is_scalar<T>::value>::type> {
static void
proc(compat::ostream &os, const compat::vector<T> &v) {
_proc<uint32_t>(os, v.size());
@@ -87,6 +89,25 @@ namespace {
}
};
+ template<typename T>
+ struct _serializer<compat::vector<T>,
+ typename std::enable_if<
+ std::is_scalar<T>::value>::type> {
+ static void
+ proc(compat::ostream &os, const compat::vector<T> &v) {
+ _proc<uint32_t>(os, v.size());
+ os.write(reinterpret_cast<const char *>(v.begin()),
+ v.size() * sizeof(T));
+ }
+
+ static void
+ proc(compat::istream &is, compat::vector<T> &v) {
+ v.reserve(_proc<uint32_t>(is));
+ is.read(reinterpret_cast<char *>(v.begin()),
+ v.size() * sizeof(T));
+ }
+ };
+
/// (De)serialize a module::section.
template<>
struct _serializer<module::section> {
--
1.9.2
From 1267038c2b0621dddc3d5c7718eed7ef2beb111b Mon Sep 17 00:00:00 2001
From: Francisco Jerez <curroje...@riseup.net>
Date: Sat, 14 Jun 2014 21:03:02 +0200
Subject: [PATCH 2/2] clover: Calculate the serialized size of a module
efficiently.
---
src/gallium/state_trackers/clover/api/program.cpp | 5 +---
src/gallium/state_trackers/clover/core/module.cpp | 32 +++++++++++++++++++++++
src/gallium/state_trackers/clover/core/module.hpp | 1 +
3 files changed, 34 insertions(+), 4 deletions(-)
diff --git a/src/gallium/state_trackers/clover/api/program.cpp b/src/gallium/state_trackers/clover/api/program.cpp
index fedc91d..a14baa3 100644
--- a/src/gallium/state_trackers/clover/api/program.cpp
+++ b/src/gallium/state_trackers/clover/api/program.cpp
@@ -190,10 +190,7 @@ clGetProgramInfo(cl_program d_prog, cl_program_info param,
case CL_PROGRAM_BINARY_SIZES:
buf.as_vector<size_t>() = map([&](const device &dev) {
- compat::ostream::buffer_t bin;
- compat::ostream s(bin);
- prog.binary(dev).serialize(s);
- return bin.size();
+ return prog.binary(dev).size();
},
prog.devices());
break;
diff --git a/src/gallium/state_trackers/clover/core/module.cpp b/src/gallium/state_trackers/clover/core/module.cpp
index 41de734..55ed91a 100644
--- a/src/gallium/state_trackers/clover/core/module.cpp
+++ b/src/gallium/state_trackers/clover/core/module.cpp
@@ -52,6 +52,13 @@ namespace {
return x;
}
+ /// Calculate the size of the specified object.
+ template<typename T>
+ void
+ _proc(module::size_t &sz, const T &x) {
+ _serializer<T>::proc(sz, x);
+ }
+
/// (De)serialize a scalar value.
template<typename T>
struct _serializer<T, typename std::enable_if<
@@ -65,6 +72,11 @@ namespace {
proc(compat::istream &is, T &x) {
is.read(reinterpret_cast<char *>(&x), sizeof(x));
}
+
+ static void
+ proc(module::size_t &sz, const T &x) {
+ sz += sizeof(x);
+ }
};
/// (De)serialize a vector.
@@ -87,6 +99,14 @@ namespace {
for (size_t i = 0; i < v.size(); i++)
new(&v[i]) T(_proc<T>(is));
}
+
+ static void
+ proc(module::size_t &sz, const compat::vector<T> &v) {
+ sz += sizeof(uint32_t);
+
+ for (size_t i = 0; i < v.size(); i++)
+ _proc<T>(sz, v[i]);
+ }
};
template<typename T>
@@ -106,6 +126,11 @@ namespace {
is.read(reinterpret_cast<char *>(v.begin()),
v.size() * sizeof(T));
}
+
+ static void
+ proc(module::size_t &sz, const compat::vector<T> &v) {
+ sz += sizeof(uint32_t) + sizeof(T) * v.size();
+ }
};
/// (De)serialize a module::section.
@@ -170,4 +195,11 @@ namespace clover {
module::deserialize(compat::istream &is) {
return _proc<module>(is);
}
+
+ module::size_t
+ module::size() const {
+ size_t sz = 0;
+ _proc(sz, *this);
+ return sz;
+ }
}
diff --git a/src/gallium/state_trackers/clover/core/module.hpp b/src/gallium/state_trackers/clover/core/module.hpp
index 4a8dbcb..18a5bfb 100644
--- a/src/gallium/state_trackers/clover/core/module.hpp
+++ b/src/gallium/state_trackers/clover/core/module.hpp
@@ -105,6 +105,7 @@ namespace clover {
void serialize(compat::ostream &os) const;
static module deserialize(compat::istream &is);
+ size_t size() const;
compat::vector<symbol> syms;
compat::vector<section> secs;
--
1.9.2
pgpjhvpfdgaag.pgp
Description: PGP signature
_______________________________________________ mesa-dev mailing list mesa-dev@lists.freedesktop.org http://lists.freedesktop.org/mailman/listinfo/mesa-dev
