add an example for a usage of stream tags

examples/cpp/19_stream_tags/.gitignore

@@ -0,0 +1,3 @@
+main
+main.o
+main_c

examples/cpp/19_stream_tags/CMakeLists.txt

@@ -0,0 +1,4 @@
+compile_cpp_example_with_modes(stream_tags main.cpp)
+
+add_custom_target(cpp_example_stream_tags_okl ALL COMMAND ${CMAKE_COMMAND} -E copy ${CMAKE_CURRENT_SOURCE_DIR}/powerOfPi2.okl powerOfPi2.okl)
+add_dependencies(examples_cpp_stream_tags cpp_example_stream_tags_okl)

examples/cpp/19_stream_tags/Makefile

@@ -0,0 +1,27 @@
+
+PROJ_DIR := $(dir $(abspath $(lastword $(MAKEFILE_LIST))))
+
+ifndef OCCA_DIR
+ include $(PROJ_DIR)/../../../scripts/build/Makefile
+else
+ include ${OCCA_DIR}/scripts/build/Makefile
+endif
+
+#---[ COMPILATION ]-------------------------------
+headers = $(wildcard $(incPath)/*.hpp) $(wildcard $(incPath)/*.tpp)
+sources = $(wildcard $(srcPath)/*.cpp)
+
+objects = $(subst $(srcPath)/,$(objPath)/,$(sources:.cpp=.o))
+
+executables: ${PROJ_DIR}/main
+
+${PROJ_DIR}/main: $(objects) $(headers) ${PROJ_DIR}/main.cpp
+ $(compiler) $(compilerFlags) -o ${PROJ_DIR}/main $(flags) $(objects) ${PROJ_DIR}/main.cpp $(paths) $(linkerFlags)
+
+$(objPath)/%.o:$(srcPath)/%.cpp $(wildcard $(subst $(srcPath)/,$(incPath)/,$(<:.cpp=.hpp))) $(wildcard $(subst $(srcPath)/,$(incPath)/,$(<:.cpp=.tpp)))
+ $(compiler) $(compilerFlags) -o $@ $(flags) -c $(paths) $<
+
+clean:
+ rm -f $(objPath)/*;
+ rm -f ${PROJ_DIR}/main;
+#=================================================

examples/cpp/19_stream_tags/README.md

@@ -0,0 +1,28 @@
+# Example: Events
+
+GPU devices introduce `streams`, which potentially allow parallel queueing of instructions
+
+`Stream tags` are used to query and manage (synchronize) those streams
+
+This example shows how to setup `occa::streamTag` to manage jobs in different streams
+
+# Compiling the Example
+
+```bash
+make
+```
+
+## Usage
+
+```
+> ./main --help
+
+Usage: ./main [OPTIONS]
+
+Example showing the use of multiple non-blocking streams in a device
+
+Options:
+ -d, --device Device properties (default: "{mode: 'CUDA', device_id: 0}")
+ -h, --help Print usage
+ -v, --verbose Compile kernels in verbose mode
+```

examples/cpp/19_stream_tags/main.cpp

@@ -0,0 +1,92 @@
+#include <iostream>
+
+#include <occa.hpp>
+
+//---[ Internal Tools ]-----------------
+// Note: These headers are not officially supported
+// Please don't rely on it outside of the occa examples
+#include <occa/internal/utils/cli.hpp>
+//======================================
+
+
+occa::json parseArgs(int argc, const char **argv);
+
+int main(int argc, const char **argv) {
+ occa::json args = parseArgs(argc, argv);
+
+ occa::setDevice(occa::json::parse(args["options/device"]));
+
+ int entries = 1<<20;
+ int block = 64;
+ int group = 1;
+
+ float *a = new float[entries];
+ for (int i = 0; i < entries; i++)
+ a[i] = 0.f;
+
+ occa::memory o_a = occa::malloc<float>(entries);
+ o_a.copyFrom(a);
+
+ occa::json kernelProps({
+ {"defines/block", block},
+ {"defines/group", group},
+ {"serial/include_std", true},
+ });
+ occa::kernel powerOfPi2 = occa::buildKernel("powerOfPi2.okl",
+ "powerOfPi2",
+ kernelProps);
+ occa::json streamProps({
+ {"nonblocking", true},
+ });
+ occa::stream stream_a = occa::createStream(streamProps);
+ occa::stream stream_b = occa::createStream(streamProps);
+
+ occa::setStream(stream_a);
+ powerOfPi2(o_a, entries);
+ occa::streamTag tag_a = occa::tagStream();
+
+ // set stream_b to wait for the job(s) to be finished in stream_a
+ occa::streamWait(stream_b, tag_a);
+
+ occa::setStream(stream_b);
+ powerOfPi2(o_a, entries);
+ occa::streamTag tag_b = occa::tagStream();
+
+ // set the device to wait for stream_b to finish
+ occa::waitFor(tag_b);
+
+ o_a.copyTo(a);
+
+ const float tol = 1e-3;
+ for (auto i = 0; i < entries; i++) {
+ if (fabs(a[i] - 3.14159) > tol) {
+ std::cerr << "Invalid output value: " << a[i] << " in " << i << std::endl;
+ return -1;
+ }
+ }
+ return 0;
+}
+
+
+occa::json parseArgs(int argc, const char **argv) {
+ occa::cli::parser parser;
+ parser
+ .withDescription(
+ "Example showing the use of multiple device streams"
+ )
+ .addOption(
+ occa::cli::option('d', "device",
+ "Device properties (default: \"{mode: 'CUDA', device_id: 0}\")")
+ .withArg()
+ .withDefaultValue("{mode: 'CUDA', device_id: 0}")
+ )
+ .addOption(
+ occa::cli::option('v', "verbose",
+ "Compile kernels in verbose mode")
+ );
+
+ occa::json args = parser.parseArgs(argc, argv);
+ occa::settings()["kernel/verbose"] = args["options/verbose"];
+
+ return args;
+}

examples/cpp/19_stream_tags/powerOfPi2.okl

@@ -0,0 +1,10 @@
+@kernel void powerOfPi2(float* x,
+ int entries) {
+ for (int g = 0; g < group; g++; @outer) {
+ for (int i = 0; i < block; ++i; @inner) {
+ for (int j=i+g*block; j < entries; j+=block*group) {
+ x[j] = pow(3.14159,x[j]);
+ }
+ }
+ }
+}