Merge branch 'master' into spokhode/dparray

Author: PokhodenkoSA

README.md

@@ -76,11 +76,17 @@ Examples
 ========
 See examples in folder `examples`.
 
-Run examples:
+Run python examples:
 ```bash
-python examples/create_sycl_queues.py
+for script in `ls examples/python/`; do echo "executing ${script}"; python examples/python/${script}; done
 ```
 
+Examples of building Cython extensions with DPC++ compiler, that interoperate with dpCtl can be found in 
+folder `cython`.
+
+Each example in `cython` folder can be built using `CC=clang CXX=dpcpp python setup.py build_ext --inplace`.
+Please refer to `run.py` script in respective folders to execute extensions.
+
 Tests
 =====
 See tests in folder `dpctl/tests`.

backends/CMakeLists.txt

@@ -77,15 +77,15 @@ add_library(
   source/dppl_sycl_queue_manager.cpp
   source/dppl_sycl_usm_interface.cpp
   source/dppl_utils.cpp
-  details/source/dppl_utils_details.cpp
+  helper/source/dppl_utils_helper.cpp
 )
 
 # Install DPPLSyclInterface
 target_include_directories(
   DPPLSyclInterface
   PRIVATE
   ${CMAKE_SOURCE_DIR}/include/
-  ${CMAKE_SOURCE_DIR}/details/include/
+  ${CMAKE_SOURCE_DIR}/helper/include/
 )
 
 if(WIN32)
@@ -125,10 +125,10 @@ foreach(HEADER ${HEADERS})
   install(FILES "${HEADER}" DESTINATION include/Support)
 endforeach()
 
-# Install all headers in details/include
-file(GLOB HEADERS "${CMAKE_SOURCE_DIR}/details/include/*.h*")
+# Install all headers in helper/include
+file(GLOB HEADERS "${CMAKE_SOURCE_DIR}/helper/include/*.h*")
 foreach(HEADER ${HEADERS})
-  install(FILES "${HEADER}" DESTINATION details/include)
+  install(FILES "${HEADER}" DESTINATION helper/include)
 endforeach()
 
 option(

backends/details/include/dppl_utils_details.h renamed to backends/helper/include/dppl_utils_helper.h

@@ -27,5 +27,5 @@
 #include <CL/sycl.hpp>
 using namespace cl::sycl;
 
-std::string DDPL_StrToDeviceType(info::device_type devTy);
-info::device_type DPPL_DeviceTypeToStr(std::string devTyStr);
+std::string DPPL_DeviceTypeToStr(info::device_type devTy);
+info::device_type DPPL_StrToDeviceType(std::string devTyStr);

backends/details/source/dppl_utils_details.cpp renamed to backends/helper/source/dppl_utils_helper.cpp

@@ -1,4 +1,4 @@
-//===------ dppl_utils_details.cpp - dpctl-C_API  ----*---- C++ -----*-----===//
+//===------ dppl_utils_helper.cpp - dpctl-C_API  ----*---- C++ -----*-----===//
 //
 //               Data Parallel Control Library (dpCtl)
 //
@@ -19,11 +19,11 @@
 //===----------------------------------------------------------------------===//
 ///
 /// \file
-/// This file implements the helper functions defined in dppl_utils_details.h.
+/// This file implements the helper functions defined in dppl_utils_helper.h.
 ///
 //===----------------------------------------------------------------------===//
 
-#include "dppl_utils_details.h"
+#include "dppl_utils_helper.h"
 #include <string>
 #include <sstream>
 
@@ -32,7 +32,7 @@ using namespace cl::sycl;
 /*!
 * Transforms enum info::device_type to string.
 */
-std::string DDPL_StrToDeviceType(info::device_type devTy)
+std::string DPPL_DeviceTypeToStr(info::device_type devTy)
 {
     std::stringstream ss;
     switch (devTy)
@@ -61,7 +61,7 @@ std::string DDPL_StrToDeviceType(info::device_type devTy)
 /*!
 * Transforms string to enum info::device_type.
 */
-info::device_type DPPL_DeviceTypeToStr(std::string devTyStr)
+info::device_type DPPL_StrToDeviceType(std::string devTyStr)
 {
     info::device_type devTy;
     if (devTyStr == "cpu") {

backends/source/dppl_sycl_context_interface.cpp

@@ -47,7 +47,11 @@ bool DPPLContext_AreEq (__dppl_keep const DPPLSyclContextRef CtxRef1,
 
 bool DPPLContext_IsHost (__dppl_keep const DPPLSyclContextRef CtxRef)
 {
-    return unwrap(CtxRef)->is_host();
+    auto Ctx = unwrap(CtxRef);
+    if (Ctx) {
+        return Ctx->is_host();
+    }
+    return false;
 }
 
 void DPPLContext_Delete (__dppl_take DPPLSyclContextRef CtxRef)

backends/source/dppl_sycl_device_interface.cpp

@@ -30,7 +30,7 @@
 #include <iostream>
 #include <cstring>
 #include <CL/sycl.hpp>                /* SYCL headers   */
-#include "../details/include/dppl_utils_details.h"
+#include "../helper/include/dppl_utils_helper.h"
 
 using namespace cl::sycl;
 
@@ -59,7 +59,7 @@ void dump_device_info (const device & Device)
     ss << std::setw(4) << " " << std::left << std::setw(16) << "Device type";
 
     auto devTy = Device.get_info<info::device::device_type>();
-    ss << DDPL_StrToDeviceType(devTy);
+    ss << DPPL_DeviceTypeToStr(devTy);
 
     std::cout << ss.str();
 }

backends/source/dppl_sycl_platform_interface.cpp

@@ -29,7 +29,7 @@
 #include <iostream>
 #include <set>
 #include <sstream>
-#include "../details/include/dppl_utils_details.h"
+#include "../helper/include/dppl_utils_helper.h"
 
 #include <CL/sycl.hpp>
 
@@ -125,7 +125,7 @@ void DPPLPlatform_DumpInfo ()
                << "Device type";
 
             auto devTy = devices[dn].get_info<info::device::device_type>();
-            ss << DDPL_StrToDeviceType(devTy);
+            ss << DPPL_DeviceTypeToStr(devTy);
         }
         std::cout << ss.str();
         ++i;
@@ -148,7 +148,12 @@ size_t DPPLPlatform_GetNumNonHostPlatforms ()
 
 size_t DPPLPlatform_GetNumNonHostBackends ()
 {
-    return get_set_of_non_hostbackends().size();
+    auto be_set = get_set_of_non_hostbackends();
+
+    if (be_set.empty())
+        return 0;
+
+    return be_set.size();
 }
 
 __dppl_give DPPLSyclBackendType *DPPLPlatform_GetListOfNonHostBackends ()

examples/cython/sycl_buffer/README.md

@@ -0,0 +1,80 @@
+#1 Example of SYCL extension working NumPy array input via SYCL buffers
+
+
+#2 Decription
+
+Cython function expecting a 2D array in C-contiguous layout that
+computes column-wise total by using SYCL oneMKL (as GEMV call with
+an all units vector).
+
+Example illustrates compiling SYCL extension, linking to oneMKL.
+
+
+#2 Compiling
+
+```
+# make sure oneAPI is activated, $ONEAPI_ROOT must be set
+CC=clang CXX=dpcpp python setup.py build_ext --inplace
+```
+
+
+#2 Running
+
+```
+# SYCL_BE=PI_OPENCL sets SYCL backend to OpenCL to avoid a
+# transient issue with MKL's using the default Level-0 backend
+(idp) [08:16:12 ansatnuc04 simple]$ SYCL_BE=PI_OPENCL ipython
+Python 3.7.7 (default, Jul 14 2020, 22:02:37)
+Type 'copyright', 'credits' or 'license' for more information
+IPython 7.17.0 -- An enhanced Interactive Python. Type '?' for help.
+
+In [1]: import syclbuffer as sb, numpy as np, dpctl
+
+In [2]: x = np.random.randn(10**4, 2500)
+
+In [3]: %time m1 = np.sum(x, axis=0)
+CPU times: user 22.3 ms, sys: 160 µs, total: 22.5 ms
+Wall time: 21.2 ms
+
+In [4]: %time m = sb.columnwise_total(x)  # first time is slower, due to JIT overhead
+CPU times: user 207 ms, sys: 36.1 ms, total: 243 ms
+Wall time: 248 ms
+
+In [5]: %time m = sb.columnwise_total(x)
+CPU times: user 8.89 ms, sys: 4.12 ms, total: 13 ms
+Wall time: 12.4 ms
+
+In [6]: %time m = sb.columnwise_total(x)
+CPU times: user 4.82 ms, sys: 8.06 ms, total: 12.9 ms
+Wall time: 12.3 ms
+```
+
+Running bench.py:
+
+```
+========== Executing warm-up ==========
+NumPy result:  [1. 1. 1. ... 1. 1. 1.]
+SYCL(Intel(R) Core(TM) i7-10710U CPU @ 1.10GHz) result: [1. 1. 1. ... 1. 1. 1.]
+SYCL(Intel(R) Gen9 HD Graphics NEO) result: [1. 1. 1. ... 1. 1. 1.]
+Times for 'opencl:cpu:0'
+[2.864787499012891, 2.690436460019555, 2.5902308400254697, 2.5802528870408423, 2.538990616973024]
+Times for 'opencl:gpu:0'
+[1.9769684099592268, 2.3491444009705447, 2.293720397981815, 2.391633405990433, 1.9465659779962152]
+Times for NumPy
+[3.4011058019823395, 3.07286038500024, 3.0390414349967614, 3.0305576199898496, 3.002687797998078]
+```
+
+Running run.py:
+
+```
+(idp) [09:14:53 ansatnuc04 sycl_buffer]$ SYCL_BE=PI_OPENCL python run.py
+Result computed by NumPy
+[  0.27170187 -23.36798583   7.31326489  -1.95121928]
+Result computed by SYCL extension
+[  0.27170187 -23.36798583   7.31326489  -1.95121928]
+
+Running on:  Intel(R) Gen9 HD Graphics NEO
+[  0.27170187 -23.36798583   7.31326489  -1.95121928]
+Running on:  Intel(R) Core(TM) i7-10710U CPU @ 1.10GHz
+[  0.27170187 -23.36798583   7.31326489  -1.95121928]
+```

examples/cython/sycl_buffer/_buffer_example.pyx

@@ -0,0 +1,28 @@
+cimport numpy as cnp
+import numpy as np
+
+cimport dpctl as c_dpctl
+import dpctl
+
+cdef extern from "use_sycl_buffer.h":
+    int c_columnwise_total(c_dpctl.DPPLSyclQueueRef q, size_t n, size_t m, double *m, double *ct) nogil
+    int c_columnwise_total_no_mkl(c_dpctl.DPPLSyclQueueRef q, size_t n, size_t m, double *m, double *ct) nogil
+
+def columnwise_total(double[:, ::1] v, method='mkl'):
+    cdef cnp.ndarray res_array = np.empty((v.shape[1],), dtype='d')
+    cdef double[::1] res_memslice = res_array
+    cdef int ret_status
+    cdef c_dpctl.SyclQueue q
+    cdef c_dpctl.DPPLSyclQueueRef q_ref
+
+    q = c_dpctl.get_current_queue()
+    q_ref = q.get_queue_ref()
+
+    if method == 'mkl':
+        with nogil:
+            ret_status = c_columnwise_total(q_ref, v.shape[0], v.shape[1], &v[0,0], &res_memslice[0])
+    else:
+        with nogil:
+            ret_status = c_columnwise_total_no_mkl(q_ref, v.shape[0], v.shape[1], &v[0,0], &res_memslice[0])        
+
+    return res_array

examples/cython/sycl_buffer/bench.py

@@ -0,0 +1,51 @@
+import dpctl
+import syclbuffer as sb
+import numpy as np
+
+X = np.full((10 ** 4, 4098), 1e-4, dtype="d")
+
+# warm-up
+print("=" * 10 + " Executing warm-up " + "=" * 10)
+print("NumPy result: ", X.sum(axis=0))
+
+dpctl.set_default_queue("opencl", "cpu", 0)
+print(
+    "SYCL({}) result: {}".format(
+        dpctl.get_current_queue().get_sycl_device().get_device_name(),
+        sb.columnwise_total(X),
+    )
+)
+
+dpctl.set_default_queue("opencl", "gpu", 0)
+print(
+    "SYCL({}) result: {}".format(
+        dpctl.get_current_queue().get_sycl_device().get_device_name(),
+        sb.columnwise_total(X),
+    )
+)
+
+import timeit
+
+print("Times for 'opencl:cpu:0'")
+print(
+    timeit.repeat(
+        stmt="sb.columnwise_total(X)",
+        setup='dpctl.set_default_queue("opencl", "cpu", 0); '
+        "sb.columnwise_total(X)",  # ensure JIT compilation is not counted
+        number=100,
+        globals=globals(),
+    )
+)
+
+print("Times for 'opencl:gpu:0'")
+print(
+    timeit.repeat(
+        stmt="sb.columnwise_total(X)",
+        setup='dpctl.set_default_queue("opencl", "gpu", 0); sb.columnwise_total(X)',
+        number=100,
+        globals=globals(),
+    )
+)
+
+print("Times for NumPy")
+print(timeit.repeat(stmt="X.sum(axis=0)", number=100, globals=globals()))