build: add basic support for IBM POWER, tested on POWER9.

Use auto-vectorization same as for ARM.

CMakeLists.txt

@@ -91,6 +91,8 @@ endif()
 if(NOT DNNL_TARGET_ARCH)
     if(CMAKE_SYSTEM_PROCESSOR MATCHES "^(aarch64.*|AARCH64.*|arm64.*|ARM64.*)")
         set(DNNL_TARGET_ARCH "AARCH64")
+    elseif(CMAKE_SYSTEM_PROCESSOR MATCHES "^(ppc64.*|PPC64.*)")
+        set(DNNL_TARGET_ARCH "PPC64")
     else()
         set(DNNL_TARGET_ARCH "X64")
     endif()

README.md

@@ -19,7 +19,8 @@ oneAPI Deep Neural Network Library (oneDNN) is an open-source cross-platform
 performance library of basic building blocks for deep learning applications.
 The library is optimized for Intel Architecture Processors, Intel Processor
 Graphics and Xe architecture-based Graphics. Support for other architectures
-such as Arm\* 64-bit Architecture (AArch64) is experimental.
+such as Arm\* 64-bit Architecture (AArch64) and OpenPOWER\* Power ISA (PPC64)
+is experimental.
 See the [System Requirements](#system-requirements) section below.
 
 oneDNN is intended for deep learning applications and framework
@@ -88,11 +89,12 @@ If the configuration you need is not available, you can
 oneDNN supports platforms based on the following architectures:
 - [Intel 64 or AMD64](https://en.wikipedia.org/wiki/X86-64),
 - [Arm 64-bit Architecture (AArch64)](https://developer.arm.com/architectures/cpu-architecture/a-profile).
+- [OpenPOWER](https://openpowerfoundation.org/) / [IBM Power ISA](https://en.wikipedia.org/wiki/Power_ISA).
 
 > **WARNING**
 >
-> Arm 64-bit Architecture (AArch64) support is **experimental** with limited
-> testing validation.
+> Arm 64-bit Architecture (AArch64) and Power ISA (PPC64) support
+> is **experimental** with limited testing validation.
 
 The library is optimized for the following CPUs:
 * Intel Atom processor with Intel SSE4.1 support
@@ -123,7 +125,7 @@ The library is optimized for the following GPUs:
 ## Requirements for Building from Source
 
 oneDNN supports systems meeting the following requirements:
-* Operating system with Intel 64 architecture support
+* Operating system with Intel 64 / Arm 64 / Power architecture support
 * C++ compiler with C++11 standard support
 * [CMake](https://cmake.org/download/) 2.8.11 or later
 * [Doxygen](http://www.doxygen.nl/download.html#srcbin) 1.8.5 or later
@@ -135,8 +137,8 @@ dependencies.
 ### CPU Engine
 
 oneDNN CPU engine is used to execute primitives on Intel Architecture
-Processors, 64-bit Arm Architecture (AArch64) processors, and compatible
-devices.
+Processors, 64-bit Arm Architecture (AArch64) processors,
+64-bit Power ISA (PPC64) processors, and compatible devices.
 
 The CPU engine is built by default and cannot be disabled at build time. The
 engine can be configured to use the OpenMP or TBB threading runtime. The

cmake/platform.cmake

@@ -107,6 +107,12 @@ elseif(UNIX OR MINGW)
              if (CMAKE_SYSTEM_PROCESSOR STREQUAL CMAKE_HOST_SYSTEM_PROCESSOR)
                  append(DEF_ARCH_OPT_FLAGS "-mcpu=native")
              endif()
+        elseif(DNNL_TARGET_ARCH STREQUAL "PPC64")
+             set(DEF_ARCH_OPT_FLAGS "-O3")
+             # For native compilation tune for the host processor
+             if (CMAKE_SYSTEM_PROCESSOR STREQUAL CMAKE_HOST_SYSTEM_PROCESSOR)
+                 append(DEF_ARCH_OPT_FLAGS "-mcpu=native")
+             endif()
         elseif(DNNL_TARGET_ARCH STREQUAL "X64")
              set(DEF_ARCH_OPT_FLAGS "-msse4.1")
         endif()
@@ -159,6 +165,13 @@ elseif(UNIX OR MINGW)
              if (CMAKE_SYSTEM_PROCESSOR STREQUAL CMAKE_HOST_SYSTEM_PROCESSOR)
                  append(DEF_ARCH_OPT_FLAGS "-mcpu=native")
              endif()
+        elseif(DNNL_TARGET_ARCH STREQUAL "PPC64")
+             set(DEF_ARCH_OPT_FLAGS "-O3")
+             # In GCC, -ftree-vectorize is turned on under -O3 since 2007.
+             # For native compilation tune for the host processor
+             if (CMAKE_SYSTEM_PROCESSOR STREQUAL CMAKE_HOST_SYSTEM_PROCESSOR)
+                 append(DEF_ARCH_OPT_FLAGS "-mcpu=native")
+             endif()
         elseif(DNNL_TARGET_ARCH STREQUAL "X64")
              set(DEF_ARCH_OPT_FLAGS "-msse4.1")
         endif()

src/cpu/README.md

@@ -43,6 +43,7 @@ header file provides a set of helpers macros that could help conditionally
 enable or disable parts of code. There the following macros defined:
 - `DNNL_X64` is 1 on x64 architecture;
 - `DNNL_AARCH64` is 1 on Arm AArch64 architecture;
+- `DNNL_PPC64` is 1 on OpenPOWER / IBM Power architecture;
 - `DNNL_ARCH_GENERIC` is 1 on other platforms.
 Only one of the macros above is defined to 1. All others are defined to 0.
 

src/cpu/platform.hpp

@@ -25,20 +25,27 @@
 // Possible architectures:
 // - DNNL_X64
 // - DNNL_AARCH64
+// - DNNL_PPC64
 // - DNNL_ARCH_GENERIC
 // Target architecture macro is set to 1, others to 0. All macros are defined.
 
-#if defined(DNNL_X64) + defined(DNNL_AARCH64) + defined(DNNL_ARCH_GENERIC) == 0
+#if defined(DNNL_X64) + defined(DNNL_AARCH64) + defined(DNNL_PPC64) \
+                + defined(DNNL_ARCH_GENERIC) \
+        == 0
 #if defined(__x86_64__) || defined(_M_X64)
 #define DNNL_X64 1
 #elif defined(__aarch64__)
 #define DNNL_AARCH64 1
+#elif defined(__powerpc64__) || defined(__PPC64__) || defined(_ARCH_PPC64)
+#define DNNL_PPC64 1
 #else
 #define DNNL_ARCH_GENERIC 1
 #endif
 #endif // defined(DNNL_X64) + ... == 0
 
-#if defined(DNNL_X64) + defined(DNNL_AARCH64) + defined(DNNL_ARCH_GENERIC) != 1
+#if defined(DNNL_X64) + defined(DNNL_AARCH64) + defined(DNNL_PPC64) \
+                + defined(DNNL_ARCH_GENERIC) \
+        != 1
 #error One and only one architecture should be defined at a time
 #endif
 
@@ -48,6 +55,9 @@
 #if !defined(DNNL_AARCH64)
 #define DNNL_AARCH64 0
 #endif
+#if !defined(DNNL_PPC64)
+#define DNNL_PPC64 0
+#endif
 #if !defined(DNNL_ARCH_GENERIC)
 #define DNNL_ARCH_GENERIC 0
 #endif
@@ -56,6 +66,7 @@
 // Equivalent to: #if DNNL_$ARCH ... #endif
 #define DNNL_X64_ONLY(...) Z_CONDITIONAL_DO(DNNL_X64, __VA_ARGS__)
 #define DNNL_AARCH64_ONLY(...) Z_CONDITIONAL_DO(DNNL_AARCH64_ONLY, __VA_ARGS__)
+#define DNNL_PPC64_ONLY(...) Z_CONDITIONAL_DO(DNNL_PPC64_ONLY, __VA_ARGS__)
 
 namespace dnnl {
 namespace impl {