mirror of
https://github.com/NVIDIA/cuda-samples.git
synced 2026-10-23 21:28:17 +08:00
Add and update samples for cuda 11.0 support
This commit is contained in:
Mahesh Doijade
360
Samples/globalToShmemAsyncCopy/Makefile
Normal file
360
Samples/globalToShmemAsyncCopy/Makefile
Normal file
@@ -0,0 +1,360 @@
|
||||
################################################################################
|
||||
# Copyright (c) 2019, NVIDIA CORPORATION. All rights reserved.
|
||||
#
|
||||
# Redistribution and use in source and binary forms, with or without
|
||||
# modification, are permitted provided that the following conditions
|
||||
# are met:
|
||||
# * Redistributions of source code must retain the above copyright
|
||||
# notice, this list of conditions and the following disclaimer.
|
||||
# * Redistributions in binary form must reproduce the above copyright
|
||||
# notice, this list of conditions and the following disclaimer in the
|
||||
# documentation and/or other materials provided with the distribution.
|
||||
# * Neither the name of NVIDIA CORPORATION nor the names of its
|
||||
# contributors may be used to endorse or promote products derived
|
||||
# from this software without specific prior written permission.
|
||||
#
|
||||
# THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS ``AS IS'' AND ANY
|
||||
# EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
|
||||
# IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR
|
||||
# PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT OWNER OR
|
||||
# CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL,
|
||||
# EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT LIMITED TO,
|
||||
# PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE, DATA, OR
|
||||
# PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY THEORY
|
||||
# OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT
|
||||
# (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE
|
||||
# OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
|
||||
#
|
||||
################################################################################
|
||||
#
|
||||
# Makefile project only supported on Mac OS X and Linux Platforms)
|
||||
#
|
||||
################################################################################
|
||||
|
||||
# Location of the CUDA Toolkit
|
||||
CUDA_PATH ?= /usr/local/cuda
|
||||
|
||||
##############################
|
||||
# start deprecated interface #
|
||||
##############################
|
||||
ifeq ($(x86_64),1)
|
||||
$(info WARNING - x86_64 variable has been deprecated)
|
||||
$(info WARNING - please use TARGET_ARCH=x86_64 instead)
|
||||
TARGET_ARCH ?= x86_64
|
||||
endif
|
||||
ifeq ($(ARMv7),1)
|
||||
$(info WARNING - ARMv7 variable has been deprecated)
|
||||
$(info WARNING - please use TARGET_ARCH=armv7l instead)
|
||||
TARGET_ARCH ?= armv7l
|
||||
endif
|
||||
ifeq ($(aarch64),1)
|
||||
$(info WARNING - aarch64 variable has been deprecated)
|
||||
$(info WARNING - please use TARGET_ARCH=aarch64 instead)
|
||||
TARGET_ARCH ?= aarch64
|
||||
endif
|
||||
ifeq ($(ppc64le),1)
|
||||
$(info WARNING - ppc64le variable has been deprecated)
|
||||
$(info WARNING - please use TARGET_ARCH=ppc64le instead)
|
||||
TARGET_ARCH ?= ppc64le
|
||||
endif
|
||||
ifneq ($(GCC),)
|
||||
$(info WARNING - GCC variable has been deprecated)
|
||||
$(info WARNING - please use HOST_COMPILER=$(GCC) instead)
|
||||
HOST_COMPILER ?= $(GCC)
|
||||
endif
|
||||
ifneq ($(abi),)
|
||||
$(error ERROR - abi variable has been removed)
|
||||
endif
|
||||
############################
|
||||
# end deprecated interface #
|
||||
############################
|
||||
|
||||
# architecture
|
||||
HOST_ARCH := $(shell uname -m)
|
||||
TARGET_ARCH ?= $(HOST_ARCH)
|
||||
ifneq (,$(filter $(TARGET_ARCH),x86_64 aarch64 sbsa ppc64le armv7l))
|
||||
ifneq ($(TARGET_ARCH),$(HOST_ARCH))
|
||||
ifneq (,$(filter $(TARGET_ARCH),x86_64 aarch64 sbsa ppc64le))
|
||||
TARGET_SIZE := 64
|
||||
else ifneq (,$(filter $(TARGET_ARCH),armv7l))
|
||||
TARGET_SIZE := 32
|
||||
endif
|
||||
else
|
||||
TARGET_SIZE := $(shell getconf LONG_BIT)
|
||||
endif
|
||||
else
|
||||
$(error ERROR - unsupported value $(TARGET_ARCH) for TARGET_ARCH!)
|
||||
endif
|
||||
|
||||
# sbsa and aarch64 systems look similar. Need to differentiate them at host level for now.
|
||||
ifeq ($(HOST_ARCH),aarch64)
|
||||
ifeq ($(CUDA_PATH)/targets/sbsa-linux,$(shell ls -1d $(CUDA_PATH)/targets/sbsa-linux))
|
||||
HOST_ARCH := sbsa
|
||||
TARGET_ARCH := sbsa
|
||||
endif
|
||||
endif
|
||||
|
||||
ifneq ($(TARGET_ARCH),$(HOST_ARCH))
|
||||
ifeq (,$(filter $(HOST_ARCH)-$(TARGET_ARCH),aarch64-armv7l x86_64-armv7l x86_64-aarch64 x86_64-sbsa x86_64-ppc64le))
|
||||
$(error ERROR - cross compiling from $(HOST_ARCH) to $(TARGET_ARCH) is not supported!)
|
||||
endif
|
||||
endif
|
||||
|
||||
# When on native aarch64 system with userspace of 32-bit, change TARGET_ARCH to armv7l
|
||||
ifeq ($(HOST_ARCH)-$(TARGET_ARCH)-$(TARGET_SIZE),aarch64-aarch64-32)
|
||||
TARGET_ARCH = armv7l
|
||||
endif
|
||||
|
||||
# operating system
|
||||
HOST_OS := $(shell uname -s 2>/dev/null | tr "[:upper:]" "[:lower:]")
|
||||
TARGET_OS ?= $(HOST_OS)
|
||||
ifeq (,$(filter $(TARGET_OS),linux darwin qnx android))
|
||||
$(error ERROR - unsupported value $(TARGET_OS) for TARGET_OS!)
|
||||
endif
|
||||
|
||||
# host compiler
|
||||
ifeq ($(TARGET_OS),darwin)
|
||||
ifeq ($(shell expr `xcodebuild -version | grep -i xcode | awk '{print $$2}' | cut -d'.' -f1` \>= 5),1)
|
||||
HOST_COMPILER ?= clang++
|
||||
endif
|
||||
else ifneq ($(TARGET_ARCH),$(HOST_ARCH))
|
||||
ifeq ($(HOST_ARCH)-$(TARGET_ARCH),x86_64-armv7l)
|
||||
ifeq ($(TARGET_OS),linux)
|
||||
HOST_COMPILER ?= arm-linux-gnueabihf-g++
|
||||
else ifeq ($(TARGET_OS),qnx)
|
||||
ifeq ($(QNX_HOST),)
|
||||
$(error ERROR - QNX_HOST must be passed to the QNX host toolchain)
|
||||
endif
|
||||
ifeq ($(QNX_TARGET),)
|
||||
$(error ERROR - QNX_TARGET must be passed to the QNX target toolchain)
|
||||
endif
|
||||
export QNX_HOST
|
||||
export QNX_TARGET
|
||||
HOST_COMPILER ?= $(QNX_HOST)/usr/bin/arm-unknown-nto-qnx6.6.0eabi-g++
|
||||
else ifeq ($(TARGET_OS),android)
|
||||
HOST_COMPILER ?= arm-linux-androideabi-g++
|
||||
endif
|
||||
else ifeq ($(TARGET_ARCH),aarch64)
|
||||
ifeq ($(TARGET_OS), linux)
|
||||
HOST_COMPILER ?= aarch64-linux-gnu-g++
|
||||
else ifeq ($(TARGET_OS),qnx)
|
||||
ifeq ($(QNX_HOST),)
|
||||
$(error ERROR - QNX_HOST must be passed to the QNX host toolchain)
|
||||
endif
|
||||
ifeq ($(QNX_TARGET),)
|
||||
$(error ERROR - QNX_TARGET must be passed to the QNX target toolchain)
|
||||
endif
|
||||
export QNX_HOST
|
||||
export QNX_TARGET
|
||||
HOST_COMPILER ?= $(QNX_HOST)/usr/bin/aarch64-unknown-nto-qnx7.0.0-g++
|
||||
else ifeq ($(TARGET_OS), android)
|
||||
HOST_COMPILER ?= aarch64-linux-android-clang++
|
||||
endif
|
||||
else ifeq ($(TARGET_ARCH),sbsa)
|
||||
HOST_COMPILER ?= aarch64-linux-gnu-g++
|
||||
else ifeq ($(TARGET_ARCH),ppc64le)
|
||||
HOST_COMPILER ?= powerpc64le-linux-gnu-g++
|
||||
endif
|
||||
endif
|
||||
HOST_COMPILER ?= g++
|
||||
NVCC := $(CUDA_PATH)/bin/nvcc -ccbin $(HOST_COMPILER)
|
||||
|
||||
# internal flags
|
||||
NVCCFLAGS := -m${TARGET_SIZE}
|
||||
CCFLAGS :=
|
||||
LDFLAGS :=
|
||||
|
||||
# build flags
|
||||
ifeq ($(TARGET_OS),darwin)
|
||||
LDFLAGS += -rpath $(CUDA_PATH)/lib
|
||||
CCFLAGS += -arch $(HOST_ARCH)
|
||||
else ifeq ($(HOST_ARCH)-$(TARGET_ARCH)-$(TARGET_OS),x86_64-armv7l-linux)
|
||||
LDFLAGS += --dynamic-linker=/lib/ld-linux-armhf.so.3
|
||||
CCFLAGS += -mfloat-abi=hard
|
||||
else ifeq ($(TARGET_OS),android)
|
||||
LDFLAGS += -pie
|
||||
CCFLAGS += -fpie -fpic -fexceptions
|
||||
endif
|
||||
|
||||
ifneq ($(TARGET_ARCH),$(HOST_ARCH))
|
||||
ifeq ($(TARGET_ARCH)-$(TARGET_OS),armv7l-linux)
|
||||
ifneq ($(TARGET_FS),)
|
||||
GCCVERSIONLTEQ46 := $(shell expr `$(HOST_COMPILER) -dumpversion` \<= 4.6)
|
||||
ifeq ($(GCCVERSIONLTEQ46),1)
|
||||
CCFLAGS += --sysroot=$(TARGET_FS)
|
||||
endif
|
||||
LDFLAGS += --sysroot=$(TARGET_FS)
|
||||
LDFLAGS += -rpath-link=$(TARGET_FS)/lib
|
||||
LDFLAGS += -rpath-link=$(TARGET_FS)/usr/lib
|
||||
LDFLAGS += -rpath-link=$(TARGET_FS)/usr/lib/arm-linux-gnueabihf
|
||||
endif
|
||||
endif
|
||||
ifeq ($(TARGET_ARCH)-$(TARGET_OS),aarch64-linux)
|
||||
ifneq ($(TARGET_FS),)
|
||||
GCCVERSIONLTEQ46 := $(shell expr `$(HOST_COMPILER) -dumpversion` \<= 4.6)
|
||||
ifeq ($(GCCVERSIONLTEQ46),1)
|
||||
CCFLAGS += --sysroot=$(TARGET_FS)
|
||||
endif
|
||||
LDFLAGS += --sysroot=$(TARGET_FS)
|
||||
LDFLAGS += -rpath-link=$(TARGET_FS)/lib -L$(TARGET_FS)/lib
|
||||
LDFLAGS += -rpath-link=$(TARGET_FS)/lib/aarch64-linux-gnu -L$(TARGET_FS)/lib/aarch64-linux-gnu
|
||||
LDFLAGS += -rpath-link=$(TARGET_FS)/usr/lib -L$(TARGET_FS)/usr/lib
|
||||
LDFLAGS += -rpath-link=$(TARGET_FS)/usr/lib/aarch64-linux-gnu -L$(TARGET_FS)/usr/lib/aarch64-linux-gnu
|
||||
LDFLAGS += --unresolved-symbols=ignore-in-shared-libs
|
||||
CCFLAGS += -isystem=$(TARGET_FS)/usr/include -I$(TARGET_FS)/usr/include
|
||||
CCFLAGS += -isystem=$(TARGET_FS)/usr/include/aarch64-linux-gnu -I$(TARGET_FS)/usr/include/aarch64-linux-gnu
|
||||
endif
|
||||
endif
|
||||
ifeq ($(TARGET_ARCH)-$(TARGET_OS),aarch64-qnx)
|
||||
CCFLAGS += -DWIN_INTERFACE_CUSTOM -I/usr/include/aarch64-qnx-gnu
|
||||
LDFLAGS += -lsocket
|
||||
LDFLAGS += -rpath=/usr/lib/aarch64-qnx-gnu -L/usr/lib/aarch64-qnx-gnu
|
||||
ifneq ($(TARGET_FS),)
|
||||
LDFLAGS += -rpath=$(TARGET_FS)/usr/lib -L $(TARGET_FS)/usr/lib
|
||||
LDFLAGS += -rpath=$(TARGET_FS)/usr/libnvidia -L $(TARGET_FS)/usr/libnvidia
|
||||
endif
|
||||
ifdef TARGET_OVERRIDE # cuda toolkit targets override
|
||||
NVCCFLAGS += -target-dir $(TARGET_OVERRIDE)
|
||||
endif
|
||||
endif
|
||||
endif
|
||||
|
||||
# Install directory of different arch
|
||||
CUDA_INSTALL_TARGET_DIR :=
|
||||
ifeq ($(TARGET_ARCH)-$(TARGET_OS),armv7l-linux)
|
||||
CUDA_INSTALL_TARGET_DIR = targets/armv7-linux-gnueabihf/
|
||||
else ifeq ($(TARGET_ARCH)-$(TARGET_OS),aarch64-linux)
|
||||
CUDA_INSTALL_TARGET_DIR = targets/aarch64-linux/
|
||||
else ifeq ($(TARGET_ARCH)-$(TARGET_OS),sbsa-linux)
|
||||
CUDA_INSTALL_TARGET_DIR = targets/sbsa-linux/
|
||||
else ifeq ($(TARGET_ARCH)-$(TARGET_OS),armv7l-android)
|
||||
CUDA_INSTALL_TARGET_DIR = targets/armv7-linux-androideabi/
|
||||
else ifeq ($(TARGET_ARCH)-$(TARGET_OS),aarch64-android)
|
||||
CUDA_INSTALL_TARGET_DIR = targets/aarch64-linux-androideabi/
|
||||
else ifeq ($(TARGET_ARCH)-$(TARGET_OS),armv7l-qnx)
|
||||
CUDA_INSTALL_TARGET_DIR = targets/ARMv7-linux-QNX/
|
||||
else ifeq ($(TARGET_ARCH)-$(TARGET_OS),aarch64-qnx)
|
||||
CUDA_INSTALL_TARGET_DIR = targets/aarch64-qnx/
|
||||
else ifeq ($(TARGET_ARCH),ppc64le)
|
||||
CUDA_INSTALL_TARGET_DIR = targets/ppc64le-linux/
|
||||
endif
|
||||
|
||||
# Debug build flags
|
||||
ifeq ($(dbg),1)
|
||||
NVCCFLAGS += -g -G
|
||||
BUILD_TYPE := debug
|
||||
else
|
||||
BUILD_TYPE := release
|
||||
endif
|
||||
|
||||
ALL_CCFLAGS :=
|
||||
ALL_CCFLAGS += $(NVCCFLAGS)
|
||||
ALL_CCFLAGS += $(EXTRA_NVCCFLAGS)
|
||||
ALL_CCFLAGS += $(addprefix -Xcompiler ,$(CCFLAGS))
|
||||
ALL_CCFLAGS += $(addprefix -Xcompiler ,$(EXTRA_CCFLAGS))
|
||||
|
||||
SAMPLE_ENABLED := 1
|
||||
|
||||
# This sample is not supported on QNX
|
||||
ifeq ($(TARGET_OS),qnx)
|
||||
$(info >>> WARNING - globalToShmemAsyncCopy is not supported on QNX - waiving sample <<<)
|
||||
SAMPLE_ENABLED := 0
|
||||
endif
|
||||
|
||||
ALL_LDFLAGS :=
|
||||
ALL_LDFLAGS += $(ALL_CCFLAGS)
|
||||
ALL_LDFLAGS += $(addprefix -Xlinker ,$(LDFLAGS))
|
||||
ALL_LDFLAGS += $(addprefix -Xlinker ,$(EXTRA_LDFLAGS))
|
||||
|
||||
# Common includes and paths for CUDA
|
||||
INCLUDES := -I../../Common
|
||||
LIBRARIES :=
|
||||
|
||||
################################################################################
|
||||
|
||||
#Detect if installed version of GCC supports required C++11
|
||||
ifeq ($(TARGET_OS),linux)
|
||||
empty :=
|
||||
space := $(empty) $(empty)
|
||||
GCCVERSIONSTRING := $(shell expr `$(HOST_COMPILER) -dumpversion`)
|
||||
#Create version number without "."
|
||||
GCCVERSION := $(shell expr `echo $(GCCVERSIONSTRING)` | cut -f1 -d.)
|
||||
GCCVERSION += $(shell expr `echo $(GCCVERSIONSTRING)` | cut -f2 -d.)
|
||||
GCCVERSION += $(shell expr `echo $(GCCVERSIONSTRING)` | cut -f3 -d.)
|
||||
# Make sure the version number has at least 3 decimals
|
||||
GCCVERSION += 00
|
||||
# Remove spaces from the version number
|
||||
GCCVERSION := $(subst $(space),$(empty),$(GCCVERSION))
|
||||
# Crop the version number to 3 decimals.
|
||||
GCCVERSION := $(shell expr `echo $(GCCVERSION)` | cut -b1-3)
|
||||
#$(warning $(GCCVERSION))
|
||||
|
||||
IS_MIN_VERSION := $(shell expr `echo $(GCCVERSION)` \>= 500)
|
||||
|
||||
ifeq ($(IS_MIN_VERSION), 1)
|
||||
$(info >>> GCC Version is greater or equal to 5.0.0 <<<)
|
||||
else
|
||||
$(info >>> Waiving build. Minimum GCC version required is 5.0.0<<<)
|
||||
SAMPLE_ENABLED := 0
|
||||
endif
|
||||
endif
|
||||
|
||||
# Gencode arguments
|
||||
ifeq ($(TARGET_ARCH),$(filter $(TARGET_ARCH),armv7l aarch64))
|
||||
SMS ?= 35 37 50 52 60 61 70 72 75 80
|
||||
else
|
||||
SMS ?= 35 37 50 52 60 61 70 75 80
|
||||
endif
|
||||
|
||||
ifeq ($(SMS),)
|
||||
$(info >>> WARNING - no SM architectures have been specified - waiving sample <<<)
|
||||
SAMPLE_ENABLED := 0
|
||||
endif
|
||||
|
||||
ifeq ($(GENCODE_FLAGS),)
|
||||
# Generate SASS code for each SM architecture listed in $(SMS)
|
||||
$(foreach sm,$(SMS),$(eval GENCODE_FLAGS += -gencode arch=compute_$(sm),code=sm_$(sm)))
|
||||
|
||||
# Generate PTX code from the highest SM architecture in $(SMS) to guarantee forward-compatibility
|
||||
HIGHEST_SM := $(lastword $(sort $(SMS)))
|
||||
ifneq ($(HIGHEST_SM),)
|
||||
GENCODE_FLAGS += -gencode arch=compute_$(HIGHEST_SM),code=compute_$(HIGHEST_SM)
|
||||
endif
|
||||
endif
|
||||
|
||||
ALL_CCFLAGS += --std=c++11
|
||||
|
||||
ifeq ($(SAMPLE_ENABLED),0)
|
||||
EXEC ?= @echo "[@]"
|
||||
endif
|
||||
|
||||
################################################################################
|
||||
|
||||
# Target rules
|
||||
all: build
|
||||
|
||||
build: globalToShmemAsyncCopy
|
||||
|
||||
check.deps:
|
||||
ifeq ($(SAMPLE_ENABLED),0)
|
||||
@echo "Sample will be waived due to the above missing dependencies"
|
||||
else
|
||||
@echo "Sample is ready - all dependencies have been met"
|
||||
endif
|
||||
|
||||
globalToShmemAsyncCopy.o:globalToShmemAsyncCopy.cu
|
||||
$(EXEC) $(NVCC) $(INCLUDES) $(ALL_CCFLAGS) $(GENCODE_FLAGS) -o $@ -c $<
|
||||
|
||||
globalToShmemAsyncCopy: globalToShmemAsyncCopy.o
|
||||
$(EXEC) $(NVCC) $(ALL_LDFLAGS) $(GENCODE_FLAGS) -o $@ $+ $(LIBRARIES)
|
||||
$(EXEC) mkdir -p ../../bin/$(TARGET_ARCH)/$(TARGET_OS)/$(BUILD_TYPE)
|
||||
$(EXEC) cp $@ ../../bin/$(TARGET_ARCH)/$(TARGET_OS)/$(BUILD_TYPE)
|
||||
|
||||
run: build
|
||||
$(EXEC) ./globalToShmemAsyncCopy
|
||||
|
||||
clean:
|
||||
rm -f globalToShmemAsyncCopy globalToShmemAsyncCopy.o
|
||||
rm -rf ../../bin/$(TARGET_ARCH)/$(TARGET_OS)/$(BUILD_TYPE)/globalToShmemAsyncCopy
|
||||
|
||||
clobber: clean
|
||||
93
Samples/globalToShmemAsyncCopy/NsightEclipse.xml
Normal file
93
Samples/globalToShmemAsyncCopy/NsightEclipse.xml
Normal file
@@ -0,0 +1,93 @@
|
||||
<?xml version="1.0" encoding="UTF-8"?>
|
||||
<!DOCTYPE entry SYSTEM "SamplesInfo.dtd">
|
||||
<entry>
|
||||
<name>globalToShmemAsyncCopy</name>
|
||||
<cflags>
|
||||
<flag>--std=c++11</flag>
|
||||
</cflags>
|
||||
<cuda_api_list>
|
||||
<toolkit>cudaEventCreate</toolkit>
|
||||
<toolkit>cudaEventRecord</toolkit>
|
||||
<toolkit>cudaEventQuery</toolkit>
|
||||
<toolkit>cudaEventDestroy</toolkit>
|
||||
<toolkit>cudaEventElapsedTime</toolkit>
|
||||
<toolkit>cudaEventSynchronize</toolkit>
|
||||
<toolkit>cudaMalloc</toolkit>
|
||||
<toolkit>cudaFree</toolkit>
|
||||
<toolkit>cudaMemcpy</toolkit>
|
||||
</cuda_api_list>
|
||||
<description><![CDATA[This sample implements matrix multiplication which uses asynchronous copy of data from global to shared memory when on compute capability 8.0 or higher. Also demonstrates arrive-wait barrier for synchronization.]]></description>
|
||||
<devicecompilation>whole</devicecompilation>
|
||||
<includepaths>
|
||||
<path>./</path>
|
||||
<path>../</path>
|
||||
<path>../../common/inc</path>
|
||||
</includepaths>
|
||||
<keyconcepts>
|
||||
<concept level="basic">CUDA Runtime API</concept>
|
||||
<concept level="basic">Linear Algebra</concept>
|
||||
<concept level="basic">CPP11 CUDA</concept>
|
||||
</keyconcepts>
|
||||
<keywords>
|
||||
<keyword>CUDA</keyword>
|
||||
<keyword>matrix multiply</keyword>
|
||||
<keyword>Async copy</keyword>
|
||||
<keyword>CPP11</keyword>
|
||||
<keyword>GCC 5.0.0</keyword>
|
||||
</keywords>
|
||||
<libraries>
|
||||
</libraries>
|
||||
<librarypaths>
|
||||
</librarypaths>
|
||||
<nsight_eclipse>true</nsight_eclipse>
|
||||
<primary_file>globalToShmemAsyncCopy.cu</primary_file>
|
||||
<required_dependencies>
|
||||
<dependency>CPP11</dependency>
|
||||
</required_dependencies>
|
||||
<scopes>
|
||||
<scope>1:CUDA Basic Topics</scope>
|
||||
<scope>3:Linear Algebra</scope>
|
||||
</scopes>
|
||||
<sm-arch>sm35</sm-arch>
|
||||
<sm-arch>sm37</sm-arch>
|
||||
<sm-arch>sm50</sm-arch>
|
||||
<sm-arch>sm52</sm-arch>
|
||||
<sm-arch>sm60</sm-arch>
|
||||
<sm-arch>sm61</sm-arch>
|
||||
<sm-arch>sm70</sm-arch>
|
||||
<sm-arch>sm72</sm-arch>
|
||||
<sm-arch>sm75</sm-arch>
|
||||
<sm-arch>sm80</sm-arch>
|
||||
<supported_envs>
|
||||
<env>
|
||||
<arch>x86_64</arch>
|
||||
<platform>linux</platform>
|
||||
</env>
|
||||
<env>
|
||||
<arch>x86_64</arch>
|
||||
<platform>macosx</platform>
|
||||
</env>
|
||||
<env>
|
||||
<arch>arm</arch>
|
||||
</env>
|
||||
<env>
|
||||
<arch>ppc64le</arch>
|
||||
<platform>linux</platform>
|
||||
</env>
|
||||
<env>
|
||||
<arch>aarch64</arch>
|
||||
<platform>linux</platform>
|
||||
</env>
|
||||
<env>
|
||||
<arch>aarch64</arch>
|
||||
<platform>qnx</platform>
|
||||
</env>
|
||||
<env>
|
||||
<platform>windows7</platform>
|
||||
</env>
|
||||
</supported_envs>
|
||||
<supported_sm_architectures>
|
||||
<include>all</include>
|
||||
</supported_sm_architectures>
|
||||
<title>Global Memory to Shared Memory Async Copy</title>
|
||||
</entry>
|
||||
74
Samples/globalToShmemAsyncCopy/README.md
Normal file
74
Samples/globalToShmemAsyncCopy/README.md
Normal file
@@ -0,0 +1,74 @@
|
||||
# globalToShmemAsyncCopy - Global Memory to Shared Memory Async Copy
|
||||
|
||||
## Description
|
||||
|
||||
This sample implements matrix multiplication which uses asynchronous copy of data from global to shared memory when on compute capability 8.0 or higher. Also demonstrates arrive-wait barrier for synchronization.
|
||||
|
||||
## Key Concepts
|
||||
|
||||
CUDA Runtime API, Linear Algebra, CPP11 CUDA
|
||||
|
||||
## Supported SM Architectures
|
||||
|
||||
[SM 3.5 ](https://developer.nvidia.com/cuda-gpus) [SM 3.7 ](https://developer.nvidia.com/cuda-gpus) [SM 5.0 ](https://developer.nvidia.com/cuda-gpus) [SM 5.2 ](https://developer.nvidia.com/cuda-gpus) [SM 6.0 ](https://developer.nvidia.com/cuda-gpus) [SM 6.1 ](https://developer.nvidia.com/cuda-gpus) [SM 7.0 ](https://developer.nvidia.com/cuda-gpus) [SM 7.2 ](https://developer.nvidia.com/cuda-gpus) [SM 7.5 ](https://developer.nvidia.com/cuda-gpus) [SM 8.0 ](https://developer.nvidia.com/cuda-gpus)
|
||||
|
||||
## Supported OSes
|
||||
|
||||
Linux, Windows
|
||||
|
||||
## Supported CPU Architecture
|
||||
|
||||
x86_64, ppc64le, armv7l, aarch64
|
||||
|
||||
## CUDA APIs involved
|
||||
|
||||
### [CUDA Runtime API](http://docs.nvidia.com/cuda/cuda-runtime-api/index.html)
|
||||
cudaEventCreate, cudaEventRecord, cudaEventQuery, cudaEventDestroy, cudaEventElapsedTime, cudaEventSynchronize, cudaMalloc, cudaFree, cudaMemcpy
|
||||
|
||||
## Dependencies needed to build/run
|
||||
[CPP11](../../README.md#cpp11)
|
||||
|
||||
## Prerequisites
|
||||
|
||||
Download and install the [CUDA Toolkit 11.0](https://developer.nvidia.com/cuda-downloads) for your corresponding platform.
|
||||
Make sure the dependencies mentioned in [Dependencies]() section above are installed.
|
||||
|
||||
## Build and Run
|
||||
|
||||
### Windows
|
||||
The Windows samples are built using the Visual Studio IDE. Solution files (.sln) are provided for each supported version of Visual Studio, using the format:
|
||||
```
|
||||
*_vs<version>.sln - for Visual Studio <version>
|
||||
```
|
||||
Each individual sample has its own set of solution files in its directory:
|
||||
|
||||
To build/examine all the samples at once, the complete solution files should be used. To build/examine a single sample, the individual sample solution files should be used.
|
||||
> **Note:** Some samples require that the Microsoft DirectX SDK (June 2010 or newer) be installed and that the VC++ directory paths are properly set up (**Tools > Options...**). Check DirectX Dependencies section for details."
|
||||
|
||||
### Linux
|
||||
The Linux samples are built using makefiles. To use the makefiles, change the current directory to the sample directory you wish to build, and run make:
|
||||
```
|
||||
$ cd <sample_dir>
|
||||
$ make
|
||||
```
|
||||
The samples makefiles can take advantage of certain options:
|
||||
* **TARGET_ARCH=<arch>** - cross-compile targeting a specific architecture. Allowed architectures are x86_64, ppc64le, armv7l, aarch64.
|
||||
By default, TARGET_ARCH is set to HOST_ARCH. On a x86_64 machine, not setting TARGET_ARCH is the equivalent of setting TARGET_ARCH=x86_64.<br/>
|
||||
`$ make TARGET_ARCH=x86_64` <br/> `$ make TARGET_ARCH=ppc64le` <br/> `$ make TARGET_ARCH=armv7l` <br/> `$ make TARGET_ARCH=aarch64` <br/>
|
||||
See [here](http://docs.nvidia.com/cuda/cuda-samples/index.html#cross-samples) for more details.
|
||||
* **dbg=1** - build with debug symbols
|
||||
```
|
||||
$ make dbg=1
|
||||
```
|
||||
* **SMS="A B ..."** - override the SM architectures for which the sample will be built, where `"A B ..."` is a space-delimited list of SM architectures. For example, to generate SASS for SM 50 and SM 60, use `SMS="50 60"`.
|
||||
```
|
||||
$ make SMS="50 60"
|
||||
```
|
||||
|
||||
* **HOST_COMPILER=<host_compiler>** - override the default g++ host compiler. See the [Linux Installation Guide](http://docs.nvidia.com/cuda/cuda-installation-guide-linux/index.html#system-requirements) for a list of supported host compilers.
|
||||
```
|
||||
$ make HOST_COMPILER=g++
|
||||
```
|
||||
|
||||
## References (for more details)
|
||||
|
||||
951
Samples/globalToShmemAsyncCopy/globalToShmemAsyncCopy.cu
Normal file
951
Samples/globalToShmemAsyncCopy/globalToShmemAsyncCopy.cu
Normal file
@@ -0,0 +1,951 @@
|
||||
/* Copyright (c) 2020, NVIDIA CORPORATION. All rights reserved.
|
||||
*
|
||||
* Redistribution and use in source and binary forms, with or without
|
||||
* modification, are permitted provided that the following conditions
|
||||
* are met:
|
||||
* * Redistributions of source code must retain the above copyright
|
||||
* notice, this list of conditions and the following disclaimer.
|
||||
* * Redistributions in binary form must reproduce the above copyright
|
||||
* notice, this list of conditions and the following disclaimer in the
|
||||
* documentation and/or other materials provided with the distribution.
|
||||
* * Neither the name of NVIDIA CORPORATION nor the names of its
|
||||
* contributors may be used to endorse or promote products derived
|
||||
* from this software without specific prior written permission.
|
||||
*
|
||||
* THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS ``AS IS'' AND ANY
|
||||
* EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
|
||||
* IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR
|
||||
* PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT OWNER OR
|
||||
* CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL,
|
||||
* EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT LIMITED TO,
|
||||
* PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE, DATA, OR
|
||||
* PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY THEORY
|
||||
* OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT
|
||||
* (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE
|
||||
* OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
|
||||
*/
|
||||
|
||||
/**
|
||||
* Matrix multiplication: C = A * B.
|
||||
*
|
||||
* This sample demonstrates implements matrix multiplication which makes use of shared memory
|
||||
* to ensure data reuse, the matrix multiplication is done using tiling approach.
|
||||
* With compute capability 8.0 or higher the CUDA kernels involved uses asynchronously copy data
|
||||
* from global to shared memory; a.k.a., async-copy.
|
||||
* This sample has been written for clarity of exposition to illustrate various CUDA programming
|
||||
* principles, not with the goal of providing the most performant generic kernel for matrix multiplication.
|
||||
*/
|
||||
|
||||
// System includes
|
||||
#include <stdio.h>
|
||||
#include <assert.h>
|
||||
|
||||
// CUDA runtime
|
||||
#include <cuda_runtime.h>
|
||||
#include <cuda_pipeline.h>
|
||||
#if __CUDA_ARCH__ >= 700
|
||||
#include <cuda_awbarrier.h>
|
||||
#endif
|
||||
|
||||
// Helper functions and utilities to work with CUDA
|
||||
#include <helper_functions.h>
|
||||
#include <helper_cuda.h>
|
||||
|
||||
namespace nvcuda_namespace = nvcuda::experimental;
|
||||
|
||||
enum kernels
|
||||
{
|
||||
AsyncCopyMultiStageLargeChunk = 0,
|
||||
AsyncCopyLargeChunk = 1,
|
||||
AsyncCopyLargeChunkAWBarrier = 2,
|
||||
AsyncCopyMultiStage = 3,
|
||||
AsyncCopySingleStage = 4,
|
||||
Naive = 5,
|
||||
NaiveLargeChunk = 6
|
||||
};
|
||||
|
||||
const char* kernelNames[] = {"AsyncCopyMultiStageLargeChunk", "AsyncCopyLargeChunk",
|
||||
"AsyncCopyLargeChunkAWBarrier", "AsyncCopyMultiStage",
|
||||
"AsyncCopySingleStage", "Naive", "NaiveLargeChunk"};
|
||||
|
||||
#define USE_CPP_API 0
|
||||
|
||||
constexpr int blockSize = 16;
|
||||
|
||||
// Multi Stage memcpy_async pipeline with large chunk copy
|
||||
template <int BLOCK_SIZE> __global__ void MatrixMulAsyncCopyMultiStageLargeChunk(float* __restrict__ C,
|
||||
const float* __restrict__ A,
|
||||
const float* __restrict__ B, int wA,
|
||||
int wB) {
|
||||
// Requires BLOCK_SIZE % 4 == 0
|
||||
|
||||
// Multi-stage pipeline version
|
||||
constexpr size_t maxPipelineStages = 4;
|
||||
|
||||
// Declaration of the shared memory array As used to
|
||||
// store the sub-matrix of A for each stage
|
||||
__shared__ float As[maxPipelineStages][BLOCK_SIZE][BLOCK_SIZE];
|
||||
|
||||
// Declaration of the shared memory array Bs used to
|
||||
// store the sub-matrix of B for each stage
|
||||
__shared__ float Bs[maxPipelineStages][BLOCK_SIZE][BLOCK_SIZE];
|
||||
|
||||
float Csub = 0.0;
|
||||
|
||||
// Index of the first sub-matrix of A processed by the block
|
||||
const int aBegin = wA * (BLOCK_SIZE) * blockIdx.y;
|
||||
|
||||
// Index of the last sub-matrix of A processed by the block
|
||||
const int aEnd = aBegin + wA - 1;
|
||||
|
||||
// Step size used to iterate through the sub-matrices of A
|
||||
int aStep = BLOCK_SIZE;
|
||||
|
||||
// Index of the first sub-matrix of B processed by the block
|
||||
const int bBegin = BLOCK_SIZE * blockIdx.x;
|
||||
|
||||
// Step size used to iterate through the sub-matrices of B
|
||||
int bStep = BLOCK_SIZE * wB;
|
||||
|
||||
const int t4x = threadIdx.x * 4 ;
|
||||
|
||||
#if USE_CPP_API
|
||||
nvcuda_namespace::pipeline pipe;
|
||||
#endif
|
||||
// Loop over all the sub-matrices of A and B
|
||||
// required to compute the block sub-matrix
|
||||
for (int a = aBegin, b = bBegin, i = 0, aStage = aBegin, bStage = bBegin, iStage = 0; a <= aEnd; a += aStep, b += bStep, ++i ) {
|
||||
// Load the matrices from device memory to shared memory; each thread loads
|
||||
// one element of each matrix
|
||||
|
||||
for ( ; aStage <= a + aStep * maxPipelineStages ; aStage += aStep, bStage += bStep, ++iStage )
|
||||
{
|
||||
if ( aStage <= aEnd && t4x < BLOCK_SIZE )
|
||||
{
|
||||
// Rotating buffer
|
||||
const int j = iStage % maxPipelineStages;
|
||||
float4 * const A4s = reinterpret_cast<float4*>(& As[j][threadIdx.y][t4x]);
|
||||
float4 * const B4s = reinterpret_cast<float4*>(& Bs[j][threadIdx.y][t4x]);
|
||||
const float4 * const A4 = reinterpret_cast<const float4*>(& A[aStage + wA * threadIdx.y + t4x]);
|
||||
const float4 * const B4 = reinterpret_cast<const float4*>(& B[aStage + wA * threadIdx.y + t4x]);
|
||||
|
||||
#if USE_CPP_API
|
||||
nvcuda_namespace::memcpy_async(*A4s,*A4, pipe);
|
||||
nvcuda_namespace::memcpy_async(*B4s,*B4, pipe);
|
||||
#else
|
||||
__pipeline_memcpy_async(A4s, A4, sizeof(float4));
|
||||
__pipeline_memcpy_async(B4s, B4, sizeof(float4));
|
||||
#endif
|
||||
}
|
||||
|
||||
#if USE_CPP_API
|
||||
pipe.commit();
|
||||
#else
|
||||
__pipeline_commit();
|
||||
#endif
|
||||
}
|
||||
#if USE_CPP_API
|
||||
pipe.wait_prior<maxPipelineStages-1>();
|
||||
#else
|
||||
__pipeline_wait_prior(maxPipelineStages-1);
|
||||
#endif
|
||||
// Synchronize to make sure the matrices are loaded
|
||||
__syncthreads();
|
||||
|
||||
// Rotating buffer
|
||||
const int j = i % maxPipelineStages;
|
||||
|
||||
// Multiply the two matrices together;
|
||||
// each thread computes one element
|
||||
// of the block sub-matrix
|
||||
for (int k = 0; k < BLOCK_SIZE; ++k) {
|
||||
Csub += As[j][threadIdx.y][k] * Bs[j][k][threadIdx.x];
|
||||
}
|
||||
|
||||
// Don't have to synchronize because
|
||||
// next iteration is loading to a different buffer
|
||||
}
|
||||
|
||||
// Write the block sub-matrix to device memory;
|
||||
// each thread writes four element
|
||||
int c = wB * BLOCK_SIZE * blockIdx.y + BLOCK_SIZE * blockIdx.x;
|
||||
C[c + wB * threadIdx.y + threadIdx.x] = Csub;
|
||||
}
|
||||
|
||||
// Single Stage memcpy_async pipeline with Large copy chunk (float4)
|
||||
template <int BLOCK_SIZE> __global__ void MatrixMulAsyncCopyLargeChunk(float* __restrict__ C,
|
||||
const float* __restrict__ A,
|
||||
const float* __restrict__ B, int wA,
|
||||
int wB) {
|
||||
// Requires BLOCK_SIZE % 4 == 0
|
||||
|
||||
// Declaration of the shared memory array As used to
|
||||
// store the sub-matrix of A
|
||||
__shared__ float As[BLOCK_SIZE][BLOCK_SIZE];
|
||||
|
||||
// Declaration of the shared memory array Bs used to
|
||||
// store the sub-matrix of B
|
||||
__shared__ float Bs[BLOCK_SIZE][BLOCK_SIZE];
|
||||
|
||||
// Index of the first sub-matrix of A processed by the block
|
||||
int aBegin = wA * BLOCK_SIZE * blockIdx.y;
|
||||
|
||||
// Index of the last sub-matrix of A processed by the block
|
||||
int aEnd = aBegin + wA - 1;
|
||||
|
||||
// Step size used to iterate through the sub-matrices of A
|
||||
int aStep = BLOCK_SIZE;
|
||||
|
||||
// Index of the first sub-matrix of B processed by the block
|
||||
int bBegin = BLOCK_SIZE * blockIdx.x;
|
||||
|
||||
// Step size used to iterate through the sub-matrices of B
|
||||
int bStep = BLOCK_SIZE * wB;
|
||||
|
||||
// Single-stage pipeline version
|
||||
float Csub = 0.0;
|
||||
|
||||
const int t4x = threadIdx.x * 4;
|
||||
#if USE_CPP_API
|
||||
nvcuda_namespace::pipeline pipe;
|
||||
#endif
|
||||
// Loop over all the sub-matrices of A and B
|
||||
// required to compute the block sub-matrix
|
||||
for (int a = aBegin, b = bBegin; a <= aEnd; a += aStep, b += bStep) {
|
||||
// Load the matrices from device memory to shared memory;
|
||||
// a subset of threads loads a contiguous chunk of elements.
|
||||
|
||||
// Previously, per-thread:
|
||||
// As[ty][tx] = A[a + wA * ty + tx];
|
||||
// Bs[ty][tx] = B[b + wB * ty + tx];
|
||||
|
||||
// Now, one fourth of the threads load four elements of each matrix
|
||||
if ( t4x < BLOCK_SIZE ) {
|
||||
float4 * const A4s = reinterpret_cast<float4*>(& As[threadIdx.y][t4x]);
|
||||
float4 * const B4s = reinterpret_cast<float4*>(& Bs[threadIdx.y][t4x]);
|
||||
const float4 * const A4 = reinterpret_cast<const float4*>(& A[a + wA * threadIdx.y + t4x]);
|
||||
const float4 * const B4 = reinterpret_cast<const float4*>(& B[a + wA * threadIdx.y + t4x]);
|
||||
|
||||
#if USE_CPP_API
|
||||
nvcuda_namespace::memcpy_async(*A4s,*A4,pipe);
|
||||
nvcuda_namespace::memcpy_async(*B4s,*B4,pipe);
|
||||
|
||||
pipe.commit_and_wait();
|
||||
#else
|
||||
__pipeline_memcpy_async(A4s, A4, sizeof(float4));
|
||||
__pipeline_memcpy_async(B4s, B4, sizeof(float4));
|
||||
|
||||
__pipeline_commit();
|
||||
__pipeline_wait_prior(0);
|
||||
#endif
|
||||
}
|
||||
|
||||
// Synchronize to make sure the matrices are loaded
|
||||
__syncthreads();
|
||||
|
||||
// Multiply the two matrices together;
|
||||
// each thread computes one element
|
||||
// of the block sub-matrix
|
||||
#pragma unroll
|
||||
for (int k = 0; k < BLOCK_SIZE; ++k) {
|
||||
Csub += As[threadIdx.y][k] * Bs[k][threadIdx.x];
|
||||
}
|
||||
|
||||
// Synchronize to make sure that the preceding
|
||||
// computation is done before loading two new
|
||||
// sub-matrices of A and B in the next iteration
|
||||
__syncthreads();
|
||||
}
|
||||
|
||||
// Write the block sub-matrix to device memory;
|
||||
// each thread writes four element
|
||||
int c = wB * BLOCK_SIZE * blockIdx.y + BLOCK_SIZE * blockIdx.x;
|
||||
C[c + wB * threadIdx.y + threadIdx.x] = Csub;
|
||||
}
|
||||
|
||||
// Single Stage memcpy_async pipeline with Large copy chunk (float4) using arrive-wait barrier
|
||||
template <int BLOCK_SIZE> __global__ void MatrixMulAsyncCopyLargeChunkAWBarrier(float* __restrict__ C,
|
||||
const float* __restrict__ A,
|
||||
const float* __restrict__ B, int wA,
|
||||
int wB) {
|
||||
#if __CUDA_ARCH__ >= 700
|
||||
// Requires BLOCK_SIZE % 4 == 0
|
||||
|
||||
// Declaration of the shared memory array As used to
|
||||
// store the sub-matrix of A
|
||||
__shared__ float As[BLOCK_SIZE][BLOCK_SIZE];
|
||||
|
||||
// Declaration of the shared memory array Bs used to
|
||||
// store the sub-matrix of B
|
||||
__shared__ float Bs[BLOCK_SIZE][BLOCK_SIZE];
|
||||
|
||||
nvcuda_namespace::pipeline pipe;
|
||||
__shared__ nvcuda_namespace::awbarrier barrier;
|
||||
|
||||
if (threadIdx.x == 0) {
|
||||
nvcuda_namespace::init(&barrier, blockDim.x*blockDim.y);
|
||||
}
|
||||
__syncthreads();
|
||||
|
||||
// Index of the first sub-matrix of A processed by the block
|
||||
int aBegin = wA * BLOCK_SIZE * blockIdx.y;
|
||||
|
||||
// Index of the last sub-matrix of A processed by the block
|
||||
int aEnd = aBegin + wA - 1;
|
||||
|
||||
// Step size used to iterate through the sub-matrices of A
|
||||
int aStep = BLOCK_SIZE;
|
||||
|
||||
// Index of the first sub-matrix of B processed by the block
|
||||
int bBegin = BLOCK_SIZE * blockIdx.x;
|
||||
|
||||
// Step size used to iterate through the sub-matrices of B
|
||||
int bStep = BLOCK_SIZE * wB;
|
||||
|
||||
float Csub = 0.0;
|
||||
|
||||
const int t4x = threadIdx.x * 4;
|
||||
|
||||
// Loop over all the sub-matrices of A and B
|
||||
// required to compute the block sub-matrix
|
||||
for (int a = aBegin, b = bBegin; a <= aEnd; a += aStep, b += bStep) {
|
||||
// Load the matrices from device memory to shared memory;
|
||||
// a subset of threads loads a contiguous chunk of elements.
|
||||
|
||||
// Now, one fourth of the threads load four elements of each matrix
|
||||
if ( t4x < BLOCK_SIZE ) {
|
||||
float4 * const A4s = reinterpret_cast<float4*>(& As[threadIdx.y][t4x]);
|
||||
float4 * const B4s = reinterpret_cast<float4*>(& Bs[threadIdx.y][t4x]);
|
||||
const float4 * const A4 = reinterpret_cast<const float4*>(& A[a + wA * threadIdx.y + t4x]);
|
||||
const float4 * const B4 = reinterpret_cast<const float4*>(& B[a + wA * threadIdx.y + t4x]);
|
||||
|
||||
nvcuda_namespace::memcpy_async(*A4s,*A4,pipe);
|
||||
nvcuda_namespace::memcpy_async(*B4s,*B4,pipe);
|
||||
|
||||
pipe.arrive_on(barrier);
|
||||
}
|
||||
|
||||
// Synchronize to make sure the matrices are loaded
|
||||
barrier.arrive_and_wait();
|
||||
|
||||
// Multiply the two matrices together;
|
||||
// each thread computes one element
|
||||
// of the block sub-matrix
|
||||
#pragma unroll
|
||||
for (int k = 0; k < BLOCK_SIZE; ++k) {
|
||||
Csub += As[threadIdx.y][k] * Bs[k][threadIdx.x];
|
||||
}
|
||||
|
||||
// Synchronize to make sure that the preceding
|
||||
// computation is done before loading two new
|
||||
// sub-matrices of A and B in the next iteration
|
||||
__syncthreads();
|
||||
}
|
||||
|
||||
// Write the block sub-matrix to device memory;
|
||||
// each thread writes four element
|
||||
int c = wB * BLOCK_SIZE * blockIdx.y + BLOCK_SIZE * blockIdx.x;
|
||||
C[c + wB * threadIdx.y + threadIdx.x] = Csub;
|
||||
#endif
|
||||
}
|
||||
|
||||
// Single Stage memcpy_async pipeline with float copy
|
||||
template <int BLOCK_SIZE> __global__ void MatrixMulAsyncCopySingleStage(float *C, const float *A,
|
||||
const float *B, int wA,
|
||||
int wB) {
|
||||
|
||||
// Declaration of the shared memory array As used to
|
||||
// store the sub-matrix of A
|
||||
__shared__ float As[BLOCK_SIZE][BLOCK_SIZE];
|
||||
|
||||
// Declaration of the shared memory array Bs used to
|
||||
// store the sub-matrix of B
|
||||
__shared__ float Bs[BLOCK_SIZE][BLOCK_SIZE];
|
||||
|
||||
// Index of the first sub-matrix of A processed by the block
|
||||
int aBegin = wA * BLOCK_SIZE * blockIdx.y;
|
||||
|
||||
// Index of the last sub-matrix of A processed by the block
|
||||
int aEnd = aBegin + wA - 1;
|
||||
|
||||
// Step size used to iterate through the sub-matrices of A
|
||||
int aStep = BLOCK_SIZE;
|
||||
|
||||
// Index of the first sub-matrix of B processed by the block
|
||||
int bBegin = BLOCK_SIZE * blockIdx.x;
|
||||
|
||||
// Step size used to iterate through the sub-matrices of B
|
||||
int bStep = BLOCK_SIZE * wB;
|
||||
|
||||
// Single-stage pipeline version
|
||||
float Csub = 0.0;
|
||||
|
||||
#if USE_CPP_API
|
||||
nvcuda_namespace::pipeline pipe;
|
||||
#endif
|
||||
// Loop over all the sub-matrices of A and B
|
||||
// required to compute the block sub-matrix
|
||||
for (int a = aBegin, b = bBegin; a <= aEnd; a += aStep, b += bStep) {
|
||||
// Load the matrices from device memory to shared memory; each thread loads
|
||||
// one element of each matrix
|
||||
{
|
||||
const float *A_float = reinterpret_cast<const float*>(A + a + wA * threadIdx.y + threadIdx.x);
|
||||
const float *B_float = reinterpret_cast<const float*>(B + b + wB * threadIdx.y + threadIdx.x);
|
||||
|
||||
#if USE_CPP_API
|
||||
|
||||
nvcuda_namespace::memcpy_async(As[threadIdx.y][threadIdx.x], *A_float, pipe);
|
||||
nvcuda_namespace::memcpy_async(Bs[threadIdx.y][threadIdx.x], *B_float, pipe);
|
||||
|
||||
pipe.commit_and_wait();
|
||||
#else
|
||||
__pipeline_memcpy_async(&As[threadIdx.y][threadIdx.x], A_float, sizeof(float));
|
||||
__pipeline_memcpy_async(&Bs[threadIdx.y][threadIdx.x], B_float, sizeof(float));
|
||||
|
||||
__pipeline_commit();
|
||||
__pipeline_wait_prior(0);
|
||||
#endif
|
||||
}
|
||||
|
||||
// Synchronize to make sure the matrices are loaded
|
||||
__syncthreads();
|
||||
|
||||
// Multiply the two matrices together;
|
||||
// each thread computes one element
|
||||
// of the block sub-matrix
|
||||
#pragma unroll
|
||||
for (int k = 0; k < BLOCK_SIZE; ++k) {
|
||||
Csub += As[threadIdx.y][k] * Bs[k][threadIdx.x];
|
||||
}
|
||||
|
||||
// Synchronize to make sure that the preceding
|
||||
// computation is done before loading two new
|
||||
// sub-matrices of A and B in the next iteration
|
||||
__syncthreads();
|
||||
}
|
||||
|
||||
// Write the block sub-matrix to device memory;
|
||||
// each thread writes four element
|
||||
int c = wB * BLOCK_SIZE * blockIdx.y + BLOCK_SIZE * blockIdx.x;
|
||||
C[c + wB * threadIdx.y + threadIdx.x] = Csub;
|
||||
}
|
||||
|
||||
// Multi Stage memcpy_async pipeline with int copy
|
||||
template <int BLOCK_SIZE> __global__ void MatrixMulAsyncCopyMultiStage(float* __restrict__ C,
|
||||
const float* __restrict__ A,
|
||||
const float* __restrict__ B, int wA,
|
||||
int wB) {
|
||||
// Multi-stage pipeline version
|
||||
constexpr size_t maxPipelineStages = 4;
|
||||
|
||||
// Declaration of the shared memory array As used to
|
||||
// store the sub-matrix of A for each stage
|
||||
__shared__ float As[maxPipelineStages][BLOCK_SIZE][BLOCK_SIZE];
|
||||
|
||||
// Declaration of the shared memory array Bs used to
|
||||
// store the sub-matrix of B for each stage
|
||||
__shared__ float Bs[maxPipelineStages][BLOCK_SIZE][BLOCK_SIZE];
|
||||
|
||||
float Csub = 0.0;
|
||||
|
||||
// Index of the first sub-matrix of A processed by the block
|
||||
const int aBegin = wA * BLOCK_SIZE * blockIdx.y;
|
||||
|
||||
// Index of the last sub-matrix of A processed by the block
|
||||
const int aEnd = aBegin + wA - 1;
|
||||
|
||||
// Step size used to iterate through the sub-matrices of A
|
||||
int aStep = BLOCK_SIZE;
|
||||
|
||||
// Index of the first sub-matrix of B processed by the block
|
||||
const int bBegin = BLOCK_SIZE * blockIdx.x;
|
||||
|
||||
// Step size used to iterate through the sub-matrices of B
|
||||
int bStep = BLOCK_SIZE * wB;
|
||||
|
||||
#if USE_CPP_API
|
||||
nvcuda_namespace::pipeline pipe;
|
||||
#endif
|
||||
// Loop over all the sub-matrices of A and B
|
||||
// required to compute the block sub-matrix
|
||||
for (int a = aBegin, b = bBegin, i = 0, aStage = aBegin, bStage = bBegin, iStage = 0; a <= aEnd; a += aStep, b += bStep, ++i ) {
|
||||
// Load the matrices from device memory to shared memory; each thread loads
|
||||
// one element of each matrix
|
||||
|
||||
for ( ; aStage <= a + aStep * maxPipelineStages ; aStage += aStep, bStage += bStep, ++iStage )
|
||||
{
|
||||
if ( aStage <= aEnd )
|
||||
{
|
||||
const float *A_float = reinterpret_cast<const float*>(A + aStage + wA * threadIdx.y + threadIdx.x);
|
||||
const float *B_float = reinterpret_cast<const float*>(B + bStage + wB * threadIdx.y + threadIdx.x);
|
||||
|
||||
// Rotating buffer
|
||||
const int j = iStage % maxPipelineStages;
|
||||
#if USE_CPP_API
|
||||
nvcuda_namespace::memcpy_async(As[j][threadIdx.y][threadIdx.x], *A_float, pipe);
|
||||
nvcuda_namespace::memcpy_async(Bs[j][threadIdx.y][threadIdx.x], *B_float, pipe);
|
||||
#else
|
||||
__pipeline_memcpy_async(&As[j][threadIdx.y][threadIdx.x], A_float, sizeof(float));
|
||||
__pipeline_memcpy_async(&Bs[j][threadIdx.y][threadIdx.x], B_float, sizeof(float));
|
||||
#endif
|
||||
}
|
||||
#if USE_CPP_API
|
||||
pipe.commit();
|
||||
#else
|
||||
__pipeline_commit();
|
||||
#endif
|
||||
}
|
||||
#if USE_CPP_API
|
||||
pipe.wait_prior<maxPipelineStages-1>();
|
||||
#else
|
||||
__pipeline_wait_prior(maxPipelineStages-1);
|
||||
#endif
|
||||
// Synchronize to make sure the matrices are loaded
|
||||
__syncthreads();
|
||||
|
||||
const int j = i % maxPipelineStages;
|
||||
|
||||
// Multiply the two matrices together;
|
||||
// each thread computes one element
|
||||
// of the block sub-matrix
|
||||
for (int k = 0; k < BLOCK_SIZE; ++k) {
|
||||
Csub += As[j][threadIdx.y][k] * Bs[j][k][threadIdx.x];
|
||||
}
|
||||
|
||||
// Don't have to synchronize because
|
||||
// next iteration is loading to a different buffer
|
||||
}
|
||||
|
||||
// Write the block sub-matrix to device memory;
|
||||
// each thread writes four element
|
||||
int c = wB * BLOCK_SIZE * blockIdx.y + BLOCK_SIZE * blockIdx.x;
|
||||
C[c + wB * threadIdx.y + threadIdx.x] = Csub;
|
||||
}
|
||||
|
||||
/**
|
||||
* Matrix multiplication (CUDA Kernel) on the device: C = A * B
|
||||
* wA is A's width and wB is B's width
|
||||
*/
|
||||
template <int BLOCK_SIZE> __global__ void MatrixMulNaive(float *C, float *A,
|
||||
float *B, int wA,
|
||||
int wB) {
|
||||
// Declaration of the shared memory array As used to
|
||||
// store the sub-matrix of A
|
||||
__shared__ float As[BLOCK_SIZE][BLOCK_SIZE];
|
||||
|
||||
// Declaration of the shared memory array Bs used to
|
||||
// store the sub-matrix of B
|
||||
__shared__ float Bs[BLOCK_SIZE][BLOCK_SIZE];
|
||||
|
||||
// Index of the first sub-matrix of A processed by the block
|
||||
int aBegin = wA * BLOCK_SIZE * blockIdx.y;
|
||||
|
||||
// Index of the last sub-matrix of A processed by the block
|
||||
int aEnd = aBegin + wA - 1;
|
||||
|
||||
// Step size used to iterate through the sub-matrices of A
|
||||
int aStep = BLOCK_SIZE;
|
||||
|
||||
// Index of the first sub-matrix of B processed by the block
|
||||
int bBegin = BLOCK_SIZE * blockIdx.x;
|
||||
|
||||
// Step size used to iterate through the sub-matrices of B
|
||||
int bStep = BLOCK_SIZE * wB;
|
||||
|
||||
// Csub is used to store the element of the block sub-matrix
|
||||
// that is computed by the thread
|
||||
float Csub = 0;
|
||||
|
||||
// Loop over all the sub-matrices of A and B
|
||||
// required to compute the block sub-matrix
|
||||
for (int a = aBegin, b = bBegin;
|
||||
a <= aEnd;
|
||||
a += aStep, b += bStep) {
|
||||
|
||||
// Load the matrices from device memory
|
||||
// to shared memory; each thread loads
|
||||
// one element of each matrix
|
||||
As[threadIdx.y][threadIdx.x] = A[a + wA * threadIdx.y + threadIdx.x];
|
||||
Bs[threadIdx.y][threadIdx.x] = B[b + wB * threadIdx.y + threadIdx.x];
|
||||
|
||||
// Synchronize to make sure the matrices are loaded
|
||||
__syncthreads();
|
||||
|
||||
// Multiply the two matrices together;
|
||||
// each thread computes one element
|
||||
// of the block sub-matrix
|
||||
#pragma unroll
|
||||
for (int k = 0; k < BLOCK_SIZE; ++k) {
|
||||
Csub += As[threadIdx.y][k] * Bs[k][threadIdx.x];
|
||||
}
|
||||
|
||||
// Synchronize to make sure that the preceding
|
||||
// computation is done before loading two new
|
||||
// sub-matrices of A and B in the next iteration
|
||||
__syncthreads();
|
||||
}
|
||||
|
||||
// Write the block sub-matrix to device memory;
|
||||
// each thread writes one element
|
||||
int c = wB * BLOCK_SIZE * blockIdx.y + BLOCK_SIZE * blockIdx.x;
|
||||
C[c + wB * threadIdx.y + threadIdx.x] = Csub;
|
||||
}
|
||||
|
||||
template <int BLOCK_SIZE> __global__ void MatrixMulNaiveLargeChunk(float *C, float *A,
|
||||
float *B, int wA,
|
||||
int wB) {
|
||||
// Declaration of the shared memory array As used to
|
||||
// store the sub-matrix of A
|
||||
__shared__ float As[BLOCK_SIZE][BLOCK_SIZE];
|
||||
|
||||
// Declaration of the shared memory array Bs used to
|
||||
// store the sub-matrix of B
|
||||
__shared__ float Bs[BLOCK_SIZE][BLOCK_SIZE];
|
||||
|
||||
int t4x = threadIdx.x * 4 ;
|
||||
|
||||
// Index of the first sub-matrix of A processed by the block
|
||||
int aBegin = wA * BLOCK_SIZE * blockIdx.y;
|
||||
|
||||
// Index of the last sub-matrix of A processed by the block
|
||||
int aEnd = aBegin + wA - 1;
|
||||
|
||||
// Step size used to iterate through the sub-matrices of A
|
||||
int aStep = BLOCK_SIZE;
|
||||
|
||||
// Index of the first sub-matrix of B processed by the block
|
||||
int bBegin = BLOCK_SIZE * blockIdx.x;
|
||||
|
||||
// Step size used to iterate through the sub-matrices of B
|
||||
int bStep = BLOCK_SIZE * wB;
|
||||
|
||||
// Csub is used to store the element of the block sub-matrix
|
||||
// that is computed by the thread
|
||||
float Csub = 0;
|
||||
|
||||
// Loop over all the sub-matrices of A and B
|
||||
// required to compute the block sub-matrix
|
||||
for (int a = aBegin, b = bBegin;
|
||||
a <= aEnd;
|
||||
a += aStep, b += bStep) {
|
||||
|
||||
// Load the matrices from device memory
|
||||
// to shared memory;
|
||||
|
||||
// One fourth of the threads load four elements of each matrix
|
||||
if ( t4x < BLOCK_SIZE ) {
|
||||
float4 * const A4s = reinterpret_cast<float4*>(& As[threadIdx.y][t4x]);
|
||||
float4 * const B4s = reinterpret_cast<float4*>(& Bs[threadIdx.y][t4x]);
|
||||
const float4 * const A4 = reinterpret_cast<float4*>(& A[a + wA * threadIdx.y + t4x]);
|
||||
const float4 * const B4 = reinterpret_cast<float4*>(& B[a + wA * threadIdx.y + t4x]);
|
||||
*A4s = *A4 ;
|
||||
*B4s = *B4 ;
|
||||
}
|
||||
|
||||
// Synchronize to make sure the matrices are loaded
|
||||
__syncthreads();
|
||||
|
||||
// Multiply the two matrices together;
|
||||
// each thread computes one element
|
||||
// of the block sub-matrix
|
||||
#pragma unroll
|
||||
for (int k = 0; k < BLOCK_SIZE; ++k) {
|
||||
Csub += As[threadIdx.y][k] * Bs[k][threadIdx.x];
|
||||
}
|
||||
|
||||
// Synchronize to make sure that the preceding
|
||||
// computation is done before loading two new
|
||||
// sub-matrices of A and B in the next iteration
|
||||
__syncthreads();
|
||||
}
|
||||
|
||||
// Write the block sub-matrix to device memory;
|
||||
// each thread writes one element
|
||||
int c = wB * BLOCK_SIZE * blockIdx.y + BLOCK_SIZE * blockIdx.x;
|
||||
C[c + wB * threadIdx.y + threadIdx.x] = Csub;
|
||||
}
|
||||
|
||||
|
||||
void ConstantInit(float *data, int size, float val) {
|
||||
for (int i = 0; i < size; ++i) {
|
||||
data[i] = val;
|
||||
}
|
||||
}
|
||||
|
||||
/**
|
||||
* Run matrix multiplication using CUDA
|
||||
*/
|
||||
int MatrixMultiply(int argc, char **argv,
|
||||
const dim3 &dimsA,
|
||||
const dim3 &dimsB,
|
||||
kernels kernel_number) {
|
||||
// Allocate host memory for matrices A and B
|
||||
unsigned int size_A = dimsA.x * dimsA.y;
|
||||
unsigned int mem_size_A = sizeof(float) * size_A;
|
||||
float *h_A = reinterpret_cast<float *>(malloc(mem_size_A));
|
||||
unsigned int size_B = dimsB.x * dimsB.y;
|
||||
unsigned int mem_size_B = sizeof(float) * size_B;
|
||||
float *h_B = reinterpret_cast<float *>(malloc(mem_size_B));
|
||||
cudaStream_t stream;
|
||||
|
||||
// Initialize host memory
|
||||
const float valB = 2.10f;
|
||||
ConstantInit(h_A, size_A, 1.0f);
|
||||
ConstantInit(h_B, size_B, valB);
|
||||
|
||||
// Allocate device memory
|
||||
float *d_A, *d_B, *d_C;
|
||||
|
||||
// Allocate host matrix C
|
||||
dim3 dimsC(dimsB.x, dimsA.y, 1);
|
||||
unsigned int mem_size_C = dimsC.x * dimsC.y * sizeof(float);
|
||||
float *h_C = reinterpret_cast<float *>(malloc(mem_size_C));
|
||||
|
||||
if (h_C == NULL) {
|
||||
fprintf(stderr, "Failed to allocate host matrix C!\n");
|
||||
exit(EXIT_FAILURE);
|
||||
}
|
||||
|
||||
checkCudaErrors(cudaMalloc(reinterpret_cast<void **>(&d_A), mem_size_A));
|
||||
checkCudaErrors(cudaMalloc(reinterpret_cast<void **>(&d_B), mem_size_B));
|
||||
checkCudaErrors(cudaMalloc(reinterpret_cast<void **>(&d_C), mem_size_C));
|
||||
// Allocate CUDA events that we'll use for timing
|
||||
cudaEvent_t start, stop;
|
||||
checkCudaErrors(cudaEventCreate(&start));
|
||||
checkCudaErrors(cudaEventCreate(&stop));
|
||||
|
||||
checkCudaErrors(cudaStreamCreateWithFlags(&stream, cudaStreamNonBlocking));
|
||||
|
||||
// copy host memory to device
|
||||
checkCudaErrors(cudaMemcpyAsync(d_A, h_A, mem_size_A, cudaMemcpyHostToDevice, stream));
|
||||
checkCudaErrors(cudaMemcpyAsync(d_B, h_B, mem_size_B, cudaMemcpyHostToDevice, stream));
|
||||
checkCudaErrors(cudaMemsetAsync(d_C, 0, mem_size_C, stream));
|
||||
|
||||
// Setup execution parameters
|
||||
dim3 threads(blockSize, blockSize);
|
||||
dim3 grid(dimsB.x / threads.x, dimsA.y / threads.y);
|
||||
|
||||
|
||||
printf("Running kernel = %d - %s\n", kernel_number, kernelNames[kernel_number]);
|
||||
// Create and start timer
|
||||
printf("Computing result using CUDA Kernel...\n");
|
||||
|
||||
// Performs warmup operation using matrixMul CUDA kernel
|
||||
switch (kernel_number)
|
||||
{
|
||||
case AsyncCopyMultiStageLargeChunk :
|
||||
default:
|
||||
MatrixMulAsyncCopyMultiStageLargeChunk<blockSize><<<grid, threads, 0, stream>>>(d_C, d_A, d_B, dimsA.x, dimsB.x);
|
||||
break;
|
||||
case AsyncCopyLargeChunk :
|
||||
MatrixMulAsyncCopyLargeChunk<blockSize><<<grid, threads, 0, stream>>>(d_C, d_A, d_B, dimsA.x, dimsB.x);
|
||||
break;
|
||||
case AsyncCopyLargeChunkAWBarrier :
|
||||
MatrixMulAsyncCopyLargeChunkAWBarrier<blockSize><<<grid, threads, 0, stream>>>(d_C, d_A, d_B, dimsA.x, dimsB.x);
|
||||
break;
|
||||
case AsyncCopyMultiStage :
|
||||
MatrixMulAsyncCopyMultiStage<blockSize><<<grid, threads, 0, stream>>>(d_C, d_A, d_B, dimsA.x, dimsB.x);
|
||||
break;
|
||||
case AsyncCopySingleStage :
|
||||
MatrixMulAsyncCopySingleStage<blockSize><<<grid, threads, 0, stream>>>(d_C, d_A, d_B, dimsA.x, dimsB.x);
|
||||
break;
|
||||
case Naive :
|
||||
MatrixMulNaive<blockSize><<<grid, threads, 0, stream>>>(d_C, d_A, d_B, dimsA.x, dimsB.x);
|
||||
break;
|
||||
case NaiveLargeChunk:
|
||||
MatrixMulNaiveLargeChunk<blockSize><<<grid, threads, 0, stream>>>(d_C, d_A, d_B, dimsA.x, dimsB.x);
|
||||
break;
|
||||
}
|
||||
|
||||
printf("done\n");
|
||||
checkCudaErrors(cudaStreamSynchronize(stream));
|
||||
|
||||
|
||||
// Execute the kernel
|
||||
int nIter = 100;
|
||||
|
||||
// Record the start event
|
||||
checkCudaErrors(cudaEventRecord(start, stream));
|
||||
|
||||
for (int j = 0; j < nIter; j++) {
|
||||
switch (kernel_number)
|
||||
{
|
||||
case AsyncCopyMultiStageLargeChunk :
|
||||
default:
|
||||
MatrixMulAsyncCopyMultiStageLargeChunk<blockSize><<<grid, threads, 0, stream>>>(d_C, d_A, d_B, dimsA.x, dimsB.x);
|
||||
break;
|
||||
case AsyncCopyLargeChunk :
|
||||
MatrixMulAsyncCopyLargeChunk<blockSize><<<grid, threads, 0, stream>>>(d_C, d_A, d_B, dimsA.x, dimsB.x);
|
||||
break;
|
||||
case AsyncCopyLargeChunkAWBarrier :
|
||||
MatrixMulAsyncCopyLargeChunkAWBarrier<blockSize><<<grid, threads, 0, stream>>>(d_C, d_A, d_B, dimsA.x, dimsB.x);
|
||||
break;
|
||||
case AsyncCopyMultiStage :
|
||||
MatrixMulAsyncCopyMultiStage<blockSize><<<grid, threads, 0, stream>>>(d_C, d_A, d_B, dimsA.x, dimsB.x);
|
||||
break;
|
||||
case AsyncCopySingleStage :
|
||||
MatrixMulAsyncCopySingleStage<blockSize><<<grid, threads, 0, stream>>>(d_C, d_A, d_B, dimsA.x, dimsB.x);
|
||||
break;
|
||||
case Naive :
|
||||
MatrixMulNaive<blockSize><<<grid, threads, 0, stream>>>(d_C, d_A, d_B, dimsA.x, dimsB.x);
|
||||
break;
|
||||
case NaiveLargeChunk:
|
||||
MatrixMulNaiveLargeChunk<blockSize><<<grid, threads, 0, stream>>>(d_C, d_A, d_B, dimsA.x, dimsB.x);
|
||||
break;
|
||||
}
|
||||
}
|
||||
|
||||
// Record the stop event
|
||||
checkCudaErrors(cudaEventRecord(stop, stream));
|
||||
|
||||
// Wait for the stop event to complete
|
||||
checkCudaErrors(cudaEventSynchronize(stop));
|
||||
|
||||
float msecTotal = 0.0f;
|
||||
checkCudaErrors(cudaEventElapsedTime(&msecTotal, start, stop));
|
||||
|
||||
// Compute and print the performance
|
||||
float msecPerMatrixMul = msecTotal / nIter;
|
||||
double flopsPerMatrixMul = 2.0 * static_cast<double>(dimsA.x) *
|
||||
static_cast<double>(dimsA.y) *
|
||||
static_cast<double>(dimsB.x);
|
||||
double gigaFlops = (flopsPerMatrixMul * 1.0e-9f) /
|
||||
(msecPerMatrixMul / 1000.0f);
|
||||
printf(
|
||||
"Performance= %.2f GFlop/s, Time= %.3f msec, Size= %.0f Ops," \
|
||||
" WorkgroupSize= %u threads/block\n",
|
||||
gigaFlops,
|
||||
msecPerMatrixMul,
|
||||
flopsPerMatrixMul,
|
||||
threads.x * threads.y);
|
||||
|
||||
// Copy result from device to host
|
||||
checkCudaErrors(cudaMemcpyAsync(h_C, d_C, mem_size_C, cudaMemcpyDeviceToHost, stream));
|
||||
checkCudaErrors(cudaStreamSynchronize(stream));
|
||||
|
||||
printf("Checking computed result for correctness: ");
|
||||
bool correct = true;
|
||||
|
||||
// test relative error by the formula
|
||||
// |<x, y>_cpu - <x,y>_gpu|/<|x|, |y|> < eps
|
||||
double eps = 1.e-6; // machine zero
|
||||
|
||||
for (int i = 0; i < static_cast<int>(dimsC.x * dimsC.y); i++) {
|
||||
double abs_err = fabs(h_C[i] - (dimsA.x * valB));
|
||||
double dot_length = dimsA.x;
|
||||
double abs_val = fabs(h_C[i]);
|
||||
double rel_err = abs_err / abs_val / dot_length;
|
||||
|
||||
if (rel_err > eps) {
|
||||
printf("Error! Matrix[%05d]=%.8f, ref=%.8f error term is > %E\n",
|
||||
i, h_C[i], dimsA.x * valB, eps);
|
||||
correct = false;
|
||||
}
|
||||
}
|
||||
|
||||
printf("%s\n", correct ? "Result = PASS" : "Result = FAIL");
|
||||
|
||||
// Clean up memory
|
||||
free(h_A);
|
||||
free(h_B);
|
||||
free(h_C);
|
||||
checkCudaErrors(cudaFree(d_A));
|
||||
checkCudaErrors(cudaFree(d_B));
|
||||
checkCudaErrors(cudaFree(d_C));
|
||||
checkCudaErrors(cudaEventDestroy(start));
|
||||
checkCudaErrors(cudaEventDestroy(stop));
|
||||
printf("\nNOTE: The CUDA Samples are not meant for performance"\
|
||||
"measurements. Results may vary when GPU Boost is enabled.\n");
|
||||
|
||||
if (correct) {
|
||||
return EXIT_SUCCESS;
|
||||
} else {
|
||||
return EXIT_FAILURE;
|
||||
}
|
||||
}
|
||||
|
||||
|
||||
int main(int argc, char **argv) {
|
||||
printf("[Matrix Multiply Using CUDA] - Starting...\n");
|
||||
|
||||
if (checkCmdLineFlag(argc, (const char **)argv, "help") ||
|
||||
checkCmdLineFlag(argc, (const char **)argv, "?")) {
|
||||
printf("Usage -device=n (n >= 0 for deviceID)\n");
|
||||
printf(" -wA=WidthA -hA=HeightA (Width x Height of Matrix A)\n");
|
||||
printf(" -wB=WidthB -hB=HeightB (Width x Height of Matrix B)\n");
|
||||
printf(" -kernel=kernel_number (0 - AsyncCopyMultiStageLargeChunk; 1 - AsyncCopyLargeChunk)\n");
|
||||
printf(" (2 - AsyncCopyLargeChunkAWBarrier; 3 - AsyncCopyMultiStage)\n");
|
||||
printf(" (4 - AsyncCopySingleStage; 5 - Naive without memcpy_async)\n");
|
||||
printf(" (6 - NaiveLargeChunk without memcpy_async)\n");
|
||||
printf(" Note: Outer matrix dimensions of A & B matrices must be equal.\n");
|
||||
|
||||
exit(EXIT_SUCCESS);
|
||||
}
|
||||
|
||||
// This will pick the best possible CUDA capable device, otherwise
|
||||
// override the device ID based on input provided at the command line
|
||||
int dev = findCudaDevice(argc, (const char **)argv);
|
||||
|
||||
int matrixBlock = 32;
|
||||
dim3 dimsA(10 * 2 * matrixBlock, 10 * 2 * matrixBlock, 1);
|
||||
dim3 dimsB(10 * 2 * matrixBlock, 10 * 2 * matrixBlock, 1);
|
||||
|
||||
// width of Matrix A
|
||||
if (checkCmdLineFlag(argc, (const char **)argv, "wA")) {
|
||||
dimsA.x = getCmdLineArgumentInt(argc, (const char **)argv, "wA");
|
||||
}
|
||||
|
||||
// height of Matrix A
|
||||
if (checkCmdLineFlag(argc, (const char **)argv, "hA")) {
|
||||
dimsA.y = getCmdLineArgumentInt(argc, (const char **)argv, "hA");
|
||||
}
|
||||
|
||||
// width of Matrix B
|
||||
if (checkCmdLineFlag(argc, (const char **)argv, "wB")) {
|
||||
dimsB.x = getCmdLineArgumentInt(argc, (const char **)argv, "wB");
|
||||
}
|
||||
|
||||
// height of Matrix B
|
||||
if (checkCmdLineFlag(argc, (const char **)argv, "hB")) {
|
||||
dimsB.y = getCmdLineArgumentInt(argc, (const char **)argv, "hB");
|
||||
}
|
||||
|
||||
if (dimsA.x != dimsB.y) {
|
||||
printf("Error: outer matrix dimensions must be equal. (%d != %d)\n",
|
||||
dimsA.x, dimsB.y);
|
||||
exit(EXIT_FAILURE);
|
||||
}
|
||||
|
||||
kernels selected_kernel = AsyncCopyMultiStageLargeChunk;
|
||||
|
||||
// kernel to run - default (AsyncCopyMultiStageLargeChunk == 0)
|
||||
if (checkCmdLineFlag(argc, (const char **)argv, "kernel")) {
|
||||
int kernel_number = getCmdLineArgumentInt(argc, (const char **)argv, "kernel");
|
||||
if (kernel_number < 7)
|
||||
{
|
||||
selected_kernel = (kernels)kernel_number;
|
||||
}
|
||||
else
|
||||
{
|
||||
printf("Error: kernel number should be between 0 to 6, you have entered %d\n", kernel_number);
|
||||
exit(EXIT_FAILURE);
|
||||
}
|
||||
|
||||
int major = 0;
|
||||
checkCudaErrors(cudaDeviceGetAttribute(&major, cudaDevAttrComputeCapabilityMajor, dev));
|
||||
if ((kernel_number == AsyncCopyLargeChunkAWBarrier) && major < 7)
|
||||
{
|
||||
printf("AsyncCopyLargeChunkAWBarrier kernel requires requires SM 7.0 or higher. Exiting...\n");
|
||||
exit(EXIT_WAIVED);
|
||||
}
|
||||
}
|
||||
|
||||
|
||||
printf("MatrixA(%d,%d), MatrixB(%d,%d)\n", dimsA.x, dimsA.y,
|
||||
dimsB.x, dimsB.y);
|
||||
|
||||
int matrix_result = MatrixMultiply(argc, argv, dimsA, dimsB, selected_kernel);
|
||||
|
||||
exit(matrix_result);
|
||||
}
|
||||
|
||||
@@ -0,0 +1,20 @@
|
||||
|
||||
Microsoft Visual Studio Solution File, Format Version 14.00
|
||||
# Visual Studio 2015
|
||||
Project("{8BC9CEB8-8B4A-11D0-8D11-00A0C91BC942}") = "globalToShmemAsyncCopy", "globalToShmemAsyncCopy_vs2015.vcxproj", "{997E0757-EA74-4A4E-A0FC-47D8C8831A15}"
|
||||
EndProject
|
||||
Global
|
||||
GlobalSection(SolutionConfigurationPlatforms) = preSolution
|
||||
Debug|x64 = Debug|x64
|
||||
Release|x64 = Release|x64
|
||||
EndGlobalSection
|
||||
GlobalSection(ProjectConfigurationPlatforms) = postSolution
|
||||
{997E0757-EA74-4A4E-A0FC-47D8C8831A15}.Debug|x64.ActiveCfg = Debug|x64
|
||||
{997E0757-EA74-4A4E-A0FC-47D8C8831A15}.Debug|x64.Build.0 = Debug|x64
|
||||
{997E0757-EA74-4A4E-A0FC-47D8C8831A15}.Release|x64.ActiveCfg = Release|x64
|
||||
{997E0757-EA74-4A4E-A0FC-47D8C8831A15}.Release|x64.Build.0 = Release|x64
|
||||
EndGlobalSection
|
||||
GlobalSection(SolutionProperties) = preSolution
|
||||
HideSolutionNode = FALSE
|
||||
EndGlobalSection
|
||||
EndGlobal
|
||||
@@ -0,0 +1,107 @@
|
||||
<?xml version="1.0" encoding="utf-8"?>
|
||||
<Project DefaultTargets="Build" ToolsVersion="4.0" xmlns="http://schemas.microsoft.com/developer/msbuild/2003">
|
||||
<PropertyGroup>
|
||||
<CUDAPropsPath Condition="'$(CUDAPropsPath)'==''">$(VCTargetsPath)\BuildCustomizations</CUDAPropsPath>
|
||||
</PropertyGroup>
|
||||
<ItemGroup Label="ProjectConfigurations">
|
||||
<ProjectConfiguration Include="Debug|x64">
|
||||
<Configuration>Debug</Configuration>
|
||||
<Platform>x64</Platform>
|
||||
</ProjectConfiguration>
|
||||
<ProjectConfiguration Include="Release|x64">
|
||||
<Configuration>Release</Configuration>
|
||||
<Platform>x64</Platform>
|
||||
</ProjectConfiguration>
|
||||
</ItemGroup>
|
||||
<PropertyGroup Label="Globals">
|
||||
<ProjectGuid>{997E0757-EA74-4A4E-A0FC-47D8C8831A15}</ProjectGuid>
|
||||
<RootNamespace>globalToShmemAsyncCopy_vs2015</RootNamespace>
|
||||
<ProjectName>globalToShmemAsyncCopy</ProjectName>
|
||||
<CudaToolkitCustomDir />
|
||||
</PropertyGroup>
|
||||
<Import Project="$(VCTargetsPath)\Microsoft.Cpp.Default.props" />
|
||||
<PropertyGroup>
|
||||
<ConfigurationType>Application</ConfigurationType>
|
||||
<CharacterSet>MultiByte</CharacterSet>
|
||||
<PlatformToolset>v140</PlatformToolset>
|
||||
</PropertyGroup>
|
||||
<PropertyGroup Condition="'$(Configuration)'=='Debug'">
|
||||
<UseDebugLibraries>true</UseDebugLibraries>
|
||||
</PropertyGroup>
|
||||
<PropertyGroup Condition="'$(Configuration)'=='Release'">
|
||||
<WholeProgramOptimization>true</WholeProgramOptimization>
|
||||
</PropertyGroup>
|
||||
<Import Project="$(VCTargetsPath)\Microsoft.Cpp.props" />
|
||||
<ImportGroup Label="ExtensionSettings">
|
||||
<Import Project="$(CUDAPropsPath)\CUDA 10.2.props" />
|
||||
</ImportGroup>
|
||||
<ImportGroup Label="PropertySheets">
|
||||
<Import Condition="exists('$(UserRootDir)\Microsoft.Cpp.$(Platform).user.props')" Label="LocalAppDataPlatform" Project="$(UserRootDir)\Microsoft.Cpp.$(Platform).user.props" />
|
||||
</ImportGroup>
|
||||
<PropertyGroup Label="UserMacros" />
|
||||
<PropertyGroup>
|
||||
<IntDir>$(Platform)/$(Configuration)/</IntDir>
|
||||
<IncludePath>$(IncludePath)</IncludePath>
|
||||
<CodeAnalysisRuleSet>AllRules.ruleset</CodeAnalysisRuleSet>
|
||||
<CodeAnalysisRules />
|
||||
<CodeAnalysisRuleAssemblies />
|
||||
</PropertyGroup>
|
||||
<PropertyGroup Condition="'$(Platform)'=='x64'">
|
||||
<OutDir>../../bin/win64/$(Configuration)/</OutDir>
|
||||
</PropertyGroup>
|
||||
<ItemDefinitionGroup>
|
||||
<ClCompile>
|
||||
<WarningLevel>Level3</WarningLevel>
|
||||
<PreprocessorDefinitions>WIN32;_MBCS;%(PreprocessorDefinitions)</PreprocessorDefinitions>
|
||||
<AdditionalIncludeDirectories>./;$(CudaToolkitDir)/include;../../Common;</AdditionalIncludeDirectories>
|
||||
</ClCompile>
|
||||
<Link>
|
||||
<SubSystem>Console</SubSystem>
|
||||
<AdditionalDependencies>cudart_static.lib;kernel32.lib;user32.lib;gdi32.lib;winspool.lib;comdlg32.lib;advapi32.lib;shell32.lib;ole32.lib;oleaut32.lib;uuid.lib;odbc32.lib;odbccp32.lib;%(AdditionalDependencies)</AdditionalDependencies>
|
||||
<AdditionalLibraryDirectories>$(CudaToolkitLibDir);</AdditionalLibraryDirectories>
|
||||
<OutputFile>$(OutDir)/globalToShmemAsyncCopy.exe</OutputFile>
|
||||
</Link>
|
||||
<CudaCompile>
|
||||
<CodeGeneration>compute_35,sm_35;compute_37,sm_37;compute_50,sm_50;compute_52,sm_52;compute_60,sm_60;compute_61,sm_61;compute_70,sm_70;compute_75,sm_75;compute_80,sm_80;</CodeGeneration>
|
||||
<AdditionalOptions>-Xcompiler "/wd 4819" %(AdditionalOptions)</AdditionalOptions>
|
||||
<Include>./;../../Common</Include>
|
||||
<Defines>WIN32</Defines>
|
||||
</CudaCompile>
|
||||
</ItemDefinitionGroup>
|
||||
<ItemDefinitionGroup Condition="'$(Configuration)'=='Debug'">
|
||||
<ClCompile>
|
||||
<Optimization>Disabled</Optimization>
|
||||
<RuntimeLibrary>MultiThreadedDebug</RuntimeLibrary>
|
||||
</ClCompile>
|
||||
<Link>
|
||||
<GenerateDebugInformation>true</GenerateDebugInformation>
|
||||
<LinkTimeCodeGeneration>Default</LinkTimeCodeGeneration>
|
||||
</Link>
|
||||
<CudaCompile>
|
||||
<Runtime>MTd</Runtime>
|
||||
<TargetMachinePlatform>64</TargetMachinePlatform>
|
||||
</CudaCompile>
|
||||
</ItemDefinitionGroup>
|
||||
<ItemDefinitionGroup Condition="'$(Configuration)'=='Release'">
|
||||
<ClCompile>
|
||||
<Optimization>MaxSpeed</Optimization>
|
||||
<RuntimeLibrary>MultiThreaded</RuntimeLibrary>
|
||||
</ClCompile>
|
||||
<Link>
|
||||
<GenerateDebugInformation>false</GenerateDebugInformation>
|
||||
<LinkTimeCodeGeneration>UseLinkTimeCodeGeneration</LinkTimeCodeGeneration>
|
||||
</Link>
|
||||
<CudaCompile>
|
||||
<Runtime>MT</Runtime>
|
||||
<TargetMachinePlatform>64</TargetMachinePlatform>
|
||||
</CudaCompile>
|
||||
</ItemDefinitionGroup>
|
||||
<ItemGroup>
|
||||
<CudaCompile Include="globalToShmemAsyncCopy.cu" />
|
||||
|
||||
</ItemGroup>
|
||||
<Import Project="$(VCTargetsPath)\Microsoft.Cpp.targets" />
|
||||
<ImportGroup Label="ExtensionTargets">
|
||||
<Import Project="$(CUDAPropsPath)\CUDA 10.2.targets" />
|
||||
</ImportGroup>
|
||||
</Project>
|
||||
@@ -0,0 +1,20 @@
|
||||
|
||||
Microsoft Visual Studio Solution File, Format Version 12.00
|
||||
# Visual Studio 2017
|
||||
Project("{8BC9CEB8-8B4A-11D0-8D11-00A0C91BC942}") = "globalToShmemAsyncCopy", "globalToShmemAsyncCopy_vs2017.vcxproj", "{997E0757-EA74-4A4E-A0FC-47D8C8831A15}"
|
||||
EndProject
|
||||
Global
|
||||
GlobalSection(SolutionConfigurationPlatforms) = preSolution
|
||||
Debug|x64 = Debug|x64
|
||||
Release|x64 = Release|x64
|
||||
EndGlobalSection
|
||||
GlobalSection(ProjectConfigurationPlatforms) = postSolution
|
||||
{997E0757-EA74-4A4E-A0FC-47D8C8831A15}.Debug|x64.ActiveCfg = Debug|x64
|
||||
{997E0757-EA74-4A4E-A0FC-47D8C8831A15}.Debug|x64.Build.0 = Debug|x64
|
||||
{997E0757-EA74-4A4E-A0FC-47D8C8831A15}.Release|x64.ActiveCfg = Release|x64
|
||||
{997E0757-EA74-4A4E-A0FC-47D8C8831A15}.Release|x64.Build.0 = Release|x64
|
||||
EndGlobalSection
|
||||
GlobalSection(SolutionProperties) = preSolution
|
||||
HideSolutionNode = FALSE
|
||||
EndGlobalSection
|
||||
EndGlobal
|
||||
@@ -0,0 +1,112 @@
|
||||
<?xml version="1.0" encoding="utf-8"?>
|
||||
<Project DefaultTargets="Build" ToolsVersion="4.0" xmlns="http://schemas.microsoft.com/developer/msbuild/2003">
|
||||
<PropertyGroup>
|
||||
<CUDAPropsPath Condition="'$(CUDAPropsPath)'==''">$(VCTargetsPath)\BuildCustomizations</CUDAPropsPath>
|
||||
</PropertyGroup>
|
||||
<ItemGroup Label="ProjectConfigurations">
|
||||
<ProjectConfiguration Include="Debug|x64">
|
||||
<Configuration>Debug</Configuration>
|
||||
<Platform>x64</Platform>
|
||||
</ProjectConfiguration>
|
||||
<ProjectConfiguration Include="Release|x64">
|
||||
<Configuration>Release</Configuration>
|
||||
<Platform>x64</Platform>
|
||||
</ProjectConfiguration>
|
||||
</ItemGroup>
|
||||
<PropertyGroup Label="Globals">
|
||||
<ProjectGuid>{997E0757-EA74-4A4E-A0FC-47D8C8831A15}</ProjectGuid>
|
||||
<RootNamespace>globalToShmemAsyncCopy_vs2017</RootNamespace>
|
||||
<ProjectName>globalToShmemAsyncCopy</ProjectName>
|
||||
<CudaToolkitCustomDir />
|
||||
</PropertyGroup>
|
||||
<PropertyGroup Condition="'$(WindowsTargetPlatformVersion)'==''">
|
||||
<LatestTargetPlatformVersion>$([Microsoft.Build.Utilities.ToolLocationHelper]::GetLatestSDKTargetPlatformVersion('Windows', '10.0'))</LatestTargetPlatformVersion>
|
||||
<WindowsTargetPlatformVersion Condition="'$(WindowsTargetPlatformVersion)' == ''">$(LatestTargetPlatformVersion)</WindowsTargetPlatformVersion>
|
||||
<TargetPlatformVersion>$(WindowsTargetPlatformVersion)</TargetPlatformVersion>
|
||||
</PropertyGroup>
|
||||
<Import Project="$(VCTargetsPath)\Microsoft.Cpp.Default.props" />
|
||||
<PropertyGroup>
|
||||
<ConfigurationType>Application</ConfigurationType>
|
||||
<CharacterSet>MultiByte</CharacterSet>
|
||||
<PlatformToolset>v141</PlatformToolset>
|
||||
</PropertyGroup>
|
||||
<PropertyGroup Condition="'$(Configuration)'=='Debug'">
|
||||
<UseDebugLibraries>true</UseDebugLibraries>
|
||||
</PropertyGroup>
|
||||
<PropertyGroup Condition="'$(Configuration)'=='Release'">
|
||||
<WholeProgramOptimization>true</WholeProgramOptimization>
|
||||
</PropertyGroup>
|
||||
<Import Project="$(VCTargetsPath)\Microsoft.Cpp.props" />
|
||||
<ImportGroup Label="ExtensionSettings">
|
||||
<Import Project="$(CUDAPropsPath)\CUDA 10.2.props" />
|
||||
</ImportGroup>
|
||||
<ImportGroup Label="PropertySheets">
|
||||
<Import Condition="exists('$(UserRootDir)\Microsoft.Cpp.$(Platform).user.props')" Label="LocalAppDataPlatform" Project="$(UserRootDir)\Microsoft.Cpp.$(Platform).user.props" />
|
||||
</ImportGroup>
|
||||
<PropertyGroup Label="UserMacros" />
|
||||
<PropertyGroup>
|
||||
<IntDir>$(Platform)/$(Configuration)/</IntDir>
|
||||
<IncludePath>$(IncludePath)</IncludePath>
|
||||
<CodeAnalysisRuleSet>AllRules.ruleset</CodeAnalysisRuleSet>
|
||||
<CodeAnalysisRules />
|
||||
<CodeAnalysisRuleAssemblies />
|
||||
</PropertyGroup>
|
||||
<PropertyGroup Condition="'$(Platform)'=='x64'">
|
||||
<OutDir>../../bin/win64/$(Configuration)/</OutDir>
|
||||
</PropertyGroup>
|
||||
<ItemDefinitionGroup>
|
||||
<ClCompile>
|
||||
<WarningLevel>Level3</WarningLevel>
|
||||
<PreprocessorDefinitions>WIN32;_MBCS;%(PreprocessorDefinitions)</PreprocessorDefinitions>
|
||||
<AdditionalIncludeDirectories>./;$(CudaToolkitDir)/include;../../Common;</AdditionalIncludeDirectories>
|
||||
</ClCompile>
|
||||
<Link>
|
||||
<SubSystem>Console</SubSystem>
|
||||
<AdditionalDependencies>cudart_static.lib;kernel32.lib;user32.lib;gdi32.lib;winspool.lib;comdlg32.lib;advapi32.lib;shell32.lib;ole32.lib;oleaut32.lib;uuid.lib;odbc32.lib;odbccp32.lib;%(AdditionalDependencies)</AdditionalDependencies>
|
||||
<AdditionalLibraryDirectories>$(CudaToolkitLibDir);</AdditionalLibraryDirectories>
|
||||
<OutputFile>$(OutDir)/globalToShmemAsyncCopy.exe</OutputFile>
|
||||
</Link>
|
||||
<CudaCompile>
|
||||
<CodeGeneration>compute_35,sm_35;compute_37,sm_37;compute_50,sm_50;compute_52,sm_52;compute_60,sm_60;compute_61,sm_61;compute_70,sm_70;compute_75,sm_75;compute_80,sm_80;</CodeGeneration>
|
||||
<AdditionalOptions>-Xcompiler "/wd 4819" %(AdditionalOptions)</AdditionalOptions>
|
||||
<Include>./;../../Common</Include>
|
||||
<Defines>WIN32</Defines>
|
||||
</CudaCompile>
|
||||
</ItemDefinitionGroup>
|
||||
<ItemDefinitionGroup Condition="'$(Configuration)'=='Debug'">
|
||||
<ClCompile>
|
||||
<Optimization>Disabled</Optimization>
|
||||
<RuntimeLibrary>MultiThreadedDebug</RuntimeLibrary>
|
||||
</ClCompile>
|
||||
<Link>
|
||||
<GenerateDebugInformation>true</GenerateDebugInformation>
|
||||
<LinkTimeCodeGeneration>Default</LinkTimeCodeGeneration>
|
||||
</Link>
|
||||
<CudaCompile>
|
||||
<Runtime>MTd</Runtime>
|
||||
<TargetMachinePlatform>64</TargetMachinePlatform>
|
||||
</CudaCompile>
|
||||
</ItemDefinitionGroup>
|
||||
<ItemDefinitionGroup Condition="'$(Configuration)'=='Release'">
|
||||
<ClCompile>
|
||||
<Optimization>MaxSpeed</Optimization>
|
||||
<RuntimeLibrary>MultiThreaded</RuntimeLibrary>
|
||||
</ClCompile>
|
||||
<Link>
|
||||
<GenerateDebugInformation>false</GenerateDebugInformation>
|
||||
<LinkTimeCodeGeneration>UseLinkTimeCodeGeneration</LinkTimeCodeGeneration>
|
||||
</Link>
|
||||
<CudaCompile>
|
||||
<Runtime>MT</Runtime>
|
||||
<TargetMachinePlatform>64</TargetMachinePlatform>
|
||||
</CudaCompile>
|
||||
</ItemDefinitionGroup>
|
||||
<ItemGroup>
|
||||
<CudaCompile Include="globalToShmemAsyncCopy.cu" />
|
||||
|
||||
</ItemGroup>
|
||||
<Import Project="$(VCTargetsPath)\Microsoft.Cpp.targets" />
|
||||
<ImportGroup Label="ExtensionTargets">
|
||||
<Import Project="$(CUDAPropsPath)\CUDA 10.2.targets" />
|
||||
</ImportGroup>
|
||||
</Project>
|
||||
@@ -0,0 +1,20 @@
|
||||
|
||||
Microsoft Visual Studio Solution File, Format Version 12.00
|
||||
# Visual Studio 2019
|
||||
Project("{8BC9CEB8-8B4A-11D0-8D11-00A0C91BC942}") = "globalToShmemAsyncCopy", "globalToShmemAsyncCopy_vs2019.vcxproj", "{997E0757-EA74-4A4E-A0FC-47D8C8831A15}"
|
||||
EndProject
|
||||
Global
|
||||
GlobalSection(SolutionConfigurationPlatforms) = preSolution
|
||||
Debug|x64 = Debug|x64
|
||||
Release|x64 = Release|x64
|
||||
EndGlobalSection
|
||||
GlobalSection(ProjectConfigurationPlatforms) = postSolution
|
||||
{997E0757-EA74-4A4E-A0FC-47D8C8831A15}.Debug|x64.ActiveCfg = Debug|x64
|
||||
{997E0757-EA74-4A4E-A0FC-47D8C8831A15}.Debug|x64.Build.0 = Debug|x64
|
||||
{997E0757-EA74-4A4E-A0FC-47D8C8831A15}.Release|x64.ActiveCfg = Release|x64
|
||||
{997E0757-EA74-4A4E-A0FC-47D8C8831A15}.Release|x64.Build.0 = Release|x64
|
||||
EndGlobalSection
|
||||
GlobalSection(SolutionProperties) = preSolution
|
||||
HideSolutionNode = FALSE
|
||||
EndGlobalSection
|
||||
EndGlobal
|
||||
@@ -0,0 +1,108 @@
|
||||
<?xml version="1.0" encoding="utf-8"?>
|
||||
<Project DefaultTargets="Build" ToolsVersion="4.0" xmlns="http://schemas.microsoft.com/developer/msbuild/2003">
|
||||
<PropertyGroup>
|
||||
<CUDAPropsPath Condition="'$(CUDAPropsPath)'==''">$(VCTargetsPath)\BuildCustomizations</CUDAPropsPath>
|
||||
</PropertyGroup>
|
||||
<ItemGroup Label="ProjectConfigurations">
|
||||
<ProjectConfiguration Include="Debug|x64">
|
||||
<Configuration>Debug</Configuration>
|
||||
<Platform>x64</Platform>
|
||||
</ProjectConfiguration>
|
||||
<ProjectConfiguration Include="Release|x64">
|
||||
<Configuration>Release</Configuration>
|
||||
<Platform>x64</Platform>
|
||||
</ProjectConfiguration>
|
||||
</ItemGroup>
|
||||
<PropertyGroup Label="Globals">
|
||||
<ProjectGuid>{997E0757-EA74-4A4E-A0FC-47D8C8831A15}</ProjectGuid>
|
||||
<RootNamespace>globalToShmemAsyncCopy_vs2019</RootNamespace>
|
||||
<ProjectName>globalToShmemAsyncCopy</ProjectName>
|
||||
<CudaToolkitCustomDir />
|
||||
</PropertyGroup>
|
||||
<Import Project="$(VCTargetsPath)\Microsoft.Cpp.Default.props" />
|
||||
<PropertyGroup>
|
||||
<ConfigurationType>Application</ConfigurationType>
|
||||
<CharacterSet>MultiByte</CharacterSet>
|
||||
<PlatformToolset>v142</PlatformToolset>
|
||||
<WindowsTargetPlatformVersion>10.0</WindowsTargetPlatformVersion>
|
||||
</PropertyGroup>
|
||||
<PropertyGroup Condition="'$(Configuration)'=='Debug'">
|
||||
<UseDebugLibraries>true</UseDebugLibraries>
|
||||
</PropertyGroup>
|
||||
<PropertyGroup Condition="'$(Configuration)'=='Release'">
|
||||
<WholeProgramOptimization>true</WholeProgramOptimization>
|
||||
</PropertyGroup>
|
||||
<Import Project="$(VCTargetsPath)\Microsoft.Cpp.props" />
|
||||
<ImportGroup Label="ExtensionSettings">
|
||||
<Import Project="$(CUDAPropsPath)\CUDA 10.2.props" />
|
||||
</ImportGroup>
|
||||
<ImportGroup Label="PropertySheets">
|
||||
<Import Condition="exists('$(UserRootDir)\Microsoft.Cpp.$(Platform).user.props')" Label="LocalAppDataPlatform" Project="$(UserRootDir)\Microsoft.Cpp.$(Platform).user.props" />
|
||||
</ImportGroup>
|
||||
<PropertyGroup Label="UserMacros" />
|
||||
<PropertyGroup>
|
||||
<IntDir>$(Platform)/$(Configuration)/</IntDir>
|
||||
<IncludePath>$(IncludePath)</IncludePath>
|
||||
<CodeAnalysisRuleSet>AllRules.ruleset</CodeAnalysisRuleSet>
|
||||
<CodeAnalysisRules />
|
||||
<CodeAnalysisRuleAssemblies />
|
||||
</PropertyGroup>
|
||||
<PropertyGroup Condition="'$(Platform)'=='x64'">
|
||||
<OutDir>../../bin/win64/$(Configuration)/</OutDir>
|
||||
</PropertyGroup>
|
||||
<ItemDefinitionGroup>
|
||||
<ClCompile>
|
||||
<WarningLevel>Level3</WarningLevel>
|
||||
<PreprocessorDefinitions>WIN32;_MBCS;%(PreprocessorDefinitions)</PreprocessorDefinitions>
|
||||
<AdditionalIncludeDirectories>./;$(CudaToolkitDir)/include;../../Common;</AdditionalIncludeDirectories>
|
||||
</ClCompile>
|
||||
<Link>
|
||||
<SubSystem>Console</SubSystem>
|
||||
<AdditionalDependencies>cudart_static.lib;kernel32.lib;user32.lib;gdi32.lib;winspool.lib;comdlg32.lib;advapi32.lib;shell32.lib;ole32.lib;oleaut32.lib;uuid.lib;odbc32.lib;odbccp32.lib;%(AdditionalDependencies)</AdditionalDependencies>
|
||||
<AdditionalLibraryDirectories>$(CudaToolkitLibDir);</AdditionalLibraryDirectories>
|
||||
<OutputFile>$(OutDir)/globalToShmemAsyncCopy.exe</OutputFile>
|
||||
</Link>
|
||||
<CudaCompile>
|
||||
<CodeGeneration>compute_35,sm_35;compute_37,sm_37;compute_50,sm_50;compute_52,sm_52;compute_60,sm_60;compute_61,sm_61;compute_70,sm_70;compute_75,sm_75;compute_80,sm_80;</CodeGeneration>
|
||||
<AdditionalOptions>-Xcompiler "/wd 4819" %(AdditionalOptions)</AdditionalOptions>
|
||||
<Include>./;../../Common</Include>
|
||||
<Defines>WIN32</Defines>
|
||||
</CudaCompile>
|
||||
</ItemDefinitionGroup>
|
||||
<ItemDefinitionGroup Condition="'$(Configuration)'=='Debug'">
|
||||
<ClCompile>
|
||||
<Optimization>Disabled</Optimization>
|
||||
<RuntimeLibrary>MultiThreadedDebug</RuntimeLibrary>
|
||||
</ClCompile>
|
||||
<Link>
|
||||
<GenerateDebugInformation>true</GenerateDebugInformation>
|
||||
<LinkTimeCodeGeneration>Default</LinkTimeCodeGeneration>
|
||||
</Link>
|
||||
<CudaCompile>
|
||||
<Runtime>MTd</Runtime>
|
||||
<TargetMachinePlatform>64</TargetMachinePlatform>
|
||||
</CudaCompile>
|
||||
</ItemDefinitionGroup>
|
||||
<ItemDefinitionGroup Condition="'$(Configuration)'=='Release'">
|
||||
<ClCompile>
|
||||
<Optimization>MaxSpeed</Optimization>
|
||||
<RuntimeLibrary>MultiThreaded</RuntimeLibrary>
|
||||
</ClCompile>
|
||||
<Link>
|
||||
<GenerateDebugInformation>false</GenerateDebugInformation>
|
||||
<LinkTimeCodeGeneration>UseLinkTimeCodeGeneration</LinkTimeCodeGeneration>
|
||||
</Link>
|
||||
<CudaCompile>
|
||||
<Runtime>MT</Runtime>
|
||||
<TargetMachinePlatform>64</TargetMachinePlatform>
|
||||
</CudaCompile>
|
||||
</ItemDefinitionGroup>
|
||||
<ItemGroup>
|
||||
<CudaCompile Include="globalToShmemAsyncCopy.cu" />
|
||||
|
||||
</ItemGroup>
|
||||
<Import Project="$(VCTargetsPath)\Microsoft.Cpp.targets" />
|
||||
<ImportGroup Label="ExtensionTargets">
|
||||
<Import Project="$(CUDAPropsPath)\CUDA 10.2.targets" />
|
||||
</ImportGroup>
|
||||
</Project>
|
||||
Reference in New Issue
Block a user