NASP/cuda-samples
85
0
Fork 0
mirror of https://github.com/NVIDIA/cuda-samples.git synced 2026-10-23 21:28:17 +08:00
Code Issues Packages Projects Releases Wiki Activity

Add and update samples with CUDA 11.2 support

Browse Source
This commit is contained in:
Mahesh Doijade
2020-12-10 01:05:32 +05:30
parent 92b0568792
commit 1cd3264681
355 changed files with 30425 additions and 7495 deletions
Download Patch File Download Diff File Expand all files Collapse all files

6
Samples/tf32TensorCoreGemm/Makefile
View File

@@ -301,11 +301,9 @@ ifeq ($(TARGET_OS),linux)
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)
IS_MIN_VERSION := $(shell expr `echo $(GCCVERSION)` \>= 50000)
ifeq ($(IS_MIN_VERSION), 1)
$(info >>> GCC Version is greater or equal to 5.0.0 <<<)
@@ -334,7 +332,7 @@ GENCODE_FLAGS += -gencode arch=compute_$(HIGHEST_SM),code=compute_$(HIGHEST_SM)
endif
endif
ALL_CCFLAGS += --std=c++11
ALL_CCFLAGS += --std=c++11 --threads 0
ifeq ($(SAMPLE_ENABLED),0)
EXEC ?= @echo "[@]"

2
Samples/tf32TensorCoreGemm/README.md
View File

@@ -27,7 +27,7 @@ cudaMalloc, cudaDeviceSynchronize, cudaFuncSetAttribute, cudaEventCreate, cudaEv
## Prerequisites
Download and install the [CUDA Toolkit 11.1](https://developer.nvidia.com/cuda-downloads) for your corresponding platform.
Download and install the [CUDA Toolkit 11.2](https://developer.nvidia.com/cuda-downloads) for your corresponding platform.
## Build and Run

100
Samples/tf32TensorCoreGemm/tf32TensorCoreGemm.cu
View File

@@ -61,7 +61,7 @@
#include <stdio.h>
#include <cuda.h>
#include <mma.h>
#include <cuda_pipeline.h>
#include <cuda/pipeline>
// helper functions and utilities to work with CUDA
#include <helper_functions.h>
@@ -69,10 +69,6 @@
// Externally configurable parameters.
// Switch for choosing cpp interface for cuda pipeline
// vs primitives interface.
#define USE_CPP_API 0
#ifndef CPU_DEBUG
// Set this to 1 to verify the correctness of the GPU-computed matrix.
#define CPU_DEBUG 0
@@ -182,7 +178,6 @@ const char* kernelNames[] = {"compute_tf32gemm_async_copy", "compute_tf32gemm",
"simple_wmma_tf32gemm"};
using namespace nvcuda;
namespace nvcuda_namespace = nvcuda::experimental;
__host__ void init_host_matrices(float *a, float *b, float *c)
{
@@ -384,7 +379,7 @@ __global__ void compute_tf32gemm(const float *A, const float *B, const float *C,
#endif
}
__global__ void compute_tf32gemm_async_copy(const float *A, const float *B, const float *C, float *D, float alpha, float beta)
__global__ void compute_tf32gemm_async_copy(const float *A, const float *B, const float *C, float *D, const float alpha, float beta)
{
#if __CUDA_ARCH__ >= 800
extern __shared__ float shmem[][CHUNK_K * K + SKEW_FLOAT];
@@ -393,23 +388,24 @@ __global__ void compute_tf32gemm_async_copy(const float *A, const float *B, cons
const unsigned int warpId = threadIdx.x / WARP_SIZE;
const unsigned int laneId = threadIdx.x % WARP_SIZE;
// Offset in shared memory from which the B matrix is stored.
const size_t shmem_idx_b_off = BLOCK_COL_TILES * M;
// This pointer is used to access the C and D matrix tiles this warp computes.
float *shmem_warp_tile_ptr = (float*)&shmem[0][0] + (warpId / BLOCK_ROW_WARPS) * SHMEM_STRIDE * N * BLOCK_ROW_WARPS + (warpId % BLOCK_ROW_WARPS) * SHMEM_OFFSET;
// This pointer is used to stream the C and D matrices block-wide tile to and from shared memory.
float *shmem_warp_stream_ptr = (float*)&shmem[0][0] + warpId * SHMEM_STRIDE * N;
// Offset in shared memory from which the B matrix is stored.
constexpr size_t shmem_idx_b_off = BLOCK_COL_TILES * M;
// Adjust the beta scaler, as it'll be multiplied by alpha at the end of
// each tile computation. Technically this is not generally correct (may result
// in a loss of precision). Zero still needs to be specially handled though.
beta /= alpha;
#if USE_CPP_API
nvcuda_namespace::pipeline pipe;
#endif
cuda::pipeline<cuda::thread_scope_thread> pipe = cuda::make_pipeline();
const auto shape4 = cuda::aligned_size_t<alignof(float4)>(sizeof(float4));
constexpr int loadStride = 2; // load 4 floats, so left-shift by 2.
// Each CTA slides along the 128 x 128 tiles from the top left corner of the matrix to the
// right and down, and selects the next tile to compute. Once there's no such tile,
// all warps in this CTA exit.
@@ -429,24 +425,14 @@ __global__ void compute_tf32gemm_async_copy(const float *A, const float *B, cons
// Stream multiple C tiles to shared memory.
#pragma unroll
for (int i = 0; i < N; i++) {
#if USE_CPP_API
nvcuda_namespace::memcpy_async(*((int4*)(shmem_warp_stream_ptr + SHMEM_STRIDE * i) + laneId),
*((int4*)(src_gmem_warp_stream_ptr + GLOBAL_MEM_STRIDE * i) + laneId),
pipe);
pipe.commit();
#else
__pipeline_memcpy_async((reinterpret_cast<int4*>(&shmem_warp_stream_ptr[(SHMEM_STRIDE * i)])) + laneId,
(reinterpret_cast<const int4*>(&src_gmem_warp_stream_ptr[(GLOBAL_MEM_STRIDE * i)])) + laneId,
sizeof(int4));
__pipeline_commit();
#endif
pipe.producer_acquire();
cuda::memcpy_async(&shmem_warp_stream_ptr[(SHMEM_STRIDE * i) + (laneId << loadStride)],
&src_gmem_warp_stream_ptr[(GLOBAL_MEM_STRIDE * i) + (laneId << loadStride)],
shape4, pipe);
pipe.producer_commit();
}
#if USE_CPP_API
pipe.wait_prior<0>();
#else
__pipeline_wait_prior(0);
#endif
// Now wait for all the above issued 8 batches to complete.
cuda::pipeline_consumer_wait_prior<0>(pipe);
__syncthreads();
// These fragments will accumulate the result of A and B matrix fragment multiplications
@@ -461,65 +447,53 @@ __global__ void compute_tf32gemm_async_copy(const float *A, const float *B, cons
const float *tile_ptr = shmem_warp_tile_ptr + i * SHMEM_STRIDE * N + j * N;
wmma::load_matrix_sync(c[i][j], tile_ptr, SHMEM_STRIDE, C_LAYOUT);
}
}
__syncthreads();
// Scale the C matrix.
#pragma unroll
for (int i = 0; i < WARP_COL_TILES; i++) {
#pragma unroll
for (int j = 0; j < WARP_ROW_TILES; j++) {
// Scale the C matrix.
#pragma unroll
for (int t = 0; t < c[i][j].num_elements; t++) {
c[i][j].x[t] *= beta;
}
}
}
pipe.consumer_release();
// sync here so that shared memory can then be used for loading A & B matrices.
__syncthreads();
// Select what warp copies what matrix to shared memory.
// Warps 0-3 copy the A matrix, warps 4-7 copy the B matrix.
const float *warp_ptr = (warpId < (WARPS_PER_BLOCK/2)) ? (&A[block_tile_i * M * K_GLOBAL] + M * K_GLOBAL * (warpId % (WARPS_PER_BLOCK/2)) * 2) :
(&B[block_tile_j * N * K_GLOBAL] + N * K_GLOBAL * (warpId % (WARPS_PER_BLOCK/2)) * 2);
constexpr int chunksPerLane = ((WARP_SIZE/2) / CHUNK_COPY_LINES_PER_WARP) * 2;
const int laneLoadElem = (laneId % CHUNK_COPY_LINE_LANES) << loadStride;
const int stridePerLaneCopy = (laneId / CHUNK_COPY_LINE_LANES);
// Go through the global K dimension by a fixed step at a time.
#pragma unroll
for (int tile_k = 0; tile_k < K_TILES; tile_k += CHUNK_K) {
// Copy slices of the A and B matrices to shared memory.
// The first half of the warps in the CTA copy the A matrix, the rest copy the B matrix.
size_t shmem_idx = warpId < (WARPS_PER_BLOCK/2) ? (M * (warpId % (WARPS_PER_BLOCK/2)) * 2) :
(N * (warpId % (WARPS_PER_BLOCK/2)) * 2 + shmem_idx_b_off);
// As for tf32 MMA M == N we use M for warp 4-7 + shmem_idx_b_off.
size_t shmem_idx = (M * (warpId % (WARPS_PER_BLOCK/2)) * 2) + ((warpId / (WARPS_PER_BLOCK/2)) * shmem_idx_b_off);
// First half of the warp copies the first row / column of the matrix,
// the second half of the warp copies the next.
const float *lane_ptr = (warp_ptr + tile_k * K + (laneId / CHUNK_COPY_LINE_LANES) * K_GLOBAL);
const float *lane_ptr = (warp_ptr + tile_k * K + stridePerLaneCopy * K_GLOBAL + laneLoadElem);
// Shift the second half of the warp to the next row / column in the shared memory.
shmem_idx += laneId / CHUNK_COPY_LINE_LANES;
shmem_idx += stridePerLaneCopy;
#pragma unroll
for(int i = 0; i < ((WARP_SIZE/2) / CHUNK_COPY_LINES_PER_WARP) * 2; i++) {
for(int i = 0; i < chunksPerLane; i++) {
// Copy 16 bytes at once in each lane.
#if USE_CPP_API
nvcuda_namespace::memcpy_async(*((int4*)&shmem[shmem_idx][0] + (laneId % CHUNK_COPY_LINE_LANES)),
*((int4*)lane_ptr + (laneId % CHUNK_COPY_LINE_LANES)), pipe);
pipe.commit();
#else
__pipeline_memcpy_async((int4*)&shmem[shmem_idx][0] + (laneId % CHUNK_COPY_LINE_LANES),
(int4*)lane_ptr + (laneId % CHUNK_COPY_LINE_LANES), sizeof(int4));
__pipeline_commit();
#endif
pipe.producer_acquire();
cuda::memcpy_async(&shmem[shmem_idx][laneLoadElem], lane_ptr, shape4, pipe);
pipe.producer_commit();
// Advance the global memory pointer and the shared memory index.
lane_ptr = lane_ptr + K_GLOBAL * CHUNK_COPY_LINES_PER_WARP;
shmem_idx += CHUNK_COPY_LINES_PER_WARP;
}
#if USE_CPP_API
pipe.wait_prior<0>();
#else
__pipeline_wait_prior(0);
#endif
cuda::pipeline_consumer_wait_prior<0>(pipe);
__syncthreads();
// Compute a grid of C matrix tiles in each warp.
@@ -558,7 +532,7 @@ __global__ void compute_tf32gemm_async_copy(const float *A, const float *B, cons
}
}
}
pipe.consumer_release();
__syncthreads();
}
@@ -586,8 +560,8 @@ __global__ void compute_tf32gemm_async_copy(const float *A, const float *B, cons
#pragma unroll
for (int i = 0; i < N; i++) {
*((int4*)(dst_gmem_warp_stream_ptr + GLOBAL_MEM_STRIDE * i) + laneId) =
*((int4*)(shmem_warp_stream_ptr + SHMEM_STRIDE * i) + laneId);
*((float4*)(dst_gmem_warp_stream_ptr + GLOBAL_MEM_STRIDE * i) + laneId) =
*((float4*)(shmem_warp_stream_ptr + SHMEM_STRIDE * i) + laneId);
}
__syncthreads();

20
Samples/tf32TensorCoreGemm/tf32TensorCoreGemm_vs2015.sln
View File

@@ -1,20 +0,0 @@

Microsoft Visual Studio Solution File, Format Version 14.00
# Visual Studio 2015
Project("{8BC9CEB8-8B4A-11D0-8D11-00A0C91BC942}") = "tf32TensorCoreGemm", "tf32TensorCoreGemm_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

107
Samples/tf32TensorCoreGemm/tf32TensorCoreGemm_vs2015.vcxproj
View File

@@ -1,107 +0,0 @@
<?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>tf32TensorCoreGemm_vs2015</RootNamespace>
<ProjectName>tf32TensorCoreGemm</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="$(CUDAPropsPath)\CUDA 11.1.props" />
</ImportGroup>
<ImportGroup Label="PropertySheets">
<Import Condition="exists('$(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>$(OutDir)/tf32TensorCoreGemm.exe</OutputFile>
</Link>
<CudaCompile>
<CodeGeneration>compute_80,sm_80;compute_86,sm_86;</CodeGeneration>
<AdditionalOptions>-Xcompiler "/wd 4819" %(AdditionalOptions)</AdditionalOptions>
<Include>./;../../Common</Include>
<Defines>WIN32</Defines>
</CudaCompile>
</ItemDefinitionGroup>
<ItemDefinitionGroup Condition="'$(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="'$(Configuration)'=='Release'">
<ClCompile>
<Optimization>MaxSpeed</Optimization>
<RuntimeLibrary>MultiThreaded</RuntimeLibrary>
</ClCompile>
<Link>
<GenerateDebugInformation>false</GenerateDebugInformation>
<LinkTimeCodeGeneration>UseLinkTimeCodeGeneration</LinkTimeCodeGeneration>
</Link>
<CudaCompile>
<Runtime>MT</Runtime>
<TargetMachinePlatform>64</TargetMachinePlatform>
</CudaCompile>
</ItemDefinitionGroup>
<ItemGroup>
<CudaCompile Include="tf32TensorCoreGemm.cu" />
</ItemGroup>
<Import Project="$(VCTargetsPath)\Microsoft.Cpp.targets" />
<ImportGroup Label="ExtensionTargets">
<Import Project="$(CUDAPropsPath)\CUDA 11.1.targets" />
</ImportGroup>
</Project>

5
Samples/tf32TensorCoreGemm/tf32TensorCoreGemm_vs2017.vcxproj
View File

@@ -38,7 +38,7 @@
</PropertyGroup>
<Import Project="$(VCTargetsPath)\Microsoft.Cpp.props" />
<ImportGroup Label="ExtensionSettings">
<Import Project="$(CUDAPropsPath)\CUDA 11.1.props" />
<Import Project="$(CUDAPropsPath)\CUDA 11.2.props" />
</ImportGroup>
<ImportGroup Label="PropertySheets">
<Import Condition="exists('$(UserRootDir)\Microsoft.Cpp.$(Platform).user.props')" Label="LocalAppDataPlatform" Project="$(UserRootDir)\Microsoft.Cpp.$(Platform).user.props" />
@@ -71,6 +71,7 @@
<AdditionalOptions>-Xcompiler "/wd 4819" %(AdditionalOptions)</AdditionalOptions>
<Include>./;../../Common</Include>
<Defines>WIN32</Defines>
<AdditionalOptions>--threads 0</AdditionalOptions>
</CudaCompile>
</ItemDefinitionGroup>
<ItemDefinitionGroup Condition="'$(Configuration)'=='Debug'">
@@ -107,6 +108,6 @@
</ItemGroup>
<Import Project="$(VCTargetsPath)\Microsoft.Cpp.targets" />
<ImportGroup Label="ExtensionTargets">
<Import Project="$(CUDAPropsPath)\CUDA 11.1.targets" />
<Import Project="$(CUDAPropsPath)\CUDA 11.2.targets" />
</ImportGroup>
</Project>

5
Samples/tf32TensorCoreGemm/tf32TensorCoreGemm_vs2019.vcxproj
View File

@@ -34,7 +34,7 @@
</PropertyGroup>
<Import Project="$(VCTargetsPath)\Microsoft.Cpp.props" />
<ImportGroup Label="ExtensionSettings">
<Import Project="$(CUDAPropsPath)\CUDA 11.1.props" />
<Import Project="$(CUDAPropsPath)\CUDA 11.2.props" />
</ImportGroup>
<ImportGroup Label="PropertySheets">
<Import Condition="exists('$(UserRootDir)\Microsoft.Cpp.$(Platform).user.props')" Label="LocalAppDataPlatform" Project="$(UserRootDir)\Microsoft.Cpp.$(Platform).user.props" />
@@ -67,6 +67,7 @@
<AdditionalOptions>-Xcompiler "/wd 4819" %(AdditionalOptions)</AdditionalOptions>
<Include>./;../../Common</Include>
<Defines>WIN32</Defines>
<AdditionalOptions>--threads 0</AdditionalOptions>
</CudaCompile>
</ItemDefinitionGroup>
<ItemDefinitionGroup Condition="'$(Configuration)'=='Debug'">
@@ -103,6 +104,6 @@
</ItemGroup>
<Import Project="$(VCTargetsPath)\Microsoft.Cpp.targets" />
<ImportGroup Label="ExtensionTargets">
<Import Project="$(CUDAPropsPath)\CUDA 11.1.targets" />
<Import Project="$(CUDAPropsPath)\CUDA 11.2.targets" />
</ImportGroup>
</Project>