Add and update samples for cuda 11.0 support

2026-10-23 21:28:17 +08:00 · 2020-05-18 22:22:06 +05:30
parent 6be514679b
commit 4f6e02970b
410 changed files with 21920 additions and 3864 deletions
--- a/Samples/warpAggregatedAtomicsCG/Makefile
+++ b/Samples/warpAggregatedAtomicsCG/Makefile
@@ -72,9 +72,9 @@ endif
 # architecture
 HOST_ARCH   := $(shell uname -m)
 TARGET_ARCH ?= $(HOST_ARCH)
-ifneq (,$(filter $(TARGET_ARCH),x86_64 aarch64 ppc64le armv7l))
+ifneq (,$(filter $(TARGET_ARCH),x86_64 aarch64 sbsa ppc64le armv7l))
    ifneq ($(TARGET_ARCH),$(HOST_ARCH))
-        ifneq (,$(filter $(TARGET_ARCH),x86_64 aarch64 ppc64le))
+        ifneq (,$(filter $(TARGET_ARCH),x86_64 aarch64 sbsa ppc64le))
            TARGET_SIZE := 64
        else ifneq (,$(filter $(TARGET_ARCH),armv7l))
            TARGET_SIZE := 32
@@ -85,8 +85,17 @@ ifneq (,$(filter $(TARGET_ARCH),x86_64 aarch64 ppc64le armv7l))
 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-ppc64le))
+    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 $(TARGET_ARCH) is not supported!)
    endif
 endif
@@ -141,6 +150,8 @@ else ifneq ($(TARGET_ARCH),$(HOST_ARCH))
        else ifeq ($(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
@@ -185,19 +196,27 @@ ifneq ($(TARGET_ARCH),$(HOST_ARCH))
                CCFLAGS += --sysroot=$(TARGET_FS)
            endif
            LDFLAGS += --sysroot=$(TARGET_FS)
-            LDFLAGS += -rpath-link=$(TARGET_FS)/lib -L $(TARGET_FS)/lib
-            LDFLAGS += -rpath-link=$(TARGET_FS)/usr/lib -L $(TARGET_FS)/usr/lib
-            LDFLAGS += -rpath-link=$(TARGET_FS)/usr/lib/aarch64-linux-gnu -L $(TARGET_FS)/usr/lib/aarch64-linux-gnu
+            LDFLAGS += -rpath-link=$(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
-            CCFLAGS += -isystem=$(TARGET_FS)/usr/include/aarch64-linux-gnu
+            CCFLAGS += -isystem=$(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 $(TARGET_OVERRIDE)
        endif
    endif
-endif
-
-ifeq ($(TARGET_OS),qnx)
-    CCFLAGS += -DWIN_INTERFACE_CUSTOM
-    LDFLAGS += -lsocket
 endif

 # Install directory of different arch
@@ -206,6 +225,8 @@ 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)
@@ -245,11 +266,38 @@ 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 $(GCCVERSION)` \>= 470)
+
+    ifeq ($(IS_MIN_VERSION), 1)
+        $(info >>> GCC Version is greater or equal to 4.7.0 <<<)
+    else
+        $(info >>> Waiving build. Minimum GCC version required is 4.7.0<<<)
+        SAMPLE_ENABLED := 0
+    endif
+endif
+
 # Gencode arguments
 ifeq ($(TARGET_ARCH),$(filter $(TARGET_ARCH),armv7l aarch64))
-SMS ?= 30 35 37 50 52 60 61 70 72 75
+SMS ?= 35 37 50 52 60 61 70 72 75 80
 else
-SMS ?= 30 35 37 50 52 60 61 70 75
+SMS ?= 35 37 50 52 60 61 70 75 80
 endif

 ifeq ($(SMS),)
@@ -268,6 +316,8 @@ GENCODE_FLAGS += -gencode arch=compute_$(HIGHEST_SM),code=compute_$(HIGHEST_SM)
 endif
 endif

+ALL_CCFLAGS += --std=c++11
+
 ifeq ($(SAMPLE_ENABLED),0)
 EXEC ?= @echo "[@]"
 endif
--- a/Samples/warpAggregatedAtomicsCG/NsightEclipse.xml
+++ b/Samples/warpAggregatedAtomicsCG/NsightEclipse.xml
@@ -2,7 +2,10 @@
 <!DOCTYPE entry SYSTEM "SamplesInfo.dtd">
 <entry>
  <name>warpAggregatedAtomicsCG</name>
-  <description><![CDATA[This sample demonstrates how using Cooperative Groups (CG) to perform warp aggregated atomics, a useful technique to improve performance when many threads atomically add to a single counter.]]></description>
+  <cflags>
+    <flag>--std=c++11</flag>
+  </cflags>
+  <description><![CDATA[This sample demonstrates how using Cooperative Groups (CG) to perform warp aggregated atomics to single and multiple counters, a useful technique to improve performance when many threads atomically add to a single or multiple counters.]]></description>
  <includepaths>
    <path>./</path>
    <path>../</path>
@@ -16,6 +19,7 @@
    <keyword>GPGPU</keyword>
    <keyword>Cooperative Groups</keyword>
    <keyword>Atomic</keyword>
+    <keyword>CPP11</keyword>
  </keywords>
  <libraries>
  </libraries>
@@ -26,16 +30,6 @@
  <scopes>
    <scope>1:CUDA Advanced Topics</scope>
  </scopes>
-  <sm-arch>sm30</sm-arch>
-  <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>
  <supported_envs>
    <env>
      <arch>x86_64</arch>
--- a/Samples/warpAggregatedAtomicsCG/README.md
+++ b/Samples/warpAggregatedAtomicsCG/README.md
@@ -2,7 +2,7 @@

 ## Description

-This sample demonstrates how using Cooperative Groups (CG) to perform warp aggregated atomics, a useful technique to improve performance when many threads atomically add to a single counter.
+This sample demonstrates how using Cooperative Groups (CG) to perform warp aggregated atomics to single and multiple counters, a useful technique to improve performance when many threads atomically add to a single or multiple counters.

 ## Key Concepts

@@ -10,11 +10,9 @@ Cooperative Groups, Atomic Intrinsics

 ## Supported SM Architectures

-[SM 3.0 ](https://developer.nvidia.com/cuda-gpus)  [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)
-
 ## Supported OSes

-Linux, Windows, MacOSX
+Linux, Windows

 ## Supported CPU Architecture

@@ -24,7 +22,7 @@ x86_64, ppc64le, armv7l, aarch64

 ## Prerequisites

-Download and install the [CUDA Toolkit 10.2](https://developer.nvidia.com/cuda-downloads) for your corresponding platform.
+Download and install the [CUDA Toolkit 11.0](https://developer.nvidia.com/cuda-downloads) for your corresponding platform.

 ## Build and Run

@@ -63,29 +61,5 @@ The samples makefiles can take advantage of certain options:
    $ make HOST_COMPILER=g++
 ```

-### Mac
-The Mac samples are built using makefiles. To use the makefiles, change directory into the sample directory you wish to build, and run make:
-```
-$ cd <sample_dir>
-$ make
-```
-
-The samples makefiles can take advantage of certain options:
-
-*  **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="A B ..."
-    ```
-
-*  **HOST_COMPILER=<host_compiler>** - override the default clang host compiler. See the [Mac Installation Guide](http://docs.nvidia.com/cuda/cuda-installation-guide-mac-os-x/index.html#system-requirements) for a list of supported host compilers.
-    ```
-    $ make HOST_COMPILER=clang
-    ```
-
 ## References (for more details)

--- a/Samples/warpAggregatedAtomicsCG/warpAggregatedAtomicsCG.cu
+++ b/Samples/warpAggregatedAtomicsCG/warpAggregatedAtomicsCG.cu
@@ -43,21 +43,17 @@ namespace cg = cooperative_groups;
 __device__ int atomicAggInc(int *counter) {
  cg::coalesced_group active = cg::coalesced_threads();

-  int mask = active.ballot(1);
-  // select the leader
-  int leader = __ffs(mask) - 1;
-
  // leader does the update
  int res = 0;
-  if (active.thread_rank() == leader) {
-    res = atomicAdd(counter, __popc(mask));
+  if (active.thread_rank() == 0) {
+    res = atomicAdd(counter, active.size());
  }

  // broadcast result
-  res = active.shfl(res, leader);
+  res = active.shfl(res, 0);

  // each thread computes its own value
-  return res + __popc(mask & ((1 << active.thread_rank()) - 1));
+  return res + active.thread_rank();
 }

 __global__ void filter_arr(int *dst, int *nres, const int *src, int n) {
@@ -68,18 +64,108 @@ __global__ void filter_arr(int *dst, int *nres, const int *src, int n) {
  }
 }

+// warp-aggregated atomic multi bucket increment
+#if __CUDA_ARCH__ >= 700
+__device__ int atomicAggIncMulti(const int bucket, int *counter)
+{
+  cg::coalesced_group active = cg::coalesced_threads();
+  // group all threads with same bucket value.
+  auto labeledGroup = cg::labeled_partition(active, bucket);
+
+  int res = 0;
+  if (labeledGroup.thread_rank() == 0)
+  {
+    res = atomicAdd(&counter[bucket], labeledGroup.size());
+  }
+
+  // broadcast result
+  res = labeledGroup.shfl(res, 0);
+
+  // each thread computes its own value
+  return res + labeledGroup.thread_rank();
+}
+#endif
+
+// Places individual value indices into its corresponding buckets.
+__global__ void mapToBuckets(const int *srcArr, int *indicesBuckets, int *bucketCounters, const int srcSize, const int numOfBuckets)
+{
+#if __CUDA_ARCH__ >= 700
+  cg::grid_group grid = cg::this_grid();
+
+  for (int i=grid.thread_rank(); i < srcSize; i += grid.size())
+  {
+    const int bucket = srcArr[i];
+    if (bucket < numOfBuckets)
+    {
+      indicesBuckets[atomicAggIncMulti(bucket, bucketCounters)] = i;
+    }
+  }
+#endif
+}
+
+int mapIndicesToBuckets(int *h_srcArr, int *d_srcArr, int numOfBuckets)
+{
+  int *d_indicesBuckets, *d_bucketCounters;
+  int *cpuBucketCounters = new int[numOfBuckets];
+  int *h_bucketCounters = new int[numOfBuckets];
+
+  memset(cpuBucketCounters, 0, sizeof(int)*numOfBuckets);
+  // Initialize each bucket counters.
+  for (int i = 0; i < numOfBuckets; i++)
+  {
+    h_bucketCounters[i] = i*NUM_ELEMS;
+  }
+
+  checkCudaErrors(cudaMalloc(&d_indicesBuckets, sizeof(int) * NUM_ELEMS * numOfBuckets));
+  checkCudaErrors(cudaMalloc(&d_bucketCounters, sizeof(int) * numOfBuckets));
+
+  checkCudaErrors(cudaMemcpy(d_bucketCounters, h_bucketCounters, sizeof(int)*numOfBuckets, cudaMemcpyHostToDevice));
+
+  dim3 dimBlock(NUM_THREADS_PER_BLOCK, 1, 1);
+  dim3 dimGrid((NUM_ELEMS / NUM_THREADS_PER_BLOCK), 1, 1);
+
+  mapToBuckets<<<dimGrid, dimBlock>>>(d_srcArr, d_indicesBuckets, d_bucketCounters, NUM_ELEMS, numOfBuckets);
+
+  checkCudaErrors(cudaMemcpy(h_bucketCounters, d_bucketCounters, sizeof(int)*numOfBuckets, cudaMemcpyDeviceToHost));
+
+  for (int i=0; i < NUM_ELEMS; i++)
+  {
+    cpuBucketCounters[h_srcArr[i]]++;
+  }
+
+  bool allMatch = true;
+  int finalElems = 0;
+  for (int i=0; i < numOfBuckets; i++)
+  {
+    finalElems += (h_bucketCounters[i] - i*NUM_ELEMS);
+    if (cpuBucketCounters[i] != (h_bucketCounters[i] - i*NUM_ELEMS))
+    {
+      allMatch = false;
+      break;
+    }
+  }
+
+  if (!allMatch && finalElems != NUM_ELEMS)
+  {
+      return EXIT_FAILURE;
+  }
+  return EXIT_SUCCESS;
+}
+
 int main(int argc, char **argv) {
  int *data_to_filter, *filtered_data, nres = 0;
  int *d_data_to_filter, *d_filtered_data, *d_nres;

+  int numOfBuckets = 5;
+
  data_to_filter = reinterpret_cast<int *>(malloc(sizeof(int) * NUM_ELEMS));

  // Generate input data.
  for (int i = 0; i < NUM_ELEMS; i++) {
-    data_to_filter[i] = rand() % 20;
+    data_to_filter[i] = rand() % numOfBuckets;
  }

-  findCudaDevice(argc, (const char **)argv);
+  int devId = findCudaDevice(argc, (const char **)argv);

  checkCudaErrors(cudaMalloc(&d_data_to_filter, sizeof(int) * NUM_ELEMS));
  checkCudaErrors(cudaMalloc(&d_filtered_data, sizeof(int) * NUM_ELEMS));
@@ -114,8 +200,18 @@ int main(int argc, char **argv) {
    }
  }

+  int major = 0;
+  checkCudaErrors(cudaDeviceGetAttribute(&major, cudaDevAttrComputeCapabilityMajor, devId));
+
+  int mapIndicesToBucketsStatus = EXIT_SUCCESS;
+  // atomicAggIncMulti require a GPU of Volta (SM7X) architecture or higher,
+  // so that it can take advantage of the new MATCH capability of Volta hardware
+  if (major >= 7) {
+    mapIndicesToBucketsStatus = mapIndicesToBuckets(data_to_filter, d_data_to_filter, numOfBuckets);
+  }
+
  printf("\nWarp Aggregated Atomics %s \n",
-         host_flt_count == nres ? "PASSED" : "FAILED");
+         (host_flt_count == nres) && (mapIndicesToBucketsStatus == EXIT_SUCCESS) ? "PASSED" : "FAILED");

  checkCudaErrors(cudaFree(d_data_to_filter));
  checkCudaErrors(cudaFree(d_filtered_data));
--- a/Samples/warpAggregatedAtomicsCG/warpAggregatedAtomicsCG_vs2015.vcxproj
+++ b/Samples/warpAggregatedAtomicsCG/warpAggregatedAtomicsCG_vs2015.vcxproj
@@ -62,7 +62,7 @@
      <OutputFile>$(OutDir)/warpAggregatedAtomicsCG.exe</OutputFile>
    </Link>
    <CudaCompile>
-      <CodeGeneration>compute_30,sm_30;compute_35,sm_35;compute_37,sm_37;compute_50,sm_50;compute_52,sm_52;compute_60,sm_60;compute_61,sm_61;compute_70,sm_70;compute_75,sm_75;</CodeGeneration>
+      <CodeGeneration>compute_35,sm_35;compute_37,sm_37;compute_50,sm_50;compute_52,sm_52;compute_60,sm_60;compute_61,sm_61;compute_70,sm_70;compute_75,sm_75;compute_80,sm_80;</CodeGeneration>
      <AdditionalOptions>-Xcompiler "/wd 4819" %(AdditionalOptions)</AdditionalOptions>
      <Include>./;../../Common</Include>
      <Defines>WIN32</Defines>
--- a/Samples/warpAggregatedAtomicsCG/warpAggregatedAtomicsCG_vs2017.vcxproj
+++ b/Samples/warpAggregatedAtomicsCG/warpAggregatedAtomicsCG_vs2017.vcxproj
@@ -67,7 +67,7 @@
      <OutputFile>$(OutDir)/warpAggregatedAtomicsCG.exe</OutputFile>
    </Link>
    <CudaCompile>
-      <CodeGeneration>compute_30,sm_30;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;</CodeGeneration>
+      <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>
--- a/Samples/warpAggregatedAtomicsCG/warpAggregatedAtomicsCG_vs2019.vcxproj
+++ b/Samples/warpAggregatedAtomicsCG/warpAggregatedAtomicsCG_vs2019.vcxproj
@@ -63,7 +63,7 @@
      <OutputFile>$(OutDir)/warpAggregatedAtomicsCG.exe</OutputFile>
    </Link>
    <CudaCompile>
-      <CodeGeneration>compute_30,sm_30;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;</CodeGeneration>
+      <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>