CUDA 13.2 samples update (#432)

- Added Python samples for CUDA Python 1.0 release
- Renamed top-level `Samples` directory to `cpp` to accommodate Python samples.

cpp/3_CUDA_Features/cudaCompressibleMemory/.vscode/c_cpp_properties.json

@@ -0,0 +1,18 @@
+{
+    "configurations": [
+        {
+            "name": "Linux",
+            "includePath": [
+                "${workspaceFolder}/**",
+                "${workspaceFolder}/../../../Common"
+            ],
+            "defines": [],
+            "compilerPath": "/usr/local/cuda/bin/nvcc",
+            "cStandard": "gnu17",
+            "cppStandard": "gnu++14",
+            "intelliSenseMode": "linux-gcc-x64",
+            "configurationProvider": "ms-vscode.makefile-tools"
+        }
+    ],
+    "version": 4
+}

cpp/3_CUDA_Features/cudaCompressibleMemory/.vscode/extensions.json

@@ -0,0 +1,7 @@
+{
+    "recommendations": [
+        "nvidia.nsight-vscode-edition",
+        "ms-vscode.cpptools",
+        "ms-vscode.makefile-tools"
+    ]
+}

cpp/3_CUDA_Features/cudaCompressibleMemory/CMakeLists.txt

@@ -0,0 +1,42 @@
+cmake_minimum_required(VERSION 3.20)
+
+list(APPEND CMAKE_MODULE_PATH "${CMAKE_CURRENT_SOURCE_DIR}/../../../cmake/Modules")
+
+project(cudaCompressibleMemory LANGUAGES C CXX CUDA)
+
+find_package(CUDAToolkit REQUIRED)
+
+set(CMAKE_POSITION_INDEPENDENT_CODE ON)
+
+set(CMAKE_CUDA_ARCHITECTURES 75 80 86 87 89 90 100 110 120)
+set(CMAKE_CUDA_FLAGS "${CMAKE_CUDA_FLAGS} -Wno-deprecated-gpu-targets")
+if(ENABLE_CUDA_DEBUG)
+    set(CMAKE_CUDA_FLAGS "${CMAKE_CUDA_FLAGS} -G")        # enable cuda-gdb (may significantly affect performance on some targets)
+else()
+    set(CMAKE_CUDA_FLAGS "${CMAKE_CUDA_FLAGS} -lineinfo") # add line information to all builds for debug tools (exclusive to -G option)
+endif()
+
+# Include directories and libraries
+include_directories(../../../Common)
+
+# Source file
+# Add target for cudaCompressibleMemory
+add_executable(cudaCompressibleMemory compMalloc.cpp saxpy.cu)
+
+target_compile_options(cudaCompressibleMemory PRIVATE $<$<COMPILE_LANGUAGE:CUDA>:--extended-lambda>)
+
+target_compile_features(cudaCompressibleMemory PRIVATE cxx_std_17 cuda_std_17)
+
+set_target_properties(cudaCompressibleMemory PROPERTIES CUDA_SEPARABLE_COMPILATION ON)
+
+target_include_directories(cudaCompressibleMemory PRIVATE
+    ${CUDAToolkit_INCLUDE_DIRS}
+)
+
+target_link_libraries(cudaCompressibleMemory PRIVATE
+    CUDA::cuda_driver
+)
+
+# Include installation configuration
+include(${CMAKE_CURRENT_SOURCE_DIR}/../../../cmake/InstallSamples.cmake)
+setup_samples_install()

cpp/3_CUDA_Features/cudaCompressibleMemory/README.md

@@ -0,0 +1,35 @@
+# cudaCompressibleMemory - CUDA Compressible Memory
+
+## Description
+
+This sample demonstrates the compressible memory allocation using cuMemMap API.
+
+## Key Concepts
+
+CUDA Driver API, Compressible Memory, MMAP
+
+## Supported SM Architectures
+
+[SM 5.0 ](https://developer.nvidia.com/cuda-gpus)  [SM 5.2 ](https://developer.nvidia.com/cuda-gpus)  [SM 5.3 ](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)  [SM 8.6 ](https://developer.nvidia.com/cuda-gpus)  [SM 8.7 ](https://developer.nvidia.com/cuda-gpus)  [SM 8.9 ](https://developer.nvidia.com/cuda-gpus)  [SM 9.0 ](https://developer.nvidia.com/cuda-gpus)
+
+## Supported OSes
+
+Linux, Windows, QNX
+
+## Supported CPU Architecture
+
+x86_64, aarch64
+
+## CUDA APIs involved
+
+### [CUDA Driver API](http://docs.nvidia.com/cuda/cuda-driver-api/index.html)
+cuMemGetAllocationPropertiesFromHandle, cuMemCreate, cuDeviceGetAttribute, cuCtxGetDevice, cuMemGetAllocationGranularity, cuMemAddressFree, cuMemUnmap, cuMemMap, cuMemRelease, cuMemAddressReserve, cuMemSetAccess
+
+### [CUDA Runtime API](http://docs.nvidia.com/cuda/cuda-runtime-api/index.html)
+cudaMemcpy, cudaEventSynchronize, cudaEventRecord, cudaEventElapsedTime, cudaOccupancyMaxPotentialBlockSize, cudaEventCreate
+
+## Prerequisites
+
+Download and install the [CUDA Toolkit](https://developer.nvidia.com/cuda-downloads) for your corresponding platform.
+
+## References (for more details)

cpp/3_CUDA_Features/cudaCompressibleMemory/compMalloc.cpp

@@ -0,0 +1,115 @@
+/* Copyright (c) 2022, 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.
+ */
+
+#include <cuda.h>
+#include <cuda_runtime_api.h>
+#include <helper_cuda.h>
+#include <stdio.h>
+#include <string.h>
+
+cudaError_t setProp(CUmemAllocationProp *prop, bool UseCompressibleMemory)
+{
+    CUdevice currentDevice;
+    if (cuCtxGetDevice(&currentDevice) != CUDA_SUCCESS)
+        return cudaErrorMemoryAllocation;
+
+    memset(prop, 0, sizeof(CUmemAllocationProp));
+    prop->type          = CU_MEM_ALLOCATION_TYPE_PINNED;
+    prop->location.type = CU_MEM_LOCATION_TYPE_DEVICE;
+    prop->location.id   = currentDevice;
+
+    if (UseCompressibleMemory)
+        prop->allocFlags.compressionType = CU_MEM_ALLOCATION_COMP_GENERIC;
+
+    return cudaSuccess;
+}
+
+cudaError_t allocateCompressible(void **adr, size_t size, bool UseCompressibleMemory)
+{
+    CUmemAllocationProp prop = {};
+    cudaError_t         err  = setProp(&prop, UseCompressibleMemory);
+    if (err != cudaSuccess)
+        return err;
+
+    size_t granularity = 0;
+    if (cuMemGetAllocationGranularity(&granularity, &prop, CU_MEM_ALLOC_GRANULARITY_MINIMUM) != CUDA_SUCCESS)
+        return cudaErrorMemoryAllocation;
+    size = ((size - 1) / granularity + 1) * granularity;
+    CUdeviceptr dptr;
+    if (cuMemAddressReserve(&dptr, size, 0, 0, 0) != CUDA_SUCCESS)
+        return cudaErrorMemoryAllocation;
+
+    CUmemGenericAllocationHandle allocationHandle;
+    if (cuMemCreate(&allocationHandle, size, &prop, 0) != CUDA_SUCCESS)
+        return cudaErrorMemoryAllocation;
+
+    // Check if cuMemCreate was able to allocate compressible memory.
+    if (UseCompressibleMemory) {
+        CUmemAllocationProp allocationProp = {};
+        cuMemGetAllocationPropertiesFromHandle(&allocationProp, allocationHandle);
+        if (allocationProp.allocFlags.compressionType != CU_MEM_ALLOCATION_COMP_GENERIC) {
+            printf("Could not allocate compressible memory... so waiving execution\n");
+            exit(EXIT_WAIVED);
+        }
+    }
+
+    if (cuMemMap(dptr, size, 0, allocationHandle, 0) != CUDA_SUCCESS)
+        return cudaErrorMemoryAllocation;
+
+    if (cuMemRelease(allocationHandle) != CUDA_SUCCESS)
+        return cudaErrorMemoryAllocation;
+
+    CUmemAccessDesc accessDescriptor;
+    accessDescriptor.location.id   = prop.location.id;
+    accessDescriptor.location.type = prop.location.type;
+    accessDescriptor.flags         = CU_MEM_ACCESS_FLAGS_PROT_READWRITE;
+
+    if (cuMemSetAccess(dptr, size, &accessDescriptor, 1) != CUDA_SUCCESS)
+        return cudaErrorMemoryAllocation;
+
+    *adr = (void *)dptr;
+    return cudaSuccess;
+}
+
+cudaError_t freeCompressible(void *ptr, size_t size, bool UseCompressibleMemory)
+{
+    CUmemAllocationProp prop = {};
+    cudaError_t         err  = setProp(&prop, UseCompressibleMemory);
+    if (err != cudaSuccess)
+        return err;
+
+    size_t granularity = 0;
+    if (cuMemGetAllocationGranularity(&granularity, &prop, CU_MEM_ALLOC_GRANULARITY_MINIMUM) != CUDA_SUCCESS)
+        return cudaErrorMemoryAllocation;
+    size = ((size - 1) / granularity + 1) * granularity;
+
+    if (ptr == NULL)
+        return cudaSuccess;
+    if (cuMemUnmap((CUdeviceptr)ptr, size) != CUDA_SUCCESS || cuMemAddressFree((CUdeviceptr)ptr, size) != CUDA_SUCCESS)
+        return cudaErrorInvalidValue;
+    return cudaSuccess;
+}

cpp/3_CUDA_Features/cudaCompressibleMemory/compMalloc.h

@@ -0,0 +1,34 @@
+/* Copyright (c) 2022, 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.
+ */
+
+#ifndef COMP_MALLOC_H
+#define COMP_MALLOC_H
+
+cudaError_t allocateCompressible(void **adr, size_t size, bool UseCompressibleMemory);
+cudaError_t freeCompressible(void *ptr, size_t size, bool UseCompressibleMemory);
+
+#endif

cpp/3_CUDA_Features/cudaCompressibleMemory/saxpy.cu

@@ -0,0 +1,192 @@
+/* Copyright (c) 2022, 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.
+ */
+
+//
+// This sample uses the compressible memory allocation if device supports it
+// and performs saxpy on it.
+// Compressible memory may give better performance if the data is amenable to
+// compression.
+
+#include <cuda.h>
+#include <stdio.h>
+#define CUDA_DRIVER_API
+#include "compMalloc.h"
+#include "helper_cuda.h"
+
+__global__ void saxpy(const float a, const float4 *x, const float4 *y, float4 *z, const size_t n)
+{
+    for (size_t i = blockIdx.x * blockDim.x + threadIdx.x; i < n; i += gridDim.x * blockDim.x) {
+        const float4 x4 = x[i];
+        const float4 y4 = y[i];
+        z[i]            = make_float4(a * x4.x + y4.x, a * x4.y + y4.y, a * x4.z + y4.z, a * x4.w + y4.w);
+    }
+}
+
+__global__ void init(float4 *x, float4 *y, const float val, const size_t n)
+{
+    const float4 val4 = make_float4(val, val, val, val);
+    for (size_t i = blockIdx.x * blockDim.x + threadIdx.x; i < n; i += gridDim.x * blockDim.x) {
+        x[i] = y[i] = val4;
+    }
+}
+
+void launchSaxpy(const float  a,
+                 float4      *x,
+                 float4      *y,
+                 float4      *z,
+                 const size_t n,
+                 const float  init_val,
+                 const bool   compressibleZbuf)
+{
+    cudaEvent_t start, stop;
+    float       ms;
+    int         blockSize;
+    int         minGridSize;
+    dim3        threads, blocks;
+
+    if (!compressibleZbuf) {
+        // We are on config where compressible buffer can only be initialized through cudaMemcpy
+        // hence, x & y buffers are allocated as compressible and initialized via cudaMemcpy
+        // whereas z buffer is allocated as non-compressible.
+        float4 *h_x = (float4 *)malloc(sizeof(float4) * n);
+        float4 *h_y = (float4 *)malloc(sizeof(float4) * n);
+        for (int i = 0; i < n; i++) {
+            h_x[i].x = h_x[i].y = h_x[i].z = h_x[i].w = init_val;
+            h_y[i].x = h_y[i].y = h_y[i].z = h_y[i].w = init_val;
+        }
+        checkCudaErrors(cudaMemcpy(x, h_x, sizeof(float4) * n, cudaMemcpyHostToDevice));
+        checkCudaErrors(cudaMemcpy(y, h_y, sizeof(float4) * n, cudaMemcpyHostToDevice));
+        free(h_x);
+        free(h_y);
+    }
+    else {
+        checkCudaErrors(cudaOccupancyMaxPotentialBlockSize(&minGridSize, &blockSize, (void *)init));
+        threads = dim3(blockSize, 1, 1);
+        blocks  = dim3(minGridSize, 1, 1);
+        init<<<blocks, threads>>>(x, y, init_val, n);
+    }
+
+    checkCudaErrors(cudaOccupancyMaxPotentialBlockSize(&minGridSize, &blockSize, (void *)saxpy));
+    threads = dim3(blockSize, 1, 1);
+    blocks  = dim3(minGridSize, 1, 1);
+
+    checkCudaErrors(cudaEventCreate(&start));
+    checkCudaErrors(cudaEventCreate(&stop));
+    checkCudaErrors(cudaEventRecord(start));
+    saxpy<<<blocks, threads>>>(a, x, y, z, n);
+    checkCudaErrors(cudaEventRecord(stop));
+    checkCudaErrors(cudaEventSynchronize(stop));
+    checkCudaErrors(cudaEventElapsedTime(&ms, start, stop));
+
+    const size_t size = n * sizeof(float4);
+    printf("Running saxpy with %d blocks x %d threads = %.3f ms %.3f TB/s\n",
+           blocks.x,
+           threads.x,
+           ms,
+           (size * 3) / ms / 1e9);
+}
+
+int main(int argc, char **argv)
+{
+    const size_t n = 10485760;
+
+    if (checkCmdLineFlag(argc, (const char **)argv, "help") || checkCmdLineFlag(argc, (const char **)argv, "?")) {
+        printf("Usage -device=n (n >= 0 for deviceID)\n");
+        exit(EXIT_SUCCESS);
+    }
+
+    findCudaDevice(argc, (const char **)argv);
+    CUdevice currentDevice;
+    checkCudaErrors(cuCtxGetDevice(&currentDevice));
+
+    // Check that the selected device supports virtual memory management
+    int vmm_supported = -1;
+    checkCudaErrors(
+        cuDeviceGetAttribute(&vmm_supported, CU_DEVICE_ATTRIBUTE_VIRTUAL_ADDRESS_MANAGEMENT_SUPPORTED, currentDevice));
+    if (vmm_supported == 0) {
+        printf("Device %d doesn't support Virtual Memory Management, waiving the execution.\n", currentDevice);
+        exit(EXIT_WAIVED);
+    }
+
+    int isCompressionAvailable;
+    checkCudaErrors(cuDeviceGetAttribute(
+        &isCompressionAvailable, CU_DEVICE_ATTRIBUTE_GENERIC_COMPRESSION_SUPPORTED, currentDevice));
+    if (isCompressionAvailable == 0) {
+        printf("Device %d doesn't support Generic memory compression, waiving the execution.\n", currentDevice);
+        exit(EXIT_WAIVED);
+    }
+
+    printf("Generic memory compression support is available\n");
+
+    int major, minor;
+    checkCudaErrors(cuDeviceGetAttribute(&major, CU_DEVICE_ATTRIBUTE_COMPUTE_CAPABILITY_MAJOR, currentDevice));
+    checkCudaErrors(cuDeviceGetAttribute(&minor, CU_DEVICE_ATTRIBUTE_COMPUTE_CAPABILITY_MINOR, currentDevice));
+    float4      *x, *y, *z;
+    const size_t size = n * sizeof(float4);
+
+    // Allocating compressible memory
+    checkCudaErrors(allocateCompressible((void **)&x, size, true));
+    checkCudaErrors(allocateCompressible((void **)&y, size, true));
+    bool compressibleZbuf = 0;
+    if ((major == 8 && minor == 0) || (major == 8 && minor == 6)) {
+        // On SM 8.0 and 8.6 GPUs compressible buffer can only be initialized
+        // through cudaMemcpy.
+        printf("allocating non-compressible Z buffer\n");
+        checkCudaErrors(allocateCompressible((void **)&z, size, false));
+        compressibleZbuf = 0;
+    }
+    else {
+        checkCudaErrors(allocateCompressible((void **)&z, size, true));
+        compressibleZbuf = 1;
+    }
+
+    printf("Running saxpy on %zu bytes of Compressible memory\n", size);
+
+    const float a        = 1.0f;
+    const float init_val = 1.0f;
+    launchSaxpy(a, x, y, z, n, init_val, compressibleZbuf);
+
+    checkCudaErrors(freeCompressible(x, size, true));
+    checkCudaErrors(freeCompressible(y, size, true));
+    checkCudaErrors(freeCompressible(z, size, true));
+
+    printf("Running saxpy on %zu bytes of Non-Compressible memory\n", size);
+    // Allocating non-compressible memory
+    checkCudaErrors(allocateCompressible((void **)&x, size, false));
+    checkCudaErrors(allocateCompressible((void **)&y, size, false));
+    checkCudaErrors(allocateCompressible((void **)&z, size, false));
+
+    launchSaxpy(a, x, y, z, n, init_val, compressibleZbuf);
+
+    checkCudaErrors(freeCompressible(x, size, false));
+    checkCudaErrors(freeCompressible(y, size, false));
+    checkCudaErrors(freeCompressible(z, size, false));
+
+    printf("\nNOTE: The CUDA Samples are not meant for performance measurements. "
+           "Results may vary when GPU Boost is enabled.\n");
+    return EXIT_SUCCESS;
+}