[INFRA] Import NVIDIA/CCCL upstream as optimization reference library

CCCL (CUDA C++ Core Libraries) provides:
- CUB: device/block/warp-level GPU primitives (reduce, scan, sort, topk)
- Thrust: high-level parallel algorithms (transform_reduce, sort, scan)
- libcudacxx: CUDA C++ standard library (atomics, barriers, memory)
- cudax: experimental features (memory resources, allocators)
- Tuning policies: per-SM hardware-specific algorithm parameters

Competition optimization vectors mapped to CCCL:
- Output TPS (83% weight): warp_reduce, block_reduce, device_topk
- Input TPS (14% weight): device_scan, block_load, prefetch
- Cache TPS (3% weight): prefix caching strategy patterns
- Memory (0.9 util): pooled/cached/buddy allocators

Source: https://github.com/NVIDIA/cccl (shallow clone, HEAD only)
License: Apache-2.0

cccl_upstream/cudax/examples/stf/1f1b.cu

@@ -0,0 +1,118 @@
+//===----------------------------------------------------------------------===//
+//
+// Part of CUDASTF in CUDA C++ Core Libraries,
+// under the Apache License v2.0 with LLVM Exceptions.
+// See https://llvm.org/LICENSE.txt for license information.
+// SPDX-License-Identifier: Apache-2.0 WITH LLVM-exception
+// SPDX-FileCopyrightText: Copyright (c) 2022-2024 NVIDIA CORPORATION & AFFILIATES.
+//
+//===----------------------------------------------------------------------===//
+
+/**
+ * @file
+ * @brief Toy example to reproduce the asynchrony of a 1F1B pipeline
+ */
+
+#include <cuda/experimental/stf.cuh>
+
+using namespace cuda::experimental::stf;
+
+__global__ void forward(slice<int>, long long int clock_cnt)
+{
+  long long int start_clock  = clock64();
+  long long int clock_offset = 0;
+  while (clock_offset < clock_cnt)
+  {
+    clock_offset = clock64() - start_clock;
+  }
+}
+
+__global__ void backward(slice<int>, long long int clock_cnt)
+{
+  long long int start_clock  = clock64();
+  long long int clock_offset = 0;
+  while (clock_offset < clock_cnt)
+  {
+    clock_offset = clock64() - start_clock;
+  }
+}
+
+int main(int argc, char** argv)
+{
+  context ctx;
+  // Use a graph context if the second argument is set and not null
+  if (argc > 2 && atoi(argv[2]))
+  {
+    ctx = graph_ctx();
+  }
+
+  int device;
+  cudaGetDevice(&device);
+
+  // cudaDevAttrClockRate: Peak clock frequency in kilohertz;
+  int clock_rate;
+  cudaDeviceGetAttribute(&clock_rate, cudaDevAttrClockRate, device);
+
+  auto occ_f = reserved::compute_occupancy(forward);
+  auto occ_b = reserved::compute_occupancy(backward);
+
+  int factor = 1;
+  if (argc > 1)
+  {
+    factor = atoi(argv[1]);
+  }
+
+  size_t num_batches = 8 * factor;
+  int num_devs       = 8;
+  int real_devs;
+  cuda_safe_call(cudaGetDeviceCount(&real_devs));
+
+  std::vector<logical_data<slice<int>>> data;
+
+  for (size_t b = 0; b < num_batches; b++)
+  {
+    auto batch_data = ctx.logical_data(shape_of<slice<int>>(1024));
+    data.push_back(batch_data);
+
+    ctx.task(exec_place::device(0), data[b].write())->*[](cudaStream_t, auto) {
+      // Init ...
+    };
+  }
+
+  cuda_safe_call(cudaStreamSynchronize(ctx.fence()));
+
+  size_t niter = 10;
+
+  for (size_t iter = 0; iter < niter; iter++)
+  {
+    for (size_t b = 0; b < num_batches; b++)
+    {
+      for (int d = 0; d < num_devs; d++)
+      {
+        ctx.task(exec_place::device(d % real_devs), data[b].rw())->*[=](cudaStream_t s, auto bd) {
+          int ms                  = 10;
+          long long int clock_cnt = (long long int) (ms * clock_rate / factor);
+          forward<<<occ_f.min_grid_size, occ_f.block_size, 0, s>>>(bd, clock_cnt);
+        };
+      }
+      //        }
+      //
+      //        for (size_t b = 0; b < num_batches; b++) {
+      for (int d = num_devs; d-- > 0;)
+      {
+        ctx.task(exec_place::device(d % real_devs), data[b].rw())->*[=](cudaStream_t s, auto bd) {
+          int ms                  = 20;
+          long long int clock_cnt = (long long int) (ms * clock_rate / factor);
+          backward<<<occ_b.min_grid_size, occ_b.block_size, 0, s>>>(bd, clock_cnt);
+        };
+      }
+    }
+
+    /* We introduce a fence because the actual pipeline would introduce
+     * some all to all communication to update coefficients */
+    cuda_safe_call(cudaStreamSynchronize(ctx.fence()));
+  }
+
+  ctx.finalize();
+  return 0;
+}