Support for CUDA Tile C++ kernels

examples/cuda-c++/vector_add_tile.cu

@@ -0,0 +1,26 @@
+#include "cuda_tile.h"
+
+__tile_global__ void vector_add_tile(float* a, float* b, float* out, int n) {
+    namespace ct = cuda::tiles;
+    using namespace ct::literals;
+
+    a   = ct::assume_aligned(a,   16_ic);
+    b   = ct::assume_aligned(b,   16_ic);
+    out = ct::assume_aligned(out, 16_ic);
+
+    // Step 1: attach a shape to each raw pointer. n is a runtime value (dynamic extent).
+    auto aSpan = ct::tensor_span{a,   ct::extents{n}};
+    auto bSpan = ct::tensor_span{b,   ct::extents{n}};
+    auto oSpan = ct::tensor_span{out, ct::extents{n}};
+
+    // Step 2: partition each span into tiles of TILE_SIZE elements (tuned by kernel_tuner).
+    constexpr auto tile = ct::integral_constant<TILE_SIZE>{};
+    auto aView = ct::partition_view{aSpan, ct::shape{tile}};
+    auto bView = ct::partition_view{bSpan, ct::shape{tile}};
+    auto oView = ct::partition_view{oSpan, ct::shape{tile}};
+
+    int  bx    = ct::bid().x;             // this block's tile-space index along .x
+    auto aTile = aView.load(bx);          // pick the bx-th tile of a
+    auto bTile = bView.load(bx);
+    oView.store(aTile + bTile, bx);       // write the tile back at the bx-th position of out
+}

examples/cuda-c++/vector_add_tile.py

@@ -0,0 +1,35 @@
+#!/usr/bin/env python
+""" This is a minimal example to tune a CUDA Tile vector add kernel """
+
+import numpy
+from kernel_tuner import tune_kernel
+
+def tune():
+
+    size = 3*2**24
+
+    a = numpy.random.randn(size).astype(numpy.float32)
+    b = numpy.random.randn(size).astype(numpy.float32)
+    c = numpy.zeros_like(b)
+    n = numpy.int32(size)
+
+    args = [a, b, c, n]
+
+    # TILE_SIZE controls how many elements each tile processes. kernel_tuner injects
+    # it as a #define so the kernel can use it as a compile-time constant.
+    # grid_x = size / TILE_SIZE (number of tiles), computed automatically by kernel_tuner.
+    tune_params = {"TILE_SIZE": [8, 16, 32, 64, 128, 256]}
+
+    answer = [None, None, a+b, None]
+
+    compiler_options = ["-enable-tile", "-std=c++20"]
+
+    results, env = tune_kernel("vector_add_tile<TILE_SIZE>", "vector_add_tile.cu", size, args,
+                       tune_params, lang="NVCUDA", compiler_options=compiler_options,
+                       answer=answer, verbose=True, grid_div_x=["TILE_SIZE"])
+
+    return results
+
+
+if __name__ == "__main__":
+    tune()