[GENERAL] Merged v1.0alpha into master. Added features are:

- A100 support via mma.16816
- Thread swizzling for conflict-free shared memory accesses without
padding
- Complete overhaul of the LLVM code generation in
codegen/selection/generator.cc to remove overengineering
- Added debugging capabilities in the Python binding
- Compilation error for kernels that spill

python/examples/tutorials/add.py

@@ -0,0 +1,39 @@
+import torch
+import triton
+
+class _add(torch.autograd.Function):
+    src = """
+__global__ void add(float* z, float* x, float* y, int N) {
+
+    int pid = get_program_id(0);
+
+    int offset[TILE] = pid * TILE + 0 ... TILE;
+    float* pz[TILE]  = z + offset;
+    float* px[TILE]  = x + offset;
+    float* py[TILE]  = y + offset;
+
+    bool check[TILE] = offset < N;
+
+    *pz = *px + *py;
+}
+    """
+
+    kernel = triton.kernel(src, defines={'TILE': 1024}, num_warps=[4])
+
+    @staticmethod
+    def forward(ctx, x, y):
+       z = torch.empty_like(x).cuda()
+       N = x.numel()
+       grid = lambda opt: (triton.cdiv(N, opt.d('TILE')),)
+       _add.kernel(z,x,y, N, grid=grid)
+       return z
+
+add = _add.apply
+
+# test
+torch.manual_seed(0)
+x = torch.rand(900).cuda()
+y = torch.rand(900).cuda()
+za = x + y
+zb = add(x, y)
+print(torch.allclose(za,zb))