Fixed kernel launch config for permute021 (#957)

Summary: make sure we have grid_z to be less than 65535. Differential Revision: D50517471
facebookincubator · Oct 21, 2023 · 4c103e9 · 4c103e9
1 parent 3c4ba48
commit 4c103e9
Show file tree

Hide file tree

Showing 2 changed files with 83 additions and 16 deletions.
diff --git a/python/aitemplate/backend/common/tensor/permute021_common.py b/python/aitemplate/backend/common/tensor/permute021_common.py
@@ -84,7 +84,36 @@
 
 {{tensor_accessor_libs}}
 
-template <typename T>
+// blockIdx.x -> ni
+// blockIdx.y -> hwi
+// blockIdx.z -> ci
+__device__ __forceinline__ void block_fn_nhc(int32_t& ni, int32_t& hwi, int32_t& ci) {
+  ni = blockIdx.x;
+  hwi = blockIdx.y;
+  ci = blockIdx.z;
+}
+
+// blockIdx.x -> ni
+// blockIdx.y -> ci
+// blockIdx.z -> hwi
+__device__ __forceinline__ void block_fn_nch(int32_t& ni, int32_t& hwi, int32_t& ci) {
+  ni = blockIdx.x;
+  ci = blockIdx.y;
+  hwi = blockIdx.z;
+}
+
+// blockIdx.x -> ci
+// blockIdx.y -> hwi
+// blockIdx.z -> ni
+__device__ __forceinline__ void block_fn_chn(int32_t& ni, int32_t& hwi, int32_t& ci) {
+  ci = blockIdx.x;
+  hwi = blockIdx.y;
+  ni = blockIdx.z;
+}
+
+using BlockFunc = void (*)(int32_t&, int32_t&, int32_t&);
+
+template <typename T, BlockFunc BLOCK_FN>
 __global__ void permute021_kernel(T *output,
                                   const T *input,
                                   const int64_t n,
@@ -101,9 +130,11 @@
   const int32_t tid  = threadIdx.y * blockDim.x + threadIdx.x;
   const int32_t wid  = tid / TILE_SIZE;
   const int32_t lid  = tid % TILE_SIZE;
-  const int32_t ni   = blockIdx.z;
-  const int32_t hwi0 = blockIdx.y * TILE_SIZE;
-  const int32_t ci0  = blockIdx.x * TILE_SIZE;
+  int32_t ni_tmp, hwi_tmp, ci_tmp;
+  BLOCK_FN(ni_tmp, hwi_tmp, ci_tmp);
+  const int32_t ni = ni_tmp;
+  const int32_t hwi0 = hwi_tmp * TILE_SIZE;
+  const int32_t ci0  = ci_tmp * TILE_SIZE;
 
   size_t input_idx = ni * hwc + (hwi0 + wid) * c + ci0;
 
@@ -172,21 +203,54 @@
   const int32_t x_dim1 = x_dims[rank-2];
   const int32_t x_dim2 = x_dims[rank-1];
 
-  const int64_t n = x_dim0;
+#define THROW_INVALID_LAUNCH_CONFIG                   \
+    throw std::runtime_error(                         \
+        std::string("invalid cuda launch config: ") + \
+        std::to_string(grid_c) + ", " +               \
+        std::to_string(grid_hw) + ", " +              \
+        std::to_string(grid_n));
+
+  const int32_t n = static_cast<int32_t>(x_dim0);
   const int32_t h = 1;
   const int32_t w = x_dim1;
   const int32_t c = x_dim2;
-  dim3 grid((c + TILE_SIZE - 1) / TILE_SIZE, (h * w + TILE_SIZE - 1) / TILE_SIZE, n);
-  dim3 block(TILE_SIZE, TILE_SIZE / CH_K);
-  permute021_kernel<{{lib_dtype}}><<<grid, block, 0, stream>>>(
-    static_cast<{{lib_dtype}}*>(out_ptr),
-    static_cast<const {{lib_dtype}}*>(in_ptr),
-    n,
-    h,
-    w,
-    c,
-    input_accessor
-  );
+  const int32_t grid_c = (c + TILE_SIZE - 1) / TILE_SIZE;
+  const int32_t grid_hw = (h * w + TILE_SIZE - 1) / TILE_SIZE;
+  const int32_t grid_n = n;
+  constexpr int32_t max_grid_z = 65535;
+  constexpr int32_t max_grid_x = 2147483647;
+  if (grid_c > max_grid_x || grid_hw > max_grid_x || grid_n > max_grid_x) {
+    THROW_INVALID_LAUNCH_CONFIG
+  }
+  if ((grid_c <= max_grid_z && grid_hw <= max_grid_z && grid_n <= max_grid_z) ||
+      (grid_c > max_grid_z && grid_hw <= max_grid_z && grid_n <= max_grid_z)) {
+    dim3 grid(grid_c, grid_hw, grid_n);
+    dim3 block(TILE_SIZE, TILE_SIZE / CH_K);
+    permute021_kernel<{{lib_dtype}}, block_fn_chn><<<grid, block, 0, stream>>>(
+        static_cast<{{lib_dtype}}*>(out_ptr),
+        static_cast<const {{lib_dtype}}*>(in_ptr),
+        n, h, w, c, input_accessor
+    );
+  } else if (grid_n > max_grid_z && grid_hw <= max_grid_z && grid_c <= max_grid_z) {
+    dim3 grid(grid_n, grid_c, grid_hw);
+    dim3 block(TILE_SIZE, TILE_SIZE / CH_K);
+    permute021_kernel<{{lib_dtype}}, block_fn_nch><<<grid, block, 0, stream>>>(
+        static_cast<{{lib_dtype}}*>(out_ptr),
+        static_cast<const {{lib_dtype}}*>(in_ptr),
+        n, h, w, c, input_accessor
+    );
+  } else if (grid_n > max_grid_z && grid_hw <= max_grid_z && grid_c <= max_grid_z) {
+    dim3 grid(grid_n, grid_hw, grid_c);
+    dim3 block(TILE_SIZE, TILE_SIZE / CH_K);
+    permute021_kernel<{{lib_dtype}}, block_fn_nhc><<<grid, block, 0, stream>>>(
+        static_cast<{{lib_dtype}}*>(out_ptr),
+        static_cast<const {{lib_dtype}}*>(in_ptr),
+        n, h, w, c, input_accessor
+    );
+  } else {
+    THROW_INVALID_LAUNCH_CONFIG
+  }
+  
 }
 } // namespace
 

diff --git a/tests/unittest/ops/test_permute021.py b/tests/unittest/ops/test_permute021.py
@@ -69,6 +69,9 @@ def _test_permute_021(
             param(3, (2, 3, 4, 384, 262), (0, 1, 2, 4, 3)),
             param(4, (IntVar([2, 3]), 384, 262), (0, 2, 1)),
             param(5, (IntVar([2, 3, 4]), 5, 384, 262), (0, 1, 3, 2)),
+            param(6, (409600, 12, 16), (0, 2, 1)),
+            param(7, (12, 409600, 16), (0, 2, 1)),
+            param(8, (12, 16, 409600), (0, 2, 1)),
         ]
     )
     def test_permute021_fp16(self, id, input_shape, dims):