|
| 1 | +from typing import Optional |
| 2 | + |
| 3 | +import tilelang.language as T |
| 4 | +import tilelang.testing |
| 5 | +import torch |
| 6 | +from tilelang.utils.target import check_hip_availability |
| 7 | + |
| 8 | +_IS_HIP_AVAILABLE = check_hip_availability() |
| 9 | +_DEFAULT_WARPS_PER_GROUP = 4 |
| 10 | + |
| 11 | + |
| 12 | +def _resolve_warp_size(warp_size: Optional[int]) -> int: |
| 13 | + if warp_size is not None: |
| 14 | + return int(warp_size) |
| 15 | + return 64 if _IS_HIP_AVAILABLE else 32 |
| 16 | + |
| 17 | + |
| 18 | +def _resolve_warps_per_group(warps_per_group: Optional[int]) -> int: |
| 19 | + if warps_per_group is not None: |
| 20 | + return int(warps_per_group) |
| 21 | + return _DEFAULT_WARPS_PER_GROUP |
| 22 | + |
| 23 | + |
| 24 | +@tilelang.jit(out_idx=[-1]) |
| 25 | +def _get_laneid_kernel(num_threads: int = 128, warp_size: Optional[int] = None): |
| 26 | + |
| 27 | + @T.prim_func |
| 28 | + def laneid_kernel(A: T.Tensor((num_threads,), "int32")): |
| 29 | + with T.Kernel(1, threads=num_threads) as _: |
| 30 | + tx = T.get_thread_binding() |
| 31 | + A[tx] = T.get_lane_idx(warp_size) |
| 32 | + |
| 33 | + return laneid_kernel |
| 34 | + |
| 35 | + |
| 36 | +@tilelang.jit(out_idx=[-1]) |
| 37 | +def _get_warp_idx_sync_kernel( |
| 38 | + num_threads: int = 128, warp_size: Optional[int] = None |
| 39 | +): |
| 40 | + |
| 41 | + @T.prim_func |
| 42 | + def warp_idx_sync_kernel(A: T.Tensor((num_threads,), "int32")): |
| 43 | + with T.Kernel(1, threads=num_threads) as _: |
| 44 | + tx = T.get_thread_binding() |
| 45 | + A[tx] = T.get_warp_idx_sync(warp_size) |
| 46 | + |
| 47 | + return warp_idx_sync_kernel |
| 48 | + |
| 49 | + |
| 50 | +@tilelang.jit(out_idx=[-1]) |
| 51 | +def _get_warp_idx_kernel(num_threads: int = 128, warp_size: Optional[int] = None): |
| 52 | + |
| 53 | + @T.prim_func |
| 54 | + def warp_idx_kernel(A: T.Tensor((num_threads,), "int32")): |
| 55 | + with T.Kernel(1, threads=num_threads) as _: |
| 56 | + tx = T.get_thread_binding() |
| 57 | + A[tx] = T.get_warp_idx(warp_size) |
| 58 | + |
| 59 | + return warp_idx_kernel |
| 60 | + |
| 61 | + |
| 62 | +@tilelang.jit(out_idx=[-1]) |
| 63 | +def _get_warp_group_idx_kernel( |
| 64 | + num_threads: int = 128, |
| 65 | + warp_size: Optional[int] = None, |
| 66 | + warps_per_group: Optional[int] = None, |
| 67 | +): |
| 68 | + |
| 69 | + @T.prim_func |
| 70 | + def warp_group_idx_kernel(A: T.Tensor((num_threads,), "int32")): |
| 71 | + with T.Kernel(1, threads=num_threads) as _: |
| 72 | + tx = T.get_thread_binding() |
| 73 | + A[tx] = T.get_warp_group_idx(warp_size, warps_per_group) |
| 74 | + |
| 75 | + return warp_group_idx_kernel |
| 76 | + |
| 77 | + |
| 78 | +@tilelang.jit(out_idx=[-1]) |
| 79 | +def _shuffle_elect_kernel( |
| 80 | + num_threads: int = 128, thread_extent: int = 64 |
| 81 | +): |
| 82 | + |
| 83 | + @T.prim_func |
| 84 | + def shuffle_elect_kernel(A: T.Tensor((num_threads,), "int32")): |
| 85 | + with T.Kernel(1, threads=num_threads) as _: |
| 86 | + tx = T.get_thread_binding() |
| 87 | + elected = T.shuffle_elect(thread_extent) |
| 88 | + A[tx] = elected |
| 89 | + |
| 90 | + return shuffle_elect_kernel |
| 91 | + |
| 92 | + |
| 93 | +def run_get_lane_id(num_threads: int = 128, warp_size: Optional[int] = None): |
| 94 | + kernel = _get_laneid_kernel(num_threads, warp_size) |
| 95 | + A = kernel() |
| 96 | + print(kernel.get_kernel_source()) |
| 97 | + print(A) |
| 98 | + expected_warp_size = _resolve_warp_size(warp_size) |
| 99 | + ref = torch.arange( |
| 100 | + num_threads, dtype=A.dtype, device=A.device |
| 101 | + ) % expected_warp_size |
| 102 | + torch.testing.assert_close(A.cpu(), ref.cpu()) |
| 103 | + return A |
| 104 | + |
| 105 | + |
| 106 | +def run_get_warp_idx_sync( |
| 107 | + num_threads: int = 128, warp_size: Optional[int] = None |
| 108 | +): |
| 109 | + kernel = _get_warp_idx_sync_kernel(num_threads, warp_size) |
| 110 | + A = kernel() |
| 111 | + print(kernel.get_kernel_source()) |
| 112 | + print(A) |
| 113 | + expected_warp_size = _resolve_warp_size(warp_size) |
| 114 | + ref = torch.arange( |
| 115 | + num_threads, dtype=A.dtype, device=A.device |
| 116 | + ) // expected_warp_size |
| 117 | + torch.testing.assert_close(A.cpu(), ref.cpu()) |
| 118 | + return A |
| 119 | + |
| 120 | + |
| 121 | +def run_get_warp_idx(num_threads: int = 128, warp_size: Optional[int] = None): |
| 122 | + kernel = _get_warp_idx_kernel(num_threads, warp_size) |
| 123 | + A = kernel() |
| 124 | + print(kernel.get_kernel_source()) |
| 125 | + print(A) |
| 126 | + expected_warp_size = _resolve_warp_size(warp_size) |
| 127 | + ref = torch.arange( |
| 128 | + num_threads, dtype=A.dtype, device=A.device |
| 129 | + ) // expected_warp_size |
| 130 | + torch.testing.assert_close(A.cpu(), ref.cpu()) |
| 131 | + return A |
| 132 | + |
| 133 | + |
| 134 | +def run_get_warp_group_idx( |
| 135 | + num_threads: int = 128, |
| 136 | + warp_size: Optional[int] = None, |
| 137 | + warps_per_group: Optional[int] = None, |
| 138 | +): |
| 139 | + kernel = _get_warp_group_idx_kernel(num_threads, warp_size, warps_per_group) |
| 140 | + A = kernel() |
| 141 | + print(kernel.get_kernel_source()) |
| 142 | + print(A) |
| 143 | + expected_warp_size = _resolve_warp_size(warp_size) |
| 144 | + expected_warps_per_group = _resolve_warps_per_group(warps_per_group) |
| 145 | + threads_per_group = expected_warp_size * expected_warps_per_group |
| 146 | + if threads_per_group <= 0: |
| 147 | + raise ValueError("threads_per_group must be positive.") |
| 148 | + ref = torch.arange( |
| 149 | + num_threads, dtype=A.dtype, device=A.device |
| 150 | + ) // threads_per_group |
| 151 | + torch.testing.assert_close(A.cpu(), ref.cpu()) |
| 152 | + return A |
| 153 | + |
| 154 | + |
| 155 | +def run_shuffle_elect( |
| 156 | + num_threads: int = 128, thread_extent: int = 64 |
| 157 | +): |
| 158 | + if thread_extent < 0: |
| 159 | + raise ValueError("thread_extent must be non-negative.") |
| 160 | + kernel = _shuffle_elect_kernel(num_threads, thread_extent) |
| 161 | + A = kernel() |
| 162 | + print(kernel.get_kernel_source()) |
| 163 | + print(A) |
| 164 | + indices = torch.arange( |
| 165 | + num_threads, device=A.device, dtype=torch.int64 |
| 166 | + ) |
| 167 | + if thread_extent == 0: |
| 168 | + mask = indices == 0 |
| 169 | + elif thread_extent > 0: |
| 170 | + mask = (indices % thread_extent) == 0 |
| 171 | + else: |
| 172 | + mask = torch.zeros_like(indices, dtype=torch.bool) |
| 173 | + ref = mask.to(dtype=A.dtype, device=A.device) |
| 174 | + torch.testing.assert_close(A.cpu(), ref.cpu()) |
| 175 | + return A |
| 176 | + |
| 177 | + |
| 178 | +@tilelang.testing.requires_cuda |
| 179 | +def test_get_lane_idx_default(): |
| 180 | + run_get_lane_id() |
| 181 | + |
| 182 | + |
| 183 | +@tilelang.testing.requires_cuda |
| 184 | +def test_get_lane_idx_custom(): |
| 185 | + run_get_lane_id(num_threads=256, warp_size=64) |
| 186 | + |
| 187 | + |
| 188 | +@tilelang.testing.requires_cuda |
| 189 | +def test_get_warp_idx_sync_default(): |
| 190 | + run_get_warp_idx_sync() |
| 191 | + |
| 192 | + |
| 193 | +@tilelang.testing.requires_cuda |
| 194 | +def test_get_warp_idx_sync_custom(): |
| 195 | + run_get_warp_idx_sync(num_threads=256, warp_size=16) |
| 196 | + |
| 197 | + |
| 198 | +@tilelang.testing.requires_cuda |
| 199 | +def test_get_warp_idx_default(): |
| 200 | + run_get_warp_idx() |
| 201 | + |
| 202 | + |
| 203 | +@tilelang.testing.requires_cuda |
| 204 | +def test_get_warp_idx_custom(): |
| 205 | + run_get_warp_idx(num_threads=320, warp_size=20) |
| 206 | + |
| 207 | + |
| 208 | +@tilelang.testing.requires_cuda |
| 209 | +def test_get_warp_group_idx_default(): |
| 210 | + run_get_warp_group_idx() |
| 211 | + |
| 212 | + |
| 213 | +@tilelang.testing.requires_cuda |
| 214 | +def test_get_warp_group_idx_custom(): |
| 215 | + run_get_warp_group_idx(num_threads=512, warp_size=32, warps_per_group=5) |
| 216 | + |
| 217 | + |
| 218 | +@tilelang.testing.requires_cuda |
| 219 | +def test_shuffle_elect_default(): |
| 220 | + run_shuffle_elect(num_threads=256, thread_extent=64) |
| 221 | + |
| 222 | + |
| 223 | +@tilelang.testing.requires_cuda |
| 224 | +def test_shuffle_elect_block_leader(): |
| 225 | + run_shuffle_elect(num_threads=128, thread_extent=0) |
| 226 | + |
| 227 | +if __name__ == "__main__": |
| 228 | + tilelang.testing.main() |
| 229 | + # run_get_lane_id() |
0 commit comments