[CI] add assert-enabled MLIR option (#78)
This deprecates the use of release-build LLVM hosted by the LLVM project, which makes debugging harder for developers. This PR implements the following solution: 1. Create LLVM release tarballs with assert enabled on our own (using Docker) 2. Host them in our own GitHub repositories 3. Use our LLVM for CI and/or development if `TRITON_USE_ASSERT_ENABLED_LLVM=1` is set.
This commit is contained in:
@@ -29,8 +29,8 @@ func @test_store_splat(%ptr: !tt.ptr<f32>) {
|
||||
%vs = tt.splat %a : (f32) -> tensor<128xf32>
|
||||
%mask = tt.splat %true : (i1) -> tensor<128xi1>
|
||||
|
||||
// CHECK: %{{.*}} = llvm.inline_asm has_side_effects asm_dialect = att operand_attrs = [] "@$0 st.global.b32 [ $1 + 0 ], { $2 };",
|
||||
// CHECK-SAME: "b,l,r" %{{.*}}, %{{.*}}, %{{.*}} : (i1, !llvm.ptr<f32, 1>, i32) -> !llvm.struct<()>
|
||||
// CHECK: %{{.*}} = llvm.inline_asm has_side_effects asm_dialect = att operand_attrs = [] "@$1 st.global.b32 [ $2 + 0 ], { $0 };",
|
||||
// CHECK-SAME: "r,b,l" %{{.*}}, %{{.*}}, %{{.*}} : (i32, i1, !llvm.ptr<f32, 1>) -> !llvm.void
|
||||
tt.store %ptrs, %vs, %mask : tensor<128xf32>
|
||||
|
||||
return
|
||||
|
@@ -183,9 +183,9 @@ module attributes {"triton_gpu.num-warps" = 4 : i32} {
|
||||
// CHECK-LABEL: basic_store
|
||||
func @basic_store(%ptrs: tensor<256x!tt.ptr<f32>, #blocked0>, %vals: tensor<256xf32, #blocked0>, %mask: tensor<256xi1, #blocked0>) {
|
||||
// CHECK: llvm.inline_asm has_side_effects asm_dialect = att
|
||||
// CHECK-SAME: st.global.b32 [ ${{.*}} + 0 ], { ${{.*}} };", "b,l,r" %{{.*}}, %{{.*}}, %{{.*}} : (i1, !llvm.ptr<f32, 1>, i32) -> !llvm.struct<()>
|
||||
// CHECK-SAME: st.global.b32 [ ${{.*}} + 0 ], { ${{.*}} };", "r,b,l" %{{.*}}, %{{.*}}, %{{.*}} : (i32, i1, !llvm.ptr<f32, 1>) -> !llvm.void
|
||||
// CHECK: llvm.inline_asm has_side_effects asm_dialect = att
|
||||
// CHECK-SAME: st.global.b32 [ ${{.*}} + 0 ], { ${{.*}} };", "b,l,r" %{{.*}}, %{{.*}}, %{{.*}} : (i1, !llvm.ptr<f32, 1>, i32) -> !llvm.struct<()>
|
||||
// CHECK-SAME: st.global.b32 [ ${{.*}} + 0 ], { ${{.*}} };", "r,b,l" %{{.*}}, %{{.*}}, %{{.*}} : (i32, i1, !llvm.ptr<f32, 1>) -> !llvm.void
|
||||
tt.store %ptrs, %vals, %mask : tensor<256xf32, #blocked0>
|
||||
return
|
||||
}
|
||||
|
Reference in New Issue
Block a user