[BACKEND] Fix crash in coalesce pass with blocked ptr #3866
Conversation
…ensorType' Signed-off-by: Tiotto, Ettore <[email protected]>
|
Is it possible to write a test? |
Signed-off-by: Tiotto, Ettore <[email protected]>
@jlebar I have added a lit test in |
test/TritonGPU/coalesce.mlir
Outdated
| #mma = #triton_gpu.nvidia_mma<{warpsPerCTA = [8, 1], CTAsPerCGA = [1, 1], CTASplitNum = [1, 1], CTAOrder = [1, 0], instrShape = [16, 256, 16]}> | ||
|
|
||
| // CHECK-LABEL: @fwd_kernel | ||
| module attributes {"triton_gpu.num-warps" = 4 : i32, "triton_gpu.threads-per-warp" = 16 : i32} { | ||
| tt.func public @fwd_kernel(%arg0: !tt.ptr<f16> , %arg1: !tt.ptr<f16> , %arg2: !tt.ptr<f16> , %arg3: f32, %arg4: !tt.ptr<f32> , %arg5: !tt.ptr<f16> , %arg6: i32 , %arg7: i32 , %arg8: i32 , %arg9: i32 , %arg10: i32 , %arg11: i32 , %arg12: i32 , %arg13: i32 , %arg14: i32 , %arg15: i32 , %arg16: i32 , %arg17: i32 , %arg18: i32, %arg19: i32, %arg20: i32 , %arg21: i32) { | ||
| %cst = arith.constant dense<0.000000e+00> : tensor<128x16xf32, #mma> | ||
| %cst_0 = arith.constant dense<0.000000e+00> : tensor<128xf32, #triton_gpu.slice<{dim = 1, parent = #mma}>> |
There was a problem hiding this comment.
can you minimize the test?
There was a problem hiding this comment.
The test was reduced (minimized) from a much larger test. I might be able to reduce it a bit more (remove some arguments). Let me try.
There was a problem hiding this comment.
The test was reduced (minimized) from a much larger test.
If we accept patches with large testcases like this, our codebase quickly becomes difficult to maintain. Indeed I've lost many days dealing with failures in tests like this.
You may need to write a testcase by hand. Coming up with clear, small testcases is a critical part of contributing to Triton.
There was a problem hiding this comment.
Done. Now the test contains just a couple of operations.
Signed-off-by: Tiotto, Ettore <[email protected]>
ThomasRaoux
changed the title
The
setCoalescedEncodingfunction can handle operations that have a 'mem access ptr' with typeRankedTensorType:Therefore the caller in
runOnOperationshould avoid calling it when the 'mem access ptr' does not haveRankedTensorType(otherwise the cast in the callee will fail).