TIR Printer bugs

[xqdan]

Thanks for participating in the TVM community! We use https://discuss.tvm.ai for any general usage questions and discussions. The issue tracker is used for actionable items such as feature proposals discussion, roadmaps, and bug tracking. You are always welcomed to post on the forum first :)

Issues that are inactive for a period of time may get closed. We adopt this policy so that we won't lose track of actionable issues that may fall at the bottom of the pile. Feel free to reopen a new one if you feel there is an additional problem that needs attention when an old one gets closed.

For bug reports, to help the developer act on the issues, please include a description of your environment, preferably a minimum script to reproduce the problem.

For feature proposals, list clear, small actionable items so we can track the progress of the change.

Hi，

I am trying to use tir printer to dump ir for each pass:

__TRACE_COUNTER__ = 0
def dumpir(module, info, is_before):
    global __TRACE_COUNTER__
    if bool(is_before) == False:
        __TRACE_COUNTER__ += 1
        pname = str(__TRACE_COUNTER__).rjust(2,'0')  + "_" + info.name + ".ir"
        with open(pname, "a") as f:
            f.write(str(module))

def tx_conv2d(input_shape, filter_shape, pad, stride):
    fmap = te.placeholder(input_shape, name="feature_map", dtype="float16")
    weight = te.placeholder(filter_shape, name="filter", dtype="float16")
    conv_res = tx_conv_compute(fmap, weight, pad, stride)
    sch = tx_conv2d_schedule(conv_res)
    with tvm.transform.PassContext(opt_level=3, trace=dumpir):
        (tvm.lower(sch, [fmap, weight, conv_res], simple_mode=True))

1, check failed for local buffer
https://github.com/apache/incubator-tvm/blob/82d157f0b83ae17fde7bbfca14110aa2f2b80b61/src/printer/tir_text_printer.cc#L196

such as: buffer(result.shared_LLB, 0x11b7170), which is not in buffer_map

2, so I comment this, then I find ir dump for 01_tir.InjectPrefetch.ir, missing buffer for a store name, which looks werid:

 attr [] "realize_scope" = "shared_LLB";
        realize(, [0:1, 0:1, 0:16, 0:32], True {
          attr [IterVar(ax0_1: int32, (nullptr), "DataPar", "")] "pragma_dma_copy" = 1;
          for (ax2_1: int32, 0, 16) {
            for (ax3_1: int32, 0, 32) {
              [0, 0, ax2_1, ax3_1] = filter[0, 0, ax2_1, ax3_1]
            }
          }

3, is this an extra ')' here ?
https://github.com/apache/incubator-tvm/blob/82d157f0b83ae17fde7bbfca14110aa2f2b80b61/src/printer/tir_text_printer.cc#L304

for load , what we print is
fmap_l1[(((i1*896) + (i2*16)) + i3)] = (float16*)feature_map.shared_LLB[(((i1*896) + (i2*16)) + i3)]) -> extra ')'

[tqchen]

cc @spectrometerHBH

[xqdan]

@spectrometerHBH you can reproduce with

diff --git a/tests/python/unittest/test_te_schedule_tensor_core.py b/tests/python/unittest/test_te_schedule_tensor_core.py
index ae2301caf..b832a0bce 100644
--- a/tests/python/unittest/test_te_schedule_tensor_core.py
+++ b/tests/python/unittest/test_te_schedule_tensor_core.py
@@ -103,14 +103,16 @@ def intrin_wmma_store_matrix(shape):
     return te.decl_tensor_intrin(C.op, intrin_func, binds={A: BA, C: BC})


-def test_tensor_core_batch_matmal():
-    if not tvm.gpu(0).exist or not tvm.runtime.enabled("cuda"):
-        print("skip because cuda is not enabled..")
-        return
-    if not nvcc.have_tensorcore(tvm.gpu(0).compute_version):
-        print("skip because gpu does not support tensor core")
-        return
+__TRACE_COUNTER__ = 0
+def dumpir(module, info, is_before):
+    global __TRACE_COUNTER__
+    if bool(is_before) == False:
+        __TRACE_COUNTER__ += 1
+        pname = str(__TRACE_COUNTER__).rjust(2,'0')  + "_" + info.name + ".ir"
+        with open(pname, "a") as f:
+            f.write(str(module))

+def test_tensor_core_batch_matmal():
     batch_size = 4
     n = 512
     m, l = n, n
@@ -195,7 +197,8 @@ def test_tensor_core_batch_matmal():
     s[C].tensorize(kernel_i, intrin_wmma_store_matrix((32, 8, 16)))
     s[CF].tensorize(_i, intrin_wmma_gemm((32, 8, 16)))

-    func = tvm.build(s, [A, B, C], 'cuda')
+    with tvm.transform.PassContext(opt_level=3, trace=dumpir):
+        func = tvm.build(s, [A, B, C], 'cuda')

     ctx = tvm.gpu(0)
     a_np = np.random.uniform(size=(batch_size, nn, ll, 32, 16)).astype(A.dtype)
@@ -384,4 +387,4 @@ def test_tensor_core_batch_conv():

 if __name__ == '__main__':
     test_tensor_core_batch_matmal()
-    test_tensor_core_batch_conv()
+    #test_tensor_core_batch_conv()

[spectrometerHBH]

Thanks for your suggestion! @xqdan

Is your codebase up-to-date? It seems to me that the source file you referenced is different from that of mainline.

[spectrometerHBH]

And #5965 #6147 are PRs related to the problems you encountered. You can try to sync your codebase and see if it works now.

[xqdan]

@spectrometerHBH thanks! I checkout latest version, looks good now, close this now.