[Issue]: Try_compute_shape in eliminate_contiguous cannot compute the cumulative layout

[MQ-mengqing]

Problem Description

When I wrote a new test with follows diff, I got terminate called after throwing an instance of 'std::runtime_error'\n what(): Not standard shape (at develop branch)

diff --git a/test/eliminate_contiguous_test.cpp b/test/eliminate_contiguous_test.cpp
index 78095b0ff..00a8babd4 100644
--- a/test/eliminate_contiguous_test.cpp
+++ b/test/eliminate_contiguous_test.cpp
@@ -27,6 +27,7 @@
 #include <migraphx/instruction.hpp>
 #include <basic_ops.hpp>
 #include <migraphx/make_op.hpp>
+#include <migraphx/op/pooling.hpp>
 
 #include <pointwise.hpp>
 #include <test.hpp>
@@ -289,4 +290,30 @@ TEST_CASE(slice_contiguous)
                m.begin(), m.end(), [](auto&& ins) { return ins.name() == "contiguous"; }) == 1);
 }
 
+TEST_CASE(chain_concat_nhwc_propagation)
+{
+    migraphx::module m;
+
+    auto s_nhwc =
+        migraphx::shape::from_permutation(migraphx::shape::float_type, {2, 3, 8, 8}, {0, 2, 3, 1});
+    auto d  = m.add_parameter("d", s_nhwc);
+    auto cd = m.add_instruction(migraphx::make_op("contiguous"), d);
+
+    auto pool_op = migraphx::make_op("pooling",
+                                     {{"mode", migraphx::op::pooling_mode::max},
+                                      {"padding", {0, 0}},
+                                      {"stride", {1, 1}},
+                                      {"lengths", {1, 1}}});
+    auto a       = m.add_instruction(pool_op, cd);
+    auto b       = m.add_instruction(pool_op, a);
+    auto c       = m.add_instruction(pool_op, b);
+    auto cc      = m.add_instruction(migraphx::make_op("concat", {{"axis", 1}}), a, b, c);
+    m.add_instruction(pass_standard_op{}, cc);
+
+    auto count = std::distance(m.begin(), m.end());
+    run_pass(m);
+    EXPECT(std::distance(m.begin(), m.end()) == count);
+    EXPECT(std::any_of(m.begin(), m.end(), [](auto&& ins) { return ins.name() == "contiguous"; }));
+}
+
 int main(int argc, const char* argv[]) { test::run(argc, argv); }

That's because try_compute_shape recursed per output and substituted only the current instruction's shape per branch, leaving every other input at its get_shape(). For consumers whose output layout is decided by find_permutation voting across multiple inputs (e.g. concat), the unchanged NCHW siblings outvoted the single NHWC change at every recursion step, so the function never observed the cumulative effect of eliminating a chain of contiguous. The replaced contiguous then surfaced NHWC at consumers that require standard layout.

Operating System

Ubuntu 24.04.4 LTS (Noble Numbat)

CPU

AMD Ryzen 7 5800H with Radeon Graphics

GPU

Other

Other

No response

ROCm Version

ROCm 6.0.0

Steps to Reproduce

No response

(Optional for Linux users) Output of /opt/rocm/bin/rocminfo --support

No response

Additional Information

No response