Lower reshape after eliminate_contiguous in a single pass

src/targets/gpu/CMakeLists.txt

@@ -225,6 +225,7 @@ add_library(migraphx_gpu
     hsa_chiplet.cpp
     kernel.cpp
     lower_device_ops.cpp
+    lower_reshape.cpp
     lowering.cpp
     loop.cpp
     lrn.cpp
@@ -237,7 +238,6 @@ add_library(migraphx_gpu
     prepare_reduce.cpp
     pooling.cpp
     problem_cache.cpp
-    propagate_reshape_layout.cpp
     rocblas.cpp
     schedule_model.cpp
     sync_device.cpp

src/targets/gpu/include/migraphx/gpu/lower_reshape.hpp

@@ -0,0 +1,89 @@
+/*
+ * The MIT License (MIT)
+ *
+ * Copyright (c) 2026 Advanced Micro Devices, Inc. All rights reserved.
+ *
+ * Permission is hereby granted, free of charge, to any person obtaining a copy
+ * of this software and associated documentation files (the "Software"), to deal
+ * in the Software without restriction, including without limitation the rights
+ * to use, copy, modify, merge, publish, distribute, sublicense, and/or sell
+ * copies of the Software, and to permit persons to whom the Software is
+ * furnished to do so, subject to the following conditions:
+ *
+ * The above copyright notice and this permission notice shall be included in
+ * all copies or substantial portions of the Software.
+ *
+ * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
+ * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
+ * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT.  IN NO EVENT SHALL THE
+ * AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
+ * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM,
+ * OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN
+ * THE SOFTWARE.
+ */
+#ifndef MIGRAPHX_GUARD_GPU_LOWER_RESHAPE_HPP
+#define MIGRAPHX_GUARD_GPU_LOWER_RESHAPE_HPP
+
+#include <migraphx/gpu/config.hpp>
+#include <string>
+
+namespace migraphx {
+inline namespace MIGRAPHX_INLINE_NS {
+
+struct module;
+
+namespace gpu {
+
+/**
+ * Lower `reshape` into a `reshape_lazy` view, inserting a repacking copy only when the
+ * input layout does not permit aliasing.
+ *
+ * A `reshape` preserves row-major logical element order, so it can be a free view over
+ * the input buffer whenever the output dims are expressible as a pure restriding of the
+ * input. Otherwise the bytes have to be moved first.
+ *
+ * This pass deliberately runs *after* `eliminate_contiguous`. Lowering inserts a
+ * `gpu::contiguous` for many ops and `eliminate_contiguous` then drops the redundant
+ * ones globally, so a reshape's real input layout is not settled until that pass has
+ * finished. Deciding here means deciding once, with the final layout in hand, instead
+ * of emitting a pessimistic copy during lowering and undoing it afterwards.
+ *
+ * For each `reshape`, in order:
+ *
+ *  1. If `reshape_lazy` can alias the input directly and lands on the element count the
+ *     `reshape` declared, emit just the `reshape_lazy`. No copy. The resulting view keeps
+ *     whatever strides aliasing produced, which need not be the ones `reshape` reported.
+ *
+ *  2. Otherwise, derive the memory order the input would need in order to be aliasable
+ *     by running the dim mapping backwards (output dims -> input dims), and emit a
+ *     `layout` copy into that order followed by the `reshape_lazy`. The `layout` is
+ *     wrapped in `gpu::precompile_op` and compiles to the existing pointwise copy
+ *     kernel, so this costs the same single kernel as a standardizing contiguous while
+ *     preserving the permutation a following op may want (e.g. NHWC into pooling).
+ *     An identity permutation is declined here and left to case 3, since a `layout`
+ *     that reorders nothing is just a standardizing copy and `gpu::contiguous` has a
+ *     prebuilt code object for it.
+ *
+ *  3. Otherwise fall back to a standardizing `gpu::contiguous` followed by the
+ *     `reshape_lazy`.
+ *
+ * Cases 1 and 2 need real strides to reason about, so they are skipped for range-based
+ * dynamic inputs, which have none; those always take case 3. Static inputs are lifted
+ * to symbolic literals so static and symbolic shapes share one code path.
+ *
+ * The 2 input form of `reshape` (`reshape(data, output_buffer)`, where the target shape
+ * is carried by a runtime-sized output buffer) is not matched and passes through
+ * un-lowered. No GPU copy op can express it: they all derive their kernel from an index
+ * space shared by source and destination, which a rank-changing copy does not have.
+ */
+struct MIGRAPHX_GPU_EXPORT lower_reshape
+{
+    std::string name() const { return "gpu::lower_reshape"; }
+    void apply(module& m) const;
+};
+
+} // namespace gpu
+} // namespace MIGRAPHX_INLINE_NS
+} // namespace migraphx
+
+#endif // MIGRAPHX_GUARD_GPU_LOWER_RESHAPE_HPP

src/targets/gpu/lower_reshape.cpp

@@ -0,0 +1,113 @@
+/*
+ * The MIT License (MIT)
+ *
+ * Copyright (c) 2026 Advanced Micro Devices, Inc. All rights reserved.
+ *
+ * Permission is hereby granted, free of charge, to any person obtaining a copy
+ * of this software and associated documentation files (the "Software"), to deal
+ * in the Software without restriction, including without limitation the rights
+ * to use, copy, modify, merge, publish, distribute, sublicense, and/or sell
+ * copies of the Software, and to permit persons to whom the Software is
+ * furnished to do so, subject to the following conditions:
+ *
+ * The above copyright notice and this permission notice shall be included in
+ * all copies or substantial portions of the Software.
+ *
+ * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
+ * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
+ * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT.  IN NO EVENT SHALL THE
+ * AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
+ * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM,
+ * OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN
+ * THE SOFTWARE.
+ */
+#include <migraphx/gpu/lower_reshape.hpp>
+#include <migraphx/matcher.hpp>
+#include <migraphx/module.hpp>
+#include <migraphx/instruction.hpp>
+#include <migraphx/make_op.hpp>
+#include <migraphx/permutation.hpp>
+#include <migraphx/reshape_dims.hpp>
+#include <migraphx/sym.hpp>
+#include <migraphx/value.hpp>
+#include <algorithm>
+
+namespace migraphx {
+inline namespace MIGRAPHX_INLINE_NS {
+namespace gpu {
+
+namespace {
+instruction_ref
+insert_precompile_op(module& m, instruction_ref pos, instruction_ref input, const operation& op)
+{
+    auto output_shape = op.compute_shape({input->get_shape()});
+    auto alloc =
+        m.insert_instruction(pos, make_op("allocate", {{"shape", to_value(output_shape)}}));
+    return m.insert_instruction(
+        pos, make_op("gpu::precompile_op", {{"op", to_value(op)}}), input, alloc);
+}
+
+instruction_ref insert_contiguous(module& m, instruction_ref pos, instruction_ref input)
+{
+    const auto& s      = input->get_shape();
+    shape output_shape = s.dynamic() ? shape{s.type(), s.dyn_dims()} : shape{s.type(), s.lens()};
+    auto alloc =
+        m.insert_instruction(pos, make_op("allocate", {{"shape", to_value(output_shape)}}));
+    return m.insert_instruction(pos, make_op("gpu::contiguous"), input, alloc);
+}
+
+struct find_reshape : match::supports_dynamic_shapes
+{
+    // Skip reshape(data, output_buffer). Every GPU copy op derives its kernel from one
+    // index space shared by source and destination, so none of them can change rank.
+    auto matcher() const { return match::name("reshape")(match::nargs(1)); }
+
+    void apply(module& m, const match::matcher_result& r) const
+    {
+        auto ins        = r.result;
+        auto dims       = ins->get_operator().to_value().at("dims");
+        auto reshape_op = make_op("reshape_lazy", {{"dims", {dims}}});
+        auto input      = ins->inputs().front();
+        const auto& s   = input->get_shape();
+
+        if(not s.dynamic() or s.symbolic())
+        {
+            auto expected    = ins->get_shape().to_symbolic();
+            auto output_dims = ins->get_shape().sym_dims();
+            auto reshaped    = reshape_dims(s.to_symbolic(), output_dims, {.lazy = true});
+            if(reshaped and sym::same_symbol(reshaped->sym_elements(), expected.sym_elements()))
+            {
+                m.replace_instruction(ins, reshape_op, {input});
+                return;
+            }
+
+            auto relayout =
+                reshape_dims(ins->get_shape().to_symbolic(), s.sym_dims(), {.lazy = true});
+            if(relayout)
+            {
+                auto perm = find_permutation(*relayout);
+                // An identity permutation means the input only needs standardizing, which is
+                // what gpu::contiguous already does. Prefer it: lower_device_ops turns it into
+                // a prebuilt code object, whereas a layout under gpu::precompile_op would pay
+                // for a jit compile to produce the same copy.
+                if(not std::is_sorted(perm.begin(), perm.end()))
+                {
+                    auto layout = insert_precompile_op(
+                        m, ins, input, make_op("layout", {{"permutation", perm}}));
+                    m.replace_instruction(ins, reshape_op, {layout});
+                    return;
+                }
+            }
+        }
+
+        auto contiguous = insert_contiguous(m, ins, input);
+        m.replace_instruction(ins, reshape_op, {contiguous});
+    }
+};
+} // namespace
+
+void lower_reshape::apply(module& m) const { match::find_matches(m, find_reshape{}); }
+
+} // namespace gpu
+} // namespace MIGRAPHX_INLINE_NS
+} // namespace migraphx

src/targets/gpu/lowering.cpp

@@ -39,9 +39,7 @@
 #include <migraphx/op/common.hpp>
 #include <migraphx/op/dot.hpp>
 #include <migraphx/op/if_op.hpp>
-#include <migraphx/op/reshape.hpp>
 #include <migraphx/op/quant_dot.hpp>
-#include <migraphx/op/reshape_lazy.hpp>
 
 #include <migraphx/gpu/context.hpp>
 #include <migraphx/gpu/lowering.hpp>
@@ -113,7 +111,6 @@ struct miopen_apply
         add_nms_op();
         add_convolution_backwards_op();
         add_select_module_op();
-        add_reshape_lazy_op();
         add_concat_past_present_op();
         add_scan_slice_op();
         add_fill_op();
@@ -632,32 +629,6 @@ struct miopen_apply
         });
     }
 
-    /**
-     *  Adds reshape lazy to reshape ops that can be aliased instead of copied.
-     *  `gpu::contiguous` are added before and after the reshape; these contiguous
-     *  instructions can be removed by the eliminate_contiguous pass.
-     */
-    void add_reshape_lazy_op()
-    {
-        apply_map.emplace("reshape", [=](instruction_ref ins) {
-            std::vector<instruction_ref> before_contiguous_args = ins->inputs();
-            auto before_alloc = insert_allocation(ins, std::prev(ins)->get_shape());
-            before_contiguous_args.push_back(before_alloc);
-            auto before_contig =
-                mod->insert_instruction(ins, make_op("gpu::contiguous"), {before_contiguous_args});
-
-            auto new_reshape_lazy = mod->insert_instruction(
-                ins,
-                make_op("reshape_lazy", {{"dims", {ins->get_operator().to_value().at("dims")}}}),
-                before_contig);
-
-            std::vector<instruction_ref> after_contiguous_args = {new_reshape_lazy};
-            auto after_alloc = insert_allocation(new_reshape_lazy, new_reshape_lazy->get_shape());
-            after_contiguous_args.push_back(after_alloc);
-            return mod->replace_instruction(ins, make_op("gpu::contiguous"), after_contiguous_args);
-        });
-    }
-
     void add_concat_past_present_op()
     {
         apply_map.emplace("concat_past_present", [=](instruction_ref ins) {