fix(vmi): preserve compact VL4 widening layout

CMakeLists.txt

@@ -8,8 +8,6 @@
 
 cmake_minimum_required(VERSION 3.20.0)
 
-include(cmake/fetch_cann_cmake.cmake)
-
 # CANN/Ascend CMake packages used by downstream PTOAS builds may still use the
 # pre-CMake-3.30 FetchContent_Populate(<declared-name>) pattern. Keep that
 # compatibility path enabled so CMake 4.x developer warnings do not become
@@ -29,6 +27,8 @@ if(POLICY CMP0175)
   set(CMAKE_POLICY_DEFAULT_CMP0175 OLD)
 endif()
 
+include(cmake/fetch_cann_cmake.cmake)
+
 # Standard Python build backends pass environment variables through to CMake
 # but do not synthesize project-specific cache arguments. Accept the existing
 # LLVM_BUILD_DIR convention directly so `pip wheel .` and `pip install .` can

cmake/fetch_cann_cmake.cmake

@@ -1,14 +1,19 @@
-# -----------------------------------------------------------------------------------------------------------
 # Copyright (c) 2026 Huawei Technologies Co., Ltd.
 # This program is free software, you can redistribute it and/or modify it under the terms and conditions of
 # CANN Open Software License Agreement Version 2.0 (the "License").
 # Please refer to the License for details. You may not use this file except in compliance with the License.
 # THIS SOFTWARE IS PROVIDED ON AN "AS IS" BASIS, WITHOUT WARRANTIES OF ANY KIND, EITHER EXPRESS OR IMPLIED,
 # INCLUDING BUT NOT LIMITED TO NON-INFRINGEMENT, MERCHANTABILITY, OR FITNESS FOR A PARTICULAR PURPOSE.
 # See LICENSE in the root of the software repository for the full text of the License.
-# -----------------------------------------------------------------------------------------------------------
 
 if(NOT PROJECT_SOURCE_DIR)
+    # Standalone builds may configure without a pre-seeded CANN third-party
+    # directory (e.g. CI). Keep FetchContent's declared SOURCE_DIR inside the
+    # writable CMake binary tree instead of resolving an empty variable to the
+    # filesystem root. Script-mode packaging still requires an explicit path.
+    if(NOT CANN_3RD_LIB_PATH AND CMAKE_BINARY_DIR)
+        set(CANN_3RD_LIB_PATH "${CMAKE_BINARY_DIR}/cann-3rd-lib")
+    endif()
     # Temporary test pin for cann/cmake MR !277. Revert this URL/ref pair
     # after validation and switch to the released cann/cmake tag once MR !277
     # is merged.

docs/isa/vmi-isa/06-convert.md

@@ -58,9 +58,9 @@
   | Attribute | Values | Valid for | Description |
   |---|---|---|---|
   | `rounding` | `"R"` (nearest-even), `"A"` (away-from-zero), `"H"` (half-up), `"Z"` (toward-zero); for the `bf16x2→f4x2` contract pair the allowed set is `"R"`,`"A"`,`"F"` (floor), `"C"` (ceil), `"Z"` (toward-zero) — `"H"` is **rejected** | fp narrowing | Rounding mode |
-  | `saturate` | `"SAT"`, `"NOSAT"` | required for fp-narrow / int-narrow; for fp→si / fp→ui the requirement follows the vcvt contract's `requiresSat` (e.g. `f16→s8` required, `f16→s32` **forbidden** — no overflow possible; same-width `bf16→f16` required, same-width `f16→bf16` **forbidden**); the `bf16x2→f4x2` narrow has `requiresSat=false` — any `saturate` is **forbidden** | `SAT` clamps to ±max of the destination type; `NOSAT` performs a direct bit truncation of the result representation. |
+  | `saturate` | `"SAT"`, `"NOSAT"` | required for fp-narrow / int-narrow; for fp→si / fp→ui the requirement follows the vcvt contract's `requiresSat` (e.g. `f16→s8` required, `f16→s32` **forbidden** — no overflow possible; same-width `bf16→f16` required, same-width `f16→bf16` **forbidden**); the `bf16x2→f4x2` narrow has `requiresSat=false` — any `saturate` is **forbidden** | For signed destinations, `SAT` clamps to `[min, max]`; for unsigned or signless destinations, it clamps to `[0, max]`. `NOSAT` performs a direct bit truncation of the result representation. |
 
-- **datatypes:** Source and destination from `{f32, f16, bf16, fp8_e4m3, fp8_e5m2, i32, i16, i8, ui32, ui16, ui8}`; packed carrier types `{!pto.bf16x2, !pto.f4E1M2x2, !pto.f4E2M1x2}` for the bf16x2↔f4x2 fp-to-fp pair (see contract `lookupVMIFpToFpContract`). `bf16x2` is **conversion-only** — it may not appear as a compute element type (`vfadd`/`vfmul`/`vcmp`/...).
+- **datatypes:** Source and destination from `{f32, f16, bf16, fp8_e4m3, fp8_e5m2, i32, i16, i8, si32, si16, si8, ui32, ui16, ui8}`; packed carrier types `{!pto.bf16x2, !pto.f4E1M2x2, !pto.f4E2M1x2}` for the bf16x2↔f4x2 fp-to-fp pair (see contract `lookupVMIFpToFpContract`). `bf16x2` is **conversion-only** — it may not appear as a compute element type (`vfadd`/`vfmul`/`vcmp`/...). Signless `iN` is treated as unsigned; use `siN` for signed conversion semantics.
 - **lowering to `pto.mi`:**
 
   | Conversion | Physical lowering | `#mi` | `dep` |
@@ -128,7 +128,6 @@
 - **notes:**
   - `vcvt` **does not change lane count** — `src.L == dst.L` always. The
     physical register count `K` changes because `bitwidth(T)` changes.
-  - Integer signedness is determined by the **element type**.
   - The `part`/`parity`/`width` axes are lowering-only; the user never writes
     `EVEN`/`ODD`/`P0..P3`.
   - Radix-4 (8↔32) is **not** a stacked predicate chain and **not** a UB

lib/PTO/Transforms/VMILayoutAssignment.cpp

@@ -29,6 +29,8 @@
 #include "llvm/ADT/SmallVector.h"
 #include "llvm/ADT/TypeSwitch.h"
 
+#include <type_traits>
+
 namespace mlir {
 namespace pto {
 #define GEN_PASS_DEF_VMILAYOUTASSIGNMENT
@@ -693,6 +695,34 @@ struct LayoutSolver {
   WalkResult addExtensionConstraint(CastOp castOp, Operation *op) {
     auto sourceType = cast<VMIVRegType>(castOp.getSource().getType());
     auto resultType = cast<VMIVRegType>(castOp.getResult().getType());
+
+    // A four-lane byte/halfword -> ui32 widening is a compact group-slot
+    // operation.  Seed the result with the group-slot carrier so propagation
+    // selects the existing group-slot extension lowering (and avoids creating
+    // an unsupported group-slot -> dense lane-stride ensure_layout).
+    if constexpr (std::is_same_v<CastOp, VMIExtUIOp>) {
+      auto sourceInteger = dyn_cast<IntegerType>(sourceType.getElementType());
+      auto resultInteger = dyn_cast<IntegerType>(resultType.getElementType());
+      unsigned sourceBits =
+          pto::getPTOStorageElemBitWidth(sourceType.getElementType());
+      unsigned resultBits =
+          pto::getPTOStorageElemBitWidth(resultType.getElementType());
+      bool compactUI32Widen =
+          sourceType.getElementCount() == 4 && sourceInteger && resultInteger &&
+          !sourceType.getLayoutAttr() && !resultType.getLayoutAttr() &&
+          (resultInteger.isUnsigned() || resultInteger.isSignless()) &&
+          (sourceBits == 8 || sourceBits == 16) &&
+          resultBits == 32;
+      if (compactUI32Widen) {
+        VMILayoutAttr compactLayout =
+            VMILayoutAttr::getGroupSlots(ctx, /*numGroups=*/4,
+                                         /*slots=*/mlir::pto::kValue8);
+        return *constraintResult(setPreferredLayout(
+            castOp.getResult(), compactLayout, op,
+            DataLayoutSeedPhase::CompactCast));
+      }
+    }
+
     FailureOr<VMICastLayoutFact> fact =
         VMILayoutSupport().getPreferredCastLayoutFact(sourceType, resultType);
     if (failed(fact)) {

packaging/ptoas-vmi/pyproject.toml.patch

@@ -1,7 +1,9 @@
 diff --git a/pyproject.toml b/pyproject.toml
 --- a/pyproject.toml
 +++ b/pyproject.toml
-@@ -18,7 +18,7 @@ build-backend = "scikit_build_core.build"
+@@ -16,11 +16,11 @@
+ build-backend = "scikit_build_core.build"
+
  [project]
 -name = "ptoas"
 -dynamic = ["version"]
@@ -10,20 +12,25 @@ diff --git a/pyproject.toml b/pyproject.toml
 +version = "0.1.6"
 +description = "PTO Assembler & Optimizer with VMI support"
  readme = "README.md"
- requires-python = ">=3.10"
+-requires-python = ">=3.7"
++requires-python = ">=3.10"
  license = "Apache-2.0"
-@@ -43,4 +43,5 @@ cmake.version = "CMakeLists.txt"
+ dependencies = [
+     "numpy",
+@@ -43,6 +43,7 @@
  ninja.version = ">=1.10"
  build.verbose = false
  editable.mode = "redirect"
 +sdist.inclusion-mode = "manual"
  install.components = ["PTOAS_Python"]
-@@ -53,3 +54,4 @@ TileOps = "lib/TileOps"
+
+ [tool.scikit-build.wheel.packages]
+@@ -53,8 +54,5 @@
+
  [tool.scikit-build.cmake.define]
  PTOAS_RELEASE_VERSION_OVERRIDE = { env = "PTOAS_RELEASE_VERSION_OVERRIDE", default = "" }
 +PTOAS_CLI_VERSION_LABEL = "vmi"
-
-@@ -56,4 +58,0 @@ PTOAS_RELEASE_VERSION_OVERRIDE = { env = "PTOAS_RELEASE_VERSION_OVERRIDE", default = "" }
+
 -[tool.scikit-build.metadata.version]
 -provider = "scikit_build_core.metadata.regex"
 -input = "CMakeLists.txt"

ptodsl/docs/user_guide/14-vmi-virtual-instruction-set.md

@@ -881,8 +881,8 @@ group_max = pto.vmi.vcmax(
 Converts the element type of `source` to the target element type. PTODSL
 infers the result vector type from the source lane count/layout and `to_dtype`.
 
-For int→int widening, the source element type must carry signedness
-(e.g. `si8`/`ui8`/`si16`/`ui16`); signless integers are rejected.
+For `vcvt`, signless integer types (`i8`/`i16`/`i32`) are treated as
+unsigned. Use `siN` types for signed conversion semantics.
 
 **Parameters**:
 

test/lit/vmi_new/vmi_to_vpto_extui_compact_vl4.pto

@@ -0,0 +1,56 @@
+// Copyright (c) 2026 Huawei Technologies Co., Ltd.
+// This program is free software, you can redistribute it and/or modify it under
+// the terms and conditions of the CANN Open Software License Agreement Version 2.0
+// (the "License"). Please refer to the License for details. You may not use
+// this file except in compliance with the License. THIS SOFTWARE IS PROVIDED ON
+// AN "AS IS" BASIS, WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express
+// or implied. See the License for the specific language governing permissions and
+// limitations under the License. See LICENSE in the root of the software repository
+// for the full text of the License.
+
+// RUN: pto-test-opt %s -vmi-lower-unified-to-legacy -vmi-mask-granularity-assignment -vmi-layout-assignment | FileCheck %s --check-prefix=ASSIGN
+// RUN: pto-test-opt %s -vmi-lower-unified-to-legacy -vmi-mask-granularity-assignment -vmi-layout-assignment -vmi-to-vpto | FileCheck %s --check-prefix=LOWER
+
+module {
+  func.func @extui_compact_vl4(
+      %src: !pto.ptr<i8, ub>, %dst: !pto.ptr<ui32, ub>, %off: index) {
+    %c1 = arith.constant 1 : index
+    %value = pto.vmi.vload %src[%off], %c1 {group = 4}
+        : !pto.ptr<i8, ub> -> !pto.vmi.vreg<4xi8>
+    %wide = pto.vmi.vcvt %value
+        : !pto.vmi.vreg<4xi8> -> !pto.vmi.vreg<4xui32>
+    pto.vmi.vstore %wide, %dst[%off], %c1 {group = 4}
+        : !pto.vmi.vreg<4xui32>, !pto.ptr<ui32, ub>
+    return
+  }
+
+  func.func @extui_compact_vl4_i16(
+      %src: !pto.ptr<i16, ub>, %dst: !pto.ptr<ui32, ub>, %off: index) {
+    %c1 = arith.constant 1 : index
+    %value = pto.vmi.vload %src[%off], %c1 {group = 4}
+        : !pto.ptr<i16, ub> -> !pto.vmi.vreg<4xi16>
+    %wide = pto.vmi.vcvt %value
+        : !pto.vmi.vreg<4xi16> -> !pto.vmi.vreg<4xui32>
+    pto.vmi.vstore %wide, %dst[%off], %c1 {group = 4}
+        : !pto.vmi.vreg<4xui32>, !pto.ptr<ui32, ub>
+    return
+  }
+}
+
+// ASSIGN-LABEL: func.func @extui_compact_vl4(
+// ASSIGN: %[[WIDE:.*]] = pto.vmi.extui %{{.*}} : !pto.vmi.vreg<4xi8, #pto.vmi.layout<num_groups = 4, slots = 8>> -> !pto.vmi.vreg<4xui32, #pto.vmi.layout<num_groups = 4, slots = 8>>
+
+// ASSIGN-LABEL: func.func @extui_compact_vl4_i16(
+// ASSIGN: %[[WIDE16:.*]] = pto.vmi.extui %{{.*}} : !pto.vmi.vreg<4xi16, #pto.vmi.layout<num_groups = 4, slots = 8>> -> !pto.vmi.vreg<4xui32, #pto.vmi.layout<num_groups = 4, slots = 8>>
+
+// LOWER-LABEL: func.func @extui_compact_vl4(
+// LOWER-COUNT-2: pto.vzunpack
+// LOWER: pto.vstus
+// LOWER-NOT: pto.vmi.
+// LOWER-NOT: unrealized_conversion_cast
+
+// LOWER-LABEL: func.func @extui_compact_vl4_i16(
+// LOWER: pto.vzunpack
+// LOWER: pto.vstus
+// LOWER-NOT: pto.vmi.
+// LOWER-NOT: unrealized_conversion_cast