Align matrix and expand emission with PTO model

docs/tileop-usage/generated/engines.md

@@ -1,6 +1,6 @@
 # LinxISA / PTO ISA v0.58.3 执行引擎
 
-架构定义的引擎类别只有 **VEC**、**TLSU**、**CUBE** 和 **SFU**。
+架构定义的引擎类别只有 **VEC**, **TLSU**, **CUBE**, and **SFU**。
 VEC 只包含逐元素操作；SFU 包含归约、广播、变换、排序以及其他需要更复杂硬件的操作。
 TEPL 仍是唯一的编译 carrier 标识。`BSTART.VEC` 和 `BSTART.SFU` 是特定引擎的汇编别名；
 inline wrapper 保留 `BSTART.TEPL`，以兼容之前的工具链源码。

include/jcore/template_asm.hpp

@@ -2657,7 +2657,7 @@ namespace pto_matmul_detail {
 
 #define PTO_MATMUL_HEADER(OPCODE, EXTRA_ATTRS)                                  \
   "BSTART.CUBE " OPCODE ", %D[DataTypeA]\n"                                      \
-  "B.DATR %D[DataTypeB], RNONE, NOSAT\n" EXTRA_ATTRS                     \
+  "B.DATR %D[DataTypeB], Zero, RNONE, NOSAT\n" EXTRA_ATTRS               \
   "B.DIM %[M], 0, ->lb0\n"                                                   \
   "B.DIM %[N], 0, ->lb1\n"                                                   \
   "B.DIM %[K], 0, ->lb2\n"
@@ -9526,7 +9526,11 @@ void TROWPROD(tile_shape_out &dst, tile_shape_in &src) {
 // TROWEXPAND: broadcast first element of each row
 template <is_tile_data_v tile_shape_out, is_tile_data_v tile_shape_in>
 void TROWEXPAND(tile_shape_out &dst, tile_shape_in &src) {
-  if constexpr (tile_shape_in::ValidCol > 0 && tile_shape_in::ValidRow > 0) {
+  static_assert(std::is_same<typename tile_shape_in::DType,
+                             typename tile_shape_out::DType>::value,
+                "TROWEXPAND: src/dst dtype must match");
+  if constexpr (tile_shape_out::ValidCol > 0 &&
+                tile_shape_out::ValidRow > 0) {
   asm volatile(
     "BSTART.TEPL 68, %D1\n"
     "B.DIM zero, %c2, ->lb0\n"
@@ -9536,9 +9540,9 @@ void TROWEXPAND(tile_shape_out &dst, tile_shape_in &src) {
     ""
     : "=Tr"(dst.data())
     : "i"(type_traits<typename tile_shape_in::DType>::TypeCode),
-      "i"(tile_shape_in::ValidCol),
-      "i"(tile_shape_in::ValidRow),
-      "i"(tile_shape_in::Cols),
+      "i"(tile_shape_out::ValidCol),
+      "i"(tile_shape_out::ValidRow),
+      "i"(tile_shape_out::Cols),
       "Tr"(src.data()),
       "i"(tile_type_traits<typename tile_shape_out::TileDType>::TilesizeCode)
   );
@@ -9552,9 +9556,9 @@ void TROWEXPAND(tile_shape_out &dst, tile_shape_in &src) {
     ""
     : "=Tr"(dst.data())
     : "i"(type_traits<typename tile_shape_in::DType>::TypeCode),
-      "r"(src.GetValidCol()),
-      "r"(src.GetValidRow()),
-      "i"(tile_shape_in::Cols),
+      "r"(dst.GetValidCol()),
+      "r"(dst.GetValidRow()),
+      "i"(tile_shape_out::Cols),
       "Tr"(src.data()),
       "i"(tile_type_traits<typename tile_shape_out::TileDType>::TilesizeCode)
   );
@@ -9792,7 +9796,11 @@ void TCOLPROD(tile_shape_out &dst, tile_shape_in &src) {
 // TCOLEXPAND: broadcast first element of each col
 template <is_tile_data_v tile_shape_out, is_tile_data_v tile_shape_in>
 void TCOLEXPAND(tile_shape_out &dst, tile_shape_in &src) {
-  if constexpr (tile_shape_in::ValidCol > 0 && tile_shape_in::ValidRow > 0) {
+  static_assert(std::is_same<typename tile_shape_in::DType,
+                             typename tile_shape_out::DType>::value,
+                "TCOLEXPAND: src/dst dtype must match");
+  if constexpr (tile_shape_out::ValidCol > 0 &&
+                tile_shape_out::ValidRow > 0) {
   asm volatile(
     "BSTART.TEPL 84, %D1\n"
     "B.DIM zero, %c2, ->lb0\n"
@@ -9802,9 +9810,9 @@ void TCOLEXPAND(tile_shape_out &dst, tile_shape_in &src) {
     ""
     : "=Tr"(dst.data())
     : "i"(type_traits<typename tile_shape_in::DType>::TypeCode),
-      "i"(tile_shape_in::ValidCol),
-      "i"(tile_shape_in::ValidRow),
-      "i"(tile_shape_in::Cols),
+      "i"(tile_shape_out::ValidCol),
+      "i"(tile_shape_out::ValidRow),
+      "i"(tile_shape_out::Cols),
       "Tr"(src.data()),
       "i"(tile_type_traits<typename tile_shape_out::TileDType>::TilesizeCode)
   );
@@ -9818,9 +9826,9 @@ void TCOLEXPAND(tile_shape_out &dst, tile_shape_in &src) {
     ""
     : "=Tr"(dst.data())
     : "i"(type_traits<typename tile_shape_in::DType>::TypeCode),
-      "r"(src.GetValidCol()),
-      "r"(src.GetValidRow()),
-      "i"(tile_shape_in::Cols),
+      "r"(dst.GetValidCol()),
+      "r"(dst.GetValidRow()),
+      "i"(tile_shape_out::Cols),
       "Tr"(src.data()),
       "i"(tile_type_traits<typename tile_shape_out::TileDType>::TilesizeCode)
   );

test/test_v058_engine_contract.py

@@ -218,7 +218,12 @@ def test_shared_tile_bindings_use_b_ios(self) -> None:
     def test_tlsu_load_store_stride_is_expressed_in_bytes(self) -> None:
         self.assertIn("GetStrideBytes", self.header)
         tlsu_doc = (ROOT / "docs" / "tileop-usage" / "tlsu" / "load-store-move" / "TLOAD.md").read_text(encoding="utf-8")
-        self.assertIn("row stride in **bytes**", tlsu_doc)
+        self.assertIn("stride", tlsu_doc)
+        self.assertRegex(
+            tlsu_doc,
+            r"row stride in \*\*bytes\*\*|\*\*字节 stride\*\*",
+        )
+        self.assertIn("B.IOR.RegSrc1", tlsu_doc)
 
     def test_tsel_binds_explicit_false_source_on_both_shape_paths(self) -> None:
         start = self.header.index("void TSEL(")
@@ -329,7 +334,7 @@ def test_local_cube_descriptor_contract_is_compile_time_guarded(self) -> None:
     def test_matrix_dtype_and_effective_shape_contract_is_centralized(self) -> None:
         tile = PTO_TILE.read_text(encoding="utf-8")
         self.assertIn(
-            '"B.DATR %D[DataTypeB], RNONE, NOSAT\\n"', self.header
+            '"B.DATR %D[DataTypeB], Zero, RNONE, NOSAT\\n"', self.header
         )
         self.assertIn("matrix_accumulator_type_code", tile)
         self.assertIn("MatrixNumericClass::Unsigned", tile)
@@ -360,6 +365,22 @@ def test_matrix_dtype_and_effective_shape_contract_is_centralized(self) -> None:
             self.assertIn("__fp8_e4m3", text)
             self.assertIn("__fp8_e8m0", text)
 
+    def test_copy_expand_describes_destination_geometry(self) -> None:
+        for mnemonic, next_comment in (
+            ("TROWEXPAND", "// TROWARGMAX:"),
+            ("TCOLEXPAND", "// TCOLARGMAX:"),
+        ):
+            start = self.header.index(
+                f"void {mnemonic}(tile_shape_out &dst, tile_shape_in &src)"
+            )
+            end = self.header.index(next_comment, start)
+            body = self.header[start:end]
+            self.assertIn('"r"(dst.GetValidCol())', body)
+            self.assertIn('"r"(dst.GetValidRow())', body)
+            self.assertIn('"i"(tile_shape_out::Cols)', body)
+            self.assertNotIn('"r"(src.GetValidCol())', body)
+            self.assertNotIn('"r"(src.GetValidRow())', body)
+
     def test_mc_gate_requires_canonical_cube_layout_names(self) -> None:
         gate = (ROOT / "test" / "tileop_api" /
                 "verify_pto0583_asm.sh").read_text()
@@ -490,7 +511,9 @@ def test_tquant_tdequant_use_datr_and_ior(self) -> None:
     def test_generated_engine_document_is_fresh(self) -> None:
         generated = ROOT / "docs" / "tileop-usage" / "generated" / "engines.md"
         self.assertTrue(generated.is_file())
-        self.assertIn("**VEC**, **TLSU**, **CUBE**, and **SFU**", generated.read_text())
+        text = generated.read_text()
+        for engine in ("VEC", "TLSU", "CUBE", "SFU"):
+            self.assertIn(f"**{engine}**", text)
 
     def test_test_harness_uses_v058_compiler_surface_without_install_mutation(self) -> None:
         makefile = (ROOT / "test" / "common" / "Makefile.common").read_text(encoding="utf-8")

test/tileop_api/src/TRowExpandMul.cpp

@@ -28,6 +28,7 @@ void test_row_vector_src1(T *dst, T *s0, T *s1) {
   TLOAD(d0, g0);
   TLOAD(d1, g1);
 
+  TROWEXPAND(d_out, d1);
   TROWEXPANDMUL(d_out, d0, d1);
   TROWEXPANDADD(d_out, d0, d1);
   TROWEXPANDSUB(d_out, d0, d1);
@@ -57,6 +58,7 @@ void test_col_vector_src1(T *dst, T *s0, T *s1) {
   TLOAD(d0, g0);
   TLOAD(d1, g1);
 
+  TCOLEXPAND(d_out, d1);
   TCOLEXPANDMUL(d_out, d0, d1);
   TCOLEXPANDADD(d_out, d0, d1);
   TCOLEXPANDSUB(d_out, d0, d1);