Feature vpto backend tinsert new4

lib/PTO/IR/PTO.cpp

@@ -3621,7 +3621,8 @@ static LogicalResult verifyTColArgReductionOpCommon(Operation *op, Type srcTy,
                                            /*requireNonZeroSrc=*/true)))
     return failure();
   Type srcElemTy = getElemTy(srcTy);
-  unsigned srcElemBits = srcElemTy ? srcElemTy.getIntOrFloatBitWidth() : 0;
+  unsigned srcElemBytes = srcElemTy ? getPTOStorageElemByteSize(srcElemTy) : 0;
+  unsigned srcElemBits = srcElemBytes * 8;
   if (!(mlir::isa<IntegerType, FloatType>(srcElemTy) &&
         (srcElemBits == 8 || srcElemBits == 16 || srcElemBits == 32)))
     return op->emitOpError(
@@ -5324,9 +5325,29 @@ mlir::LogicalResult mlir::pto::TInsertOp::verify() {
       return success();
     }
 
+    // A5 acc->vec path (L0C->UB, ND/NZ modes in pto-isa).
+    if (*srcSpace == pto::AddressSpace::ACC && *dstSpace == pto::AddressSpace::VEC) {
+      if (!isColMajorRowMajorNZ(srcTb))
+        return emitOpError(
+            "expects A5 acc->vec tinsert src to use blayout=col_major and slayout=row_major");
+      bool dstIsND = isRowMajorNoneBoxND(dstTb);
+      bool dstIsNZ = isColMajorRowMajorNZ(dstTb);
+      if (!dstIsND && !dstIsNZ)
+        return emitOpError(
+            "expects A5 acc->vec tinsert dst to use ND (row_major/none_box) or NZ (col_major/row_major) layout");
+      bool okTypes = (srcElem.isF32() &&
+                      (dstElem.isF16() || dstElem.isBF16() || dstElem.isF32())) ||
+                     (srcElem.isInteger(32) && dstElem.isInteger(32));
+      if (!okTypes)
+        return emitOpError(
+            "expects A5 acc->vec tinsert element types to be "
+            "(src=f32,dst=f16/bf16/f32) or (src=i32,dst=i32)");
+      return success();
+    }
+
     return emitOpError(
         "expects A5 tinsert to use a supported src/dst loc pair: "
-        "acc->mat, vec->mat, or vec->vec");
+        "acc->mat, acc->vec, vec->mat, or vec->vec");
   };
   return dispatchVerifierByArch(getOperation(), verifyA2A3, verifyA5);
 }
@@ -5696,8 +5717,8 @@ llvm::LogicalResult mlir::pto::TGatherOp::verify() {
     if (!srcSpace || !dstSpace || *srcSpace != pto::AddressSpace::VEC ||
         *dstSpace != pto::AddressSpace::VEC)
       return emitOpError("expects src and dst to be in the vec address space");
-    unsigned srcElemBytes = srcElem.getIntOrFloatBitWidth() / 8;
-    unsigned dstElemBytes = dstElem.getIntOrFloatBitWidth() / 8;
+    unsigned srcElemBytes = getPTOStorageElemByteSize(srcElem);
+    unsigned dstElemBytes = getPTOStorageElemByteSize(dstElem);
     if (srcElemBytes != dstElemBytes)
       return emitOpError("expects src and dst element sizes to match");
 
@@ -9457,7 +9478,7 @@ mlir::LogicalResult mlir::pto::TTransOp::verify() {
       if (srcTb.getBLayoutValueI32() != static_cast<int32_t>(pto::BLayout::RowMajor))
         return emitOpError() << "expects A2/A3 transpose src to use the row_major blayout";
     }
-    unsigned elemBytes = srcElem.getIntOrFloatBitWidth() / 8;
+    unsigned elemBytes = getPTOStorageElemByteSize(srcElem);
     if (elemBytes != 1 && elemBytes != 2 && elemBytes != 4)
       return emitOpError() << "expects transpose element size to be 1, 2, or 4 bytes";
     auto isAllowedWidthType = [&](Type ty) {
@@ -9484,7 +9505,7 @@ mlir::LogicalResult mlir::pto::TTransOp::verify() {
     Type dstElem = getElemTy(dstTy);
     if (!srcElem || !tmpElem || !dstElem || srcElem != dstElem || srcElem != tmpElem)
       return emitOpError() << "expects src, tmp, and dst to have the same element type";
-    unsigned elemBytes = srcElem.getIntOrFloatBitWidth() / 8;
+    unsigned elemBytes = getPTOStorageElemByteSize(srcElem);
     if (elemBytes != 1 && elemBytes != 2 && elemBytes != 4)
       return emitOpError() << "expects transpose element size to be 1, 2, or 4 bytes";
     auto isAllowedWidthType = [&](Type ty) {

lib/PTO/IR/VPTO.cpp

@@ -2625,12 +2625,21 @@ static ParseResult parseStructuredAccStoreClauses(
     if (seenClause) {
       if (failed(parser.parseOptionalComma()))
         return success();
+    } else {
+      // First iteration: optionally consume the comma between operands and
+      // clauses. For some callers (MteL0cUbOp after dst_mode parsing with
+      // parseOptionalComma already consumed), the comma may already be gone.
+      (void)parser.parseOptionalComma();
     }
     StringRef keyword;
-    if (parser.parseKeyword(&keyword)) {
+    if (failed(parser.parseOptionalKeyword(&keyword))) {
+      // No keyword found. If this is the first iteration (no clauses yet),
+      // that means there are zero clauses — return success without error.
+      // On subsequent iterations, a keyword is expected after the comma.
       if (!seenClause)
         return success();
-      return failure();
+      return parser.emitError(parser.getCurrentLocation(),
+                              "expected mte_l0c clause keyword");
     }
     seenClause = true;
 
@@ -6996,7 +7005,7 @@ ParseResult MteL0cL1Op::parse(OpAsmParser &parser, OperationState &result) {
       parseRequiredOperandWithComma(parser, m) ||
       parseRequiredOperandWithComma(parser, n) ||
       parseRequiredOperandWithComma(parser, srcStride) ||
-      parseRequiredOperandWithComma(parser, dstStride) ||
+      parser.parseOperand(dstStride) ||
       parseStructuredAccStoreClauses(parser, state) ||
       parser.parseOptionalAttrDict(result.attributes) || parser.parseColon())
     return failure();
@@ -7164,7 +7173,7 @@ ParseResult MteL0cGmOp::parse(OpAsmParser &parser, OperationState &result) {
       parseRequiredOperandWithComma(parser, srcStride) ||
       parseRequiredOperandWithComma(parser, dstStride) ||
       parseRequiredOperandWithComma(parser, sid) ||
-      parseRequiredOperandWithComma(parser, l2CacheCtrl) ||
+      parser.parseOperand(l2CacheCtrl) ||
       parseStructuredAccStoreClauses(parser, state) ||
       parser.parseOptionalAttrDict(result.attributes) || parser.parseColon())
     return failure();

lib/PTO/Transforms/FoldTileBufIntrinsics.cpp

@@ -95,7 +95,12 @@ static std::optional<TileHandleInfo> resolveTileHandle(Value tileBuf,
 
   user->emitError("FoldTileBufIntrinsics: expected tile_buf to be defined by "
                   "the active materialized tile-handle bridge "
-                  "(pto.alloc_tile or pto.materialize_tile)");
+                  "(pto.alloc_tile or pto.materialize_tile)")
+      << "; got value of type: " << tileBuf.getType()
+      << "; defined by: "
+      << (tileBuf.getDefiningOp()
+              ? tileBuf.getDefiningOp()->getName().getStringRef()
+              : StringRef("block-argument"));
   return std::nullopt;
 }
 

lib/PTO/Transforms/GraphSyncSolver/MemInfo.cpp

@@ -11,6 +11,7 @@
 
 #include "PTO/Transforms/GraphSyncSolver/MemInfo.h"
 #include "PTO/IR/PTO.h"
+#include "PTO/IR/PTOTypeUtils.h"
 #include "../Utils.h"
 #include "mlir/Dialect/Arith/IR/Arith.h"
 #include "mlir/IR/BuiltinTypeInterfaces.h"
@@ -29,11 +30,11 @@ static std::optional<int64_t> getBufferBitSize(Value value) {
     return ShapedType::kDynamic;
   }
   Type elementType = shaped.getElementType();
-  auto bitWidth = elementType.getIntOrFloatBitWidth();
-  if (bitWidth == 0) {
+  unsigned byteSize = getPTOStorageElemByteSize(elementType);
+  if (byteSize == 0) {
     return ShapedType::kDynamic;
   }
-  return shaped.getNumElements() * bitWidth;
+  return shaped.getNumElements() * byteSize * 8;
 }
 
 llvm::SmallVector<int64_t> getAddresses(const llvm::SmallVector<Value> &addrs) {

lib/PTO/Transforms/InsertSync/PTOIRTranslator.cpp

@@ -12,6 +12,7 @@
 // See LICENSE in the root of the software repository for the full text of the License.
 
 #include "PTO/Transforms/InsertSync/PTOIRTranslator.h"
+#include "PTO/IR/PTOTypeUtils.h"
 #include "mlir/Dialect/Arith/IR/Arith.h"
 #include "mlir/Dialect/SCF/IR/SCF.h"
 #include "mlir/Dialect/MemRef/IR/MemRef.h"
@@ -20,7 +21,6 @@
 #include "llvm/Support/FormatVariadic.h"
 #include "mlir/Dialect/MemRef/IR/MemRef.h"
 #include "mlir/IR/Matchers.h"
-// [P0 新增] 引入副作用接口和 PTO 接口
 #include "mlir/Interfaces/SideEffectInterfaces.h"
 
 #define DEBUG_TYPE "pto-ir-translator"
@@ -35,7 +35,7 @@ static std::pair<int64_t, int64_t> getStaticOffsetAndSize(Operation *op, Value s
   auto srcType = dyn_cast<MemRefType>(src.getType());
   if (!srcType) return {0, 0};
 
-  int64_t elemSize = srcType.getElementType().getIntOrFloatBitWidth() / 8;
+  int64_t elemSize = getPTOStorageElemByteSize(srcType.getElementType());
   if (elemSize == 0) elemSize = 1;
 
   // === Case 1: memref.subview ===
@@ -235,7 +235,7 @@ LogicalResult PTOIRTranslator::UpdateAllocTileOpMemInfo(pto::AllocTileOp op) {
     }
 
     if (isStatic) {
-      int64_t elemSize = tileType.getElementType().getIntOrFloatBitWidth() / 8;
+      int64_t elemSize = getPTOStorageElemByteSize(tileType.getElementType());
       int64_t numElements = 1;
       for (auto dim : shape) numElements *= dim;
       sizeInBytes = numElements * elemSize;
@@ -280,7 +280,7 @@ LogicalResult PTOIRTranslator::UpdatePointerCastOpMemInfo(pto::PointerCastOp op)
 
   uint64_t sizeInBytes = 0;
   if (memRefType.hasStaticShape()) {
-    int64_t elemSize = memRefType.getElementType().getIntOrFloatBitWidth() / 8;
+    int64_t elemSize = getPTOStorageElemByteSize(memRefType.getElementType());
     int64_t numElements = 1;
     for (auto dim : memRefType.getShape()) numElements *= dim;
     sizeInBytes = numElements * elemSize;
@@ -314,7 +314,7 @@ PTOIRTranslator::UpdateDeclareTileMemRefOpMemInfo(pto::DeclareTileMemRefOp op) {
 
   uint64_t sizeInBytes = 0;
   if (memRefType.hasStaticShape()) {
-    int64_t elemSize = memRefType.getElementType().getIntOrFloatBitWidth() / 8;
+    int64_t elemSize = getPTOStorageElemByteSize(memRefType.getElementType());
     if (elemSize == 0)
       elemSize = 1;
 
@@ -593,8 +593,8 @@ LogicalResult PTOIRTranslator::UpdateMemrefAllocOpMemInfo(memref::AllocOp op) {
   // 1. 计算大小 (Bytes)
   uint64_t sizeInBytes = 0;
   if (memRefType.hasStaticShape()) {
-    int64_t elemSize = memRefType.getElementType().getIntOrFloatBitWidth() / 8;
-    if (elemSize == 0) elemSize = 1; // bool case
+    int64_t elemSize = getPTOStorageElemByteSize(memRefType.getElementType());
+    if (elemSize == 0) elemSize = 1;
 
     int64_t numElements = 1;
     for (auto dim : memRefType.getShape()) numElements *= dim;

lib/PTO/Transforms/VPTOExpandWrapperOps.cpp

@@ -710,10 +710,9 @@ struct LoadCbufToCbControl {
 static FailureOr<LoadCbufToCbControl>
 deriveLoadCbufToCbControl(Location loc, Value k, Value n, Type elementType,
                           bool transpose, PatternRewriter &rewriter) {
-  unsigned elemBitWidth = elementType.getIntOrFloatBitWidth();
-  if (elemBitWidth == 0 || (elemBitWidth % 8) != 0)
+  uint64_t elemBytes = pto::getPTOStorageElemByteSize(elementType);
+  if (elemBytes == 0)
     return failure();
-  uint64_t elemBytes = elemBitWidth / 8;
 
   auto constant = [&](uint64_t value) -> Value {
     return rewriter.create<arith::ConstantIntOp>(loc, value, 64);
@@ -749,10 +748,9 @@ deriveLoadCbufToCbControl(Location loc, Value k, Value n, Type elementType,
 static FailureOr<LoadCbufToCbControl>
 deriveLoadCbufToCaControl(Location loc, Value m, Value k, Type elementType,
                           bool transpose, PatternRewriter &rewriter) {
-  unsigned elemBitWidth = elementType.getIntOrFloatBitWidth();
-  if (elemBitWidth == 0 || (elemBitWidth % 8) != 0)
+  uint64_t elemBytes = pto::getPTOStorageElemByteSize(elementType);
+  if (elemBytes == 0)
     return failure();
-  uint64_t elemBytes = elemBitWidth / 8;
 
   auto constant = [&](uint64_t value) -> Value {
     return rewriter.create<arith::ConstantIntOp>(loc, value, 64);
@@ -1324,10 +1322,9 @@ struct ExpandRightLoadMxPattern : public OpRewritePattern<pto::MteL1L0bMxOp> {
     if (!sourceType)
       return rewriter.notifyMatchFailure(op, "expected typed L1 source");
 
-    unsigned elemBitWidth = sourceType.getElementType().getIntOrFloatBitWidth();
-    if (elemBitWidth == 0 || (elemBitWidth % 8) != 0)
+    uint64_t elemBytes = pto::getPTOStorageElemByteSize(sourceType.getElementType());
+    if (elemBytes == 0)
       return rewriter.notifyMatchFailure(op, "unsupported element type");
-    uint64_t elemBytes = elemBitWidth / 8;
 
     auto constant = [&](uint64_t value) -> Value {
       return rewriter.create<arith::ConstantIntOp>(loc, value, 64);

lib/PTO/Transforms/VPTOLLVMEmitter.cpp

@@ -4833,11 +4833,10 @@ class LowerLoadCbufToCaMxOpPattern final
       return rewriter.notifyMatchFailure(op, "failed to map cbuf/ca pointer spaces");
 
     Type sourceElemType = cast<pto::PtrType>(op.getSource().getType()).getElementType();
-    unsigned elemBitWidth = sourceElemType.getIntOrFloatBitWidth();
-    if (elemBitWidth == 0 || (elemBitWidth % 8) != 0)
+    unsigned elemBytes = getPTOStorageElemByteSize(sourceElemType);
+    if (elemBytes == 0)
       return rewriter.notifyMatchFailure(op,
                                          "unsupported load_cbuf_to_ca_mx element type");
-    uint64_t elemBytes = elemBitWidth / 8;
     Location loc = op.getLoc();
     auto constant = [&](uint64_t value) -> Value {
       return rewriter.create<arith::ConstantIntOp>(loc, value, 64);
@@ -4911,8 +4910,8 @@ class LowerLoadCbufToCbMxOpPattern final
       return rewriter.notifyMatchFailure(op, "failed to map cbuf/cb pointer spaces");
 
     Type sourceElemType = cast<pto::PtrType>(op.getSource().getType()).getElementType();
-    unsigned elemBitWidth = sourceElemType.getIntOrFloatBitWidth();
-    if (elemBitWidth == 0 || (elemBitWidth % 8) != 0)
+    unsigned elemBytes = getPTOStorageElemByteSize(sourceElemType);
+    if (elemBytes == 0)
       return rewriter.notifyMatchFailure(op,
                                          "unsupported load_cbuf_to_cb_mx element type");
     FailureOr<Value> config0 =