From addb94039669b8094dea22885334e42a7909b761 Mon Sep 17 00:00:00 2001
From: jimmychou <1049859649@qq.com>
Date: Mon, 25 May 2026 16:48:05 +0800
Subject: [PATCH 1/6] Add precision template attrs for vector math ops

---
 include/PTO/IR/PTOAttrs.td                    | 104 +++++++++++
 include/PTO/IR/PTOOps.td                      |  30 ++-
 lib/PTO/Transforms/PTOToEmitC.cpp             | 173 +++++++++++++++++-
 lib/PTO/Transforms/PTOViewToMemref.cpp        |   9 +-
 .../lit/pto/tcolexpanddiv_precision_emitc.pto |  19 ++
 test/lit/pto/tdiv_precision_emitc.pto         |  21 +++
 test/lit/pto/texp_precision_emitc.pto         |  18 ++
 test/lit/pto/tfmod_precision_emitc.pto        |  19 ++
 test/lit/pto/tlog_precision_emitc.pto         |  18 ++
 test/lit/pto/trecip_precision_emitc.pto       |  18 ++
 test/lit/pto/trem_precision_emitc.pto         |  20 ++
 .../lit/pto/trowexpanddiv_precision_emitc.pto |  20 ++
 test/lit/pto/trsqrt_precision_emitc.pto       |  19 ++
 test/lit/pto/tsqrt_precision_emitc.pto        |  18 ++
 14 files changed, 483 insertions(+), 23 deletions(-)
 create mode 100644 test/lit/pto/tcolexpanddiv_precision_emitc.pto
 create mode 100644 test/lit/pto/tdiv_precision_emitc.pto
 create mode 100644 test/lit/pto/texp_precision_emitc.pto
 create mode 100644 test/lit/pto/tfmod_precision_emitc.pto
 create mode 100644 test/lit/pto/tlog_precision_emitc.pto
 create mode 100644 test/lit/pto/trecip_precision_emitc.pto
 create mode 100644 test/lit/pto/trem_precision_emitc.pto
 create mode 100644 test/lit/pto/trowexpanddiv_precision_emitc.pto
 create mode 100644 test/lit/pto/trsqrt_precision_emitc.pto
 create mode 100644 test/lit/pto/tsqrt_precision_emitc.pto

diff --git a/include/PTO/IR/PTOAttrs.td b/include/PTO/IR/PTOAttrs.td
index fc1e0ca2f..79b3ff227 100644
--- a/include/PTO/IR/PTOAttrs.td
+++ b/include/PTO/IR/PTOAttrs.td
@@ -424,6 +424,110 @@ def PTO_SaturationModeAttr : EnumAttr<PTO_Dialect, PTO_SaturationModeEnum, "satu
   let summary = "saturation mode attribute";
 }
 //===----------------------------------------------------------------------===//
+// TDIV template controls
+//===----------------------------------------------------------------------===//
+
+def PTO_DivPrecisionEnum : PTO_I32Enum<
+  "DivPrecision", "PTO TDIV precision mode", [
+    I32EnumAttrCase<"Default", 0, "default">,
+    I32EnumAttrCase<"HighPrecision", 1, "high_precision">
+  ]>;
+
+def PTO_DivPrecisionAttr : EnumAttr<PTO_Dialect, PTO_DivPrecisionEnum, "div_precision"> {
+  let summary = "TDIV precision mode attribute";
+}
+//===----------------------------------------------------------------------===//
+// TEXP template controls
+//===----------------------------------------------------------------------===//
+
+def PTO_ExpPrecisionEnum : PTO_I32Enum<
+  "ExpPrecision", "PTO TEXP precision mode", [
+    I32EnumAttrCase<"Default", 0, "default">,
+    I32EnumAttrCase<"HighPrecision", 1, "high_precision">
+  ]>;
+
+def PTO_ExpPrecisionAttr : EnumAttr<PTO_Dialect, PTO_ExpPrecisionEnum, "exp_precision"> {
+  let summary = "TEXP precision mode attribute";
+}
+//===----------------------------------------------------------------------===//
+// TLOG template controls
+//===----------------------------------------------------------------------===//
+
+def PTO_LogPrecisionEnum : PTO_I32Enum<
+  "LogPrecision", "PTO TLOG precision mode", [
+    I32EnumAttrCase<"Default", 0, "default">,
+    I32EnumAttrCase<"HighPrecision", 1, "high_precision">
+  ]>;
+
+def PTO_LogPrecisionAttr : EnumAttr<PTO_Dialect, PTO_LogPrecisionEnum, "log_precision"> {
+  let summary = "TLOG precision mode attribute";
+}
+//===----------------------------------------------------------------------===//
+// TRECIP template controls
+//===----------------------------------------------------------------------===//
+
+def PTO_RecipPrecisionEnum : PTO_I32Enum<
+  "RecipPrecision", "PTO TRECIP precision mode", [
+    I32EnumAttrCase<"Default", 0, "default">,
+    I32EnumAttrCase<"HighPrecision", 1, "high_precision">
+  ]>;
+
+def PTO_RecipPrecisionAttr : EnumAttr<PTO_Dialect, PTO_RecipPrecisionEnum, "recip_precision"> {
+  let summary = "TRECIP precision mode attribute";
+}
+//===----------------------------------------------------------------------===//
+// TREM template controls
+//===----------------------------------------------------------------------===//
+
+def PTO_RemPrecisionEnum : PTO_I32Enum<
+  "RemPrecision", "PTO TREM precision mode", [
+    I32EnumAttrCase<"Default", 0, "default">,
+    I32EnumAttrCase<"HighPrecision", 1, "high_precision">
+  ]>;
+
+def PTO_RemPrecisionAttr : EnumAttr<PTO_Dialect, PTO_RemPrecisionEnum, "rem_precision"> {
+  let summary = "TREM precision mode attribute";
+}
+//===----------------------------------------------------------------------===//
+// TRSQRT template controls
+//===----------------------------------------------------------------------===//
+
+def PTO_RsqrtPrecisionEnum : PTO_I32Enum<
+  "RsqrtPrecision", "PTO TRSQRT precision mode", [
+    I32EnumAttrCase<"Default", 0, "default">,
+    I32EnumAttrCase<"HighPrecision", 1, "high_precision">
+  ]>;
+
+def PTO_RsqrtPrecisionAttr : EnumAttr<PTO_Dialect, PTO_RsqrtPrecisionEnum, "rsqrt_precision"> {
+  let summary = "TRSQRT precision mode attribute";
+}
+//===----------------------------------------------------------------------===//
+// TSQRT template controls
+//===----------------------------------------------------------------------===//
+
+def PTO_SqrtPrecisionEnum : PTO_I32Enum<
+  "SqrtPrecision", "PTO TSQRT precision mode", [
+    I32EnumAttrCase<"Default", 0, "default">,
+    I32EnumAttrCase<"HighPrecision", 1, "high_precision">
+  ]>;
+
+def PTO_SqrtPrecisionAttr : EnumAttr<PTO_Dialect, PTO_SqrtPrecisionEnum, "sqrt_precision"> {
+  let summary = "TSQRT precision mode attribute";
+}
+//===----------------------------------------------------------------------===//
+// TFMOD template controls
+//===----------------------------------------------------------------------===//
+
+def PTO_FmodPrecisionEnum : PTO_I32Enum<
+  "FmodPrecision", "PTO TFMOD precision mode", [
+    I32EnumAttrCase<"Default", 0, "default">,
+    I32EnumAttrCase<"HighPrecision", 1, "high_precision">
+  ]>;
+
+def PTO_FmodPrecisionAttr : EnumAttr<PTO_Dialect, PTO_FmodPrecisionEnum, "fmod_precision"> {
+  let summary = "TFMOD precision mode attribute";
+}
+//===----------------------------------------------------------------------===//
 // TStore template controls
 //===----------------------------------------------------------------------===//
 
diff --git a/include/PTO/IR/PTOOps.td b/include/PTO/IR/PTOOps.td
index 344b399a6..a944db149 100644
--- a/include/PTO/IR/PTOOps.td
+++ b/include/PTO/IR/PTOOps.td
@@ -3425,7 +3425,8 @@ def TColExpandDivOp : PTO_TOp<"tcolexpanddiv", [
   let arguments = (ins
     PTODpsType:$src0,
     PTODpsType:$src1,
-    PTODpsType:$dst
+    PTODpsType:$dst,
+    DefaultValuedAttr<PTO_DivPrecisionAttr, "::mlir::pto::DivPrecision::Default">:$precisionType
   );
 
   let results = (outs);
@@ -3804,7 +3805,8 @@ def TDivOp : PTO_TOp<"tdiv", [
   let arguments = (ins
     PTODpsType:$src0,
     PTODpsType:$src1,
-    PTODpsType:$dst
+    PTODpsType:$dst,
+    DefaultValuedAttr<PTO_DivPrecisionAttr, "::mlir::pto::DivPrecision::Default">:$precisionType
   );
 
   let results = (outs);
@@ -3860,7 +3862,8 @@ def TFModOp : PTO_TOp<"tfmod", [
   let arguments = (ins
     PTODpsType:$src0,
     PTODpsType:$src1,
-    PTODpsType:$dst
+    PTODpsType:$dst,
+    DefaultValuedAttr<PTO_FmodPrecisionAttr, "::mlir::pto::FmodPrecision::Default">:$precisionType
   );
 
   let results = (outs);
@@ -3917,7 +3920,8 @@ def TExpOp : PTO_TOp<"texp", [
 
   let arguments = (ins
     PTODpsType:$src,
-    PTODpsType:$dst
+    PTODpsType:$dst,
+    DefaultValuedAttr<PTO_ExpPrecisionAttr, "::mlir::pto::ExpPrecision::Default">:$precisionType
   );
 
   let results = (outs);
@@ -4355,7 +4359,8 @@ def TLogOp : PTO_TOp<"tlog", [
 
   let arguments = (ins
     PTODpsType:$src,
-    PTODpsType:$dst
+    PTODpsType:$dst,
+    DefaultValuedAttr<PTO_LogPrecisionAttr, "::mlir::pto::LogPrecision::Default">:$precisionType
   );
 
   let results = (outs);
@@ -5118,7 +5123,8 @@ def TRecipOp: PTO_TOp<"trecip", [
 
   let arguments = (ins
     PTODpsType:$src,
-    PTODpsType:$dst
+    PTODpsType:$dst,
+    DefaultValuedAttr<PTO_RecipPrecisionAttr, "::mlir::pto::RecipPrecision::Default">:$precisionType
   );
 
   let results = (outs);
@@ -5182,7 +5188,8 @@ def TRemOp: PTO_TOp<"trem", [
     PTODpsType:$src0,
     PTODpsType:$src1,
     PTODpsType:$tmp,
-    PTODpsType:$dst
+    PTODpsType:$dst,
+    DefaultValuedAttr<PTO_RemPrecisionAttr, "::mlir::pto::RemPrecision::Default">:$precisionType
   );
 
   let results = (outs);
@@ -5311,7 +5318,8 @@ def TRowExpandDivOp: PTO_TOp<"trowexpanddiv", [
     PTODpsType:$src0,
     PTODpsType:$src1,
     Optional<PTODpsType>:$tmp,
-    PTODpsType:$dst
+    PTODpsType:$dst,
+    DefaultValuedAttr<PTO_DivPrecisionAttr, "::mlir::pto::DivPrecision::Default">:$precisionType
   );
 
   let results = (outs);
@@ -5702,7 +5710,8 @@ def TRsqrtOp: PTO_TOp<"trsqrt", [
   let arguments = (ins
     PTODpsType:$src,
     PTODpsType:$dst,
-    Optional<PTODpsType>:$tmp
+    Optional<PTODpsType>:$tmp,
+    DefaultValuedAttr<PTO_RsqrtPrecisionAttr, "::mlir::pto::RsqrtPrecision::Default">:$precisionType
   );
 
   let results = (outs);
@@ -6012,7 +6021,8 @@ def TSqrtOp: PTO_TOp<"tsqrt", [
 
   let arguments = (ins
     PTODpsType:$src,
-    PTODpsType:$dst
+    PTODpsType:$dst,
+    DefaultValuedAttr<PTO_SqrtPrecisionAttr, "::mlir::pto::SqrtPrecision::Default">:$precisionType
   );
 
   let results = (outs);
diff --git a/lib/PTO/Transforms/PTOToEmitC.cpp b/lib/PTO/Transforms/PTOToEmitC.cpp
index b3f0c6bbd..235f64d66 100644
--- a/lib/PTO/Transforms/PTOToEmitC.cpp
+++ b/lib/PTO/Transforms/PTOToEmitC.cpp
@@ -7683,15 +7683,31 @@ struct PTOColExpandDivToEmitC : public OpConversionPattern<pto::TColExpandDivOp>
   LogicalResult matchAndRewrite(pto::TColExpandDivOp op, OpAdaptor adaptor,
                                 ConversionPatternRewriter &rewriter) const override {
     auto loc = op.getLoc();
+    auto *ctx = rewriter.getContext();
 
     Value src0 = peelUnrealized(adaptor.getSrc0());
     Value src1 = peelUnrealized(adaptor.getSrc1());
     Value dst = peelUnrealized(adaptor.getDst());
 
+    ArrayAttr templateArgs;
+    if (op.getPrecisionType() != pto::DivPrecision::Default) {
+      StringRef precisionTok;
+      switch (op.getPrecisionType()) {
+      case pto::DivPrecision::Default:
+        precisionTok = "pto::DivAlgorithm::DEFAULT";
+        break;
+      case pto::DivPrecision::HighPrecision:
+        precisionTok = "pto::DivAlgorithm::HIGH_PRECISION";
+        break;
+      }
+      templateArgs = rewriter.getArrayAttr(
+          {emitc::OpaqueAttr::get(ctx, precisionTok)});
+    }
+
     rewriter.create<emitc::CallOpaqueOp>(
         loc, TypeRange{}, "TCOLEXPANDDIV",
         /*args=*/ArrayAttr{},
-        /*templateArgs=*/ArrayAttr{},
+        /*templateArgs=*/templateArgs,
         /*operands=*/ValueRange{dst, src0, src1});
 
     rewriter.eraseOp(op);
@@ -8136,10 +8152,26 @@ struct PTODivToTDIV : public OpConversionPattern<pto::TDivOp> {
     Value src0 = peelUnrealized(adaptor.getSrc0());
     Value src1 = peelUnrealized(adaptor.getSrc1());
     Value dst  = peelUnrealized(adaptor.getDst());
+    auto *ctx = rewriter.getContext();
+
+    ArrayAttr templateArgs;
+    if (op.getPrecisionType() != pto::DivPrecision::Default) {
+      StringRef precisionTok;
+      switch (op.getPrecisionType()) {
+      case pto::DivPrecision::Default:
+        precisionTok = "pto::DivAlgorithm::DEFAULT";
+        break;
+      case pto::DivPrecision::HighPrecision:
+        precisionTok = "pto::DivAlgorithm::HIGH_PRECISION";
+        break;
+      }
+      templateArgs = rewriter.getArrayAttr(
+          {emitc::OpaqueAttr::get(ctx, precisionTok)});
+    }
 
     rewriter.create<emitc::CallOpaqueOp>(
         op.getLoc(), TypeRange{}, "TDIV",
-        ArrayAttr{}, ArrayAttr{},
+        ArrayAttr{}, templateArgs,
         ValueRange{dst, src0, src1});
 
     rewriter.eraseOp(op);
@@ -8210,13 +8242,29 @@ struct PTOExpToEmitC : public OpConversionPattern<pto::TExpOp> {
   LogicalResult matchAndRewrite(pto::TExpOp op, OpAdaptor adaptor,
                                 ConversionPatternRewriter &rewriter) const override {
     auto loc = op.getLoc();
+    auto *ctx = rewriter.getContext();
 
     Value src = peelUnrealized(adaptor.getSrc());
     Value dst = peelUnrealized(adaptor.getDst());
 
+    ArrayAttr templateArgs;
+    if (op.getPrecisionType() != pto::ExpPrecision::Default) {
+      StringRef precisionTok;
+      switch (op.getPrecisionType()) {
+      case pto::ExpPrecision::Default:
+        precisionTok = "pto::ExpAlgorithm::DEFAULT";
+        break;
+      case pto::ExpPrecision::HighPrecision:
+        precisionTok = "pto::ExpAlgorithm::HIGH_PRECISION";
+        break;
+      }
+      templateArgs = rewriter.getArrayAttr(
+          {emitc::OpaqueAttr::get(ctx, precisionTok)});
+    }
+
     rewriter.create<emitc::CallOpaqueOp>(
         loc, TypeRange{}, "TEXP",
-        ArrayAttr{}, ArrayAttr{},
+        ArrayAttr{}, templateArgs,
         ValueRange{dst, src});
 
     rewriter.eraseOp(op);
@@ -8567,14 +8615,29 @@ struct PTOLogToEmitC : public OpConversionPattern<pto::TLogOp> {
   LogicalResult matchAndRewrite(pto::TLogOp op, OpAdaptor adaptor,
                                 ConversionPatternRewriter &rewriter) const override {
     auto loc = op.getLoc();
+    auto *ctx = rewriter.getContext();
 
     Value src = peelUnrealized(adaptor.getSrc());
     Value dst = peelUnrealized(adaptor.getDst());
 
     SmallVector<Value, 2> operands{dst, src};
+    ArrayAttr templateArgs;
+    if (op.getPrecisionType() != pto::LogPrecision::Default) {
+      StringRef precisionTok;
+      switch (op.getPrecisionType()) {
+      case pto::LogPrecision::Default:
+        precisionTok = "pto::LogAlgorithm::DEFAULT";
+        break;
+      case pto::LogPrecision::HighPrecision:
+        precisionTok = "pto::LogAlgorithm::HIGH_PRECISION";
+        break;
+      }
+      templateArgs = rewriter.getArrayAttr(
+          {emitc::OpaqueAttr::get(ctx, precisionTok)});
+    }
     rewriter.create<emitc::CallOpaqueOp>(
         loc, TypeRange{}, "TLOG",
-        /*args=*/ArrayAttr{}, /*templateArgs=*/ArrayAttr{},
+        /*args=*/ArrayAttr{}, /*templateArgs=*/templateArgs,
         /*operands=*/operands);
 
     rewriter.eraseOp(op);
@@ -9374,14 +9437,29 @@ struct PTORecipToEmitC : public OpConversionPattern<pto::TRecipOp> {
   LogicalResult matchAndRewrite(pto::TRecipOp op, OpAdaptor adaptor,
                                 ConversionPatternRewriter &rewriter) const override {
     auto loc = op.getLoc();
+    auto *ctx = rewriter.getContext();
 
     Value src = peelUnrealized(adaptor.getSrc());
     Value dst = peelUnrealized(adaptor.getDst());
 
     SmallVector<Value, 2> operands{dst, src};
+    ArrayAttr templateArgs;
+    if (op.getPrecisionType() != pto::RecipPrecision::Default) {
+      StringRef precisionTok;
+      switch (op.getPrecisionType()) {
+      case pto::RecipPrecision::Default:
+        precisionTok = "pto::RecipAlgorithm::DEFAULT";
+        break;
+      case pto::RecipPrecision::HighPrecision:
+        precisionTok = "pto::RecipAlgorithm::HIGH_PRECISION";
+        break;
+      }
+      templateArgs = rewriter.getArrayAttr(
+          {emitc::OpaqueAttr::get(ctx, precisionTok)});
+    }
     rewriter.create<emitc::CallOpaqueOp>(
         loc, TypeRange{}, "TRECIP",
-        /*args=*/ArrayAttr{}, /*templateArgs=*/ArrayAttr{},
+        /*args=*/ArrayAttr{}, /*templateArgs=*/templateArgs,
         /*operands=*/operands);
 
     rewriter.eraseOp(op);
@@ -9422,15 +9500,30 @@ struct PTORemToEmitC : public OpConversionPattern<pto::TRemOp> {
   LogicalResult matchAndRewrite(pto::TRemOp op, OpAdaptor adaptor,
                                 ConversionPatternRewriter &rewriter) const override {
     auto loc = op.getLoc();
+    auto *ctx = rewriter.getContext();
 
     Value src0 = peelUnrealized(adaptor.getSrc0());
     Value src1 = peelUnrealized(adaptor.getSrc1());
     Value tmp  = peelUnrealized(adaptor.getTmp());
     Value dst  = peelUnrealized(adaptor.getDst());
     SmallVector<Value, 4> operands{dst, src0, src1, tmp};
+    ArrayAttr templateArgs;
+    if (op.getPrecisionType() != pto::RemPrecision::Default) {
+      StringRef precisionTok;
+      switch (op.getPrecisionType()) {
+      case pto::RemPrecision::Default:
+        precisionTok = "pto::RemAlgorithm::DEFAULT";
+        break;
+      case pto::RemPrecision::HighPrecision:
+        precisionTok = "pto::RemAlgorithm::HIGH_PRECISION";
+        break;
+      }
+      templateArgs = rewriter.getArrayAttr(
+          {emitc::OpaqueAttr::get(ctx, precisionTok)});
+    }
     rewriter.create<emitc::CallOpaqueOp>(
         loc, TypeRange{}, "TREM",
-        /*args=*/ArrayAttr{}, /*templateArgs=*/ArrayAttr{},
+        /*args=*/ArrayAttr{}, /*templateArgs=*/templateArgs,
         /*operands=*/operands);
 
     rewriter.eraseOp(op);
@@ -9444,15 +9537,30 @@ struct PTOFModToEmitC : public OpConversionPattern<pto::TFModOp> {
   LogicalResult matchAndRewrite(pto::TFModOp op, OpAdaptor adaptor,
                                 ConversionPatternRewriter &rewriter) const override {
     auto loc = op.getLoc();
+    auto *ctx = rewriter.getContext();
 
     Value src0 = peelUnrealized(adaptor.getSrc0());
     Value src1 = peelUnrealized(adaptor.getSrc1());
     Value dst  = peelUnrealized(adaptor.getDst());
 
     SmallVector<Value, 3> operands{dst, src0, src1};
+    ArrayAttr templateArgs;
+    if (op.getPrecisionType() != pto::FmodPrecision::Default) {
+      StringRef precisionTok;
+      switch (op.getPrecisionType()) {
+      case pto::FmodPrecision::Default:
+        precisionTok = "pto::FmodAlgorithm::DEFAULT";
+        break;
+      case pto::FmodPrecision::HighPrecision:
+        precisionTok = "pto::FmodAlgorithm::HIGH_PRECISION";
+        break;
+      }
+      templateArgs = rewriter.getArrayAttr(
+          {emitc::OpaqueAttr::get(ctx, precisionTok)});
+    }
     rewriter.create<emitc::CallOpaqueOp>(
         loc, TypeRange{}, "TFMOD",
-        /*args=*/ArrayAttr{}, /*templateArgs=*/ArrayAttr{},
+        /*args=*/ArrayAttr{}, /*templateArgs=*/templateArgs,
         /*operands=*/operands);
 
     rewriter.eraseOp(op);
@@ -9764,6 +9872,7 @@ struct PTORowExpandDivToEmitC : public OpConversionPattern<pto::TRowExpandDivOp>
   LogicalResult matchAndRewrite(pto::TRowExpandDivOp op, OpAdaptor adaptor,
                                 ConversionPatternRewriter &rewriter) const override {
     auto loc = op.getLoc();
+    auto *ctx = rewriter.getContext();
 
     Value src0 = peelUnrealized(adaptor.getSrc0());
     Value src1 = peelUnrealized(adaptor.getSrc1());
@@ -9775,9 +9884,23 @@ struct PTORowExpandDivToEmitC : public OpConversionPattern<pto::TRowExpandDivOp>
       operands.assign({dst, src0, src1, tmp});
     else
       operands.assign({dst, src0, src1});
+    ArrayAttr templateArgs;
+    if (op.getPrecisionType() != pto::DivPrecision::Default) {
+      StringRef precisionTok;
+      switch (op.getPrecisionType()) {
+      case pto::DivPrecision::Default:
+        precisionTok = "pto::DivAlgorithm::DEFAULT";
+        break;
+      case pto::DivPrecision::HighPrecision:
+        precisionTok = "pto::DivAlgorithm::HIGH_PRECISION";
+        break;
+      }
+      templateArgs = rewriter.getArrayAttr(
+          {emitc::OpaqueAttr::get(ctx, precisionTok)});
+    }
     rewriter.create<emitc::CallOpaqueOp>(
         loc, TypeRange{}, "TROWEXPANDDIV",
-        /*args=*/ArrayAttr{}, /*templateArgs=*/ArrayAttr{},
+        /*args=*/ArrayAttr{}, /*templateArgs=*/templateArgs,
         /*operands=*/operands);
 
     rewriter.eraseOp(op);
@@ -10054,15 +10177,30 @@ struct PTORsqrtToEmitC : public OpConversionPattern<pto::TRsqrtOp> {
   LogicalResult matchAndRewrite(pto::TRsqrtOp op, OpAdaptor adaptor,
                                 ConversionPatternRewriter &rewriter) const override {
     auto loc = op.getLoc();
+    auto *ctx = rewriter.getContext();
 
     Value src = peelUnrealized(adaptor.getSrc());
     Value dst = peelUnrealized(adaptor.getDst());
     SmallVector<Value, 3> operands{dst, src};
     if (Value tmp = adaptor.getTmp())
       operands.push_back(peelUnrealized(tmp));
+    ArrayAttr templateArgs;
+    if (op.getPrecisionType() != pto::RsqrtPrecision::Default) {
+      StringRef precisionTok;
+      switch (op.getPrecisionType()) {
+      case pto::RsqrtPrecision::Default:
+        precisionTok = "pto::RsqrtAlgorithm::DEFAULT";
+        break;
+      case pto::RsqrtPrecision::HighPrecision:
+        precisionTok = "pto::RsqrtAlgorithm::HIGH_PRECISION";
+        break;
+      }
+      templateArgs = rewriter.getArrayAttr(
+          {emitc::OpaqueAttr::get(ctx, precisionTok)});
+    }
     rewriter.create<emitc::CallOpaqueOp>(
         loc, TypeRange{}, "TRSQRT",
-        /*args=*/ArrayAttr{}, /*templateArgs=*/ArrayAttr{},
+        /*args=*/ArrayAttr{}, /*templateArgs=*/templateArgs,
         /*operands=*/operands);
 
     rewriter.eraseOp(op);
@@ -10298,14 +10436,29 @@ struct PTOSqrtSToEmitC : public OpConversionPattern<pto::TSqrtOp> {
   LogicalResult matchAndRewrite(pto::TSqrtOp op, OpAdaptor adaptor,
                                 ConversionPatternRewriter &rewriter) const override {
     auto loc = op.getLoc();
+    auto *ctx = rewriter.getContext();
 
     Value src = peelUnrealized(adaptor.getSrc());
     Value dst = peelUnrealized(adaptor.getDst());
 
     SmallVector<Value, 4> operands{dst, src};
+    ArrayAttr templateArgs;
+    if (op.getPrecisionType() != pto::SqrtPrecision::Default) {
+      StringRef precisionTok;
+      switch (op.getPrecisionType()) {
+      case pto::SqrtPrecision::Default:
+        precisionTok = "pto::SqrtAlgorithm::DEFAULT";
+        break;
+      case pto::SqrtPrecision::HighPrecision:
+        precisionTok = "pto::SqrtAlgorithm::HIGH_PRECISION";
+        break;
+      }
+      templateArgs = rewriter.getArrayAttr(
+          {emitc::OpaqueAttr::get(ctx, precisionTok)});
+    }
     rewriter.create<emitc::CallOpaqueOp>(
         loc, TypeRange{}, "TSQRT",
-        /*args=*/ArrayAttr{}, /*templateArgs=*/ArrayAttr{},
+        /*args=*/ArrayAttr{}, /*templateArgs=*/templateArgs,
         /*operands=*/operands);
 
     rewriter.eraseOp(op);
diff --git a/lib/PTO/Transforms/PTOViewToMemref.cpp b/lib/PTO/Transforms/PTOViewToMemref.cpp
index c21669b81..63849ef2a 100644
--- a/lib/PTO/Transforms/PTOViewToMemref.cpp
+++ b/lib/PTO/Transforms/PTOViewToMemref.cpp
@@ -1864,7 +1864,8 @@ struct PTOViewToMemrefPass
         IRRewriter rewriter(ctx);
         rewriter.setInsertionPoint(op);
         rewriter.replaceOpWithNewOp<pto::TExpOp>(
-            op, TypeRange{}, op->getOperand(0), op->getOperand(1));
+            op, TypeRange{}, op->getOperand(0), op->getOperand(1),
+            op.getPrecisionTypeAttr());
       }
 
       // --- TMulOp [Src, Scalar, Dst] ---
@@ -2694,7 +2695,8 @@ struct PTOViewToMemrefPass
             TypeRange{},
             src0,
             src1,
-            dst);
+            dst,
+            op.getPrecisionTypeAttr());
       }
 
       DefaultInlineVector<mlir::pto::TDivSOp> divsops;
@@ -3059,7 +3061,8 @@ struct PTOViewToMemrefPass
             op,
             TypeRange{},
             src,
-            dst);
+            dst,
+            op.getPrecisionTypeAttr());
       }
 
       DefaultInlineVector<mlir::pto::TLReluOp> lreluops;
diff --git a/test/lit/pto/tcolexpanddiv_precision_emitc.pto b/test/lit/pto/tcolexpanddiv_precision_emitc.pto
new file mode 100644
index 000000000..ea46d994f
--- /dev/null
+++ b/test/lit/pto/tcolexpanddiv_precision_emitc.pto
@@ -0,0 +1,19 @@
+// RUN: ptoas --pto-arch=a5 %s 2>&1 | FileCheck %s --check-prefix=A5
+
+module {
+  func.func @tcolexpanddiv_precision_emitc() {
+    %src0 = pto.alloc_tile : !pto.tile_buf<loc=vec, dtype=f32, rows=4, cols=16, v_row=4, v_col=16, blayout=row_major, slayout=none_box, fractal=512, pad=0>
+    %src1 = pto.alloc_tile : !pto.tile_buf<loc=vec, dtype=f32, rows=1, cols=16, v_row=1, v_col=16, blayout=row_major, slayout=none_box, fractal=512, pad=0>
+    %dst = pto.alloc_tile : !pto.tile_buf<loc=vec, dtype=f32, rows=4, cols=16, v_row=4, v_col=16, blayout=row_major, slayout=none_box, fractal=512, pad=0>
+
+    pto.tcolexpanddiv ins(%src0, %src1 : !pto.tile_buf<loc=vec, dtype=f32, rows=4, cols=16, v_row=4, v_col=16, blayout=row_major, slayout=none_box, fractal=512, pad=0>, !pto.tile_buf<loc=vec, dtype=f32, rows=1, cols=16, v_row=1, v_col=16, blayout=row_major, slayout=none_box, fractal=512, pad=0>)
+                       outs(%dst : !pto.tile_buf<loc=vec, dtype=f32, rows=4, cols=16, v_row=4, v_col=16, blayout=row_major, slayout=none_box, fractal=512, pad=0>)
+    pto.tcolexpanddiv ins(%src0, %src1 : !pto.tile_buf<loc=vec, dtype=f32, rows=4, cols=16, v_row=4, v_col=16, blayout=row_major, slayout=none_box, fractal=512, pad=0>, !pto.tile_buf<loc=vec, dtype=f32, rows=1, cols=16, v_row=1, v_col=16, blayout=row_major, slayout=none_box, fractal=512, pad=0>)
+                       outs(%dst : !pto.tile_buf<loc=vec, dtype=f32, rows=4, cols=16, v_row=4, v_col=16, blayout=row_major, slayout=none_box, fractal=512, pad=0>)
+                       {precisionType = #pto<div_precision high_precision>}
+    return
+  }
+}
+
+// A5: TCOLEXPANDDIV([[VDST:v[0-9]+]], [[VSRC0:v[0-9]+]], [[VSRC1:v[0-9]+]]);
+// A5: TCOLEXPANDDIV<pto::DivAlgorithm::HIGH_PRECISION>([[VDST]], [[VSRC0]], [[VSRC1]]);
diff --git a/test/lit/pto/tdiv_precision_emitc.pto b/test/lit/pto/tdiv_precision_emitc.pto
new file mode 100644
index 000000000..16aad222d
--- /dev/null
+++ b/test/lit/pto/tdiv_precision_emitc.pto
@@ -0,0 +1,21 @@
+// RUN: ptoas --pto-arch=a5 %s 2>&1 | FileCheck %s --check-prefix=A5
+// RUN: ptoas --pto-arch=a5 --enable-insert-sync %s 2>&1 | FileCheck %s --check-prefix=A5
+
+module {
+  func.func @tdiv_precision_emitc() {
+    %src0 = pto.alloc_tile : !pto.tile_buf<loc=vec, dtype=f32, rows=1, cols=16, v_row=1, v_col=16, blayout=row_major, slayout=none_box, fractal=512, pad=0>
+    %src1 = pto.alloc_tile : !pto.tile_buf<loc=vec, dtype=f32, rows=1, cols=16, v_row=1, v_col=16, blayout=row_major, slayout=none_box, fractal=512, pad=0>
+    %dst = pto.alloc_tile : !pto.tile_buf<loc=vec, dtype=f32, rows=1, cols=16, v_row=1, v_col=16, blayout=row_major, slayout=none_box, fractal=512, pad=0>
+
+    pto.tdiv ins(%src0, %src1 : !pto.tile_buf<loc=vec, dtype=f32, rows=1, cols=16, v_row=1, v_col=16, blayout=row_major, slayout=none_box, fractal=512, pad=0>, !pto.tile_buf<loc=vec, dtype=f32, rows=1, cols=16, v_row=1, v_col=16, blayout=row_major, slayout=none_box, fractal=512, pad=0>)
+             outs(%dst : !pto.tile_buf<loc=vec, dtype=f32, rows=1, cols=16, v_row=1, v_col=16, blayout=row_major, slayout=none_box, fractal=512, pad=0>)
+
+    pto.tdiv ins(%src0, %src1 : !pto.tile_buf<loc=vec, dtype=f32, rows=1, cols=16, v_row=1, v_col=16, blayout=row_major, slayout=none_box, fractal=512, pad=0>, !pto.tile_buf<loc=vec, dtype=f32, rows=1, cols=16, v_row=1, v_col=16, blayout=row_major, slayout=none_box, fractal=512, pad=0>)
+             outs(%dst : !pto.tile_buf<loc=vec, dtype=f32, rows=1, cols=16, v_row=1, v_col=16, blayout=row_major, slayout=none_box, fractal=512, pad=0>)
+             {precisionType = #pto<div_precision high_precision>}
+    return
+  }
+}
+
+// A5: TDIV([[VDST:v[0-9]+]], [[VSRC0:v[0-9]+]], [[VSRC1:v[0-9]+]]);
+// A5: TDIV<pto::DivAlgorithm::HIGH_PRECISION>([[VDST]], [[VSRC0]], [[VSRC1]]);
diff --git a/test/lit/pto/texp_precision_emitc.pto b/test/lit/pto/texp_precision_emitc.pto
new file mode 100644
index 000000000..68f300115
--- /dev/null
+++ b/test/lit/pto/texp_precision_emitc.pto
@@ -0,0 +1,18 @@
+// RUN: ptoas --pto-arch=a5 %s 2>&1 | FileCheck %s --check-prefix=A5
+
+module {
+  func.func @texp_precision_emitc() {
+    %src = pto.alloc_tile : !pto.tile_buf<loc=vec, dtype=f32, rows=1, cols=16, v_row=1, v_col=16, blayout=row_major, slayout=none_box, fractal=512, pad=0>
+    %dst = pto.alloc_tile : !pto.tile_buf<loc=vec, dtype=f32, rows=1, cols=16, v_row=1, v_col=16, blayout=row_major, slayout=none_box, fractal=512, pad=0>
+
+    pto.texp ins(%src : !pto.tile_buf<loc=vec, dtype=f32, rows=1, cols=16, v_row=1, v_col=16, blayout=row_major, slayout=none_box, fractal=512, pad=0>)
+             outs(%dst : !pto.tile_buf<loc=vec, dtype=f32, rows=1, cols=16, v_row=1, v_col=16, blayout=row_major, slayout=none_box, fractal=512, pad=0>)
+    pto.texp ins(%src : !pto.tile_buf<loc=vec, dtype=f32, rows=1, cols=16, v_row=1, v_col=16, blayout=row_major, slayout=none_box, fractal=512, pad=0>)
+             outs(%dst : !pto.tile_buf<loc=vec, dtype=f32, rows=1, cols=16, v_row=1, v_col=16, blayout=row_major, slayout=none_box, fractal=512, pad=0>)
+             {precisionType = #pto<exp_precision high_precision>}
+    return
+  }
+}
+
+// A5: TEXP([[VDST:v[0-9]+]], [[VSRC:v[0-9]+]]);
+// A5: TEXP<pto::ExpAlgorithm::HIGH_PRECISION>([[VDST]], [[VSRC]]);
diff --git a/test/lit/pto/tfmod_precision_emitc.pto b/test/lit/pto/tfmod_precision_emitc.pto
new file mode 100644
index 000000000..bbd94ebc2
--- /dev/null
+++ b/test/lit/pto/tfmod_precision_emitc.pto
@@ -0,0 +1,19 @@
+// RUN: ptoas --pto-arch=a5 %s 2>&1 | FileCheck %s --check-prefix=A5
+
+module {
+  func.func @tfmod_precision_emitc() {
+    %src0 = pto.alloc_tile : !pto.tile_buf<loc=vec, dtype=f32, rows=1, cols=16, v_row=1, v_col=16, blayout=row_major, slayout=none_box, fractal=512, pad=0>
+    %src1 = pto.alloc_tile : !pto.tile_buf<loc=vec, dtype=f32, rows=1, cols=16, v_row=1, v_col=16, blayout=row_major, slayout=none_box, fractal=512, pad=0>
+    %dst = pto.alloc_tile : !pto.tile_buf<loc=vec, dtype=f32, rows=1, cols=16, v_row=1, v_col=16, blayout=row_major, slayout=none_box, fractal=512, pad=0>
+
+    pto.tfmod ins(%src0, %src1 : !pto.tile_buf<loc=vec, dtype=f32, rows=1, cols=16, v_row=1, v_col=16, blayout=row_major, slayout=none_box, fractal=512, pad=0>, !pto.tile_buf<loc=vec, dtype=f32, rows=1, cols=16, v_row=1, v_col=16, blayout=row_major, slayout=none_box, fractal=512, pad=0>)
+              outs(%dst : !pto.tile_buf<loc=vec, dtype=f32, rows=1, cols=16, v_row=1, v_col=16, blayout=row_major, slayout=none_box, fractal=512, pad=0>)
+    pto.tfmod ins(%src0, %src1 : !pto.tile_buf<loc=vec, dtype=f32, rows=1, cols=16, v_row=1, v_col=16, blayout=row_major, slayout=none_box, fractal=512, pad=0>, !pto.tile_buf<loc=vec, dtype=f32, rows=1, cols=16, v_row=1, v_col=16, blayout=row_major, slayout=none_box, fractal=512, pad=0>)
+              outs(%dst : !pto.tile_buf<loc=vec, dtype=f32, rows=1, cols=16, v_row=1, v_col=16, blayout=row_major, slayout=none_box, fractal=512, pad=0>)
+              {precisionType = #pto<fmod_precision high_precision>}
+    return
+  }
+}
+
+// A5: TFMOD([[VDST:v[0-9]+]], [[VSRC0:v[0-9]+]], [[VSRC1:v[0-9]+]]);
+// A5: TFMOD<pto::FmodAlgorithm::HIGH_PRECISION>([[VDST]], [[VSRC0]], [[VSRC1]]);
diff --git a/test/lit/pto/tlog_precision_emitc.pto b/test/lit/pto/tlog_precision_emitc.pto
new file mode 100644
index 000000000..f50d1aefb
--- /dev/null
+++ b/test/lit/pto/tlog_precision_emitc.pto
@@ -0,0 +1,18 @@
+// RUN: ptoas --pto-arch=a5 %s 2>&1 | FileCheck %s --check-prefix=A5
+
+module {
+  func.func @tlog_precision_emitc() {
+    %src = pto.alloc_tile : !pto.tile_buf<loc=vec, dtype=f32, rows=1, cols=16, v_row=1, v_col=16, blayout=row_major, slayout=none_box, fractal=512, pad=0>
+    %dst = pto.alloc_tile : !pto.tile_buf<loc=vec, dtype=f32, rows=1, cols=16, v_row=1, v_col=16, blayout=row_major, slayout=none_box, fractal=512, pad=0>
+
+    pto.tlog ins(%src : !pto.tile_buf<loc=vec, dtype=f32, rows=1, cols=16, v_row=1, v_col=16, blayout=row_major, slayout=none_box, fractal=512, pad=0>)
+             outs(%dst : !pto.tile_buf<loc=vec, dtype=f32, rows=1, cols=16, v_row=1, v_col=16, blayout=row_major, slayout=none_box, fractal=512, pad=0>)
+    pto.tlog ins(%src : !pto.tile_buf<loc=vec, dtype=f32, rows=1, cols=16, v_row=1, v_col=16, blayout=row_major, slayout=none_box, fractal=512, pad=0>)
+             outs(%dst : !pto.tile_buf<loc=vec, dtype=f32, rows=1, cols=16, v_row=1, v_col=16, blayout=row_major, slayout=none_box, fractal=512, pad=0>)
+             {precisionType = #pto<log_precision high_precision>}
+    return
+  }
+}
+
+// A5: TLOG([[VDST:v[0-9]+]], [[VSRC:v[0-9]+]]);
+// A5: TLOG<pto::LogAlgorithm::HIGH_PRECISION>([[VDST]], [[VSRC]]);
diff --git a/test/lit/pto/trecip_precision_emitc.pto b/test/lit/pto/trecip_precision_emitc.pto
new file mode 100644
index 000000000..830d03d0e
--- /dev/null
+++ b/test/lit/pto/trecip_precision_emitc.pto
@@ -0,0 +1,18 @@
+// RUN: ptoas --pto-arch=a5 %s 2>&1 | FileCheck %s --check-prefix=A5
+
+module {
+  func.func @trecip_precision_emitc() {
+    %src = pto.alloc_tile : !pto.tile_buf<loc=vec, dtype=f32, rows=1, cols=16, v_row=1, v_col=16, blayout=row_major, slayout=none_box, fractal=512, pad=0>
+    %dst = pto.alloc_tile : !pto.tile_buf<loc=vec, dtype=f32, rows=1, cols=16, v_row=1, v_col=16, blayout=row_major, slayout=none_box, fractal=512, pad=0>
+
+    pto.trecip ins(%src : !pto.tile_buf<loc=vec, dtype=f32, rows=1, cols=16, v_row=1, v_col=16, blayout=row_major, slayout=none_box, fractal=512, pad=0>)
+               outs(%dst : !pto.tile_buf<loc=vec, dtype=f32, rows=1, cols=16, v_row=1, v_col=16, blayout=row_major, slayout=none_box, fractal=512, pad=0>)
+    pto.trecip ins(%src : !pto.tile_buf<loc=vec, dtype=f32, rows=1, cols=16, v_row=1, v_col=16, blayout=row_major, slayout=none_box, fractal=512, pad=0>)
+               outs(%dst : !pto.tile_buf<loc=vec, dtype=f32, rows=1, cols=16, v_row=1, v_col=16, blayout=row_major, slayout=none_box, fractal=512, pad=0>)
+               {precisionType = #pto<recip_precision high_precision>}
+    return
+  }
+}
+
+// A5: TRECIP([[VDST:v[0-9]+]], [[VSRC:v[0-9]+]]);
+// A5: TRECIP<pto::RecipAlgorithm::HIGH_PRECISION>([[VDST]], [[VSRC]]);
diff --git a/test/lit/pto/trem_precision_emitc.pto b/test/lit/pto/trem_precision_emitc.pto
new file mode 100644
index 000000000..462867bd4
--- /dev/null
+++ b/test/lit/pto/trem_precision_emitc.pto
@@ -0,0 +1,20 @@
+// RUN: ptoas --pto-arch=a5 %s 2>&1 | FileCheck %s --check-prefix=A5
+
+module {
+  func.func @trem_precision_emitc() {
+    %src0 = pto.alloc_tile : !pto.tile_buf<loc=vec, dtype=f32, rows=1, cols=16, v_row=1, v_col=16, blayout=row_major, slayout=none_box, fractal=512, pad=0>
+    %src1 = pto.alloc_tile : !pto.tile_buf<loc=vec, dtype=f32, rows=1, cols=16, v_row=1, v_col=16, blayout=row_major, slayout=none_box, fractal=512, pad=0>
+    %tmp = pto.alloc_tile : !pto.tile_buf<loc=vec, dtype=f32, rows=1, cols=16, v_row=1, v_col=16, blayout=row_major, slayout=none_box, fractal=512, pad=0>
+    %dst = pto.alloc_tile : !pto.tile_buf<loc=vec, dtype=f32, rows=1, cols=16, v_row=1, v_col=16, blayout=row_major, slayout=none_box, fractal=512, pad=0>
+
+    pto.trem ins(%src0, %src1, %tmp : !pto.tile_buf<loc=vec, dtype=f32, rows=1, cols=16, v_row=1, v_col=16, blayout=row_major, slayout=none_box, fractal=512, pad=0>, !pto.tile_buf<loc=vec, dtype=f32, rows=1, cols=16, v_row=1, v_col=16, blayout=row_major, slayout=none_box, fractal=512, pad=0>, !pto.tile_buf<loc=vec, dtype=f32, rows=1, cols=16, v_row=1, v_col=16, blayout=row_major, slayout=none_box, fractal=512, pad=0>)
+             outs(%dst : !pto.tile_buf<loc=vec, dtype=f32, rows=1, cols=16, v_row=1, v_col=16, blayout=row_major, slayout=none_box, fractal=512, pad=0>)
+    pto.trem ins(%src0, %src1, %tmp : !pto.tile_buf<loc=vec, dtype=f32, rows=1, cols=16, v_row=1, v_col=16, blayout=row_major, slayout=none_box, fractal=512, pad=0>, !pto.tile_buf<loc=vec, dtype=f32, rows=1, cols=16, v_row=1, v_col=16, blayout=row_major, slayout=none_box, fractal=512, pad=0>, !pto.tile_buf<loc=vec, dtype=f32, rows=1, cols=16, v_row=1, v_col=16, blayout=row_major, slayout=none_box, fractal=512, pad=0>)
+             outs(%dst : !pto.tile_buf<loc=vec, dtype=f32, rows=1, cols=16, v_row=1, v_col=16, blayout=row_major, slayout=none_box, fractal=512, pad=0>)
+             {precisionType = #pto<rem_precision high_precision>}
+    return
+  }
+}
+
+// A5: TREM([[VDST:v[0-9]+]], [[VSRC0:v[0-9]+]], [[VSRC1:v[0-9]+]], [[VTMP:v[0-9]+]]);
+// A5: TREM<pto::RemAlgorithm::HIGH_PRECISION>([[VDST]], [[VSRC0]], [[VSRC1]], [[VTMP]]);
diff --git a/test/lit/pto/trowexpanddiv_precision_emitc.pto b/test/lit/pto/trowexpanddiv_precision_emitc.pto
new file mode 100644
index 000000000..6056f708f
--- /dev/null
+++ b/test/lit/pto/trowexpanddiv_precision_emitc.pto
@@ -0,0 +1,20 @@
+// RUN: ptoas --pto-arch=a5 %s 2>&1 | FileCheck %s --check-prefix=A5
+
+module {
+  func.func @trowexpanddiv_precision_emitc() {
+    %src0 = pto.alloc_tile : !pto.tile_buf<loc=vec, dtype=f32, rows=4, cols=16, v_row=4, v_col=16, blayout=row_major, slayout=none_box, fractal=512, pad=0>
+    %src1 = pto.alloc_tile : !pto.tile_buf<loc=vec, dtype=f32, rows=4, cols=8, v_row=4, v_col=8, blayout=row_major, slayout=none_box, fractal=512, pad=0>
+    %tmp = pto.alloc_tile : !pto.tile_buf<loc=vec, dtype=f32, rows=4, cols=16, v_row=4, v_col=16, blayout=row_major, slayout=none_box, fractal=512, pad=0>
+    %dst = pto.alloc_tile : !pto.tile_buf<loc=vec, dtype=f32, rows=4, cols=16, v_row=4, v_col=16, blayout=row_major, slayout=none_box, fractal=512, pad=0>
+
+    pto.trowexpanddiv ins(%src0, %src1 : !pto.tile_buf<loc=vec, dtype=f32, rows=4, cols=16, v_row=4, v_col=16, blayout=row_major, slayout=none_box, fractal=512, pad=0>, !pto.tile_buf<loc=vec, dtype=f32, rows=4, cols=8, v_row=4, v_col=8, blayout=row_major, slayout=none_box, fractal=512, pad=0>)
+                       outs(%dst : !pto.tile_buf<loc=vec, dtype=f32, rows=4, cols=16, v_row=4, v_col=16, blayout=row_major, slayout=none_box, fractal=512, pad=0>)
+    pto.trowexpanddiv ins(%src0, %src1, %tmp : !pto.tile_buf<loc=vec, dtype=f32, rows=4, cols=16, v_row=4, v_col=16, blayout=row_major, slayout=none_box, fractal=512, pad=0>, !pto.tile_buf<loc=vec, dtype=f32, rows=4, cols=8, v_row=4, v_col=8, blayout=row_major, slayout=none_box, fractal=512, pad=0>, !pto.tile_buf<loc=vec, dtype=f32, rows=4, cols=16, v_row=4, v_col=16, blayout=row_major, slayout=none_box, fractal=512, pad=0>)
+                       outs(%dst : !pto.tile_buf<loc=vec, dtype=f32, rows=4, cols=16, v_row=4, v_col=16, blayout=row_major, slayout=none_box, fractal=512, pad=0>)
+                       {precisionType = #pto<div_precision high_precision>}
+    return
+  }
+}
+
+// A5: TROWEXPANDDIV([[VDST0:v[0-9]+]], [[VSRC00:v[0-9]+]], [[VSRC01:v[0-9]+]]);
+// A5: TROWEXPANDDIV<pto::DivAlgorithm::HIGH_PRECISION>([[VDST1:v[0-9]+]], [[VSRC10:v[0-9]+]], [[VSRC11:v[0-9]+]], [[VTMP:v[0-9]+]]);
diff --git a/test/lit/pto/trsqrt_precision_emitc.pto b/test/lit/pto/trsqrt_precision_emitc.pto
new file mode 100644
index 000000000..0ff89ca39
--- /dev/null
+++ b/test/lit/pto/trsqrt_precision_emitc.pto
@@ -0,0 +1,19 @@
+// RUN: ptoas --pto-arch=a5 %s 2>&1 | FileCheck %s --check-prefix=A5
+
+module {
+  func.func @trsqrt_precision_emitc() {
+    %src = pto.alloc_tile : !pto.tile_buf<loc=vec, dtype=f32, rows=1, cols=16, v_row=1, v_col=16, blayout=row_major, slayout=none_box, fractal=512, pad=0>
+    %dst = pto.alloc_tile : !pto.tile_buf<loc=vec, dtype=f32, rows=1, cols=16, v_row=1, v_col=16, blayout=row_major, slayout=none_box, fractal=512, pad=0>
+    %tmp = pto.alloc_tile : !pto.tile_buf<loc=vec, dtype=f32, rows=1, cols=16, v_row=1, v_col=16, blayout=row_major, slayout=none_box, fractal=512, pad=0>
+
+    pto.trsqrt ins(%src : !pto.tile_buf<loc=vec, dtype=f32, rows=1, cols=16, v_row=1, v_col=16, blayout=row_major, slayout=none_box, fractal=512, pad=0>)
+               outs(%dst : !pto.tile_buf<loc=vec, dtype=f32, rows=1, cols=16, v_row=1, v_col=16, blayout=row_major, slayout=none_box, fractal=512, pad=0>)
+    pto.trsqrt ins(%src, %tmp : !pto.tile_buf<loc=vec, dtype=f32, rows=1, cols=16, v_row=1, v_col=16, blayout=row_major, slayout=none_box, fractal=512, pad=0>, !pto.tile_buf<loc=vec, dtype=f32, rows=1, cols=16, v_row=1, v_col=16, blayout=row_major, slayout=none_box, fractal=512, pad=0>)
+               outs(%dst : !pto.tile_buf<loc=vec, dtype=f32, rows=1, cols=16, v_row=1, v_col=16, blayout=row_major, slayout=none_box, fractal=512, pad=0>)
+               {precisionType = #pto<rsqrt_precision high_precision>}
+    return
+  }
+}
+
+// A5: TRSQRT([[VDST0:v[0-9]+]], [[VSRC0:v[0-9]+]]);
+// A5: TRSQRT<pto::RsqrtAlgorithm::HIGH_PRECISION>([[VDST1:v[0-9]+]], [[VSRC1:v[0-9]+]], [[VTMP:v[0-9]+]]);
diff --git a/test/lit/pto/tsqrt_precision_emitc.pto b/test/lit/pto/tsqrt_precision_emitc.pto
new file mode 100644
index 000000000..92a30a2ad
--- /dev/null
+++ b/test/lit/pto/tsqrt_precision_emitc.pto
@@ -0,0 +1,18 @@
+// RUN: ptoas --pto-arch=a5 %s 2>&1 | FileCheck %s --check-prefix=A5
+
+module {
+  func.func @tsqrt_precision_emitc() {
+    %src = pto.alloc_tile : !pto.tile_buf<loc=vec, dtype=f32, rows=1, cols=16, v_row=1, v_col=16, blayout=row_major, slayout=none_box, fractal=512, pad=0>
+    %dst = pto.alloc_tile : !pto.tile_buf<loc=vec, dtype=f32, rows=1, cols=16, v_row=1, v_col=16, blayout=row_major, slayout=none_box, fractal=512, pad=0>
+
+    pto.tsqrt ins(%src : !pto.tile_buf<loc=vec, dtype=f32, rows=1, cols=16, v_row=1, v_col=16, blayout=row_major, slayout=none_box, fractal=512, pad=0>)
+              outs(%dst : !pto.tile_buf<loc=vec, dtype=f32, rows=1, cols=16, v_row=1, v_col=16, blayout=row_major, slayout=none_box, fractal=512, pad=0>)
+    pto.tsqrt ins(%src : !pto.tile_buf<loc=vec, dtype=f32, rows=1, cols=16, v_row=1, v_col=16, blayout=row_major, slayout=none_box, fractal=512, pad=0>)
+              outs(%dst : !pto.tile_buf<loc=vec, dtype=f32, rows=1, cols=16, v_row=1, v_col=16, blayout=row_major, slayout=none_box, fractal=512, pad=0>)
+              {precisionType = #pto<sqrt_precision high_precision>}
+    return
+  }
+}
+
+// A5: TSQRT([[VDST:v[0-9]+]], [[VSRC:v[0-9]+]]);
+// A5: TSQRT<pto::SqrtAlgorithm::HIGH_PRECISION>([[VDST]], [[VSRC]]);

From e2f75cf853606bde467effa8d0641df74258c8a0 Mon Sep 17 00:00:00 2001
From: jimmychou <1049859649@qq.com>
Date: Tue, 26 May 2026 09:32:18 +0800
Subject: [PATCH 2/6] Add AccPhase support for matmul and gemv ops

---
 include/PTO/IR/PTOAttrs.td                 | 14 +++++++++
 include/PTO/IR/PTOOps.td                   | 18 ++++++++----
 test/lit/pto/tgemv_accphase_emitc.pto      | 34 ++++++++++++++++++++++
 test/lit/pto/tgemv_mx_accphase_emitc.pto   | 19 ++++++++++++
 test/lit/pto/tmatmul_accphase_emitc.pto    | 34 ++++++++++++++++++++++
 test/lit/pto/tmatmul_mx_accphase_emitc.pto | 21 +++++++++++++
 6 files changed, 134 insertions(+), 6 deletions(-)
 create mode 100644 test/lit/pto/tgemv_accphase_emitc.pto
 create mode 100644 test/lit/pto/tgemv_mx_accphase_emitc.pto
 create mode 100644 test/lit/pto/tmatmul_accphase_emitc.pto
 create mode 100644 test/lit/pto/tmatmul_mx_accphase_emitc.pto

diff --git a/include/PTO/IR/PTOAttrs.td b/include/PTO/IR/PTOAttrs.td
index 79b3ff227..c6a100455 100644
--- a/include/PTO/IR/PTOAttrs.td
+++ b/include/PTO/IR/PTOAttrs.td
@@ -528,6 +528,20 @@ def PTO_FmodPrecisionAttr : EnumAttr<PTO_Dialect, PTO_FmodPrecisionEnum, "fmod_p
   let summary = "TFMOD precision mode attribute";
 }
 //===----------------------------------------------------------------------===//
+// Matmul/Gemv accumulate phase controls
+//===----------------------------------------------------------------------===//
+
+def PTO_AccPhaseEnum : PTO_I32Enum<
+  "AccPhase", "PTO matmul/gemv accumulate phase", [
+    I32EnumAttrCase<"Unspecified", 0, "unspecified">,
+    I32EnumAttrCase<"Partial", 2, "partial">,
+    I32EnumAttrCase<"Final", 3, "final">
+  ]>;
+
+def PTO_AccPhaseAttr : EnumAttr<PTO_Dialect, PTO_AccPhaseEnum, "acc_phase"> {
+  let summary = "matmul/gemv accumulate phase attribute";
+}
+//===----------------------------------------------------------------------===//
 // TStore template controls
 //===----------------------------------------------------------------------===//
 
diff --git a/include/PTO/IR/PTOOps.td b/include/PTO/IR/PTOOps.td
index a944db149..863ed8c51 100644
--- a/include/PTO/IR/PTOOps.td
+++ b/include/PTO/IR/PTOOps.td
@@ -817,7 +817,8 @@ def TMatmulBiasOp : PTO_TOp<"tmatmul.bias", [
     PTODpsType:$a,
     PTODpsType:$b,
     PTODpsType:$bias,
-    PTODpsType:$dst
+    PTODpsType:$dst,
+    DefaultValuedAttr<PTO_AccPhaseAttr, "::mlir::pto::AccPhase::Unspecified">:$accPhase
   );
 
   let results = (outs Optional<AnyRankedTensor>:$result);
@@ -856,7 +857,8 @@ def TMatmulMxOp : PTO_TOp<"tmatmul.mx", [
     PTODpsType:$a_scale,
     PTODpsType:$b,
     PTODpsType:$b_scale,
-    PTODpsType:$dst);
+    PTODpsType:$dst,
+    DefaultValuedAttr<PTO_AccPhaseAttr, "::mlir::pto::AccPhase::Unspecified">:$accPhase);
 
   let results = (outs Optional<AnyRankedTensor>:$result);
   let hasVerifier = 1;
@@ -1031,7 +1033,8 @@ def TGemvOp : PTO_TOp<"tgemv", [
   let arguments = (ins
     PTODpsType:$lhs,
     PTODpsType:$rhs,
-    PTODpsType:$dst
+    PTODpsType:$dst,
+    DefaultValuedAttr<PTO_AccPhaseAttr, "::mlir::pto::AccPhase::Unspecified">:$accPhase
   );
 
   let results = (outs
@@ -1067,7 +1070,8 @@ def TGemvAccOp : PTO_TOp<"tgemv.acc", [
     PTODpsType:$acc_in,
     PTODpsType:$lhs,
     PTODpsType:$rhs,
-    PTODpsType:$dst
+    PTODpsType:$dst,
+    DefaultValuedAttr<PTO_AccPhaseAttr, "::mlir::pto::AccPhase::Unspecified">:$accPhase
   );
 
   let results = (outs
@@ -1102,7 +1106,8 @@ def TGemvBiasOp : PTO_TOp<"tgemv.bias", [
     PTODpsType :$a,
     PTODpsType :$b,
     PTODpsType :$bias,
-    PTODpsType :$dst
+    PTODpsType :$dst,
+    DefaultValuedAttr<PTO_AccPhaseAttr, "::mlir::pto::AccPhase::Unspecified">:$accPhase
   );
 
   let results = (outs Optional<AnyRankedTensor>:$result);
@@ -1137,7 +1142,8 @@ def TGemvMxOp : PTO_TOp<"tgemv.mx", [
     PTODpsType:$a_scale,
     PTODpsType:$b,
     PTODpsType:$b_scale,
-    PTODpsType:$dst
+    PTODpsType:$dst,
+    DefaultValuedAttr<PTO_AccPhaseAttr, "::mlir::pto::AccPhase::Unspecified">:$accPhase
   );
 
   let results = (outs Optional<AnyRankedTensor>:$result);
diff --git a/test/lit/pto/tgemv_accphase_emitc.pto b/test/lit/pto/tgemv_accphase_emitc.pto
new file mode 100644
index 000000000..a7aa3ce88
--- /dev/null
+++ b/test/lit/pto/tgemv_accphase_emitc.pto
@@ -0,0 +1,34 @@
+// RUN: ptoas --pto-arch=a5 %s 2>&1 | FileCheck %s --check-prefix=A5
+
+module attributes {pto.target_arch = "a5"} {
+  func.func @tgemv_accphase_emitc() attributes {pto.kernel_kind = #pto.kernel_kind<cube>} {
+    %lhs = pto.alloc_tile : !pto.tile_buf<loc=left, dtype=f16, rows=1, cols=128, v_row=1, v_col=128, blayout=col_major, slayout=row_major, fractal=512, pad=0>
+    %rhs = pto.alloc_tile : !pto.tile_buf<loc=right, dtype=f16, rows=128, cols=16, v_row=128, v_col=16, blayout=row_major, slayout=col_major, fractal=512, pad=0>
+    %acc_in = pto.alloc_tile : !pto.tile_buf<loc=acc, dtype=f32, rows=1, cols=16, v_row=1, v_col=16, blayout=col_major, slayout=row_major, fractal=1024, pad=0>
+    %bias = pto.alloc_tile : !pto.tile_buf<loc=bias, dtype=f32, rows=1, cols=16, v_row=1, v_col=16, blayout=row_major, slayout=none_box, fractal=512, pad=0>
+    %dst = pto.alloc_tile : !pto.tile_buf<loc=acc, dtype=f32, rows=1, cols=16, v_row=1, v_col=16, blayout=col_major, slayout=row_major, fractal=1024, pad=0>
+
+    pto.tgemv ins(%lhs, %rhs : !pto.tile_buf<loc=left, dtype=f16, rows=1, cols=128, v_row=1, v_col=128, blayout=col_major, slayout=row_major, fractal=512, pad=0>, !pto.tile_buf<loc=right, dtype=f16, rows=128, cols=16, v_row=128, v_col=16, blayout=row_major, slayout=col_major, fractal=512, pad=0>) outs(%dst : !pto.tile_buf<loc=acc, dtype=f32, rows=1, cols=16, v_row=1, v_col=16, blayout=col_major, slayout=row_major, fractal=1024, pad=0>)
+    pto.tgemv ins(%lhs, %rhs : !pto.tile_buf<loc=left, dtype=f16, rows=1, cols=128, v_row=1, v_col=128, blayout=col_major, slayout=row_major, fractal=512, pad=0>, !pto.tile_buf<loc=right, dtype=f16, rows=128, cols=16, v_row=128, v_col=16, blayout=row_major, slayout=col_major, fractal=512, pad=0>) outs(%dst : !pto.tile_buf<loc=acc, dtype=f32, rows=1, cols=16, v_row=1, v_col=16, blayout=col_major, slayout=row_major, fractal=1024, pad=0>) {accPhase = #pto<acc_phase partial>}
+    pto.tgemv ins(%lhs, %rhs : !pto.tile_buf<loc=left, dtype=f16, rows=1, cols=128, v_row=1, v_col=128, blayout=col_major, slayout=row_major, fractal=512, pad=0>, !pto.tile_buf<loc=right, dtype=f16, rows=128, cols=16, v_row=128, v_col=16, blayout=row_major, slayout=col_major, fractal=512, pad=0>) outs(%dst : !pto.tile_buf<loc=acc, dtype=f32, rows=1, cols=16, v_row=1, v_col=16, blayout=col_major, slayout=row_major, fractal=1024, pad=0>) {accPhase = #pto<acc_phase final>}
+
+    pto.tgemv.acc ins(%acc_in, %lhs, %rhs : !pto.tile_buf<loc=acc, dtype=f32, rows=1, cols=16, v_row=1, v_col=16, blayout=col_major, slayout=row_major, fractal=1024, pad=0>, !pto.tile_buf<loc=left, dtype=f16, rows=1, cols=128, v_row=1, v_col=128, blayout=col_major, slayout=row_major, fractal=512, pad=0>, !pto.tile_buf<loc=right, dtype=f16, rows=128, cols=16, v_row=128, v_col=16, blayout=row_major, slayout=col_major, fractal=512, pad=0>) outs(%dst : !pto.tile_buf<loc=acc, dtype=f32, rows=1, cols=16, v_row=1, v_col=16, blayout=col_major, slayout=row_major, fractal=1024, pad=0>)
+    pto.tgemv.acc ins(%acc_in, %lhs, %rhs : !pto.tile_buf<loc=acc, dtype=f32, rows=1, cols=16, v_row=1, v_col=16, blayout=col_major, slayout=row_major, fractal=1024, pad=0>, !pto.tile_buf<loc=left, dtype=f16, rows=1, cols=128, v_row=1, v_col=128, blayout=col_major, slayout=row_major, fractal=512, pad=0>, !pto.tile_buf<loc=right, dtype=f16, rows=128, cols=16, v_row=128, v_col=16, blayout=row_major, slayout=col_major, fractal=512, pad=0>) outs(%dst : !pto.tile_buf<loc=acc, dtype=f32, rows=1, cols=16, v_row=1, v_col=16, blayout=col_major, slayout=row_major, fractal=1024, pad=0>) {accPhase = #pto<acc_phase partial>}
+    pto.tgemv.acc ins(%acc_in, %lhs, %rhs : !pto.tile_buf<loc=acc, dtype=f32, rows=1, cols=16, v_row=1, v_col=16, blayout=col_major, slayout=row_major, fractal=1024, pad=0>, !pto.tile_buf<loc=left, dtype=f16, rows=1, cols=128, v_row=1, v_col=128, blayout=col_major, slayout=row_major, fractal=512, pad=0>, !pto.tile_buf<loc=right, dtype=f16, rows=128, cols=16, v_row=128, v_col=16, blayout=row_major, slayout=col_major, fractal=512, pad=0>) outs(%dst : !pto.tile_buf<loc=acc, dtype=f32, rows=1, cols=16, v_row=1, v_col=16, blayout=col_major, slayout=row_major, fractal=1024, pad=0>) {accPhase = #pto<acc_phase final>}
+
+    pto.tgemv.bias ins(%lhs, %rhs, %bias : !pto.tile_buf<loc=left, dtype=f16, rows=1, cols=128, v_row=1, v_col=128, blayout=col_major, slayout=row_major, fractal=512, pad=0>, !pto.tile_buf<loc=right, dtype=f16, rows=128, cols=16, v_row=128, v_col=16, blayout=row_major, slayout=col_major, fractal=512, pad=0>, !pto.tile_buf<loc=bias, dtype=f32, rows=1, cols=16, v_row=1, v_col=16, blayout=row_major, slayout=none_box, fractal=512, pad=0>) outs(%dst : !pto.tile_buf<loc=acc, dtype=f32, rows=1, cols=16, v_row=1, v_col=16, blayout=col_major, slayout=row_major, fractal=1024, pad=0>)
+    pto.tgemv.bias ins(%lhs, %rhs, %bias : !pto.tile_buf<loc=left, dtype=f16, rows=1, cols=128, v_row=1, v_col=128, blayout=col_major, slayout=row_major, fractal=512, pad=0>, !pto.tile_buf<loc=right, dtype=f16, rows=128, cols=16, v_row=128, v_col=16, blayout=row_major, slayout=col_major, fractal=512, pad=0>, !pto.tile_buf<loc=bias, dtype=f32, rows=1, cols=16, v_row=1, v_col=16, blayout=row_major, slayout=none_box, fractal=512, pad=0>) outs(%dst : !pto.tile_buf<loc=acc, dtype=f32, rows=1, cols=16, v_row=1, v_col=16, blayout=col_major, slayout=row_major, fractal=1024, pad=0>) {accPhase = #pto<acc_phase partial>}
+    pto.tgemv.bias ins(%lhs, %rhs, %bias : !pto.tile_buf<loc=left, dtype=f16, rows=1, cols=128, v_row=1, v_col=128, blayout=col_major, slayout=row_major, fractal=512, pad=0>, !pto.tile_buf<loc=right, dtype=f16, rows=128, cols=16, v_row=128, v_col=16, blayout=row_major, slayout=col_major, fractal=512, pad=0>, !pto.tile_buf<loc=bias, dtype=f32, rows=1, cols=16, v_row=1, v_col=16, blayout=row_major, slayout=none_box, fractal=512, pad=0>) outs(%dst : !pto.tile_buf<loc=acc, dtype=f32, rows=1, cols=16, v_row=1, v_col=16, blayout=col_major, slayout=row_major, fractal=1024, pad=0>) {accPhase = #pto<acc_phase final>}
+    return
+  }
+}
+
+// A5: TGEMV([[D0:v[0-9]+]], [[L0:v[0-9]+]], [[R0:v[0-9]+]]);
+// A5: TGEMV<pto::AccPhase::Partial>([[D0]], [[L0]], [[R0]]);
+// A5: TGEMV<pto::AccPhase::Final>([[D0]], [[L0]], [[R0]]);
+// A5: TGEMV_ACC([[D1:v[0-9]+]], [[CIN:v[0-9]+]], [[L1:v[0-9]+]], [[R1:v[0-9]+]]);
+// A5: TGEMV_ACC<pto::AccPhase::Partial>([[D1]], [[CIN]], [[L1]], [[R1]]);
+// A5: TGEMV_ACC<pto::AccPhase::Final>([[D1]], [[CIN]], [[L1]], [[R1]]);
+// A5: TGEMV_BIAS([[D2:v[0-9]+]], [[L2:v[0-9]+]], [[R2:v[0-9]+]], [[B2:v[0-9]+]]);
+// A5: TGEMV_BIAS<pto::AccPhase::Partial>([[D2]], [[L2]], [[R2]], [[B2]]);
+// A5: TGEMV_BIAS<pto::AccPhase::Final>([[D2]], [[L2]], [[R2]], [[B2]]);
diff --git a/test/lit/pto/tgemv_mx_accphase_emitc.pto b/test/lit/pto/tgemv_mx_accphase_emitc.pto
new file mode 100644
index 000000000..f8c5d5876
--- /dev/null
+++ b/test/lit/pto/tgemv_mx_accphase_emitc.pto
@@ -0,0 +1,19 @@
+// RUN: ptoas --pto-arch=a5 %s 2>&1 | FileCheck %s --check-prefix=A5
+
+module attributes {pto.target_arch = "a5"} {
+  func.func @tgemv_mx_accphase_emitc(%a : !pto.tile_buf<loc=left, dtype=f8E4M3FN, rows=1, cols=128, v_row=1, v_col=128, blayout=col_major, slayout=row_major, fractal=512, pad=0>,
+                                     %b : !pto.tile_buf<loc=right, dtype=f8E4M3FN, rows=128, cols=32, v_row=128, v_col=32, blayout=row_major, slayout=col_major, fractal=512, pad=0>) attributes {pto.kernel_kind = #pto.kernel_kind<cube>} {
+    %a_scale = pto.alloc_tile : !pto.tile_buf<loc=scaling, dtype=f16, rows=1, cols=128, v_row=1, v_col=128, blayout=row_major, slayout=row_major, fractal=512, pad=0>
+    %b_scale = pto.alloc_tile : !pto.tile_buf<loc=scaling, dtype=f16, rows=128, cols=32, v_row=128, v_col=32, blayout=row_major, slayout=row_major, fractal=512, pad=0>
+    %dst = pto.alloc_tile : !pto.tile_buf<loc=acc, dtype=f32, rows=1, cols=32, v_row=1, v_col=32, blayout=col_major, slayout=row_major, fractal=1024, pad=0>
+
+    pto.tgemv.mx ins(%a, %a_scale, %b, %b_scale : !pto.tile_buf<loc=left, dtype=f8E4M3FN, rows=1, cols=128, v_row=1, v_col=128, blayout=col_major, slayout=row_major, fractal=512, pad=0>, !pto.tile_buf<loc=scaling, dtype=f16, rows=1, cols=128, v_row=1, v_col=128, blayout=row_major, slayout=row_major, fractal=512, pad=0>, !pto.tile_buf<loc=right, dtype=f8E4M3FN, rows=128, cols=32, v_row=128, v_col=32, blayout=row_major, slayout=col_major, fractal=512, pad=0>, !pto.tile_buf<loc=scaling, dtype=f16, rows=128, cols=32, v_row=128, v_col=32, blayout=row_major, slayout=row_major, fractal=512, pad=0>) outs(%dst : !pto.tile_buf<loc=acc, dtype=f32, rows=1, cols=32, v_row=1, v_col=32, blayout=col_major, slayout=row_major, fractal=1024, pad=0>)
+    pto.tgemv.mx ins(%a, %a_scale, %b, %b_scale : !pto.tile_buf<loc=left, dtype=f8E4M3FN, rows=1, cols=128, v_row=1, v_col=128, blayout=col_major, slayout=row_major, fractal=512, pad=0>, !pto.tile_buf<loc=scaling, dtype=f16, rows=1, cols=128, v_row=1, v_col=128, blayout=row_major, slayout=row_major, fractal=512, pad=0>, !pto.tile_buf<loc=right, dtype=f8E4M3FN, rows=128, cols=32, v_row=128, v_col=32, blayout=row_major, slayout=col_major, fractal=512, pad=0>, !pto.tile_buf<loc=scaling, dtype=f16, rows=128, cols=32, v_row=128, v_col=32, blayout=row_major, slayout=row_major, fractal=512, pad=0>) outs(%dst : !pto.tile_buf<loc=acc, dtype=f32, rows=1, cols=32, v_row=1, v_col=32, blayout=col_major, slayout=row_major, fractal=1024, pad=0>) {accPhase = #pto<acc_phase partial>}
+    pto.tgemv.mx ins(%a, %a_scale, %b, %b_scale : !pto.tile_buf<loc=left, dtype=f8E4M3FN, rows=1, cols=128, v_row=1, v_col=128, blayout=col_major, slayout=row_major, fractal=512, pad=0>, !pto.tile_buf<loc=scaling, dtype=f16, rows=1, cols=128, v_row=1, v_col=128, blayout=row_major, slayout=row_major, fractal=512, pad=0>, !pto.tile_buf<loc=right, dtype=f8E4M3FN, rows=128, cols=32, v_row=128, v_col=32, blayout=row_major, slayout=col_major, fractal=512, pad=0>, !pto.tile_buf<loc=scaling, dtype=f16, rows=128, cols=32, v_row=128, v_col=32, blayout=row_major, slayout=row_major, fractal=512, pad=0>) outs(%dst : !pto.tile_buf<loc=acc, dtype=f32, rows=1, cols=32, v_row=1, v_col=32, blayout=col_major, slayout=row_major, fractal=1024, pad=0>) {accPhase = #pto<acc_phase final>}
+    return
+  }
+}
+
+// A5: TGEMV_MX([[D0:v[0-9]+]], [[A0:v[0-9]+]], [[AS0:v[0-9]+]], [[B0:v[0-9]+]], [[BS0:v[0-9]+]]);
+// A5: TGEMV_MX<pto::AccPhase::Partial>([[D1:v[0-9]+]], [[A1:v[0-9]+]], [[AS1:v[0-9]+]], [[B1:v[0-9]+]], [[BS1:v[0-9]+]]);
+// A5: TGEMV_MX<pto::AccPhase::Final>([[D2:v[0-9]+]], [[A2:v[0-9]+]], [[AS2:v[0-9]+]], [[B2:v[0-9]+]], [[BS2:v[0-9]+]]);
diff --git a/test/lit/pto/tmatmul_accphase_emitc.pto b/test/lit/pto/tmatmul_accphase_emitc.pto
new file mode 100644
index 000000000..ee02bb197
--- /dev/null
+++ b/test/lit/pto/tmatmul_accphase_emitc.pto
@@ -0,0 +1,34 @@
+// RUN: ptoas --pto-arch=a5 %s 2>&1 | FileCheck %s --check-prefix=A5
+
+module attributes {pto.target_arch = "a5"} {
+  func.func @tmatmul_accphase_emitc() attributes {pto.kernel_kind = #pto.kernel_kind<cube>} {
+    %lhs = pto.alloc_tile : !pto.tile_buf<loc=left, dtype=f16, rows=32, cols=32, v_row=32, v_col=32, blayout=row_major, slayout=row_major, fractal=512, pad=0>
+    %rhs = pto.alloc_tile : !pto.tile_buf<loc=right, dtype=f16, rows=32, cols=32, v_row=32, v_col=32, blayout=row_major, slayout=col_major, fractal=512, pad=0>
+    %acc_in = pto.alloc_tile : !pto.tile_buf<loc=acc, dtype=f32, rows=32, cols=32, v_row=32, v_col=32, blayout=col_major, slayout=row_major, fractal=1024, pad=0>
+    %bias = pto.alloc_tile : !pto.tile_buf<loc=bias, dtype=f32, rows=1, cols=32, v_row=1, v_col=32, blayout=row_major, slayout=none_box, fractal=512, pad=0>
+    %dst = pto.alloc_tile : !pto.tile_buf<loc=acc, dtype=f32, rows=32, cols=32, v_row=32, v_col=32, blayout=col_major, slayout=row_major, fractal=1024, pad=0>
+
+    pto.tmatmul ins(%lhs, %rhs : !pto.tile_buf<loc=left, dtype=f16, rows=32, cols=32, v_row=32, v_col=32, blayout=row_major, slayout=row_major, fractal=512, pad=0>, !pto.tile_buf<loc=right, dtype=f16, rows=32, cols=32, v_row=32, v_col=32, blayout=row_major, slayout=col_major, fractal=512, pad=0>) outs(%dst : !pto.tile_buf<loc=acc, dtype=f32, rows=32, cols=32, v_row=32, v_col=32, blayout=col_major, slayout=row_major, fractal=1024, pad=0>)
+    pto.tmatmul ins(%lhs, %rhs : !pto.tile_buf<loc=left, dtype=f16, rows=32, cols=32, v_row=32, v_col=32, blayout=row_major, slayout=row_major, fractal=512, pad=0>, !pto.tile_buf<loc=right, dtype=f16, rows=32, cols=32, v_row=32, v_col=32, blayout=row_major, slayout=col_major, fractal=512, pad=0>) outs(%dst : !pto.tile_buf<loc=acc, dtype=f32, rows=32, cols=32, v_row=32, v_col=32, blayout=col_major, slayout=row_major, fractal=1024, pad=0>) {accPhase = #pto<acc_phase partial>}
+    pto.tmatmul ins(%lhs, %rhs : !pto.tile_buf<loc=left, dtype=f16, rows=32, cols=32, v_row=32, v_col=32, blayout=row_major, slayout=row_major, fractal=512, pad=0>, !pto.tile_buf<loc=right, dtype=f16, rows=32, cols=32, v_row=32, v_col=32, blayout=row_major, slayout=col_major, fractal=512, pad=0>) outs(%dst : !pto.tile_buf<loc=acc, dtype=f32, rows=32, cols=32, v_row=32, v_col=32, blayout=col_major, slayout=row_major, fractal=1024, pad=0>) {accPhase = #pto<acc_phase final>}
+
+    pto.tmatmul.acc ins(%acc_in, %lhs, %rhs : !pto.tile_buf<loc=acc, dtype=f32, rows=32, cols=32, v_row=32, v_col=32, blayout=col_major, slayout=row_major, fractal=1024, pad=0>, !pto.tile_buf<loc=left, dtype=f16, rows=32, cols=32, v_row=32, v_col=32, blayout=row_major, slayout=row_major, fractal=512, pad=0>, !pto.tile_buf<loc=right, dtype=f16, rows=32, cols=32, v_row=32, v_col=32, blayout=row_major, slayout=col_major, fractal=512, pad=0>) outs(%dst : !pto.tile_buf<loc=acc, dtype=f32, rows=32, cols=32, v_row=32, v_col=32, blayout=col_major, slayout=row_major, fractal=1024, pad=0>)
+    pto.tmatmul.acc ins(%acc_in, %lhs, %rhs : !pto.tile_buf<loc=acc, dtype=f32, rows=32, cols=32, v_row=32, v_col=32, blayout=col_major, slayout=row_major, fractal=1024, pad=0>, !pto.tile_buf<loc=left, dtype=f16, rows=32, cols=32, v_row=32, v_col=32, blayout=row_major, slayout=row_major, fractal=512, pad=0>, !pto.tile_buf<loc=right, dtype=f16, rows=32, cols=32, v_row=32, v_col=32, blayout=row_major, slayout=col_major, fractal=512, pad=0>) outs(%dst : !pto.tile_buf<loc=acc, dtype=f32, rows=32, cols=32, v_row=32, v_col=32, blayout=col_major, slayout=row_major, fractal=1024, pad=0>) {accPhase = #pto<acc_phase partial>}
+    pto.tmatmul.acc ins(%acc_in, %lhs, %rhs : !pto.tile_buf<loc=acc, dtype=f32, rows=32, cols=32, v_row=32, v_col=32, blayout=col_major, slayout=row_major, fractal=1024, pad=0>, !pto.tile_buf<loc=left, dtype=f16, rows=32, cols=32, v_row=32, v_col=32, blayout=row_major, slayout=row_major, fractal=512, pad=0>, !pto.tile_buf<loc=right, dtype=f16, rows=32, cols=32, v_row=32, v_col=32, blayout=row_major, slayout=col_major, fractal=512, pad=0>) outs(%dst : !pto.tile_buf<loc=acc, dtype=f32, rows=32, cols=32, v_row=32, v_col=32, blayout=col_major, slayout=row_major, fractal=1024, pad=0>) {accPhase = #pto<acc_phase final>}
+
+    pto.tmatmul.bias ins(%lhs, %rhs, %bias : !pto.tile_buf<loc=left, dtype=f16, rows=32, cols=32, v_row=32, v_col=32, blayout=row_major, slayout=row_major, fractal=512, pad=0>, !pto.tile_buf<loc=right, dtype=f16, rows=32, cols=32, v_row=32, v_col=32, blayout=row_major, slayout=col_major, fractal=512, pad=0>, !pto.tile_buf<loc=bias, dtype=f32, rows=1, cols=32, v_row=1, v_col=32, blayout=row_major, slayout=none_box, fractal=512, pad=0>) outs(%dst : !pto.tile_buf<loc=acc, dtype=f32, rows=32, cols=32, v_row=32, v_col=32, blayout=col_major, slayout=row_major, fractal=1024, pad=0>)
+    pto.tmatmul.bias ins(%lhs, %rhs, %bias : !pto.tile_buf<loc=left, dtype=f16, rows=32, cols=32, v_row=32, v_col=32, blayout=row_major, slayout=row_major, fractal=512, pad=0>, !pto.tile_buf<loc=right, dtype=f16, rows=32, cols=32, v_row=32, v_col=32, blayout=row_major, slayout=col_major, fractal=512, pad=0>, !pto.tile_buf<loc=bias, dtype=f32, rows=1, cols=32, v_row=1, v_col=32, blayout=row_major, slayout=none_box, fractal=512, pad=0>) outs(%dst : !pto.tile_buf<loc=acc, dtype=f32, rows=32, cols=32, v_row=32, v_col=32, blayout=col_major, slayout=row_major, fractal=1024, pad=0>) {accPhase = #pto<acc_phase partial>}
+    pto.tmatmul.bias ins(%lhs, %rhs, %bias : !pto.tile_buf<loc=left, dtype=f16, rows=32, cols=32, v_row=32, v_col=32, blayout=row_major, slayout=row_major, fractal=512, pad=0>, !pto.tile_buf<loc=right, dtype=f16, rows=32, cols=32, v_row=32, v_col=32, blayout=row_major, slayout=col_major, fractal=512, pad=0>, !pto.tile_buf<loc=bias, dtype=f32, rows=1, cols=32, v_row=1, v_col=32, blayout=row_major, slayout=none_box, fractal=512, pad=0>) outs(%dst : !pto.tile_buf<loc=acc, dtype=f32, rows=32, cols=32, v_row=32, v_col=32, blayout=col_major, slayout=row_major, fractal=1024, pad=0>) {accPhase = #pto<acc_phase final>}
+    return
+  }
+}
+
+// A5: TMATMUL([[D0:v[0-9]+]], [[L0:v[0-9]+]], [[R0:v[0-9]+]]);
+// A5: TMATMUL<pto::AccPhase::Partial>([[D0]], [[L0]], [[R0]]);
+// A5: TMATMUL<pto::AccPhase::Final>([[D0]], [[L0]], [[R0]]);
+// A5: TMATMUL_ACC([[D1:v[0-9]+]], [[CIN:v[0-9]+]], [[L1:v[0-9]+]], [[R1:v[0-9]+]]);
+// A5: TMATMUL_ACC<pto::AccPhase::Partial>([[D1]], [[CIN]], [[L1]], [[R1]]);
+// A5: TMATMUL_ACC<pto::AccPhase::Final>([[D1]], [[CIN]], [[L1]], [[R1]]);
+// A5: TMATMUL_BIAS([[D2:v[0-9]+]], [[L2:v[0-9]+]], [[R2:v[0-9]+]], [[B2:v[0-9]+]]);
+// A5: TMATMUL_BIAS<pto::AccPhase::Partial>([[D2]], [[L2]], [[R2]], [[B2]]);
+// A5: TMATMUL_BIAS<pto::AccPhase::Final>([[D2]], [[L2]], [[R2]], [[B2]]);
diff --git a/test/lit/pto/tmatmul_mx_accphase_emitc.pto b/test/lit/pto/tmatmul_mx_accphase_emitc.pto
new file mode 100644
index 000000000..bde097042
--- /dev/null
+++ b/test/lit/pto/tmatmul_mx_accphase_emitc.pto
@@ -0,0 +1,21 @@
+// RUN: ptoas --pto-arch=a5 %s 2>&1 | FileCheck %s --check-prefix=A5
+
+module attributes {pto.target_arch = "a5"} {
+  func.func @tmatmul_mx_accphase_emitc(%a : !pto.tile_buf<loc=left, dtype=f8E4M3FN, rows=16, cols=32, v_row=16, v_col=32, blayout=col_major, slayout=row_major, fractal=512, pad=0>,
+                                       %b : !pto.tile_buf<loc=right, dtype=f8E4M3FN, rows=32, cols=32, v_row=32, v_col=32, blayout=row_major, slayout=col_major, fractal=512, pad=0>) attributes {pto.kernel_kind = #pto.kernel_kind<cube>} {
+    %a_scale = pto.alloc_tile : !pto.tile_buf<loc=scaling, dtype=f16, rows=16, cols=32, v_row=16, v_col=32, blayout=row_major, slayout=none_box, fractal=32, pad=0>
+    %b_scale = pto.alloc_tile : !pto.tile_buf<loc=scaling, dtype=f16, rows=32, cols=32, v_row=32, v_col=32, blayout=row_major, slayout=none_box, fractal=32, pad=0>
+    %dst0 = pto.alloc_tile : !pto.tile_buf<loc=acc, dtype=f32, rows=16, cols=32, v_row=16, v_col=32, blayout=col_major, slayout=row_major, fractal=512, pad=0>
+    %dst1 = pto.alloc_tile : !pto.tile_buf<loc=acc, dtype=f32, rows=16, cols=32, v_row=16, v_col=32, blayout=col_major, slayout=row_major, fractal=512, pad=0>
+    %dst2 = pto.alloc_tile : !pto.tile_buf<loc=acc, dtype=f32, rows=16, cols=32, v_row=16, v_col=32, blayout=col_major, slayout=row_major, fractal=512, pad=0>
+
+    pto.tmatmul.mx ins(%a, %a_scale, %b, %b_scale : !pto.tile_buf<loc=left, dtype=f8E4M3FN, rows=16, cols=32, v_row=16, v_col=32, blayout=col_major, slayout=row_major, fractal=512, pad=0>, !pto.tile_buf<loc=scaling, dtype=f16, rows=16, cols=32, v_row=16, v_col=32, blayout=row_major, slayout=none_box, fractal=32, pad=0>, !pto.tile_buf<loc=right, dtype=f8E4M3FN, rows=32, cols=32, v_row=32, v_col=32, blayout=row_major, slayout=col_major, fractal=512, pad=0>, !pto.tile_buf<loc=scaling, dtype=f16, rows=32, cols=32, v_row=32, v_col=32, blayout=row_major, slayout=none_box, fractal=32, pad=0>) outs(%dst0 : !pto.tile_buf<loc=acc, dtype=f32, rows=16, cols=32, v_row=16, v_col=32, blayout=col_major, slayout=row_major, fractal=512, pad=0>)
+    pto.tmatmul.mx ins(%a, %a_scale, %b, %b_scale : !pto.tile_buf<loc=left, dtype=f8E4M3FN, rows=16, cols=32, v_row=16, v_col=32, blayout=col_major, slayout=row_major, fractal=512, pad=0>, !pto.tile_buf<loc=scaling, dtype=f16, rows=16, cols=32, v_row=16, v_col=32, blayout=row_major, slayout=none_box, fractal=32, pad=0>, !pto.tile_buf<loc=right, dtype=f8E4M3FN, rows=32, cols=32, v_row=32, v_col=32, blayout=row_major, slayout=col_major, fractal=512, pad=0>, !pto.tile_buf<loc=scaling, dtype=f16, rows=32, cols=32, v_row=32, v_col=32, blayout=row_major, slayout=none_box, fractal=32, pad=0>) outs(%dst1 : !pto.tile_buf<loc=acc, dtype=f32, rows=16, cols=32, v_row=16, v_col=32, blayout=col_major, slayout=row_major, fractal=512, pad=0>) {accPhase = #pto<acc_phase partial>}
+    pto.tmatmul.mx ins(%a, %a_scale, %b, %b_scale : !pto.tile_buf<loc=left, dtype=f8E4M3FN, rows=16, cols=32, v_row=16, v_col=32, blayout=col_major, slayout=row_major, fractal=512, pad=0>, !pto.tile_buf<loc=scaling, dtype=f16, rows=16, cols=32, v_row=16, v_col=32, blayout=row_major, slayout=none_box, fractal=32, pad=0>, !pto.tile_buf<loc=right, dtype=f8E4M3FN, rows=32, cols=32, v_row=32, v_col=32, blayout=row_major, slayout=col_major, fractal=512, pad=0>, !pto.tile_buf<loc=scaling, dtype=f16, rows=32, cols=32, v_row=32, v_col=32, blayout=row_major, slayout=none_box, fractal=32, pad=0>) outs(%dst2 : !pto.tile_buf<loc=acc, dtype=f32, rows=16, cols=32, v_row=16, v_col=32, blayout=col_major, slayout=row_major, fractal=512, pad=0>) {accPhase = #pto<acc_phase final>}
+    return
+  }
+}
+
+// A5: TMATMUL_MX([[D0:v[0-9]+]], [[A0:v[0-9]+]], [[AS0:v[0-9]+]], [[B0:v[0-9]+]], [[BS0:v[0-9]+]]);
+// A5: TMATMUL_MX<pto::AccPhase::Partial>([[D1:v[0-9]+]], [[A1:v[0-9]+]], [[AS1:v[0-9]+]], [[B1:v[0-9]+]], [[BS1:v[0-9]+]]);
+// A5: TMATMUL_MX<pto::AccPhase::Final>([[D2:v[0-9]+]], [[A2:v[0-9]+]], [[AS2:v[0-9]+]], [[B2:v[0-9]+]], [[BS2:v[0-9]+]]);

From fca3aa98c4d403d9e12a5b873a4d43ca159fe371 Mon Sep 17 00:00:00 2001
From: jimmychou <1049859649@qq.com>
Date: Tue, 26 May 2026 12:20:46 +0800
Subject: [PATCH 3/6] fix(pto): keep void EmitC calls void in matmul/gemv
 lowering

---
 lib/PTO/Transforms/PTOToEmitC.cpp | 6 +++---
 1 file changed, 3 insertions(+), 3 deletions(-)

diff --git a/lib/PTO/Transforms/PTOToEmitC.cpp b/lib/PTO/Transforms/PTOToEmitC.cpp
index 235f64d66..bc6b39ad4 100644
--- a/lib/PTO/Transforms/PTOToEmitC.cpp
+++ b/lib/PTO/Transforms/PTOToEmitC.cpp
@@ -9698,10 +9698,10 @@ static void replaceOrEraseWithOpaqueCall(Operation *op,
                                         StringRef callee,
                                         ArrayRef<Value> args,
                                         ConversionPatternRewriter &rewriter) {
-  TypeRange resultTypes = op->getResultTypes();
   auto call = rewriter.create<emitc::CallOpaqueOp>(
-      op->getLoc(), resultTypes, callee, ArrayAttr{}, ArrayAttr{}, ValueRange(args));
-  if (resultTypes.empty())
+      op->getLoc(), TypeRange{}, callee, ArrayAttr{}, ArrayAttr{},
+      ValueRange(args));
+  if (op->getNumResults() == 0)
     rewriter.eraseOp(op);
   else
     rewriter.replaceOp(op, call.getResults());

From 880cfcda51eb36b9dce84bd200bd860cf91940a8 Mon Sep 17 00:00:00 2001
From: jimmychou <1049859649@qq.com>
Date: Tue, 26 May 2026 14:53:49 +0800
Subject: [PATCH 4/6] fix(pto): reuse upstream AccPhase attr definition

---
 include/PTO/IR/PTOAttrs.td | 14 --------------
 1 file changed, 14 deletions(-)

diff --git a/include/PTO/IR/PTOAttrs.td b/include/PTO/IR/PTOAttrs.td
index c6a100455..79b3ff227 100644
--- a/include/PTO/IR/PTOAttrs.td
+++ b/include/PTO/IR/PTOAttrs.td
@@ -528,20 +528,6 @@ def PTO_FmodPrecisionAttr : EnumAttr<PTO_Dialect, PTO_FmodPrecisionEnum, "fmod_p
   let summary = "TFMOD precision mode attribute";
 }
 //===----------------------------------------------------------------------===//
-// Matmul/Gemv accumulate phase controls
-//===----------------------------------------------------------------------===//
-
-def PTO_AccPhaseEnum : PTO_I32Enum<
-  "AccPhase", "PTO matmul/gemv accumulate phase", [
-    I32EnumAttrCase<"Unspecified", 0, "unspecified">,
-    I32EnumAttrCase<"Partial", 2, "partial">,
-    I32EnumAttrCase<"Final", 3, "final">
-  ]>;
-
-def PTO_AccPhaseAttr : EnumAttr<PTO_Dialect, PTO_AccPhaseEnum, "acc_phase"> {
-  let summary = "matmul/gemv accumulate phase attribute";
-}
-//===----------------------------------------------------------------------===//
 // TStore template controls
 //===----------------------------------------------------------------------===//
 

From 849a30193820538055f299acd7bd7dfd9282fb50 Mon Sep 17 00:00:00 2001
From: jimmychou <1049859649@qq.com>
Date: Tue, 26 May 2026 19:37:01 +0800
Subject: [PATCH 5/6] fix(pto): complete accphase emitc lowering

---
 lib/PTO/IR/PTO.cpp                       | 87 ++++++++++++++++++++----
 lib/PTO/Transforms/PTOToEmitC.cpp        | 76 +++++++++++----------
 lib/PTO/Transforms/PTOViewToMemref.cpp   | 12 ++--
 test/lit/pto/tgemv_mx_accphase_emitc.pto | 15 ++--
 test/lit/pto/tmatmul_acc_phase_emitc.pto |  8 +--
 5 files changed, 134 insertions(+), 64 deletions(-)

diff --git a/lib/PTO/IR/PTO.cpp b/lib/PTO/IR/PTO.cpp
index 5427ac36e..d3d557a84 100644
--- a/lib/PTO/IR/PTO.cpp
+++ b/lib/PTO/IR/PTO.cpp
@@ -4121,6 +4121,63 @@ static LogicalResult verifyGemvTileOperands(Operation *op, Type lhsTy, Type rhsT
   return failure();
 }
 
+static LogicalResult verifyA5MxGemvTileOperands(Operation *op, Type lhsTy,
+                                                Type rhsTy, Type dstTy) {
+  if (failed(verifyTileBufCommon(op, lhsTy, "lhs", /*allowLowPrecision=*/true)) ||
+      failed(verifyTileBufCommon(op, rhsTy, "rhs", /*allowLowPrecision=*/true)) ||
+      failed(verifyAccTileCommon(op, dstTy, "dst")))
+    return failure();
+
+  auto lhsSpace = getPTOMemorySpaceEnum(lhsTy);
+  auto rhsSpace = getPTOMemorySpaceEnum(rhsTy);
+  auto dstSpace = getPTOMemorySpaceEnum(dstTy);
+  if (!lhsSpace || !rhsSpace || !dstSpace)
+    return op->emitOpError("expects lhs, rhs, and dst to have explicit address spaces");
+  if (*lhsSpace != pto::AddressSpace::LEFT || *rhsSpace != pto::AddressSpace::RIGHT ||
+      *dstSpace != pto::AddressSpace::ACC)
+    return op->emitOpError(
+        "expects lhs, rhs, and dst to use the left, right, and acc address spaces");
+
+  auto lhsValid = getValidShapeVec(lhsTy);
+  auto rhsValid = getValidShapeVec(rhsTy);
+  auto dstValid = getValidShapeVec(dstTy);
+  if (lhsValid[0] != ShapedType::kDynamic && lhsValid[0] != 1)
+    return op->emitOpError("expects lhs valid_shape[0] to be 1 for tgemv");
+  if (dstValid[0] != ShapedType::kDynamic && dstValid[0] != 1)
+    return op->emitOpError("expects dst valid_shape[0] to be 1 for tgemv");
+  if (lhsValid[1] != ShapedType::kDynamic && rhsValid[0] != ShapedType::kDynamic &&
+      lhsValid[1] != rhsValid[0])
+    return op->emitOpError()
+           << "expects lhs valid_shape[1] to equal rhs valid_shape[0], but got "
+           << lhsValid[1] << " vs " << rhsValid[0];
+  if (rhsValid[1] != ShapedType::kDynamic && dstValid[1] != ShapedType::kDynamic &&
+      rhsValid[1] != dstValid[1])
+    return op->emitOpError()
+           << "expects rhs valid_shape[1] to equal dst valid_shape[1], but got "
+           << rhsValid[1] << " vs " << dstValid[1];
+
+  auto lhsTb = dyn_cast<pto::TileBufType>(lhsTy);
+  auto rhsTb = dyn_cast<pto::TileBufType>(rhsTy);
+  auto dstTb = dyn_cast<pto::TileBufType>(dstTy);
+  if (!lhsTb || !rhsTb || !dstTb)
+    return success();
+
+  if (lhsTb.getBLayoutValueI32() != static_cast<int32_t>(pto::BLayout::ColMajor))
+    return op->emitOpError("expects lhs to use the col_major blayout on A5");
+  if (rhsTb.getBLayoutValueI32() != static_cast<int32_t>(pto::BLayout::RowMajor))
+    return op->emitOpError("expects rhs to use the row_major blayout on A5");
+  if (dstTb.getBLayoutValueI32() != static_cast<int32_t>(pto::BLayout::ColMajor))
+    return op->emitOpError("expects dst to use the col_major blayout on A5");
+
+  if (lhsTb.getSLayoutValueI32() != static_cast<int32_t>(pto::SLayout::RowMajor))
+    return op->emitOpError("expects lhs to use the row_major slayout on A5");
+  if (rhsTb.getSLayoutValueI32() != static_cast<int32_t>(pto::SLayout::ColMajor))
+    return op->emitOpError("expects rhs to use the col_major slayout on A5");
+  if (dstTb.getSLayoutValueI32() != static_cast<int32_t>(pto::SLayout::RowMajor))
+    return op->emitOpError("expects dst to use the row_major slayout on A5");
+  return success();
+}
+
 static LogicalResult verifyMatBiasTileA2A3(Operation *op, Type biasTy, Type dstTy,
                                            bool requireFloatBias) {
   if (failed(verifyTileBufCommon(op, biasTy, "bias")))
@@ -5625,7 +5682,9 @@ static bool isA5Fp8LikeType(Type ty) {
 }
 
 static bool isA5MxInputType(Type ty) {
-  return isA5Fp8LikeType(ty);
+  if (auto ft = dyn_cast<FloatType>(ty))
+    return ft.isFloat8E4M3FN();
+  return false;
 }
 
 static LogicalResult verifyA5MxTypeTriple(Operation *op, Type lhsTy, Type rhsTy,
@@ -5635,10 +5694,12 @@ static LogicalResult verifyA5MxTypeTriple(Operation *op, Type lhsTy, Type rhsTy,
   Type rhsElem = getElemTy(rhsTy);
   Type dstElem = getElemTy(dstTy);
 
-  if (!isA5MxInputType(lhsElem) || !isA5MxInputType(rhsElem))
-    return op->emitOpError()
-           << "expects A5 mx operands " << lhsName << " and " << rhsName
-           << " to use fp8 element types";
+  if (!isA5MxInputType(lhsElem))
+    return op->emitOpError() << lhsName << ": dtype " << lhsElem
+                             << " is not supported by this op yet";
+  if (!isA5MxInputType(rhsElem))
+    return op->emitOpError() << rhsName << ": dtype " << rhsElem
+                             << " is not supported by this op yet";
 
   if (!dstElem.isF32())
     return op->emitOpError()
@@ -6609,11 +6670,11 @@ LogicalResult TGemvMxOp::verify() {
                                              getA().getType(), "a_scale", "a")) ||
         failed(verifyScaleTileMatchesOperand(*this, getBScale().getType(),
                                              getB().getType(), "b_scale", "b")) ||
-        failed(verifyGemvTileOperands(*this, getA().getType(), getB().getType(),
-                                      getDst().getType())))
+        failed(verifyA5MxGemvTileOperands(*this, getA().getType(), getB().getType(),
+                                          getDst().getType())))
       return failure();
     if (failed(verifyA5MxTypeTriple(*this, getA().getType(), getB().getType(),
-                                    getDst().getType(), "a", "b", "dst")))
+                                    getDst().getType(), "lhs", "rhs", "dst")))
       return failure();
     return verifyMatmulLike(*this, getA().getType(), getB().getType(),
                             getDst().getType());
@@ -6635,7 +6696,7 @@ LogicalResult TGemvMxAccOp::verify() {
                                       getDst().getType())))
       return failure();
     if (failed(verifyA5MxTypeTriple(*this, getA().getType(), getB().getType(),
-                                    getDst().getType(), "a", "b", "dst")))
+                                    getDst().getType(), "lhs", "rhs", "dst")))
       return failure();
     if (failed(verifyTileBufSameElemType(*this, getCIn().getType(),
                                              getDst().getType(), "c_in", "dst")) ||
@@ -6663,7 +6724,7 @@ LogicalResult TGemvMxBiasOp::verify() {
                                  /*requireFloatBias=*/true)))
       return failure();
     if (failed(verifyA5MxTypeTriple(*this, getA().getType(), getB().getType(),
-                                    getDst().getType(), "a", "b", "dst")))
+                                    getDst().getType(), "lhs", "rhs", "dst")))
       return failure();
     auto biasShape = getShapeVec(getBias().getType());
     auto dstShape = getShapeVec(getDst().getType());
@@ -6710,7 +6771,7 @@ LogicalResult TMatmulMxOp::verify() {
     if (failed(verifyA2A3()))
       return failure();
     return verifyA5MxTypeTriple(*this, getA().getType(), getB().getType(),
-                                getDst().getType(), "a", "b", "dst");
+                                getDst().getType(), "lhs", "rhs", "dst");
   };
   return dispatchVerifierByArch(getOperation(), verifyA2A3, verifyA5);
 }
@@ -6727,7 +6788,7 @@ LogicalResult TMatmulMxAccOp::verify() {
     if (failed(verifyA2A3()))
       return failure();
     if (failed(verifyA5MxTypeTriple(*this, getA().getType(), getB().getType(),
-                                    getDst().getType(), "a", "b", "dst")))
+                                    getDst().getType(), "lhs", "rhs", "dst")))
       return failure();
     if (failed(verifyTileBufSameElemType(*this, getCIn().getType(),
                                              getDst().getType(), "c_in", "dst")) ||
@@ -6754,7 +6815,7 @@ LogicalResult TMatmulMxBiasOp::verify() {
     if (failed(verifyA2A3()))
       return failure();
     return verifyA5MxTypeTriple(*this, getA().getType(), getB().getType(),
-                                getDst().getType(), "a", "b", "dst");
+                                getDst().getType(), "lhs", "rhs", "dst");
   };
   return dispatchVerifierByArch(getOperation(), verifyA2A3, verifyA5);
 }
diff --git a/lib/PTO/Transforms/PTOToEmitC.cpp b/lib/PTO/Transforms/PTOToEmitC.cpp
index bc6b39ad4..e9e65050a 100644
--- a/lib/PTO/Transforms/PTOToEmitC.cpp
+++ b/lib/PTO/Transforms/PTOToEmitC.cpp
@@ -4520,8 +4520,8 @@ struct PTOTStoreToTSTORE : public OpConversionPattern<pto::TStoreOp> {
 // Render `pto.tmatmul` as one of three forms depending on the optional
 // `acc_phase` attribute:
 //   * absent / Unspecified  -> `TMATMUL(dst, lhs, rhs)`
-//   * Partial               -> `TMATMUL<AccPhase::Partial>(dst, lhs, rhs)`
-//   * Final                 -> `TMATMUL<AccPhase::Final>(dst, lhs, rhs)`
+//   * Partial               -> `TMATMUL<pto::AccPhase::Partial>(dst, lhs, rhs)`
+//   * Final                 -> `TMATMUL<pto::AccPhase::Final>(dst, lhs, rhs)`
 // The Unspecified default keeps backward compatibility with all upstream IR
 // that does not yet emit an explicit phase attribute.
 static ArrayAttr buildAccPhaseTemplateArgs(ConversionPatternRewriter &rewriter,
@@ -4531,10 +4531,10 @@ static ArrayAttr buildAccPhaseTemplateArgs(ConversionPatternRewriter &rewriter,
   case pto::AccPhase::Unspecified:
     return ArrayAttr{};
   case pto::AccPhase::Partial:
-    tmpl = "AccPhase::Partial";
+    tmpl = "pto::AccPhase::Partial";
     break;
   case pto::AccPhase::Final:
-    tmpl = "AccPhase::Final";
+    tmpl = "pto::AccPhase::Final";
     break;
   }
   if (tmpl.empty())
@@ -4585,10 +4585,12 @@ struct PTOTGemvToTGEMV : public OpConversionPattern<pto::TGemvOp> {
     Value rhs = peelUnrealized(adaptor.getRhs()); // B (Vector)
     Value dst = peelUnrealized(adaptor.getDst()); // C (Result)
 
-    // 2. 直接生成函数调用 TGEMV(dst, lhs, rhs)
+    ArrayAttr templateArgs =
+        buildAccPhaseTemplateArgs(rewriter, op.getAccPhase());
+
     rewriter.create<emitc::CallOpaqueOp>(
         op.getLoc(), TypeRange{}, "TGEMV",
-        ArrayAttr{}, ArrayAttr{},
+        /*args=*/ArrayAttr{}, /*template_args=*/templateArgs,
         ValueRange{dst, lhs, rhs});
 
     // 3. 处理 Op 替换/删除
@@ -4618,10 +4620,12 @@ struct PTOTGemvAccToTGEMVACC : public OpConversionPattern<pto::TGemvAccOp> {
     Value rhs   = peelUnrealized(adaptor.getRhs());   // B (Vector)
     Value dst   = peelUnrealized(adaptor.getDst());   // AccNew
 
-    // 2. 直接生成函数调用 TGEMV_ACC(dst, accIn, lhs, rhs)
+    ArrayAttr templateArgs =
+        buildAccPhaseTemplateArgs(rewriter, op.getAccPhase());
+
     rewriter.create<emitc::CallOpaqueOp>(
         op.getLoc(), TypeRange{}, "TGEMV_ACC",
-        ArrayAttr{}, ArrayAttr{},
+        /*args=*/ArrayAttr{}, /*template_args=*/templateArgs,
         ValueRange{dst, accIn, lhs, rhs});
 
     // 3. 处理 Op 替换/删除
@@ -9694,25 +9698,13 @@ struct PTORowExpandExpdifToEmitC
 // PTOConvert.cpp  (add lowering + patterns.add for TROWEXPANDDIV DPS/memref op)
 //===----------------------------------------------------------------------===//
 // Helper: replace or erase based on whether op has results.
-static void replaceOrEraseWithOpaqueCall(Operation *op,
-                                        StringRef callee,
-                                        ArrayRef<Value> args,
-                                        ConversionPatternRewriter &rewriter) {
-  auto call = rewriter.create<emitc::CallOpaqueOp>(
-      op->getLoc(), TypeRange{}, callee, ArrayAttr{}, ArrayAttr{},
-      ValueRange(args));
-  if (op->getNumResults() == 0)
-    rewriter.eraseOp(op);
-  else
-    rewriter.replaceOp(op, call.getResults());
-}
-
 static void replaceOrEraseWithOpaqueCallAndReturnDst(Operation *op, Value dst,
                                                      StringRef callee,
                                                      ArrayRef<Value> args,
+                                                     ArrayAttr templateArgs,
                                                      ConversionPatternRewriter &rewriter) {
   rewriter.create<emitc::CallOpaqueOp>(
-      op->getLoc(), TypeRange{}, callee, ArrayAttr{}, ArrayAttr{}, ValueRange(args));
+      op->getLoc(), TypeRange{}, callee, ArrayAttr{}, templateArgs, ValueRange(args));
   if (op->getNumResults() == 1)
     rewriter.replaceOp(op, dst);
   else
@@ -9731,8 +9723,10 @@ struct PTOTGemvBiasToTGEMV_BIAS
     Value bias = peelUnrealized(adaptor.getBias());
     Value dst  = peelUnrealized(adaptor.getDst());
 
-    replaceOrEraseWithOpaqueCall(op.getOperation(), "TGEMV_BIAS",
-                                {dst, a, b, bias}, rewriter);
+    ArrayAttr templateArgs =
+        buildAccPhaseTemplateArgs(rewriter, op.getAccPhase());
+    replaceOrEraseWithOpaqueCallAndReturnDst(op.getOperation(), dst, "TGEMV_BIAS",
+                                             {dst, a, b, bias}, templateArgs, rewriter);
     return success();
   }
 };
@@ -9749,8 +9743,11 @@ struct PTOTGemvMXToTGEMV_MX
     Value bScale  = peelUnrealized(adaptor.getBScale());
     Value dst     = peelUnrealized(adaptor.getDst());
 
+    ArrayAttr templateArgs =
+        buildAccPhaseTemplateArgs(rewriter, op.getAccPhase());
     replaceOrEraseWithOpaqueCallAndReturnDst(op.getOperation(), dst, "TGEMV_MX",
-                                             {dst, a, aScale, b, bScale}, rewriter);
+                                             {dst, a, aScale, b, bScale}, templateArgs,
+                                             rewriter);
     return success();
   }
 };
@@ -9769,7 +9766,8 @@ struct PTOTGemvMXAccToTGEMV_MX
     Value dst     = peelUnrealized(adaptor.getDst());
 
     replaceOrEraseWithOpaqueCallAndReturnDst(op.getOperation(), dst, "TGEMV_MX",
-                                             {dst, cIn, a, aScale, b, bScale}, rewriter);
+                                             {dst, cIn, a, aScale, b, bScale}, ArrayAttr{},
+                                             rewriter);
     return success();
   }
 };
@@ -9788,7 +9786,8 @@ struct PTOTGemvMXBiasToTGEMV_MX
     Value dst     = peelUnrealized(adaptor.getDst());
 
     replaceOrEraseWithOpaqueCallAndReturnDst(op.getOperation(), dst, "TGEMV_MX",
-                                             {dst, a, aScale, b, bScale, bias}, rewriter);
+                                             {dst, a, aScale, b, bScale, bias}, ArrayAttr{},
+                                             rewriter);
     return success();
   }
 };
@@ -9804,8 +9803,10 @@ struct PTOTMatmulBiasToTMATMUL_BIAS
     Value bias = peelUnrealized(adaptor.getBias());
     Value dst  = peelUnrealized(adaptor.getDst());
 
-    replaceOrEraseWithOpaqueCall(op.getOperation(), "TMATMUL_BIAS",
-                                {dst, a, b, bias}, rewriter);
+    ArrayAttr templateArgs =
+        buildAccPhaseTemplateArgs(rewriter, op.getAccPhase());
+    replaceOrEraseWithOpaqueCallAndReturnDst(op.getOperation(), dst, "TMATMUL_BIAS",
+                                             {dst, a, b, bias}, templateArgs, rewriter);
     return success();
   }
 };
@@ -9822,8 +9823,11 @@ struct PTOTMatmulMXToTMATMUL_MX
     Value bScale  = peelUnrealized(adaptor.getBScale());
     Value dst     = peelUnrealized(adaptor.getDst());
 
-    replaceOrEraseWithOpaqueCall(op.getOperation(), "TMATMUL_MX",
-                                {dst, a, aScale, b, bScale}, rewriter);
+    ArrayAttr templateArgs =
+        buildAccPhaseTemplateArgs(rewriter, op.getAccPhase());
+    replaceOrEraseWithOpaqueCallAndReturnDst(op.getOperation(), dst, "TMATMUL_MX",
+                                             {dst, a, aScale, b, bScale}, templateArgs,
+                                             rewriter);
     return success();
   }
 };
@@ -9841,8 +9845,9 @@ struct PTOTMatmulMXAccToTMATMUL_MX_ACC
     Value bScale  = peelUnrealized(adaptor.getBScale());
     Value dst     = peelUnrealized(adaptor.getDst());
 
-    replaceOrEraseWithOpaqueCall(op.getOperation(), "TMATMUL_MX",
-                                {dst, cIn, a, aScale, b, bScale}, rewriter);
+    replaceOrEraseWithOpaqueCallAndReturnDst(op.getOperation(), dst, "TMATMUL_MX",
+                                             {dst, cIn, a, aScale, b, bScale}, ArrayAttr{},
+                                             rewriter);
     return success();
   }
 };
@@ -9860,8 +9865,9 @@ struct PTOTMatmulMXBiasToTMATMUL_MX_BIAS
     Value bias    = peelUnrealized(adaptor.getBias());
     Value dst     = peelUnrealized(adaptor.getDst());
 
-    replaceOrEraseWithOpaqueCall(op.getOperation(), "TMATMUL_MX",
-                                {dst, a, aScale, b, bScale, bias}, rewriter);
+    replaceOrEraseWithOpaqueCallAndReturnDst(op.getOperation(), dst, "TMATMUL_MX",
+                                             {dst, a, aScale, b, bScale, bias}, ArrayAttr{},
+                                             rewriter);
     return success();
   }
 };
diff --git a/lib/PTO/Transforms/PTOViewToMemref.cpp b/lib/PTO/Transforms/PTOViewToMemref.cpp
index 63849ef2a..f17d2e99d 100644
--- a/lib/PTO/Transforms/PTOViewToMemref.cpp
+++ b/lib/PTO/Transforms/PTOViewToMemref.cpp
@@ -1940,7 +1940,7 @@ struct PTOViewToMemrefPass
           op, TypeRange{}, 
           op->getOperand(0), op->getOperand(1),
           op->getOperand(kThirdOperandIndex),
-          op->getOperand(kFourthOperandIndex));
+          op->getOperand(kFourthOperandIndex), op.getAccPhaseAttr());
       }
 
       // --- TMatmulMxOp---
@@ -1954,7 +1954,7 @@ struct PTOViewToMemrefPass
           op->getOperand(0), op->getOperand(1),
           op->getOperand(kThirdOperandIndex),
           op->getOperand(kFourthOperandIndex),
-          op->getOperand(kFifthOperandIndex));
+          op->getOperand(kFifthOperandIndex), op.getAccPhaseAttr());
       }
 
       // --- TMatmulMxAccOp  ---
@@ -1999,7 +1999,7 @@ struct PTOViewToMemrefPass
         Value dst = op->getOperand(kThirdOperandIndex);
 
         rewriter.replaceOpWithNewOp<pto::TGemvOp>(
-          op, TypeRange{}, lhs, rhs, dst);
+          op, TypeRange{}, lhs, rhs, dst, op.getAccPhaseAttr());
       }
 
       // --- TGemvAccOp [Acc, Lhs, Rhs, Dst] ---
@@ -2012,7 +2012,7 @@ struct PTOViewToMemrefPass
           op, TypeRange{}, 
           op->getOperand(0), op->getOperand(1),
           op->getOperand(kThirdOperandIndex),
-          op->getOperand(kFourthOperandIndex));
+          op->getOperand(kFourthOperandIndex), op.getAccPhaseAttr());
       }
 
       // --- TGemvBiasOp [Acc, Lhs, Rhs, Bias, Dst] ---
@@ -2025,7 +2025,7 @@ struct PTOViewToMemrefPass
           op, TypeRange{}, 
           op->getOperand(0), op->getOperand(1),
           op->getOperand(kThirdOperandIndex),
-          op->getOperand(kFourthOperandIndex));
+          op->getOperand(kFourthOperandIndex), op.getAccPhaseAttr());
       }
 
       // --- TGemvMxOp [A, AScale, B, BScale, Dst] ---
@@ -2039,7 +2039,7 @@ struct PTOViewToMemrefPass
           op->getOperand(0), op->getOperand(1),
           op->getOperand(kThirdOperandIndex),
           op->getOperand(kFourthOperandIndex),
-          op->getOperand(kFifthOperandIndex));
+          op->getOperand(kFifthOperandIndex), op.getAccPhaseAttr());
       }
 
       // --- TGemvMxAccOp [CIn, A, AScale, B, BScale, Dst] ---
diff --git a/test/lit/pto/tgemv_mx_accphase_emitc.pto b/test/lit/pto/tgemv_mx_accphase_emitc.pto
index f8c5d5876..7a434746a 100644
--- a/test/lit/pto/tgemv_mx_accphase_emitc.pto
+++ b/test/lit/pto/tgemv_mx_accphase_emitc.pto
@@ -1,15 +1,18 @@
 // RUN: ptoas --pto-arch=a5 %s 2>&1 | FileCheck %s --check-prefix=A5
 
 module attributes {pto.target_arch = "a5"} {
-  func.func @tgemv_mx_accphase_emitc(%a : !pto.tile_buf<loc=left, dtype=f8E4M3FN, rows=1, cols=128, v_row=1, v_col=128, blayout=col_major, slayout=row_major, fractal=512, pad=0>,
-                                     %b : !pto.tile_buf<loc=right, dtype=f8E4M3FN, rows=128, cols=32, v_row=128, v_col=32, blayout=row_major, slayout=col_major, fractal=512, pad=0>) attributes {pto.kernel_kind = #pto.kernel_kind<cube>} {
+  func.func @tgemv_mx_accphase_emitc() attributes {pto.kernel_kind = #pto.kernel_kind<cube>} {
+    %a = pto.alloc_tile : !pto.tile_buf<loc=left, dtype=f8E4M3FN, rows=1, cols=128, v_row=1, v_col=128, blayout=col_major, slayout=row_major, fractal=512, pad=0>
+    %b = pto.alloc_tile : !pto.tile_buf<loc=right, dtype=f8E4M3FN, rows=128, cols=32, v_row=128, v_col=32, blayout=row_major, slayout=col_major, fractal=512, pad=0>
     %a_scale = pto.alloc_tile : !pto.tile_buf<loc=scaling, dtype=f16, rows=1, cols=128, v_row=1, v_col=128, blayout=row_major, slayout=row_major, fractal=512, pad=0>
     %b_scale = pto.alloc_tile : !pto.tile_buf<loc=scaling, dtype=f16, rows=128, cols=32, v_row=128, v_col=32, blayout=row_major, slayout=row_major, fractal=512, pad=0>
-    %dst = pto.alloc_tile : !pto.tile_buf<loc=acc, dtype=f32, rows=1, cols=32, v_row=1, v_col=32, blayout=col_major, slayout=row_major, fractal=1024, pad=0>
+    %dst0 = pto.alloc_tile : !pto.tile_buf<loc=acc, dtype=f32, rows=1, cols=32, v_row=1, v_col=32, blayout=col_major, slayout=row_major, fractal=1024, pad=0>
+    %dst1 = pto.alloc_tile : !pto.tile_buf<loc=acc, dtype=f32, rows=1, cols=32, v_row=1, v_col=32, blayout=col_major, slayout=row_major, fractal=1024, pad=0>
+    %dst2 = pto.alloc_tile : !pto.tile_buf<loc=acc, dtype=f32, rows=1, cols=32, v_row=1, v_col=32, blayout=col_major, slayout=row_major, fractal=1024, pad=0>
 
-    pto.tgemv.mx ins(%a, %a_scale, %b, %b_scale : !pto.tile_buf<loc=left, dtype=f8E4M3FN, rows=1, cols=128, v_row=1, v_col=128, blayout=col_major, slayout=row_major, fractal=512, pad=0>, !pto.tile_buf<loc=scaling, dtype=f16, rows=1, cols=128, v_row=1, v_col=128, blayout=row_major, slayout=row_major, fractal=512, pad=0>, !pto.tile_buf<loc=right, dtype=f8E4M3FN, rows=128, cols=32, v_row=128, v_col=32, blayout=row_major, slayout=col_major, fractal=512, pad=0>, !pto.tile_buf<loc=scaling, dtype=f16, rows=128, cols=32, v_row=128, v_col=32, blayout=row_major, slayout=row_major, fractal=512, pad=0>) outs(%dst : !pto.tile_buf<loc=acc, dtype=f32, rows=1, cols=32, v_row=1, v_col=32, blayout=col_major, slayout=row_major, fractal=1024, pad=0>)
-    pto.tgemv.mx ins(%a, %a_scale, %b, %b_scale : !pto.tile_buf<loc=left, dtype=f8E4M3FN, rows=1, cols=128, v_row=1, v_col=128, blayout=col_major, slayout=row_major, fractal=512, pad=0>, !pto.tile_buf<loc=scaling, dtype=f16, rows=1, cols=128, v_row=1, v_col=128, blayout=row_major, slayout=row_major, fractal=512, pad=0>, !pto.tile_buf<loc=right, dtype=f8E4M3FN, rows=128, cols=32, v_row=128, v_col=32, blayout=row_major, slayout=col_major, fractal=512, pad=0>, !pto.tile_buf<loc=scaling, dtype=f16, rows=128, cols=32, v_row=128, v_col=32, blayout=row_major, slayout=row_major, fractal=512, pad=0>) outs(%dst : !pto.tile_buf<loc=acc, dtype=f32, rows=1, cols=32, v_row=1, v_col=32, blayout=col_major, slayout=row_major, fractal=1024, pad=0>) {accPhase = #pto<acc_phase partial>}
-    pto.tgemv.mx ins(%a, %a_scale, %b, %b_scale : !pto.tile_buf<loc=left, dtype=f8E4M3FN, rows=1, cols=128, v_row=1, v_col=128, blayout=col_major, slayout=row_major, fractal=512, pad=0>, !pto.tile_buf<loc=scaling, dtype=f16, rows=1, cols=128, v_row=1, v_col=128, blayout=row_major, slayout=row_major, fractal=512, pad=0>, !pto.tile_buf<loc=right, dtype=f8E4M3FN, rows=128, cols=32, v_row=128, v_col=32, blayout=row_major, slayout=col_major, fractal=512, pad=0>, !pto.tile_buf<loc=scaling, dtype=f16, rows=128, cols=32, v_row=128, v_col=32, blayout=row_major, slayout=row_major, fractal=512, pad=0>) outs(%dst : !pto.tile_buf<loc=acc, dtype=f32, rows=1, cols=32, v_row=1, v_col=32, blayout=col_major, slayout=row_major, fractal=1024, pad=0>) {accPhase = #pto<acc_phase final>}
+    pto.tgemv.mx ins(%a, %a_scale, %b, %b_scale : !pto.tile_buf<loc=left, dtype=f8E4M3FN, rows=1, cols=128, v_row=1, v_col=128, blayout=col_major, slayout=row_major, fractal=512, pad=0>, !pto.tile_buf<loc=scaling, dtype=f16, rows=1, cols=128, v_row=1, v_col=128, blayout=row_major, slayout=row_major, fractal=512, pad=0>, !pto.tile_buf<loc=right, dtype=f8E4M3FN, rows=128, cols=32, v_row=128, v_col=32, blayout=row_major, slayout=col_major, fractal=512, pad=0>, !pto.tile_buf<loc=scaling, dtype=f16, rows=128, cols=32, v_row=128, v_col=32, blayout=row_major, slayout=row_major, fractal=512, pad=0>) outs(%dst0 : !pto.tile_buf<loc=acc, dtype=f32, rows=1, cols=32, v_row=1, v_col=32, blayout=col_major, slayout=row_major, fractal=1024, pad=0>)
+    pto.tgemv.mx ins(%a, %a_scale, %b, %b_scale : !pto.tile_buf<loc=left, dtype=f8E4M3FN, rows=1, cols=128, v_row=1, v_col=128, blayout=col_major, slayout=row_major, fractal=512, pad=0>, !pto.tile_buf<loc=scaling, dtype=f16, rows=1, cols=128, v_row=1, v_col=128, blayout=row_major, slayout=row_major, fractal=512, pad=0>, !pto.tile_buf<loc=right, dtype=f8E4M3FN, rows=128, cols=32, v_row=128, v_col=32, blayout=row_major, slayout=col_major, fractal=512, pad=0>, !pto.tile_buf<loc=scaling, dtype=f16, rows=128, cols=32, v_row=128, v_col=32, blayout=row_major, slayout=row_major, fractal=512, pad=0>) outs(%dst1 : !pto.tile_buf<loc=acc, dtype=f32, rows=1, cols=32, v_row=1, v_col=32, blayout=col_major, slayout=row_major, fractal=1024, pad=0>) {accPhase = #pto<acc_phase partial>}
+    pto.tgemv.mx ins(%a, %a_scale, %b, %b_scale : !pto.tile_buf<loc=left, dtype=f8E4M3FN, rows=1, cols=128, v_row=1, v_col=128, blayout=col_major, slayout=row_major, fractal=512, pad=0>, !pto.tile_buf<loc=scaling, dtype=f16, rows=1, cols=128, v_row=1, v_col=128, blayout=row_major, slayout=row_major, fractal=512, pad=0>, !pto.tile_buf<loc=right, dtype=f8E4M3FN, rows=128, cols=32, v_row=128, v_col=32, blayout=row_major, slayout=col_major, fractal=512, pad=0>, !pto.tile_buf<loc=scaling, dtype=f16, rows=128, cols=32, v_row=128, v_col=32, blayout=row_major, slayout=row_major, fractal=512, pad=0>) outs(%dst2 : !pto.tile_buf<loc=acc, dtype=f32, rows=1, cols=32, v_row=1, v_col=32, blayout=col_major, slayout=row_major, fractal=1024, pad=0>) {accPhase = #pto<acc_phase final>}
     return
   }
 }
diff --git a/test/lit/pto/tmatmul_acc_phase_emitc.pto b/test/lit/pto/tmatmul_acc_phase_emitc.pto
index 868a991dc..41dad0049 100644
--- a/test/lit/pto/tmatmul_acc_phase_emitc.pto
+++ b/test/lit/pto/tmatmul_acc_phase_emitc.pto
@@ -49,8 +49,8 @@ module {
 }
 
 // A3: TMATMUL([[C:v[0-9]+]], [[A:v[0-9]+]], [[B:v[0-9]+]]);
-// A3: TMATMUL<AccPhase::Partial>([[C]], [[A]], [[B]]);
-// A3: TMATMUL<AccPhase::Final>([[C]], [[A]], [[B]]);
+// A3: TMATMUL<pto::AccPhase::Partial>([[C]], [[A]], [[B]]);
+// A3: TMATMUL<pto::AccPhase::Final>([[C]], [[A]], [[B]]);
 // A3: TMATMUL_ACC([[C]], [[C]], [[A]], [[B]]);
-// A3: TMATMUL_ACC<AccPhase::Partial>([[C]], [[C]], [[A]], [[B]]);
-// A3: TMATMUL_ACC<AccPhase::Final>([[C]], [[C]], [[A]], [[B]]);
+// A3: TMATMUL_ACC<pto::AccPhase::Partial>([[C]], [[C]], [[A]], [[B]]);
+// A3: TMATMUL_ACC<pto::AccPhase::Final>([[C]], [[C]], [[A]], [[B]]);

From 89dc65404639deb94534fe45371b9965188c7e9f Mon Sep 17 00:00:00 2001
From: jimmychou <1049859649@qq.com>
Date: Wed, 27 May 2026 17:38:47 +0800
Subject: [PATCH 6/6] fix(pto): complete MX accphase coverage

---
 include/PTO/IR/PTOOps.td                      |  6 +++--
 lib/PTO/IR/PTO.cpp                            | 22 +++++++++++++++++--
 lib/PTO/Transforms/PTOToEmitC.cpp             |  8 +++++--
 lib/PTO/Transforms/PTOViewToMemref.cpp        |  4 ++--
 test/lit/pto/tgemv_mx_accphase_emitc.pto      | 10 +++++++++
 .../pto/tgemv_mx_verify_invalid_shape_a5.pto  | 15 +++++++++++++
 test/lit/pto/tmatmul_mx_accphase_emitc.pto    | 10 +++++++++
 7 files changed, 67 insertions(+), 8 deletions(-)
 create mode 100644 test/lit/pto/tgemv_mx_verify_invalid_shape_a5.pto

diff --git a/include/PTO/IR/PTOOps.td b/include/PTO/IR/PTOOps.td
index 863ed8c51..99b00c28d 100644
--- a/include/PTO/IR/PTOOps.td
+++ b/include/PTO/IR/PTOOps.td
@@ -894,7 +894,8 @@ def TMatmulMxAccOp : PTO_TOp<"tmatmul.mx.acc", [
     PTODpsType:$a_scale,
     PTODpsType:$b,
     PTODpsType:$b_scale,
-    PTODpsType:$dst);
+    PTODpsType:$dst,
+    DefaultValuedAttr<PTO_AccPhaseAttr, "::mlir::pto::AccPhase::Unspecified">:$accPhase);
 
   let results = (outs Optional<AnyRankedTensor>:$result);
   let hasVerifier = 1;
@@ -1177,7 +1178,8 @@ def TGemvMxAccOp : PTO_TOp<"tgemv.mx.acc", [
     PTODpsType:$a_scale,
     PTODpsType:$b,
     PTODpsType:$b_scale,
-    PTODpsType:$dst
+    PTODpsType:$dst,
+    DefaultValuedAttr<PTO_AccPhaseAttr, "::mlir::pto::AccPhase::Unspecified">:$accPhase
   );
 
   let results = (outs Optional<AnyRankedTensor>:$result);
diff --git a/lib/PTO/IR/PTO.cpp b/lib/PTO/IR/PTO.cpp
index d3d557a84..e9d4b3f73 100644
--- a/lib/PTO/IR/PTO.cpp
+++ b/lib/PTO/IR/PTO.cpp
@@ -4138,9 +4138,27 @@ static LogicalResult verifyA5MxGemvTileOperands(Operation *op, Type lhsTy,
     return op->emitOpError(
         "expects lhs, rhs, and dst to use the left, right, and acc address spaces");
 
+  auto lhsShape = getMatmulLogicalShapeVec(lhsTy);
+  auto rhsShape = getMatmulLogicalShapeVec(rhsTy);
+  auto dstShape = getMatmulLogicalShapeVec(dstTy);
+  if ((lhsShape[0] != dstShape[0] || rhsShape[1] != dstShape[1] ||
+       lhsShape[1] != rhsShape[0]))
+    return op->emitOpError(
+        "expects static matmul tile shapes lhs[M,K], rhs[K,N], and dst[M,N]");
+
   auto lhsValid = getValidShapeVec(lhsTy);
   auto rhsValid = getValidShapeVec(rhsTy);
   auto dstValid = getValidShapeVec(dstTy);
+  if (lhsValid.size() == 2 && rhsValid.size() == 2) {
+    int64_t m = lhsValid[0];
+    int64_t k = lhsValid[1];
+    int64_t n = rhsValid[1];
+    if ((m != ShapedType::kDynamic && (m < 1 || m > 4095)) ||
+        (k != ShapedType::kDynamic && (k < 1 || k > 4095)) ||
+        (n != ShapedType::kDynamic && (n < 1 || n > 4095)))
+      return op->emitOpError("expects m, k, and n valid sizes to be in [1, 4095]");
+  }
+
   if (lhsValid[0] != ShapedType::kDynamic && lhsValid[0] != 1)
     return op->emitOpError("expects lhs valid_shape[0] to be 1 for tgemv");
   if (dstValid[0] != ShapedType::kDynamic && dstValid[0] != 1)
@@ -6692,8 +6710,8 @@ LogicalResult TGemvMxAccOp::verify() {
                                              getA().getType(), "a_scale", "a")) ||
         failed(verifyScaleTileMatchesOperand(*this, getBScale().getType(),
                                              getB().getType(), "b_scale", "b")) ||
-        failed(verifyGemvTileOperands(*this, getA().getType(), getB().getType(),
-                                      getDst().getType())))
+        failed(verifyA5MxGemvTileOperands(*this, getA().getType(), getB().getType(),
+                                          getDst().getType())))
       return failure();
     if (failed(verifyA5MxTypeTriple(*this, getA().getType(), getB().getType(),
                                     getDst().getType(), "lhs", "rhs", "dst")))
diff --git a/lib/PTO/Transforms/PTOToEmitC.cpp b/lib/PTO/Transforms/PTOToEmitC.cpp
index e9e65050a..4d2b0f6a5 100644
--- a/lib/PTO/Transforms/PTOToEmitC.cpp
+++ b/lib/PTO/Transforms/PTOToEmitC.cpp
@@ -9765,8 +9765,10 @@ struct PTOTGemvMXAccToTGEMV_MX
     Value bScale  = peelUnrealized(adaptor.getBScale());
     Value dst     = peelUnrealized(adaptor.getDst());
 
+    ArrayAttr templateArgs =
+        buildAccPhaseTemplateArgs(rewriter, op.getAccPhase());
     replaceOrEraseWithOpaqueCallAndReturnDst(op.getOperation(), dst, "TGEMV_MX",
-                                             {dst, cIn, a, aScale, b, bScale}, ArrayAttr{},
+                                             {dst, cIn, a, aScale, b, bScale}, templateArgs,
                                              rewriter);
     return success();
   }
@@ -9845,8 +9847,10 @@ struct PTOTMatmulMXAccToTMATMUL_MX_ACC
     Value bScale  = peelUnrealized(adaptor.getBScale());
     Value dst     = peelUnrealized(adaptor.getDst());
 
+    ArrayAttr templateArgs =
+        buildAccPhaseTemplateArgs(rewriter, op.getAccPhase());
     replaceOrEraseWithOpaqueCallAndReturnDst(op.getOperation(), dst, "TMATMUL_MX",
-                                             {dst, cIn, a, aScale, b, bScale}, ArrayAttr{},
+                                             {dst, cIn, a, aScale, b, bScale}, templateArgs,
                                              rewriter);
     return success();
   }
diff --git a/lib/PTO/Transforms/PTOViewToMemref.cpp b/lib/PTO/Transforms/PTOViewToMemref.cpp
index f17d2e99d..075d37a7c 100644
--- a/lib/PTO/Transforms/PTOViewToMemref.cpp
+++ b/lib/PTO/Transforms/PTOViewToMemref.cpp
@@ -1969,7 +1969,7 @@ struct PTOViewToMemrefPass
           op->getOperand(kThirdOperandIndex),
           op->getOperand(kFourthOperandIndex),
           op->getOperand(kFifthOperandIndex),
-          op->getOperand(kSixthOperandIndex));
+          op->getOperand(kSixthOperandIndex), op.getAccPhaseAttr());
       }
 
       // --- TMatmulMxBiasOp ---
@@ -2054,7 +2054,7 @@ struct PTOViewToMemrefPass
           op->getOperand(kThirdOperandIndex),
           op->getOperand(kFourthOperandIndex),
           op->getOperand(kFifthOperandIndex),
-          op->getOperand(kSixthOperandIndex));
+          op->getOperand(kSixthOperandIndex), op.getAccPhaseAttr());
       }
 
       // --- TGemvMxBiasOp [A, AScale, B, BScale, Bias, Dst] ---
diff --git a/test/lit/pto/tgemv_mx_accphase_emitc.pto b/test/lit/pto/tgemv_mx_accphase_emitc.pto
index 7a434746a..dc1b70601 100644
--- a/test/lit/pto/tgemv_mx_accphase_emitc.pto
+++ b/test/lit/pto/tgemv_mx_accphase_emitc.pto
@@ -6,13 +6,20 @@ module attributes {pto.target_arch = "a5"} {
     %b = pto.alloc_tile : !pto.tile_buf<loc=right, dtype=f8E4M3FN, rows=128, cols=32, v_row=128, v_col=32, blayout=row_major, slayout=col_major, fractal=512, pad=0>
     %a_scale = pto.alloc_tile : !pto.tile_buf<loc=scaling, dtype=f16, rows=1, cols=128, v_row=1, v_col=128, blayout=row_major, slayout=row_major, fractal=512, pad=0>
     %b_scale = pto.alloc_tile : !pto.tile_buf<loc=scaling, dtype=f16, rows=128, cols=32, v_row=128, v_col=32, blayout=row_major, slayout=row_major, fractal=512, pad=0>
+    %c_in = pto.alloc_tile : !pto.tile_buf<loc=acc, dtype=f32, rows=1, cols=32, v_row=1, v_col=32, blayout=col_major, slayout=row_major, fractal=1024, pad=0>
     %dst0 = pto.alloc_tile : !pto.tile_buf<loc=acc, dtype=f32, rows=1, cols=32, v_row=1, v_col=32, blayout=col_major, slayout=row_major, fractal=1024, pad=0>
     %dst1 = pto.alloc_tile : !pto.tile_buf<loc=acc, dtype=f32, rows=1, cols=32, v_row=1, v_col=32, blayout=col_major, slayout=row_major, fractal=1024, pad=0>
     %dst2 = pto.alloc_tile : !pto.tile_buf<loc=acc, dtype=f32, rows=1, cols=32, v_row=1, v_col=32, blayout=col_major, slayout=row_major, fractal=1024, pad=0>
+    %dst3 = pto.alloc_tile : !pto.tile_buf<loc=acc, dtype=f32, rows=1, cols=32, v_row=1, v_col=32, blayout=col_major, slayout=row_major, fractal=1024, pad=0>
+    %dst4 = pto.alloc_tile : !pto.tile_buf<loc=acc, dtype=f32, rows=1, cols=32, v_row=1, v_col=32, blayout=col_major, slayout=row_major, fractal=1024, pad=0>
+    %dst5 = pto.alloc_tile : !pto.tile_buf<loc=acc, dtype=f32, rows=1, cols=32, v_row=1, v_col=32, blayout=col_major, slayout=row_major, fractal=1024, pad=0>
 
     pto.tgemv.mx ins(%a, %a_scale, %b, %b_scale : !pto.tile_buf<loc=left, dtype=f8E4M3FN, rows=1, cols=128, v_row=1, v_col=128, blayout=col_major, slayout=row_major, fractal=512, pad=0>, !pto.tile_buf<loc=scaling, dtype=f16, rows=1, cols=128, v_row=1, v_col=128, blayout=row_major, slayout=row_major, fractal=512, pad=0>, !pto.tile_buf<loc=right, dtype=f8E4M3FN, rows=128, cols=32, v_row=128, v_col=32, blayout=row_major, slayout=col_major, fractal=512, pad=0>, !pto.tile_buf<loc=scaling, dtype=f16, rows=128, cols=32, v_row=128, v_col=32, blayout=row_major, slayout=row_major, fractal=512, pad=0>) outs(%dst0 : !pto.tile_buf<loc=acc, dtype=f32, rows=1, cols=32, v_row=1, v_col=32, blayout=col_major, slayout=row_major, fractal=1024, pad=0>)
     pto.tgemv.mx ins(%a, %a_scale, %b, %b_scale : !pto.tile_buf<loc=left, dtype=f8E4M3FN, rows=1, cols=128, v_row=1, v_col=128, blayout=col_major, slayout=row_major, fractal=512, pad=0>, !pto.tile_buf<loc=scaling, dtype=f16, rows=1, cols=128, v_row=1, v_col=128, blayout=row_major, slayout=row_major, fractal=512, pad=0>, !pto.tile_buf<loc=right, dtype=f8E4M3FN, rows=128, cols=32, v_row=128, v_col=32, blayout=row_major, slayout=col_major, fractal=512, pad=0>, !pto.tile_buf<loc=scaling, dtype=f16, rows=128, cols=32, v_row=128, v_col=32, blayout=row_major, slayout=row_major, fractal=512, pad=0>) outs(%dst1 : !pto.tile_buf<loc=acc, dtype=f32, rows=1, cols=32, v_row=1, v_col=32, blayout=col_major, slayout=row_major, fractal=1024, pad=0>) {accPhase = #pto<acc_phase partial>}
     pto.tgemv.mx ins(%a, %a_scale, %b, %b_scale : !pto.tile_buf<loc=left, dtype=f8E4M3FN, rows=1, cols=128, v_row=1, v_col=128, blayout=col_major, slayout=row_major, fractal=512, pad=0>, !pto.tile_buf<loc=scaling, dtype=f16, rows=1, cols=128, v_row=1, v_col=128, blayout=row_major, slayout=row_major, fractal=512, pad=0>, !pto.tile_buf<loc=right, dtype=f8E4M3FN, rows=128, cols=32, v_row=128, v_col=32, blayout=row_major, slayout=col_major, fractal=512, pad=0>, !pto.tile_buf<loc=scaling, dtype=f16, rows=128, cols=32, v_row=128, v_col=32, blayout=row_major, slayout=row_major, fractal=512, pad=0>) outs(%dst2 : !pto.tile_buf<loc=acc, dtype=f32, rows=1, cols=32, v_row=1, v_col=32, blayout=col_major, slayout=row_major, fractal=1024, pad=0>) {accPhase = #pto<acc_phase final>}
+    pto.tgemv.mx.acc ins(%c_in, %a, %a_scale, %b, %b_scale : !pto.tile_buf<loc=acc, dtype=f32, rows=1, cols=32, v_row=1, v_col=32, blayout=col_major, slayout=row_major, fractal=1024, pad=0>, !pto.tile_buf<loc=left, dtype=f8E4M3FN, rows=1, cols=128, v_row=1, v_col=128, blayout=col_major, slayout=row_major, fractal=512, pad=0>, !pto.tile_buf<loc=scaling, dtype=f16, rows=1, cols=128, v_row=1, v_col=128, blayout=row_major, slayout=row_major, fractal=512, pad=0>, !pto.tile_buf<loc=right, dtype=f8E4M3FN, rows=128, cols=32, v_row=128, v_col=32, blayout=row_major, slayout=col_major, fractal=512, pad=0>, !pto.tile_buf<loc=scaling, dtype=f16, rows=128, cols=32, v_row=128, v_col=32, blayout=row_major, slayout=row_major, fractal=512, pad=0>) outs(%dst3 : !pto.tile_buf<loc=acc, dtype=f32, rows=1, cols=32, v_row=1, v_col=32, blayout=col_major, slayout=row_major, fractal=1024, pad=0>)
+    pto.tgemv.mx.acc ins(%c_in, %a, %a_scale, %b, %b_scale : !pto.tile_buf<loc=acc, dtype=f32, rows=1, cols=32, v_row=1, v_col=32, blayout=col_major, slayout=row_major, fractal=1024, pad=0>, !pto.tile_buf<loc=left, dtype=f8E4M3FN, rows=1, cols=128, v_row=1, v_col=128, blayout=col_major, slayout=row_major, fractal=512, pad=0>, !pto.tile_buf<loc=scaling, dtype=f16, rows=1, cols=128, v_row=1, v_col=128, blayout=row_major, slayout=row_major, fractal=512, pad=0>, !pto.tile_buf<loc=right, dtype=f8E4M3FN, rows=128, cols=32, v_row=128, v_col=32, blayout=row_major, slayout=col_major, fractal=512, pad=0>, !pto.tile_buf<loc=scaling, dtype=f16, rows=128, cols=32, v_row=128, v_col=32, blayout=row_major, slayout=row_major, fractal=512, pad=0>) outs(%dst4 : !pto.tile_buf<loc=acc, dtype=f32, rows=1, cols=32, v_row=1, v_col=32, blayout=col_major, slayout=row_major, fractal=1024, pad=0>) {accPhase = #pto<acc_phase partial>}
+    pto.tgemv.mx.acc ins(%c_in, %a, %a_scale, %b, %b_scale : !pto.tile_buf<loc=acc, dtype=f32, rows=1, cols=32, v_row=1, v_col=32, blayout=col_major, slayout=row_major, fractal=1024, pad=0>, !pto.tile_buf<loc=left, dtype=f8E4M3FN, rows=1, cols=128, v_row=1, v_col=128, blayout=col_major, slayout=row_major, fractal=512, pad=0>, !pto.tile_buf<loc=scaling, dtype=f16, rows=1, cols=128, v_row=1, v_col=128, blayout=row_major, slayout=row_major, fractal=512, pad=0>, !pto.tile_buf<loc=right, dtype=f8E4M3FN, rows=128, cols=32, v_row=128, v_col=32, blayout=row_major, slayout=col_major, fractal=512, pad=0>, !pto.tile_buf<loc=scaling, dtype=f16, rows=128, cols=32, v_row=128, v_col=32, blayout=row_major, slayout=row_major, fractal=512, pad=0>) outs(%dst5 : !pto.tile_buf<loc=acc, dtype=f32, rows=1, cols=32, v_row=1, v_col=32, blayout=col_major, slayout=row_major, fractal=1024, pad=0>) {accPhase = #pto<acc_phase final>}
     return
   }
 }
@@ -20,3 +27,6 @@ module attributes {pto.target_arch = "a5"} {
 // A5: TGEMV_MX([[D0:v[0-9]+]], [[A0:v[0-9]+]], [[AS0:v[0-9]+]], [[B0:v[0-9]+]], [[BS0:v[0-9]+]]);
 // A5: TGEMV_MX<pto::AccPhase::Partial>([[D1:v[0-9]+]], [[A1:v[0-9]+]], [[AS1:v[0-9]+]], [[B1:v[0-9]+]], [[BS1:v[0-9]+]]);
 // A5: TGEMV_MX<pto::AccPhase::Final>([[D2:v[0-9]+]], [[A2:v[0-9]+]], [[AS2:v[0-9]+]], [[B2:v[0-9]+]], [[BS2:v[0-9]+]]);
+// A5: TGEMV_MX([[D3:v[0-9]+]], [[C3:v[0-9]+]], [[A3:v[0-9]+]], [[AS3:v[0-9]+]], [[B3:v[0-9]+]], [[BS3:v[0-9]+]]);
+// A5: TGEMV_MX<pto::AccPhase::Partial>([[D4:v[0-9]+]], [[C4:v[0-9]+]], [[A4:v[0-9]+]], [[AS4:v[0-9]+]], [[B4:v[0-9]+]], [[BS4:v[0-9]+]]);
+// A5: TGEMV_MX<pto::AccPhase::Final>([[D5:v[0-9]+]], [[C5:v[0-9]+]], [[A5:v[0-9]+]], [[AS5:v[0-9]+]], [[B5:v[0-9]+]], [[BS5:v[0-9]+]]);
diff --git a/test/lit/pto/tgemv_mx_verify_invalid_shape_a5.pto b/test/lit/pto/tgemv_mx_verify_invalid_shape_a5.pto
new file mode 100644
index 000000000..a0f4da6fe
--- /dev/null
+++ b/test/lit/pto/tgemv_mx_verify_invalid_shape_a5.pto
@@ -0,0 +1,15 @@
+// RUN: not ptoas --pto-arch=a5 %s 2>&1 | FileCheck %s
+
+module attributes {pto.target_arch = "a5"} {
+  func.func @tgemv_mx_invalid_static_shape(%a : !pto.tile_buf<loc=left, dtype=f8E4M3FN, rows=1, cols=128, v_row=1, v_col=128, blayout=col_major, slayout=row_major, fractal=512, pad=0>,
+                                           %b : !pto.tile_buf<loc=right, dtype=f8E4M3FN, rows=64, cols=32, v_row=64, v_col=32, blayout=row_major, slayout=col_major, fractal=512, pad=0>) attributes {pto.kernel_kind = #pto.kernel_kind<cube>} {
+    %a_scale = pto.alloc_tile : !pto.tile_buf<loc=scaling, dtype=f16, rows=1, cols=128, v_row=1, v_col=128, blayout=row_major, slayout=row_major, fractal=512, pad=0>
+    %b_scale = pto.alloc_tile : !pto.tile_buf<loc=scaling, dtype=f16, rows=64, cols=32, v_row=64, v_col=32, blayout=row_major, slayout=row_major, fractal=512, pad=0>
+    %dst = pto.alloc_tile : !pto.tile_buf<loc=acc, dtype=f32, rows=1, cols=32, v_row=1, v_col=32, blayout=col_major, slayout=row_major, fractal=1024, pad=0>
+
+    pto.tgemv.mx ins(%a, %a_scale, %b, %b_scale : !pto.tile_buf<loc=left, dtype=f8E4M3FN, rows=1, cols=128, v_row=1, v_col=128, blayout=col_major, slayout=row_major, fractal=512, pad=0>, !pto.tile_buf<loc=scaling, dtype=f16, rows=1, cols=128, v_row=1, v_col=128, blayout=row_major, slayout=row_major, fractal=512, pad=0>, !pto.tile_buf<loc=right, dtype=f8E4M3FN, rows=64, cols=32, v_row=64, v_col=32, blayout=row_major, slayout=col_major, fractal=512, pad=0>, !pto.tile_buf<loc=scaling, dtype=f16, rows=64, cols=32, v_row=64, v_col=32, blayout=row_major, slayout=row_major, fractal=512, pad=0>) outs(%dst : !pto.tile_buf<loc=acc, dtype=f32, rows=1, cols=32, v_row=1, v_col=32, blayout=col_major, slayout=row_major, fractal=1024, pad=0>)
+    return
+  }
+}
+
+// CHECK: error: 'pto.tgemv.mx' op expects static matmul tile shapes lhs[M,K], rhs[K,N], and dst[M,N]
diff --git a/test/lit/pto/tmatmul_mx_accphase_emitc.pto b/test/lit/pto/tmatmul_mx_accphase_emitc.pto
index bde097042..d62c8de84 100644
--- a/test/lit/pto/tmatmul_mx_accphase_emitc.pto
+++ b/test/lit/pto/tmatmul_mx_accphase_emitc.pto
@@ -5,13 +5,20 @@ module attributes {pto.target_arch = "a5"} {
                                        %b : !pto.tile_buf<loc=right, dtype=f8E4M3FN, rows=32, cols=32, v_row=32, v_col=32, blayout=row_major, slayout=col_major, fractal=512, pad=0>) attributes {pto.kernel_kind = #pto.kernel_kind<cube>} {
     %a_scale = pto.alloc_tile : !pto.tile_buf<loc=scaling, dtype=f16, rows=16, cols=32, v_row=16, v_col=32, blayout=row_major, slayout=none_box, fractal=32, pad=0>
     %b_scale = pto.alloc_tile : !pto.tile_buf<loc=scaling, dtype=f16, rows=32, cols=32, v_row=32, v_col=32, blayout=row_major, slayout=none_box, fractal=32, pad=0>
+    %c_in = pto.alloc_tile : !pto.tile_buf<loc=acc, dtype=f32, rows=16, cols=32, v_row=16, v_col=32, blayout=col_major, slayout=row_major, fractal=512, pad=0>
     %dst0 = pto.alloc_tile : !pto.tile_buf<loc=acc, dtype=f32, rows=16, cols=32, v_row=16, v_col=32, blayout=col_major, slayout=row_major, fractal=512, pad=0>
     %dst1 = pto.alloc_tile : !pto.tile_buf<loc=acc, dtype=f32, rows=16, cols=32, v_row=16, v_col=32, blayout=col_major, slayout=row_major, fractal=512, pad=0>
     %dst2 = pto.alloc_tile : !pto.tile_buf<loc=acc, dtype=f32, rows=16, cols=32, v_row=16, v_col=32, blayout=col_major, slayout=row_major, fractal=512, pad=0>
+    %dst3 = pto.alloc_tile : !pto.tile_buf<loc=acc, dtype=f32, rows=16, cols=32, v_row=16, v_col=32, blayout=col_major, slayout=row_major, fractal=512, pad=0>
+    %dst4 = pto.alloc_tile : !pto.tile_buf<loc=acc, dtype=f32, rows=16, cols=32, v_row=16, v_col=32, blayout=col_major, slayout=row_major, fractal=512, pad=0>
+    %dst5 = pto.alloc_tile : !pto.tile_buf<loc=acc, dtype=f32, rows=16, cols=32, v_row=16, v_col=32, blayout=col_major, slayout=row_major, fractal=512, pad=0>
 
     pto.tmatmul.mx ins(%a, %a_scale, %b, %b_scale : !pto.tile_buf<loc=left, dtype=f8E4M3FN, rows=16, cols=32, v_row=16, v_col=32, blayout=col_major, slayout=row_major, fractal=512, pad=0>, !pto.tile_buf<loc=scaling, dtype=f16, rows=16, cols=32, v_row=16, v_col=32, blayout=row_major, slayout=none_box, fractal=32, pad=0>, !pto.tile_buf<loc=right, dtype=f8E4M3FN, rows=32, cols=32, v_row=32, v_col=32, blayout=row_major, slayout=col_major, fractal=512, pad=0>, !pto.tile_buf<loc=scaling, dtype=f16, rows=32, cols=32, v_row=32, v_col=32, blayout=row_major, slayout=none_box, fractal=32, pad=0>) outs(%dst0 : !pto.tile_buf<loc=acc, dtype=f32, rows=16, cols=32, v_row=16, v_col=32, blayout=col_major, slayout=row_major, fractal=512, pad=0>)
     pto.tmatmul.mx ins(%a, %a_scale, %b, %b_scale : !pto.tile_buf<loc=left, dtype=f8E4M3FN, rows=16, cols=32, v_row=16, v_col=32, blayout=col_major, slayout=row_major, fractal=512, pad=0>, !pto.tile_buf<loc=scaling, dtype=f16, rows=16, cols=32, v_row=16, v_col=32, blayout=row_major, slayout=none_box, fractal=32, pad=0>, !pto.tile_buf<loc=right, dtype=f8E4M3FN, rows=32, cols=32, v_row=32, v_col=32, blayout=row_major, slayout=col_major, fractal=512, pad=0>, !pto.tile_buf<loc=scaling, dtype=f16, rows=32, cols=32, v_row=32, v_col=32, blayout=row_major, slayout=none_box, fractal=32, pad=0>) outs(%dst1 : !pto.tile_buf<loc=acc, dtype=f32, rows=16, cols=32, v_row=16, v_col=32, blayout=col_major, slayout=row_major, fractal=512, pad=0>) {accPhase = #pto<acc_phase partial>}
     pto.tmatmul.mx ins(%a, %a_scale, %b, %b_scale : !pto.tile_buf<loc=left, dtype=f8E4M3FN, rows=16, cols=32, v_row=16, v_col=32, blayout=col_major, slayout=row_major, fractal=512, pad=0>, !pto.tile_buf<loc=scaling, dtype=f16, rows=16, cols=32, v_row=16, v_col=32, blayout=row_major, slayout=none_box, fractal=32, pad=0>, !pto.tile_buf<loc=right, dtype=f8E4M3FN, rows=32, cols=32, v_row=32, v_col=32, blayout=row_major, slayout=col_major, fractal=512, pad=0>, !pto.tile_buf<loc=scaling, dtype=f16, rows=32, cols=32, v_row=32, v_col=32, blayout=row_major, slayout=none_box, fractal=32, pad=0>) outs(%dst2 : !pto.tile_buf<loc=acc, dtype=f32, rows=16, cols=32, v_row=16, v_col=32, blayout=col_major, slayout=row_major, fractal=512, pad=0>) {accPhase = #pto<acc_phase final>}
+    pto.tmatmul.mx.acc ins(%c_in, %a, %a_scale, %b, %b_scale : !pto.tile_buf<loc=acc, dtype=f32, rows=16, cols=32, v_row=16, v_col=32, blayout=col_major, slayout=row_major, fractal=512, pad=0>, !pto.tile_buf<loc=left, dtype=f8E4M3FN, rows=16, cols=32, v_row=16, v_col=32, blayout=col_major, slayout=row_major, fractal=512, pad=0>, !pto.tile_buf<loc=scaling, dtype=f16, rows=16, cols=32, v_row=16, v_col=32, blayout=row_major, slayout=none_box, fractal=32, pad=0>, !pto.tile_buf<loc=right, dtype=f8E4M3FN, rows=32, cols=32, v_row=32, v_col=32, blayout=row_major, slayout=col_major, fractal=512, pad=0>, !pto.tile_buf<loc=scaling, dtype=f16, rows=32, cols=32, v_row=32, v_col=32, blayout=row_major, slayout=none_box, fractal=32, pad=0>) outs(%dst3 : !pto.tile_buf<loc=acc, dtype=f32, rows=16, cols=32, v_row=16, v_col=32, blayout=col_major, slayout=row_major, fractal=512, pad=0>)
+    pto.tmatmul.mx.acc ins(%c_in, %a, %a_scale, %b, %b_scale : !pto.tile_buf<loc=acc, dtype=f32, rows=16, cols=32, v_row=16, v_col=32, blayout=col_major, slayout=row_major, fractal=512, pad=0>, !pto.tile_buf<loc=left, dtype=f8E4M3FN, rows=16, cols=32, v_row=16, v_col=32, blayout=col_major, slayout=row_major, fractal=512, pad=0>, !pto.tile_buf<loc=scaling, dtype=f16, rows=16, cols=32, v_row=16, v_col=32, blayout=row_major, slayout=none_box, fractal=32, pad=0>, !pto.tile_buf<loc=right, dtype=f8E4M3FN, rows=32, cols=32, v_row=32, v_col=32, blayout=row_major, slayout=col_major, fractal=512, pad=0>, !pto.tile_buf<loc=scaling, dtype=f16, rows=32, cols=32, v_row=32, v_col=32, blayout=row_major, slayout=none_box, fractal=32, pad=0>) outs(%dst4 : !pto.tile_buf<loc=acc, dtype=f32, rows=16, cols=32, v_row=16, v_col=32, blayout=col_major, slayout=row_major, fractal=512, pad=0>) {accPhase = #pto<acc_phase partial>}
+    pto.tmatmul.mx.acc ins(%c_in, %a, %a_scale, %b, %b_scale : !pto.tile_buf<loc=acc, dtype=f32, rows=16, cols=32, v_row=16, v_col=32, blayout=col_major, slayout=row_major, fractal=512, pad=0>, !pto.tile_buf<loc=left, dtype=f8E4M3FN, rows=16, cols=32, v_row=16, v_col=32, blayout=col_major, slayout=row_major, fractal=512, pad=0>, !pto.tile_buf<loc=scaling, dtype=f16, rows=16, cols=32, v_row=16, v_col=32, blayout=row_major, slayout=none_box, fractal=32, pad=0>, !pto.tile_buf<loc=right, dtype=f8E4M3FN, rows=32, cols=32, v_row=32, v_col=32, blayout=row_major, slayout=col_major, fractal=512, pad=0>, !pto.tile_buf<loc=scaling, dtype=f16, rows=32, cols=32, v_row=32, v_col=32, blayout=row_major, slayout=none_box, fractal=32, pad=0>) outs(%dst5 : !pto.tile_buf<loc=acc, dtype=f32, rows=16, cols=32, v_row=16, v_col=32, blayout=col_major, slayout=row_major, fractal=512, pad=0>) {accPhase = #pto<acc_phase final>}
     return
   }
 }
@@ -19,3 +26,6 @@ module attributes {pto.target_arch = "a5"} {
 // A5: TMATMUL_MX([[D0:v[0-9]+]], [[A0:v[0-9]+]], [[AS0:v[0-9]+]], [[B0:v[0-9]+]], [[BS0:v[0-9]+]]);
 // A5: TMATMUL_MX<pto::AccPhase::Partial>([[D1:v[0-9]+]], [[A1:v[0-9]+]], [[AS1:v[0-9]+]], [[B1:v[0-9]+]], [[BS1:v[0-9]+]]);
 // A5: TMATMUL_MX<pto::AccPhase::Final>([[D2:v[0-9]+]], [[A2:v[0-9]+]], [[AS2:v[0-9]+]], [[B2:v[0-9]+]], [[BS2:v[0-9]+]]);
+// A5: TMATMUL_MX([[D3:v[0-9]+]], [[C3:v[0-9]+]], [[A3:v[0-9]+]], [[AS3:v[0-9]+]], [[B3:v[0-9]+]], [[BS3:v[0-9]+]]);
+// A5: TMATMUL_MX<pto::AccPhase::Partial>([[D4:v[0-9]+]], [[C4:v[0-9]+]], [[A4:v[0-9]+]], [[AS4:v[0-9]+]], [[B4:v[0-9]+]], [[BS4:v[0-9]+]]);
+// A5: TMATMUL_MX<pto::AccPhase::Final>([[D5:v[0-9]+]], [[C5:v[0-9]+]], [[A5:v[0-9]+]], [[AS5:v[0-9]+]], [[B5:v[0-9]+]], [[BS5:v[0-9]+]]);