Expose configurable frontend pipe slot_num

docs/PTO_IR_manual.md

@@ -8346,6 +8346,10 @@ frontend/framework generated IR. The detailed design document is:
   function.
 - `slot_size` is expressed in bytes and uses the pre-split logical pipe-entry
   size.
+- `slot_num` is an optional compile-time integer attribute on
+  `pto.aic_initialize_pipe` / `pto.aiv_initialize_pipe`. It controls the GM
+  ring FIFO depth and defaults to `8` for `dir_mask = 1/2` or `4` for
+  `dir_mask = 3`.
 - `local_slot_num` is an optional compile-time integer attribute on
   `pto.aic_initialize_pipe` / `pto.aiv_initialize_pipe`.
   On A2/A3 it overrides the default consumer-side local FIFO slot count only
@@ -8369,9 +8373,10 @@ frontend/framework generated IR. The detailed design document is:
   (`pto.initialize_l2g2l_pipe`). It does not implicitly execute `pto.tstore` or
   `pto.tload`; callers move data explicitly before `tpush` or after `tpop`.
 - When every transfer op bound to one pipe id uses a global entry, the pipe is
-  a global-only GM FIFO. Its frontend initialize op carries only
-  `gm_slot_tensor`; `gm_slot_buffer`, `c2v_consumer_buf`, `v2c_consumer_buf`, `local_slot_num`,
-  `pto.reserve_buffer`, and `pto.import_reserved_buffer` are not used.
+  a global-only GM FIFO. Its frontend initialize op carries `gm_slot_tensor`
+  and may carry `slot_num`; `gm_slot_buffer`, `c2v_consumer_buf`,
+  `v2c_consumer_buf`, `local_slot_num`, `pto.reserve_buffer`, and
+  `pto.import_reserved_buffer` are not used.
 - For global entries, the matched initialize op's `gm_slot_tensor` describes
   one FIFO slot entry, not the full multi-slot FIFO buffer. Its dtype, shape,
   stride, and layout must match the `tensor_view` returned by `talloc` /
@@ -8505,7 +8510,7 @@ this op.
 
 ```mlir
 // A2/A3 (with GM slot buffer):
-pto.aic_initialize_pipe {id = 0, dir_mask = 1, slot_size = 1024, local_slot_num = 1}
+pto.aic_initialize_pipe {id = 0, dir_mask = 1, slot_size = 1024, slot_num = 2, local_slot_num = 1}
   (gm_slot_buffer = %gm_buf : !pto.ptr<f32>,
    c2v_consumer_buf = %c2v_import : i32,
    v2c_consumer_buf = %c0_i32 : i32)
@@ -8529,6 +8534,8 @@ pto.aic_initialize_pipe {id = 0, dir_mask = 1, slot_size = 1024, nosplit = true}
   the same function
 - `dir_mask`: communication direction encoding
 - `slot_size`: logical slot size in bytes
+- `slot_num`: optional GM ring FIFO slot count; omitted defaults to `8` for
+  `dir_mask = 1/2` or `4` for `dir_mask = 3`
 - `local_slot_num`: optional A2/A3-only local FIFO slot count override for the
   lowered `pto.initialize_l2g2l_pipe`; omitted for global-only GM FIFO
 - `nosplit`: optional compile-time boolean controlling no-split pipe mode
@@ -8551,12 +8558,12 @@ pto.aic_initialize_pipe {id = 0, dir_mask = 1, slot_size = 1024, nosplit = true}
 - Must appear in Cube kernels
 - Multiple `pto.aic_initialize_pipe` ops are allowed in one Cube function, but
   `id` must be unique among frontend initialize ops in that function
+- If `slot_num` is present, it must be greater than `0`
 - If `local_slot_num` is present, it must be greater than `0` and no greater
-  than the legacy slot count implied by `dir_mask`
-  (`8` for `dir_mask = 1/2`, `4` for `dir_mask = 3`)
+  than the effective `slot_num`
 - A global-only GM FIFO initialize carries only `gm_slot_tensor`; it must not
   carry `gm_slot_buffer`, `local_slot_num`, `c2v_consumer_buf`, or
-  `v2c_consumer_buf`
+  `v2c_consumer_buf`; it may carry `slot_num`
 - For global-only GM FIFO, `slot_size` must match the byte size of
   `gm_slot_tensor`
 - Global-entry `talloc` / `tpush` / `tpop` / `tfree` entry types must match the
@@ -8576,7 +8583,7 @@ pto.aic_initialize_pipe {id = 0, dir_mask = 1, slot_size = 1024, nosplit = true}
 
 ```mlir
 // A2/A3 (with GM slot buffer):
-pto.aiv_initialize_pipe {id = 0, dir_mask = 1, slot_size = 1024, local_slot_num = 1}
+pto.aiv_initialize_pipe {id = 0, dir_mask = 1, slot_size = 1024, slot_num = 2, local_slot_num = 1}
   (gm_slot_buffer = %gm_buf : !pto.ptr<f32>,
    c2v_consumer_buf = %c2v_local : i32,
    v2c_consumer_buf = %c0_i32 : i32)

docs/designs/ptoas-tpush-tpop-design.md

@@ -390,7 +390,7 @@ func.func @vector_kernel(%gm_slot_buffer : !pto.ptr<f32>,
 - 启用 local address planning 的编译流程：`reserve_buffer` 只允许 `auto = true`
 - 跳过 local address planning 的编译流程：`reserve_buffer` 只允许 `auto = false` 且显式提供 `base`
 - `import_reserved_buffer` 必须能在 `peer_func` 中找到同名 `reserve_buffer`
-- global-only GM FIFO 的 initialize 只提供 `gm_slot_tensor`，不提供 `gm_slot_buffer`、`local_slot_num`、`c2v_consumer_buf`、`v2c_consumer_buf`，且不要求成对的 `reserve_buffer` / `import_reserved_buffer`
+- global-only GM FIFO 的 initialize 只提供 `gm_slot_tensor`（可附带 `slot_num`），不提供 `gm_slot_buffer`、`local_slot_num`、`c2v_consumer_buf`、`v2c_consumer_buf`，且不要求成对的 `reserve_buffer` / `import_reserved_buffer`
 
 ## 4. 核心约定
 
@@ -516,7 +516,7 @@ DIR_BOTH 示例：
   - 表示 GM 路径下 consumer 侧 local slot buffer 的槽数，仅在存在 local FIFO buffer 的 tile-entry 路径有意义
   - 仅在通过 GM 传递时对底层 `TPipe` 模板参数有意义，不改变 GM FIFO 的 `slot_num`
   - 存在 local FIFO buffer 且缺省时，默认值等于该内部 pipe 的 `slot_num`
-  - 因此当前固定规则下：
+  - 因此前端未显式指定 `slot_num` 时：
     - `DIR_MASK=1/2` 直接 lowering 时，`local_slot_num = 8`
     - `DIR_MASK=3` 单条 DIR_BOTH pipe，`local_slot_num = 4`
   - global-only GM FIFO 不携带 `local_slot_num`
@@ -658,10 +658,10 @@ pto.tfree(%entry, %pipe : !pto.tensor_view<128x512xf32>, !pto.pipe) {split = 0}
 #### A2/A3
 
 - `pto.aic_initialize_pipe` 和 `pto.aiv_initialize_pipe` lower 为 `pto.initialize_l2g2l_pipe`
-- 若前端 init 只提供 `gm_slot_tensor`，则 lower 为只携带 `gm_slot_tensor` 的 global-only GM FIFO；不补 `local_slot_num`，不生成 local consumer address operand，也不依赖 `reserve_buffer` / `import_reserved_buffer`
+- 若前端 init 只提供 `gm_slot_tensor`（可附带 `slot_num`），则 lower 为只携带 `gm_slot_tensor` 的 global-only GM FIFO；不补 `local_slot_num`，不生成 local consumer address operand，也不依赖 `reserve_buffer` / `import_reserved_buffer`
 - 若前端提供了 consumer 侧 local FIFO buffer，且提供了 `local_slot_num`，则直接转发到 lowered
   `pto.initialize_l2g2l_pipe`
-- 若前端提供了 consumer 侧 local FIFO buffer 但未提供更具体信息，lowering 默认补上 `local_slot_num = slot_num`
+- 若前端提供了 consumer 侧 local FIFO buffer 但未提供 `local_slot_num`，lowering 默认补上 `local_slot_num = slot_num`
 
 #### A5
 
@@ -670,17 +670,17 @@ pto.tfree(%entry, %pipe : !pto.tensor_view<128x512xf32>, !pto.pipe) {split = 0}
 ### 6.2 `DIR_MASK=1/2`
 
 - 只生成一条内部 pipe
-- `slot_num = 8`
-- 对带 consumer 侧 local FIFO buffer 的 `initialize_l2g2l_pipe`，默认 `local_slot_num = 8`
+- `slot_num` 缺省为 `8`，也可由前端显式指定
+- 对带 consumer 侧 local FIFO buffer 的 `initialize_l2g2l_pipe`，默认 `local_slot_num = slot_num`
 - 若前端显式提供 `local_slot_num`，则使用显式值
 - global-only GM FIFO 不携带 `local_slot_num`，地址/descriptor 操作数只有 `gm_slot_tensor`
 
 ### 6.3 `DIR_MASK=3`
 
 前端一个 init op 生成**单条** DIR_BOTH 内部 pipe：
 
-- `%pipe`：`dir_mask = 3`，`slot_num = 4`
-- 若 lowering 为带 consumer 侧 local FIFO buffer 的 `initialize_l2g2l_pipe`，默认 `local_slot_num = 4`
+- `%pipe`：`dir_mask = 3`，`slot_num` 缺省为 `4`，也可由前端显式指定
+- 若 lowering 为带 consumer 侧 local FIFO buffer 的 `initialize_l2g2l_pipe`，默认 `local_slot_num = slot_num`
 - 若前端显式提供 `local_slot_num`，则使用显式值
 
 地址选择规则：
@@ -995,7 +995,7 @@ pass 在模块级按两步执行：
 - 方向相关 op 只能出现在合法 kernel 中
 - 前端数据传输 op 的 `split` 必须是合法的编译期常量属性
 - `global` entry 形式的 `talloc_to_*` / `tpush_to_*` / `tpop_from_*` / `tfree_from_*` 只能绑定到 GM FIFO pipe（A2/A3 `initialize_l2g2l_pipe` 路径）
-- 绑定到 global-only GM FIFO 的 initialize 只允许携带 `gm_slot_tensor`，不得携带 `gm_slot_buffer`、`local_slot_num`、`c2v_consumer_buf`、`v2c_consumer_buf`；该路径不要求 `reserve_buffer` / `import_reserved_buffer`
+- 绑定到 global-only GM FIFO 的 initialize 只允许携带 `gm_slot_tensor`（可附带 `slot_num`），不得携带 `gm_slot_buffer`、`local_slot_num`、`c2v_consumer_buf`、`v2c_consumer_buf`；该路径不要求 `reserve_buffer` / `import_reserved_buffer`
 - `gm_slot_tensor` 本身描述单个 slot entry；其字节数必须匹配 `slot_size`
 - `talloc_to_*` / `tpop_from_*` 返回的 `tensor_view` 类型必须匹配 `gm_slot_tensor`
 - `global` entry 的 dtype、shape 与 stride/layout 必须足以生成底层 `GlobalTensor<RawDType, Shape, Stride, Layout>` 类型
@@ -1008,11 +1008,12 @@ pass 在模块级按两步执行：
 内部 verifier 负责检查：
 
 - `slot_size > 0`
-- `slot_num` 只允许 `8` 或 `4`
-- `DIR_MASK=1/2` 时，`slot_num` 必须与单向/双向 lowering 规则一致
+- `slot_num >= 1`
+- legacy 前端 `pto.aic_initialize_pipe` / `pto.aiv_initialize_pipe` 可显式提供
+  `slot_num`；缺省时 `DIR_MASK=1/2` 使用 `8`，`DIR_MASK=3` 使用 `4`
 - `local_slot_num` 若出现，可出现在 `pto.initialize_l2g2l_pipe` 或 legacy 前端
   `pto.aic_initialize_pipe` / `pto.aiv_initialize_pipe` 上，且必须大于 `0`
-  且不大于其对应 lowering 规则下的 `slot_num`；global-only GM FIFO 不携带 `local_slot_num`
+  且不大于其有效 `slot_num`；global-only GM FIFO 不携带 `local_slot_num`
 - `flag_base` 若出现，必须满足基本合法性；是否已填写以及具体分配值由 flag 分配保证
 - `pto.initialize_l2g2l_pipe` 必须提供 `gm_addr` 或 `gm_slot_tensor`；只有存在 consumer 侧 local FIFO buffer 时才提供 `local_addr` / `peer_local_addr`
 - `pto.initialize_l2l_pipe` 必须提供 `local_addr`

include/PTO/IR/PTOOps.td

@@ -1505,6 +1505,7 @@ def AicInitializePipeOp : PTO_Op<"aic_initialize_pipe",
       DefaultValuedOptionalAttr<I32Attr, "0">:$id,
       I8Attr:$dir_mask,
       I32Attr:$slot_size,
+      OptionalAttr<I32Attr>:$slot_num,
       OptionalAttr<I32Attr>:$local_slot_num,
       OptionalAttr<BoolAttr>:$nosplit,
       Optional<PtrType>:$gm_slot_buffer,
@@ -1526,6 +1527,7 @@ def AivInitializePipeOp : PTO_Op<"aiv_initialize_pipe",
       DefaultValuedOptionalAttr<I32Attr, "0">:$id,
       I8Attr:$dir_mask,
       I32Attr:$slot_size,
+      OptionalAttr<I32Attr>:$slot_num,
       OptionalAttr<I32Attr>:$local_slot_num,
       OptionalAttr<BoolAttr>:$nosplit,
       Optional<PtrType>:$gm_slot_buffer,

lib/PTO/IR/PTO.cpp

@@ -11457,6 +11457,7 @@ static ParseResult parseFrontendInitializePipeOp(OpAsmParser &parser,
   bool sawId = false;
   bool sawDirMask = false;
   bool sawSlotSize = false;
+  bool sawSlotNum = false;
   bool sawLocalSlotNum = false;
   bool sawNoSplit = false;
 
@@ -11495,6 +11496,15 @@ static ParseResult parseFrontendInitializePipeOp(OpAsmParser &parser,
                                 "slot_size", attrs))
         return failure();
       sawSlotSize = true;
+    } else if (keyword == "slot_num") {
+      if (sawSlotNum)
+        return parser.emitError(parser.getCurrentLocation(),
+                                "duplicate 'slot_num' clause");
+      IntegerAttr slotNumAttr;
+      if (parser.parseAttribute(slotNumAttr, parser.getBuilder().getI32Type(),
+                                "slot_num", attrs))
+        return failure();
+      sawSlotNum = true;
     } else if (keyword == "local_slot_num") {
       if (sawLocalSlotNum)
         return parser.emitError(parser.getCurrentLocation(),
@@ -11632,6 +11642,8 @@ static void printFrontendInitializePipeOp(InitOpT op, OpAsmPrinter &p) {
     printClause("id", op.getId());
   printClause("dir_mask", static_cast<int32_t>(op.getDirMask()));
   printClause("slot_size", op.getSlotSize());
+  if (auto slotNumAttr = op.getSlotNumAttr())
+    printClause("slot_num", slotNumAttr.getInt());
   if (auto localSlotNumAttr = op.getLocalSlotNumAttr())
     printClause("local_slot_num", localSlotNumAttr.getInt());
   if (auto noSplitAttr = op.getNosplitAttr())
@@ -11658,7 +11670,8 @@ static void printFrontendInitializePipeOp(InitOpT op, OpAsmPrinter &p) {
   p << ")";
   p.printOptionalAttrDict(
       op->getAttrs(),
-      /*elidedAttrs=*/{"id", "dir_mask", "slot_size", "local_slot_num",
+      /*elidedAttrs=*/{"id", "dir_mask", "slot_size", "slot_num",
+                       "local_slot_num",
                        "nosplit", "operandSegmentSizes"});
 }
 
@@ -11744,6 +11757,12 @@ static LogicalResult verifyFrontendInitCommon(InitOpT op,
     return op.emitOpError("expects 'dir_mask' to be 1, 2, or 3");
   if (op.getSlotSize() <= 0)
     return op.emitOpError("expects 'slot_size' to be greater than 0");
+  int32_t slotNum = dirMask == 3 ? 4 : 8;
+  if (auto slotNumAttr = op.getSlotNumAttr()) {
+    slotNum = slotNumAttr.getInt();
+    if (slotNum <= 0)
+      return op.emitOpError("expects 'slot_num' to be greater than 0");
+  }
 
   bool hasGlobalSlotTensor = static_cast<bool>(op.getGmSlotTensor());
   bool hasC2vConsumerBuf = static_cast<bool>(op.getC2vConsumerBuf());
@@ -11779,11 +11798,10 @@ static LogicalResult verifyFrontendInitCommon(InitOpT op,
     int32_t localSlotNum = localSlotNumAttr.getInt();
     if (localSlotNum <= 0)
       return op.emitOpError("expects 'local_slot_num' to be greater than 0");
-    int32_t loweredSlotNum = dirMask == 3 ? 4 : 8;
-    if (localSlotNum > loweredSlotNum) {
+    if (localSlotNum > slotNum) {
       return op.emitOpError()
-             << "expects 'local_slot_num' to be less than or equal to "
-             << loweredSlotNum << " for dir_mask = " << static_cast<int>(dirMask);
+             << "expects 'local_slot_num' to be less than or equal to slot_num ("
+             << slotNum << ") for dir_mask = " << static_cast<int>(dirMask);
     }
   }
 
@@ -12060,8 +12078,8 @@ static LogicalResult verifyPipeShape(Operation *op, int8_t dirMask, int32_t slot
     return op->emitOpError("expects 'dir_mask' to be 1, 2, or 3");
   if (slotSize <= 0)
     return op->emitOpError("expects 'slot_size' to be greater than 0");
-  if (slotNum != 4 && slotNum != 8)
-    return op->emitOpError("expects 'slot_num' to be 4 or 8");
+  if (slotNum <= 0)
+    return op->emitOpError("expects 'slot_num' to be greater than 0");
   if (flagBase && *flagBase < 0)
     return op->emitOpError("expects 'flag_base' to be non-negative when present");
   if (flagBase) {

lib/PTO/Transforms/PTOLowerFrontendPipeOpsPass.cpp

@@ -65,6 +65,15 @@ static void propagateFrontendIdAttr(InitOpT initOp, Operation *pipeOp,
                   rewriter.getI32IntegerAttr(initOp.getId()));
 }
 
+template <typename InitOpT>
+static int32_t getFrontendSlotNum(InitOpT initOp) {
+  if (auto slotNumAttr = initOp.getSlotNumAttr())
+    return slotNumAttr.getInt();
+  return initOp.getDirMask() == kBidirectionalDirMask
+             ? kBidirectionalSlotNum
+             : kSingleDirectionSlotNum;
+}
+
 static std::optional<int64_t> getStaticIndexLikeValue(Value value) {
   if (auto cst = value.getDefiningOp<arith::ConstantIndexOp>())
     return cst.value();
@@ -166,9 +175,10 @@ static FailureOr<FrontendPipeHandles>
 lowerSingleDirectionFrontendInit(InitOpT initOp, IRRewriter &rewriter,
                                  PTOArch arch, Type pipeTy, int8_t dirMask,
                                  Value localAddr) {
+  int32_t slotNum = getFrontendSlotNum(initOp);
   auto pipeOr =
-      createFrontendPipe(initOp, rewriter, arch, pipeTy, dirMask,
-                         kSingleDirectionSlotNum, localAddr);
+      createFrontendPipe(initOp, rewriter, arch, pipeTy, dirMask, slotNum,
+                         localAddr);
   if (failed(pipeOr))
     return failure();
 
@@ -190,9 +200,9 @@ template <typename InitOpT>
 static FailureOr<FrontendPipeHandles>
 lowerBidirectionalFrontendInit(InitOpT initOp, IRRewriter &rewriter,
                                PTOArch arch, Type pipeTy) {
+  int32_t slotNum = getFrontendSlotNum(initOp);
   auto pipeOr = createFrontendPipe(initOp, rewriter, arch, pipeTy,
-                                   kBidirectionalDirMask,
-                                   kBidirectionalSlotNum,
+                                   kBidirectionalDirMask, slotNum,
                                    initOp.getC2vConsumerBuf(),
                                    initOp.getV2cConsumerBuf());
   if (failed(pipeOr))

test/lit/pto/tpush_tpop_frontend_local_slot_num_invalid.pto

@@ -12,4 +12,4 @@ module {
   }
 }
 
-// CHECK: error: 'pto.aic_initialize_pipe' op expects 'local_slot_num' to be less than or equal to 4 for dir_mask = 3
+// CHECK: error: 'pto.aic_initialize_pipe' op expects 'local_slot_num' to be less than or equal to slot_num (4) for dir_mask = 3

test/lit/pto/tpush_tpop_frontend_slot_num_a3.pto

@@ -0,0 +1,49 @@
+// RUN: ptoas --pto-arch=a3 %s 2>&1 | FileCheck %s --check-prefix=A3
+
+module {
+  func.func @cube_kernel(%gm_slot_buffer: !pto.ptr<f32>)
+      attributes {pto.kernel_kind = #pto.kernel_kind<cube>} {
+    %c0_i32 = arith.constant 0 : i32
+    %v2c_local = pto.reserve_buffer {
+      name = "v2c_fifo",
+      size = 2048,
+      location = #pto.address_space<mat>,
+      auto = true
+    } -> i32
+    pto.aic_initialize_pipe {id = 0, dir_mask = 2, slot_size = 1024, slot_num = 2}
+      (gm_slot_buffer = %gm_slot_buffer : !pto.ptr<f32>,
+       c2v_consumer_buf = %c0_i32 : i32,
+       v2c_consumer_buf = %v2c_local : i32)
+
+    %recv_tile = pto.tpop_from_aiv {id = 0, split = 0}
+      -> !pto.tile_buf<loc=mat, dtype=f32, rows=16, cols=16, v_row=16, v_col=16, blayout=col_major, slayout=row_major, fractal=512, pad=0>
+    pto.tfree_from_aiv {id = 0, split = 0}
+    return
+  }
+
+  func.func @vector_kernel(%gm_slot_buffer: !pto.ptr<f32>)
+      attributes {pto.kernel_kind = #pto.kernel_kind<vector>} {
+    %c0_i32 = arith.constant 0 : i32
+    %v2c_import = pto.import_reserved_buffer {
+      name = "v2c_fifo",
+      peer_func = @cube_kernel
+    } -> i32
+    pto.aiv_initialize_pipe {id = 0, dir_mask = 2, slot_size = 1024, slot_num = 2}
+      (gm_slot_buffer = %gm_slot_buffer : !pto.ptr<f32>,
+       c2v_consumer_buf = %c0_i32 : i32,
+       v2c_consumer_buf = %v2c_import : i32)
+
+    %vec_tile = pto.alloc_tile : !pto.tile_buf<loc=vec, dtype=f32, rows=16, cols=16, v_row=16, v_col=16, blayout=row_major, slayout=none_box, fractal=512, pad=0>
+    pto.tpush_to_aic(%vec_tile : !pto.tile_buf<loc=vec, dtype=f32, rows=16, cols=16, v_row=16, v_col=16, blayout=row_major, slayout=none_box, fractal=512, pad=0>) {id = 0, split = 0}
+    return
+  }
+}
+
+// A3-LABEL: AICORE void cube_kernel(__gm__ float*
+// A3: auto {{v[0-9]+}} = TPipe<0, Direction::DIR_V2C, 1024, 2, 2, true>(
+// A3: TPOP<TPipe<0, Direction::DIR_V2C, 1024, 2, 2, true>
+// A3: TFREE<TPipe<0, Direction::DIR_V2C, 1024, 2, 2, true>, TileSplitAxis::TILE_NO_SPLIT>(
+
+// A3-LABEL: AICORE void vector_kernel(__gm__ float*
+// A3: auto {{v[0-9]+}} = TPipe<0, Direction::DIR_V2C, 1024, 2, 2, true>(
+// A3: TPUSH<TPipe<0, Direction::DIR_V2C, 1024, 2, 2, true>

test/lit/pto/tpush_tpop_frontend_slot_num_invalid.pto

@@ -0,0 +1,15 @@
+// RUN: not ptoas --pto-arch=a3 %s 2>&1 | FileCheck %s
+
+module {
+  func.func @cube_kernel(%gm_slot_buffer: !pto.ptr<f32>)
+      attributes {pto.kernel_kind = #pto.kernel_kind<cube>} {
+    %c0_i32 = arith.constant 0 : i32
+    pto.aic_initialize_pipe {id = 0, dir_mask = 1, slot_size = 1024, slot_num = 0}
+      (gm_slot_buffer = %gm_slot_buffer : !pto.ptr<f32>,
+       c2v_consumer_buf = %c0_i32 : i32,
+       v2c_consumer_buf = %c0_i32 : i32)
+    return
+  }
+}
+
+// CHECK: error: 'pto.aic_initialize_pipe' op expects 'slot_num' to be greater than 0

test/lit/pto/tpush_tpop_frontend_slot_num_local_invalid.pto

@@ -0,0 +1,15 @@
+// RUN: not ptoas --pto-arch=a3 %s 2>&1 | FileCheck %s
+
+module {
+  func.func @cube_kernel(%gm_slot_buffer: !pto.ptr<f32>)
+      attributes {pto.kernel_kind = #pto.kernel_kind<cube>} {
+    %c0_i32 = arith.constant 0 : i32
+    pto.aic_initialize_pipe {id = 0, dir_mask = 2, slot_size = 1024, slot_num = 2, local_slot_num = 3}
+      (gm_slot_buffer = %gm_slot_buffer : !pto.ptr<f32>,
+       c2v_consumer_buf = %c0_i32 : i32,
+       v2c_consumer_buf = %c0_i32 : i32)
+    return
+  }
+}
+
+// CHECK: error: 'pto.aic_initialize_pipe' op expects 'local_slot_num' to be less than or equal to slot_num (2) for dir_mask = 2