Merge pull request #1069 from InfiniTensor/issue/1031_T1_1_15

【算子比赛2025秋】T1-1-15

src/infiniop/ops/reciprocal/operator.cc

+#include "../../operator.h"
+#include "../../handle.h"
+#include "infiniop/ops/reciprocal.h"
+
+#ifdef ENABLE_CPU_API
+#include "cpu/reciprocal_cpu.h"
+#endif
+#if defined(ENABLE_NVIDIA_API) || defined(ENABLE_ILUVATAR_API) || defined(ENABLE_QY_API)
+#include "nvidia/reciprocal_nvidia.cuh"
+#endif
+#ifdef ENABLE_METAX_API
+#include "metax/reciprocal_metax.h"
+#endif
+#ifdef ENABLE_KUNLUN_API
+#include "kunlun/reciprocal_kunlun.h"
+#endif
+#ifdef ENABLE_CAMBRICON_API
+#include "bang/reciprocal_bang.h"
+#endif
+#ifdef ENABLE_MOORE_API
+#include "moore/reciprocal_moore.h"
+#endif
+
+__INFINI_C infiniStatus_t infiniopCreateReciprocalDescriptor(
+    infiniopHandle_t handle,
+    infiniopReciprocalDescriptor_t *desc_ptr,
+    infiniopTensorDescriptor_t y_desc,
+    infiniopTensorDescriptor_t x_desc) {
+
+#define CREATE(CASE, NAMESPACE)                                                   \
+    case CASE:                                                                    \
+        return op::reciprocal::NAMESPACE::Descriptor::create(                     \
+            handle,                                                               \
+            reinterpret_cast<op::reciprocal::NAMESPACE::Descriptor **>(desc_ptr), \
+            y_desc,                                                               \
+            {x_desc})
+
+    switch (handle->device) {
+
+#ifdef ENABLE_CPU_API
+        CREATE(INFINI_DEVICE_CPU, cpu);
+#endif
+#ifdef ENABLE_NVIDIA_API
+        CREATE(INFINI_DEVICE_NVIDIA, nvidia);
+#endif
+#ifdef ENABLE_ILUVATAR_API
+        CREATE(INFINI_DEVICE_ILUVATAR, nvidia);
+#endif
+#ifdef ENABLE_QY_API
+        CREATE(INFINI_DEVICE_QY, nvidia);
+#endif
+#ifdef ENABLE_METAX_API
+        CREATE(INFINI_DEVICE_METAX, metax);
+#endif
+#ifdef ENABLE_KUNLUN_API
+        CREATE(INFINI_DEVICE_KUNLUN, kunlun);
+#endif
+#ifdef ENABLE_CAMBRICON_API
+        CREATE(INFINI_DEVICE_CAMBRICON, bang);
+#endif
+#ifdef ENABLE_MOORE_API
+        CREATE(INFINI_DEVICE_MOORE, moore);
+#endif
+
+    default:
+        return INFINI_STATUS_DEVICE_TYPE_NOT_SUPPORTED;
+    }
+
+#undef CREATE
+}
+
+__INFINI_C infiniStatus_t infiniopGetReciprocalWorkspaceSize(infiniopReciprocalDescriptor_t desc, size_t *size) {
+
+#define GET(CASE, NAMESPACE)                                                                      \
+    case CASE:                                                                                    \
+        *size = reinterpret_cast<op::reciprocal::NAMESPACE::Descriptor *>(desc)->workspaceSize(); \
+        return INFINI_STATUS_SUCCESS
+
+    switch (desc->device_type) {
+#ifdef ENABLE_CPU_API
+        GET(INFINI_DEVICE_CPU, cpu);
+#endif
+#ifdef ENABLE_NVIDIA_API
+        GET(INFINI_DEVICE_NVIDIA, nvidia);
+#endif
+#ifdef ENABLE_ILUVATAR_API
+        GET(INFINI_DEVICE_ILUVATAR, nvidia);
+#endif
+#ifdef ENABLE_QY_API
+        GET(INFINI_DEVICE_QY, nvidia);
+#endif
+#ifdef ENABLE_METAX_API
+        GET(INFINI_DEVICE_METAX, metax);
+#endif
+#ifdef ENABLE_KUNLUN_API
+        GET(INFINI_DEVICE_KUNLUN, kunlun);
+#endif
+#ifdef ENABLE_CAMBRICON_API
+        GET(INFINI_DEVICE_CAMBRICON, bang);
+#endif
+#ifdef ENABLE_MOORE_API
+        GET(INFINI_DEVICE_MOORE, moore);
+#endif
+    default:
+        return INFINI_STATUS_DEVICE_TYPE_NOT_SUPPORTED;
+    }
+#undef GET
+
+    return INFINI_STATUS_DEVICE_TYPE_NOT_SUPPORTED;
+}
+
+__INFINI_C infiniStatus_t infiniopReciprocal(
+    infiniopReciprocalDescriptor_t desc,
+    void *workspace,
+    size_t workspace_size,
+    void *y,
+    const void *x,
+    void *stream) {
+
+#define CALCULATE(CASE, NAMESPACE)                                                   \
+    case CASE:                                                                       \
+        return reinterpret_cast<const op::reciprocal::NAMESPACE::Descriptor *>(desc) \
+            ->calculate(workspace, workspace_size, y, {x}, stream)
+
+    switch (desc->device_type) {
+
+#ifdef ENABLE_CPU_API
+        CALCULATE(INFINI_DEVICE_CPU, cpu);
+#endif
+#ifdef ENABLE_NVIDIA_API
+        CALCULATE(INFINI_DEVICE_NVIDIA, nvidia);
+#endif
+#ifdef ENABLE_ILUVATAR_API
+        CALCULATE(INFINI_DEVICE_ILUVATAR, nvidia);
+#endif
+#ifdef ENABLE_QY_API
+        CALCULATE(INFINI_DEVICE_QY, nvidia);
+#endif
+#ifdef ENABLE_METAX_API
+        CALCULATE(INFINI_DEVICE_METAX, metax);
+#endif
+#ifdef ENABLE_KUNLUN_API
+        CALCULATE(INFINI_DEVICE_KUNLUN, kunlun);
+#endif
+#ifdef ENABLE_CAMBRICON_API
+        CALCULATE(INFINI_DEVICE_CAMBRICON, bang);
+#endif
+#ifdef ENABLE_MOORE_API
+        CALCULATE(INFINI_DEVICE_MOORE, moore);
+#endif
+
+    default:
+        return INFINI_STATUS_DEVICE_TYPE_NOT_SUPPORTED;
+    }
+
+#undef CALCULATE
+}
+
+__INFINI_C infiniStatus_t
+infiniopDestroyReciprocalDescriptor(infiniopReciprocalDescriptor_t desc) {
+
+#define DELETE(CASE, NAMESPACE)                                                       \
+    case CASE:                                                                        \
+        delete reinterpret_cast<const op::reciprocal::NAMESPACE::Descriptor *>(desc); \
+        return INFINI_STATUS_SUCCESS
+
+    switch (desc->device_type) {
+
+#ifdef ENABLE_CPU_API
+        DELETE(INFINI_DEVICE_CPU, cpu);
+#endif
+#ifdef ENABLE_NVIDIA_API
+        DELETE(INFINI_DEVICE_NVIDIA, nvidia);
+#endif
+#ifdef ENABLE_ILUVATAR_API
+        DELETE(INFINI_DEVICE_ILUVATAR, nvidia);
+#endif
+#ifdef ENABLE_QY_API
+        DELETE(INFINI_DEVICE_QY, nvidia);
+#endif
+#ifdef ENABLE_METAX_API
+        DELETE(INFINI_DEVICE_METAX, metax);
+#endif
+#ifdef ENABLE_KUNLUN_API
+        DELETE(INFINI_DEVICE_KUNLUN, kunlun);
+#endif
+#ifdef ENABLE_CAMBRICON_API
+        DELETE(INFINI_DEVICE_CAMBRICON, bang);
+#endif
+#ifdef ENABLE_MOORE_API
+        DELETE(INFINI_DEVICE_MOORE, moore);
+#endif
+
+    default:
+        return INFINI_STATUS_DEVICE_TYPE_NOT_SUPPORTED;
+    }
+
+#undef DELETE
+}

test/infinicore/ops/addcmul.py

@@ -119,9 +119,9 @@ class OpTest(BaseOperatorTest):
     def torch_operator(self, *args, **kwargs):
         return torch.addcmul(*args, **kwargs)
 
-    # def infinicore_operator(self, *args, **kwargs):
-    #     """InfiniCore implementation (operator not yet available)."""
-    #     return infinicore.addcmul(*args, **kwargs)
+    def infinicore_operator(self, *args, **kwargs):
+        """InfiniCore implementation (operator not yet available)."""
+        return infinicore.addcmul(*args, **kwargs)
 
 
 def main():

test/infinicore/ops/atanh.py

@@ -97,9 +97,9 @@ class OpTest(BaseOperatorTest):
     def torch_operator(self, *args, **kwargs):
         return torch.atanh(*args, **kwargs)
 
-    # def infinicore_operator(self, *args, **kwargs):
-    #     """InfiniCore implementation (operator not yet available)."""
-    #     return infinicore.atanh(*args, **kwargs)
+    def infinicore_operator(self, *args, **kwargs):
+        """InfiniCore atanh implementation"""
+        return infinicore.atanh(*args, **kwargs)
 
 
 def main():

test/infinicore/ops/binary_cross_entropy_with_logits.py

@@ -80,9 +80,9 @@ class OpTest(BaseOperatorTest):
     def torch_operator(self, *args, **kwargs):
         return torch.nn.functional.binary_cross_entropy_with_logits(*args, **kwargs)
 
-    # def infinicore_operator(self, *args, **kwargs):
-    #     """InfiniCore implementation (operator not yet available)."""
-    #     return infinicore.nn.functional.binary_cross_entropy_with_logits(*args, **kwargs)
+    def infinicore_operator(self, *args, **kwargs):
+        """InfiniCore implementation (operator not yet available)."""
+        return infinicore.nn.functional.binary_cross_entropy_with_logits(*args, **kwargs)
 
 
 def main():

test/infinicore/ops/cdist.py

@@ -63,9 +63,9 @@ class OpTest(BaseOperatorTest):
     def torch_operator(self, *args, **kwargs):
         return torch.cdist(*args, **kwargs)
 
-    # def infinicore_operator(self, *args, **kwargs):
-    #     """InfiniCore implementation (operator not yet available)."""
-    #     return infinicore.cdist(*args, **kwargs)
+    def infinicore_operator(self, *args, **kwargs):
+        """InfiniCore implementation (operator not yet available)."""
+        return infinicore.cdist(*args, **kwargs)
 
 
 def main():

test/infinicore/ops/reciprocal.py

@@ -89,9 +89,9 @@ class OpTest(BaseOperatorTest):
     def torch_operator(self, *args, **kwargs):
         return torch.reciprocal(*args, **kwargs)
 
-    # def infinicore_operator(self, *args, **kwargs):
-    #     """InfiniCore implementation (operator not yet available)."""
-    #     return infinicore.reciprocal(*args, **kwargs)
+    def infinicore_operator(self, *args, **kwargs):
+        """InfiniCore implementation (operator not yet available)."""
+        return infinicore.reciprocal(*args, **kwargs)
 
 
 def main():

test/infiniop/addcmul.py

+import torch
+import ctypes
+from ctypes import c_uint64, c_float
+from libinfiniop import (
+    LIBINFINIOP,
+    TestTensor,
+    get_test_devices,
+    check_error,
+    test_operator,
+    get_args,
+    debug,
+    get_tolerance,
+    profile_operation,
+    TestWorkspace,
+    InfiniDtype,
+    InfiniDtypeNames,
+    InfiniDeviceNames,
+    infiniopOperatorDescriptor_t,
+)
+from enum import Enum, auto
+
+# ==============================================================================
+# Configuration
+# ==============================================================================
+_TEST_CASES_ = [
+    # shape, input_stride, t1_stride, t2_stride
+    ((3, 3), None, None, None),
+    ((32, 512), None, None, None),
+    ((32, 512), (1024, 1), (1024, 1), (1024, 1)),
+    ((16, 32, 64), None, None, None),
+    ((8, 1, 1024), None, None, None), # 包含广播形状的潜在测试
+]
+
+_TENSOR_DTYPES = [InfiniDtype.F16, InfiniDtype.BF16, InfiniDtype.F32]
+
+_TOLERANCE_MAP = {
+    InfiniDtype.F16: {"atol": 1e-3, "rtol": 1e-2},
+    InfiniDtype.BF16: {"atol": 5e-3, "rtol": 5e-2},
+    InfiniDtype.F32: {"atol": 1e-5, "rtol": 1e-5},
+}
+
+class Inplace(Enum):
+    OUT_OF_PLACE = auto()
+    INPLACE_INPUT = auto()
+
+_INPLACE = [Inplace.OUT_OF_PLACE, Inplace.INPLACE_INPUT]
+_VALUES = [1.0, 0.5, -2.0] # 测试不同的 value 系数
+
+_TEST_CASES = [
+    test_case + (inplace_item, value)
+    for test_case in _TEST_CASES_
+    for inplace_item in _INPLACE
+    for value in _VALUES
+]
+
+DEBUG = False
+PROFILE = False
+NUM_PRERUN = 10
+NUM_ITERATIONS = 100
+
+def test(
+    handle,
+    device,
+    shape,
+    input_stride=None,
+    t1_stride=None,
+    t2_stride=None,
+    inplace=Inplace.OUT_OF_PLACE,
+    value=1.0,
+    dtype=InfiniDtype.F16,
+    sync=None,
+):
+    print(
+        f"Testing Addcmul on {InfiniDeviceNames[device]} with shape:{shape} value:{value} dtype:{InfiniDtypeNames[dtype]} inplace:{inplace}"
+    )
+
+    # 准备输入 Tensor
+    input_tensor = TestTensor(shape, input_stride, dtype, device)
+    t1 = TestTensor(shape, t1_stride, dtype, device)
+    t2 = TestTensor(shape, t2_stride, dtype, device)
+
+    # 使用 PyTorch 计算参考答案
+    # out = input + value * t1 * t2
+    ans = torch.addcmul(input_tensor.torch_tensor(), t1.torch_tensor(), t2.torch_tensor(), value=value)
+
+    if inplace == Inplace.INPLACE_INPUT:
+        out = input_tensor
+    else:
+        out = TestTensor(shape, None, dtype, device)
+
+    if sync is not None:
+        sync()
+
+    descriptor = infiniopOperatorDescriptor_t()
+    # 注意：根据之前的定义，Create 接口接收 value
+    check_error(
+        LIBINFINIOP.infiniopCreateAddcmulDescriptor(
+            handle, 
+            ctypes.byref(descriptor), 
+            out.descriptor, 
+            input_tensor.descriptor, 
+            t1.descriptor, 
+            t2.descriptor, 
+            c_float(value)
+        )
+    )
+
+    # 销毁临时描述符以防内核错误引用
+    for t in [input_tensor, t1, t2, out]:
+        t.destroy_desc()
+
+    workspace_size = c_uint64(0)
+    check_error(
+        LIBINFINIOP.infiniopGetAddcmulWorkspaceSize(
+            descriptor, ctypes.byref(workspace_size)
+        )
+    )
+    workspace = TestWorkspace(workspace_size.value, input_tensor.device)
+
+    def lib_addcmul():
+        check_error(
+            LIBINFINIOP.infiniopAddcmul(
+                descriptor,
+                workspace.data(),
+                workspace_size.value,
+                out.data(),
+                input_tensor.data(),
+                t1.data(),
+                t2.data(),
+                None,
+            )
+        )
+
+    lib_addcmul()
+
+    if sync is not None:
+        sync()
+
+    atol, rtol = get_tolerance(_TOLERANCE_MAP, dtype)
+    if DEBUG:
+        debug(out.actual_tensor(), ans, atol=atol, rtol=rtol)
+    
+    assert torch.allclose(out.actual_tensor(), ans, atol=atol, rtol=rtol)
+
+    if PROFILE:
+        profile_operation("PyTorch", lambda: torch.addcmul(input_tensor.torch_tensor(), t1.torch_tensor(), t2.torch_tensor(), value=value), device, NUM_PRERUN, NUM_ITERATIONS)
+        profile_operation("    lib", lambda: lib_addcmul(), device, NUM_PRERUN, NUM_ITERATIONS)
+
+    check_error(LIBINFINIOP.infiniopDestroyAddcmulDescriptor(descriptor))
+
+
+if __name__ == "__main__":
+    args = get_args()
+
+    DEBUG = args.debug
+    PROFILE = args.profile
+    NUM_PRERUN = args.num_prerun
+    NUM_ITERATIONS = args.num_iterations
+
+    for device in get_test_devices(args):
+        test_operator(device, test, _TEST_CASES, _TENSOR_DTYPES)
+
+    print("\033[92mAddcmul tests passed!\033[0m")

test/infiniop/atanh.py

+import torch
+import ctypes
+from ctypes import c_uint64
+from libinfiniop import (
+    LIBINFINIOP,
+    TestTensor,
+    get_test_devices,
+    check_error,
+    test_operator,
+    get_args,
+    debug,
+    get_tolerance,
+    profile_operation,
+    TestWorkspace,
+    InfiniDtype,
+    InfiniDtypeNames,
+    InfiniDeviceNames,
+    infiniopOperatorDescriptor_t,
+)
+from enum import Enum, auto
+
+# ==============================================================================
+# Configuration
+# ==============================================================================
+_TEST_CASES_ = [
+    # shape, a_stride, y_stride
+    ((13, 4), None, None),
+    ((13, 4), (10, 1), (10, 1)),
+    ((13, 4, 4), None, None),
+    ((13, 4, 4), (20, 4, 1), (20, 4, 1)),
+    ((16, 5632), None, None),
+    ((4, 4, 5632), (45056, 5632, 1), (45056, 5632, 1)),
+]
+
+class Inplace(Enum):
+    OUT_OF_PLACE = auto()
+    INPLACE_A = auto()
+
+_INPLACE = [
+    Inplace.OUT_OF_PLACE,
+    Inplace.INPLACE_A,
+]
+
+_TEST_CASES = [
+    test_case + (inplace_item,)
+    for test_case in _TEST_CASES_
+    for inplace_item in _INPLACE
+]
+
+# atanh typically supports floating point types
+_TENSOR_DTYPES = [InfiniDtype.F16, InfiniDtype.F32, InfiniDtype.BF16]
+
+_TOLERANCE_MAP = {
+    InfiniDtype.F16: {"atol": 1e-3, "rtol": 1e-3},
+    InfiniDtype.F32: {"atol": 1e-6, "rtol": 1e-6},
+    InfiniDtype.BF16: {"atol": 1e-2, "rtol": 1e-2},
+}
+
+DEBUG = False
+PROFILE = False
+NUM_PRERUN = 10
+NUM_ITERATIONS = 1000
+
+def atanh_torch(y, a):
+    torch.atanh(a, out=y)
+
+def test(
+    handle,
+    device,
+    shape,
+    a_stride=None,
+    y_stride=None,
+    inplace=Inplace.OUT_OF_PLACE,
+    dtype=torch.float16,
+    sync=None,
+):
+    # Initialize input tensor
+    a = TestTensor(shape, a_stride, dtype, device)
+    
+    # Crucial: clamp values to (-1, 1) to avoid NaN/Inf for atanh
+    with torch.no_grad():
+        a.torch_tensor().clamp_(-0.99, 0.99)
+        # Keep underlying data in sync for all devices (including CPU)
+        a.actual_tensor().copy_(a.torch_tensor())
+
+    if inplace == Inplace.INPLACE_A:
+        if a_stride != y_stride:
+            return
+        y = a
+    else:
+        y = TestTensor(shape, y_stride, dtype, device, mode="ones")
+
+    if y.is_broadcast():
+        return
+
+    print(
+        f"Testing Atanh on {InfiniDeviceNames[device]} with shape:{shape} a_stride:{a_stride} y_stride:{y_stride} "
+        f"dtype:{InfiniDtypeNames[dtype]} inplace:{inplace}"
+    )
+
+    # Reference calculation
+    atanh_torch(y.torch_tensor(), a.torch_tensor())
+
+    if sync is not None:
+        sync()
+
+    # Create descriptor
+    descriptor = infiniopOperatorDescriptor_t()
+    check_error(
+        LIBINFINIOP.infiniopCreateAtanhDescriptor(
+            handle,
+            ctypes.byref(descriptor),
+            y.descriptor,
+            a.descriptor,
+        )
+    )
+
+    # Invalidate descriptors to ensure kernel uses its own internal state
+    for tensor in [a, y]:
+        tensor.destroy_desc()
+
+    # Workspace management
+    workspace_size = c_uint64(0)
+    check_error(
+        LIBINFINIOP.infiniopGetAtanhWorkspaceSize(
+            descriptor, ctypes.byref(workspace_size)
+        )
+    )
+    workspace = TestWorkspace(workspace_size.value, y.device)
+
+    def lib_atanh():
+        check_error(
+            LIBINFINIOP.infiniopAtanh(
+                descriptor,
+                workspace.data(),
+                workspace.size(),
+                y.data(),
+                a.data(),
+                None,
+            )
+        )
+
+    # Run library function
+    lib_atanh()
+
+    # Validate results
+    atol, rtol = get_tolerance(_TOLERANCE_MAP, dtype)
+    if DEBUG:
+        debug(y.actual_tensor(), y.torch_tensor(), atol=atol, rtol=rtol)
+    
+    assert torch.allclose(y.actual_tensor(), y.torch_tensor(), atol=atol, rtol=rtol)
+
+    # Profiling
+    if PROFILE:
+        profile_operation("PyTorch", lambda: atanh_torch(y.torch_tensor(), a.torch_tensor()), device, NUM_PRERUN, NUM_ITERATIONS)
+        profile_operation("   lib", lambda: lib_atanh(), device, NUM_PRERUN, NUM_ITERATIONS)
+        
+    check_error(LIBINFINIOP.infiniopDestroyAtanhDescriptor(descriptor))
+
+if __name__ == "__main__":
+    args = get_args()
+
+    DEBUG = args.debug
+    PROFILE = args.profile
+    NUM_PRERUN = args.num_prerun
+    NUM_ITERATIONS = args.num_iterations
+
+    for device in get_test_devices(args):
+        test_operator(device, test, _TEST_CASES, _TENSOR_DTYPES)
+
+    print("\033[92mAtanh Test passed!\033[0m")

test/infiniop/binary_cross_entropy_with_logits.py

+import torch
+import ctypes
+from ctypes import c_uint64, c_float, c_char_p
+from libinfiniop import (
+    LIBINFINIOP,
+    TestTensor,
+    get_test_devices,
+    check_error,
+    test_operator,
+    get_args,
+    get_tolerance,
+    profile_operation,
+    TestWorkspace,
+    InfiniDtype,
+    InfiniDtypeNames,
+    InfiniDeviceNames,
+    infiniopOperatorDescriptor_t,
+)
+
+# ==============================================================================
+# Configuration
+# ==============================================================================
+# 测试场景：(shape, has_weight, has_pos_weight, reduction)
+_TEST_CASES_DATA = [
+    ((4, 5), False, False, "none"),
+    ((8, 8), True, False, "sum"),
+    ((32, 512), False, True, "mean"),
+    ((16, 32, 64), True, True, "mean"), 
+]
+
+_TENSOR_DTYPES = [InfiniDtype.F16, InfiniDtype.BF16, InfiniDtype.F32]
+
+_TOLERANCE_MAP = {
+    InfiniDtype.F16: {"atol": 1e-2, "rtol": 5e-2},
+    InfiniDtype.BF16: {"atol": 1e-2, "rtol": 5e-2},
+    InfiniDtype.F32: {"atol": 1e-5, "rtol": 1e-5},
+}
+
+_REDUCTIONS = ["none", "mean", "sum"]
+
+_REDUCTION_MAP = {
+    "none": 0,  # INFINIOP_REDUCTION_NONE
+    "mean": 1,  # INFINIOP_REDUCTION_MEAN
+    "sum": 2,   # INFINIOP_REDUCTION_SUM
+}
+
+# 生成最终测试用例组合
+_TEST_CASES = _TEST_CASES_DATA 
+
+DEBUG = False
+PROFILE = False
+
+def test(
+    handle,
+    device,
+    shape,
+    has_weight=False,
+    has_pos_weight=False,
+    reduction="none",
+    dtype=InfiniDtype.F16,
+    sync=None,
+):
+    print(
+        f"Testing BCEWithLogits on {InfiniDeviceNames[device]} shape:{shape} "
+        f"weight:{has_weight} pos_weight:{has_pos_weight} reduction:{reduction} dtype:{InfiniDtypeNames[dtype]}"
+    )
+
+    # 1. 准备输入 Tensor
+    input_tensor = TestTensor(shape, None, dtype, device)
+    target = TestTensor(shape, None, dtype, device)
+    
+    weight = TestTensor(shape, None, dtype, device) if has_weight else None
+    # pos_weight 通常在最后一维广播，形状为 (C,)
+    pos_weight_shape = (shape[-1],)
+    pos_weight = TestTensor(pos_weight_shape, None, dtype, device) if has_pos_weight else None
+
+    # 2. 使用 PyTorch 计算参考答案
+    torch_input = input_tensor.torch_tensor()
+    torch_target = target.torch_tensor()
+    torch_weight = weight.torch_tensor() if has_weight else None
+    torch_pos_weight = pos_weight.torch_tensor() if has_pos_weight else None
+
+    ans = torch.nn.functional.binary_cross_entropy_with_logits(
+        torch_input, 
+        torch_target, 
+        weight=torch_weight, 
+        pos_weight=torch_pos_weight, 
+        reduction=reduction
+    )
+
+    # 3. 准备输出 Tensor (根据 reduction 确定形状)
+    out_shape = () if reduction != "none" else shape
+    out = TestTensor(out_shape, None, dtype, device)
+
+    if sync is not None:
+        sync()
+
+    # 4. 创建描述符并执行
+    descriptor = infiniopOperatorDescriptor_t()
+    
+    # 模拟 C 接口调用
+    reduction_enum = _REDUCTION_MAP[reduction]
+    check_error(
+        LIBINFINIOP.infiniopCreateBCEWithLogitsDescriptor(
+            handle,
+            ctypes.byref(descriptor),
+            out.descriptor,
+            input_tensor.descriptor,
+            target.descriptor,
+            weight.descriptor if has_weight else None,
+            pos_weight.descriptor if has_pos_weight else None,
+            reduction_enum  # 传入归约方式枚举值，对应 infiniopReduction_t
+        )
+    )
+
+    # 销毁临时描述符
+    for t in [input_tensor, target, out]:
+        t.destroy_desc()
+    if weight: weight.destroy_desc()
+    if pos_weight: pos_weight.destroy_desc()
+
+    workspace_size = c_uint64(0)
+    check_error(LIBINFINIOP.infiniopGetBCEWithLogitsWorkspaceSize(descriptor, ctypes.byref(workspace_size)))
+    workspace = TestWorkspace(workspace_size.value, device)
+
+    def lib_op():
+        check_error(
+            LIBINFINIOP.infiniopBCEWithLogits(
+                descriptor,
+                workspace.data(),
+                workspace_size.value,
+                out.data(),
+                input_tensor.data(),
+                target.data(),
+                weight.data() if has_weight else None,
+                pos_weight.data() if has_pos_weight else None,
+                None,
+            )
+        )
+
+    lib_op()
+
+    if sync is not None:
+        sync()
+
+    # 5. 验证结果
+    atol, rtol = get_tolerance(_TOLERANCE_MAP, dtype)
+    assert torch.allclose(out.actual_tensor(), ans, atol=atol, rtol=rtol)
+
+    if PROFILE:
+        profile_operation("PyTorch", lambda: torch.nn.functional.binary_cross_entropy_with_logits(
+            torch_input, torch_target, weight=torch_weight, pos_weight=torch_pos_weight, reduction=reduction
+        ), device)
+        profile_operation("   lib", lib_op, device)
+
+    check_error(LIBINFINIOP.infiniopDestroyBCEWithLogitsDescriptor(descriptor))
+
+if __name__ == "__main__":
+    args = get_args()
+    for device in get_test_devices(args):
+        test_operator(device, test, _TEST_CASES, _TENSOR_DTYPES)
+    print("\033[92mBCEWithLogits tests passed!\033[0m")

test/infiniop/cdist.py

+import torch
+import ctypes
+from ctypes import c_uint64, c_float, c_double
+from libinfiniop import (
+    LIBINFINIOP,
+    TestTensor,
+    get_test_devices,
+    check_error,
+    test_operator,
+    get_args,
+    debug,
+    get_tolerance,
+    profile_operation,
+    TestWorkspace,
+    InfiniDtype,
+    InfiniDtypeNames,
+    InfiniDeviceNames,
+    infiniopOperatorDescriptor_t,
+)
+
+# ==============================================================================
+# Configuration
+# ==============================================================================
+# 格式: (M, N, D, x1_stride, x2_stride)
+# x1: (M, D), x2: (N, D), out: (M, N)
+_TEST_CASES_DATA = [
+    (5, 6, 3, None, None),
+    (32, 64, 128, None, None),
+    (32, 64, 128, (256, 1), (256, 1)), # 测试带步长的输入
+    (10, 7, 5, None, None),
+]
+
+_TENSOR_DTYPES = [InfiniDtype.F32] # cdist 通常对精度敏感，初测建议用 F32
+
+_TOLERANCE_MAP = {
+    InfiniDtype.F32: {"atol": 1e-5, "rtol": 1e-4},
+}
+
+_P_VALUES = [1.0, 2.0, float("inf")] # 不同的 p 范数测试
+
+_TEST_CASES = [
+    test_case + (p_val,)
+    for test_case in _TEST_CASES_DATA
+    for p_val in _P_VALUES
+]
+
+DEBUG = False
+PROFILE = False
+NUM_PRERUN = 10
+NUM_ITERATIONS = 100
+
+def test(
+    handle,
+    device,
+    M, N, D,
+    x1_stride=None,
+    x2_stride=None,
+    p=2.0,
+    dtype=InfiniDtype.F32,
+    sync=None,
+):
+    print(
+        f"Testing Cdist on {InfiniDeviceNames[device]} with M:{M}, N:{N}, D:{D}, p:{p}, dtype:{InfiniDtypeNames[dtype]}"
+    )
+
+    # 1. 准备输入输出形状
+    x1_shape = (M, D)
+    x2_shape = (N, D)
+    out_shape = (M, N)
+
+    # 2. 准备输入 Tensor
+    x1 = TestTensor(x1_shape, x1_stride, dtype, device)
+    x2 = TestTensor(x2_shape, x2_stride, dtype, device)
+    out = TestTensor(out_shape, None, dtype, device)
+
+    # 3. 使用 PyTorch 计算参考答案
+    # torch.cdist 要求输入至少是 2D
+    ans = torch.cdist(x1.torch_tensor(), x2.torch_tensor(), p=p)
+
+    if sync is not None:
+        sync()
+
+    # 4. 创建算子描述符
+    descriptor = infiniopOperatorDescriptor_t()
+    # 注意：这里假设 C 接口名为 infiniopCreateCdistDescriptor
+    check_error(
+        LIBINFINIOP.infiniopCreateCdistDescriptor(
+            handle, 
+            ctypes.byref(descriptor), 
+            out.descriptor, 
+            x1.descriptor, 
+            x2.descriptor, 
+            c_double(p) # 通常 p 使用 double 或 float 传递
+        )
+    )
+
+    # 销毁临时描述符以防内核错误引用（沿用 addcmul 风格）
+    for t in [x1, x2, out]:
+        t.destroy_desc()
+
+    # 5. Workspace 准备
+    workspace_size = c_uint64(0)
+    check_error(
+        LIBINFINIOP.infiniopGetCdistWorkspaceSize(
+            descriptor, ctypes.byref(workspace_size)
+        )
+    )
+    workspace = TestWorkspace(workspace_size.value, x1.device)
+
+    # 6. 执行函数定义
+    def lib_cdist():
+        check_error(
+            LIBINFINIOP.infiniopCdist(
+                descriptor,
+                workspace.data(),
+                workspace_size.value,
+                out.data(),
+                x1.data(),
+                x2.data(),
+                None, # stream
+            )
+        )
+
+    # 7. 运行
+    lib_cdist()
+
+    if sync is not None:
+        sync()
+
+    # 8. 验证结果
+    atol, rtol = get_tolerance(_TOLERANCE_MAP, dtype)
+    if DEBUG:
+        debug(out.actual_tensor(), ans, atol=atol, rtol=rtol)
+    
+    assert torch.allclose(out.actual_tensor(), ans, atol=atol, rtol=rtol)
+
+    # 9. 性能分析
+    if PROFILE:
+        profile_operation("PyTorch", lambda: torch.cdist(x1.torch_tensor(), x2.torch_tensor(), p=p), device, NUM_PRERUN, NUM_ITERATIONS)
+        profile_operation("   lib", lambda: lib_cdist(), device, NUM_PRERUN, NUM_ITERATIONS)
+
+    check_error(LIBINFINIOP.infiniopDestroyCdistDescriptor(descriptor))
+
+
+if __name__ == "__main__":
+    args = get_args()
+
+    DEBUG = args.debug
+    PROFILE = args.profile
+    NUM_PRERUN = args.num_prerun
+    NUM_ITERATIONS = args.num_iterations
+
+    for device in get_test_devices(args):
+        test_operator(device, test, _TEST_CASES, _TENSOR_DTYPES)
+
+    print("\033[92mCdist tests passed!\033[0m")

test/infiniop/libinfiniop/op_register.py

@@ -4,8 +4,7 @@ from .structs import (
     infiniopOperatorDescriptor_t,
 )
 
-from ctypes import c_int32, c_void_p, c_size_t, POINTER, c_float
-
+from ctypes import c_int32, c_void_p, c_size_t, POINTER, c_float, c_double, c_uint64
 
 class OpRegister:
     registry = []
@@ -20,6 +19,36 @@ class OpRegister:
         for op in cls.registry:
             op(lib)
 
+@OpRegister.operator
+def atanh_(lib):
+    lib.infiniopCreateAtanhDescriptor.restype = c_int32
+    lib.infiniopCreateAtanhDescriptor.argtypes = [
+        infiniopHandle_t,
+        POINTER(infiniopOperatorDescriptor_t),
+        infiniopTensorDescriptor_t,
+        infiniopTensorDescriptor_t,
+    ]
+
+    lib.infiniopGetAtanhWorkspaceSize.restype = c_int32
+    lib.infiniopGetAtanhWorkspaceSize.argtypes = [
+        infiniopOperatorDescriptor_t,
+        POINTER(c_size_t),
+    ]
+
+    lib.infiniopAtanh.restype = c_int32
+    lib.infiniopAtanh.argtypes = [
+        infiniopOperatorDescriptor_t,
+        c_void_p,  # workspace
+        c_size_t,  # workspace_size
+        c_void_p,  # y_data
+        c_void_p,  # a_data
+        c_void_p,  # stream
+    ]
+
+    lib.infiniopDestroyAtanhDescriptor.restype = c_int32
+    lib.infiniopDestroyAtanhDescriptor.argtypes = [
+        infiniopOperatorDescriptor_t,
+    ]
 
 @OpRegister.operator
 def add_(lib):
@@ -55,6 +84,156 @@ def add_(lib):
     ]
 
 @OpRegister.operator
+def addcmul_(lib):
+    lib.infiniopCreateAddcmulDescriptor.restype = c_int32
+    lib.infiniopCreateAddcmulDescriptor.argtypes = [
+        infiniopHandle_t,                       # handle
+        POINTER(infiniopOperatorDescriptor_t),  # desc_ptr
+        infiniopTensorDescriptor_t,             # out_desc
+        infiniopTensorDescriptor_t,             # input_desc
+        infiniopTensorDescriptor_t,             # t1_desc
+        infiniopTensorDescriptor_t,             # t2_desc
+        c_float,                                # value (标量系数)
+    ]
+
+    lib.infiniopGetAddcmulWorkspaceSize.restype = c_int32
+    lib.infiniopGetAddcmulWorkspaceSize.argtypes = [
+        infiniopOperatorDescriptor_t,           # descriptor
+        POINTER(c_size_t),                      # size_ptr
+    ]
+
+    lib.infiniopAddcmul.restype = c_int32
+    lib.infiniopAddcmul.argtypes = [
+        infiniopOperatorDescriptor_t,           # descriptor
+        c_void_p,                               # workspace
+        c_size_t,                               # workspace_size
+        c_void_p,                               # out_ptr
+        c_void_p,                               # input_ptr
+        c_void_p,                               # t1_ptr
+        c_void_p,                               # t2_ptr
+        c_void_p,                               # stream
+    ]
+
+    lib.infiniopDestroyAddcmulDescriptor.restype = c_int32
+    lib.infiniopDestroyAddcmulDescriptor.argtypes = [
+        infiniopOperatorDescriptor_t,           # descriptor
+    ]
+    
+@OpRegister.operator
+def cdist_(lib):
+    # 1. 创建描述符接口
+    # 接口通常接收 handle, 输出 desc, 两个输入 desc, 以及范数 p
+    lib.infiniopCreateCdistDescriptor.restype = c_int32
+    lib.infiniopCreateCdistDescriptor.argtypes = [
+        infiniopHandle_t,                       # handle
+        POINTER(infiniopOperatorDescriptor_t),  # desc_ptr
+        infiniopTensorDescriptor_t,             # y_desc (输出)
+        infiniopTensorDescriptor_t,             # x1_desc
+        infiniopTensorDescriptor_t,             # x2_desc
+        c_double,                               # p (范数阶数)
+    ]
+
+    # 2. 获取 Workspace 大小接口
+    lib.infiniopGetCdistWorkspaceSize.restype = c_int32
+    lib.infiniopGetCdistWorkspaceSize.argtypes = [
+        infiniopOperatorDescriptor_t,           # descriptor
+        POINTER(c_size_t),                      # size_ptr
+    ]
+
+    # 3. 执行算子接口
+    lib.infiniopCdist.restype = c_int32
+    lib.infiniopCdist.argtypes = [
+        infiniopOperatorDescriptor_t,           # descriptor
+        c_void_p,                               # workspace
+        c_size_t,                               # workspace_size
+        c_void_p,                               # y_ptr
+        c_void_p,                               # x1_ptr
+        c_void_p,                               # x2_ptr
+        c_void_p,                               # stream
+    ]
+
+    # 4. 销毁描述符接口
+    lib.infiniopDestroyCdistDescriptor.restype = c_int32
+    lib.infiniopDestroyCdistDescriptor.argtypes = [
+        infiniopOperatorDescriptor_t,           # descriptor
+    ]
+
+@OpRegister.operator
+def binary_cross_entropy_with_logits_(lib):
+    # 1. 创建描述符 (Descriptor Creation)
+    lib.infiniopCreateBCEWithLogitsDescriptor.restype = c_int32
+    lib.infiniopCreateBCEWithLogitsDescriptor.argtypes = [
+        infiniopHandle_t,                        # handle
+        POINTER(infiniopOperatorDescriptor_t),   # desc_ptr
+        infiniopTensorDescriptor_t,              # out_desc
+        infiniopTensorDescriptor_t,              # input_desc (logits)
+        infiniopTensorDescriptor_t,              # target_desc
+        infiniopTensorDescriptor_t,              # weight_desc (可选，不可用则传 NULL)
+        infiniopTensorDescriptor_t,              # pos_weight_desc (可选，不可用则传 NULL)
+        c_int32                                  # reduction (0:none, 1:mean, 2:sum)
+    ]
+
+    # 2. 获取工作空间大小 (Workspace Size)
+    lib.infiniopGetBCEWithLogitsWorkspaceSize.restype = c_int32
+    lib.infiniopGetBCEWithLogitsWorkspaceSize.argtypes = [
+        infiniopOperatorDescriptor_t,            # descriptor
+        POINTER(c_size_t),                       # size_ptr
+    ]
+
+    # 3. 执行算子 (Execution)
+    lib.infiniopBCEWithLogits.restype = c_int32
+    lib.infiniopBCEWithLogits.argtypes = [
+        infiniopOperatorDescriptor_t,            # descriptor
+        c_void_p,                                # workspace
+        c_size_t,                                # workspace_size
+        c_void_p,                                # out_ptr
+        c_void_p,                                # input_ptr (logits)
+        c_void_p,                                # target_ptr
+        c_void_p,                                # weight_ptr (可选)
+        c_void_p,                                # pos_weight_ptr (可选)
+        c_void_p,                                # stream
+    ]
+
+    # 4. 销毁描述符 (Destruction)
+    lib.infiniopDestroyBCEWithLogitsDescriptor.restype = c_int32
+    lib.infiniopDestroyBCEWithLogitsDescriptor.argtypes = [
+        infiniopOperatorDescriptor_t,            # descriptor
+    ]
+
+@OpRegister.operator
+def reciprocal_(lib):
+    lib.infiniopCreateReciprocalDescriptor.restype = c_int32
+    lib.infiniopCreateReciprocalDescriptor.argtypes = [
+        infiniopHandle_t,
+        POINTER(infiniopOperatorDescriptor_t),
+        infiniopTensorDescriptor_t, # Output descriptor
+        infiniopTensorDescriptor_t, # Input descriptor
+    ]
+
+    # 获取工作空间大小接口
+    lib.infiniopGetReciprocalWorkspaceSize.restype = c_int32
+    lib.infiniopGetReciprocalWorkspaceSize.argtypes = [
+        infiniopOperatorDescriptor_t,
+        POINTER(c_size_t),
+    ]
+
+    # 最后的 c_void_p 通常对应 stream 或其他异步句柄，保持一致即可
+    lib.infiniopReciprocal.restype = c_int32
+    lib.infiniopReciprocal.argtypes = [
+        infiniopOperatorDescriptor_t,
+        c_void_p, # Workspace pointer
+        c_size_t, # Workspace size
+        c_void_p, # Output data pointer
+        c_void_p, # Input data pointer
+        c_void_p, # Stream pointer (optional)
+    ]
+
+    # 销毁描述符接口
+    lib.infiniopDestroyReciprocalDescriptor.restype = c_int32
+    lib.infiniopDestroyReciprocalDescriptor.argtypes = [
+        infiniopOperatorDescriptor_t,
+    ]
+@OpRegister.operator
 def equal_(lib):
     # =========================================================
     # 1. 注册 Create 函数

test/infiniop/reciprocal.py

+import torch
+import ctypes
+from ctypes import c_uint64
+from libinfiniop import (
+    LIBINFINIOP,
+    TestTensor,
+    get_test_devices,
+    check_error,
+    test_operator,
+    get_args,
+    debug,
+    get_tolerance,
+    profile_operation,
+    TestWorkspace,
+    InfiniDtype,
+    InfiniDtypeNames,
+    InfiniDeviceNames,
+    infiniopOperatorDescriptor_t,
+)
+from enum import Enum, auto
+
+# ==============================================================================
+# Configuration
+# ==============================================================================
+_TEST_CASES_ = [
+    # shape, input_stride, output_stride
+    ((13, 4), None, None),
+    ((13, 4), (10, 1), (10, 1)),
+    ((16, 5632), None, None),
+    ((16, 5632), (13312, 1), (13312, 1)),
+    ((13, 16, 2), (128, 4, 1), (64, 4, 1)),
+    ((4, 4, 5632), None, None),
+]
+
+class Inplace(Enum):
+    OUT_OF_PLACE = auto()
+    INPLACE = auto()
+
+_INPLACE = [
+    Inplace.OUT_OF_PLACE,
+    Inplace.INPLACE,
+]
+
+_TEST_CASES = [
+    test_case + (inplace_item,)
+    for test_case in _TEST_CASES_
+    for inplace_item in _INPLACE
+]
+
+# Reciprocal usually outputs floats; Integer types are often not supported or special-cased
+_TENSOR_DTYPES = [InfiniDtype.F16, InfiniDtype.F32, InfiniDtype.BF16]
+
+_TOLERANCE_MAP = {
+    InfiniDtype.F16: {"atol": 1e-3, "rtol": 1e-3},
+    InfiniDtype.F32: {"atol": 1e-5, "rtol": 1e-5},
+    InfiniDtype.BF16: {"atol": 1e-2, "rtol": 1e-2},
+}
+
+DEBUG = False
+PROFILE = False
+NUM_PRERUN = 10
+NUM_ITERATIONS = 1000
+
+def reciprocal(y, x):
+    torch.reciprocal(x, out=y)
+
+def test(
+    handle,
+    device,
+    shape,
+    in_stride=None,
+    out_stride=None,
+    inplace=Inplace.OUT_OF_PLACE,
+    dtype=torch.float16,
+    sync=None,
+):
+    # Initialize input 'x'
+    # Use 'random' mode but ensure values are not near zero to avoid infinity
+    x = TestTensor(shape, in_stride, dtype, device)
+    
+    if inplace == Inplace.INPLACE:
+        if in_stride != out_stride:
+            return
+        y = x
+    else:
+        y = TestTensor(shape, out_stride, dtype, device)
+
+    if y.is_broadcast():
+        return
+
+    print(
+        f"Testing Reciprocal on {InfiniDeviceNames[device]} with shape:{shape} "
+        f"in_stride:{in_stride} out_stride:{out_stride} "
+        f"dtype:{InfiniDtypeNames[dtype]} inplace:{inplace}"
+    )
+
+    # Calculate ground truth using PyTorch
+    reciprocal(y.torch_tensor(), x.torch_tensor())
+
+    if sync is not None:
+        sync()
+
+    # Create Descriptor
+    descriptor = infiniopOperatorDescriptor_t()
+    check_error(
+        LIBINFINIOP.infiniopCreateReciprocalDescriptor(
+            handle,
+            ctypes.byref(descriptor),
+            y.descriptor,
+            x.descriptor,
+        )
+    )
+
+    # Invalidate descriptors as per framework requirement
+    for tensor in [x, y]:
+        tensor.destroy_desc()
+
+    # Workspace allocation
+    workspace_size = c_uint64(0)
+    check_error(
+        LIBINFINIOP.infiniopGetReciprocalWorkspaceSize(
+            descriptor, ctypes.byref(workspace_size)
+        )
+    )
+    workspace = TestWorkspace(workspace_size.value, y.device)
+
+    def lib_reciprocal():
+        check_error(
+            LIBINFINIOP.infiniopReciprocal(
+                descriptor,
+                workspace.data(),
+                workspace.size(),
+                y.data(),
+                x.data(),
+                None,
+            )
+        )
+
+    lib_reciprocal()
+
+    # Verification
+    atol, rtol = get_tolerance(_TOLERANCE_MAP, dtype)
+    if DEBUG:
+        debug(y.actual_tensor(), y.torch_tensor(), atol=atol, rtol=rtol)
+    assert torch.allclose(y.actual_tensor(), y.torch_tensor(), atol=atol, rtol=rtol)
+
+    # Profiling
+    if PROFILE:
+        profile_operation("PyTorch", lambda: reciprocal(y.torch_tensor(), x.torch_tensor()), device, NUM_PRERUN, NUM_ITERATIONS)
+        profile_operation("   lib", lambda: lib_reciprocal(), device, NUM_PRERUN, NUM_ITERATIONS)
+        
+    check_error(LIBINFINIOP.infiniopDestroyReciprocalDescriptor(descriptor))
+
+if __name__ == "__main__":
+    args = get_args()
+
+    DEBUG = args.debug
+    PROFILE = args.profile
+    NUM_PRERUN = args.num_prerun
+    NUM_ITERATIONS = args.num_iterations
+
+    for device in get_test_devices(args):
+        test_operator(device, test, _TEST_CASES, _TENSOR_DTYPES)
+
+    print("\033[92mTest passed!\033[0m")