Merge remote-tracking branch 'origin/main' into issue/150

src/infiniop/devices/kunlun/kunlun_common.h → src/infiniop/devices/kunlun/kunlun_kernel_common.h

-#ifndef __INFINIOP_KUNLUN_COMMON_H__
-#define __INFINIOP_KUNLUN_COMMON_H__
+#ifndef __INFINIOP_KUNLUN_KERNEL_COMMON_H__
+#define __INFINIOP_KUNLUN_KERNEL_COMMON_H__
 
 // This header file will only be include by .xpu file
+#include "kunlun_kernel_dtype.h"
 #include "xpu/kernel/xtdk.h"
 #include "xpu/kernel/xtdk_math.h"
 #include "xpu/kernel/xtdk_simd.h"
 #include "xpu/runtime.h"
 
+namespace device::kunlun::kernel {
 // Get mask for kunlun xpu 512bit register calculation
 // if data is not enough to 512bit, padding zero and use
 // mask to identify real data
@@ -26,6 +28,37 @@ inline __device__ void atomicAddF32(__shared_ptr__ float *ptr, float value) {
     }
 }
 
+inline __device__ size_t indexToReducedOffset(
+    size_t flat_index,
+    size_t ndim,
+    const _ptrdiff_t *broadcasted_strides,
+    const _ptrdiff_t *target_strides) {
+
+    size_t res = 0;
+    for (size_t i = 0; i < ndim; ++i) {
+        res += flat_index / broadcasted_strides[i].value * target_strides[i].value;
+        flat_index %= broadcasted_strides[i].value;
+        mfence();
+    }
+    return res;
+}
+
+inline __device__ size_t indexToOffset(
+    size_t flat_index,
+    size_t ndim,
+    const _size_t *shape,
+    const _ptrdiff_t *strides) {
+
+    size_t res = 0;
+    for (size_t i = ndim; i-- > 0;) {
+        res += (flat_index % shape[i].value) * strides[i].value;
+        flat_index /= shape[i].value;
+        mfence();
+    }
+    return res;
+}
+
+} // namespace device::kunlun::kernel
 // TODO: atomicAddF16
 // TODO: atomicAddI8
 #endif

src/infiniop/devices/kunlun/kunlun_kernel_dtype.h

+#ifndef __INFINIOP_KUNLUN_DTYPE_H__
+#define __INFINIOP_KUNLUN_DTYPE_H__
+
+#include "xpu/kernel/xtdk.h"
+#include "xpu/kernel/xtdk_math.h"
+#include "xpu/kernel/xtdk_simd.h"
+#include "xpu/runtime.h"
+
+// kunlun ptrdiff_t* is used to save ptrdiff_t array
+// copied from host
+typedef struct _ptrdiff_t {
+    long value;   // 32 bit
+    long padding; // 32 bit
+} _ptrdiff_t;
+
+// same as ptrdiff
+typedef struct _size_t {
+    size_t value;
+    size_t padding;
+} _size_t;
+
+#endif

src/infiniop/devices/maca/maca_kernel_common.h

+#define INFINIOP_MACA_KERNEL __global__ void
+// Posible maximum number of threads per block for MACA architectures
+// Used for picking correct kernel launch configuration
+#define MACA_BLOCK_SIZE_1024 1024
+#define MACA_BLOCK_SIZE_512 512
+
+#define CHECK_MACA(API) CHECK_INTERNAL(API, hcSuccess)
+
+namespace device::maca {
+
+// return the memory offset of original tensor, given the flattened index of broadcasted tensor
+__forceinline__ __device__ __host__ size_t
+indexToReducedOffset(
+    size_t flat_index,
+    size_t ndim,
+    const ptrdiff_t *broadcasted_strides,
+    const ptrdiff_t *target_strides) {
+    size_t res = 0;
+    for (size_t i = 0; i < ndim; ++i) {
+        res += flat_index / broadcasted_strides[i] * target_strides[i];
+        flat_index %= broadcasted_strides[i];
+    }
+    return res;
+}
+
+// get the memory offset of the given element in a tensor given its flat index
+__forceinline__ __device__ __host__ size_t
+indexToOffset(
+    size_t flat_index,
+    size_t ndim,
+    const size_t *shape,
+    const ptrdiff_t *strides) {
+    size_t res = 0;
+    for (size_t i = ndim; i-- > 0;) {
+        res += (flat_index % shape[i]) * strides[i];
+        flat_index /= shape[i];
+    }
+    return res;
+}
+} // namespace device::maca
+
+#ifdef ENABLE_MACA_API
+#include <maca_fp16.h>
+__forceinline__ __device__ float
+exp_(const float val) {
+    return expf(val);
+}
+
+__forceinline__ __device__ long double
+exp_(const long double val) {
+    return expl(val);
+}
+
+__forceinline__ __device__ double
+exp_(const double val) {
+    return exp(val);
+}
+
+__forceinline__ __device__ __half
+exp_(const __half x) {
+    return hexp(x);
+}
+#endif

src/infiniop/elementwise/kunlun/elementwise_kunlun.h

+#ifndef __INFINIOP_ELEMENTWISE_KUNLUN_H__
+#define __INFINIOP_ELEMENTWISE_KUNLUN_H__
+
+#include "../../../utils.h"
+#include "../../devices/kunlun/kunlun_handle.h"
+#include "elementwise_kunlun_api.h"
+
+namespace op::elementwise::kunlun {
+
+struct DeviceImpl::Opaque {
+    std::shared_ptr<device::kunlun::Handle::Internal> internal;
+
+    Opaque(const std::shared_ptr<device::kunlun::Handle::Internal> &internal_)
+        : internal(internal_) {}
+
+    template <size_t N, typename Op, typename Tdata, typename... Args>
+    infiniStatus_t calculateImpl(const op::elementwise::ElementwiseInfo &info,
+                                 void *workspace,
+                                 void *output,
+                                 const std::vector<const void *> &inputs,
+                                 kunlunStream_t stream,
+                                 Args &&...args) {
+
+        auto output_size = info.getOutputSize();
+        if (output_size == 0) {
+            return INFINI_STATUS_SUCCESS;
+        }
+
+        // Device pointers
+        const void **d_inputs_arr = nullptr;
+        const bool *d_input_contiguous = nullptr;
+        const bool *d_input_broadcasted = nullptr;
+        const size_t *d_output_shape = nullptr;
+        const ptrdiff_t *d_output_strides = nullptr;
+        const size_t *d_input_shapes = nullptr;
+        const ptrdiff_t *d_input_strides = nullptr;
+
+        CHECK_STATUS(infoToDevice<N>(info, workspace, inputs.data(), d_inputs_arr,
+                                     d_input_contiguous, d_input_broadcasted,
+                                     d_output_shape, d_output_strides,
+                                     d_input_shapes, d_input_strides));
+
+        Op::template launch<Tdata>(
+            output_size,
+            info.getNdim(),
+            info.isOutputContiguous(),
+            reinterpret_cast<const void *>(d_input_contiguous),
+            reinterpret_cast<const void *>(d_input_broadcasted),
+            reinterpret_cast<const void *>(d_output_shape),
+            reinterpret_cast<const void *>(d_input_shapes),
+            reinterpret_cast<const void *>(d_output_strides),
+            reinterpret_cast<const void *>(d_input_strides),
+            output,
+            reinterpret_cast<const void *const *>(d_inputs_arr),
+            stream,
+            args...);
+
+        return INFINI_STATUS_SUCCESS;
+    }
+
+private:
+    template <size_t N>
+    infiniStatus_t infoToDevice(
+        const op::elementwise::ElementwiseInfo &info,
+        void *workspace,
+        const void *const *h_inputs_arr,
+        const void **&d_inputs_arr,
+        const bool *&d_input_contiguous,
+        const bool *&d_input_broadcasted,
+        const size_t *&d_output_shape,
+        const ptrdiff_t *&d_output_strides,
+        const size_t *&d_input_shapes,
+        const ptrdiff_t *&d_input_strides) const {
+
+        constexpr auto input_size = N;
+        const auto ndim = info.getNdim();
+        constexpr auto input_arr_size = N * sizeof(*h_inputs_arr);
+        const int8_t *info_meta_start = info.getMetaStart();
+        const int8_t *d_meta_start = reinterpret_cast<int8_t *>(workspace) + input_arr_size;
+
+        // copy the input pointer array and meta to device
+        CHECK_KUNLUN(xpu_memcpy(workspace, h_inputs_arr, input_arr_size, XPU_HOST_TO_DEVICE));
+        CHECK_KUNLUN(xpu_memcpy((void *)d_meta_start, info_meta_start, info.getMetaMemSize(), XPU_HOST_TO_DEVICE));
+
+        // offset/assign the pointers
+        d_inputs_arr = reinterpret_cast<const void **>(workspace);
+        d_output_shape = reinterpret_cast<const size_t *>(d_meta_start);
+        d_output_strides = reinterpret_cast<const ptrdiff_t *>(d_output_shape + ndim);
+        d_input_shapes = reinterpret_cast<const size_t *>(d_output_strides + ndim);
+        d_input_strides = reinterpret_cast<const ptrdiff_t *>(d_input_shapes + input_size * ndim);
+        d_input_contiguous = reinterpret_cast<const bool *>(d_input_strides + input_size * ndim);
+        d_input_broadcasted = reinterpret_cast<const bool *>(d_input_contiguous + input_size);
+
+        return INFINI_STATUS_SUCCESS;
+    }
+};
+
+template <typename... Args>
+utils::Result<DeviceImpl *> DeviceImpl::create(Args &&...args) {
+    auto opaque = std::make_shared<Opaque>(std::forward<Args>(args)...);
+    return utils::Result<DeviceImpl *>(new DeviceImpl(opaque));
+}
+
+template <typename Op, typename Tdata, typename... Args>
+infiniStatus_t DeviceImpl::calculate(const op::elementwise::ElementwiseInfo &info,
+                                     void *workspace,
+                                     void *output,
+                                     const std::vector<const void *> &inputs,
+                                     void *stream,
+                                     Args &&...args) {
+    constexpr size_t N = Op::num_inputs;
+    return _opaque->calculateImpl<N, Op, Tdata>(
+        info, workspace, output, inputs,
+        reinterpret_cast<kunlunStream_t>(stream),
+        std::forward<Args>(args)...);
+}
+} // namespace op::elementwise::kunlun
+
+// Template for kunlun kernel interface declaration
+#define LAUNCH_ELEMENTWISE_KERNEL(OpName)       \
+    template <typename Tdata, typename... Args> \
+    void launch##OpName##Kernel(                \
+        size_t output_size,                     \
+        size_t ndim,                            \
+        bool output_contiguous,                 \
+        const void *input_contiguous,           \
+        const void *input_broadcasted,          \
+        const void *output_shape,               \
+        const void *input_shapes,               \
+        const void *output_strides,             \
+        const void *input_strides,              \
+        void *output,                           \
+        const void *const *inputs,              \
+        XPUStream stream,                       \
+        Args... args);
+
+#endif

src/infiniop/elementwise/kunlun/elementwise_kunlun_api.h

+#ifndef __INFINIOP_ELEMENTWISE_KUNLUN_API_H__
+#define __INFINIOP_ELEMENTWISE_KUNLUN_API_H__
+
+#include "../elementwise.h"
+
+namespace op::elementwise::kunlun {
+
+class DeviceImpl final {
+    struct Opaque;
+    std::shared_ptr<Opaque> _opaque;
+
+    DeviceImpl(std::shared_ptr<Opaque> opaque) : _opaque(std::move(opaque)) {}
+
+public:
+    ~DeviceImpl() = default;
+
+    template <typename... Args>
+    static utils::Result<DeviceImpl *> create(Args &&...args);
+
+    template <typename Op, typename Tdata, typename... Args>
+    infiniStatus_t calculate(
+        const op::elementwise::ElementwiseInfo &info,
+        void *workspace,
+        void *output,
+        const std::vector<const void *> &inputs,
+        void *stream,
+        Args &&...args);
+};
+
+} // namespace op::elementwise::kunlun
+
+#define CREATE_ELEMENTWISE_KUNLUN_DESCRIPTOR(HANDLE, DTYPE, OUT_DESC, INPUT_DESC_VEC)          \
+                                                                                               \
+    auto info_result = op::elementwise::ElementwiseInfo::create(OUT_DESC, INPUT_DESC_VEC);     \
+    CHECK_RESULT(info_result);                                                                 \
+    auto info = info_result.take();                                                            \
+    auto workspace_size = info.getMetaMemSize() + info.getInputSize() * sizeof(void *);        \
+                                                                                               \
+    auto device_impl_result = op::elementwise::kunlun::DeviceImpl::create(HANDLE->internal()); \
+    CHECK_RESULT(device_impl_result);                                                          \
+                                                                                               \
+    *desc_ptr = new Descriptor(                                                                \
+        DTYPE,                                                                                 \
+        std::move(info),                                                                       \
+        std::move(device_impl_result.take()),                                                  \
+        workspace_size,                                                                        \
+        HANDLE->device,                                                                        \
+        HANDLE->device_id);
+
+#endif

src/infiniop/elementwise/kunlun/elementwise_kunlun_kernel.h

+#ifndef __INFINIOP_ELEMENTWISE_KUNLUN_XPU__
+#define __INFINIOP_ELEMENTWISE_KUNLUN_XPU__
+
+#include "../../devices/kunlun/kunlun_kernel_common.h"
+
+using namespace device::kunlun::kernel;
+
+/**
+ * @brief Computes input tile offset
+ */
+struct InputIndexer {
+    size_t idx;
+    size_t ndim;
+    const bool *input_contiguous;
+    const bool *input_broadcasted;
+    const _size_t *input_shapes;
+    const _ptrdiff_t *input_strides;
+    const _ptrdiff_t *output_strides;
+
+    __device__ size_t operator()(size_t input_id) const {
+        return input_contiguous[input_id]
+                 ? idx
+                 : (input_broadcasted[input_id]
+                        ? indexToReducedOffset(idx, ndim, output_strides, input_strides + input_id * ndim)
+                        : indexToOffset(idx, ndim, input_shapes + input_id * ndim, input_strides + input_id * ndim));
+    }
+};
+
+/**
+ * @brief Computes the output index in memory, accounting for strides if non-contiguous.
+ *
+ * @param idx            Linear index.
+ * @param is_contiguous  Whether the output tensor is contiguous.
+ * @param ndim           Number of dimensions.
+ * @param shape          Shape of the output tensor.
+ * @param strides        Strides of the output tensor.
+ * @return               Memory offset index.
+ */
+inline __device__ size_t
+getOutputIndex(size_t idx,
+               bool is_contiguous,
+               size_t ndim,
+               const _size_t *shape,
+               const _ptrdiff_t *strides) {
+    return is_contiguous ? idx : indexToOffset(idx, ndim, shape, strides);
+}
+
+template <size_t N, typename Op, typename Tdata, typename... Args>
+__device__ void launchOp(
+    __global_ptr__ Tdata **typed_inputs, // gm pointer
+    __global_ptr__ Tdata *output,        // gm pointer output
+    Tdata *inputs_buf,                   // local mem buffer
+    size_t *input_indexes,
+    size_t output_index,
+    Args... args) {
+
+    static_assert(N == Op::num_inputs, "template N is not equal to Op::num_inputs!\n");
+
+#pragma unroll
+    // Copy inputs to buf
+    for (size_t i = 0; i < N; i++) {
+        auto gm = typed_inputs[i] + input_indexes[i];
+        auto lm = inputs_buf + i;
+        GM2LM_ASYNC(gm, lm, 1 * sizeof(Tdata));
+    }
+    mfence();
+
+    // Calculate elementwise
+    // Inputs save all operands
+    Tdata out = Op{}(inputs_buf, args...);
+
+    // Copy out to gm
+    LM2GM_ASYNC(&out, output + output_index, 1 * sizeof(Tdata));
+    mfence();
+}
+
+template <size_t N, typename Op, typename Tdata, typename... Args>
+__global__ void elementwiseKernel(
+    size_t output_size,
+    size_t ndim,
+    bool output_contiguous,
+    const bool *input_contiguous_gm,
+    const bool *input_broadcasted_gm,
+    const _size_t *output_shape_gm,
+    const _size_t *input_shapes_gm,
+    const _ptrdiff_t *output_strides_gm,
+    const _ptrdiff_t *input_strides_gm,
+    Tdata *output,
+    const void *const *inputs,
+    Args... args) {
+
+    int cid = core_id();
+    int ncores = core_num();
+    if (cid >= ncores) {
+        return;
+    }
+    int thread_id = ncores * cluster_id() + cid;
+    int nthreads = ncores * cluster_num();
+
+    // Cast input gm pointer type
+    auto typed_inputs = reinterpret_cast<const __global_ptr__ Tdata *const __global_ptr__ *>(inputs);
+
+    const int BUFF_SIZE = 64;
+    // Input data cache
+    __local__ Tdata inputs_buf[N];
+    // Input contiguous/broadcasted flags
+    __local__ bool input_contiguous[N];
+    __local__ bool input_broadcasted[N];
+    // Input shape/strides
+    __local__ _size_t input_shapes[N * ndim];
+    __local__ _ptrdiff_t input_strides[N * ndim];
+    // Output shape/strides
+    __local__ _size_t output_shape[ndim];
+    __local__ _ptrdiff_t output_strides[ndim];
+    // Inputs gm ptr buf
+    __local__ __global_ptr__ Tdata *typed_inputs_ptr[N];
+
+    // Load from gm
+    GM2LM_ASYNC(input_contiguous_gm, input_contiguous, N * sizeof(bool));
+    GM2LM_ASYNC(input_broadcasted_gm, input_broadcasted, N * sizeof(bool));
+    GM2LM_ASYNC(input_shapes_gm, input_shapes, N * ndim * sizeof(_size_t));
+    GM2LM_ASYNC(input_strides_gm, input_strides, N * ndim * sizeof(_ptrdiff_t));
+    GM2LM_ASYNC(output_shape_gm, output_shape, ndim * sizeof(_size_t));
+    GM2LM_ASYNC(output_strides_gm, output_strides, ndim * sizeof(_ptrdiff_t));
+    GM2LM_ASYNC(typed_inputs, typed_inputs_ptr, N * sizeof(__global_ptr__ Tdata *));
+    mfence();
+
+    int len_per_loop = min(BUFF_SIZE, roundup_div(output_size, nthreads));
+
+    for (int start = thread_id * len_per_loop; start < output_size; start += nthreads * len_per_loop) {
+        size_t read_len = min(len_per_loop, output_size - start);
+        for (int idx = start; idx < start + read_len; ++idx) {
+            size_t out_idx = getOutputIndex(static_cast<size_t>(idx), output_contiguous,
+                                            ndim, output_shape, output_strides);
+            InputIndexer indexer{static_cast<size_t>(idx), ndim, input_contiguous, input_broadcasted,
+                                 input_shapes, input_strides, output_strides};
+            // Get index offset for every operand
+            size_t indexes[N];
+            for (size_t i = 0; i < N; i++) {
+                indexes[i] = indexer(i);
+            }
+            // Launch operater
+            launchOp<N, Op, Tdata>(&typed_inputs_ptr[0], output, inputs_buf, indexes, out_idx, args...);
+        }
+    }
+    sync_cluster();
+}
+
+#define LAUNCH_ELEMENTWISE_KERNEL_IMPL(OpName, Op)                       \
+    template <typename Tdata, typename... Args>                          \
+    void launch##OpName##Kernel(                                         \
+        size_t output_size,                                              \
+        size_t ndim,                                                     \
+        bool output_contiguous,                                          \
+        const void *input_contiguous,                                    \
+        const void *input_broadcasted,                                   \
+        const void *output_shape,                                        \
+        const void *input_shapes,                                        \
+        const void *output_strides,                                      \
+        const void *input_strides,                                       \
+        void *output,                                                    \
+        const void *const *inputs,                                       \
+        XPUStream stream,                                                \
+        Args... args) {                                                  \
+        elementwiseKernel<Op::num_inputs, Op, Tdata><<<8, 64, stream>>>( \
+            output_size, ndim, output_contiguous,                        \
+            reinterpret_cast<const bool *>(input_contiguous),            \
+            reinterpret_cast<const bool *>(input_broadcasted),           \
+            reinterpret_cast<const _size_t *>(output_shape),             \
+            reinterpret_cast<const _size_t *>(input_shapes),             \
+            reinterpret_cast<const _ptrdiff_t *>(output_strides),        \
+            reinterpret_cast<const _ptrdiff_t *>(input_strides),         \
+            reinterpret_cast<Tdata *>(output), inputs, args...);         \
+    }
+
+#define LAUNCH_ELEMENTWISE_KERNEL_INSTANTIATE(OpName, T, ...) \
+    template void launch##OpName##Kernel<T, ##__VA_ARGS__>(   \
+        size_t output_size,                                   \
+        size_t ndim,                                          \
+        bool output_contiguous,                               \
+        const void *input_contiguous,                         \
+        const void *input_broadcasted,                        \
+        const void *output_shape,                             \
+        const void *input_shapes,                             \
+        const void *output_strides,                           \
+        const void *input_strides,                            \
+        void *output,                                         \
+        const void *const *inputs,                            \
+        XPUStream stream,                                     \
+        ##__VA_ARGS__);
+
+#endif

src/infiniop/elementwise/maca/elementwise_maca.h

+#ifndef __INFINIOP_ELEMENTWISE_MACA_H__
+#define __INFINIOP_ELEMENTWISE_MACA_H__
+
+#include "../../../utils.h"
+#include "../../devices/maca/common_maca.h"
+#include "../../devices/maca/maca_kernel_common.h"
+#include "elementwise_maca_api.h"
+
+namespace op::elementwise::maca {
+template <typename T>
+__device__ __forceinline__ const T *typedInputPtr(const void *ptr) {
+    return reinterpret_cast<const T *>(ptr);
+}
+
+__device__ __forceinline__ size_t getOutputIndex(size_t idx, bool is_contiguous, size_t ndim,
+                                                 const size_t *shape, const ptrdiff_t *strides) {
+    return is_contiguous ? idx : device::maca::indexToOffset(idx, ndim, shape, strides);
+}
+
+struct InputIndexer {
+    size_t idx;
+    size_t ndim;
+    const bool *input_contiguous;
+    const bool *input_broadcasted;
+    const size_t *input_shapes;
+    const ptrdiff_t *input_strides;
+    const ptrdiff_t *output_strides;
+
+    __device__ __forceinline__ size_t operator()(size_t input_id) const {
+        return input_contiguous[input_id]
+                 ? idx
+                 : (input_broadcasted[input_id]
+                        ? device::maca::indexToReducedOffset(idx, ndim, output_strides, input_strides + input_id * ndim)
+                        : device::maca::indexToOffset(idx, ndim, input_shapes + input_id * ndim, input_strides + input_id * ndim));
+    }
+};
+
+template <typename F, size_t... Is>
+__device__ __forceinline__ void unpackInputsAndApply(F &&f, std::index_sequence<Is...>) {
+    f(std::integral_constant<size_t, Is>{}...);
+}
+
+template <size_t N, typename Op, typename Tdata, typename... Args>
+INFINIOP_MACA_KERNEL elementwiseKernel(
+    size_t output_size,
+    size_t ndim,
+    bool output_contiguous,
+    const bool *__restrict__ input_contiguous,
+    const bool *__restrict__ input_broadcasted,
+    const size_t *__restrict__ output_shape,
+    const size_t *__restrict__ input_shapes,
+    const ptrdiff_t *__restrict__ output_strides,
+    const ptrdiff_t *__restrict__ input_strides,
+    Tdata *output,
+    const void *const *inputs,
+    size_t offset,
+    Args... args) {
+
+    size_t idx = blockIdx.x * blockDim.x + threadIdx.x + offset;
+
+    if (idx < output_size) {
+        const Tdata *const *typed_inputs = reinterpret_cast<const Tdata *const *>(inputs);
+        size_t out_idx = getOutputIndex(idx, output_contiguous, ndim, output_shape, output_strides);
+        InputIndexer indexer{idx, ndim, input_contiguous, input_broadcasted, input_shapes, input_strides, output_strides};
+
+        unpackInputsAndApply(
+            [&](auto... Is) {
+                output[out_idx] = Op{}(typed_inputs[Is.value][indexer(Is.value)]..., std::forward<Args>(args)...);
+            },
+            std::make_index_sequence<N>{});
+    }
+}
+
+template <typename Op, typename Tout, typename... Tin>
+INFINIOP_MACA_KERNEL elementwiseKernel(
+    size_t output_size,
+    size_t ndim,
+    bool output_contiguous,
+    const bool *__restrict__ input_contiguous,
+    const bool *__restrict__ input_broadcasted,
+    const size_t *__restrict__ output_shape,
+    const size_t *__restrict__ input_shapes,
+    const ptrdiff_t *__restrict__ output_strides,
+    const ptrdiff_t *__restrict__ input_strides,
+    Tout *output,
+    const void *const *__restrict__ inputs,
+    size_t offset) {
+
+    size_t idx = blockIdx.x * blockDim.x + threadIdx.x + offset;
+
+    if (idx < output_size) {
+        size_t out_idx = getOutputIndex(idx, output_contiguous, ndim, output_shape, output_strides);
+        InputIndexer indexer{idx, ndim, input_contiguous, input_broadcasted, input_shapes, input_strides, output_strides};
+
+        unpackInputsAndApply(
+            [&](auto... Is) {
+                output[out_idx] = Op{}.template operator()<Tout, Tin...>(
+                    (typedInputPtr<Tin>(inputs[Is.value])[indexer(Is.value)])...);
+            },
+            std::index_sequence_for<Tin...>{});
+    }
+}
+
+struct DeviceImpl::Opaque {
+    std::shared_ptr<device::maca::Handle::Internal> internal;
+
+    Opaque(const std::shared_ptr<device::maca::Handle::Internal> &internal)
+        : internal(internal) {}
+
+    template <uint32_t BLOCK_SIZE, size_t N, typename Op, typename Tdata, typename... Args>
+    infiniStatus_t calculateImpl(const op::elementwise::ElementwiseInfo &info,
+                                 void *workspace,
+                                 void *output,
+                                 const std::vector<const void *> &inputs,
+                                 hcStream_t stream,
+                                 Args &&...args) {
+        return launchElementwiseKernel<BLOCK_SIZE, N>(
+            info, workspace,
+            reinterpret_cast<Tdata *>(output), inputs,
+            elementwiseKernel<N, Op, Tdata, Args...>,
+            stream,
+            std::forward<Args>(args)...);
+    }
+
+    template <uint32_t BLOCK_SIZE, size_t N, typename Op, typename Tout, typename... Tin, typename... Args,
+              std::enable_if_t<(sizeof...(Tin) == Op::num_inputs), int> = 0>
+    infiniStatus_t calculateImpl(const op::elementwise::ElementwiseInfo &info,
+                                 void *workspace,
+                                 void *output,
+                                 const std::vector<const void *> &inputs,
+                                 hcStream_t stream,
+                                 Args &&...args) {
+        return launchElementwiseKernel<BLOCK_SIZE, N>(
+            info, workspace,
+            reinterpret_cast<Tout *>(output), inputs,
+            elementwiseKernel<Op, Tout, Tin...>,
+            stream);
+    }
+
+private:
+    template <size_t N>
+    infiniStatus_t infoToDevice(
+        const op::elementwise::ElementwiseInfo &info,
+        void *workspace,
+        const void *const *h_inputs_arr,
+        const void **&d_inputs_arr,
+        const bool *&d_input_contiguous,
+        const bool *&d_input_broadcasted,
+        const size_t *&d_output_shape,
+        const ptrdiff_t *&d_output_strides,
+        const size_t *&d_input_shapes,
+        const ptrdiff_t *&d_input_strides,
+        hcStream_t stream) const {
+
+        constexpr auto input_size = N;
+        const auto ndim = info.getNdim();
+        constexpr auto input_arr_size = N * sizeof(*h_inputs_arr);
+        const int8_t *info_meta_start = info.getMetaStart();
+        const int8_t *d_meta_start = reinterpret_cast<int8_t *>(workspace) + input_arr_size;
+
+        // copy the input pointer array and meta to device
+        CHECK_MACA(hcMemcpyAsync(workspace, h_inputs_arr, input_arr_size, hcMemcpyHostToDevice, stream));
+        CHECK_MACA(hcMemcpyAsync((void *)d_meta_start, info_meta_start, info.getMetaMemSize(), hcMemcpyHostToDevice, stream));
+
+        // offset/assign the pointers
+        d_inputs_arr = reinterpret_cast<const void **>(workspace);
+        d_output_shape = reinterpret_cast<const size_t *>(d_meta_start);
+        d_output_strides = reinterpret_cast<const ptrdiff_t *>(d_output_shape + ndim);
+        d_input_shapes = reinterpret_cast<const size_t *>(d_output_strides + ndim);
+        d_input_strides = reinterpret_cast<const ptrdiff_t *>(d_input_shapes + input_size * ndim);
+        d_input_contiguous = reinterpret_cast<const bool *>(d_input_strides + input_size * ndim);
+        d_input_broadcasted = reinterpret_cast<const bool *>(d_input_contiguous + input_size);
+
+        return INFINI_STATUS_SUCCESS;
+    }
+
+    template <uint32_t BLOCK_SIZE, size_t N, typename KernelFunc, typename Tout, typename... Args>
+    infiniStatus_t launchElementwiseKernel(
+        const op::elementwise::ElementwiseInfo &info,
+        void *workspace,
+        Tout *output,
+        const std::vector<const void *> &inputs,
+        KernelFunc kernel_func,
+        hcStream_t stream,
+        Args &&...args) {
+
+        auto output_size = info.getOutputSize();
+        if (output_size == 0) {
+            return INFINI_STATUS_SUCCESS;
+        }
+
+        // Device pointers
+        const void **d_inputs_arr = nullptr;
+        const bool *d_input_contiguous = nullptr;
+        const bool *d_input_broadcasted = nullptr;
+        const size_t *d_output_shape = nullptr;
+        const ptrdiff_t *d_output_strides = nullptr;
+        const size_t *d_input_shapes = nullptr;
+        const ptrdiff_t *d_input_strides = nullptr;
+
+        CHECK_STATUS(infoToDevice<N>(info, workspace, inputs.data(), d_inputs_arr,
+                                     d_input_contiguous, d_input_broadcasted,
+                                     d_output_shape, d_output_strides,
+                                     d_input_shapes, d_input_strides, stream));
+
+        dim3 blockDims(std::min(BLOCK_SIZE, static_cast<uint32_t>(internal->maxThreadsPerBlock())));
+        dim3 gridDims(std::min(uint32_t(CEIL_DIV(output_size, blockDims.x)), static_cast<uint32_t>(internal->gridSizeX())));
+        size_t step = gridDims.x * blockDims.x;
+
+        for (size_t i = 0; i < output_size; i += step) {
+            kernel_func<<<gridDims, blockDims, 0, stream>>>(
+                output_size, info.getNdim(), info.isOutputContiguous(),
+                d_input_contiguous, d_input_broadcasted,
+                d_output_shape, d_input_shapes,
+                d_output_strides, d_input_strides,
+                output, reinterpret_cast<const void **>(d_inputs_arr),
+                i, std::forward<Args>(args)...);
+        }
+
+        return INFINI_STATUS_SUCCESS;
+    }
+};
+
+template <typename... Args>
+utils::Result<DeviceImpl *> DeviceImpl::create(Args &&...args) {
+    auto opaque = std::make_shared<Opaque>(std::forward<Args>(args)...);
+    return utils::Result<DeviceImpl *>(new DeviceImpl(opaque));
+}
+
+/* Invoke elementwise operation for different input types */
+template <uint32_t BLOCK_SIZE, typename Op, typename Tout, typename... Tin, typename... Args,
+          std::enable_if_t<(sizeof...(Tin) == Op::num_inputs), int>>
+infiniStatus_t DeviceImpl::calculate(const op::elementwise::ElementwiseInfo &info,
+                                     void *workspace,
+                                     void *output,
+                                     const std::vector<const void *> &inputs,
+                                     void *stream,
+                                     Args &&...args) {
+    constexpr size_t N = Op::num_inputs;
+    static_assert(sizeof...(Tin) == N, "Input type count mismatch");
+    return _opaque->calculateImpl<BLOCK_SIZE, N, Op, Tout, Tin...>(
+        info, workspace, output, inputs,
+        reinterpret_cast<hcStream_t>(stream),
+        std::forward<Args>(args)...);
+}
+
+/* Invoke elementwise operation when all inputs have the same dtype */
+template <uint32_t BLOCK_SIZE, typename Op, typename Tdata, typename... Args>
+infiniStatus_t DeviceImpl::calculate(const op::elementwise::ElementwiseInfo &info,
+                                     void *workspace,
+                                     void *output,
+                                     const std::vector<const void *> &inputs,
+                                     void *stream,
+                                     Args &&...args) {
+    constexpr size_t N = Op::num_inputs;
+    return _opaque->calculateImpl<BLOCK_SIZE, N, Op, Tdata>(
+        info, workspace, output, inputs,
+        reinterpret_cast<hcStream_t>(stream),
+        std::forward<Args>(args)...);
+}
+
+} // namespace op::elementwise::maca
+
+#endif

src/infiniop/elementwise/maca/elementwise_maca_api.h

+#ifndef __INFINIOP_ELEMENTWISE_MACA_API_H__
+#define __INFINIOP_ELEMENTWISE_MACA_API_H__
+
+#include "../elementwise.h"
+
+namespace op::elementwise::maca {
+
+class DeviceImpl final {
+    struct Opaque;
+    std::shared_ptr<Opaque> _opaque;
+
+    DeviceImpl(std::shared_ptr<Opaque> opaque) : _opaque(std::move(opaque)) {}
+
+public:
+    ~DeviceImpl() = default;
+
+    template <typename... Args>
+    static utils::Result<DeviceImpl *> create(Args &&...args);
+
+    template <uint32_t BLOCK_SIZE, typename Op, typename Tdata, typename... Args>
+    infiniStatus_t calculate(
+        const op::elementwise::ElementwiseInfo &info,
+        void *workspace,
+        void *output,
+        const std::vector<const void *> &inputs,
+        void *stream,
+        Args &&...args);
+
+    template <uint32_t BLOCK_SIZE, typename Op, typename Tout, typename... Tin,
+              typename... Args,
+              std::enable_if_t<(sizeof...(Tin) == Op::num_inputs), int> = 0>
+    infiniStatus_t calculate(
+        const op::elementwise::ElementwiseInfo &info,
+        void *workspace,
+        void *output,
+        const std::vector<const void *> &inputs,
+        void *stream,
+        Args &&...args);
+};
+} // namespace op::elementwise::maca
+#define CREATE_ELEMENTWISE_MACA_DESCRIPTOR(HANDLE, DTYPE, OUT_DESC, INPUT_DESC_VEC)          \
+                                                                                             \
+    auto info_result = op::elementwise::ElementwiseInfo::create(OUT_DESC, INPUT_DESC_VEC);   \
+    CHECK_RESULT(info_result);                                                               \
+    auto info = info_result.take();                                                          \
+    auto workspace_size = info.getMetaMemSize() + info.getInputSize() * sizeof(void *);      \
+                                                                                             \
+    auto device_impl_result = op::elementwise::maca::DeviceImpl::create(HANDLE->internal()); \
+    CHECK_RESULT(device_impl_result);                                                        \
+                                                                                             \
+    *desc_ptr = new Descriptor(                                                              \
+        DTYPE,                                                                               \
+        std::move(info),                                                                     \
+        std::move(device_impl_result.take()),                                                \
+        workspace_size,                                                                      \
+        HANDLE->device,                                                                      \
+        HANDLE->device_id);
+
+#endif // __INFINIOP_ELEMENTWISE_MACA_API_H__

src/infiniop/ops/attention/attention.h

+#ifndef ATTENTION_H
+#define ATTENTION_H
+
+#include "../../operator.h"
+#include "info.h"
+
+#define DESCRIPTOR(NAMESPACE)                                    \
+                                                                 \
+    namespace op::attention::NAMESPACE {                         \
+    class Descriptor final : public InfiniopDescriptor {         \
+        struct Opaque;                                           \
+        Opaque *_opaque;                                         \
+        size_t _workspace_size;                                  \
+                                                                 \
+        Descriptor(                                              \
+            Opaque *opaque,                                      \
+            size_t workspace_size,                               \
+            infiniDevice_t device_type,                          \
+            int device_id)                                       \
+            : InfiniopDescriptor{device_type, device_id},        \
+              _opaque(opaque),                                   \
+              _workspace_size(workspace_size) {}                 \
+                                                                 \
+    public:                                                      \
+        ~Descriptor();                                           \
+                                                                 \
+        size_t workspaceSize() const { return _workspace_size; } \
+                                                                 \
+        static infiniStatus_t create(                            \
+            infiniopHandle_t handle,                             \
+            Descriptor **desc_ptr,                               \
+            infiniopTensorDescriptor_t y_desc,                   \
+            infiniopTensorDescriptor_t x_desc);                  \
+    };                                                           \
+    }
+
+#endif // ATTENTION_H

src/infiniop/ops/attention/operator.cc

+#include "../../operator.h"
+#include "../../../utils.h"
+#include "../../../utils/check.h"
+#include "../../handle.h"
+#include "../../tensor.h"
+#include "infiniop/ops/attention.h"
+#include "infiniop/ops/causal_softmax.h"
+#include "infiniop/ops/gemm.h"
+#include "infiniop/ops/rearrange.h"
+
+#include <cmath>
+#include <cstdint>
+
+struct InfiniopAttentionDescriptor {
+    InfiniopDescriptor _super;
+    infiniopRearrangeDescriptor_t rearrange_desc_k;
+    infiniopRearrangeDescriptor_t rearrange_desc_v;
+    infiniopRearrangeDescriptor_t rearrange_desc_q;
+    infiniopRearrangeDescriptor_t rearrange_desc_out;
+    infiniopGemmDescriptor_t matmul_desc1;
+    infiniopGemmDescriptor_t matmul_desc2;
+    infiniopCausalSoftmaxDescriptor_t softmax_desc;
+    size_t workspace_size;
+    size_t op_workspace_offset;
+    size_t op_workspace_size;
+    size_t q_cont_offset;
+    size_t att_score_offset;
+    size_t att_val_offset;
+    size_t k_cache_offset;
+    size_t v_cache_offset;
+    float qk_alpha;
+};
+
+__C __export infiniStatus_t infiniopCreateAttentionDescriptor(infiniopHandle_t handle,
+                                                              infiniopAttentionDescriptor_t *desc_ptr,
+                                                              infiniopTensorDescriptor_t out_desc,
+                                                              infiniopTensorDescriptor_t q_desc,
+                                                              infiniopTensorDescriptor_t k_desc,
+                                                              infiniopTensorDescriptor_t v_desc,
+                                                              infiniopTensorDescriptor_t k_cache_desc,
+                                                              infiniopTensorDescriptor_t v_cache_desc,
+                                                              size_t pos) {
+    if (out_desc->ndim() != 3 || q_desc->ndim() != 3 || k_desc->ndim() != 3 || v_desc->ndim() != 3 || k_cache_desc->ndim() != 3 || v_cache_desc->ndim() != 3) {
+        return INFINI_STATUS_BAD_TENSOR_SHAPE;
+    }
+
+    if (!out_desc->isContiguous(0, 2)) {
+        return INFINI_STATUS_BAD_TENSOR_STRIDES;
+    }
+
+    if (q_desc->strides()[2] != 1 || k_desc->strides()[2] != 1 || v_desc->strides()[2] != 1 || k_cache_desc->strides()[2] != 1 || v_cache_desc->strides()[2] != 1) {
+        return INFINI_STATUS_BAD_TENSOR_STRIDES;
+    }
+
+    size_t n_q_head = q_desc->shape()[0];
+    size_t seq_len = q_desc->shape()[1];
+    size_t head_dim = q_desc->shape()[2];
+    size_t hidden_size = n_q_head * head_dim;
+    size_t n_kv_head = k_desc->shape()[0];
+    size_t total_seq_len = seq_len + pos;
+    size_t n_group = n_q_head / n_kv_head;
+    size_t alignment = 256;
+
+    if (out_desc->shape()[0] != seq_len || out_desc->shape()[1] != n_q_head || out_desc->shape()[2] != head_dim) {
+        return INFINI_STATUS_BAD_PARAM;
+    }
+
+    // k: [n_kv_head, seq_len, head_dim]
+    if (k_desc->shape()[0] != n_kv_head || k_desc->shape()[1] != seq_len || k_desc->shape()[2] != head_dim) {
+        return INFINI_STATUS_BAD_PARAM;
+    }
+
+    // v: [n_kv_head, seq_len, head_dim]
+    if (v_desc->shape()[0] != n_kv_head || v_desc->shape()[1] != seq_len || v_desc->shape()[2] != head_dim) {
+        return INFINI_STATUS_BAD_PARAM;
+    }
+
+    // k_cache: [n_kv_head, _, head_dim]
+    if (k_cache_desc->shape()[0] != n_kv_head || k_cache_desc->shape()[1] < total_seq_len || k_cache_desc->shape()[2] != head_dim) {
+        return INFINI_STATUS_BAD_PARAM;
+    }
+
+    // v_cache: [n_kv_head, _, head_dim]
+    if (v_cache_desc->shape()[0] != n_kv_head || v_cache_desc->shape()[1] < total_seq_len || v_cache_desc->shape()[2] != head_dim) {
+        return INFINI_STATUS_BAD_PARAM;
+    }
+
+    // Rearrange k into k_cache
+    infiniopTensorDescriptor_t dst_k_desc;
+    CHECK_STATUS(infiniopCreateTensorDescriptor(&dst_k_desc, 3, k_desc->shape().data(), k_cache_desc->strides().data(), k_cache_desc->dtype()));
+    infiniopRearrangeDescriptor_t rearrange_desc_k;
+    CHECK_STATUS(infiniopCreateRearrangeDescriptor(handle, &rearrange_desc_k, dst_k_desc, k_desc));
+
+    // Rearrange v into v_cache
+    infiniopTensorDescriptor_t dst_v_desc;
+    CHECK_STATUS(infiniopCreateTensorDescriptor(&dst_v_desc, 3, v_desc->shape().data(), v_cache_desc->strides().data(), v_cache_desc->dtype()));
+    infiniopRearrangeDescriptor_t rearrange_desc_v;
+    CHECK_STATUS(infiniopCreateRearrangeDescriptor(handle, &rearrange_desc_v, dst_v_desc, v_desc));
+
+    infiniopRearrangeDescriptor_t rearrange_desc_q = nullptr;
+    size_t q_cont_size = 0;
+    infiniopTensorDescriptor_t rearranged_q_desc;
+    // Rearrange q into contiguous
+    if (!q_desc->isContiguous(0, 1)) {
+        CHECK_STATUS(infiniopCreateTensorDescriptor(&rearranged_q_desc, 3, q_desc->shape().data(), nullptr, q_desc->dtype()));
+        q_cont_size = utils::align(rearranged_q_desc->numel() * infiniSizeOf(rearranged_q_desc->dtype()), alignment);
+        rearrange_desc_q = new InfiniopDescriptor;
+        CHECK_STATUS(infiniopCreateRearrangeDescriptor(handle, &rearrange_desc_q, rearranged_q_desc, q_desc));
+    }
+
+    // Matmul1: q * full_k
+    //      q: [n_q_head, seq_len, head_dim] -> [n_kv_head, n_group *seq_len, head_dim]
+    infiniopTensorDescriptor_t reshaped_q_desc;
+    CHECK_STATUS(infiniopCreateTensorDescriptor(&reshaped_q_desc, 3, q_desc->shape().data(), nullptr, q_desc->dtype()));
+    TRANSFORM_TENSOR_DESC(reshaped_q_desc, dimSplit(0, {n_kv_head, n_group}));
+    TRANSFORM_TENSOR_DESC(reshaped_q_desc, dimMerge(1, 2));
+    //      full_k: [n_kv_head, head_dim, total_seq_len]
+    infiniopTensorDescriptor_t full_k_desc;
+    size_t full_k_shape[3] = {n_kv_head, total_seq_len, head_dim};
+    CHECK_STATUS(infiniopCreateTensorDescriptor(&full_k_desc, 3, full_k_shape, k_cache_desc->strides().data(), k_cache_desc->dtype()));
+    TRANSFORM_TENSOR_DESC(full_k_desc, dimPermute({0, 2, 1}));
+    //      qk: [n_kv_head, n_group * seq_len, total_seq_len]
+    infiniopTensorDescriptor_t qk_desc;
+    size_t qk_shape[3] = {n_kv_head, n_group * seq_len, total_seq_len};
+    CHECK_STATUS(infiniopCreateTensorDescriptor(&qk_desc, 3, qk_shape, nullptr, q_desc->dtype()));
+    //      matmul1_desc
+    //          qk_alpha
+    float qk_alpha = 1 / sqrt(head_dim);
+    infiniopGemmDescriptor_t matmul1_desc;
+    CHECK_STATUS(infiniopCreateGemmDescriptor(handle, &matmul1_desc, qk_desc, reshaped_q_desc, full_k_desc));
+    //      matmul1 workspace size
+    size_t matmul1_workspace_size;
+    CHECK_STATUS(infiniopGetGemmWorkspaceSize(matmul1_desc, &matmul1_workspace_size));
+    matmul1_workspace_size = utils::align(matmul1_workspace_size, alignment);
+    //      attention score tensor size
+    size_t attn_score_size = utils::align(qk_desc->numel() * infiniSizeOf(qk_desc->dtype()), alignment);
+
+    // CausalSoftmax: softmax(qk)
+    //      qk: [n_kv_head, n_group * seq_len, total_seq_len] -> [n_q_head, seq_len, total_seq_len]
+    TRANSFORM_TENSOR_DESC(qk_desc, dimSplit(1, {n_group, seq_len}));
+    TRANSFORM_TENSOR_DESC(qk_desc, dimMerge(0, 1));
+    infiniopCausalSoftmaxDescriptor_t softmax_desc;
+    CHECK_STATUS(infiniopCreateCausalSoftmaxDescriptor(handle, &softmax_desc, qk_desc, qk_desc));
+    //      softmax workspace size
+    size_t softmax_workspace_size;
+    CHECK_STATUS(infiniopGetCausalSoftmaxWorkspaceSize(softmax_desc, &softmax_workspace_size));
+    softmax_workspace_size = utils::align(softmax_workspace_size, alignment);
+
+    // Matmul2: softmax(qk) * full_v
+    //      softmax(qk): [n_q_head, seq_len, total_seq_len] -> [n_kv_head, n_group * seq_len, total_seq_len]
+    //      full_v: [n_kv_head, total_seq_len, head_dim]
+    TRANSFORM_TENSOR_DESC(qk_desc, dimSplit(0, {n_kv_head, n_group}));
+    TRANSFORM_TENSOR_DESC(qk_desc, dimMerge(1, 2));
+    infiniopTensorDescriptor_t full_v_desc;
+    size_t full_v_shape[3] = {n_kv_head, total_seq_len, head_dim};
+    CHECK_STATUS(infiniopCreateTensorDescriptor(&full_v_desc, 3, full_v_shape, v_cache_desc->strides().data(), v_cache_desc->dtype()));
+    //      temp_out: [n_kv_head, n_group * seq_len, head_dim]
+    infiniopTensorDescriptor_t att_val_desc;
+    size_t temp_out_shape[3] = {n_kv_head, n_group * seq_len, head_dim};
+    CHECK_STATUS(infiniopCreateTensorDescriptor(&att_val_desc, 3, temp_out_shape, nullptr, q_desc->dtype()));
+    //      matmul2_desc
+    infiniopGemmDescriptor_t matmul2_desc;
+    CHECK_STATUS(infiniopCreateGemmDescriptor(handle, &matmul2_desc, att_val_desc, qk_desc, full_v_desc));
+    //      matmul2 workspace size
+    size_t matmul2_workspace_size;
+    CHECK_STATUS(infiniopGetGemmWorkspaceSize(matmul2_desc, &matmul2_workspace_size));
+    matmul2_workspace_size = utils::align(matmul2_workspace_size, alignment);
+    //      attention value tensor size
+    size_t att_val_size = utils::align(att_val_desc->numel() * infiniSizeOf(att_val_desc->dtype()), alignment);
+
+    // Rearrange temp_out into out
+    //      out: [seq_len, n_q_head, head_dim]
+    //      temp_out: [n_kv_head, n_group * seq_len, head_dim] -> [n_q_head, seq_len, head_dim] -> [seq_len, n_q_head, head_dim]
+    TRANSFORM_TENSOR_DESC(att_val_desc, dimSplit(1, {n_group, seq_len}));
+    TRANSFORM_TENSOR_DESC(att_val_desc, dimMerge(0, 1));
+    TRANSFORM_TENSOR_DESC(att_val_desc, dimPermute({1, 0, 2}));
+    infiniopRearrangeDescriptor_t rearrange_desc_out;
+    CHECK_STATUS(infiniopCreateRearrangeDescriptor(handle, &rearrange_desc_out, out_desc, att_val_desc));
+
+    // workspace size
+    size_t op_workspace_size = utils::align(std::max(std::max(matmul1_workspace_size, matmul2_workspace_size), softmax_workspace_size), alignment);
+    size_t temp_tensors_size = attn_score_size + std::max(q_cont_size, att_val_size);
+    size_t workspace_size = temp_tensors_size + op_workspace_size;
+
+    // k_cache_offset
+    size_t k_cache_offset = 0;
+    if (pos > 0) {
+        k_cache_offset = pos * k_cache_desc->getByteStrides()[1];
+    }
+
+    // v_cache_offset
+    size_t v_cache_offset = 0;
+    if (pos > 0) {
+        v_cache_offset = pos * v_cache_desc->getByteStrides()[1];
+    }
+
+    // create attention descriptor
+    *(InfiniopAttentionDescriptor **)desc_ptr = new InfiniopAttentionDescriptor{
+        {handle->device, handle->device_id},
+        rearrange_desc_k,
+        rearrange_desc_v,
+        rearrange_desc_q,
+        rearrange_desc_out,
+        matmul1_desc,
+        matmul2_desc,
+        softmax_desc,
+        workspace_size,
+        temp_tensors_size,
+        op_workspace_size,
+        attn_score_size,
+        0,
+        attn_score_size,
+        k_cache_offset,
+        v_cache_offset,
+        1.f / std::sqrt(float(head_dim)),
+    };
+
+    return INFINI_STATUS_SUCCESS;
+}
+
+__C __export infiniStatus_t infiniopGetAttentionWorkspaceSize(infiniopAttentionDescriptor_t desc, size_t *size) {
+    *size = ((InfiniopAttentionDescriptor *)desc)->workspace_size;
+    return INFINI_STATUS_SUCCESS;
+}
+
+__C __export infiniStatus_t infiniopAttention(infiniopAttentionDescriptor_t desc_,
+                                              void *workspace_,
+                                              size_t workspace_size_,
+                                              void *out,
+                                              void const *q,
+                                              void const *k,
+                                              void const *v,
+                                              void *k_cache,
+                                              void *v_cache,
+                                              void *stream) {
+    auto desc = (InfiniopAttentionDescriptor *)desc_;
+    if (workspace_size_ < desc->workspace_size) {
+        return INFINI_STATUS_INSUFFICIENT_WORKSPACE; // STATUS_MEMORY_NOT_ALLOCATED
+    }
+    void *workspace = (char *)workspace_ + desc->op_workspace_offset;
+    size_t workspace_size = desc->op_workspace_size;
+    void *att_score = (char *)workspace_ + desc->att_score_offset;
+    void *att_val = (char *)workspace_ + desc->att_val_offset;
+    void const *q_ = q;
+    // concat k and v to k_cache and v_cache
+    CHECK_STATUS(infiniopRearrange(desc->rearrange_desc_k,
+                                   (char *)k_cache + desc->k_cache_offset, k, stream));
+
+    CHECK_STATUS(infiniopRearrange(desc->rearrange_desc_v,
+                                   (char *)v_cache + desc->v_cache_offset, v, stream));
+
+    // rearrange q into contiguous
+    if (desc->rearrange_desc_q) {
+        void *q_cont = (char *)workspace_ + desc->q_cont_offset;
+        CHECK_STATUS(infiniopRearrange(desc->rearrange_desc_q, q_cont, q, stream));
+        q_ = q_cont;
+    }
+
+    // matmul1: q * full_k
+    CHECK_STATUS(infiniopGemm(desc->matmul_desc1,
+                              workspace, workspace_size,
+                              att_score, q_, k_cache, desc->qk_alpha, 0.0, stream));
+    // softmax(qk)
+    CHECK_STATUS(infiniopCausalSoftmax(desc->softmax_desc,
+                                       workspace, workspace_size,
+                                       att_score, att_score, stream));
+    // matmul2: softmax(qk) * full_v
+    CHECK_STATUS(infiniopGemm(desc->matmul_desc2,
+                              workspace, workspace_size,
+                              att_val, att_score, v_cache, 1.0, 0.0, stream));
+    // rearrange out
+    CHECK_STATUS(infiniopRearrange(desc->rearrange_desc_out, out, att_val, stream));
+
+    return INFINI_STATUS_SUCCESS;
+}
+
+__C __export infiniStatus_t infiniopDestroyAttentionDescriptor(infiniopAttentionDescriptor_t desc_) {
+    auto desc = (InfiniopAttentionDescriptor *)desc_;
+    if (desc->rearrange_desc_q) {
+        CHECK_STATUS(infiniopDestroyRearrangeDescriptor(desc->rearrange_desc_q));
+    }
+    CHECK_STATUS(infiniopDestroyRearrangeDescriptor(desc->rearrange_desc_k));
+    CHECK_STATUS(infiniopDestroyRearrangeDescriptor(desc->rearrange_desc_v));
+    CHECK_STATUS(infiniopDestroyRearrangeDescriptor(desc->rearrange_desc_out));
+    CHECK_STATUS(infiniopDestroyGemmDescriptor(desc->matmul_desc1));
+    CHECK_STATUS(infiniopDestroyGemmDescriptor(desc->matmul_desc2));
+    CHECK_STATUS(infiniopDestroyCausalSoftmaxDescriptor(desc->softmax_desc));
+    delete desc;
+
+    return INFINI_STATUS_SUCCESS;
+}

src/infiniop/ops/causal_softmax/ascend/causal_softmax_aclnn.cc → src/infiniop/ops/causal_softmax/ascend/causal_softmax_ascend.cc

-#include "causal_softmax_aclnn.h"
+#include "causal_softmax_ascend.h"
 #include "../../../devices/ascend/common_ascend.h"
 #include <aclnnop/aclnn_masked_fill_tensor.h>
 #include <aclnnop/aclnn_softmax.h>
@@ -12,6 +12,8 @@ struct Descriptor::Opaque {
     aclnnTensorDescriptor_t value;
     void *mask_addr;
     void *value_addr;
+    uint64_t workspacesize;
+    aclOpExecutor *executor;
 
     ~Opaque() {
         delete x;
@@ -21,6 +23,9 @@ struct Descriptor::Opaque {
 
         aclrtFree(mask_addr);
         aclrtFree(value_addr);
+
+        // Delete useless executor
+        aclDestroyAclOpExecutor(executor);
     }
 };
 
@@ -92,18 +97,18 @@ infiniStatus_t Descriptor::create(
     aclTensor *tvalue = value->tensor;
 
     CHECK_ACL(aclnnInplaceMaskedFillTensorGetWorkspaceSize(tx, tmask, tvalue, &workspacesize_mask, &mask_executor));
-    int64_t dim = 2;
 
+    int64_t dim = 2;
     CHECK_ACL(aclnnSoftmaxGetWorkspaceSize(tx, dim, ty, &workspacesize_softmax, &executor));
+    // set executor reusable
+    aclSetAclOpExecutorRepeatable(executor);
 
-    // Create the descriptor
-    size_t all_workspacesize = workspacesize_softmax + workspacesize_mask;
-    *desc_ptr = new Descriptor(new Opaque{x, mask, y, value, mask_addr, value_addr},
-                               std::move(info), all_workspacesize, handle_ascend->device, handle_ascend->device_id);
+    // Create the descripto
+    size_t all_workspacesize = std::max(workspacesize_softmax, workspacesize_mask);
 
-    // Delete useless executor
-    aclDestroyAclOpExecutor(executor);
-    aclDestroyAclOpExecutor(mask_executor);
+    *desc_ptr = new Descriptor(new Opaque{x, mask, y, value, mask_addr, value_addr,
+                                          workspacesize_softmax, executor},
+                               std::move(info), all_workspacesize, handle_ascend->device, handle_ascend->device_id);
 
     return INFINI_STATUS_SUCCESS;
 }
@@ -116,23 +121,18 @@ infiniStatus_t Descriptor::calculate(void *workspace, size_t workspace_size, voi
     auto ty = _opaque->y->tensor;
     auto tmask = _opaque->mask->tensor;
     auto tvalue = _opaque->value->tensor;
-    aclOpExecutor *executor = nullptr;
     aclOpExecutor *mask_executor = nullptr;
-    size_t workspacesize_softmax = 0;
     size_t workspacesize_mask = 0;
-    int64_t dim = 2;
 
     AclSetTensorAddr(mask_executor, 0, tx, (void *)x);
     AclSetTensorAddr(mask_executor, 1, tmask, _opaque->mask_addr);
     AclSetTensorAddr(mask_executor, 2, tvalue, _opaque->value_addr);
     CHECK_ACL(aclnnInplaceMaskedFillTensorGetWorkspaceSize(tx, tmask, tvalue, &workspacesize_mask, &mask_executor));
     CHECK_ACL(aclnnInplaceMaskedFillTensor(workspace, workspacesize_mask, mask_executor, stream));
-    CHECK_ACL(aclrtSynchronizeStream(stream));
 
-    AclSetTensorAddr(executor, 0, tx, (void *)x);
-    AclSetTensorAddr(executor, 1, ty, y);
-    CHECK_ACL(aclnnSoftmaxGetWorkspaceSize(tx, dim, ty, &workspacesize_softmax, &executor));
-    CHECK_ACL(aclnnSoftmax(workspace, workspacesize_softmax, executor, stream));
+    AclSetTensorAddr(_opaque->executor, 0, tx, (void *)x);
+    AclSetTensorAddr(_opaque->executor, 1, ty, y);
+    CHECK_ACL(aclnnSoftmax(workspace, _opaque->workspacesize, _opaque->executor, stream));
 
     return INFINI_STATUS_SUCCESS;
 }

src/infiniop/ops/causal_softmax/cpu/causal_softmax_cpu.cc

@@ -48,7 +48,7 @@ infiniStatus_t causal_softmax(const CausalSoftmaxInfo *info, T *y, const T *x) {
             if constexpr (std::is_same<T, fp16_t>::value) {
                 y_[j * info->y_stride_j] = utils::cast<fp16_t>(utils::cast<float>(y_[j * info->y_stride_j]) / sum);
             } else {
-                y_[j * info->y_stride_j] = y_[y_offset + j * info->y_stride_j] / sum;
+                y_[j * info->y_stride_j] = y_[j * info->y_stride_j] / sum;
             }
         }
     }

src/infiniop/ops/causal_softmax/cuda/causal_softmax_kernel.cuh

@@ -18,7 +18,7 @@ INFINIOP_CUDA_KERNEL causalSoftmax(
 
     // [Reduce] Find max value in each row and store in shared memory
     __shared__ Tdata max_;
-    Tdata max_0 = op::common_cuda::reduce_op::max<BLOCK_SIZE, Tdata>(x, width);
+    Tdata max_0 = op::common_cuda::reduce_op::max<BLOCK_SIZE, Tdata>(x, width - height + 1 + blockIdx.x);
     if (threadIdx.x == 0) {
         max_ = max_0;
     }

src/infiniop/ops/causal_softmax/operator.cc

@@ -9,7 +9,7 @@
 #include "cuda/causal_softmax_cuda.cuh"
 #endif
 #ifdef ENABLE_ASCEND_API
-#include "ascend/causal_softmax_aclnn.h"
+#include "ascend/causal_softmax_ascend.h"
 #endif
 
 __C infiniStatus_t infiniopCreateCausalSoftmaxDescriptor(

src/infiniop/ops/clip/cpu/clip_cpu.cc

+#include "clip_cpu.h"
+
+namespace op::clip::cpu {
+
+Descriptor::~Descriptor() = default;
+
+infiniStatus_t Descriptor::create(
+    infiniopHandle_t handle_,
+    Descriptor **desc_ptr,
+    infiniopTensorDescriptor_t out_desc,
+    std::vector<infiniopTensorDescriptor_t> input_desc_vec) {
+
+    auto handle = reinterpret_cast<device::cpu::Handle *>(handle_);
+    auto dtype = out_desc->dtype();
+
+    const auto &in_desc = input_desc_vec.at(0);
+    const auto &min_desc = input_desc_vec.at(1);
+    const auto &max_desc = input_desc_vec.at(2);
+    const auto &out_shape = out_desc->shape();
+    const auto &in_shape = in_desc->shape();
+    const auto &min_shape = min_desc->shape();
+    const auto &max_shape = max_desc->shape();
+
+    CHECK_DTYPE(dtype, INFINI_DTYPE_F16, INFINI_DTYPE_F32, INFINI_DTYPE_F64);
+    CHECK_SAME_SHAPE(out_shape, in_shape);
+    CHECK_SAME_SHAPE(out_shape, min_shape);
+    CHECK_SAME_SHAPE(out_shape, max_shape);
+
+    CREATE_ELEMENTWISE_CPU_DESCRIPTOR(handle, dtype, out_desc, input_desc_vec);
+
+    return INFINI_STATUS_SUCCESS;
+}
+
+infiniStatus_t Descriptor::calculate(
+    void *workspace,
+    size_t workspace_size,
+    void *output,
+    std::vector<const void *> inputs,
+    void *stream) const {
+
+    switch (_dtype) {
+    case INFINI_DTYPE_F16:
+        return _device_info->calculate<ClipOp, fp16_t>(_info, output, inputs, stream);
+    case INFINI_DTYPE_F32:
+        return _device_info->calculate<ClipOp, float>(_info, output, inputs, stream);
+    case INFINI_DTYPE_F64:
+        return _device_info->calculate<ClipOp, double>(_info, output, inputs, stream);
+    default:
+        return INFINI_STATUS_BAD_TENSOR_DTYPE;
+    }
+
+    return INFINI_STATUS_SUCCESS;
+}
+} // namespace op::clip::cpu

src/infiniop/ops/clip/cpu/clip_cpu.h

+#ifndef __CLIP_CPU_H__
+#define __CLIP_CPU_H__
+
+#include "../../../elementwise/cpu/elementwise_cpu.h"
+#include "infiniop/ops/clip.h"
+
+ELEMENTWISE_DESCRIPTOR(clip, cpu)
+
+namespace op::clip::cpu {
+
+typedef struct ClipOp {
+public:
+    static constexpr size_t num_inputs = 3;
+
+    template <typename T>
+    T operator()(const T &x, const T &min_val, const T &max_val) const {
+        return std::max(std::min(x, max_val), min_val);
+    }
+} ClipOp;
+
+} // namespace op::clip::cpu
+
+#endif // __CLIP_CPU_H__

src/infiniop/ops/clip/cuda/clip_cuda.cu

+#include "clip_cuda.cuh"
+#include "clip_cuda_internal.cuh"
+
+namespace op::clip::cuda {
+
+Descriptor::~Descriptor() = default;
+
+infiniStatus_t Descriptor::create(
+    infiniopHandle_t handle_,
+    Descriptor **desc_ptr,
+    infiniopTensorDescriptor_t out_desc,
+    std::vector<infiniopTensorDescriptor_t> input_desc_vec) {
+
+    auto handle = reinterpret_cast<device::cuda::Handle *>(handle_);
+    auto dtype = out_desc->dtype();
+
+    const auto &in_desc = input_desc_vec.at(0);
+    const auto &min_desc = input_desc_vec.at(1);
+    const auto &max_desc = input_desc_vec.at(2);
+    const auto &out_shape = out_desc->shape();
+    const auto &in_shape = in_desc->shape();
+    const auto &min_shape = min_desc->shape();
+    const auto &max_shape = max_desc->shape();
+
+    CHECK_DTYPE(dtype, INFINI_DTYPE_F16, INFINI_DTYPE_F32, INFINI_DTYPE_F64);
+    CHECK_SAME_SHAPE(out_shape, in_shape);
+    CHECK_SAME_SHAPE(out_shape, min_shape);
+    CHECK_SAME_SHAPE(out_shape, max_shape);
+
+    CREATE_ELEMENTWISE_CUDA_DESCRIPTOR(handle, dtype, out_desc, input_desc_vec);
+
+    return INFINI_STATUS_SUCCESS;
+}
+
+infiniStatus_t Descriptor::calculate(
+    void *workspace,
+    size_t workspace_size,
+    void *output,
+    std::vector<const void *> inputs,
+    void *stream) const {
+
+    if (workspace_size < _workspace_size) {
+        return INFINI_STATUS_INSUFFICIENT_WORKSPACE;
+    }
+
+    switch (_dtype) {
+    case INFINI_DTYPE_F16:
+        return _device_info->calculate<256, ClipOp, half>(_info, workspace, output, inputs, stream);
+    case INFINI_DTYPE_F32:
+        return _device_info->calculate<256, ClipOp, float>(_info, workspace, output, inputs, stream);
+    case INFINI_DTYPE_F64:
+        return _device_info->calculate<256, ClipOp, double>(_info, workspace, output, inputs, stream);
+    default:
+        return INFINI_STATUS_BAD_TENSOR_DTYPE;
+    }
+
+    return INFINI_STATUS_SUCCESS;
+}
+} // namespace op::clip::cuda

src/infiniop/ops/clip/cuda/clip_cuda.cuh

+#ifndef __CLIP_CUDA_API_H__
+#define __CLIP_CUDA_API_H__
+
+#include "../../../elementwise/cuda/elementwise_cuda_api.cuh"
+#include "infiniop/ops/clip.h"
+
+ELEMENTWISE_DESCRIPTOR(clip, cuda)
+
+#endif // __CLIP_CUDA_API_H__

src/infiniop/ops/clip/cuda/clip_cuda_internal.cuh

+#ifndef __CLIP_CUDA_H__
+#define __CLIP_CUDA_H__
+
+#include "../../../elementwise/cuda/elementwise_cuda.cuh"
+#include <cuda_fp16.h>
+
+namespace op::clip::cuda {
+
+typedef struct ClipOp {
+public:
+    static constexpr size_t num_inputs = 3;
+
+    template <typename T>
+    __device__ __forceinline__ T operator()(const T &x, const T &min_val, const T &max_val) const {
+        if constexpr (std::is_same_v<T, half2>) {
+            return __hmax2(__hmin2(x, max_val), min_val);
+        } else if constexpr (std::is_same_v<T, half>) {
+            return __hmax(__hmin(x, max_val), min_val);
+        } else if constexpr (std::is_same_v<T, float>) {
+            return fmaxf(fminf(x, max_val), min_val);
+        } else if constexpr (std::is_same_v<T, double>) {
+            return fmax(fmin(x, max_val), min_val);
+        } else {
+            return std::max(std::min(x, max_val), min_val);
+        }
+    }
+} ClipOp;
+} // namespace op::clip::cuda
+
+#endif // __CLIP_CUDA_H__