asc_get_mask_spr

产品支持情况

• Ascend 950PR/Ascend 950DT：支持

• Atlas A3 训练系列产品/Atlas A3 推理系列产品：不支持

• Atlas A2 训练系列产品/Atlas A2 推理系列产品：不支持

• Atlas 200I/500 A2 推理产品：不支持

• Atlas 推理系列产品AI Core：不支持

• Atlas 推理系列产品Vector Core：不支持

• Atlas 训练系列产品：不支持

功能说明

头文件路径为："c_api/reg_compute/compute/reg_mask.h"。

从SPR（Special Purpose Register）中读取mask值，并按数据类型格式化后返回。

• 对于b16数据类型：从SPR读取mask共128bit数据，将每个bit扩展为2bit，拼成256bit掩码寄存器返回。
• 对于b32数据类型：从SPR读取mask共64bit数据，将每个bit扩展为4bit，拼成256bit掩码寄存器返回。

函数原型

C

__simd_callee__ inline vector_bool asc_get_mask_spr_b16()
__simd_callee__ inline vector_bool asc_get_mask_spr_b32()

参数说明

无

返回值说明

vector_bool，掩码寄存器。

约束说明

• 本接口仅在AIV上生效，非AIV调用直接返回。
• 本接口需在Vector Function（__simd_vf__标记的函数）内调用。

调用示例

将代码保存为example.asc后，可通过bisheng命令编译运行，其中--npu-arch参数需根据实际产品型号指定对应的NPU架构，具体产品与NPU架构的映射关系请参考__NPU_ARCH__。

以Ascend 950PR/Ascend 950DT产品（对应NPU架构为dav-3510）为例，编译运行命令如下：

Bash

bisheng example.asc -o main --npu-arch=dav-3510&& ./main

C++

#include <cstdint>
#include <iostream>
#include <vector>
#include "c_api/asc_simd.h"
#include "acl/acl.h"

namespace {
template <typename T>
void print_data(const char* label, const std::vector<T>& values)
{
    std::cout << label << ":";
    const size_t count = values.size() < 8 ? values.size() : 8;
    for (size_t i = 0; i < count; ++i) std::cout << ' ' << +values[i];
    if (values.size() > count) std::cout << " ...";
    std::cout << std::endl;
}

template <typename T>
bool compare_data(const std::vector<T>& actual, const std::vector<T>& expected, double tolerance = 0.0)
{
    if (actual.size() != expected.size()) return false;
    for (size_t i = 0; i < actual.size(); ++i) {
        if (actual[i] == expected[i]) continue;
        const double diff = static_cast<double>(actual[i]) - static_cast<double>(expected[i]);
        if (diff > tolerance || diff < -tolerance) return false;
    }
    return true;
}

template <typename T>
bool compare_range_data(const std::vector<T>& actual, const std::vector<T>& expected,
    size_t begin, size_t count, double tolerance = 0.0)
{
    if (begin + count > actual.size() || begin + count > expected.size()) return false;
    for (size_t i = begin; i < begin + count; ++i) {
        if (actual[i] == expected[i]) continue;
        const double diff = static_cast<double>(actual[i]) - static_cast<double>(expected[i]);
        if (diff > tolerance || diff < -tolerance) return false;
    }
    return true;
}

constexpr uint32_t BUFFER_BYTES = 768;

__simd_vf__ inline void exercise(__ubuf__ uint8_t* output, __ubuf__ uint8_t* input)
{
    uint32_t count16 = 128;
    vector_bool seed16 = asc_update_mask_b16(count16);
    vector_bool mask16 = asc_get_mask_spr_b16();
    vector_uint16_t value16;
    asc_duplicate_scalar(value16, static_cast<uint16_t>(1), seed16);
    asc_storealign(reinterpret_cast<__ubuf__ uint16_t*>(output), value16, mask16);
    uint32_t count32 = 64;
    vector_bool seed32 = asc_update_mask_b32(count32);
    vector_bool mask32 = asc_get_mask_spr_b32();
    vector_uint32_t value32;
    asc_duplicate_scalar(value32, static_cast<uint32_t>(1), seed32);
    asc_storealign(reinterpret_cast<__ubuf__ uint32_t*>(output + 256), value32, mask32);
}

__global__ __vector__ void asc_get_mask_spr_kernel(__gm__ uint8_t* output, __gm__ uint8_t* input)
{
    asc_init();
    __ubuf__ uint8_t output_local[BUFFER_BYTES];
    __ubuf__ uint8_t input_local[BUFFER_BYTES];
    asc_copy_gm2ub_align(input_local, input, BUFFER_BYTES);
    asc_copy_gm2ub_align(output_local, input, BUFFER_BYTES);
    asc_sync_notify(PIPE_MTE2, PIPE_V, EVENT_ID0);
    asc_sync_wait(PIPE_MTE2, PIPE_V, EVENT_ID0);
    exercise(output_local, input_local);
    asc_sync_notify(PIPE_V, PIPE_MTE3, EVENT_ID0);
    asc_sync_wait(PIPE_V, PIPE_MTE3, EVENT_ID0);
    asc_copy_ub2gm_align(output, output_local, BUFFER_BYTES);
    asc_sync();
}
} // namespace

int main()
{
    std::vector<uint8_t> input(BUFFER_BYTES), output(BUFFER_BYTES, 0xff), golden(BUFFER_BYTES);
    for (uint32_t i = 0; i < BUFFER_BYTES; ++i) input[i] = static_cast<uint8_t>(i % 251);
    for (uint32_t i = 0; i < 256; i += sizeof(uint16_t)) golden[i] = 1;
    for (uint32_t i = 256; i < 512; i += sizeof(uint32_t)) golden[i] = 1;
    std::copy(input.begin() + 512, input.end(), golden.begin() + 512);
    aclInit(nullptr);
    aclrtSetDevice(0);
    uint8_t* input_device = nullptr;
    aclrtMalloc(reinterpret_cast<void**>(&input_device), (BUFFER_BYTES) * sizeof(uint8_t),
        ACL_MEM_MALLOC_HUGE_FIRST);
    uint8_t* output_device = nullptr;
    aclrtMalloc(reinterpret_cast<void**>(&output_device), (BUFFER_BYTES) * sizeof(uint8_t),
        ACL_MEM_MALLOC_HUGE_FIRST);
    aclrtMemcpy(input_device, input.size() * sizeof(uint8_t), input.data(), input.size() * sizeof(uint8_t),
        ACL_MEMCPY_HOST_TO_DEVICE);
    asc_get_mask_spr_kernel<<<1, 0>>>(output_device, input_device);
    aclrtSynchronizeDevice();
    aclrtMemcpy(output.data(), output.size() * sizeof(uint8_t), output_device, output.size() * sizeof(uint8_t),
        ACL_MEMCPY_DEVICE_TO_HOST);
    print_data("Input bytes", input);
    print_data("Output bytes", output);
    print_data("Golden bytes", golden);
    const bool passed = compare_data(output, golden);
    std::cout << (passed ? "[Success] asc_get_mask_spr_b16/asc_get_mask_spr_b32 passed." : "[Failed] asc_get_mask_spr_b16/asc_get_mask_spr_b32 failed.") << std::endl;
    aclrtFree(input_device);
    aclrtFree(output_device);
    aclrtResetDevice(0);
    aclFinalize();
    return passed ? 0 : 1;
}