Method enqueue

plugin/skipLayerNormPlugin/skipLayerNormPlugin.cpp:257–343

}

int32_t SkipLayerNormPluginDynamic::enqueue(PluginTensorDesc const* inputDesc, PluginTensorDesc const* outputDesc,
    void const* const* inputs, void* const* outputs, void* workspace, cudaStream_t stream) noexcept
{
    int32_t status = -1;
    try
    {
        PLUGIN_VALIDATE(inputs != nullptr);
        PLUGIN_VALIDATE(outputs != nullptr);
        int32_t const inputVolume = volume(inputDesc[0].dims);
        DataType iType = inputDesc->type;

        // Our plugin outputs only one tensor
        // Launch CUDA kernel wrapper and save its return value
        if (iType == DataType::kFLOAT)
        {
            auto const* const input = static_cast<float const*>(inputs[0]);
            auto const* const skip = static_cast<float const*>(inputs[1]);
            auto* output = static_cast<float*>(outputs[0]);
            auto const* const bias = static_cast<float const*>(mBiasDev.get());
            auto const* const beta = static_cast<float const*>(mBetaDev.get());
            auto const* const gamma = static_cast<float const*>(mGammaDev.get());
            if (mHasBias)
            {
                status = computeSkipLayerNorm<float, true>(
                    stream, static_cast<int32_t>(mLd), inputVolume, input, skip, beta, gamma, output, bias);
            }
            else
            {
                status = computeSkipLayerNorm<float, false>(
                    stream, static_cast<int32_t>(mLd), inputVolume, input, skip, beta, gamma, output, bias);
            }
        }
        else if (iType == DataType::kHALF)
        {
            auto const* const input = static_cast<half const*>(inputs[0]);
            auto const* const skip = static_cast<half const*>(inputs[1]);
            auto* output = static_cast<half*>(outputs[0]);
            auto const* const bias = static_cast<half const*>(mBiasDev.get());
            auto const* const beta = static_cast<half const*>(mBetaDev.get());
            auto const* const gamma = static_cast<half const*>(mGammaDev.get());
            if (mHasBias)
            {
                status = computeSkipLayerNorm<half, true>(
                    stream, static_cast<int32_t>(mLd), inputVolume, input, skip, beta, gamma, output, bias);
            }
            else
            {
                status = computeSkipLayerNorm<half, false>(
                    stream, static_cast<int32_t>(mLd), inputVolume, input, skip, beta, gamma, output, bias);
            }
        }
        else if (iType == DataType::kINT8)
        {
            float const dqScaleIn = inputDesc[0].scale;
            float const dqScaleSkip = inputDesc[1].scale;
            PLUGIN_VALIDATE(outputDesc[0].scale != 0.0F);
            float const qScale = 1.F / outputDesc[0].scale;
            auto const* const input = static_cast<int8_t const*>(inputs[0]);