chore: import upstream snapshot with attribution

2026-07-13 13:33:03 +08:00
commit 5b57521aa1
8226 changed files with 3425766 additions and 0 deletions
@@ -0,0 +1,121 @@
+//
+//  VulkanTopKV2.cpp
+//  MNN
+//
+//  Vulkan image-mode implementation of TopKV2, directly operating on images.
+//
+
+#include "VulkanTopKV2.hpp"
+#include "core/Macro.h"
+#include "core/TensorUtils.hpp"
+
+namespace MNN {
+
+struct GpuTopKV2Param {
+    int rowSize;
+    int k;
+    int numRows;
+    int C4; // UP_DIV(rowSize, 4)
+};
+
+VulkanTopKV2::VulkanTopKV2(const Op* op, Backend* bn, int k) : VulkanBasicExecution(bn) {
+    auto vkBn = (VulkanBackend *)backend();
+
+    mK = k;
+    mLargest = true;
+    auto param = op->main_as_TopKV2();
+    if (nullptr != param) {
+        mLargest = param->largest();
+    }
+
+    std::vector<VkDescriptorType> types{
+        VK_DESCRIPTOR_TYPE_STORAGE_IMAGE,        // output values (float image)
+        VK_DESCRIPTOR_TYPE_STORAGE_IMAGE,        // output indices (int image)
+        VK_DESCRIPTOR_TYPE_COMBINED_IMAGE_SAMPLER, // input (float sampler)
+        VK_DESCRIPTOR_TYPE_UNIFORM_BUFFER,       // params
+    };
+
+    if (mLargest) {
+        mPipeline = vkBn->getPipeline("glsl_topkv2_SORT_DESC_comp", types);
+    } else {
+        mPipeline = vkBn->getPipeline("glsl_topkv2_comp", types);
+    }
+
+    mGpuParam.reset(new VulkanBuffer(vkBn->getMemoryPool(), false, sizeof(GpuTopKV2Param), nullptr, VK_BUFFER_USAGE_UNIFORM_BUFFER_BIT));
+    mDescriptorSet.reset(mPipeline->createSet());
+}
+
+VulkanTopKV2::~VulkanTopKV2() {
+}
+
+ErrorCode VulkanTopKV2::onEncode(const std::vector<Tensor*>& inputs, const std::vector<Tensor*>& outputs,
+                                  const VulkanCommandPool::Buffer* cmdBuffer) {
+    auto vkBn = (VulkanBackend*)backend();
+    auto input = inputs[0];
+    auto outputValue = outputs[0];
+    auto outputIndex = outputs[1];
+
+    const int rowSize = input->length(input->dimensions() - 1);
+    if (rowSize <= 0) {
+        MNN_PRINT("VulkanTopKV2: rowSize <= 0 (%d), skip execution with output uninitialized\n", rowSize);
+        return NO_ERROR;
+    }
+    const int numRows = input->elementSize() / rowSize;
+    const int k = mK;
+
+    // Set GPU params
+    auto topkParam = reinterpret_cast<GpuTopKV2Param*>(mGpuParam->map());
+    topkParam->rowSize = rowSize;
+    topkParam->k = k;
+    topkParam->numRows = numRows;
+    topkParam->C4 = UP_DIV(rowSize, 4);
+    mGpuParam->unmap();
+
+    // Get VulkanTensor image handles
+    auto vkInput = reinterpret_cast<VulkanTensor*>(input->deviceId());
+    auto vkOutValue = reinterpret_cast<VulkanTensor*>(outputValue->deviceId());
+    auto vkOutIndex = reinterpret_cast<VulkanTensor*>(outputIndex->deviceId());
+
+    // Write descriptor set with images directly
+    mDescriptorSet.reset(mPipeline->createSet());
+    mDescriptorSet->writeImage(vkOutValue->image()->view(), vkBn->getCommonSampler()->get(),
+                               VK_IMAGE_LAYOUT_GENERAL, 0);
+    mDescriptorSet->writeImage(vkOutIndex->image()->view(), vkBn->getCommonSampler()->get(),
+                               VK_IMAGE_LAYOUT_GENERAL, 1);
+    mDescriptorSet->writeImage(vkInput->image()->view(), vkBn->getCommonSampler()->get(),
+                               VK_IMAGE_LAYOUT_SHADER_READ_ONLY_OPTIMAL, 2);
+    mDescriptorSet->writeBuffer(mGpuParam->buffer(), 3, mGpuParam->size());
+
+    // Barriers
+    vkOutValue->image()->barrierWrite(cmdBuffer->get());
+    vkOutIndex->image()->barrierWrite(cmdBuffer->get());
+    vkInput->image()->barrierRead(cmdBuffer->get());
+
+    // Dispatch: x=1 (workgroup handles 128 threads internally), y=numRows, z=1
+    mPipeline->bind(cmdBuffer->get(), mDescriptorSet->get());
+    vkCmdDispatch(cmdBuffer->get(), 1, numRows, 1);
+
+    return NO_ERROR;
+}
+
+class VulkanTopKV2Creator : public VulkanBackend::Creator {
+public:
+    virtual VulkanBasicExecution* onCreate(const std::vector<Tensor*>& inputs, const std::vector<Tensor*>& outputs, const MNN::Op* op,
+                                Backend* backend) const override {
+        if (inputs.size() < 2 || outputs.size() != 2) {
+            return nullptr;
+        }
+        if (TensorUtils::getDescribe(inputs[0])->dimensionFormat == MNN_DATA_FORMAT_NC4HW4) {
+            return nullptr;
+        }
+        const int k = inputs[1]->host<int32_t>()[0];
+        return new VulkanTopKV2(op, backend, k);
+    }
+};
+
+static bool gResistor = []() {
+    VulkanBackend::addCreator(OpType_TopKV2, new VulkanTopKV2Creator);
+    return true;
+}();
+
+}