wehub-resource-sync
120 lines
4.8 KiB
Plaintext
120 lines
4.8 KiB
Plaintext
//
|
|
// MetalROIAlign.mm
|
|
// MNN
|
|
//
|
|
|
|
#import "MetalROIAlign.hpp"
|
|
#import "MetalBackend.hpp"
|
|
#import "MNNMetalContext.h"
|
|
#import "core/Macro.h"
|
|
#import "core/TensorUtils.hpp"
|
|
|
|
#if MNN_METAL_ENABLED
|
|
namespace MNN {
|
|
|
|
MetalROIAlign::MetalROIAlign(Backend *backend, const MNN::Op *op)
|
|
: MetalExecution(backend) {
|
|
auto param = op->main_as_RoiParameters();
|
|
mPooledWidth = param->pooledWidth();
|
|
mPooledHeight = param->pooledHeight();
|
|
mSpatialScale = param->spatialScale();
|
|
mSamplingRatio = param->samplingRatio();
|
|
mAligned = param->aligned();
|
|
mPoolType = param->poolType();
|
|
|
|
auto mtbn = static_cast<MetalBackend *>(backend);
|
|
auto context = (__bridge MNNMetalContext *)mtbn->context();
|
|
// 12 fields: 11 int + 1 float (spatial_scale at index 8)
|
|
mConstBuffer = [context newDeviceBuffer:12 * sizeof(int) access:CPUWriteOnly];
|
|
}
|
|
|
|
ErrorCode MetalROIAlign::onResize(const std::vector<Tensor *> &inputs,
|
|
const std::vector<Tensor *> &outputs) {
|
|
auto backend = static_cast<MetalBackend *>(this->backend());
|
|
auto context = (__bridge MNNMetalContext *)backend->context();
|
|
auto input = inputs[0];
|
|
auto output = outputs[0];
|
|
|
|
int inputWidth = input->width();
|
|
int inputHeight = input->height();
|
|
int inputBatch = input->batch();
|
|
int channel = input->channel();
|
|
int numSlice = UP_DIV(channel, 4);
|
|
int numROI = output->batch();
|
|
bool hasBatchIndices = (inputs.size() >= 3);
|
|
|
|
// Select pipeline based on pool type
|
|
NSString *kernelName = (mPoolType == PoolType_MAXPOOL) ? @"roi_align_max" : @"roi_align_avg";
|
|
mPipeline = [context pipelineWithName:kernelName fp16:backend->useFp16InsteadFp32()];
|
|
|
|
// Fill constant buffer (matches roi_align_shape struct in shader)
|
|
auto cst = (int *)mConstBuffer.contents;
|
|
cst[0] = inputWidth; // input_width
|
|
cst[1] = inputHeight; // input_height
|
|
cst[2] = inputWidth * inputHeight; // input_size (H*W)
|
|
cst[3] = inputBatch; // input_batch
|
|
cst[4] = mPooledWidth; // output_width (pooled_width)
|
|
cst[5] = mPooledHeight; // output_height (pooled_height)
|
|
cst[6] = mPooledWidth * mPooledHeight; // output_size
|
|
cst[7] = numROI; // num_roi
|
|
((float *)cst)[8] = mSpatialScale; // spatial_scale (float)
|
|
cst[9] = mSamplingRatio; // sampling_ratio
|
|
cst[10] = mAligned ? 1 : 0; // aligned
|
|
cst[11] = hasBatchIndices ? 1 : 0; // has_batch_indices
|
|
|
|
// Compute threadgroup size
|
|
// Grid: (pooledWidth, pooledHeight, numROI * numSlice)
|
|
auto threads = MTLSizeMake(mPooledWidth, mPooledHeight, numROI * numSlice);
|
|
mThreads = [context computeBestGroupAndLocal:mPipeline threads:threads];
|
|
|
|
return NO_ERROR;
|
|
}
|
|
|
|
void MetalROIAlign::onEncode(const std::vector<Tensor *> &inputs,
|
|
const std::vector<Tensor *> &outputs,
|
|
id<MTLComputeCommandEncoder> encoder) {
|
|
auto input = inputs[0];
|
|
auto rois = inputs[1];
|
|
auto output = outputs[0];
|
|
|
|
[encoder setComputePipelineState:mPipeline];
|
|
MetalBackend::setTensor(input, encoder, 0); // buffer(0) = input feature map (ftype4, NC4HW4)
|
|
MetalBackend::setTensor(rois, encoder, 1); // buffer(1) = ROI data (ftype, linear)
|
|
MetalBackend::setTensor(output, encoder, 2); // buffer(2) = output (ftype4, NC4HW4)
|
|
[encoder setBuffer:mConstBuffer offset:0 atIndex:3]; // buffer(3) = constants
|
|
|
|
// buffer(4) = batch indices (optional)
|
|
if (inputs.size() >= 3) {
|
|
MetalBackend::setTensor(inputs[2], encoder, 4);
|
|
} else {
|
|
// Set a dummy buffer when batch_indices not provided (shader won't use it)
|
|
[encoder setBuffer:mConstBuffer offset:0 atIndex:4];
|
|
}
|
|
|
|
[encoder dispatchThreadgroups:mThreads.first
|
|
threadsPerThreadgroup:mThreads.second];
|
|
}
|
|
|
|
class MetalROIAlignCreator : public MetalBackend::Creator {
|
|
public:
|
|
virtual Execution *onCreate(const std::vector<Tensor *> &inputs,
|
|
const MNN::Op *op, Backend *backend,
|
|
const std::vector<Tensor *> &outputs) const override {
|
|
auto param = op->main_as_RoiParameters();
|
|
if (param == nullptr) {
|
|
MNN_ERROR("MetalROIAlign: RoiParameters is null\n");
|
|
return nullptr;
|
|
}
|
|
// Only support forward inference (outputGrad=false)
|
|
if (param->outputGrad()) {
|
|
MNN_ERROR("MetalROIAlign: outputGrad=true not supported on Metal\n");
|
|
return nullptr;
|
|
}
|
|
return new MetalROIAlign(backend, op);
|
|
}
|
|
};
|
|
|
|
REGISTER_METAL_OP_CREATOR(MetalROIAlignCreator, OpType_ROIAlign);
|
|
|
|
} // namespace MNN
|
|
#endif /* MNN_METAL_ENABLED */ |