wehub-resource-sync
133 lines
4.9 KiB
Plaintext
Executable File
133 lines
4.9 KiB
Plaintext
Executable File
//
|
|
// MetalScale.mm
|
|
// MNN
|
|
//
|
|
// Created by MNN on 2019/01/30.
|
|
// Copyright © 2018, Alibaba Group Holding Limited
|
|
//
|
|
|
|
#import "backend/metal/MetalScale.hpp"
|
|
#import "backend/metal/MNNMetalContext.h"
|
|
#import "core/Macro.h"
|
|
#import "backend/metal/MetalBackend.hpp"
|
|
|
|
#if MNN_METAL_ENABLED
|
|
namespace MNN {
|
|
static const char* shader_MetalScale_metal =
|
|
"struct scale_shape {\n"
|
|
" int size;\n"
|
|
" int steps;\n"
|
|
" int batch;\n"
|
|
" int offset;\n"
|
|
"};\n"
|
|
"kernel void scale_ca(const device M4 *in [[buffer(0)]],\n"
|
|
" device M4 *out [[buffer(1)]],\n"
|
|
" constant scale_shape &s [[buffer(2)]],\n"
|
|
" const device float4 *scalesbias[[buffer(3)]],\n"
|
|
" uint2 gid [[thread_position_in_grid]]) {\n"
|
|
" if ((int)gid.x >= s.size || (int)gid.y >= s.steps*s.batch) return;\n"
|
|
" int z=gid.y/s.batch;\n"
|
|
" int offset=s.offset;\n"
|
|
" float4 scale=scalesbias[z];\n"
|
|
"#ifdef BIAS\n"
|
|
" float4 bias=scalesbias[z+offset];\n"
|
|
" out[int(gid.y)*s.size+int(gid.x)] = (M4)((float4)in[int(gid.y)*s.size+int(gid.x)]*scale+bias);\n"
|
|
"#else \n"
|
|
" out[int(gid.y)*s.size+int(gid.x)] = (M4)((float4)in[int(gid.y)*s.size+int(gid.x)]*scale);\n"
|
|
"#endif \n"
|
|
"}\n"
|
|
;
|
|
|
|
MetalScale::MetalScale(Backend *backend, const Scale *scale) : MetalExecution(backend) {
|
|
auto mtbn = static_cast<MetalBackend *>(backend);
|
|
auto bufferAlloc = mtbn->getStaticBufferPool();
|
|
auto context = (__bridge MNNMetalContext *)static_cast<MetalBackend *>(backend)->context();
|
|
auto channel4 = UP_DIV(scale->channels(), 4) * 4;
|
|
mBiasOffset = channel4 / 4;
|
|
auto rt = (MetalRuntime*)mtbn->runtime();
|
|
MTLCompileOptions *option = [[MTLCompileOptions alloc] init];
|
|
auto dic = [NSMutableDictionary dictionaryWithCapacity:0];
|
|
std::vector<std::string> keys = {"scale"};
|
|
if (mtbn->useFp16InsteadFp32()) {
|
|
[dic setValue:@"half4" forKey:@"M4"];
|
|
keys.emplace_back("fp16");
|
|
} else {
|
|
[dic setValue:@"float4" forKey:@"M4"];
|
|
keys.emplace_back("fp32");
|
|
}
|
|
int scaleBiasNumber = 1;
|
|
if (nullptr != scale->biasData()) {
|
|
[dic setValue:@"1" forKey:@"BIAS"];
|
|
keys.emplace_back("BIAS");
|
|
scaleBiasNumber = 2;
|
|
}
|
|
option.preprocessorMacros = dic;
|
|
auto pipeline = rt->findPipeline(keys);
|
|
if (nil == pipeline) {
|
|
// Rebuild Pipeline
|
|
pipeline = mtbn->makeComputePipelineWithSourceOption(shader_MetalScale_metal, "scale_ca", option);
|
|
rt->insertPipeline(keys, pipeline);
|
|
}
|
|
mPipeline = pipeline;
|
|
mScaleBias = bufferAlloc->alloc(scaleBiasNumber * channel4 * sizeof(float));
|
|
if (mScaleBias.first == nullptr) {
|
|
mValid = false;
|
|
return;
|
|
}
|
|
auto scalePtr = MetalBackend::getMemPtr(mScaleBias);
|
|
::memset(scalePtr, 0, channel4 * sizeof(float));
|
|
::memcpy(scalePtr, scale->scaleData()->data(), scale->channels() * sizeof(float));
|
|
if (nullptr != scale->biasData()) {
|
|
auto biasPtr = scalePtr + channel4 * sizeof(float);
|
|
::memset(biasPtr, 0, channel4 * sizeof(float));
|
|
::memcpy(biasPtr, scale->biasData()->data(), scale->channels() * sizeof(float));
|
|
}
|
|
mConst = [context newDeviceBuffer:4 * sizeof(int) access:CPUWriteOnly];
|
|
}
|
|
MetalScale::~MetalScale() {
|
|
auto mtbn = static_cast<MetalBackend *>(backend());
|
|
auto bufferAlloc = mtbn->getStaticBufferPool();
|
|
if (nullptr != mScaleBias.first) {
|
|
bufferAlloc->free(mScaleBias);
|
|
}
|
|
}
|
|
|
|
ErrorCode MetalScale::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 output = outputs[0];
|
|
|
|
// shape
|
|
int w = output->width();
|
|
int h = output->height();
|
|
int c = output->channel();
|
|
int z = UP_DIV(c, 4);
|
|
((int *)mConst.contents)[0] = w*h;
|
|
((int *)mConst.contents)[1] = z;
|
|
((int *)mConst.contents)[2] = output->batch();
|
|
((int *)mConst.contents)[3] = mBiasOffset;
|
|
mThreads = [context computeBestGroupAndLocal:mPipeline threads:MTLSizeMake(w*h, z * outputs[0]->batch(), 1)];
|
|
return NO_ERROR;
|
|
}
|
|
|
|
void MetalScale::onEncode(const std::vector<Tensor *> &inputs, const std::vector<Tensor *> &outputs, id<MTLComputeCommandEncoder> encoder) {
|
|
auto input = inputs[0], output = outputs[0];
|
|
[encoder setComputePipelineState:mPipeline];
|
|
MetalBackend::setTensor(input, encoder, 0);
|
|
MetalBackend::setTensor(output, encoder, 1);
|
|
[encoder setBuffer:mConst offset:0 atIndex:2];
|
|
MetalBackend::setMem(mScaleBias, encoder, 3);
|
|
[encoder dispatchThreadgroups:mThreads.first threadsPerThreadgroup:mThreads.second];
|
|
|
|
}
|
|
|
|
class MetalScaleCreator : public MetalBackend::Creator {
|
|
public:
|
|
virtual Execution *onCreate(const std::vector<Tensor *> &inputs, const MNN::Op *op, Backend *backend, const std::vector<Tensor *>& outputs) const {
|
|
return new MetalScale(backend, op->main_as_Scale());
|
|
}
|
|
};
|
|
REGISTER_METAL_OP_CREATOR(MetalScaleCreator, OpType_Scale);
|
|
} // namespace MNN
|
|
#endif /* MNN_METAL_ENABLED */
|