mmr/tkDNN
Archived
1
0
Fork 0
Code Issues 46 Pull Requests 12 Releases 4 Activity

tensorrt7 support for ipluginv2

Browse Source
This commit is contained in:
perseusdg
2021-10-18 18:07:57 +05:30
parent 18cc6abbb6
commit 54e7af11ed
30 changed files with 313 additions and 64 deletions
Download Patch File Download Diff File Expand all files Collapse all files
src/pluginsRT/ActivationLeakyRT.cpp
+13 -2
View File
@@ -42,6 +42,7 @@ size_t ActivationLeakyRT::getWorkspaceSize(int maxBatchSize) const NOEXCEPT {
return 0;
}
#if NV_TENSORRT_MAJOR > 7
int ActivationLeakyRT::enqueue(int batchSize, const void *const *inputs, void *const *outputs, void *workspace,
cudaStream_t stream) NOEXCEPT {
activationLEAKYForward(
@@ -51,6 +52,17 @@ int ActivationLeakyRT::enqueue(int batchSize, const void *const *inputs, void *c
return 0;
}
#elif NV_TENSORRT_MAJOR == 7
int32_t ActivationLeakyRT::enqueue(int32_t batchSize, const void *const *inputs, void **outputs, void *workspace,
cudaStream_t stream) {
activationLEAKYForward(
(dnnType *) reinterpret_cast<const dnnType *>(inputs[0]),
reinterpret_cast<dnnType *>(outputs[0]), batchSize * size, slope,
stream);
return 0;
}
#endif
size_t ActivationLeakyRT::getSerializationSize() const NOEXCEPT {
return 1 * sizeof(int) + 1 * sizeof(float);
@@ -94,8 +106,7 @@ IPluginV2* ActivationLeakyRT::clone() const NOEXCEPT {
}
ActivationLeakyRTPluginCreator::ActivationLeakyRTPluginCreator() {
mPluginAttributes.emplace_back(
PluginField("slope", nullptr, PluginFieldType::kFLOAT32, 1));
mPluginAttributes.clear();
mFC.nbFields = mPluginAttributes.size();
mFC.fields = mPluginAttributes.data();
}
src/pluginsRT/ActivationLogisticRT.cpp
+9 -1
View File
@@ -38,13 +38,21 @@ void ActivationLogisticRT::terminate() NOEXCEPT {}
size_t ActivationLogisticRT::getWorkspaceSize(int maxBatchSize) const NOEXCEPT {
return 0;
}
#if NV_TENSORRT_MAJOR > 7
int ActivationLogisticRT::enqueue(int batchSize, const void *const *inputs, void *const *outputs, void *workspace,
cudaStream_t stream) NOEXCEPT {
activationLOGISTICForward((dnnType *) reinterpret_cast<const dnnType *>(inputs[0]),
reinterpret_cast<dnnType *>(outputs[0]), batchSize * size, stream);
return 0;
}
#elif NV_TENSORRT_MAJOR == 7
int32_t ActivationLogisticRT::enqueue(int32_t batchSize, const void *const *inputs, void **outputs, void *workspace,
cudaStream_t stream) {
activationLOGISTICForward((dnnType *) reinterpret_cast<const dnnType *>(inputs[0]),
reinterpret_cast<dnnType *>(outputs[0]), batchSize * size, stream);
return 0;
}
#endif
size_t ActivationLogisticRT::getSerializationSize() const NOEXCEPT {
return 1 * sizeof(int);
src/pluginsRT/ActivationMishRT.cpp
+12
View File
@@ -38,12 +38,22 @@ void ActivationMishRT::terminate() NOEXCEPT {}
size_t ActivationMishRT::getWorkspaceSize(int maxBatchSize) const NOEXCEPT { return 0; }
#if NV_TENSORRT_MAJOR > 7
int ActivationMishRT::enqueue(int batchSize, const void *const *inputs, void *const *outputs, void *workspace,
cudaStream_t stream) NOEXCEPT {
activationMishForward((dnnType *) reinterpret_cast<const dnnType *>(inputs[0]),
reinterpret_cast<dnnType *>(outputs[0]), batchSize * size, stream);
return 0;
}
#elif NV_TENSORRT_MAJOR == 7
int32_t ActivationMishRT::enqueue(int32_t batchSize, const void *const *inputs, void **outputs, void *workspace,
cudaStream_t stream) {
activationMishForward((dnnType *) reinterpret_cast<const dnnType *>(inputs[0]),
reinterpret_cast<dnnType *>(outputs[0]), batchSize * size, stream);
return 0;
}
#endif
size_t ActivationMishRT::getSerializationSize() const NOEXCEPT {
return 1 * sizeof(int);
}
@@ -80,6 +90,8 @@ IPluginV2 *ActivationMishRT::clone() const NOEXCEPT {
return p;
}
ActivationMishRTPluginCreator::ActivationMishRTPluginCreator() {
mPluginAttributes.clear();
mFC.nbFields = mPluginAttributes.size();
src/pluginsRT/ActivationReLUCeilingRT.cpp
+10 -1
View File
@@ -40,11 +40,20 @@ size_t ActivationReLUCeiling::getWorkspaceSize(int maxBatchSize) const NOEXCEPT
return 0;
}
#if NV_TENSORRT_MAJOR > 7
int ActivationReLUCeiling::enqueue(int batchSize, const void *const *inputs, void *const *outputs, void *workspace,cudaStream_t stream) NOEXCEPT {
activationReLUCeilingForward((dnnType *) reinterpret_cast<const dnnType *>(inputs[0]),
reinterpret_cast<dnnType *>(outputs[0]), batchSize * size, ceiling, stream);
return 0;
}
#elif NV_TENSORRT_MAJOR == 7
int32_t ActivationReLUCeiling::enqueue(int32_t batchSize, const void *const *inputs, void **outputs, void *workspace,
cudaStream_t stream) {
activationReLUCeilingForward((dnnType *) reinterpret_cast<const dnnType *>(inputs[0]),
reinterpret_cast<dnnType *>(outputs[0]), batchSize * size, ceiling, stream);
return 0;
}
#endif
size_t ActivationReLUCeiling::getSerializationSize() const NOEXCEPT {
return 1 * sizeof(int) + 1 * sizeof(float);
@@ -86,7 +95,7 @@ void ActivationReLUCeiling::setPluginNamespace(const char *pluginNamespace) NOEX
}
ActivationReLUCeilingPluginCreator::ActivationReLUCeilingPluginCreator() {
mPluginAttributes.emplace_back(PluginField("ceiling", nullptr, PluginFieldType::kFLOAT32, 1));
mPluginAttributes.clear();
mFC.nbFields = mPluginAttributes.size();
mFC.fields = mPluginAttributes.data();
}
src/pluginsRT/DeformableConvRT.cpp
+32 -18
View File
@@ -137,6 +137,7 @@ void DeformableConvRT::terminate() NOEXCEPT {}
size_t DeformableConvRT::getWorkspaceSize(int maxBatchSize) const NOEXCEPT {return 0;}
#if NV_TENSORRT_MAJOR > 7
int DeformableConvRT::enqueue(int batchSize, const void *const *inputs, void *const *outputs, void *workspace,
cudaStream_t stream) NOEXCEPT {
dnnType *srcData = (dnnType*)reinterpret_cast<const dnnType*>(inputs[0]);
@@ -165,6 +166,36 @@ int DeformableConvRT::enqueue(int batchSize, const void *const *inputs, void *co
}
return 0;
}
#elif NV_TENSORRT_MAJOR == 7
int32_t DeformableConvRT::enqueue(int32_t batchSize, const void *const *inputs, void **outputs, void *workspace,
cudaStream_t stream) {
dnnType *srcData = (dnnType*)reinterpret_cast<const dnnType*>(inputs[0]);
dnnType *output_conv = (dnnType*)reinterpret_cast<const dnnType*>(inputs[1]);
// split conv2d outputs into offset to mask
for(int b=0; b<batchSize; b++) {
checkCuda(cudaMemcpy(offset, output_conv + b * 3 * chunk_dim, 2*chunk_dim*sizeof(dnnType), cudaMemcpyDeviceToDevice));
checkCuda(cudaMemcpy(mask, output_conv + b * 3 * chunk_dim + 2*chunk_dim, chunk_dim*sizeof(dnnType), cudaMemcpyDeviceToDevice));
// kernel sigmoid
activationSIGMOIDForward(mask, mask, chunk_dim);
// deformable convolution
dcnV2CudaForward(stat, handle,
srcData, data_d,
bias2_d, ones_d1,
offset, mask,
reinterpret_cast<dnnType*>(outputs[0]), ones_d2,
kh, kw,
sh, sw,
ph, pw,
1, 1,
deformableGroup, b,
i_n, i_c, i_h, i_w,
o_n, o_c, o_h, o_w,
chunk_dim);
}
return 0;
}
#endif
size_t DeformableConvRT::getSerializationSize() const NOEXCEPT {
return 16 * sizeof(int) + chunk_dim * 3 * sizeof(dnnType) + (i_c * o_c * kh * kw * 1 ) * sizeof(dnnType) +
@@ -251,25 +282,8 @@ IPluginV2 *DeformableConvRT::clone() const NOEXCEPT {
return p;
}
DeformableConvRTPluginCreator::DeformableConvRTPluginCreator() {
mPluginAttributes.emplace_back(PluginField("chunk_dim",nullptr,PluginFieldType::kINT32,1));
mPluginAttributes.emplace_back(PluginField("kh",nullptr,PluginFieldType::kINT32,1));
mPluginAttributes.emplace_back(PluginField("kw",nullptr,PluginFieldType::kINT32,1));
mPluginAttributes.emplace_back(PluginField("sh",nullptr,PluginFieldType::kINT32,1));
mPluginAttributes.emplace_back(PluginField("sw",nullptr,PluginFieldType::kINT32,1));
mPluginAttributes.emplace_back(PluginField("ph",nullptr,PluginFieldType::kINT32,1));
mPluginAttributes.emplace_back(PluginField("pw",nullptr,PluginFieldType::kINT32,1));
mPluginAttributes.emplace_back(PluginField("deformableGroup",nullptr,PluginFieldType::kINT32,1));
mPluginAttributes.emplace_back(PluginField("i_n",nullptr,PluginFieldType::kINT32,1));
mPluginAttributes.emplace_back(PluginField("i_c",nullptr,PluginFieldType::kINT32,1));
mPluginAttributes.emplace_back(PluginField("i_h",nullptr,PluginFieldType::kINT32,1));
mPluginAttributes.emplace_back(PluginField("i_w",nullptr,PluginFieldType::kINT32,1));
mPluginAttributes.emplace_back(PluginField("o_n",nullptr,PluginFieldType::kINT32,1));
mPluginAttributes.emplace_back(PluginField("o_c",nullptr,PluginFieldType::kINT32,1));
mPluginAttributes.emplace_back(PluginField("o_h",nullptr,PluginFieldType::kINT32,1));
mPluginAttributes.emplace_back(PluginField("o_w",nullptr,PluginFieldType::kINT32,1));
mPluginAttributes.emplace_back(PluginField("defRT",nullptr,PluginFieldType::kUNKNOWN,1));
mPluginAttributes.clear();
mFC.nbFields = mPluginAttributes.size();
mFC.fields = mPluginAttributes.data();
}
src/pluginsRT/FlattenConcatRT.cpp
+19 -1
View File
@@ -54,6 +54,7 @@ size_t FlattenConcatRT::getWorkspaceSize(int maxBatchSize) const NOEXCEPT {
return 0;
}
#if NV_TENSORRT_MAJOR > 7
int FlattenConcatRT::enqueue(int batchSize, const void *const *inputs, void *const *outputs, void *workspace,
cudaStream_t stream) NOEXCEPT {
dnnType *srcData = (dnnType*)reinterpret_cast<const dnnType*>(inputs[0]);
@@ -69,6 +70,24 @@ int FlattenConcatRT::enqueue(int batchSize, const void *const *inputs, void *con
}
return 0;
}
#elif NV_TENSORRT_MAJOR == 7
int32_t FlattenConcatRT::enqueue(int32_t batchSize, const void *const *inputs, void **outputs, void *workspace,
cudaStream_t stream) {
dnnType *srcData = (dnnType*)reinterpret_cast<const dnnType*>(inputs[0]);
dnnType *dstData = reinterpret_cast<dnnType*>(outputs[0]);
checkCuda( cudaMemcpyAsync(dstData, srcData, batchSize*rows*cols*sizeof(dnnType), cudaMemcpyDeviceToDevice, stream));
checkERROR( cublasSetStream(handle, stream) );
for(int i=0; i<batchSize; i++) {
float const alpha(1.0);
float const beta(0.0);
int offset = i*rows*cols;
checkERROR( cublasSgeam( handle, CUBLAS_OP_T, CUBLAS_OP_N, rows, cols, &alpha, srcData + offset, cols, &beta, srcData + offset, rows, dstData + offset, rows ));
}
return 0;
}
#endif
size_t FlattenConcatRT::getSerializationSize() const NOEXCEPT {
return 5*sizeof(int);
@@ -114,7 +133,6 @@ IPluginV2 *FlattenConcatRT::clone() const NOEXCEPT {
return p;
}
FlattenConcatRTPluginCreator::FlattenConcatRTPluginCreator() {
mPluginAttributes.clear();
mFC.nbFields = mPluginAttributes.size();
src/pluginsRT/MaxPoolingSizeRT.cpp
+12 -9
View File
@@ -54,6 +54,7 @@ size_t MaxPoolFixedSizeRT::getWorkspaceSize(int maxBatchSize) const NOEXCEPT {
return 0;
}
#if NV_TENSORRT_MAJOR > 7
int MaxPoolFixedSizeRT::enqueue(int batchSize, const void *const *inputs, void *const *outputs, void *workspace,
cudaStream_t stream) NOEXCEPT {
dnnType *srcData = (dnnType*)reinterpret_cast<const dnnType*>(inputs[0]);
@@ -61,6 +62,16 @@ int MaxPoolFixedSizeRT::enqueue(int batchSize, const void *const *inputs, void *
MaxPoolingForward(srcData, dstData, batchSize, this->c, this->h, this->w, this->stride_H, this->stride_W, this->winSize, this->padding, stream);
return 0;
}
#elif NV_TENSORRT_MAJOR == 7
int32_t MaxPoolFixedSizeRT::enqueue(int32_t batchSize, const void *const *inputs, void **outputs, void *workspace,
cudaStream_t stream) {
dnnType *srcData = (dnnType*)reinterpret_cast<const dnnType*>(inputs[0]);
dnnType *dstData = reinterpret_cast<dnnType*>(outputs[0]);
MaxPoolingForward(srcData, dstData, batchSize, this->c, this->h, this->w, this->stride_H, this->stride_W, this->winSize, this->padding, stream);
return 0;
}
#endif
size_t MaxPoolFixedSizeRT::getSerializationSize() const NOEXCEPT {
return 8*sizeof(int);
@@ -110,16 +121,8 @@ IPluginV2 *MaxPoolFixedSizeRT::clone() const NOEXCEPT {
return p;
}
MaxPoolFixedSizeRTPluginCreator::MaxPoolFixedSizeRTPluginCreator() {
mPluginAttributes.emplace_back(PluginField("c",nullptr,PluginFieldType::kINT32,1));
mPluginAttributes.emplace_back(PluginField("h",nullptr,PluginFieldType::kINT32,1));
mPluginAttributes.emplace_back(PluginField("w",nullptr,PluginFieldType::kINT32,1));
mPluginAttributes.emplace_back(PluginField("n",nullptr,PluginFieldType::kINT32,1));
mPluginAttributes.emplace_back(PluginField("stride_H",nullptr,PluginFieldType::kINT32,1));
mPluginAttributes.emplace_back(PluginField("stride_W",nullptr,PluginFieldType::kINT32,1));
mPluginAttributes.emplace_back(PluginField("winSize",nullptr,PluginFieldType::kINT32,1));
mPluginAttributes.emplace_back(PluginField("padding",nullptr,PluginFieldType::kINT32,1));
mPluginAttributes.clear();
mFC.nbFields = mPluginAttributes.size();
mFC.fields = mPluginAttributes.data();
}
src/pluginsRT/RegionRT.cpp
+28 -4
View File
@@ -43,6 +43,7 @@ void RegionRT::terminate() NOEXCEPT {}
size_t RegionRT::getWorkspaceSize(int maxBatchSize) const NOEXCEPT { return 0; }
#if NV_TENSORRT_MAJOR > 7
int RegionRT::enqueue(int batchSize, const void *const *inputs, void *const *outputs, void *workspace,
cudaStream_t stream) NOEXCEPT {
dnnType *srcData = (dnnType*)reinterpret_cast<const dnnType*>(inputs[0]);
@@ -68,6 +69,32 @@ int RegionRT::enqueue(int batchSize, const void *const *inputs, void *const *out
return 0;
}
#elif NV_TENSORRT_MAJOR == 7
int32_t RegionRT::enqueue(int32_t batchSize, const void *const *inputs, void **outputs, void *workspace, cudaStream_t stream) {
dnnType *srcData = (dnnType*)reinterpret_cast<const dnnType*>(inputs[0]);
dnnType *dstData = reinterpret_cast<dnnType*>(outputs[0]);
checkCuda( cudaMemcpyAsync(dstData, srcData, batchSize*c*h*w*sizeof(dnnType), cudaMemcpyDeviceToDevice, stream));
for (int b = 0; b < batchSize; ++b){
for(int n = 0; n < num; ++n){
int index = entry_index(b, n*w*h, 0);
activationLOGISTICForward(srcData + index, dstData + index, 2*w*h, stream);
index = entry_index(b, n*w*h, coords);
activationLOGISTICForward(srcData + index, dstData + index, w*h, stream);
}
}
//softmax start
int index = entry_index(0, 0, coords + 1);
softmaxForward( srcData + index, classes, batchSize*num,
(c*h*w)/num,
w*h, 1, w*h, 1, dstData + index, stream);
return 0;
}
#endif
size_t RegionRT::getSerializationSize() const NOEXCEPT {
return 6*sizeof(int);
@@ -112,11 +139,8 @@ IPluginV2 *RegionRT::clone() const NOEXCEPT {
return p;
}
RegionRTPluginCreator::RegionRTPluginCreator() {
mPluginAttributes.emplace_back(PluginField("classes",nullptr,PluginFieldType::kINT32,1));
mPluginAttributes.emplace_back(PluginField("coords",nullptr,PluginFieldType::kINT32,1));
mPluginAttributes.emplace_back(PluginField("num",nullptr,PluginFieldType::kINT32,1));
mPluginAttributes.clear();
mFC.nbFields = mPluginAttributes.size();
mFC.fields = mPluginAttributes.data();
}
src/pluginsRT/ReorgRT.cpp
+11 -2
View File
@@ -43,12 +43,22 @@ size_t ReorgRT::getWorkspaceSize(int maxBatchSize) const NOEXCEPT {
return 0;
}
#if NV_TENSORRT_MAJOR > 7
int ReorgRT::enqueue(int batchSize, const void *const *inputs, void *const *outputs, void *workspace,cudaStream_t stream) NOEXCEPT {
reorgForward((dnnType*)reinterpret_cast<const dnnType*>(inputs[0]),
reinterpret_cast<dnnType*>(outputs[0]),
batchSize, c, h, w, stride, stream);
return 0;
}
#elif NV_TENSORRT_MAJOR == 7
int32_t ReorgRT::enqueue(int32_t batchSize, const void *const *inputs, void **outputs, void *workspace, cudaStream_t stream) {
reorgForward((dnnType*)reinterpret_cast<const dnnType*>(inputs[0]),
reinterpret_cast<dnnType*>(outputs[0]),
batchSize, c, h, w, stride, stream);
return 0;
}
#endif
size_t ReorgRT::getSerializationSize() const NOEXCEPT {
return 4*sizeof(int);
@@ -93,9 +103,8 @@ IPluginV2 *ReorgRT::clone() const NOEXCEPT {
return p;
}
ReorgRTPluginCreator::ReorgRTPluginCreator() {
mPluginAttributes.emplace_back(PluginField("stride",nullptr,PluginFieldType::kINT32,1));
mPluginAttributes.clear();
mFC.nbFields = mPluginAttributes.size();
mFC.fields = mPluginAttributes.data();
}
src/pluginsRT/ReshapeRT.cpp
+12 -2
View File
@@ -43,6 +43,7 @@ size_t ReshapeRT::getWorkspaceSize(int maxBatchSize) const NOEXCEPT {
return 0;
}
#if NV_TENSORRT_MAJOR > 7
int ReshapeRT::enqueue(int batchSize, const void *const *inputs, void *const *outputs, void *workspace,
cudaStream_t stream) NOEXCEPT {
dnnType *srcData = (dnnType*)reinterpret_cast<const dnnType*>(inputs[0]);
@@ -51,6 +52,16 @@ int ReshapeRT::enqueue(int batchSize, const void *const *inputs, void *const *ou
checkCuda( cudaMemcpyAsync(dstData, srcData, batchSize*c*h*w*sizeof(dnnType), cudaMemcpyDeviceToDevice, stream));
return 0;
}
#elif NV_TENSORRT_MAJOR == 7
int32_t ReshapeRT::enqueue(int32_t batchSize, const void *const *inputs, void **outputs, void *workspace, cudaStream_t stream) {
dnnType *srcData = (dnnType*)reinterpret_cast<const dnnType*>(inputs[0]);
dnnType *dstData = reinterpret_cast<dnnType*>(outputs[0]);
checkCuda( cudaMemcpyAsync(dstData, srcData, batchSize*c*h*w*sizeof(dnnType), cudaMemcpyDeviceToDevice, stream));
return 0;
}
#endif
size_t ReshapeRT::getSerializationSize() const NOEXCEPT {
return 4*sizeof(int);
@@ -96,9 +107,8 @@ IPluginV2 *ReshapeRT::clone() const NOEXCEPT {
return p;
}
ReshapeRTPluginCreator::ReshapeRTPluginCreator() {
mPluginAttributes.emplace_back(PluginField("new_dim",nullptr,PluginFieldType::kUNKNOWN,1));
mPluginAttributes.clear();
mFC.nbFields = mPluginAttributes.size();
mFC.fields = mPluginAttributes.data();
}
src/pluginsRT/ResizeLayerRT.cpp
+11 -4
View File
@@ -47,6 +47,7 @@ void ResizeLayerRT::terminate() NOEXCEPT {}
size_t ResizeLayerRT::getWorkspaceSize(int maxBatchSize) const NOEXCEPT { return 0; }
#if NV_TENSORRT_MAJOR > 7
int ResizeLayerRT::enqueue(int batchSize, const void *const *inputs, void *const *outputs, void *workspace,
cudaStream_t stream) NOEXCEPT {
resizeForward((dnnType*)reinterpret_cast<const dnnType*>(inputs[0]),
@@ -54,6 +55,15 @@ int ResizeLayerRT::enqueue(int batchSize, const void *const *inputs, void *const
batchSize, i_c, i_h, i_w, o_c, o_h, o_w, stream);
return 0;
}
#elif NV_TENSORRT_MAJOR == 7
int32_t ResizeLayerRT::enqueue(int32_t batchSize, const void *const *inputs, void **outputs, void *workspace,
cudaStream_t stream) {
resizeForward((dnnType*)reinterpret_cast<const dnnType*>(inputs[0]),
reinterpret_cast<dnnType*>(outputs[0]),
batchSize, i_c, i_h, i_w, o_c, o_h, o_w, stream);
return 0;
}
#endif
size_t ResizeLayerRT::getSerializationSize() const NOEXCEPT {
return 6*sizeof(int);
@@ -101,11 +111,8 @@ IPluginV2 *ResizeLayerRT::clone() const NOEXCEPT {
return p;
}
ResizeLayerRTPluginCreator::ResizeLayerRTPluginCreator() {
mPluginAttributes.emplace_back(PluginField("o_c",nullptr,PluginFieldType::kINT32,1));
mPluginAttributes.emplace_back(PluginField("o_h",nullptr,PluginFieldType::kINT32,1));
mPluginAttributes.emplace_back(PluginField("o_w",nullptr,PluginFieldType::kINT32,1));
mPluginAttributes.clear();
mFC.nbFields = mPluginAttributes.size();
mFC.fields = mPluginAttributes.data();
}
src/pluginsRT/RouteRT.cpp
+18 -3
View File
@@ -60,6 +60,7 @@ size_t RouteRT::getWorkspaceSize(int maxBatchSize) const NOEXCEPT {
return 0;
}
#if NV_TENSORRT_MAJOR > 7
int RouteRT::enqueue(int batchSize, const void *const *inputs, void *const *outputs, void *workspace,
cudaStream_t stream) NOEXCEPT {
dnnType *dstData = reinterpret_cast<dnnType*>(outputs[0]);
@@ -75,6 +76,22 @@ int RouteRT::enqueue(int batchSize, const void *const *inputs, void *const *outp
}
return 0;
}
#elif NV_TENSORRT_MAJOR == 7
int32_t RouteRT::enqueue(int32_t batchSize, const void *const *inputs, void **outputs, void *workspace, cudaStream_t stream) {
dnnType *dstData = reinterpret_cast<dnnType*>(outputs[0]);
for(int b=0; b<batchSize; b++) {
int offset = 0;
for(int i=0; i<in; i++) {
dnnType *input = (dnnType*)reinterpret_cast<const dnnType*>(inputs[i]);
int in_dim = c_in[i]*h*w;
int part_in_dim = in_dim / this->groups;
checkCuda( cudaMemcpyAsync(dstData + b*c*w*h + offset, input + b*c*w*h*groups + this->group_id*part_in_dim, part_in_dim*sizeof(dnnType), cudaMemcpyDeviceToDevice, stream) );
offset += part_in_dim;
}
}
return 0;
}
#endif
size_t RouteRT::getSerializationSize() const NOEXCEPT {
return (6+MAX_INPUTS)*sizeof(int);
@@ -124,10 +141,8 @@ IPluginV2 *RouteRT::clone() const NOEXCEPT {
return p;
}
RouteRTPluginCreator::RouteRTPluginCreator() {
mPluginAttributes.emplace_back(PluginField("groups",nullptr,PluginFieldType::kINT32,1));
mPluginAttributes.emplace_back(PluginField("group_id",nullptr,PluginFieldType::kINT32,1));
mPluginAttributes.clear();
mFC.nbFields = mPluginAttributes.size();
mFC.fields = mPluginAttributes.data();
}
src/pluginsRT/ShortcutRT.cpp
+16 -3
View File
@@ -50,6 +50,7 @@ void ShortcutRT::terminate() NOEXCEPT {}
size_t ShortcutRT::getWorkspaceSize(int maxBatchSize) const NOEXCEPT { return 0; }
#if NV_TENSORRT_MAJOR > 7
int ShortcutRT::enqueue(int batchSize, const void *const *inputs, void *const *outputs, void *workspace,
cudaStream_t stream) NOEXCEPT {
dnnType *srcData = (dnnType*)reinterpret_cast<const dnnType*>(inputs[0]);
@@ -61,6 +62,20 @@ int ShortcutRT::enqueue(int batchSize, const void *const *inputs, void *const *o
return 0;
}
#elif NV_TENSORRT_MAJOR == 7
int32_t ShortcutRT::enqueue(int32_t batchSize, const void *const *inputs, void **outputs, void *workspace,
cudaStream_t stream) {
dnnType *srcData = (dnnType*)reinterpret_cast<const dnnType*>(inputs[0]);
dnnType *srcDataBack = (dnnType*)reinterpret_cast<const dnnType*>(inputs[1]);
dnnType *dstData = reinterpret_cast<dnnType*>(outputs[0]);
checkCuda( cudaMemcpyAsync(dstData, srcData, batchSize*c*h*w*sizeof(dnnType), cudaMemcpyDeviceToDevice, stream));
shortcutForward(srcDataBack, dstData, batchSize, c, h, w, 1, batchSize, bc, bh, bw, 1, mul, stream);
return 0;
}
#endif
size_t ShortcutRT::getSerializationSize() const NOEXCEPT {
return 6*sizeof(int) + sizeof(bool);
@@ -108,10 +123,8 @@ IPluginV2 *ShortcutRT::clone() const NOEXCEPT {
return p;
}
ShortcutRTPluginCreator::ShortcutRTPluginCreator() {
mPluginAttributes.emplace_back(PluginField("bDim",nullptr,PluginFieldType::kUNKNOWN,1));
mPluginAttributes.emplace_back(PluginField("mul",nullptr,PluginFieldType::kUNKNOWN,1));
mPluginAttributes.clear();
mFC.nbFields = mPluginAttributes.size();
mFC.fields = mPluginAttributes.data();
}
src/pluginsRT/UpsampleRT.cpp
+13 -2
View File
@@ -44,6 +44,7 @@ size_t UpsampleRT::getWorkspaceSize(int maxBatchSize) const NOEXCEPT {
return 0;
}
#if NV_TENSORRT_MAJOR > 7
int UpsampleRT::enqueue(int batchSize, const void *const *inputs, void *const *outputs, void *workspace,
cudaStream_t stream) NOEXCEPT {
dnnType *srcData = (dnnType*)reinterpret_cast<const dnnType*>(inputs[0]);
@@ -53,6 +54,17 @@ int UpsampleRT::enqueue(int batchSize, const void *const *inputs, void *const *o
upsampleForward(srcData, dstData, batchSize, c, h, w, stride, 1, 1, stream);
return 0;
}
#elif NV_TENSORRT_MAJOR == 7
int32_t UpsampleRT::enqueue(int32_t batchSize, const void *const *inputs, void **outputs, void *workspace,
cudaStream_t stream) {
dnnType *srcData = (dnnType*)reinterpret_cast<const dnnType*>(inputs[0]);
dnnType *dstData = reinterpret_cast<dnnType*>(outputs[0]);
fill(dstData, batchSize*c*h*w*stride*stride, 0.0, stream);
upsampleForward(srcData, dstData, batchSize, c, h, w, stride, 1, 1, stream);
return 0;
}
#endif
size_t UpsampleRT::getSerializationSize() const NOEXCEPT {
return 4*sizeof(int);
@@ -97,9 +109,8 @@ IPluginV2 *UpsampleRT::clone() const NOEXCEPT {
return p;
}
UpsampleRTPluginCreator::UpsampleRTPluginCreator() {
mPluginAttributes.emplace_back(PluginField("stride",nullptr,PluginFieldType::kINT32,1));
mPluginAttributes.clear();
mFC.nbFields = mPluginAttributes.size();
mFC.fields = mPluginAttributes.data();
}
src/pluginsRT/YoloRT.cpp
+34 -9
View File
@@ -85,6 +85,7 @@ size_t YoloRT::getWorkspaceSize(int maxBatchSize) const NOEXCEPT {
return 0;
}
#if NV_TENSORRT_MAJOR > 7
int YoloRT::enqueue(int batchSize, const void *const *inputs, void *const *outputs, void *workspace,
cudaStream_t stream) NOEXCEPT {
dnnType *srcData = (dnnType *) reinterpret_cast<const dnnType *>(inputs[0]);
@@ -115,6 +116,38 @@ int YoloRT::enqueue(int batchSize, const void *const *inputs, void *const *outpu
//std::cout<<"YOLO END\n";
return 0;
}
#elif NV_TENSORRT_MAJOR == 7
int32_t YoloRT::enqueue(int32_t batchSize, const void *const *inputs, void **outputs, void *workspace, cudaStream_t stream) {
dnnType *srcData = (dnnType *) reinterpret_cast<const dnnType *>(inputs[0]);
dnnType *dstData = reinterpret_cast<dnnType *>(outputs[0]);
checkCuda(cudaMemcpyAsync(dstData, srcData, batchSize * c * h * w * sizeof(dnnType), cudaMemcpyDeviceToDevice,
stream));
for (int b = 0; b < batchSize; ++b) {
for (int n = 0; n < n_masks; ++n) {
int index = entry_index(b, n * w * h, 0);
if (new_coords == 1) {
if (this->scaleXY != 1)
scalAdd(dstData + index, 2 * w * h, this->scaleXY, -0.5 * (this->scaleXY - 1), 1);
} else {
activationLOGISTICForward(srcData + index, dstData + index, 2 * w * h, stream); //x,y
if (this->scaleXY != 1)
scalAdd(dstData + index, 2 * w * h, this->scaleXY, -0.5 * (this->scaleXY - 1), 1);
index = entry_index(b, n * w * h, 4);
activationLOGISTICForward(srcData + index, dstData + index, (1 + classes) * w * h, stream);
}
}
}
//std::cout<<"YOLO END\n";
return 0;
}
#endif
size_t YoloRT::getSerializationSize() const NOEXCEPT {
return 8 * sizeof(int) + 2 * sizeof(float) + n_masks * sizeof(dnnType) + num * n_masks * 2 * sizeof(dnnType) +
@@ -182,16 +215,8 @@ IPluginV2 *YoloRT::clone() const NOEXCEPT {
return p;
}
YoloRTPluginCreator::YoloRTPluginCreator() {
mPluginAttributes.emplace_back(PluginField("classes",nullptr,PluginFieldType::kINT32,1));
mPluginAttributes.emplace_back(PluginField("num",nullptr,PluginFieldType::kINT32,1));
mPluginAttributes.emplace_back(PluginField("yolo",nullptr,PluginFieldType::kUNKNOWN,1));
mPluginAttributes.emplace_back(PluginField("numMasks",nullptr,PluginFieldType::kINT32,1));
mPluginAttributes.emplace_back(PluginField("scaleXY",nullptr,PluginFieldType::kFLOAT32,1));
mPluginAttributes.emplace_back(PluginField("nmsThresh",nullptr,PluginFieldType::kFLOAT32,1));
mPluginAttributes.emplace_back(PluginField("nmsKind",nullptr,PluginFieldType::kINT32,1));
mPluginAttributes.emplace_back(PluginField("newCoords",nullptr,PluginFieldType::kINT32,1));
mPluginAttributes.clear();
mFC.nbFields = mPluginAttributes.size();
mFC.fields = mPluginAttributes.data();
}