[IE CLDNN] DispatchData refactoring (openvinotoolkit#2508)

mryzhov · Dec 15, 2020 · d00bf25 · d00bf25
1 parent af8f09e
commit d00bf25
Show file tree

Hide file tree

Showing 298 changed files with 2,529 additions and 3,223 deletions.
diff --git a/...hirdparty/clDNN/kernel_selector/core/actual_kernels/activation/activation_kernel_base.cpp b/...hirdparty/clDNN/kernel_selector/core/actual_kernels/activation/activation_kernel_base.cpp
@@ -23,31 +23,21 @@ namespace kernel_selector {
 ActivationKernelBase::DispatchData ActivationKernelBase::SetDefault(const activation_params& arg) const {
     const auto& out = arg.output;
 
-    DispatchData runInfo;
-    std::vector<size_t> global;
-    std::vector<size_t> local;
+    DispatchData dispatchData;
     if (out.GetLayout() == DataLayout::yxfb) {
-        global = {out.Feature().v * out.Batch().v, out.X().v, out.Y().v};
-        local = GetOptimalLocalWorkGroupSizes(global, arg.engineInfo);
+        dispatchData.gws = {out.Feature().v * out.Batch().v, out.X().v, out.Y().v};
+        dispatchData.lws = GetOptimalLocalWorkGroupSizes(dispatchData.gws, arg.engineInfo);
     } else if (out.GetLayout() == DataLayout::b_fs_yx_fsv16) {
-        global = {Align(out.Feature().v, 16) * out.Batch().v, out.X().v, out.Y().v};
-        local = {16, 1, 1};
+        dispatchData.gws = {Align(out.Feature().v, 16) * out.Batch().v, out.X().v, out.Y().v};
+        dispatchData.lws = {16, 1, 1};
     } else {
-        global = {out.X().v, out.Y().v * out.Z().v, out.Feature().v * out.Batch().v};
-        local = GetOptimalLocalWorkGroupSizes(global, arg.engineInfo);
+        dispatchData.gws = {out.X().v, out.Y().v * out.Z().v, out.Feature().v * out.Batch().v};
+        dispatchData.lws = GetOptimalLocalWorkGroupSizes(dispatchData.gws, arg.engineInfo);
     }
 
-    runInfo.gws0 = global[0];
-    runInfo.gws1 = global[1];
-    runInfo.gws2 = global[2];
-    runInfo.lws0 = local[0];
-    runInfo.lws1 = local[1];
-    runInfo.lws2 = local[2];
+    dispatchData.efficiency = DONT_USE_IF_HAVE_SOMETHING_ELSE;
 
-    runInfo.efficiency = DONT_USE_IF_HAVE_SOMETHING_ELSE;
-    runInfo.fp16UnitUsed = out.GetDType() == Datatype::F16;
-
-    return runInfo;
+    return dispatchData;
 }
 
 JitConstants ActivationKernelBase::GetJitConstants(const activation_params& params, DispatchData) const {
@@ -94,20 +84,20 @@ KernelsData ActivationKernelBase::GetCommonKernelsData(const Params& params, con
     activation_params& newParams = *static_cast<activation_params*>(kd.params.get());
     const std::string kernel_id = GetEntryPoint(kernelName, params.layerID, options);
 
-    auto runInfo = SetDefault(newParams);
-    auto cldnn_jit = GetJitConstants(newParams, runInfo);
+    auto dispatchData = SetDefault(newParams);
+    auto cldnn_jit = GetJitConstants(newParams, dispatchData);
     auto entry_point = GetEntryPoint(kernelName, newParams.layerID, options);
     auto jit = CreateJit(kernelName, cldnn_jit, entry_point);
 
     auto& kernel = kd.kernels[0];
-    FillCLKernelData(kernel, runInfo, params.engineInfo, kernelName, jit, entry_point,
+    FillCLKernelData(kernel, dispatchData, params.engineInfo, kernelName, jit, entry_point,
                      DEFAULT, false, false, 1, GetFusedPrimitiveInputsCount(params));
 
     if (!newParams.inputActivationParams.empty()) {
         kernel.arguments.push_back({ArgumentDescriptor::Types::SLOPE, 0});
     }
 
-    kd.estimatedTime = runInfo.efficiency;
+    kd.estimatedTime = dispatchData.efficiency;
 
     return {kd};
 }

diff --git a/.../thirdparty/clDNN/kernel_selector/core/actual_kernels/activation/activation_kernel_base.h b/.../thirdparty/clDNN/kernel_selector/core/actual_kernels/activation/activation_kernel_base.h
@@ -65,7 +65,7 @@ class ActivationKernelBase : public common_kernel_base {
 
 protected:
     bool Validate(const Params& p, const optional_params& o) const override;
-    virtual JitConstants GetJitConstants(const activation_params& params, DispatchData kd) const;
+    virtual JitConstants GetJitConstants(const activation_params& params, DispatchData dispatchData) const;
     virtual DispatchData SetDefault(const activation_params& arg) const;
     KernelsData GetCommonKernelsData(const Params& params, const optional_params& options) const;
 };

diff --git a/...thirdparty/clDNN/kernel_selector/core/actual_kernels/activation/activation_kernel_opt.cpp b/...thirdparty/clDNN/kernel_selector/core/actual_kernels/activation/activation_kernel_opt.cpp
@@ -38,24 +38,16 @@ ParamsKey ActivationKernelOpt::GetSupportedKey() const {
 }
 
 ActivationKernelOpt::Parent::DispatchData ActivationKernelOpt::SetDefault(const activation_params& params) const {
-    auto runInfo = Parent::SetDefault(params);
+    auto dispatchData = Parent::SetDefault(params);
 
     const auto totalSize = params.inputs[0].LogicalSize();
 
-    std::vector<size_t> global = {totalSize / NUM_COLS_WI};
-    std::vector<size_t> local = GetOptimalLocalWorkGroupSizes(global, params.engineInfo);
+    dispatchData.gws = { totalSize / NUM_COLS_WI, 1, 1 };
+    dispatchData.lws = GetOptimalLocalWorkGroupSizes(dispatchData.gws, params.engineInfo);
 
-    runInfo.gws0 = global[0];
-    runInfo.gws1 = 1;
-    runInfo.gws2 = 1;
+    dispatchData.efficiency = FORCE_PRIORITY_6;
 
-    runInfo.lws0 = local[0];
-    runInfo.lws1 = 1;
-    runInfo.lws2 = 1;
-
-    runInfo.efficiency = FORCE_PRIORITY_6;
-
-    return runInfo;
+    return dispatchData;
 }
 
 bool ActivationKernelOpt::Validate(const Params& p, const optional_params& o) const {
@@ -87,8 +79,8 @@ bool ActivationKernelOpt::Validate(const Params& p, const optional_params& o) co
     return true;
 }
 
-JitConstants ActivationKernelOpt::GetJitConstants(const activation_params& params, DispatchData kd) const {
-    auto jit = ActivationKernelBase::GetJitConstants(params, kd);
+JitConstants ActivationKernelOpt::GetJitConstants(const activation_params& params, DispatchData dispatchData) const {
+    auto jit = ActivationKernelBase::GetJitConstants(params, dispatchData);
     auto input_dt = params.inputs[0].GetDType();
 
     jit.AddConstant(MakeJitConstant("NUM_COLS_WI", NUM_COLS_WI));

diff --git a/...e/thirdparty/clDNN/kernel_selector/core/actual_kernels/activation/activation_kernel_opt.h b/...e/thirdparty/clDNN/kernel_selector/core/actual_kernels/activation/activation_kernel_opt.h
@@ -33,7 +33,7 @@ class ActivationKernelOpt : public ActivationKernelBase {
     static const int NUM_COLS_WI = 4;
     DispatchData SetDefault(const activation_params& arg) const override;
     bool Validate(const Params& p, const optional_params& o) const override;
-    JitConstants GetJitConstants(const activation_params& params, DispatchData kd) const override;
+    JitConstants GetJitConstants(const activation_params& params, DispatchData dispatchData) const override;
     std::vector<FusedOpType> GetSupportedFusedOps() const override {
         return {FusedOpType::QUANTIZE,
                 FusedOpType::SCALE,

diff --git a/...thirdparty/clDNN/kernel_selector/core/actual_kernels/activation/activation_kernel_ref.cpp b/...thirdparty/clDNN/kernel_selector/core/actual_kernels/activation/activation_kernel_ref.cpp
@@ -41,8 +41,8 @@ ParamsKey ActivationKernelRef::GetSupportedKey() const {
     return k;
 }
 
-JitConstants ActivationKernelRef::GetJitConstants(const activation_params& params, DispatchData kd) const {
-    auto jit = ActivationKernelBase::GetJitConstants(params, kd);
+JitConstants ActivationKernelRef::GetJitConstants(const activation_params& params, DispatchData dispatchData) const {
+    auto jit = ActivationKernelBase::GetJitConstants(params, dispatchData);
     auto input_dt = params.inputs[0].GetDType();
 
     if (!params.fused_ops.empty()) {

diff --git a/...e/thirdparty/clDNN/kernel_selector/core/actual_kernels/activation/activation_kernel_ref.h b/...e/thirdparty/clDNN/kernel_selector/core/actual_kernels/activation/activation_kernel_ref.h
@@ -27,7 +27,7 @@ class ActivationKernelRef : public ActivationKernelBase {
 
     KernelsData GetKernelsData(const Params& params, const optional_params& options) const override;
     ParamsKey GetSupportedKey() const override;
-    JitConstants GetJitConstants(const activation_params& params, DispatchData kd) const override;
+    JitConstants GetJitConstants(const activation_params& params, DispatchData dispatchData) const override;
     std::vector<FusedOpType> GetSupportedFusedOps() const override {
         return {FusedOpType::QUANTIZE,
                 FusedOpType::SCALE,

diff --git a/...rdparty/clDNN/kernel_selector/core/actual_kernels/arg_max_min/arg_max_min_kernel_axis.cpp b/...rdparty/clDNN/kernel_selector/core/actual_kernels/arg_max_min/arg_max_min_kernel_axis.cpp
@@ -86,22 +86,11 @@ KernelsData ArgMaxMinKernelAxis::GetKernelsData(const Params& params, const opti
     }
     const arg_max_min_params& orgParams = static_cast<const arg_max_min_params&>(params);
 
-    DispatchData runInfo;
-    runInfo.fp16UnitUsed = orgParams.inputs[0].GetDType() == Datatype::F16;
-
     size_t sort_size = orgParams.argMaxMinSortType == ArgMaxMinSortType::VALUE ? getSortSize(orgParams) : 1;
 
-    std::vector<size_t> local, global;
-    global = { Align(getOperationNumber(orgParams), 32), sort_size, 1 };
-    local = GetOptimalLocalWorkGroupSizes(global, params.engineInfo);
-
-    runInfo.gws0 = global[0];
-    runInfo.gws1 = global[1];
-    runInfo.gws2 = global[2];
-
-    runInfo.lws0 = local[0];
-    runInfo.lws1 = local[1];
-    runInfo.lws2 = local[2];
+    DispatchData dispatchData;
+    dispatchData.gws = { Align(getOperationNumber(orgParams), 32), sort_size, 1 };
+    dispatchData.lws = GetOptimalLocalWorkGroupSizes(dispatchData.gws, params.engineInfo);
 
     KernelData kd = KernelData::Default<arg_max_min_params>(params);
 
@@ -110,7 +99,7 @@ KernelsData ArgMaxMinKernelAxis::GetKernelsData(const Params& params, const opti
     auto jit = CreateJit(kernelName, cldnn_jit, entry_point);
 
     auto& kernel = kd.kernels[0];
-    FillCLKernelData(kernel, runInfo, params.engineInfo, kernelName, jit, entry_point);
+    FillCLKernelData(kernel, dispatchData, params.engineInfo, kernelName, jit, entry_point);
 
     if (orgParams.outputs_num == 2) {
         kernel.arguments.push_back({ArgumentDescriptor::Types::INPUT, 1});

diff --git a/...rdparty/clDNN/kernel_selector/core/actual_kernels/arg_max_min/arg_max_min_kernel_base.cpp b/...rdparty/clDNN/kernel_selector/core/actual_kernels/arg_max_min/arg_max_min_kernel_base.cpp
@@ -1,5 +1,5 @@
 /*
-// Copyright (c) 2018 Intel Corporation
+// Copyright (c) 2018-2020 Intel Corporation
 //
 // Licensed under the Apache License, Version 2.0 (the "License");
 // you may not use this file except in compliance with the License.
@@ -37,20 +37,12 @@ JitConstants ArgMaxMinKernelBase::GetJitConstants(const arg_max_min_params& para
 }
 
 ArgMaxMinKernelBase::DispatchData ArgMaxMinKernelBase::SetDefault(const arg_max_min_params& params) const {
-    DispatchData kd;
+    DispatchData dispatchData;
 
-    kd.fp16UnitUsed = params.inputs[0].GetDType() == Datatype::F16;
+    dispatchData.gws = { 128, params.inputs[0].Batch().v, 1 };
+    dispatchData.lws = { 128, 1, 1 };
 
-    // Determine global work sizes.
-    kd.gws0 = 128;
-    kd.gws1 = params.inputs[0].Batch().v;
-    kd.gws2 = 1;
-
-    kd.lws0 = 128;
-    kd.lws1 = 1;
-    kd.lws2 = 1;
-
-    return kd;
+    return dispatchData;
 }
 
 KernelsData ArgMaxMinKernelBase::GetCommonKernelsData(const Params& params, const optional_params& options, float estimatedTime) const {
@@ -60,7 +52,7 @@ KernelsData ArgMaxMinKernelBase::GetCommonKernelsData(const Params& params, cons
 
     const arg_max_min_params& orgParams = static_cast<const arg_max_min_params&>(params);
 
-    DispatchData runInfo = SetDefault(orgParams);
+    DispatchData dispatchData = SetDefault(orgParams);
 
     KernelData kd = KernelData::Default<arg_max_min_params>(params);
 
@@ -69,7 +61,7 @@ KernelsData ArgMaxMinKernelBase::GetCommonKernelsData(const Params& params, cons
     auto jit = CreateJit(kernelName, cldnn_jit, entry_point);
 
     auto& kernel = kd.kernels[0];
-    FillCLKernelData(kernel, runInfo, params.engineInfo, kernelName, jit, entry_point);
+    FillCLKernelData(kernel, dispatchData, params.engineInfo, kernelName, jit, entry_point);
 
     kd.estimatedTime = estimatedTime;
 

diff --git a/...irdparty/clDNN/kernel_selector/core/actual_kernels/arg_max_min/arg_max_min_kernel_opt.cpp b/...irdparty/clDNN/kernel_selector/core/actual_kernels/arg_max_min/arg_max_min_kernel_opt.cpp
@@ -1,5 +1,5 @@
 /*
-// Copyright (c) 2018 Intel Corporation
+// Copyright (c) 2018-2020 Intel Corporation
 //
 // Licensed under the Apache License, Version 2.0 (the "License");
 // you may not use this file except in compliance with the License.
@@ -37,11 +37,11 @@ KernelsData ArgMaxMinKernelOpt::GetKernelsData(const Params& params, const optio
 
     const arg_max_min_params& orgParams = static_cast<const arg_max_min_params&>(params);
 
-    int topK = orgParams.topK;
-    long size = (long)(orgParams.inputs[0].X().v * orgParams.inputs[0].Y().v * orgParams.inputs[0].Feature().v) / 8;
-    long outSize = size / 16 * topK;
+    size_t topK = orgParams.topK;
+    size_t size = (size_t)(orgParams.inputs[0].X().v * orgParams.inputs[0].Y().v * orgParams.inputs[0].Feature().v) / 8;
+    size_t outSize = size / 16 * topK;
     int kernelAmount = 1;
-    for (; outSize > 128; outSize = (long)((outSize / 128 + 1) * topK)) {
+    for (; outSize > 128; outSize = (size_t)((outSize / 128 + 1) * topK)) {
         kernelAmount++;
     }
     KernelData kd = KernelData::Default<arg_max_min_params>(params, kernelAmount);
@@ -57,22 +57,15 @@ KernelsData ArgMaxMinKernelOpt::GetKernelsData(const Params& params, const optio
         newParams.inputs[0] = input;
 
         auto& kernel = kd.kernels[i];
-        DispatchData runInfo = SetDefault(newParams);
+        DispatchData dispatchData = SetDefault(newParams);
         auto cldnnJit = GetJitConstants(newParams);
         auto entryPoint = GetEntryPoint(kernelName, newParams.layerID, options);
         auto jit = CreateJit(kernelName, cldnnJit, entryPoint);
 
-        runInfo.fp16UnitUsed = orgParams.inputs[0].GetDType() == Datatype::F16;
+        dispatchData.gws = { Align(size, 16), orgParams.inputs[0].Batch().v, 1 };
+        dispatchData.lws = { 16, 1, 1 };
 
-        runInfo.gws0 = Align(size, 16);
-        runInfo.gws1 = orgParams.inputs[0].Batch().v;  // B
-        runInfo.gws2 = 1;
-
-        runInfo.lws0 = 16;
-        runInfo.lws1 = 1;
-        runInfo.lws2 = 1;
-
-        FillCLKernelData(kernel, runInfo, params.engineInfo, kernelName, jit, entryPoint);
+        FillCLKernelData(kernel, dispatchData, params.engineInfo, kernelName, jit, entryPoint);
         size = (size / 128 + 1) * topK;
     }
 

diff --git a/...N/kernel_selector/core/actual_kernels/average_unpooling/average_unpooling_kernel_base.cpp b/...N/kernel_selector/core/actual_kernels/average_unpooling/average_unpooling_kernel_base.cpp
@@ -1,5 +1,5 @@
 /*
-// Copyright (c) 2018 Intel Corporation
+// Copyright (c) 2018-2020 Intel Corporation
 //
 // Licensed under the Apache License, Version 2.0 (the "License");
 // you may not use this file except in compliance with the License.
@@ -38,32 +38,30 @@ AverageUnpoolingKernelBase::DispatchData AverageUnpoolingKernelBase::SetDefault(
     const average_unpooling_params& params) const {
     const auto& input = params.inputs[0];
 
-    DispatchData kd;
+    DispatchData dispatchData;
 
     if (input.GetLayout() == DataLayout::bfyx || input.GetLayout() == DataLayout::byxf) {
         // Determine global work sizes.
-        kd.gws2 = input.Batch().v * input.Feature().v;  // B, F
-        kd.gws0 = Align(input.X().v, 32);               // X
-        kd.gws1 = input.Y().v;                          // Y
+        dispatchData.gws =  { Align(input.X().v, 32),               // X
+                              input.Y().v,                          // Y
+                              input.Batch().v * input.Feature().v,  // B, F
+                            };
 
-        kd.lws0 = 32;
-        kd.lws1 = 1;
-        kd.lws2 = 1;
+        dispatchData.lws = { 32, 1, 1 };
     } else {
         // Determine global work sizes.
-        kd.gws0 = input.Batch().v * input.Feature().v;  // B, F
-        kd.gws1 = input.X().v;                          // X
-        kd.gws2 = input.Y().v;                          // Y
-
-        kd.lws0 = std::min(std::max(kd.gws0, static_cast<size_t>(1)), static_cast<size_t>(32));
-        while (kd.gws0 % kd.lws0 != 0) {
-            --kd.lws0;
+        dispatchData.gws = { input.Batch().v * input.Feature().v,  // B, F
+                             input.X().v,                          // X
+                             input.Y().v };                        // Y
+
+        dispatchData.lws = {1, 1, 1};
+        dispatchData.lws[0] = std::min(std::max(dispatchData.gws[0], static_cast<size_t>(1)), static_cast<size_t>(32));
+        while (dispatchData.gws[0] % dispatchData.lws[0] != 0) {
+            --dispatchData.lws[0];
         }
-        kd.lws1 = 1;
-        kd.lws2 = 1;
     }
 
-    return kd;
+    return dispatchData;
 }
 
 KernelsData AverageUnpoolingKernelBase::GetCommonKernelsData(const Params& params,
@@ -75,7 +73,7 @@ KernelsData AverageUnpoolingKernelBase::GetCommonKernelsData(const Params& param
 
     const average_unpooling_params& orgParams = static_cast<const average_unpooling_params&>(params);
 
-    DispatchData runInfo = SetDefault(orgParams);
+    DispatchData dispatchData = SetDefault(orgParams);
 
     KernelData kd = KernelData::Default<average_unpooling_params>(params);
 
@@ -84,10 +82,10 @@ KernelsData AverageUnpoolingKernelBase::GetCommonKernelsData(const Params& param
     auto jit = CreateJit(kernelName, cldnn_jit, entry_point);
 
     auto& kernel = kd.kernels[0];
-    FillCLKernelData(kernel, runInfo, params.engineInfo, kernelName, jit, entry_point);
+    FillCLKernelData(kernel, dispatchData, params.engineInfo, kernelName, jit, entry_point);
 
     kd.estimatedTime = estimatedTime;
 
     return {kd};
 }
-}  // namespace kernel_selector
+}  // namespace kernel_selector