[GPU] Add only necessary LOOP unroll params for jit constants

Author: kelvinchoi-intel

src/plugins/intel_gpu/src/kernel_selector/kernels/convolution/convolution_kernel_base.cpp

@@ -25,7 +25,7 @@ bool ConvolutionKernelBase::Validate(const Params& p) const {
     return true;
 }
 
-JitConstants ConvolutionKernelBase::GetJitConstants(const convolution_params& params, const DispatchData& dispatchData) const {
+JitConstants ConvolutionKernelBase::GetJitConstants(const convolution_params& params, const DispatchData& dispatchData, bool LoopUnrollParams) const {
     JitConstants mem_consts = WeightBiasKernelBase::GetJitConstants(params);
     mem_consts.Merge(GetFusedPrimitivesJitConstants(params, dispatchData));
     const auto& padding = params.padding_begin;
@@ -83,19 +83,21 @@ JitConstants ConvolutionKernelBase::GetJitConstants(const convolution_params& pa
         }
     }
 
-    std::vector<uint32_t> unrollLoopParams{params.filterSize.x,
-                                           params.filterSize.y,
-                                           (uint32_t)dispatchData.gemmStyle.globalWorkSizeDX,
-                                           (uint32_t)dispatchData.gemmStyle.globalWorkSizeDY,
-                                           (uint32_t)dispatchData.gemmStyle.globalWorkSizeDZ,
-                                           (uint32_t)dispatchData.gemmStyle.subBlockDimM,
-                                           (uint32_t)dispatchData.gemmStyle.subBlockDimK,
-                                           (uint32_t)dispatchData.gemmStyle.subBlockDimN};
+    if (LoopUnrollParams) {
+        std::vector<uint32_t> unrollLoopParams{params.filterSize.x,
+                                            params.filterSize.y,
+                                            (uint32_t)dispatchData.gemmStyle.globalWorkSizeDX,
+                                            (uint32_t)dispatchData.gemmStyle.globalWorkSizeDY,
+                                            (uint32_t)dispatchData.gemmStyle.globalWorkSizeDZ,
+                                            (uint32_t)dispatchData.gemmStyle.subBlockDimM,
+                                            (uint32_t)dispatchData.gemmStyle.subBlockDimK,
+                                            (uint32_t)dispatchData.gemmStyle.subBlockDimN};
 
-    auto loopCount = *std::max_element(unrollLoopParams.begin(), unrollLoopParams.end());
+        auto loopCount = *std::max_element(unrollLoopParams.begin(), unrollLoopParams.end());
 
-    JitConstants mem_consts_loop = MakeLoopUnrollParamsJitConstants(loopCount);
-    mem_consts.Merge(mem_consts_loop);
+        JitConstants mem_consts_loop = MakeLoopUnrollParamsJitConstants(loopCount);
+        mem_consts.Merge(mem_consts_loop);
+    }
 
     return mem_consts;
 }

src/plugins/intel_gpu/src/kernel_selector/kernels/convolution/convolution_kernel_base.h

@@ -36,10 +36,8 @@ class ConvolutionKernelBase : public WeightBiasKernelBase {
             size_t globalWorkSizeDZ;
         };
 
-        union {
-            CLDNNStyle cldnnStyle;
-            GEMMStyle gemmStyle;
-        };
+        CLDNNStyle cldnnStyle;
+        GEMMStyle gemmStyle;
     };
 
     std::string GetAutoTuneOptions(int autoTuneIndex) const;
@@ -53,7 +51,7 @@ class ConvolutionKernelBase : public WeightBiasKernelBase {
     virtual bool NeedPaddedInput() const { return false; }
     bool Validate(const Params& p) const override;
     using WeightBiasKernelBase::GetJitConstants;
-    virtual JitConstants GetJitConstants(const convolution_params& params, const DispatchData& dispatchData) const;
+    virtual JitConstants GetJitConstants(const convolution_params& params, const DispatchData& dispatchData, bool LoopUnrollParams = false) const;
     virtual JitConstants GetFusedPrimitivesJitConstants(const convolution_params& params, const DispatchData& dispatchData) const;
     virtual DispatchData SetDefault(const convolution_params& params, int autoTuneIndex = -1) const;
     static bool CheckWorkGroups(const DispatchData&);

src/plugins/intel_gpu/src/kernel_selector/kernels/convolution/convolution_kernel_bfyx_direct_10_12_16.cpp

@@ -33,7 +33,7 @@ DeviceFeaturesKey ConvolutionKernel_bfyx_Direct_10_10_12::get_required_device_fe
 
 JitConstants ConvolutionKernel_bfyx_Direct_10_10_12::GetJitConstants(const convolution_params& cp,
                                                                      const DispatchData& dispatchData) const {
-    JitConstants jit = Parent::GetJitConstants(cp, dispatchData);
+    JitConstants jit = Parent::GetJitConstants(cp, dispatchData, true);
 
     jit.AddConstants({
         MakeJitConstant("ALIGNED_OFM", RoundUp(cp.outputs[0].Feature().v / cp.groups, dispatchData.gemmStyle.subBlockDimN) * cp.groups),

src/plugins/intel_gpu/src/kernel_selector/kernels/convolution/convolution_kernel_bfyx_gemm_like.cpp

@@ -44,7 +44,7 @@ std::string ConvolutionKernel_bfyx_GEMMLike::GetKernelName(const convolution_par
 
 JitConstants ConvolutionKernel_bfyx_GEMMLike::GetJitConstants(const convolution_params& params,
                                                               const DispatchData& dispatchData) const {
-    JitConstants jit = Parent::GetJitConstants(params, dispatchData);
+    JitConstants jit = Parent::GetJitConstants(params, dispatchData, true);
 
     jit.AddConstants({
         MakeJitConstant("ALIGNED_OFM_PER_GROUP", RoundUp(params.outputs[0].Feature().v / params.groups, dispatchData.gemmStyle.subBlockDimN)),

src/plugins/intel_gpu/src/kernel_selector/kernels/convolution/convolution_kernel_bfyx_os_iyx_osv16.cpp

@@ -218,7 +218,7 @@ JitConstants ConvolutionKernel_bfyx_os_iyx_osv16::GetJitConstants(const convolut
     const size_t of_threads_per_batch = RoundUp(of_maps_per_group, sub_group_size);
     size_t leftovers = of_threads_per_batch - of_maps_per_group;
 
-    auto jit = Parent::GetJitConstants(params, dispatchData);
+    auto jit = Parent::GetJitConstants(params, dispatchData, true);
 
     if (!params.fused_ops.empty()) {
         auto input_dt = GetUnitType(params);