Merge pull request openvinotoolkit#8 from EunsooSheenIntel/taylor/per…

…mute_opt_new

Add GetBestLwsFromGws in PermuteKernel_tile_8x8

inference-engine/thirdparty/clDNN/kernel_selector/core/actual_kernels/permute/permute_kernel_tile_8x8.cpp

@@ -17,7 +17,10 @@
 #include "kernel_selector_utils.h"
 #include <string>
 #include <functional>
+#include <cmath>
+
 #define CEIL_DIV(A, B) ((A + B - 1)/(B))
+
 namespace kernel_selector {
 ParamsKey PermuteKernel_tile_8x8::GetSupportedKey() const {
     ParamsKey k;
@@ -159,39 +162,72 @@ JitConstants PermuteKernel_tile_8x8::GetJitConstants(const permute_params& param
     return jit;
 }
 
+static std::vector<size_t> GetBestLwsFromGws(const std::vector<size_t>& gws, const size_t tile_height, const size_t tile_width)
+{
+    std::vector<size_t> lws{1,1,1};
+
+    // SLM size: float32 * tile_height * tile_width * work_items <= 64K
+    size_t max_num_work_items = std::min(256lu, 65536 / (4 * tile_height * tile_width));
+
+    // set lws[0] first
+    size_t max_divider = static_cast<size_t>(std::sqrt(gws[0])+1);
+    for (size_t divider = 1; divider <= max_divider; ++divider)
+    {
+        if (gws[0] % divider == 0)
+        {
+            const size_t lws0 = gws[0] / divider;
+            if (lws0 <= max_num_work_items)
+            {
+                lws[0] = std::max(lws[0], lws0);
+            }
+            if (divider <= max_num_work_items)
+            {
+                lws[0] = std::max(lws[0], divider);
+            }
+        }
+    }
+
+    // set lws[2]
+    max_num_work_items /= lws[0];
+    max_divider = static_cast<size_t>(std::sqrt(gws[2])+1);
+    for (size_t divider = 1; divider <= max_divider; ++divider)
+    {
+        if (gws[2] % divider == 0)
+        {
+            const size_t lws2 = gws[2] / divider;
+            if (lws2 <= max_num_work_items)
+            {
+                lws[2] = std::max(lws[2], lws2);
+            }
+            if (divider <= max_num_work_items)
+            {
+                lws[2] = std::max(lws[2], divider);
+            }
+        }
+    }
+    return lws;
+}
+
 CommonDispatchData PermuteKernel_tile_8x8::SetDefault(const permute_params& params) const {
     CommonDispatchData dispatchData;
     const auto& in =  params.inputs[0];
     const auto& tile_w = params.tile_w;
     const auto& tile_h = params.tile_h;
     switch (in.GetLayout()) {
         case DataLayout::bfyx:
-            // for f800, y64, x64 
-            // gws : 64/8, 64, 800
-            //dispatchData.gws = {in.X().v / tile_w, in.Y().v, (in.Feature().v / tile_h) * in.Batch().v};
-            //dispatchData.lws = {2, 1, 50}; // TODO
-
-            // for f800, y64, x64 
-            // gws : 64/8, 64, 810
             dispatchData.gws = {CEIL_DIV(in.X().v , tile_w), in.Y().v, CEIL_DIV(in.Feature().v, tile_h) * in.Batch().v};
-            dispatchData.lws = {2, 1, 51}; // TODO
             break;
         case DataLayout::bfzyx:
             dispatchData.gws = {CEIL_DIV(in.X().v , tile_w), in.Y().v * in.Z().v, CEIL_DIV(in.Feature().v, tile_h) * in.Batch().v};
-            dispatchData.lws = {64, 1, 2}; // TODO
-//            dispatchData.lws = {128, 1, 2}; // TODO
-//            dispatchData.lws = {64, 1, 2}; // TODO
-//            dispatchData.lws = {3, 1, 2}; // TODO
             break;
         case DataLayout::bfwzyx:
             dispatchData.gws = {CEIL_DIV(in.X().v , tile_w), in.Y().v * in.Z().v * in.W().v, CEIL_DIV(in.Feature().v, tile_h) * in.Batch().v};
-            dispatchData.lws = {64, 1, 2}; // TODO
             break;
         default:
             throw std::runtime_error("Unsupported combination\n");
             break;
     }
-
+    dispatchData.lws = GetBestLwsFromGws(dispatchData.gws, tile_h, tile_w);
     return dispatchData;
 }