[GPU] Fix bfyx conv degradation for gen9 (#8911)

* Fix for bfyx conv performance issue.
PR 6399 was for wide palne + shallow featuremap cases.
If the plane is too small so that the increased computation due toe the alignment for block_size is relatively large, performance drops.

* Fix for bfyx conv performance issue. Use subgroup size 16 when output feature size is >= 8 instead of > 8

* Set bfyx convolution subgroup size to be 16 when output_feature size is 8 and the dtype is fp16

inference-engine/thirdparty/clDNN/kernel_selector/core/actual_kernels/convolution/convolution_kernel_bfyx_os_iyx_osv16.cpp

@@ -80,7 +80,7 @@ static std::pair<size_t, size_t> get_bfyx_req_input_block_dims(size_t output_blo
     return std::make_pair(input_block_array_size, input_block_read_width);
 }
 
-static void shrink_blocks_to_output_size(size_t output_x, size_t output_y, size_t& block_x, size_t& block_y) {
+static void shrink_blocks_to_output_size(size_t output_x, size_t output_y, size_t& block_x, size_t& block_y, size_t sub_group_size) {
     // how many elements we will compute in each dimension
     size_t computed_x = Align(output_x, block_x);
     size_t computed_y = Align(output_y, block_y);
@@ -94,8 +94,10 @@ static void shrink_blocks_to_output_size(size_t output_x, size_t output_y, size_
     block_x -= unused_x / simds_x;
     block_y -= unused_y / simds_y;
 
-    block_x = Align(block_x, 2);
-    block_y = Align(block_y, 2);
+    if (simds_x * simds_y >= sub_group_size) {
+        block_x = Align(block_x, 2);
+        block_y = Align(block_y, 2);
+    }
 }
 
 ConvolutionKernel_bfyx_os_iyx_osv16::AutoTuneOption ConvolutionKernel_bfyx_os_iyx_osv16::GetAutoTuneOptions(
@@ -145,7 +147,7 @@ ConvolutionKernel_bfyx_os_iyx_osv16::AutoTuneOption ConvolutionKernel_bfyx_os_iy
     // if this is not 1x1 batch1 case then shrink filters, other way we're memory bound and it's best to use 16x1 block
     // sizes
     if (cp.filterSize.x != 1 || cp.filterSize.y != 1 || cp.output.Batch().v != 1) {
-        shrink_blocks_to_output_size(cp.output.X().v, cp.output.Y().v, option.blockWidth, option.blockHeight);
+        shrink_blocks_to_output_size(cp.output.X().v, cp.output.Y().v, option.blockWidth, option.blockHeight, sub_group_size);
     }
     return option;
 }

inference-engine/thirdparty/clDNN/kernel_selector/core/actual_kernels/convolution/convolution_kernel_bfyx_os_iyx_osv16.h

@@ -40,7 +40,12 @@ protected:
             // Smaller # EU tends to be computation bounds.
             // In such case, using larger worksize will result in larger computational inefficiency
             // w.r.t the unalined output feature
-            return (params.output.Feature().v > 8 || !IsSIMDSizeSupported(params.engineInfo, 8)) ? 16 : 8;
+            if (params.output.Feature().v > 8 || !IsSIMDSizeSupported(params.engineInfo, 8)
+               || (params.output.GetDType() == Datatype::F16) && params.output.Feature().v == 8) {
+                return 16;
+            } else {
+                return 8;
+            }
         } else {
             return 16;
         }