[GPU] Use parallel sum reduction in RMS BFYX OPT kernel (openvinotool…

…kit#25896) ### Details: - Use parallel sum reduction for RMS BFYX OPT kernel - Improve heuristics ### Tickets: - 148937 Co-authored-by: Pavel Durandin <[email protected]>
nashez · Aug 6, 2024 · decdac6 · decdac6
1 parent 5ec4375
commit decdac6
Show file tree

Hide file tree

Showing 2 changed files with 9 additions and 6 deletions.
diff --git a/src/plugins/intel_gpu/src/kernel_selector/cl_kernels/rms_gpu_bfyx_opt.cl b/src/plugins/intel_gpu/src/kernel_selector/cl_kernels/rms_gpu_bfyx_opt.cl
@@ -92,12 +92,15 @@ KERNEL(rms_gpu_bfyx_opt)(
         slm_buf[get_sub_group_id()] = rms;
 
     barrier(CLK_LOCAL_MEM_FENCE);
-    if (in_data_idx == 0) {
-        for (uint i = 1; i < get_num_sub_groups(); ++i)
-        {
-            rms += slm_buf[i];
+    for (uint offset = get_num_sub_groups() / 2; offset > 0; offset /= 2) {
+        if (in_data_idx < offset) {
+            slm_buf[in_data_idx] += slm_buf[in_data_idx + offset];
         }
-        rms = rms / data_size;
+        barrier(CLK_LOCAL_MEM_FENCE);
+    }
+
+    if (in_data_idx == 0) {
+        rms = slm_buf[0] / data_size;
         slm_buf[0] = native_powr(sqrt(rms + TO_ACCUMULATOR_TYPE(EPSILON)), -1);
     }
     barrier(CLK_LOCAL_MEM_FENCE);

diff --git a/src/plugins/intel_gpu/src/kernel_selector/kernels/rms/rms_kernel_bfyx_opt.cpp b/src/plugins/intel_gpu/src/kernel_selector/kernels/rms/rms_kernel_bfyx_opt.cpp
@@ -120,7 +120,7 @@ RMSKernelBase::DispatchData RMSKernelBfyxOpt::SetDefault(const rms_params& param
 
         dispatchData.itemsNum = dispatchData.dataSize;
         // Compute maximum possible LWS that does not exceed device capabilities and optimizes number of global memory reads
-        while ((dispatchData.itemsNum > 32 || dispatchData.lws[0] < dispatchData.itemsNum) && (2 * dispatchData.lws[0] <= max_lws)) {
+        while ((dispatchData.itemsNum > 8 || dispatchData.lws[0] < dispatchData.itemsNum) && (2 * dispatchData.lws[0] <= max_lws)) {
             dispatchData.lws[0] *= 2;
             dispatchData.itemsNum /= 2;
         }