Merge branch 'improve-gpu-sched' into bert-opt-layernorm

src/targets/gpu/compile_hip_code_object.cpp

@@ -138,16 +138,16 @@ compute_global_for(context& ctx, std::size_t n, std::size_t over)
         std::size_t groups     = (n + local - 1) / local;
         std::size_t max_blocks = max_global / local;
         std::size_t nglobal    = std::min(max_blocks * over, groups) * local;
-        return nglobal;
+        return std::min(nglobal, n);
     };
 }
 
 std::size_t compute_block_size(std::size_t n, std::size_t max_block_size)
 {
-    size_t block_size = 128;
-    while(block_size <= max_block_size and block_size <= n)
-        block_size *= 2;
-    return block_size / 2;
+    const std::size_t min_block_size  = 64;
+    const std::size_t base_block_size = 32;
+    auto block_size                   = (((n - 1) / base_block_size + 1)) * base_block_size;
+    return std::min(std::max(min_block_size, block_size), max_block_size);
 }
 
 operation compile_hip_code_object(const std::string& content, hip_compile_options options)