Replace with powf (#2896)

Summary:

Replace `pow` with `powf` for better performance

Differential Revision: D58560455

Author: spcyppt

fbgemm_gpu/codegen/genscript/jinja_environment.py

@@ -338,7 +338,7 @@ def compute_global_weight_decay(is_global_weight_decay_kernel: bool) -> str:
     if is_global_weight_decay_kernel:
         return """
         const auto prev_iter = prev_iter_dev[linear_index];
-        const auto global_weight_decay = prev_iter == 0 ? 1 : max(gwd_lower_bound, std::pow(weight_decay_base, iter - prev_iter - 1));
+        const auto global_weight_decay = prev_iter == 0 ? 1 : max(gwd_lower_bound, powf(weight_decay_base, iter - prev_iter - 1));
         if (threadIdx.x == 0) {
             prev_iter_dev[linear_index] = iter;
         }

fbgemm_gpu/codegen/training/forward/embedding_forward_split_kernel_template.cu

@@ -188,7 +188,7 @@ using namespace fbgemm_gpu;
         {%- if is_gwd_kernel %}
         // if l > L or prev_iter == 0, global_weight_decay = 1
         const auto prev_it = prev_iter[idx];
-        const auto global_weight_decay = (l > L || prev_it == 0) ? 1 : max(gwd_lower_bound, std::pow(weight_decay_base, iter - prev_it - 1));
+        const auto global_weight_decay = (l > L || prev_it == 0) ? 1 : max(gwd_lower_bound, powf(weight_decay_base, iter - prev_it - 1));
         {%- endif %}
 
         {%- if weighted %}