Fix accuracy issue in sum_OdO kernel

Jiashi Li · Jiashi Li · commit c7590278ce1b · 2025-08-14T09:37:44.000+08:00
diff --git a/csrc/sm100/kernel/fmha_kernel_bwd_sum_OdO.hpp b/csrc/sm100/kernel/fmha_kernel_bwd_sum_OdO.hpp
@@ -140,7 +140,7 @@ struct FmhaKernelBwdSumOdO {
         *reinterpret_cast<Vec*>(value_dO) = *reinterpret_cast<const Vec*>(&ptr_dO_bhq[idx_d]);
 
         for (int v = 0; v < kElementsPerLoad; v++) {
-          acc += value_O[v] * value_dO[v];
+          acc += ElementAcc(value_O[v]) * ElementAcc(value_dO[v]);
         }
       }
 

Original file line number	Diff line number	Diff line change
`@@ -140,7 +140,7 @@ struct FmhaKernelBwdSumOdO {`
`140`	`140`	`reinterpret_cast<Vec>(value_dO) = reinterpret_cast<const Vec>(&ptr_dO_bhq[idx_d]);`
`141`	`141`
`142`	`142`	`for (int v = 0; v < kElementsPerLoad; v++) {`
`143`		`- acc += value_O[v] * value_dO[v];`
	`143`	`+ acc += ElementAcc(value_O[v]) * ElementAcc(value_dO[v]);`
`144`	`144`	`}`
`145`	`145`	`}`
`146`	`146`