fix win32 cu12 (#59852)

wwbitejotunn · web-flow · commit 5be87ba0649c · 2023-12-09T15:58:14.000+08:00
diff --git a/paddle/phi/kernels/funcs/weight_only_gemv.cu b/paddle/phi/kernels/funcs/weight_only_gemv.cu
@@ -729,7 +729,7 @@ __global__ void weight_only_batched_gemv_multi_warp(const int8_t* qweight,
               *reinterpret_cast<HALF_2_TYPE*>(in_v + y),
               v);
         }
-        accumulator[b] += v.x + v.y;
+        accumulator[b] = accumulator[b] + static_cast<T>(v.x + v.y);
       } else {
 #pragma unroll
         for (int x = 0; x < NPerBlock / 2; ++x) {

Original file line number	Diff line number	Diff line change
`@@ -729,7 +729,7 @@ __global__ void weight_only_batched_gemv_multi_warp(const int8_t* qweight,`
`729`	`729`	`reinterpret_cast<HALF_2_TYPE>(in_v + y),`
`730`	`730`	`v);`
`731`	`731`	`}`
`732`		`- accumulator[b] += v.x + v.y;`
	`732`	`+ accumulator[b] = accumulator[b] + static_cast<T>(v.x + v.y);`
`733`	`733`	`} else {`
`734`	`734`	`#pragma unroll`
`735`	`735`	`for (int x = 0; x < NPerBlock / 2; ++x) {`