diff --git a/src/RandomX/aes_cuda.hpp b/src/RandomX/aes_cuda.hpp
index 65b54e2..af0f8ec 100644
--- a/src/RandomX/aes_cuda.hpp
+++ b/src/RandomX/aes_cuda.hpp
@@ -793,12 +793,11 @@ __global__ void hashAes1Rx4(const void* input, void* hash, uint32_t batch_size)
 
 #define ITER(m) \
 	{ \
-		uint32_t k[4], y[4]; \
-		*(uint4*)(k) = p[m * 4]; \
-		y[0] = t0[get_byte(x[0], 0)] ^ t1[get_byte(x[1], s1)] ^ t2[get_byte(x[2], 16)] ^ t3[get_byte(x[3], s3)] ^ k[0]; \
-		y[1] = t0[get_byte(x[1], 0)] ^ t1[get_byte(x[2], s1)] ^ t2[get_byte(x[3], 16)] ^ t3[get_byte(x[0], s3)] ^ k[1]; \
-		y[2] = t0[get_byte(x[2], 0)] ^ t1[get_byte(x[3], s1)] ^ t2[get_byte(x[0], 16)] ^ t3[get_byte(x[1], s3)] ^ k[2]; \
-		y[3] = t0[get_byte(x[3], 0)] ^ t1[get_byte(x[0], s1)] ^ t2[get_byte(x[1], 16)] ^ t3[get_byte(x[2], s3)] ^ k[3]; \
+		uint32_t y[4]; \
+		y[0] = t0[get_byte(x[0], 0)] ^ t1[get_byte(x[1], s1)] ^ t2[get_byte(x[2], 16)] ^ t3[get_byte(x[3], s3)] ^ p[m * 4].x; \
+		y[1] = t0[get_byte(x[1], 0)] ^ t1[get_byte(x[2], s1)] ^ t2[get_byte(x[3], 16)] ^ t3[get_byte(x[0], s3)] ^ p[m * 4].y; \
+		y[2] = t0[get_byte(x[2], 0)] ^ t1[get_byte(x[3], s1)] ^ t2[get_byte(x[0], 16)] ^ t3[get_byte(x[1], s3)] ^ p[m * 4].z; \
+		y[3] = t0[get_byte(x[3], 0)] ^ t1[get_byte(x[0], s1)] ^ t2[get_byte(x[1], 16)] ^ t3[get_byte(x[2], s3)] ^ p[m * 4].w; \
 		x[0] = y[0]; \
 		x[1] = y[1]; \
 		x[2] = y[2]; \