diff --git a/libvmaf/src/feature/cuda/integer_adm_cuda.c b/libvmaf/src/feature/cuda/integer_adm_cuda.c
index cdf9f482c..9dec3c11c 100644
--- a/libvmaf/src/feature/cuda/integer_adm_cuda.c
+++ b/libvmaf/src/feature/cuda/integer_adm_cuda.c
@@ -149,7 +149,6 @@ void adm_dwt2_s123_combined_device(AdmStateCuda *s,const int32_t *d_i4_scale, in
     const int BLOCK_Y = (h + 1) / 2;
     
     void * args_vert[] = {&d_i4_scale, &tmp_buf, &w, &h, &img_stride, &*p};
-    const int num_threads = 128;
     switch (scale) {
         case 1:
             CHECK_CUDA(cuLaunchKernel(s->func_dwt_s123_combined_vert_kernel_0_0_int32_t,
diff --git a/libvmaf/src/feature/cuda/integer_psnr_cuda.c b/libvmaf/src/feature/cuda/integer_psnr_cuda.c
index d121c7c74..d29bdabcb 100644
--- a/libvmaf/src/feature/cuda/integer_psnr_cuda.c
+++ b/libvmaf/src/feature/cuda/integer_psnr_cuda.c
@@ -23,9 +23,8 @@
 #include <stddef.h>
 #include <string.h>
 
+#include "common.h"
 #include "cuda_helper.cuh"
-#include "libvmaf/vmaf_cuda_state.h"
-#include "picture_cuda.h"
 
 #include "feature_collector.h"
 #include "feature_extractor.h"
@@ -231,13 +230,8 @@ static int extract_fex_cuda(VmafFeatureExtractor *fex,
     CHECK_CUDA(cuEventRecord(s->event, vmaf_cuda_picture_get_stream(ref_pic)));
     // This event ensures the input buffer is consumed
     CHECK_CUDA(cuStreamWaitEvent(s->str, s->event, CU_EVENT_WAIT_DEFAULT));
-    // CHECK_CUDA(cuCtxPushCurrent(fex->cu_state->ctx));
-    // CHECK_CUDA(cuEventDestroy(s->event));
-    // CHECK_CUDA(cuEventCreate(&s->event, CU_EVENT_DEFAULT));
-    // CHECK_CUDA(cuCtxPopCurrent(NULL));
     
-    // Download sad
-    // CHECK_CUDA(cuStreamSynchronize(s->host_stream));
+    // Download SSE
     CHECK_CUDA(cuMemcpyDtoHAsync(s->sse_host, (CUdeviceptr)s->sse_device->data,
                 sizeof(uint64_t) * 3, s->str));
     CHECK_CUDA(cuEventRecord(s->finished, s->str));
diff --git a/libvmaf/src/libvmaf.c b/libvmaf/src/libvmaf.c
index cc167376f..6e0574aca 100644
--- a/libvmaf/src/libvmaf.c
+++ b/libvmaf/src/libvmaf.c
@@ -593,7 +593,7 @@ static int translate_picture_device(VmafContext *vmaf, VmafPicture *pic,
         return err;
     }
 
-    err = vmaf_cuda_picture_download_async(pic, pic_host, 0x3);
+    err = vmaf_cuda_picture_download_async(pic, pic_host, 0xF);
     if (err) {
         vmaf_log(VMAF_LOG_LEVEL_ERROR,
                  "problem moving cuda pic into host buffer\n");