diff --git a/gfx/common/vulkan_common.c b/gfx/common/vulkan_common.c
index c52d3e1d82..88b8dda6b7 100644
--- a/gfx/common/vulkan_common.c
+++ b/gfx/common/vulkan_common.c
@@ -1746,6 +1746,22 @@ static bool vulkan_context_init_device(gfx_ctx_vulkan_data_t *vk)
    vk->emulate_mailbox = vk->fullscreen;
 #endif
 
+   /* If we're emulating mailbox, stick to using fences rather than semaphores.
+    * Avoids some really weird driver bugs. */
+   if (!vk->emulate_mailbox)
+   {
+      if (vk->context.gpu_properties.deviceType == VK_PHYSICAL_DEVICE_TYPE_INTEGRATED_GPU)
+      {
+         vk->use_wsi_semaphore = true;
+         RARCH_LOG("[Vulkan]: Using semaphores for WSI acquire.\n");
+      }
+      else
+      {
+         vk->use_wsi_semaphore = false;
+         RARCH_LOG("[Vulkan]: Using fences for WSI acquire.\n");
+      }
+   }
+
    RARCH_LOG("[Vulkan]: Using GPU: %s\n", vk->context.gpu_properties.deviceName);
 
    {
@@ -2555,10 +2571,28 @@ static void vulkan_destroy_swapchain(gfx_ctx_vulkan_data_t *vk)
       if (vk->context.swapchain_fences[i] != VK_NULL_HANDLE)
          vkDestroyFence(vk->context.device,
                vk->context.swapchain_fences[i], NULL);
+      if (vk->context.swapchain_recycled_semaphores[i] != VK_NULL_HANDLE)
+         vkDestroySemaphore(vk->context.device,
+               vk->context.swapchain_recycled_semaphores[i], NULL);
+      if (vk->context.swapchain_wait_semaphores[i] != VK_NULL_HANDLE)
+         vkDestroySemaphore(vk->context.device,
+               vk->context.swapchain_wait_semaphores[i], NULL);
    }
 
-   memset(vk->context.swapchain_semaphores, 0, sizeof(vk->context.swapchain_semaphores));
-   memset(vk->context.swapchain_fences, 0, sizeof(vk->context.swapchain_fences));
+   if (vk->context.swapchain_acquire_semaphore != VK_NULL_HANDLE)
+      vkDestroySemaphore(vk->context.device,
+            vk->context.swapchain_acquire_semaphore, NULL);
+   vk->context.swapchain_acquire_semaphore = VK_NULL_HANDLE;
+
+   memset(vk->context.swapchain_semaphores, 0,
+         sizeof(vk->context.swapchain_semaphores));
+   memset(vk->context.swapchain_recycled_semaphores, 0,
+         sizeof(vk->context.swapchain_recycled_semaphores));
+   memset(vk->context.swapchain_wait_semaphores, 0,
+         sizeof(vk->context.swapchain_wait_semaphores));
+   memset(vk->context.swapchain_fences, 0,
+         sizeof(vk->context.swapchain_fences));
+   vk->context.num_recycled_acquire_semaphores = 0;
 }
 
 void vulkan_present(gfx_ctx_vulkan_data_t *vk, unsigned index)
@@ -2673,6 +2707,12 @@ void vulkan_context_destroy(gfx_ctx_vulkan_data_t *vk,
    }
 }
 
+static void vulkan_recycle_acquire_semaphore(struct vulkan_context *ctx, VkSemaphore sem)
+{
+   assert(ctx->num_recycled_acquire_semaphores < VULKAN_MAX_SWAPCHAIN_IMAGES);
+   ctx->swapchain_recycled_semaphores[ctx->num_recycled_acquire_semaphores++] = sem;
+}
+
 static void vulkan_acquire_clear_fences(gfx_ctx_vulkan_data_t *vk)
 {
    unsigned i;
@@ -2685,11 +2725,31 @@ static void vulkan_acquire_clear_fences(gfx_ctx_vulkan_data_t *vk)
          vk->context.swapchain_fences[i] = VK_NULL_HANDLE;
       }
       vk->context.swapchain_fences_signalled[i] = false;
+
+      if (vk->context.swapchain_wait_semaphores[i])
+         vulkan_recycle_acquire_semaphore(&vk->context, vk->context.swapchain_wait_semaphores[i]);
+      vk->context.swapchain_wait_semaphores[i] = VK_NULL_HANDLE;
    }
 
    vk->context.current_frame_index = 0;
 }
 
+static VkSemaphore vulkan_get_wsi_acquire_semaphore(struct vulkan_context *ctx)
+{
+   if (ctx->num_recycled_acquire_semaphores == 0)
+   {
+      VkSemaphoreCreateInfo sem_info = { VK_STRUCTURE_TYPE_SEMAPHORE_CREATE_INFO };
+      vkCreateSemaphore(ctx->device, &sem_info, NULL,
+            &ctx->swapchain_recycled_semaphores[ctx->num_recycled_acquire_semaphores++]);
+   }
+
+   VkSemaphore sem =
+      ctx->swapchain_recycled_semaphores[--ctx->num_recycled_acquire_semaphores];
+   ctx->swapchain_recycled_semaphores[ctx->num_recycled_acquire_semaphores] =
+      VK_NULL_HANDLE;
+   return sem;
+}
+
 static void vulkan_acquire_wait_fences(gfx_ctx_vulkan_data_t *vk)
 {
    VkFenceCreateInfo fence_info =
@@ -2712,6 +2772,10 @@ static void vulkan_acquire_wait_fences(gfx_ctx_vulkan_data_t *vk)
    else
       vkCreateFence(vk->context.device, &fence_info, NULL, next_fence);
    vk->context.swapchain_fences_signalled[index] = false;
+
+   if (vk->context.swapchain_wait_semaphores[index] != VK_NULL_HANDLE)
+       vulkan_recycle_acquire_semaphore(&vk->context, vk->context.swapchain_wait_semaphores[index]);
+   vk->context.swapchain_wait_semaphores[index] = VK_NULL_HANDLE;
 }
 
 static void vulkan_create_wait_fences(gfx_ctx_vulkan_data_t *vk)
@@ -2734,7 +2798,8 @@ void vulkan_acquire_next_image(gfx_ctx_vulkan_data_t *vk)
 {
    unsigned index;
    VkResult err;
-   VkFence fence;
+   VkFence fence = VK_NULL_HANDLE;
+   VkSemaphore semaphore = VK_NULL_HANDLE;
    VkFenceCreateInfo fence_info   =
    { VK_STRUCTURE_TYPE_FENCE_CREATE_INFO };
    VkSemaphoreCreateInfo sem_info =
@@ -2756,6 +2821,7 @@ retry:
       {
          /* We still don't have a swapchain, so just fake it ... */
          vk->context.current_swapchain_index = 0;
+         vk->context.current_frame_index = 0;
          vulkan_acquire_clear_fences(vk);
          vulkan_acquire_wait_fences(vk);
          vk->context.invalid_swapchain = true;
@@ -2772,14 +2838,17 @@ retry:
        * MAILBOX would do. */
       err   = vulkan_emulated_mailbox_acquire_next_image(
             &vk->mailbox, &vk->context.current_swapchain_index);
-      fence = VK_NULL_HANDLE;
    }
    else
    {
-      vkCreateFence(vk->context.device, &fence_info, NULL, &fence);
+      if (vk->use_wsi_semaphore)
+          semaphore = vulkan_get_wsi_acquire_semaphore(&vk->context);
+      else
+          vkCreateFence(vk->context.device, &fence_info, NULL, &fence);
+
       err = vkAcquireNextImageKHR(vk->context.device,
             vk->swapchain, UINT64_MAX,
-            VK_NULL_HANDLE, fence, &vk->context.current_swapchain_index);
+            semaphore, fence, &vk->context.current_swapchain_index);
 
 #ifdef ANDROID
       /* VK_SUBOPTIMAL_KHR can be returned on Android 10 
@@ -2796,9 +2865,27 @@ retry:
       if (fence != VK_NULL_HANDLE)
          vkWaitForFences(vk->context.device, 1, &fence, true, UINT64_MAX);
       vk->context.has_acquired_swapchain = true;
+
+      if (vk->context.swapchain_acquire_semaphore)
+      {
+#ifdef HAVE_THREADS
+         slock_lock(vk->context.queue_lock);
+#endif
+         RARCH_LOG("[Vulkan]: Destroying stale acquire semaphore.\n");
+         vkDeviceWaitIdle(vk->context.device);
+         vkDestroySemaphore(vk->context.device, vk->context.swapchain_acquire_semaphore, NULL);
+#ifdef HAVE_THREADS
+         slock_unlock(vk->context.queue_lock);
+#endif
+      }
+      vk->context.swapchain_acquire_semaphore = semaphore;
    }
    else
+   {
       vk->context.has_acquired_swapchain = false;
+      if (semaphore)
+         vulkan_recycle_acquire_semaphore(&vk->context, semaphore);
+   }
 
 #ifdef WSI_HARDENING_TEST
    trigger_spurious_error_vkresult(&err);
@@ -2809,9 +2896,7 @@ retry:
 
    if (err == VK_NOT_READY || err == VK_TIMEOUT)
    {
-      /* Just pretend we have a swapchain index, round-robin style. */
-      vk->context.current_swapchain_index =
-         (vk->context.current_swapchain_index + 1) % vk->context.num_swapchain_images;
+      /* Do nothing. */
    }
    else if (err == VK_ERROR_OUT_OF_DATE_KHR || err == VK_SUBOPTIMAL_KHR)
    {
diff --git a/gfx/common/vulkan_common.h b/gfx/common/vulkan_common.h
index d44babf57c..475d6afb7a 100644
--- a/gfx/common/vulkan_common.h
+++ b/gfx/common/vulkan_common.h
@@ -118,6 +118,11 @@ typedef struct vulkan_context
    VkSemaphore swapchain_semaphores[VULKAN_MAX_SWAPCHAIN_IMAGES];
    VkFormat swapchain_format;
 
+   VkSemaphore swapchain_acquire_semaphore;
+   unsigned num_recycled_acquire_semaphores;
+   VkSemaphore swapchain_recycled_semaphores[VULKAN_MAX_SWAPCHAIN_IMAGES];
+   VkSemaphore swapchain_wait_semaphores[VULKAN_MAX_SWAPCHAIN_IMAGES];
+
    slock_t *queue_lock;
    retro_vulkan_destroy_device_t destroy_device;
 
@@ -154,6 +159,10 @@ typedef struct gfx_ctx_vulkan_data
    bool created_new_swapchain;
    bool emulate_mailbox;
    bool emulating_mailbox;
+   /* If set, prefer a path where we use
+    * semaphores instead of fences for vkAcquireNextImageKHR.
+    * Helps workaround certain performance issues on some drivers. */
+   bool use_wsi_semaphore;
    vulkan_context_t context;
    VkSurfaceKHR vk_surface;
    VkSwapchainKHR swapchain;
@@ -421,7 +430,7 @@ typedef struct vk
 
       struct retro_hw_render_interface_vulkan iface;
       const struct retro_vulkan_image *image;
-      const VkSemaphore *semaphores;
+      VkSemaphore *semaphores;
       VkSemaphore signal_semaphore;
       VkPipelineStageFlags *wait_dst_stages;
       VkCommandBuffer *cmd;
diff --git a/gfx/drivers/vulkan.c b/gfx/drivers/vulkan.c
index 12b90e1e05..db07f7e311 100644
--- a/gfx/drivers/vulkan.c
+++ b/gfx/drivers/vulkan.c
@@ -909,6 +909,7 @@ static void vulkan_deinit_static_resources(vk_t *vk)
          vk->staging_pool, NULL);
    free(vk->hw.cmd);
    free(vk->hw.wait_dst_stages);
+   free(vk->hw.semaphores);
 
    for (i = 0; i < VULKAN_MAX_SWAPCHAIN_IMAGES; i++)
       if (vk->readback.staging[i].memory != VK_NULL_HANDLE)
@@ -1002,21 +1003,27 @@ static void vulkan_set_image(void *handle,
 
    vk->hw.image          = image;
    vk->hw.num_semaphores = num_semaphores;
-   vk->hw.semaphores     = semaphores;
 
    if (num_semaphores > 0)
    {
-      VkPipelineStageFlags *stage_flags = (VkPipelineStageFlags*)
-         realloc(vk->hw.wait_dst_stages,
-            sizeof(VkPipelineStageFlags) * vk->hw.num_semaphores);
+      /* Allocate one extra in case we need to use WSI acquire semaphores. */
+      VkPipelineStageFlags *stage_flags = (VkPipelineStageFlags*)realloc(vk->hw.wait_dst_stages,
+            sizeof(VkPipelineStageFlags) * (vk->hw.num_semaphores + 1));
+
+      VkSemaphore *new_semaphores = (VkSemaphore*)realloc(vk->hw.semaphores,
+            sizeof(VkSemaphore) * (vk->hw.num_semaphores + 1));
 
       /* If this fails, we're screwed anyways. */
-      retro_assert(stage_flags);
+      retro_assert(stage_flags && new_semaphores);
 
       vk->hw.wait_dst_stages = stage_flags;
+      vk->hw.semaphores = new_semaphores;
 
       for (i = 0; i < vk->hw.num_semaphores; i++)
+      {
          vk->hw.wait_dst_stages[i] = VK_PIPELINE_STAGE_FRAGMENT_SHADER_BIT;
+         vk->hw.semaphores[i] = semaphores[i];
+      }
 
       vk->hw.valid_semaphore = true;
       vk->hw.src_queue_family = src_queue_family;
@@ -1629,7 +1636,7 @@ static void vulkan_inject_black_frame(vk_t *vk, video_frame_info_t *video_info,
    vulkan_image_layout_transition(vk, vk->cmd, backbuffer->image,
          VK_IMAGE_LAYOUT_UNDEFINED, VK_IMAGE_LAYOUT_TRANSFER_DST_OPTIMAL,
          0, VK_ACCESS_TRANSFER_WRITE_BIT,
-         VK_PIPELINE_STAGE_TOP_OF_PIPE_BIT,
+         VK_PIPELINE_STAGE_COLOR_ATTACHMENT_OUTPUT_BIT,
          VK_PIPELINE_STAGE_TRANSFER_BIT);
 
    vkCmdClearColorImage(vk->cmd, backbuffer->image, VK_IMAGE_LAYOUT_TRANSFER_DST_OPTIMAL,
@@ -1645,12 +1652,28 @@ static void vulkan_inject_black_frame(vk_t *vk, video_frame_info_t *video_info,
 
    submit_info.commandBufferCount   = 1;
    submit_info.pCommandBuffers      = &vk->cmd;
-   if (vk->context->swapchain_semaphores[swapchain_index] != VK_NULL_HANDLE)
+   if (vk->context->has_acquired_swapchain &&
+         vk->context->swapchain_semaphores[swapchain_index] != VK_NULL_HANDLE)
    {
       submit_info.signalSemaphoreCount = 1;
       submit_info.pSignalSemaphores = &vk->context->swapchain_semaphores[swapchain_index];
    }
 
+   if (vk->context->has_acquired_swapchain &&
+         vk->context->swapchain_acquire_semaphore != VK_NULL_HANDLE)
+   {
+      static const VkPipelineStageFlags wait_stage =
+         VK_PIPELINE_STAGE_COLOR_ATTACHMENT_OUTPUT_BIT;
+
+      assert(!vk->context->swapchain_wait_semaphores[frame_index]);
+      vk->context->swapchain_wait_semaphores[frame_index] =
+         vk->context->swapchain_acquire_semaphore;
+      vk->context->swapchain_acquire_semaphore = VK_NULL_HANDLE;
+      submit_info.waitSemaphoreCount = 1;
+      submit_info.pWaitSemaphores = &vk->context->swapchain_wait_semaphores[frame_index];
+      submit_info.pWaitDstStageMask = &wait_stage;
+   }
+
 #ifdef HAVE_THREADS
    slock_lock(vk->context->queue_lock);
 #endif
@@ -1922,11 +1945,11 @@ static bool vulkan_frame(void *data, const void *frame,
       clear_color.color.float32[2] = 0.0f;
       clear_color.color.float32[3] = 0.0f;
 
-      /* Prepare backbuffer for rendering. We don't use WSI semaphores here. */
+      /* Prepare backbuffer for rendering. */
       vulkan_image_layout_transition(vk, vk->cmd, backbuffer->image,
             VK_IMAGE_LAYOUT_UNDEFINED, VK_IMAGE_LAYOUT_COLOR_ATTACHMENT_OPTIMAL,
             0, VK_ACCESS_COLOR_ATTACHMENT_WRITE_BIT | VK_ACCESS_COLOR_ATTACHMENT_READ_BIT,
-            VK_PIPELINE_STAGE_TOP_OF_PIPE_BIT,
+            VK_PIPELINE_STAGE_COLOR_ATTACHMENT_OUTPUT_BIT,
             VK_PIPELINE_STAGE_COLOR_ATTACHMENT_OUTPUT_BIT);
 
       /* Begin render pass and set up viewport */
@@ -2023,7 +2046,7 @@ static bool vulkan_frame(void *data, const void *frame,
             VK_IMAGE_LAYOUT_TRANSFER_SRC_OPTIMAL,
             VK_ACCESS_COLOR_ATTACHMENT_WRITE_BIT,
             VK_ACCESS_TRANSFER_READ_BIT,
-            VK_PIPELINE_STAGE_ALL_GRAPHICS_BIT,
+            VK_PIPELINE_STAGE_COLOR_ATTACHMENT_OUTPUT_BIT,
             VK_PIPELINE_STAGE_TRANSFER_BIT);
 
       vulkan_readback(vk);
@@ -2049,8 +2072,8 @@ static bool vulkan_frame(void *data, const void *frame,
             VK_IMAGE_LAYOUT_COLOR_ATTACHMENT_OPTIMAL,
             VK_IMAGE_LAYOUT_PRESENT_SRC_KHR,
             VK_ACCESS_COLOR_ATTACHMENT_WRITE_BIT,
-            VK_ACCESS_MEMORY_READ_BIT,
-            VK_PIPELINE_STAGE_ALL_GRAPHICS_BIT,
+            0,
+            VK_PIPELINE_STAGE_COLOR_ATTACHMENT_OUTPUT_BIT,
             VK_PIPELINE_STAGE_BOTTOM_OF_PIPE_BIT);
    }
 
@@ -2097,6 +2120,35 @@ static bool vulkan_frame(void *data, const void *frame,
 
       /* Consume the semaphores. */
       vk->hw.valid_semaphore = false;
+
+      /* We allocated space for this. */
+      if (vk->context->has_acquired_swapchain &&
+            vk->context->swapchain_acquire_semaphore != VK_NULL_HANDLE)
+      {
+         assert(!vk->context->swapchain_wait_semaphores[frame_index]);
+         vk->context->swapchain_wait_semaphores[frame_index] =
+            vk->context->swapchain_acquire_semaphore;
+         vk->context->swapchain_acquire_semaphore = VK_NULL_HANDLE;
+
+         vk->hw.semaphores[submit_info.waitSemaphoreCount] = vk->context->swapchain_wait_semaphores[frame_index];
+         vk->hw.wait_dst_stages[submit_info.waitSemaphoreCount] = VK_PIPELINE_STAGE_COLOR_ATTACHMENT_OUTPUT_BIT;
+         submit_info.waitSemaphoreCount++;
+      }
+   }
+   else if (vk->context->has_acquired_swapchain &&
+         vk->context->swapchain_acquire_semaphore != VK_NULL_HANDLE)
+   {
+      static const VkPipelineStageFlags wait_stage =
+         VK_PIPELINE_STAGE_COLOR_ATTACHMENT_OUTPUT_BIT;
+
+      assert(!vk->context->swapchain_wait_semaphores[frame_index]);
+      vk->context->swapchain_wait_semaphores[frame_index] =
+         vk->context->swapchain_acquire_semaphore;
+      vk->context->swapchain_acquire_semaphore = VK_NULL_HANDLE;
+
+      submit_info.waitSemaphoreCount = 1;
+      submit_info.pWaitSemaphores = &vk->context->swapchain_wait_semaphores[frame_index];
+      submit_info.pWaitDstStageMask = &wait_stage;
    }
 
    submit_info.signalSemaphoreCount = 0;