(vulkan) vulkan_common optimizations

Author: LibretroAdmin

gfx/common/vulkan_common.c

@@ -1344,7 +1344,9 @@ static void vulkan_acquire_clear_fences(gfx_ctx_vulkan_data_t *vk)
       {
          struct vulkan_context *ctx = &vk->context;
          VkSemaphore sem            = vk->context.swapchain_wait_semaphores[i];
-         assert(ctx->num_recycled_acquire_semaphores < VULKAN_MAX_SWAPCHAIN_IMAGES);
+#ifdef DEBUG
+         retro_assert(ctx->num_recycled_acquire_semaphores < VULKAN_MAX_SWAPCHAIN_IMAGES);
+#endif
          ctx->swapchain_recycled_semaphores[ctx->num_recycled_acquire_semaphores++] = sem;
       }
       vk->context.swapchain_wait_semaphores[i] = VK_NULL_HANDLE;
@@ -1406,7 +1408,9 @@ static void vulkan_acquire_wait_fences(gfx_ctx_vulkan_data_t *vk)
    {
       struct vulkan_context *ctx = &vk->context;
       VkSemaphore sem            = vk->context.swapchain_wait_semaphores[index];
-      assert(ctx->num_recycled_acquire_semaphores < VULKAN_MAX_SWAPCHAIN_IMAGES);
+#ifdef DEBUG
+      retro_assert(ctx->num_recycled_acquire_semaphores < VULKAN_MAX_SWAPCHAIN_IMAGES);
+#endif
       ctx->swapchain_recycled_semaphores[ctx->num_recycled_acquire_semaphores++] = sem;
    }
    vk->context.swapchain_wait_semaphores[index] = VK_NULL_HANDLE;
@@ -1469,8 +1473,10 @@ bool vulkan_buffer_chain_alloc(const struct vulkan_context *context,
 
       chain->current = chain->current->next;
       chain->offset  = 0;
+#ifdef DEBUG
       /* This cannot possibly fail. */
       retro_assert(vulkan_buffer_chain_suballoc(chain, len, range));
+#endif
    }
    return true;
 }
@@ -1597,15 +1603,16 @@ VkDescriptorSet vulkan_descriptor_manager_alloc(
       }
 
       manager->current->next = vulkan_alloc_descriptor_pool(device, manager);
+#ifdef DEBUG
       retro_assert(manager->current->next);
+#endif
 
       manager->current = manager->current->next;
       manager->count   = 0;
    }
    return manager->current->sets[manager->count++];
 }
 
-
 bool vulkan_surface_create(gfx_ctx_vulkan_data_t *vk,
       enum vulkan_wsi_type type,
       void *display, void *surface,
@@ -1872,7 +1879,9 @@ void vulkan_acquire_next_image(gfx_ctx_vulkan_data_t *vk)
       }
    }
 
+#ifdef DEBUG
    retro_assert(!(vk->context.flags & VK_CTX_FLAG_HAS_ACQUIRED_SWAPCHAIN));
+#endif
 
    if (vk->flags & VK_DATA_FLAG_EMULATING_MAILBOX)
    {
@@ -1903,16 +1912,16 @@ void vulkan_acquire_next_image(gfx_ctx_vulkan_data_t *vk)
          vkWaitForFences(vk->context.device, 1, &fence, true, UINT64_MAX);
       vk->context.flags |= VK_CTX_FLAG_HAS_ACQUIRED_SWAPCHAIN;
 
-      if (vk->context.swapchain_acquire_semaphore)
+      /* Recycle the previous acquire semaphore rather than stalling the device.
+       * vkDeviceWaitIdle here was a significant hot-path bottleneck. */
+      if (vk->context.swapchain_acquire_semaphore != VK_NULL_HANDLE)
       {
-#ifdef HAVE_THREADS
-         slock_lock(vk->context.queue_lock);
-#endif
-         vkDeviceWaitIdle(vk->context.device);
-         vkDestroySemaphore(vk->context.device, vk->context.swapchain_acquire_semaphore, NULL);
-#ifdef HAVE_THREADS
-         slock_unlock(vk->context.queue_lock);
+         struct vulkan_context *ctx = &vk->context;
+#ifdef DEBUG
+         retro_assert(ctx->num_recycled_acquire_semaphores < VULKAN_MAX_SWAPCHAIN_IMAGES);
 #endif
+         ctx->swapchain_recycled_semaphores[ctx->num_recycled_acquire_semaphores++] =
+            vk->context.swapchain_acquire_semaphore;
       }
       vk->context.swapchain_acquire_semaphore = semaphore;
    }
@@ -1923,7 +1932,9 @@ void vulkan_acquire_next_image(gfx_ctx_vulkan_data_t *vk)
       {
          struct vulkan_context *ctx = &vk->context;
          VkSemaphore sem            = semaphore;
-         assert(ctx->num_recycled_acquire_semaphores < VULKAN_MAX_SWAPCHAIN_IMAGES);
+#ifdef DEBUG
+         retro_assert(ctx->num_recycled_acquire_semaphores < VULKAN_MAX_SWAPCHAIN_IMAGES);
+#endif
          ctx->swapchain_recycled_semaphores[ctx->num_recycled_acquire_semaphores++] = sem;
       }
    }

gfx/common/vulkan_common.h

@@ -146,60 +146,40 @@
    vkUpdateDescriptorSets(_device, 1, &write, 0, NULL); \
 }
 
+/*
+ * Vertex layout: two triangles forming a quad (CCW winding).
+ *   v0 (TL) -- v2/v4 (TR)
+ *   |        /        |
+ *   v1/v5 (BL) -- v3 (BR)
+ */
 #define VULKAN_WRITE_QUAD_VBO(pv, _x, _y, _width, _height, _tex_x, _tex_y, _tex_width, _tex_height, vulkan_color) \
 { \
    float r        = (vulkan_color)->r; \
    float g        = (vulkan_color)->g; \
    float b        = (vulkan_color)->b; \
    float a        = (vulkan_color)->a; \
-   pv[0].x        = (_x)     + 0.0f * (_width); \
-   pv[0].y        = (_y)     + 0.0f * (_height); \
-   pv[0].tex_x    = (_tex_x) + 0.0f * (_tex_width); \
-   pv[0].tex_y    = (_tex_y) + 0.0f * (_tex_height); \
-   pv[0].color.r  = r; \
-   pv[0].color.g  = g; \
-   pv[0].color.b  = b; \
-   pv[0].color.a  = a; \
-   pv[1].x        = (_x)     + 0.0f * (_width); \
-   pv[1].y        = (_y)     + 1.0f * (_height); \
-   pv[1].tex_x    = (_tex_x) + 0.0f * (_tex_width); \
-   pv[1].tex_y    = (_tex_y) + 1.0f * (_tex_height); \
-   pv[1].color.r  = r; \
-   pv[1].color.g  = g; \
-   pv[1].color.b  = b; \
-   pv[1].color.a  = a; \
-   pv[2].x        = (_x)     + 1.0f * (_width); \
-   pv[2].y        = (_y)     + 0.0f * (_height); \
-   pv[2].tex_x    = (_tex_x) + 1.0f * (_tex_width); \
-   pv[2].tex_y    = (_tex_y) + 0.0f * (_tex_height); \
-   pv[2].color.r  = r; \
-   pv[2].color.g  = g; \
-   pv[2].color.b  = b; \
-   pv[2].color.a  = a; \
-   pv[3].x        = (_x)     + 1.0f * (_width); \
-   pv[3].y        = (_y)     + 1.0f * (_height); \
-   pv[3].tex_x    = (_tex_x) + 1.0f * (_tex_width); \
-   pv[3].tex_y    = (_tex_y) + 1.0f * (_tex_height); \
-   pv[3].color.r  = r; \
-   pv[3].color.g  = g; \
-   pv[3].color.b  = b; \
-   pv[3].color.a  = a; \
-   pv[4].x        = (_x)     + 1.0f * (_width); \
-   pv[4].y        = (_y)     + 0.0f * (_height); \
-   pv[4].tex_x    = (_tex_x) + 1.0f * (_tex_width); \
-   pv[4].tex_y    = (_tex_y) + 0.0f * (_tex_height); \
-   pv[4].color.r  = r; \
-   pv[4].color.g  = g; \
-   pv[4].color.b  = b; \
-   pv[4].color.a  = a; \
-   pv[5].x        = (_x)     + 0.0f * (_width); \
-   pv[5].y        = (_y)     + 1.0f * (_height); \
-   pv[5].tex_x    = (_tex_x) + 0.0f * (_tex_width); \
-   pv[5].tex_y    = (_tex_y) + 1.0f * (_tex_height); \
-   pv[5].color.r  = r; \
-   pv[5].color.g  = g; \
-   pv[5].color.b  = b; \
-   pv[5].color.a  = a; \
+   float _x1      = (float)(_x); \
+   float _y1      = (float)(_y); \
+   float _x2      = _x1 + (float)(_width); \
+   float _y2      = _y1 + (float)(_height); \
+   float _tx1     = (float)(_tex_x); \
+   float _ty1     = (float)(_tex_y); \
+   float _tx2     = _tx1 + (float)(_tex_width); \
+   float _ty2     = _ty1 + (float)(_tex_height); \
+   /* Triangle 1: TL, BL, TR */ \
+   pv[0].x = _x1; pv[0].y = _y1; pv[0].tex_x = _tx1; pv[0].tex_y = _ty1; \
+   pv[0].color.r = r; pv[0].color.g = g; pv[0].color.b = b; pv[0].color.a = a; \
+   pv[1].x = _x1; pv[1].y = _y2; pv[1].tex_x = _tx1; pv[1].tex_y = _ty2; \
+   pv[1].color.r = r; pv[1].color.g = g; pv[1].color.b = b; pv[1].color.a = a; \
+   pv[2].x = _x2; pv[2].y = _y1; pv[2].tex_x = _tx2; pv[2].tex_y = _ty1; \
+   pv[2].color.r = r; pv[2].color.g = g; pv[2].color.b = b; pv[2].color.a = a; \
+   /* Triangle 2: TR, BL, BR */ \
+   pv[3].x = _x2; pv[3].y = _y2; pv[3].tex_x = _tx2; pv[3].tex_y = _ty2; \
+   pv[3].color.r = r; pv[3].color.g = g; pv[3].color.b = b; pv[3].color.a = a; \
+   pv[4].x = _x2; pv[4].y = _y1; pv[4].tex_x = _tx2; pv[4].tex_y = _ty1; \
+   pv[4].color.r = r; pv[4].color.g = g; pv[4].color.b = b; pv[4].color.a = a; \
+   pv[5].x = _x1; pv[5].y = _y2; pv[5].tex_x = _tx1; pv[5].tex_y = _ty2; \
+   pv[5].color.r = r; pv[5].color.g = g; pv[5].color.b = b; pv[5].color.a = a; \
 }
 
 /* We don't have to sync against previous TRANSFER,