wined3d: Set an array of UAVs as a single CS operation.

dlls/d3d11/device.c

@@ -1087,20 +1087,17 @@ static void STDMETHODCALLTYPE d3d11_device_context_OMSetRenderTargets(ID3D11Devi
 }
 
 static void STDMETHODCALLTYPE d3d11_device_context_OMSetRenderTargetsAndUnorderedAccessViews(
-        ID3D11DeviceContext1 *iface, UINT render_target_view_count,
-        ID3D11RenderTargetView *const *render_target_views, ID3D11DepthStencilView *depth_stencil_view,
-        UINT unordered_access_view_start_slot, UINT unordered_access_view_count,
-        ID3D11UnorderedAccessView *const *unordered_access_views, const UINT *initial_counts)
+        ID3D11DeviceContext1 *iface, UINT render_target_view_count, ID3D11RenderTargetView *const *render_target_views,
+        ID3D11DepthStencilView *depth_stencil_view, UINT uav_start_idx, UINT uav_count,
+        ID3D11UnorderedAccessView *const *uavs, const UINT *initial_counts)
 {
     struct d3d11_device_context *context = impl_from_ID3D11DeviceContext1(iface);
     unsigned int i;
 
     TRACE("iface %p, render_target_view_count %u, render_target_views %p, depth_stencil_view %p, "
-            "unordered_access_view_start_slot %u, unordered_access_view_count %u, unordered_access_views %p, "
-            "initial_counts %p.\n",
+            "uav_start_idx %u, uav_count %u, uavs %p, initial_counts %p.\n",
             iface, render_target_view_count, render_target_views, depth_stencil_view,
-            unordered_access_view_start_slot, unordered_access_view_count, unordered_access_views,
-            initial_counts);
+            uav_start_idx, uav_count, uavs, initial_counts);
 
     if (render_target_view_count != D3D11_KEEP_RENDER_TARGETS_AND_DEPTH_STENCIL)
     {
@@ -1108,28 +1105,31 @@ static void STDMETHODCALLTYPE d3d11_device_context_OMSetRenderTargetsAndUnordere
                 depth_stencil_view);
     }
 
-    if (unordered_access_view_count != D3D11_KEEP_UNORDERED_ACCESS_VIEWS)
+    if (uav_count != D3D11_KEEP_UNORDERED_ACCESS_VIEWS)
     {
-        wined3d_mutex_lock();
-        for (i = 0; i < unordered_access_view_start_slot; ++i)
+        struct wined3d_unordered_access_view *wined3d_views[D3D11_PS_CS_UAV_REGISTER_COUNT] = {0};
+        unsigned int wined3d_initial_counts[D3D11_PS_CS_UAV_REGISTER_COUNT];
+
+        if (!wined3d_bound_range(uav_start_idx, uav_count, ARRAY_SIZE(wined3d_views)))
         {
-            wined3d_device_context_set_unordered_access_view(context->wined3d_context,
-                    WINED3D_PIPELINE_GRAPHICS, i, NULL, ~0u);
+            WARN("View count %u exceeds limit; ignoring call.\n", uav_count);
+            return;
         }
-        for (i = 0; i < unordered_access_view_count; ++i)
-        {
-            struct d3d11_unordered_access_view *view
-                    = unsafe_impl_from_ID3D11UnorderedAccessView(unordered_access_views[i]);
 
-            wined3d_device_context_set_unordered_access_view(context->wined3d_context,
-                    WINED3D_PIPELINE_GRAPHICS, unordered_access_view_start_slot + i,
-                    view ? view->wined3d_view : NULL, initial_counts ? initial_counts[i] : ~0u);
-        }
-        for (; unordered_access_view_start_slot + i < D3D11_PS_CS_UAV_REGISTER_COUNT; ++i)
+        memset(wined3d_initial_counts, 0xff, sizeof(wined3d_initial_counts));
+
+        for (i = 0; i < uav_count; ++i)
         {
-            wined3d_device_context_set_unordered_access_view(context->wined3d_context,
-                    WINED3D_PIPELINE_GRAPHICS, unordered_access_view_start_slot + i, NULL, ~0u);
+            struct d3d11_unordered_access_view *view =
+                    unsafe_impl_from_ID3D11UnorderedAccessView(uavs[i]);
+
+            wined3d_views[uav_start_idx + i] = view ? view->wined3d_view : NULL;
+            wined3d_initial_counts[uav_start_idx + i] = initial_counts ? initial_counts[i] : ~0u;
         }
+
+        wined3d_mutex_lock();
+        wined3d_device_context_set_unordered_access_views(context->wined3d_context, WINED3D_PIPELINE_GRAPHICS,
+                0, ARRAY_SIZE(wined3d_views), wined3d_views, wined3d_initial_counts);
         wined3d_mutex_unlock();
     }
 }
@@ -1651,19 +1651,28 @@ static void STDMETHODCALLTYPE d3d11_device_context_CSSetUnorderedAccessViews(ID3
         UINT start_slot, UINT view_count, ID3D11UnorderedAccessView *const *views, const UINT *initial_counts)
 {
     struct d3d11_device_context *context = impl_from_ID3D11DeviceContext1(iface);
+    struct wined3d_unordered_access_view *wined3d_views[D3D11_PS_CS_UAV_REGISTER_COUNT];
     unsigned int i;
 
     TRACE("iface %p, start_slot %u, view_count %u, views %p, initial_counts %p.\n",
             iface, start_slot, view_count, views, initial_counts);
 
-    wined3d_mutex_lock();
+    if (view_count > ARRAY_SIZE(wined3d_views))
+    {
+        WARN("View count %u exceeds limit; ignoring call.\n", view_count);
+        return;
+    }
+
     for (i = 0; i < view_count; ++i)
     {
         struct d3d11_unordered_access_view *view = unsafe_impl_from_ID3D11UnorderedAccessView(views[i]);
 
-        wined3d_device_context_set_unordered_access_view(context->wined3d_context, WINED3D_PIPELINE_COMPUTE,
-                start_slot + i, view ? view->wined3d_view : NULL, initial_counts ? initial_counts[i] : ~0u);
+        wined3d_views[i] = view ? view->wined3d_view : NULL;
     }
+
+    wined3d_mutex_lock();
+    wined3d_device_context_set_unordered_access_views(context->wined3d_context, WINED3D_PIPELINE_COMPUTE,
+            start_slot, view_count, wined3d_views, initial_counts);
     wined3d_mutex_unlock();
 }
 

dlls/wined3d/cs.c

@@ -123,7 +123,7 @@ enum wined3d_cs_op
     WINED3D_CS_OP_SET_CONSTANT_BUFFERS,
     WINED3D_CS_OP_SET_TEXTURE,
     WINED3D_CS_OP_SET_SHADER_RESOURCE_VIEWS,
-    WINED3D_CS_OP_SET_UNORDERED_ACCESS_VIEW,
+    WINED3D_CS_OP_SET_UNORDERED_ACCESS_VIEWS,
     WINED3D_CS_OP_SET_SAMPLERS,
     WINED3D_CS_OP_SET_SHADER,
     WINED3D_CS_OP_SET_BLEND_STATE,
@@ -319,13 +319,17 @@ struct wined3d_cs_set_shader_resource_views
     struct wined3d_shader_resource_view *views[1];
 };
 
-struct wined3d_cs_set_unordered_access_view
+struct wined3d_cs_set_unordered_access_views
 {
     enum wined3d_cs_op opcode;
     enum wined3d_pipeline pipeline;
-    unsigned int view_idx;
+    unsigned int start_idx;
+    unsigned int count;
+    struct
+    {
         struct wined3d_unordered_access_view *view;
         unsigned int initial_count;
+    } uavs[1];
 };
 
 struct wined3d_cs_set_samplers
@@ -604,7 +608,7 @@ static const char *debug_cs_op(enum wined3d_cs_op op)
         WINED3D_TO_STR(WINED3D_CS_OP_SET_CONSTANT_BUFFERS);
         WINED3D_TO_STR(WINED3D_CS_OP_SET_TEXTURE);
         WINED3D_TO_STR(WINED3D_CS_OP_SET_SHADER_RESOURCE_VIEWS);
-        WINED3D_TO_STR(WINED3D_CS_OP_SET_UNORDERED_ACCESS_VIEW);
+        WINED3D_TO_STR(WINED3D_CS_OP_SET_UNORDERED_ACCESS_VIEWS);
         WINED3D_TO_STR(WINED3D_CS_OP_SET_SAMPLERS);
         WINED3D_TO_STR(WINED3D_CS_OP_SET_SHADER);
         WINED3D_TO_STR(WINED3D_CS_OP_SET_BLEND_STATE);
@@ -1702,37 +1706,49 @@ void wined3d_device_context_emit_set_shader_resource_views(struct wined3d_device
     wined3d_device_context_submit(context, WINED3D_CS_QUEUE_DEFAULT);
 }
 
-static void wined3d_cs_exec_set_unordered_access_view(struct wined3d_cs *cs, const void *data)
+static void wined3d_cs_exec_set_unordered_access_views(struct wined3d_cs *cs, const void *data)
 {
-    const struct wined3d_cs_set_unordered_access_view *op = data;
-    struct wined3d_unordered_access_view *prev;
+    const struct wined3d_cs_set_unordered_access_views *op = data;
+    unsigned int i;
 
-    prev = cs->state.unordered_access_view[op->pipeline][op->view_idx];
-    cs->state.unordered_access_view[op->pipeline][op->view_idx] = op->view;
+    for (i = 0; i < op->count; ++i)
+    {
+        struct wined3d_unordered_access_view *prev = cs->state.unordered_access_view[op->pipeline][op->start_idx + i];
+        struct wined3d_unordered_access_view *view = op->uavs[i].view;
+        unsigned int initial_count = op->uavs[i].initial_count;
+
+        cs->state.unordered_access_view[op->pipeline][op->start_idx + i] = view;
 
-    if (op->view)
-        InterlockedIncrement(&op->view->resource->bind_count);
+        if (view)
+            InterlockedIncrement(&view->resource->bind_count);
         if (prev)
             InterlockedDecrement(&prev->resource->bind_count);
 
-    if (op->view && op->initial_count != ~0u)
-        wined3d_unordered_access_view_set_counter(op->view, op->initial_count);
+        if (view && initial_count != ~0u)
+            wined3d_unordered_access_view_set_counter(view, initial_count);
+    }
 
     device_invalidate_state(cs->c.device, STATE_UNORDERED_ACCESS_VIEW_BINDING(op->pipeline));
 }
 
-void wined3d_device_context_emit_set_unordered_access_view(struct wined3d_device_context *context,
-        enum wined3d_pipeline pipeline, unsigned int view_idx, struct wined3d_unordered_access_view *view,
-        unsigned int initial_count)
+void wined3d_device_context_emit_set_unordered_access_views(struct wined3d_device_context *context,
+        enum wined3d_pipeline pipeline, unsigned int start_idx, unsigned int count,
+        struct wined3d_unordered_access_view *const *views, const unsigned int *initial_counts)
 {
-    struct wined3d_cs_set_unordered_access_view *op;
+    struct wined3d_cs_set_unordered_access_views *op;
+    unsigned int i;
 
-    op = wined3d_device_context_require_space(context, sizeof(*op), WINED3D_CS_QUEUE_DEFAULT);
-    op->opcode = WINED3D_CS_OP_SET_UNORDERED_ACCESS_VIEW;
+    op = wined3d_device_context_require_space(context,
+            offsetof(struct wined3d_cs_set_unordered_access_views, uavs[count]), WINED3D_CS_QUEUE_DEFAULT);
+    op->opcode = WINED3D_CS_OP_SET_UNORDERED_ACCESS_VIEWS;
     op->pipeline = pipeline;
-    op->view_idx = view_idx;
-    op->view = view;
-    op->initial_count = initial_count;
+    op->start_idx = start_idx;
+    op->count = count;
+    for (i = 0; i < count; ++i)
+    {
+        op->uavs[i].view = views[i];
+        op->uavs[i].initial_count = initial_counts ? initial_counts[i] : ~0u;
+    }
 
     wined3d_device_context_submit(context, WINED3D_CS_QUEUE_DEFAULT);
 }
@@ -2917,7 +2933,7 @@ static void (* const wined3d_cs_op_handlers[])(struct wined3d_cs *cs, const void
     /* WINED3D_CS_OP_SET_CONSTANT_BUFFERS        */ wined3d_cs_exec_set_constant_buffers,
     /* WINED3D_CS_OP_SET_TEXTURE                 */ wined3d_cs_exec_set_texture,
     /* WINED3D_CS_OP_SET_SHADER_RESOURCE_VIEWS   */ wined3d_cs_exec_set_shader_resource_views,
-    /* WINED3D_CS_OP_SET_UNORDERED_ACCESS_VIEW   */ wined3d_cs_exec_set_unordered_access_view,
+    /* WINED3D_CS_OP_SET_UNORDERED_ACCESS_VIEWS  */ wined3d_cs_exec_set_unordered_access_views,
     /* WINED3D_CS_OP_SET_SAMPLERS                */ wined3d_cs_exec_set_samplers,
     /* WINED3D_CS_OP_SET_SHADER                  */ wined3d_cs_exec_set_shader,
     /* WINED3D_CS_OP_SET_BLEND_STATE             */ wined3d_cs_exec_set_blend_state,

dlls/wined3d/device.c

@@ -1703,13 +1703,8 @@ void CDECL wined3d_device_context_set_state(struct wined3d_device_context *conte
     }
 
     for (i = 0; i < WINED3D_PIPELINE_COUNT; ++i)
-    {
-        for (j = 0; j < MAX_UNORDERED_ACCESS_VIEWS; ++j)
-        {
-            wined3d_device_context_emit_set_unordered_access_view(context, i, j,
-                    state->unordered_access_view[i][j], ~0);
-        }
-    }
+        wined3d_device_context_emit_set_unordered_access_views(context, i, 0, MAX_UNORDERED_ACCESS_VIEWS,
+                state->unordered_access_view[i], NULL);
 
     wined3d_device_context_push_constants(context, WINED3D_PUSH_CONSTANTS_VS_F,
             0, WINED3D_MAX_VS_CONSTS_F, state->vs_consts_f);
@@ -2063,31 +2058,37 @@ void CDECL wined3d_device_context_set_samplers(struct wined3d_device_context *co
     }
 }
 
-void CDECL wined3d_device_context_set_unordered_access_view(struct wined3d_device_context *context,
-        enum wined3d_pipeline pipeline, unsigned int idx, struct wined3d_unordered_access_view *uav,
-        unsigned int initial_count)
+void CDECL wined3d_device_context_set_unordered_access_views(struct wined3d_device_context *context,
+        enum wined3d_pipeline pipeline, unsigned int start_idx, unsigned int count,
+        struct wined3d_unordered_access_view *const *uavs, const unsigned int *initial_counts)
 {
     struct wined3d_state *state = context->state;
-    struct wined3d_unordered_access_view *prev;
+    unsigned int i;
 
-    TRACE("context %p, pipeline %#x, idx %u, uav %p, initial_count %u.\n", context, pipeline, idx, uav, initial_count);
+    TRACE("context %p, pipeline %#x, start_idx %u, count %u, uavs %p, initial_counts %p.\n",
+            context, pipeline, start_idx, count, uavs, initial_counts);
 
-    if (idx >= MAX_UNORDERED_ACCESS_VIEWS)
+    if (!wined3d_bound_range(start_idx, count, MAX_UNORDERED_ACCESS_VIEWS))
     {
-        WARN("Invalid UAV index %u.\n", idx);
+        WARN("Invalid UAV index %u, count %u.\n", start_idx, count);
         return;
     }
 
-    prev = state->unordered_access_view[pipeline][idx];
-    if (uav == prev && initial_count == ~0u)
+    if (!memcmp(uavs, &state->unordered_access_view[pipeline][start_idx], count * sizeof(*uavs)) && !initial_counts)
         return;
 
+    wined3d_device_context_emit_set_unordered_access_views(context, pipeline, start_idx, count, uavs, initial_counts);
+    for (i = 0; i < count; ++i)
+    {
+        struct wined3d_unordered_access_view *prev = state->unordered_access_view[pipeline][start_idx + i];
+        struct wined3d_unordered_access_view *uav = uavs[i];
+
         if (uav)
             wined3d_unordered_access_view_incref(uav);
-    state->unordered_access_view[pipeline][idx] = uav;
-    wined3d_device_context_emit_set_unordered_access_view(context, pipeline, idx, uav, initial_count);
+        state->unordered_access_view[pipeline][start_idx + i] = uav;
         if (prev)
             wined3d_unordered_access_view_decref(prev);
+    }
 }
 
 static void wined3d_device_context_unbind_srv_for_rtv(struct wined3d_device_context *context,

dlls/wined3d/wined3d.spec

@@ -138,7 +138,7 @@
 @ cdecl wined3d_device_context_set_state(ptr ptr)
 @ cdecl wined3d_device_context_set_stream_output(ptr long ptr long)
 @ cdecl wined3d_device_context_set_stream_source(ptr long ptr long long)
-@ cdecl wined3d_device_context_set_unordered_access_view(ptr long long ptr long)
+@ cdecl wined3d_device_context_set_unordered_access_views(ptr long long long ptr ptr)
 @ cdecl wined3d_device_context_set_vertex_declaration(ptr ptr)
 @ cdecl wined3d_device_context_set_viewports(ptr long ptr)
 @ cdecl wined3d_device_context_unmap(ptr ptr long)

dlls/wined3d/wined3d_private.h

@@ -4856,9 +4856,9 @@ void wined3d_device_context_emit_set_texture_state(struct wined3d_device_context
         enum wined3d_texture_stage_state state, unsigned int value) DECLSPEC_HIDDEN;
 void wined3d_device_context_emit_set_transform(struct wined3d_device_context *context,
         enum wined3d_transform_state state, const struct wined3d_matrix *matrix) DECLSPEC_HIDDEN;
-void wined3d_device_context_emit_set_unordered_access_view(struct wined3d_device_context *context,
-        enum wined3d_pipeline pipeline, unsigned int view_idx, struct wined3d_unordered_access_view *view,
-        unsigned int initial_count) DECLSPEC_HIDDEN;
+void wined3d_device_context_emit_set_unordered_access_views(struct wined3d_device_context *context,
+        enum wined3d_pipeline pipeline, unsigned int start_idx, unsigned int count,
+        struct wined3d_unordered_access_view *const *views, const unsigned int *initial_count) DECLSPEC_HIDDEN;
 void wined3d_device_context_emit_set_vertex_declaration(struct wined3d_device_context *context,
         struct wined3d_vertex_declaration *declaration) DECLSPEC_HIDDEN;
 void wined3d_device_context_emit_set_viewports(struct wined3d_device_context *context, unsigned int viewport_count,

include/wine/wined3d.h

@@ -2527,9 +2527,9 @@ void __cdecl wined3d_device_context_set_stream_output(struct wined3d_device_cont
         struct wined3d_buffer *buffer, unsigned int offset);
 HRESULT __cdecl wined3d_device_context_set_stream_source(struct wined3d_device_context *context,
         unsigned int stream_idx, struct wined3d_buffer *buffer, unsigned int offset, unsigned int stride);
-void __cdecl wined3d_device_context_set_unordered_access_view(struct wined3d_device_context *context,
-        enum wined3d_pipeline pipeline, unsigned int idx, struct wined3d_unordered_access_view *uav,
-        unsigned int initial_count);
+void __cdecl wined3d_device_context_set_unordered_access_views(struct wined3d_device_context *context,
+        enum wined3d_pipeline pipeline, unsigned int start_idx, unsigned int count,
+        struct wined3d_unordered_access_view *const *uavs, const unsigned int *initial_counts);
 void __cdecl wined3d_device_context_set_vertex_declaration(struct wined3d_device_context *context,
         struct wined3d_vertex_declaration *declaration);
 void __cdecl wined3d_device_context_set_viewports(struct wined3d_device_context *context, unsigned int viewport_count,