nvc0: add support for real ARB_multi_draw_indirect

The draw groups are now split up into groups of 32 if there's a
non-packed stride, or in groups of 400-500 if the draw data is packed.

Signed-off-by: Ilia Mirkin <imirkin@alum.mit.edu>

4 files changed, 47 insertions, 18 deletions

diff --git a/src/gallium/drivers/nouveau/nvc0/nvc0_query_hw.c b/src/gallium/drivers/nouveau/nvc0/nvc0_query_hw.c
index 8021a65dc46..1bed0162baf 100644
--- a/src/gallium/drivers/nouveau/nvc0/nvc0_query_hw.c
+++ b/src/gallium/drivers/nouveau/nvc0/nvc0_query_hw.c
@@ -470,8 +470,6 @@ nvc0_hw_query_pushbuf_submit(struct nouveau_pushbuf *push,
 {
    struct nvc0_hw_query *hq = nvc0_hw_query(q);
 
-#define NVC0_IB_ENTRY_1_NO_PREFETCH (1 << (31 - 8))
-
    PUSH_REFN(push, hq->bo, NOUVEAU_BO_RD | NOUVEAU_BO_GART);
    nouveau_pushbuf_data(push, hq->bo, hq->offset + result_offset, 4 |
                         NVC0_IB_ENTRY_1_NO_PREFETCH);
diff --git a/src/gallium/drivers/nouveau/nvc0/nvc0_screen.c b/src/gallium/drivers/nouveau/nvc0/nvc0_screen.c
index a184e8fdd87..86bd8632d0b 100644
--- a/src/gallium/drivers/nouveau/nvc0/nvc0_screen.c
+++ b/src/gallium/drivers/nouveau/nvc0/nvc0_screen.c
@@ -186,6 +186,7 @@ nvc0_screen_get_param(struct pipe_screen *pscreen, enum pipe_cap param)
    case PIPE_CAP_CLEAR_TEXTURE:
    case PIPE_CAP_DRAW_PARAMETERS:
    case PIPE_CAP_TGSI_PACK_HALF_FLOAT:
+   case PIPE_CAP_MULTI_DRAW_INDIRECT:
       return 1;
    case PIPE_CAP_SEAMLESS_CUBE_MAP_PER_TEXTURE:
       return (class_3d >= NVE4_3D_CLASS) ? 1 : 0;
@@ -208,7 +209,6 @@ nvc0_screen_get_param(struct pipe_screen *pscreen, enum pipe_cap param)
    case PIPE_CAP_VERTEXID_NOBASE:
    case PIPE_CAP_RESOURCE_FROM_USER_MEMORY:
    case PIPE_CAP_DEVICE_RESET_STATUS_QUERY:
-   case PIPE_CAP_MULTI_DRAW_INDIRECT:
    case PIPE_CAP_MULTI_DRAW_INDIRECT_PARAMS:
       return 0;
 
diff --git a/src/gallium/drivers/nouveau/nvc0/nvc0_vbo.c b/src/gallium/drivers/nouveau/nvc0/nvc0_vbo.c
index 55aeb806288..1d889b9db0d 100644
--- a/src/gallium/drivers/nouveau/nvc0/nvc0_vbo.c
+++ b/src/gallium/drivers/nouveau/nvc0/nvc0_vbo.c
@@ -807,8 +807,8 @@ nvc0_draw_indirect(struct nvc0_context *nvc0, const struct pipe_draw_info *info)
 {
    struct nouveau_pushbuf *push = nvc0->base.pushbuf;
    struct nv04_resource *buf = nv04_resource(info->indirect);
-   unsigned size;
-   const uint32_t offset = buf->offset + info->indirect_offset;
+   unsigned size, macro, count = info->indirect_count, drawid = info->drawid;
+   uint32_t offset = buf->offset + info->indirect_offset;
 
    /* must make FIFO wait for engines idle before continuing to process */
    if (buf->fence_wr && !nouveau_fence_signalled(buf->fence_wr))
@@ -820,13 +820,11 @@ nvc0_draw_indirect(struct nvc0_context *nvc0, const struct pipe_draw_info *info)
    PUSH_DATAh(push, nvc0->screen->uniform_bo->offset + (5 << 16) + (0 << 9));
    PUSH_DATA (push, nvc0->screen->uniform_bo->offset + (5 << 16) + (0 << 9));
 
-   nouveau_pushbuf_space(push, 8, 0, 1);
-   PUSH_REFN(push, buf->bo, NOUVEAU_BO_RD | buf->domain);
    if (info->indexed) {
       assert(nvc0->idxbuf.buffer);
       assert(nouveau_resource_mapped_by_gpu(nvc0->idxbuf.buffer));
-      size = 5 * 4;
-      BEGIN_1IC0(push, NVC0_3D(MACRO_DRAW_ELEMENTS_INDIRECT), 3 + size / 4);
+      size = 5;
+      macro = NVC0_3D_MACRO_DRAW_ELEMENTS_INDIRECT;
    } else {
       if (nvc0->state.index_bias) {
          /* index_bias is implied 0 if !info->indexed (really ?) */
@@ -834,15 +832,47 @@ nvc0_draw_indirect(struct nvc0_context *nvc0, const struct pipe_draw_info *info)
          IMMED_NVC0(push, NVC0_3D(VERTEX_ID_BASE), 0);
          nvc0->state.index_bias = 0;
       }
-      size = 4 * 4;
-      BEGIN_1IC0(push, NVC0_3D(MACRO_DRAW_ARRAYS_INDIRECT), 3 + size / 4);
-   }
-   PUSH_DATA(push, nvc0_prim_gl(info->mode));
-   PUSH_DATA(push, info->drawid);
-   PUSH_DATA(push, 1);
-#define NVC0_IB_ENTRY_1_NO_PREFETCH (1 << (31 - 8))
-   nouveau_pushbuf_data(push,
-                        buf->bo, offset, NVC0_IB_ENTRY_1_NO_PREFETCH | size);
+      size = 4;
+      macro = NVC0_3D_MACRO_DRAW_ARRAYS_INDIRECT;
+   }
+
+   /* If the stride is not the natural stride, we have to stick a separate
+    * push data reference for each draw. Otherwise it can all go in as one.
+    * Of course there is a maximum packet size, so we have to break things up
+    * along those borders as well.
+    */
+   while (count) {
+      unsigned draws = count, pushes, i;
+      if (info->indirect_stride == size * 4) {
+         draws = MIN2(draws, (NV04_PFIFO_MAX_PACKET_LEN - 4) / size);
+         pushes = 1;
+      } else {
+         draws = MIN2(draws, 32);
+         pushes = draws;
+      }
+
+      nouveau_pushbuf_space(push, 8, 0, pushes);
+      PUSH_REFN(push, buf->bo, NOUVEAU_BO_RD | buf->domain);
+      PUSH_DATA(push, NVC0_FIFO_PKHDR_1I(0, macro, 3 + draws * size));
+      PUSH_DATA(push, nvc0_prim_gl(info->mode));
+      PUSH_DATA(push, drawid);
+      PUSH_DATA(push, draws);
+      if (pushes == 1) {
+         nouveau_pushbuf_data(push,
+                              buf->bo, offset,
+                              NVC0_IB_ENTRY_1_NO_PREFETCH | (size * 4 * draws));
+         offset += draws * info->indirect_stride;
+      } else {
+         for (i = 0; i < pushes; i++) {
+            nouveau_pushbuf_data(push,
+                                 buf->bo, offset,
+                                 NVC0_IB_ENTRY_1_NO_PREFETCH | (size * 4));
+            offset += info->indirect_stride;
+         }
+      }
+      count -= draws;
+      drawid += draws;
+   }
 }
 
 static inline void
diff --git a/src/gallium/drivers/nouveau/nvc0/nvc0_winsys.h b/src/gallium/drivers/nouveau/nvc0/nvc0_winsys.h
index 4ea8ca3cfa2..79abe78b77a 100644
--- a/src/gallium/drivers/nouveau/nvc0/nvc0_winsys.h
+++ b/src/gallium/drivers/nouveau/nvc0/nvc0_winsys.h
@@ -68,6 +68,7 @@ PUSH_REFN(struct nouveau_pushbuf *push, struct nouveau_bo *bo, uint32_t flags)
 #define SUBC_SW(m) 7, (m)
 
 #define NVC0_3D_SERIALIZE NV50_GRAPH_SERIALIZE
+#define NVC0_IB_ENTRY_1_NO_PREFETCH (1 << (31 - 8))
 
 static inline uint32_t
 NVC0_FIFO_PKHDR_SQ(int subc, int mthd, unsigned size)