nvc0: adjust indirect draw macros to handle multiple draws at once

These are still invoked one at a time, but the underlying macro can
handle multiple draws.

Signed-off-by: Ilia Mirkin <[email protected]>

3 files changed, 101 insertions, 52 deletions

diff --git a/src/gallium/drivers/nouveau/nvc0/mme/com9097.mme b/src/gallium/drivers/nouveau/nvc0/mme/com9097.mme
index 8c647d0c66c..35355edf2e7 100644
--- a/src/gallium/drivers/nouveau/nvc0/mme/com9097.mme
+++ b/src/gallium/drivers/nouveau/nvc0/mme/com9097.mme
@@ -230,29 +230,43 @@ locn_0f_ts:
  * Forcefully sets VERTEX_ID_BASE to the value of VB_ELEMENT_BASE.
  *
  * arg     = mode
- * parm[0] = count
- * parm[1] = instance_count
- * parm[2] = start
- * parm[3] = index_bias
- * parm[4] = start_instance
+ * parm[0] = start_drawid
+ * parm[1] = numparams
+ * parm[2 + 5n + 0] = count
+ * parm[2 + 5n + 1] = instance_count
+ * parm[2 + 5n + 2] = start
+ * parm[2 + 5n + 3] = index_bias
+ * parm[2 + 5n + 4] = start_instance
+ *
+ * SCRATCH[0] = saved VB_ELEMENT_BASE
+ * SCRATCH[1] = saved VB_INSTANCE_BASE
  */
 .section #mme9097_draw_elts_indirect
+   read $r6 0x50d /* VB_ELEMENT_BASE */
+   read $r7 0x50e /* VB_INSTANCE_BASE */
+   maddr 0x1d00
+   send $r6 /* SCRATCH[0] = VB_ELEMENT_BASE */
+   send $r7 /* SCRATCH[1] = VB_INSTANCE_BASE */
+   parm $r6 /* start_drawid */
+   parm $r7 /* numparams */
+dei_draw_again:
    parm $r3 /* count */
    parm $r2 /* instance_count */
    parm $r4 maddr 0x5f7 /* INDEX_BATCH_FIRST, start */
    parm $r4 send $r4 /* index_bias, send start */
-   maddr 0x8e4 /* CB_DATA */
+   maddr 0x18e3 /* CB_POS */
+   send 0x180 /* 256 + 128 */
    braz $r2 #dei_end
    parm $r5 send $r4 /* start_instance, send index_bias */
    send $r5 /* send start_instance */
-   read $r6 0x50d /* VB_ELEMENT_BASE */
-   read $r7 0x50e /* VB_INSTANCE_BASE */
+   send $r6 /* draw id */
    maddr 0x150d /* VB_ELEMENT,INSTANCE_BASE */
    send $r4
    send $r5
    maddr 0x446
    send $r4
    mov $r4 0x1
+   mov $r1 (extrinsrt $r1 0x0 0 1 26) /* clear INSTANCE_NEXT */
 dei_again:
    maddr 0x586 /* VERTEX_BEGIN_GL */
    send $r1 /* mode */
@@ -262,48 +276,61 @@ dei_again:
    maddrsend 0x585 /* VERTEX_END_GL */
    branz $r2 #dei_again
    mov $r1 (extrinsrt $r1 $r4 0 1 26) /* set INSTANCE_NEXT */
+dei_end:
+   mov $r7 (add $r7 -1)
+   branz $r7 #dei_draw_again
+   mov $r6 (add $r6 1)
+   read $r6 0xd00
+   read $r7 0xd01
    maddr 0x150d /* VB_ELEMENT,INSTANCE_BASE */
    send $r6
    send $r7
    exit maddr 0x446
    send $r6
-dei_end:
-   exit
-   nop
 
 /* NVC0_3D_MACRO_DRAW_ARRAYS_INDIRECT:
  *
  * NOTE: Saves and restores VB_INSTANCE_BASE.
  *
  * arg     = mode
- * parm[0] = count
- * parm[1] = instance_count
- * parm[2] = start
- * parm[3] = start_instance
+ * parm[0] = start_drawid
+ * parm[1] = numparams
+ * parm[2 + 4n + 0] = count
+ * parm[2 + 4n + 1] = instance_count
+ * parm[2 + 4n + 2] = start
+ * parm[2 + 4n + 3] = start_instance
  */
 .section #mme9097_draw_arrays_indirect
+   read $r5 0x50e /* VB_INSTANCE_BASE */
+   parm $r6 /* start_drawid */
+   parm $r7 /* numparams */
+dai_draw_again:
    parm $r2 /* count */
    parm $r3 /* instance_count */
    parm $r4 maddr 0x35d /* VERTEX_BUFFER_FIRST, start */
    braz $r3 #dai_end
    parm $r4 send $r4 /* start_instance */
-   maddrsend 0x8e4 /* CB_DATA, send 0 as base_vertex */
+   maddr 0x18e3 /* CB_POS */
+   send 0x180 /* 256 + 128 */
+   send 0x0 /* send 0 as base_vertex */
    send $r4 /* send start_instance */
-   read $r6 0x50e /* VB_INSTANCE_BASE */
+   send $r6 /* draw id */
    maddr 0x50e /* VB_INSTANCE_BASE */
-   mov $r5 0x1
    send $r4
+   mov $r4 0x1
+   mov $r1 (extrinsrt $r1 0x0 0 1 26) /* clear INSTANCE_NEXT */
 dai_again:
    maddr 0x586 /* VERTEX_BEGIN_GL */
    send $r1 /* mode */
    maddr 0x35e /* VERTEX_BUFFER_COUNT */
    send $r2
-   mov $r3 (sub $r3 $r5)
+   mov $r3 (sub $r3 $r4)
    maddrsend 0x585 /* VERTEX_END_GL */
    branz $r3 #dai_again
-   mov $r1 (extrinsrt $r1 $r5 0 1 26) /* set INSTANCE_NEXT */
-   exit maddr 0x50e /* VB_INSTANCE_BASE to restore */
-   send $r6
+   mov $r1 (extrinsrt $r1 $r4 0 1 26) /* set INSTANCE_NEXT */
 dai_end:
-   exit
-   nop
+   mov $r7 (add $r7 -1)
+   branz $r7 #dai_draw_again
+   mov $r6 (add $r6 1)
+   exit maddr 0x50e /* VB_INSTANCE_BASE to restore */
+   send $r5
diff --git a/src/gallium/drivers/nouveau/nvc0/mme/com9097.mme.h b/src/gallium/drivers/nouveau/nvc0/mme/com9097.mme.h
index acad303ce60..0aebeeb6e66 100644
--- a/src/gallium/drivers/nouveau/nvc0/mme/com9097.mme.h
+++ b/src/gallium/drivers/nouveau/nvc0/mme/com9097.mme.h
@@ -125,24 +125,33 @@ uint32_t mme9097_tep_select[] = {
 };
 
 uint32_t mme9097_draw_elts_indirect[] = {
+	0x01434615,
+/* 0x0007: dei_draw_again */
+	0x01438715,
+	0x07400021,
+	0x00003041,
+	0x00003841,
+	0x00000601,
+/* 0x0018: dei_again */
+	0x00000701,
 	0x00000301,
+/* 0x0020: dei_end */
 	0x00000201,
 	0x017dc451,
 	0x00002431,
-/* 0x0010: dei_again */
-	0x02390021,
-	0x00061007,
+	0x0638c021,
+	0x00600041,
+	0x0004d007,
 	0x00002531,
-/* 0x001d: dei_end */
 	0x00002841,
-	0x01434615,
-	0x01438715,
+	0x00003041,
 	0x05434021,
 	0x00002041,
 	0x00002841,
 	0x01118021,
 	0x00002041,
 	0x00004411,
+	0xd0400912,
 	0x01618021,
 	0x00000841,
 	0x017e0021,
@@ -151,39 +160,50 @@ uint32_t mme9097_draw_elts_indirect[] = {
 	0x01614071,
 	0xfffe9017,
 	0xd0410912,
+	0xffffff11,
+	0xfff9b817,
+	0x00007611,
+	0x03400615,
+	0x03404715,
 	0x05434021,
 	0x00003041,
 	0x00003841,
 	0x011180a1,
 	0x00003041,
-	0x00000091,
-	0x00000011,
 };
 
 uint32_t mme9097_draw_arrays_indirect[] = {
+/* 0x0003: dai_draw_again */
+	0x01438515,
+	0x00000601,
+	0x00000701,
 	0x00000201,
+/* 0x0011: dai_again */
 	0x00000301,
-/* 0x000b: dai_again */
 	0x00d74451,
-	0x00049807,
+/* 0x0019: dai_end */
+	0x0004d807,
 	0x00002431,
-/* 0x0015: dai_end */
-	0x02390071,
+	0x0638c021,
+	0x00600041,
+	0x00000041,
 	0x00002041,
-	0x01438615,
+	0x00003041,
 	0x01438021,
-	0x00004511,
 	0x00002041,
+	0x00004411,
+	0xd0400912,
 	0x01618021,
 	0x00000841,
 	0x00d78021,
 	0x00001041,
-	0x00055b10,
+	0x00051b10,
 	0x01614071,
 	0xfffe9817,
-	0xd0414912,
+	0xd0410912,
+	0xffffff11,
+	0xfffa7817,
+	0x00007611,
 	0x014380a1,
-	0x00003041,
-	0x00000091,
-	0x00000011,
+	0x00002841,
 };
diff --git a/src/gallium/drivers/nouveau/nvc0/nvc0_vbo.c b/src/gallium/drivers/nouveau/nvc0/nvc0_vbo.c
index 251753357eb..55aeb806288 100644
--- a/src/gallium/drivers/nouveau/nvc0/nvc0_vbo.c
+++ b/src/gallium/drivers/nouveau/nvc0/nvc0_vbo.c
@@ -819,8 +819,6 @@ nvc0_draw_indirect(struct nvc0_context *nvc0, const struct pipe_draw_info *info)
    PUSH_DATA (push, 512);
    PUSH_DATAh(push, nvc0->screen->uniform_bo->offset + (5 << 16) + (0 << 9));
    PUSH_DATA (push, nvc0->screen->uniform_bo->offset + (5 << 16) + (0 << 9));
-   BEGIN_NVC0(push, NVC0_3D(CB_POS), 1);
-   PUSH_DATA (push, 256 + 128);
 
    nouveau_pushbuf_space(push, 8, 0, 1);
    PUSH_REFN(push, buf->bo, NOUVEAU_BO_RD | buf->domain);
@@ -828,7 +826,7 @@ nvc0_draw_indirect(struct nvc0_context *nvc0, const struct pipe_draw_info *info)
       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), 1 + size / 4);
+      BEGIN_1IC0(push, NVC0_3D(MACRO_DRAW_ELEMENTS_INDIRECT), 3 + size / 4);
    } else {
       if (nvc0->state.index_bias) {
          /* index_bias is implied 0 if !info->indexed (really ?) */
@@ -837,9 +835,11 @@ nvc0_draw_indirect(struct nvc0_context *nvc0, const struct pipe_draw_info *info)
          nvc0->state.index_bias = 0;
       }
       size = 4 * 4;
-      BEGIN_1IC0(push, NVC0_3D(MACRO_DRAW_ARRAYS_INDIRECT), 1 + size / 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);
@@ -913,11 +913,13 @@ nvc0_draw_vbo(struct pipe_context *pipe, const struct pipe_draw_info *info)
       PUSH_DATA (push, 512);
       PUSH_DATAh(push, nvc0->screen->uniform_bo->offset + (5 << 16) + (0 << 9));
       PUSH_DATA (push, nvc0->screen->uniform_bo->offset + (5 << 16) + (0 << 9));
-      BEGIN_1IC0(push, NVC0_3D(CB_POS), 1 + 3);
-      PUSH_DATA (push, 256 + 128);
-      PUSH_DATA (push, info->index_bias);
-      PUSH_DATA (push, info->start_instance);
-      PUSH_DATA (push, info->drawid);
+      if (!info->indirect) {
+         BEGIN_1IC0(push, NVC0_3D(CB_POS), 1 + 3);
+         PUSH_DATA (push, 256 + 128);
+         PUSH_DATA (push, info->index_bias);
+         PUSH_DATA (push, info->start_instance);
+         PUSH_DATA (push, info->drawid);
+      }
    }
 
    push->kick_notify = nvc0_draw_vbo_kick_notify;