nvc0: fix submission of VertexID and EdgeFlag in push mode

NOTE: This is a candidate for the 8.0 branch.

9 files changed, 84 insertions, 21 deletions

diff --git a/src/gallium/drivers/nv50/codegen/nv50_ir.cpp b/src/gallium/drivers/nv50/codegen/nv50_ir.cpp
index a8f05290df2..7d9b7e2b760 100644
--- a/src/gallium/drivers/nv50/codegen/nv50_ir.cpp
+++ b/src/gallium/drivers/nv50/codegen/nv50_ir.cpp
@@ -938,6 +938,7 @@ nv50_ir_init_prog_info(struct nv50_ir_prog_info *info)
    }
    info->io.clipDistance = 0xff;
    info->io.pointSize = 0xff;
+   info->io.vertexId = 0xff;
    info->io.edgeFlagIn = 0xff;
    info->io.edgeFlagOut = 0xff;
    info->io.fragDepth = 0xff;
diff --git a/src/gallium/drivers/nv50/codegen/nv50_ir_driver.h b/src/gallium/drivers/nv50/codegen/nv50_ir_driver.h
index 73fb023c378..a37a29a99a6 100644
--- a/src/gallium/drivers/nv50/codegen/nv50_ir_driver.h
+++ b/src/gallium/drivers/nv50/codegen/nv50_ir_driver.h
@@ -155,6 +155,7 @@ struct nv50_ir_prog_info
       uint8_t cullDistanceMask;  /* clip distance mode (1 bit per output) */
       int8_t genUserClip;        /* request user clip planes for ClipVertex */
       uint8_t pointSize;         /* output index for PointSize */
+      uint8_t vertexId;          /* system value index of VertexID */
       uint8_t edgeFlagIn;
       uint8_t edgeFlagOut;
       uint8_t fragDepth;         /* output index of FragDepth */
diff --git a/src/gallium/drivers/nv50/codegen/nv50_ir_from_tgsi.cpp b/src/gallium/drivers/nv50/codegen/nv50_ir_from_tgsi.cpp
index a1c8f84d7ed..c104dbe9670 100644
--- a/src/gallium/drivers/nv50/codegen/nv50_ir_from_tgsi.cpp
+++ b/src/gallium/drivers/nv50/codegen/nv50_ir_from_tgsi.cpp
@@ -864,6 +864,13 @@ bool Source::scanDeclaration(const struct tgsi_full_declaration *decl)
       }
       break;
    case TGSI_FILE_SYSTEM_VALUE:
+      switch (sn) {
+      case TGSI_SEMANTIC_VERTEXID:
+         info->io.vertexId = first;
+         break;
+      default:
+         break;
+      }
       for (i = first; i <= last; ++i, ++si) {
          info->sv[i].sn = sn;
          info->sv[i].si = si;
diff --git a/src/gallium/drivers/nvc0/nvc0_3d.xml.h b/src/gallium/drivers/nvc0/nvc0_3d.xml.h
index 8d6ea7e2745..6a1dff70f76 100644
--- a/src/gallium/drivers/nvc0/nvc0_3d.xml.h
+++ b/src/gallium/drivers/nvc0/nvc0_3d.xml.h
@@ -913,6 +913,11 @@ WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE SOFTWARE.
 #define NVC0_3D_VERTEX_BEGIN_GL_INSTANCE_NEXT			0x04000000
 #define NVC0_3D_VERTEX_BEGIN_GL_INSTANCE_CONT			0x08000000
 
+#define NVC0_3D_VERTEX_ID_REPLACE				0x0000161c
+#define NVC0_3D_VERTEX_ID_REPLACE_ENABLE			0x00000001
+#define NVC0_3D_VERTEX_ID_REPLACE_SOURCE__MASK			0x00000ff0
+#define NVC0_3D_VERTEX_ID_REPLACE_SOURCE__SHIFT			4
+
 #define NVC0_3D_VERTEX_DATA					0x00001640
 
 #define NVC0_3D_PRIM_RESTART_ENABLE				0x00001644
diff --git a/src/gallium/drivers/nvc0/nvc0_context.h b/src/gallium/drivers/nvc0/nvc0_context.h
index af95d1ab6aa..b8f0d92b6b4 100644
--- a/src/gallium/drivers/nvc0/nvc0_context.h
+++ b/src/gallium/drivers/nvc0/nvc0_context.h
@@ -134,9 +134,6 @@ struct nvc0_context {
    struct draw_context *draw;
 };
 
-#define NVC0_USING_EDGEFLAG(ctx) \
-   ((ctx)->vertprog->vp.edgeflag < PIPE_MAX_ATTRIBS)
-
 static INLINE struct nvc0_context *
 nvc0_context(struct pipe_context *pipe)
 {
diff --git a/src/gallium/drivers/nvc0/nvc0_program.c b/src/gallium/drivers/nvc0/nvc0_program.c
index 813008cd537..cff76fe67f3 100644
--- a/src/gallium/drivers/nvc0/nvc0_program.c
+++ b/src/gallium/drivers/nvc0/nvc0_program.c
@@ -107,7 +107,7 @@ nvc0_vp_assign_input_slots(struct nv50_ir_prog_info *info)
 
    for (n = 0, i = 0; i < info->numInputs; ++i) {
       switch (info->in[i].sn) {
-      case TGSI_SEMANTIC_INSTANCEID:
+      case TGSI_SEMANTIC_INSTANCEID: /* for SM4 only, in TGSI they're SVs */
       case TGSI_SEMANTIC_VERTEXID:
          info->in[i].mask = 0x1;
          info->in[i].slot[0] =
@@ -580,7 +580,11 @@ nvc0_program_translate(struct nvc0_program *prog)
    prog->relocs = info->bin.relocData;
    prog->max_gpr = MAX2(4, (info->bin.maxGPR + 1));
 
-   prog->vp.edgeflag = PIPE_MAX_ATTRIBS;
+   prog->vp.need_vertex_id = info->io.vertexId < PIPE_MAX_SHADER_INPUTS;
+
+   if (info->io.edgeFlagOut < PIPE_MAX_ATTRIBS)
+      info->out[info->io.edgeFlagOut].mask = 0; /* for headergen */
+   prog->vp.edgeflag = info->io.edgeFlagIn;
 
    switch (prog->type) {
    case PIPE_SHADER_VERTEX:
diff --git a/src/gallium/drivers/nvc0/nvc0_program.h b/src/gallium/drivers/nvc0/nvc0_program.h
index c384ef534d5..6eb8c96d60a 100644
--- a/src/gallium/drivers/nvc0/nvc0_program.h
+++ b/src/gallium/drivers/nvc0/nvc0_program.h
@@ -37,8 +37,9 @@ struct nvc0_program {
    struct {
       uint32_t clip_mode; /* clip/cull selection */
       uint8_t clip_enable; /* mask of defined clip planes */
-      uint8_t edgeflag;
       uint8_t num_ucps; /* also set to max if ClipDistance is used */
+      uint8_t edgeflag; /* attribute index of edgeflag input */
+      boolean need_vertex_id;
    } vp;
    struct {
       uint8_t early_z;
diff --git a/src/gallium/drivers/nvc0/nvc0_push.c b/src/gallium/drivers/nvc0/nvc0_push.c
index 238671d721c..412186c2345 100644
--- a/src/gallium/drivers/nvc0/nvc0_push.c
+++ b/src/gallium/drivers/nvc0/nvc0_push.c
@@ -21,6 +21,7 @@ struct push_context {
    struct translate *translate;
 
    boolean primitive_restart;
+   boolean need_vertex_id;
    uint32_t prim;
    uint32_t restart_index;
    uint32_t instance_id;
@@ -42,22 +43,23 @@ init_push_context(struct nvc0_context *nvc0, struct push_context *ctx)
    ctx->chan = nvc0->screen->base.channel;
    ctx->translate = nvc0->vertex->translate;
 
+   if (likely(nvc0->vertex->num_elements < 32))
+      ctx->need_vertex_id = nvc0->vertprog->vp.need_vertex_id;
+   else
+      ctx->need_vertex_id = FALSE;
+
+   ctx->edgeflag.buffer = -1;
    ctx->edgeflag.value = 0.5f;
 
-   if (NVC0_USING_EDGEFLAG(nvc0)) {
+   if (unlikely(nvc0->vertprog->vp.edgeflag < PIPE_MAX_ATTRIBS)) {
       ve = &nvc0->vertex->element[nvc0->vertprog->vp.edgeflag].pipe;
-
       ctx->edgeflag.buffer = ve->vertex_buffer_index;
       ctx->edgeflag.offset = ve->src_offset;
-
       ctx->packet_vertex_limit = 1;
    } else {
-      ctx->edgeflag.buffer = -1;
-      ctx->edgeflag.offset = 0;
-      ctx->edgeflag.stride = 0;
-      ctx->edgeflag.data = NULL;
-
       ctx->packet_vertex_limit = nvc0->vertex->vtx_per_packet_max;
+      if (unlikely(ctx->need_vertex_id))
+         ctx->packet_vertex_limit = 1;
    }
 
    ctx->vertex_words = nvc0->vertex->vtx_size;
@@ -74,6 +76,17 @@ set_edgeflag(struct push_context *ctx, unsigned vtx_id)
    }
 }
 
+static INLINE void
+set_vertexid(struct push_context *ctx, uint32_t vtx_id)
+{
+#if 0
+   BEGIN_RING(ctx->chan, RING_3D(VERTEX_ID), 1); /* broken on nvc0 */
+#else
+   BEGIN_RING(ctx->chan, RING_3D(VERTEX_DATA), 1); /* as last attribute */
+#endif
+   OUT_RING  (ctx->chan, vtx_id);
+}
+
 static INLINE unsigned
 prim_restart_search_i08(uint8_t *elts, unsigned push, uint8_t index)
 {
@@ -117,7 +130,7 @@ emit_vertices_i08(struct push_context *ctx, unsigned start, unsigned count)
       if (ctx->primitive_restart)
          nr = prim_restart_search_i08(elts, push, ctx->restart_index);
 
-      if (unlikely(ctx->edgeflag.buffer >= 0) && nr)
+      if (unlikely(ctx->edgeflag.buffer >= 0) && likely(nr))
          set_edgeflag(ctx, elts[0]);
 
       size = ctx->vertex_words * nr;
@@ -126,8 +139,11 @@ emit_vertices_i08(struct push_context *ctx, unsigned start, unsigned count)
 
       ctx->translate->run_elts8(ctx->translate, elts, nr, ctx->instance_id,
                                 ctx->chan->cur);
-
       ctx->chan->cur += size;
+
+      if (unlikely(ctx->need_vertex_id) && likely(size))
+         set_vertexid(ctx, elts[0]);
+
       count -= nr;
       elts += nr;
 
@@ -155,7 +171,7 @@ emit_vertices_i16(struct push_context *ctx, unsigned start, unsigned count)
       if (ctx->primitive_restart)
          nr = prim_restart_search_i16(elts, push, ctx->restart_index);
 
-      if (unlikely(ctx->edgeflag.buffer >= 0) && nr)
+      if (unlikely(ctx->edgeflag.buffer >= 0) && likely(nr))
          set_edgeflag(ctx, elts[0]);
 
       size = ctx->vertex_words * nr;
@@ -164,8 +180,11 @@ emit_vertices_i16(struct push_context *ctx, unsigned start, unsigned count)
 
       ctx->translate->run_elts16(ctx->translate, elts, nr, ctx->instance_id,
                                  ctx->chan->cur);
-
       ctx->chan->cur += size;
+
+      if (unlikely(ctx->need_vertex_id))
+         set_vertexid(ctx, elts[0]);
+
       count -= nr;
       elts += nr;
 
@@ -193,7 +212,7 @@ emit_vertices_i32(struct push_context *ctx, unsigned start, unsigned count)
       if (ctx->primitive_restart)
          nr = prim_restart_search_i32(elts, push, ctx->restart_index);
 
-      if (unlikely(ctx->edgeflag.buffer >= 0) && nr)
+      if (unlikely(ctx->edgeflag.buffer >= 0) && likely(nr))
          set_edgeflag(ctx, elts[0]);
 
       size = ctx->vertex_words * nr;
@@ -202,8 +221,11 @@ emit_vertices_i32(struct push_context *ctx, unsigned start, unsigned count)
 
       ctx->translate->run_elts(ctx->translate, elts, nr, ctx->instance_id,
                                ctx->chan->cur);
-
       ctx->chan->cur += size;
+
+      if (unlikely(ctx->need_vertex_id))
+         set_vertexid(ctx, elts[0]);
+
       count -= nr;
       elts += nr;
 
@@ -233,6 +255,10 @@ emit_vertices_seq(struct push_context *ctx, unsigned start, unsigned count)
       ctx->translate->run(ctx->translate, start, push, ctx->instance_id,
                           ctx->chan->cur);
       ctx->chan->cur += size;
+
+      if (unlikely(ctx->need_vertex_id))
+         set_vertexid(ctx, start);
+
       count -= push;
       start += push;
    }
@@ -326,6 +352,16 @@ nvc0_push_vbo(struct nvc0_context *nvc0, const struct pipe_draw_info *info)
    ctx.instance_id = info->start_instance;
    ctx.prim = nvc0_prim_gl(info->mode);
 
+   if (unlikely(ctx.need_vertex_id)) {
+      const unsigned a = nvc0->vertex->num_elements;
+      BEGIN_RING(ctx.chan, RING_3D(VERTEX_ATTRIB_FORMAT(a)), 1);
+      OUT_RING  (ctx.chan, (a << NVC0_3D_VERTEX_ATTRIB_FORMAT_BUFFER__SHIFT) |
+                 NVC0_3D_VERTEX_ATTRIB_FORMAT_TYPE_FLOAT |
+                 NVC0_3D_VERTEX_ATTRIB_FORMAT_SIZE_32);
+      BEGIN_RING(ctx.chan, RING_3D(VERTEX_ID_REPLACE), 1);
+      OUT_RING  (ctx.chan, (((0x80 + a * 0x10) / 4) << 4) | 1);
+   }
+
    while (inst_count--) {
       BEGIN_RING(ctx.chan, RING_3D(VERTEX_BEGIN_GL), 1);
       OUT_RING  (ctx.chan, ctx.prim);
@@ -355,6 +391,16 @@ nvc0_push_vbo(struct nvc0_context *nvc0, const struct pipe_draw_info *info)
    if (unlikely(ctx.edgeflag.value == 0.0f))
       IMMED_RING(ctx.chan, RING_3D(EDGEFLAG_ENABLE), 1);
 
+   if (unlikely(ctx.need_vertex_id)) {
+      const unsigned a = nvc0->vertex->num_elements;
+      IMMED_RING(ctx.chan, RING_3D(VERTEX_ID_REPLACE), 0);
+      BEGIN_RING(ctx.chan, RING_3D(VERTEX_ATTRIB_FORMAT(a)), 1);
+      OUT_RING  (ctx.chan,
+                 NVC0_3D_VERTEX_ATTRIB_FORMAT_CONST |
+                 NVC0_3D_VERTEX_ATTRIB_FORMAT_TYPE_FLOAT |
+                 NVC0_3D_VERTEX_ATTRIB_FORMAT_SIZE_32);
+   }
+
    if (info->indexed)
       nouveau_resource_unmap(nv04_resource(nvc0->idxbuf.buffer));
 
diff --git a/src/gallium/drivers/nvc0/nvc0_vbo.c b/src/gallium/drivers/nvc0/nvc0_vbo.c
index 7cf69910e6a..3e95d509f99 100644
--- a/src/gallium/drivers/nvc0/nvc0_vbo.c
+++ b/src/gallium/drivers/nvc0/nvc0_vbo.c
@@ -263,7 +263,8 @@ nvc0_vertex_arrays_validate(struct nvc0_context *nvc0)
    struct nvc0_vertex_element *ve;
    unsigned i;
 
-   if (unlikely(vertex->need_conversion || NVC0_USING_EDGEFLAG(nvc0))) {
+   if (unlikely(vertex->need_conversion) ||
+       unlikely(nvc0->vertprog->vp.edgeflag < PIPE_MAX_ATTRIBS)) {
       nvc0->vbo_fifo = ~0;
       nvc0->vbo_user = 0;
    } else {