nvc0: support edge flags

4 files changed, 88 insertions, 14 deletions

diff --git a/src/gallium/drivers/nvc0/nvc0_context.h b/src/gallium/drivers/nvc0/nvc0_context.h
index aac358e142..102997e4fc 100644
--- a/src/gallium/drivers/nvc0/nvc0_context.h
+++ b/src/gallium/drivers/nvc0/nvc0_context.h
@@ -136,6 +136,9 @@ 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 9895f2141b..3c59213176 100644
--- a/src/gallium/drivers/nvc0/nvc0_program.c
+++ b/src/gallium/drivers/nvc0/nvc0_program.c
@@ -202,6 +202,8 @@ nvc0_varying_location(unsigned sn, unsigned si)
       return 0x40;
    case TGSI_SEMANTIC_FACE:
       return 0x3fc;
+   case TGSI_SEMANTIC_EDGEFLAG: /* doesn't exist, set value like for an sreg */
+      return 0xf00;
       /*
    case TGSI_SEMANTIC_CLIP_DISTANCE:
       return 0x2c0 + (si * 4);
@@ -290,6 +292,8 @@ prog_decl(struct nvc0_translation_info *ti,
                   ti->output_loc[i][c] = si * 4 + c;
             }
          } else {
+            if (sn == TGSI_SEMANTIC_EDGEFLAG)
+               ti->edgeflag_out = i;
             for (c = 0; c < 4; ++c)
                ti->output_loc[i][c] = nvc0_varying_location(sn, si) + c * 4;
             /* for TFB_VARYING_LOCS: */
@@ -427,6 +431,8 @@ nvc0_vp_gp_gen_header(struct nvc0_program *vp, struct nvc0_translation_info *ti)
 
    for (i = 0; i <= ti->scan.file_max[TGSI_FILE_OUTPUT]; ++i) {
       a = (ti->output_loc[i][0] - 0x40) / 4;
+      if (ti->output_loc[i][0] >= 0xf00)
+         continue;
       for (c = 0; c < 4; ++c, ++a) {
          if (!ti->output_access[i][c])
             continue;
@@ -670,6 +676,8 @@ nvc0_program_translate(struct nvc0_program *prog)
 
    ti->edgeflag_out = PIPE_MAX_SHADER_OUTPUTS;
 
+   prog->vp.edgeflag = PIPE_MAX_ATTRIBS;
+
    if (prog->type == PIPE_SHADER_VERTEX && prog->vp.num_ucps)
       ti->append_ucp = TRUE;
 
diff --git a/src/gallium/drivers/nvc0/nvc0_push.c b/src/gallium/drivers/nvc0/nvc0_push.c
index 68544c90d2..2e9f4c1092 100644
--- a/src/gallium/drivers/nvc0/nvc0_push.c
+++ b/src/gallium/drivers/nvc0/nvc0_push.c
@@ -15,9 +15,6 @@ struct push_context {
 
    void *idxbuf;
 
-   float edgeflag;
-   int edgeflag_attr;
-
    uint32_t vertex_words;
    uint32_t packet_vertex_limit;
 
@@ -27,8 +24,56 @@ struct push_context {
    uint32_t prim;
    uint32_t restart_index;
    uint32_t instance_id;
+
+   struct {
+      int buffer;
+      float value;
+      uint8_t *data;
+      unsigned offset;
+      unsigned stride;
+   } edgeflag;
 };
 
+static void
+init_push_context(struct nvc0_context *nvc0, struct push_context *ctx)
+{
+   struct pipe_vertex_element *ve;
+
+   ctx->chan = nvc0->screen->base.channel;
+   ctx->translate = nvc0->vertex->translate;
+
+   ctx->edgeflag.value = 0.5f;
+
+   if (NVC0_USING_EDGEFLAG(nvc0)) {
+      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;
+   }
+
+   ctx->vertex_words = nvc0->vertex->vtx_size;
+}
+
+static INLINE void
+set_edgeflag(struct push_context *ctx, unsigned vtx_id)
+{
+   float f = *(float *)(ctx->edgeflag.data + vtx_id * ctx->edgeflag.stride);
+
+   if (ctx->edgeflag.value != f) {
+      ctx->edgeflag.value = f;
+      IMMED_RING(ctx->chan, RING_3D(EDGEFLAG_ENABLE), f ? 1 : 0);
+   }
+}
+
 static INLINE unsigned
 prim_restart_search_i08(uint8_t *elts, unsigned push, uint8_t index)
 {
@@ -62,7 +107,7 @@ prim_restart_search_i32(uint32_t *elts, unsigned push, uint32_t index)
 static void
 emit_vertices_i08(struct push_context *ctx, unsigned start, unsigned count)
 {
-   uint8_t *elts = (uint8_t *)ctx->idxbuf + start;
+   uint8_t *restrict elts = (uint8_t *)ctx->idxbuf + start;
 
    while (count) {
       unsigned push = MIN2(count, ctx->packet_vertex_limit);
@@ -72,6 +117,9 @@ 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)
+         set_edgeflag(ctx, elts[0]);
+
       size = ctx->vertex_words * nr;
 
       BEGIN_RING_NI(ctx->chan, RING_3D(VERTEX_DATA), size);
@@ -97,7 +145,7 @@ emit_vertices_i08(struct push_context *ctx, unsigned start, unsigned count)
 static void
 emit_vertices_i16(struct push_context *ctx, unsigned start, unsigned count)
 {
-   uint16_t *elts = (uint16_t *)ctx->idxbuf + start;
+   uint16_t *restrict elts = (uint16_t *)ctx->idxbuf + start;
 
    while (count) {
       unsigned push = MIN2(count, ctx->packet_vertex_limit);
@@ -107,6 +155,9 @@ 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)
+         set_edgeflag(ctx, elts[0]);
+
       size = ctx->vertex_words * nr;
 
       BEGIN_RING_NI(ctx->chan, RING_3D(VERTEX_DATA), size);
@@ -132,7 +183,7 @@ emit_vertices_i16(struct push_context *ctx, unsigned start, unsigned count)
 static void
 emit_vertices_i32(struct push_context *ctx, unsigned start, unsigned count)
 {
-   uint32_t *elts = (uint32_t *)ctx->idxbuf + start;
+   uint32_t *restrict elts = (uint32_t *)ctx->idxbuf + start;
 
    while (count) {
       unsigned push = MIN2(count, ctx->packet_vertex_limit);
@@ -142,6 +193,9 @@ 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)
+         set_edgeflag(ctx, elts[0]);
+
       size = ctx->vertex_words * nr;
 
       BEGIN_RING_NI(ctx->chan, RING_3D(VERTEX_DATA), size);
@@ -171,6 +225,9 @@ emit_vertices_seq(struct push_context *ctx, unsigned start, unsigned count)
       unsigned push = MIN2(count, ctx->packet_vertex_limit);
       unsigned size = ctx->vertex_words * push;
 
+      if (unlikely(ctx->edgeflag.buffer >= 0))
+         set_edgeflag(ctx, start);
+
       BEGIN_RING_NI(ctx->chan, RING_3D(VERTEX_DATA), size);
 
       ctx->translate->run(ctx->translate, start, push, ctx->instance_id,
@@ -219,10 +276,7 @@ nvc0_push_vbo(struct nvc0_context *nvc0, const struct pipe_draw_info *info)
    unsigned inst = info->instance_count;
    boolean apply_bias = info->indexed && info->index_bias;
 
-   ctx.chan = nvc0->screen->base.channel;
-   ctx.translate = nvc0->vertex->translate;
-   ctx.packet_vertex_limit = nvc0->vertex->vtx_per_packet_max;
-   ctx.vertex_words = nvc0->vertex->vtx_size;
+   init_push_context(nvc0, &ctx);
 
    for (i = 0; i < nvc0->num_vtxbufs; ++i) {
       uint8_t *data;
@@ -236,12 +290,18 @@ nvc0_push_vbo(struct nvc0_context *nvc0, const struct pipe_draw_info *info)
          data += info->index_bias * vb->stride;
 
       ctx.translate->set_buffer(ctx.translate, i, data, vb->stride, ~0);
+
+      if (unlikely(i == ctx.edgeflag.buffer)) {
+         ctx.edgeflag.data = data + ctx.edgeflag.offset;
+         ctx.edgeflag.stride = vb->stride;
+      }
    }
 
    if (info->indexed) {
-      ctx.idxbuf = nouveau_resource_map_offset(&nvc0->base,
-                                               nv04_resource(nvc0->idxbuf.buffer),
-                                               nvc0->idxbuf.offset, NOUVEAU_BO_RD);
+      ctx.idxbuf =
+         nouveau_resource_map_offset(&nvc0->base,
+                                     nv04_resource(nvc0->idxbuf.buffer),
+                                     nvc0->idxbuf.offset, NOUVEAU_BO_RD);
       if (!ctx.idxbuf)
          return;
       index_size = nvc0->idxbuf.index_size;
@@ -283,6 +343,9 @@ nvc0_push_vbo(struct nvc0_context *nvc0, const struct pipe_draw_info *info)
       ctx.prim |= NVC0_3D_VERTEX_BEGIN_GL_INSTANCE_NEXT;
    }
 
+   if (unlikely(ctx.edgeflag.value == 0.0f))
+      IMMED_RING(ctx.chan, RING_3D(EDGEFLAG_ENABLE), 1);
+
    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 e7e7ce7dc2..3698043029 100644
--- a/src/gallium/drivers/nvc0/nvc0_vbo.c
+++ b/src/gallium/drivers/nvc0/nvc0_vbo.c
@@ -265,7 +265,7 @@ nvc0_vertex_arrays_validate(struct nvc0_context *nvc0)
    struct nvc0_vertex_element *ve;
    unsigned i;
 
-   if (unlikely(vertex->need_conversion)) {
+   if (unlikely(vertex->need_conversion || NVC0_USING_EDGEFLAG(nvc0))) {
       nvc0->vbo_fifo = ~0;
       nvc0->vbo_user = 0;
    } else {