radeonsi: add gfx10 helpers for determining whether edgeflags are enabled

They will return false when culling lines.

Acked-by: Pierre-Eric Pelloux-Prayer <pierre-eric.pelloux-prayer@amd.com>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/13048>

src/gallium/drivers/radeonsi/gfx10_shader_ngg.c

@@ -114,7 +114,8 @@ bool gfx10_ngg_export_prim_early(struct si_shader *shader)
 
    assert(shader->key.as_ngg && !shader->key.as_es);
 
-   return sel->info.stage != MESA_SHADER_GEOMETRY && !sel->info.writes_edgeflag;
+   return sel->info.stage != MESA_SHADER_GEOMETRY &&
+          !gfx10_ngg_writes_user_edgeflags(shader);
 }
 
 void gfx10_ngg_build_sendmsg_gs_alloc_req(struct si_shader_context *ctx)
@@ -146,7 +147,7 @@ void gfx10_ngg_build_export_prim(struct si_shader_context *ctx, LLVMValueRef use
          /* This is only used with NGG culling, which returns the NGG
           * passthrough prim export encoding.
           */
-         if (ctx->shader->selector->info.writes_edgeflag) {
+         if (gfx10_ngg_writes_user_edgeflags(ctx->shader)) {
             unsigned all_bits_no_edgeflags = ~SI_NGG_PRIM_EDGE_FLAG_BITS;
             LLVMValueRef edgeflags = LLVMConstInt(ctx->ac.i32, all_bits_no_edgeflags, 0);
 
@@ -179,8 +180,7 @@ void gfx10_ngg_build_export_prim(struct si_shader_context *ctx, LLVMValueRef use
 
       prim.isnull = ctx->ac.i1false;
 
-      if (ctx->stage == MESA_SHADER_VERTEX &&
-          !ctx->shader->selector->info.base.vs.blit_sgprs_amd)
+      if (gfx10_edgeflags_have_effect(ctx->shader))
          prim.edgeflags = ac_pack_edgeflags_for_export(&ctx->ac, &ctx->args);
       else
          prim.edgeflags = ctx->ac.i32_0;
@@ -188,7 +188,7 @@ void gfx10_ngg_build_export_prim(struct si_shader_context *ctx, LLVMValueRef use
       for (unsigned i = 0; i < prim.num_vertices; ++i)
          prim.index[i] = si_unpack_param(ctx, ctx->args.gs_vtx_offset[i / 2], (i & 1) * 16, 16);
 
-      if (ctx->shader->selector->info.writes_edgeflag) {
+      if (gfx10_ngg_writes_user_edgeflags(ctx->shader)) {
          LLVMValueRef edgeflags = ctx->ac.i32_0;
 
          for (unsigned i = 0; i < prim.num_vertices; ++i) {
@@ -603,7 +603,7 @@ static unsigned ngg_nogs_vertex_size(struct si_shader *shader)
     * used for padding to reduce LDS bank conflicts. */
    if (shader->selector->so.num_outputs)
       lds_vertex_size = 4 * shader->selector->info.num_outputs + 1;
-   if (shader->selector->info.writes_edgeflag)
+   if (gfx10_ngg_writes_user_edgeflags(shader))
       lds_vertex_size = MAX2(lds_vertex_size, 1);
 
    /* LDS size for passing data from GS to ES.
@@ -1161,7 +1161,7 @@ void gfx10_emit_ngg_culling_epilogue(struct ac_shader_abi *abi)
       prim.num_vertices = num_vertices;
       prim.isnull = ctx->ac.i1false;
 
-      if (ctx->stage == MESA_SHADER_VERTEX)
+      if (gfx10_edgeflags_have_effect(shader))
          prim.edgeflags = ac_pack_edgeflags_for_export(&ctx->ac, &ctx->args);
       else
          prim.edgeflags = ctx->ac.i32_0;
@@ -1272,7 +1272,7 @@ void gfx10_emit_ngg_culling_epilogue(struct ac_shader_abi *abi)
    }
 
    /* These two also use LDS. */
-   if (sel->info.writes_edgeflag ||
+   if (gfx10_ngg_writes_user_edgeflags(shader) ||
        (ctx->stage == MESA_SHADER_VERTEX && shader->key.mono.u.vs_export_prim_id))
       ac_build_s_barrier(&ctx->ac);
 
@@ -1297,7 +1297,7 @@ void gfx10_emit_ngg_epilogue(struct ac_shader_abi *abi)
 
    LLVMValueRef vertex_ptr = NULL;
 
-   if (sel->so.num_outputs || sel->info.writes_edgeflag)
+   if (sel->so.num_outputs || gfx10_ngg_writes_user_edgeflags(ctx->shader))
       vertex_ptr = ngg_nogs_vertex_ptr(ctx, get_thread_id_in_tg(ctx));
 
    for (unsigned i = 0; i < info->num_outputs; i++) {
@@ -1318,7 +1318,7 @@ void gfx10_emit_ngg_epilogue(struct ac_shader_abi *abi)
       }
 
       /* Store the edgeflag at the end (if streamout is enabled) */
-      if (info->output_semantic[i] == VARYING_SLOT_EDGE && sel->info.writes_edgeflag) {
+      if (info->output_semantic[i] == VARYING_SLOT_EDGE && gfx10_ngg_writes_user_edgeflags(ctx->shader)) {
          LLVMValueRef edgeflag = LLVMBuildLoad(builder, addrs[4 * i], "");
          /* The output is a float, but the hw expects a 1-bit integer. */
          edgeflag = LLVMBuildFPToUI(ctx->ac.builder, edgeflag, ctx->ac.i32, "");
@@ -1331,7 +1331,7 @@ void gfx10_emit_ngg_epilogue(struct ac_shader_abi *abi)
    }
 
    bool unterminated_es_if_block =
-      !sel->so.num_outputs && !sel->info.writes_edgeflag &&
+      !sel->so.num_outputs && !gfx10_ngg_writes_user_edgeflags(ctx->shader) &&
       !ctx->screen->use_ngg_streamout && /* no query buffer */
       (ctx->stage != MESA_SHADER_VERTEX || !ctx->shader->key.mono.u.vs_export_prim_id);
 
@@ -1373,7 +1373,7 @@ void gfx10_emit_ngg_epilogue(struct ac_shader_abi *abi)
 
    LLVMValueRef user_edgeflags[3] = {};
 
-   if (sel->info.writes_edgeflag) {
+   if (gfx10_ngg_writes_user_edgeflags(ctx->shader)) {
       assert(!unterminated_es_if_block);
 
       /* Streamout already inserted the barrier, so don't insert it again. */
@@ -1401,7 +1401,7 @@ void gfx10_emit_ngg_epilogue(struct ac_shader_abi *abi)
       assert(!unterminated_es_if_block);
 
       /* Streamout and edge flags use LDS. Make it idle, so that we can reuse it. */
-      if (sel->so.num_outputs || sel->info.writes_edgeflag)
+      if (sel->so.num_outputs || gfx10_ngg_writes_user_edgeflags(ctx->shader))
          ac_build_s_barrier(&ctx->ac);
 
       ac_build_ifcc(&ctx->ac, is_gs_thread, 5400);

src/gallium/drivers/radeonsi/si_shader.h

@@ -960,6 +960,21 @@ static inline bool si_shader_uses_bindless_images(struct si_shader_selector *sel
    return selector ? selector->info.uses_bindless_images : false;
 }
 
+static inline bool gfx10_edgeflags_have_effect(struct si_shader *shader)
+{
+   if (shader->selector->info.stage == MESA_SHADER_VERTEX &&
+       !shader->selector->info.base.vs.blit_sgprs_amd)
+      return true;
+
+   return false;
+}
+
+static inline bool gfx10_ngg_writes_user_edgeflags(struct si_shader *shader)
+{
+   return gfx10_edgeflags_have_effect(shader) &&
+          shader->selector->info.writes_edgeflag;
+}
+
 #ifdef __cplusplus
 }
 #endif

src/gallium/drivers/radeonsi/si_state_shaders.c

@@ -1193,7 +1193,7 @@ static void gfx10_shader_ngg(struct si_screen *sscreen, struct si_shader *shader
     * for the GL_LINE polygon mode to skip rendering lines on inner edges.
     */
    if (gs_info->uses_invocationid ||
-       (gs_stage == MESA_SHADER_VERTEX && !gfx10_is_ngg_passthrough(shader)))
+       (gfx10_edgeflags_have_effect(shader) && !gfx10_is_ngg_passthrough(shader)))
       gs_vgpr_comp_cnt = 3; /* VGPR3 contains InvocationID, edge flags. */
    else if ((gs_stage == MESA_SHADER_GEOMETRY && gs_info->uses_primid) ||
             (gs_stage == MESA_SHADER_VERTEX && shader->key.mono.u.vs_export_prim_id))
@@ -1276,13 +1276,12 @@ static void gfx10_shader_ngg(struct si_screen *sscreen, struct si_shader *shader
       S_028B90_CNT(gs_num_invocations) | S_028B90_ENABLE(gs_num_invocations > 1) |
       S_028B90_EN_MAX_VERT_OUT_PER_GS_INSTANCE(shader->ngg.max_vert_out_per_gs_instance);
 
-   /* Always output hw-generated edge flags and pass them via the prim
+   /* Output hw-generated edge flags if needed and pass them via the prim
     * export to prevent drawing lines on internal edges of decomposed
-    * primitives (such as quads) with polygon mode = lines. Only VS needs
-    * this.
+    * primitives (such as quads) with polygon mode = lines.
     */
    shader->ctx_reg.ngg.pa_cl_ngg_cntl =
-      S_028838_INDEX_BUF_EDGE_FLAG_ENA(gs_stage == MESA_SHADER_VERTEX) |
+      S_028838_INDEX_BUF_EDGE_FLAG_ENA(gfx10_edgeflags_have_effect(shader)) |
       /* Reuse for NGG. */
       S_028838_VERTEX_REUSE_DEPTH(sscreen->info.chip_class >= GFX10_3 ? 30 : 0);
    shader->pa_cl_vs_out_cntl = si_get_vs_out_cntl(shader->selector, shader, true);