radeonsi: use si_nir_clamp_vertex_color

Clamp vertex color in nir. Now only GS copy shader use
si_vertex_color_clamping, so move it there. It will be
completely removed after we switch to nir GS copy shader.

Acked-by: Pierre-Eric Pelloux-Prayer <pierre-eric.pelloux-prayer@amd.com>
Reviewed-by: Marek Olšák <marek.olsak@amd.com>
Signed-off-by: Qiang Yu <yuq825@gmail.com>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/19429>

diff --git a/src/gallium/drivers/radeonsi/si_shader.c b/src/gallium/drivers/radeonsi/si_shader.c
index 513c101..f6b4f84 100644 (file)
--- a/src/gallium/drivers/radeonsi/si_shader.c
+++ b/src/gallium/drivers/radeonsi/si_shader.c
@@ -1726,6 +1726,15 @@ struct nir_shader *si_get_nir_shader(struct si_shader *shader, bool *free_nir,
    if (sel->stage == MESA_SHADER_FRAGMENT && key->ps.mono.point_smoothing)
       NIR_PASS(progress, nir, nir_lower_point_smooth);
 
+   bool is_last_vgt_stage =
+      (sel->stage == MESA_SHADER_VERTEX ||
+       sel->stage == MESA_SHADER_TESS_EVAL ||
+       (sel->stage == MESA_SHADER_GEOMETRY && shader->key.ge.as_ngg)) &&
+      !shader->key.ge.as_ls && !shader->key.ge.as_es;
+
+   if (is_last_vgt_stage)
+      NIR_PASS(progress, nir, si_nir_clamp_vertex_color);
+
    if (progress)
       si_nir_opts(sel->screen, nir, true);
 

diff --git a/src/gallium/drivers/radeonsi/si_shader_llvm_gs.c b/src/gallium/drivers/radeonsi/si_shader_llvm_gs.c
index 0daea37..3632671 100644 (file)
--- a/src/gallium/drivers/radeonsi/si_shader_llvm_gs.c
+++ b/src/gallium/drivers/radeonsi/si_shader_llvm_gs.c
@@ -375,6 +375,71 @@ void si_preload_gs_rings(struct si_shader_context *ctx)
    }
 }
 
+/**
+ * Vertex color clamping.
+ *
+ * This uses a state constant loaded in a user data SGPR and
+ * an IF statement is added that clamps all colors if the constant
+ * is true.
+ */
+static void si_vertex_color_clamping(struct si_shader_context *ctx,
+                                     struct si_shader_output_values *outputs, unsigned noutput)
+{
+   LLVMValueRef addr[SI_MAX_VS_OUTPUTS][4];
+   bool has_colors = false;
+
+   /* Store original colors to alloca variables. */
+   for (unsigned i = 0; i < noutput; i++) {
+      if (outputs[i].semantic != VARYING_SLOT_COL0 &&
+          outputs[i].semantic != VARYING_SLOT_COL1 &&
+          outputs[i].semantic != VARYING_SLOT_BFC0 &&
+          outputs[i].semantic != VARYING_SLOT_BFC1)
+         continue;
+
+      for (unsigned j = 0; j < 4; j++)
+         addr[i][j] = ac_build_alloca_init(&ctx->ac, outputs[i].values[j], "");
+
+      has_colors = true;
+   }
+
+   if (!has_colors)
+      return;
+
+   /* The state is in the first bit of the user SGPR. */
+   LLVMValueRef cond = GET_FIELD(ctx, VS_STATE_CLAMP_VERTEX_COLOR);
+   cond = LLVMBuildTrunc(ctx->ac.builder, cond, ctx->ac.i1, "");
+
+   ac_build_ifcc(&ctx->ac, cond, 6502);
+
+   /* Store clamped colors to alloca variables within the conditional block. */
+   for (unsigned i = 0; i < noutput; i++) {
+      if (outputs[i].semantic != VARYING_SLOT_COL0 &&
+          outputs[i].semantic != VARYING_SLOT_COL1 &&
+          outputs[i].semantic != VARYING_SLOT_BFC0 &&
+          outputs[i].semantic != VARYING_SLOT_BFC1)
+         continue;
+
+      for (unsigned j = 0; j < 4; j++) {
+         LLVMBuildStore(ctx->ac.builder, ac_build_clamp(&ctx->ac, outputs[i].values[j]),
+                        addr[i][j]);
+      }
+   }
+   ac_build_endif(&ctx->ac, 6502);
+
+   /* Load clamped colors */
+   for (unsigned i = 0; i < noutput; i++) {
+      if (outputs[i].semantic != VARYING_SLOT_COL0 &&
+          outputs[i].semantic != VARYING_SLOT_COL1 &&
+          outputs[i].semantic != VARYING_SLOT_BFC0 &&
+          outputs[i].semantic != VARYING_SLOT_BFC1)
+         continue;
+
+      for (unsigned j = 0; j < 4; j++) {
+         outputs[i].values[j] = LLVMBuildLoad2(ctx->ac.builder, ctx->ac.f32, addr[i][j], "");
+      }
+   }
+}
+
 /* Generate code for the hardware VS shader stage to go with a geometry shader */
 struct si_shader *si_generate_gs_copy_shader(struct si_screen *sscreen,
                                              struct ac_llvm_compiler *compiler,
@@ -496,8 +561,10 @@ struct si_shader *si_generate_gs_copy_shader(struct si_screen *sscreen,
          si_llvm_emit_streamout(&ctx, outputs, gsinfo->num_outputs, stream);
       }
 
-      if (stream == 0)
+      if (stream == 0) {
+         si_vertex_color_clamping(&ctx, outputs, gsinfo->num_outputs);
          si_llvm_build_vs_exports(&ctx, NULL, outputs, gsinfo->num_outputs);
+      }
 
       LLVMBuildBr(builder, end_bb);
    }

diff --git a/src/gallium/drivers/radeonsi/si_shader_llvm_vs.c b/src/gallium/drivers/radeonsi/si_shader_llvm_vs.c
index 950daf4..7e0ce7c 100644 (file)
--- a/src/gallium/drivers/radeonsi/si_shader_llvm_vs.c
+++ b/src/gallium/drivers/radeonsi/si_shader_llvm_vs.c
@@ -501,71 +501,6 @@ static void si_llvm_init_vs_export_args(struct si_shader_context *ctx, const LLV
 }
 
 /**
- * Vertex color clamping.
- *
- * This uses a state constant loaded in a user data SGPR and
- * an IF statement is added that clamps all colors if the constant
- * is true.
- */
-static void si_vertex_color_clamping(struct si_shader_context *ctx,
-                                     struct si_shader_output_values *outputs, unsigned noutput)
-{
-   LLVMValueRef addr[SI_MAX_VS_OUTPUTS][4];
-   bool has_colors = false;
-
-   /* Store original colors to alloca variables. */
-   for (unsigned i = 0; i < noutput; i++) {
-      if (outputs[i].semantic != VARYING_SLOT_COL0 &&
-          outputs[i].semantic != VARYING_SLOT_COL1 &&
-          outputs[i].semantic != VARYING_SLOT_BFC0 &&
-          outputs[i].semantic != VARYING_SLOT_BFC1)
-         continue;
-
-      for (unsigned j = 0; j < 4; j++)
-         addr[i][j] = ac_build_alloca_init(&ctx->ac, outputs[i].values[j], "");
-
-      has_colors = true;
-   }
-
-   if (!has_colors)
-      return;
-
-   /* The state is in the first bit of the user SGPR. */
-   LLVMValueRef cond = GET_FIELD(ctx, VS_STATE_CLAMP_VERTEX_COLOR);
-   cond = LLVMBuildTrunc(ctx->ac.builder, cond, ctx->ac.i1, "");
-
-   ac_build_ifcc(&ctx->ac, cond, 6502);
-
-   /* Store clamped colors to alloca variables within the conditional block. */
-   for (unsigned i = 0; i < noutput; i++) {
-      if (outputs[i].semantic != VARYING_SLOT_COL0 &&
-          outputs[i].semantic != VARYING_SLOT_COL1 &&
-          outputs[i].semantic != VARYING_SLOT_BFC0 &&
-          outputs[i].semantic != VARYING_SLOT_BFC1)
-         continue;
-
-      for (unsigned j = 0; j < 4; j++) {
-         LLVMBuildStore(ctx->ac.builder, ac_build_clamp(&ctx->ac, outputs[i].values[j]),
-                        addr[i][j]);
-      }
-   }
-   ac_build_endif(&ctx->ac, 6502);
-
-   /* Load clamped colors */
-   for (unsigned i = 0; i < noutput; i++) {
-      if (outputs[i].semantic != VARYING_SLOT_COL0 &&
-          outputs[i].semantic != VARYING_SLOT_COL1 &&
-          outputs[i].semantic != VARYING_SLOT_BFC0 &&
-          outputs[i].semantic != VARYING_SLOT_BFC1)
-         continue;
-
-      for (unsigned j = 0; j < 4; j++) {
-         outputs[i].values[j] = LLVMBuildLoad2(ctx->ac.builder, ctx->ac.f32, addr[i][j], "");
-      }
-   }
-}
-
-/**
  * Generate export instructions for hardware VS shader stage or NGG GS stage
  * (position and parameter data only).
  *
@@ -584,8 +519,6 @@ void si_llvm_build_vs_exports(struct si_shader_context *ctx, LLVMValueRef num_ex
                             shader->selector->info.culldist_mask;
    int i;
 
-   si_vertex_color_clamping(ctx, outputs, noutput);
-
    /* Build position exports. */
    for (i = 0; i < noutput; i++) {
       switch (outputs[i].semantic) {