ac/llvm: skip s_barrier if tess patches don't cross a wave boundary

If tess patches are wholly in one wave, "s_waitcnt lgkm(0)" is sufficient.

Reviewed-by: Pierre-Eric Pelloux-Prayer <pierre-eric.pelloux-prayer@amd.com>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/16304>

src/amd/llvm/ac_nir_to_llvm.c

@@ -3928,6 +3928,11 @@ static void visit_intrinsic(struct ac_nir_context *ctx, nir_intrinsic_instr *ins
       break;
    }
    case nir_intrinsic_control_barrier:
+      /* If output patches are wholly in one wave, we don't need a barrier. */
+      if (ctx->stage == MESA_SHADER_TESS_CTRL &&
+          ctx->ac.wave_size % ctx->info->tess.tcs_vertices_out == 0)
+         break;
+
       ac_build_s_barrier(&ctx->ac, ctx->stage);
       break;
    case nir_intrinsic_shared_atomic_add:

src/gallium/drivers/radeonsi/si_shader.c

@@ -1971,6 +1971,14 @@ void si_get_tcs_epilog_key(struct si_shader *shader, union si_shader_part_key *k
    memset(key, 0, sizeof(*key));
    key->tcs_epilog.wave32 = shader->wave_size == 32;
    key->tcs_epilog.states = shader->key.ge.part.tcs.epilog;
+
+   /* If output patches are wholly in one wave, we don't need a barrier.
+    * The fixed-func TCS doesn't set tcs_vertices_out, but it won't use a barrier
+    * anyway because tess levels are always defined in all invocations there.
+    */
+   key->tcs_epilog.noop_s_barrier =
+      shader->selector->info.base.tess.tcs_vertices_out &&
+      shader->wave_size % shader->selector->info.base.tess.tcs_vertices_out == 0;
 }
 
 /**

src/gallium/drivers/radeonsi/si_shader.h

@@ -601,6 +601,7 @@ union si_shader_part_key {
    struct {
       struct si_tcs_epilog_bits states;
       unsigned wave32 : 1;
+      unsigned noop_s_barrier : 1;
    } tcs_epilog;
    struct {
       struct si_ps_prolog_bits states;

src/gallium/drivers/radeonsi/si_shader_llvm.c

@@ -1017,6 +1017,16 @@ bool si_llvm_translate_nir(struct si_shader_context *ctx, struct si_shader *shad
              shader->selector->info.base.inputs_read &
              ~shader->selector->info.tcs_vgpr_only_inputs) {
             ac_build_waitcnt(&ctx->ac, AC_WAIT_LGKM);
+
+            /* If both input and output patches are wholly in one wave, we don't need a barrier.
+             * That's true when both VS and TCS have the same number of patch vertices and
+             * the wave size is a multiple of the number of patch vertices.
+             *
+             * The fixed-func TCS doesn't set tcs_vertices_out.
+             */
+            if (!shader->key.ge.opt.same_patch_vertices ||
+                (sel->info.base.tess.tcs_vertices_out &&
+                 ctx->ac.wave_size % sel->info.base.tess.tcs_vertices_out != 0))
                ac_build_s_barrier(&ctx->ac, ctx->stage);
          }
       } else if (ctx->stage == MESA_SHADER_GEOMETRY && !shader->key.ge.as_ngg) {

src/gallium/drivers/radeonsi/si_shader_llvm_tess.c

@@ -671,8 +671,8 @@ static void si_copy_tcs_inputs(struct si_shader_context *ctx)
    }
 }
 
-static void si_write_tess_factors(struct si_shader_context *ctx, LLVMValueRef rel_patch_id,
-                                  LLVMValueRef invocation_id,
+static void si_write_tess_factors(struct si_shader_context *ctx, union si_shader_part_key *key,
+                                  LLVMValueRef rel_patch_id, LLVMValueRef invocation_id,
                                   LLVMValueRef tcs_out_current_patch_data_offset,
                                   LLVMValueRef invoc0_tf_outer[4], LLVMValueRef invoc0_tf_inner[2])
 {
@@ -685,6 +685,8 @@ static void si_write_tess_factors(struct si_shader_context *ctx, LLVMValueRef re
    /* Add a barrier before loading tess factors from LDS. */
    if (!shader->key.ge.part.tcs.epilog.invoc0_tess_factors_are_def) {
       ac_build_waitcnt(&ctx->ac, AC_WAIT_LGKM);
+
+      if (!key->tcs_epilog.noop_s_barrier)
          ac_build_s_barrier(&ctx->ac, ctx->stage);
    }
 
@@ -1075,7 +1077,7 @@ void si_llvm_build_tcs_epilog(struct si_shader_context *ctx, union si_shader_par
    for (unsigned i = 0; i < 6; i++)
       invoc0_tess_factors[i] = ac_get_arg(&ctx->ac, tess_factors[i]);
 
-   si_write_tess_factors(ctx, ac_get_arg(&ctx->ac, rel_patch_id),
+   si_write_tess_factors(ctx, key, ac_get_arg(&ctx->ac, rel_patch_id),
                          ac_get_arg(&ctx->ac, invocation_id),
                          ac_get_arg(&ctx->ac, tcs_out_current_patch_data_offset),
                          invoc0_tess_factors, invoc0_tess_factors + 4);