aco: Skip 2nd read of merged wave info when TCS in/out vertices are equal.

When TCS has an equal number of input and output, it means that the
number of VS and TCS invocations (LS and HS) are the same; and that
the HS invocations operate on the same vertices as the LS.

When this is the case, this commit removes the else-if between
the merged VS and TCS halves, making it possible to schedule
and optimize the code accross the two halves.

Totals:
SGPRS: 5577367 -> 5581735 (0.08 %)
VGPRS: 3958592 -> 3960752 (0.05 %)
Code Size: 254867144 -> 254838244 (-0.01 %) bytes
Max Waves: 1053887 -> 1053747 (-0.01 %)

Totals from affected shaders:
SGPRS: 29032 -> 33400 (15.05 %)
VGPRS: 35664 -> 37824 (6.06 %)
Code Size: 1979028 -> 1950128 (-1.46 %) bytes
Max Waves: 7310 -> 7170 (-1.92 %)

Signed-off-by: Timur Kristóf <timur.kristof@gmail.com>
Reviewed-by: Rhys Perry <pendingchaos02@gmail.com>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/4165>

diff --git a/src/amd/compiler/aco_instruction_selection.cpp b/src/amd/compiler/aco_instruction_selection.cpp
index 340c6383646b6653844a81250a4f5502ef868952..b6add9ae7f3eb11589eb0c10c04170e767f70d9e 100644 (file)
--- a/src/amd/compiler/aco_instruction_selection.cpp
+++ b/src/amd/compiler/aco_instruction_selection.cpp
@@ -9752,6 +9752,7 @@ void select_program(Program *program,
                     struct radv_shader_args *args)
 {
    isel_context ctx = setup_isel_context(program, shader_count, shaders, config, args, false);
+   if_context ic_merged_wave_info;
 
    for (unsigned i = 0; i < shader_count; i++) {
       nir_shader *nir = shaders[i];
@@ -9778,14 +9779,15 @@ void select_program(Program *program,
                            (nir->info.stage == MESA_SHADER_TESS_EVAL &&
                             ctx.stage == tess_eval_geometry_gs));
 
-      if_context ic;
-      if (shader_count >= 2 && !empty_shader) {
+      bool check_merged_wave_info = ctx.tcs_in_out_eq ? i == 0 : (shader_count >= 2 && !empty_shader);
+      bool endif_merged_wave_info = ctx.tcs_in_out_eq ? i == 1 : check_merged_wave_info;
+      if (check_merged_wave_info) {
          Builder bld(ctx.program, ctx.block);
          Temp count = bld.sop2(aco_opcode::s_bfe_u32, bld.def(s1), bld.def(s1, scc), get_arg(&ctx, args->merged_wave_info), Operand((8u << 16) | (i * 8u)));
          Temp thread_id = emit_mbcnt(&ctx, bld.def(v1));
          Temp cond = bld.vopc(aco_opcode::v_cmp_gt_u32, bld.hint_vcc(bld.def(bld.lm)), count, thread_id);
 
-         begin_divergent_if_then(&ctx, &ic, cond);
+         begin_divergent_if_then(&ctx, &ic_merged_wave_info, cond);
       }
 
       if (i) {
@@ -9821,9 +9823,9 @@ void select_program(Program *program,
       if (ctx.stage == fragment_fs)
          create_fs_exports(&ctx);
 
-      if (shader_count >= 2 && !empty_shader) {
-         begin_divergent_if_else(&ctx, &ic);
-         end_divergent_if(&ctx, &ic);
+      if (endif_merged_wave_info) {
+         begin_divergent_if_else(&ctx, &ic_merged_wave_info);
+         end_divergent_if(&ctx, &ic_merged_wave_info);
       }
 
       ralloc_free(ctx.divergent_vals);

diff --git a/src/amd/compiler/aco_instruction_selection_setup.cpp b/src/amd/compiler/aco_instruction_selection_setup.cpp
index ff32aa3dc84d0bc7da1258a7d413cfdca85da61e..4ff1de72113401bedd1127473a46b1bdd97bc0cf 100644 (file)
--- a/src/amd/compiler/aco_instruction_selection_setup.cpp
+++ b/src/amd/compiler/aco_instruction_selection_setup.cpp
@@ -100,6 +100,7 @@ struct isel_context {
    unsigned tcs_tess_lvl_in_loc;
    uint32_t tcs_num_inputs;
    uint32_t tcs_num_patches;
+   bool tcs_in_out_eq = false;
 
    /* VS, FS or GS output information */
    output_state outputs;
@@ -899,6 +900,15 @@ setup_tcs_variables(isel_context *ctx, nir_shader *nir)
       unreachable("Unsupported TCS shader stage");
    }
 
+   /* When the number of TCS input and output vertices are the same (typically 3):
+    * - There is an equal amount of LS and HS invocations
+    * - In case of merged LSHS shaders, the LS and HS halves of the shader
+    *   always process the exact same vertex. We can use this knowledge to optimize them.
+    */
+   ctx->tcs_in_out_eq =
+      ctx->stage == vertex_tess_control_hs &&
+      ctx->args->options->key.tcs.input_vertices == nir->info.tess.tcs_vertices_out;
+
    ctx->tcs_num_patches = get_tcs_num_patches(
                              ctx->args->options->key.tcs.input_vertices,
                              nir->info.tess.tcs_vertices_out,