intel/cs: Rework the way thread local ID is handled
authorJason Ekstrand <jason.ekstrand@intel.com>
Sat, 30 Sep 2017 00:57:32 +0000 (17:57 -0700)
committerJason Ekstrand <jason.ekstrand@intel.com>
Tue, 7 Nov 2017 18:37:52 +0000 (10:37 -0800)
Previously, brw_nir_lower_intrinsics added the param and then emitted a
load_uniform intrinsic to load it directly.  This commit switches things
over to use a specific NIR intrinsic for the thread id.  The one thing I
don't like about this approach is that we have to copy thread_local_id
over to the new visitor in import_uniforms.

Reviewed-by: Iago Toral Quiroga <itoral@igalia.com>
src/compiler/nir/nir_intrinsics.h
src/intel/compiler/brw_fs.cpp
src/intel/compiler/brw_fs.h
src/intel/compiler/brw_fs_nir.cpp
src/intel/compiler/brw_nir.h
src/intel/compiler/brw_nir_lower_cs_intrinsics.c

index cefd18be904c44ffc4bf3d7ed6650e0235b5b8c3..47022dd135bf8a02dcf7ea191ab6985d2e89e5ca 100644 (file)
@@ -364,6 +364,9 @@ SYSTEM_VALUE(blend_const_color_a_float, 1, 0, xx, xx, xx)
 SYSTEM_VALUE(blend_const_color_rgba8888_unorm, 1, 0, xx, xx, xx)
 SYSTEM_VALUE(blend_const_color_aaaa8888_unorm, 1, 0, xx, xx, xx)
 
+/* Intel specific system values */
+SYSTEM_VALUE(intel_thread_local_id, 1, 0, xx, xx, xx)
+
 /**
  * Barycentric coordinate intrinsics.
  *
index 68a47bac84186ff768d3f4a6511d85cdd05c3d8d..c0b6047b9547f84d6c51044d4d9610a6603d91cb 100644 (file)
@@ -996,6 +996,7 @@ fs_visitor::import_uniforms(fs_visitor *v)
    this->push_constant_loc = v->push_constant_loc;
    this->pull_constant_loc = v->pull_constant_loc;
    this->uniforms = v->uniforms;
+   this->thread_local_id = v->thread_local_id;
 }
 
 void
@@ -6834,8 +6835,7 @@ brw_compile_cs(const struct brw_compiler *compiler, void *log_data,
 {
    nir_shader *shader = nir_shader_clone(mem_ctx, src_shader);
    shader = brw_nir_apply_sampler_key(shader, compiler, &key->tex, true);
-
-   brw_nir_lower_cs_intrinsics(shader, prog_data);
+   brw_nir_lower_cs_intrinsics(shader);
    shader = brw_postprocess_nir(shader, compiler, true);
 
    prog_data->local_size[0] = shader->info.cs.local_size[0];
index da3259323ec351a615a23337ce462e7ea65fb32c..f51a4d8889bd651b1ed2583782b52caaed5b078b 100644 (file)
@@ -315,6 +315,7 @@ public:
     */
    int *push_constant_loc;
 
+   fs_reg thread_local_id;
    fs_reg frag_depth;
    fs_reg frag_stencil;
    fs_reg sample_mask;
index 04b6e5119a20c7e02a94e0d083d5781c96eb9525..77d8bae4db60755bc9a944d8d436a7c0960b852b 100644 (file)
@@ -88,6 +88,16 @@ fs_visitor::nir_setup_uniforms()
    }
 
    uniforms = nir->num_uniforms / 4;
+
+   if (stage == MESA_SHADER_COMPUTE) {
+      /* Add a uniform for the thread local id.  It must be the last uniform
+       * on the list.
+       */
+      assert(uniforms == prog_data->nr_params);
+      uint32_t *param = brw_stage_prog_data_add_params(prog_data, 1);
+      *param = BRW_PARAM_BUILTIN_THREAD_LOCAL_ID;
+      thread_local_id = fs_reg(UNIFORM, uniforms++, BRW_REGISTER_TYPE_UD);
+   }
 }
 
 static bool
@@ -3412,6 +3422,10 @@ fs_visitor::nir_emit_cs_intrinsic(const fs_builder &bld,
       cs_prog_data->uses_barrier = true;
       break;
 
+   case nir_intrinsic_load_intel_thread_local_id:
+      bld.MOV(retype(dest, BRW_REGISTER_TYPE_UD), thread_local_id);
+      break;
+
    case nir_intrinsic_load_local_invocation_id:
    case nir_intrinsic_load_work_group_id: {
       gl_system_value sv = nir_system_value_from_intrinsic(instr->intrinsic);
index 1493b742e42e6e474b4eab8e8fa74967ac9054d4..3e407122681a69e8c8e0783528ec901d03370af4 100644 (file)
@@ -95,8 +95,7 @@ void brw_nir_analyze_boolean_resolves(nir_shader *nir);
 nir_shader *brw_preprocess_nir(const struct brw_compiler *compiler,
                                nir_shader *nir);
 
-bool brw_nir_lower_cs_intrinsics(nir_shader *nir,
-                                 struct brw_cs_prog_data *prog_data);
+bool brw_nir_lower_cs_intrinsics(nir_shader *nir);
 void brw_nir_lower_vs_inputs(nir_shader *nir,
                              bool use_legacy_snorm_formula,
                              const uint8_t *vs_attrib_wa_flags);
index d27727624c64a9ed251ad3b7ca742f21799f8432..07d2dccd0412e2a2184298055022fd068abf258c 100644 (file)
 
 struct lower_intrinsics_state {
    nir_shader *nir;
-   struct brw_cs_prog_data *prog_data;
    nir_function_impl *impl;
    bool progress;
    nir_builder builder;
-   int thread_local_id_index;
+   unsigned local_workgroup_size;
 };
 
-static nir_ssa_def *
-read_thread_local_id(struct lower_intrinsics_state *state)
-{
-   struct brw_cs_prog_data *prog_data = state->prog_data;
-   nir_builder *b = &state->builder;
-   nir_shader *nir = state->nir;
-   const unsigned *sizes = nir->info.cs.local_size;
-   const unsigned group_size = sizes[0] * sizes[1] * sizes[2];
-
-   /* Some programs have local_size dimensions so small that the thread local
-    * ID will always be 0.
-    */
-   if (group_size <= 8)
-      return nir_imm_int(b, 0);
-
-   if (state->thread_local_id_index == -1) {
-      state->thread_local_id_index = prog_data->base.nr_params;
-      uint32_t *param = brw_stage_prog_data_add_params(&prog_data->base, 1);
-      *param = BRW_PARAM_BUILTIN_THREAD_LOCAL_ID;
-      nir->num_uniforms += 4;
-   }
-   unsigned id_index = state->thread_local_id_index;
-
-   nir_intrinsic_instr *load =
-      nir_intrinsic_instr_create(nir, nir_intrinsic_load_uniform);
-   load->num_components = 1;
-   load->src[0] = nir_src_for_ssa(nir_imm_int(b, 0));
-   nir_ssa_dest_init(&load->instr, &load->dest, 1, 32, NULL);
-   nir_intrinsic_set_base(load, id_index * sizeof(uint32_t));
-   nir_intrinsic_set_range(load, sizeof(uint32_t));
-   nir_builder_instr_insert(b, &load->instr);
-   return &load->dest.ssa;
-}
-
 static bool
 lower_cs_intrinsics_convert_block(struct lower_intrinsics_state *state,
                                   nir_block *block)
@@ -91,7 +56,12 @@ lower_cs_intrinsics_convert_block(struct lower_intrinsics_state *state,
           *    gl_LocalInvocationIndex =
           *       cs_thread_local_id + subgroup_invocation;
           */
-         nir_ssa_def *thread_local_id = read_thread_local_id(state);
+         nir_ssa_def *thread_local_id;
+         if (state->local_workgroup_size <= 8)
+            thread_local_id = nir_imm_int(b, 0);
+         else
+            thread_local_id = nir_load_intel_thread_local_id(b);
+
          nir_ssa_def *channel = nir_load_subgroup_invocation(b);
          sysval = nir_iadd(b, channel, thread_local_id);
          break;
@@ -157,8 +127,7 @@ lower_cs_intrinsics_convert_impl(struct lower_intrinsics_state *state)
 }
 
 bool
-brw_nir_lower_cs_intrinsics(nir_shader *nir,
-                            struct brw_cs_prog_data *prog_data)
+brw_nir_lower_cs_intrinsics(nir_shader *nir)
 {
    assert(nir->info.stage == MESA_SHADER_COMPUTE);
 
@@ -166,9 +135,9 @@ brw_nir_lower_cs_intrinsics(nir_shader *nir,
    struct lower_intrinsics_state state;
    memset(&state, 0, sizeof(state));
    state.nir = nir;
-   state.prog_data = prog_data;
-
-   state.thread_local_id_index = -1;
+   state.local_workgroup_size = nir->info.cs.local_size[0] *
+                                nir->info.cs.local_size[1] *
+                                nir->info.cs.local_size[2];
 
    do {
       state.progress = false;