X-Git-Url: https://git.libre-soc.org/?a=blobdiff_plain;f=src%2Fcompiler%2Fnir%2Fnir_lower_io.c;h=df91febd68dd1f5fa7af78e98c7c34ff6959a182;hb=4925347ec5d469574f1280e972940cfea345ed6e;hp=d9af8bf3c7a5bbd507446f1f4784e222ef13ad0d;hpb=3124ce699bb3844e793f00e00bfbea5c91744f90;p=mesa.git

diff --git a/src/compiler/nir/nir_lower_io.c b/src/compiler/nir/nir_lower_io.c
index d9af8bf3c7a..df91febd68d 100644
--- a/src/compiler/nir/nir_lower_io.c
+++ b/src/compiler/nir/nir_lower_io.c
@@ -36,9 +36,9 @@
 
 struct lower_io_state {
    nir_builder builder;
-   void *mem_ctx;
    int (*type_size)(const struct glsl_type *type);
-   nir_variable_mode mode;
+   nir_variable_mode modes;
+   nir_lower_io_options options;
 };
 
 void
@@ -49,7 +49,7 @@ nir_assign_var_locations(struct exec_list *var_list, unsigned *size,
 
    nir_foreach_variable(var, var_list) {
       /*
-       * UBO's have their own address spaces, so don't count them towards the
+       * UBOs have their own address spaces, so don't count them towards the
        * number of global uniforms
        */
       if ((var->data.mode == nir_var_uniform || var->data.mode == nir_var_shader_storage) &&
@@ -64,32 +64,31 @@ nir_assign_var_locations(struct exec_list *var_list, unsigned *size,
 }
 
 /**
- * Returns true if we're processing a stage whose inputs are arrays indexed
- * by a vertex number (such as geometry shader inputs).
+ * Return true if the given variable is a per-vertex input/output array.
+ * (such as geometry shader inputs).
  */
-static bool
-is_per_vertex_input(struct lower_io_state *state, nir_variable *var)
+bool
+nir_is_per_vertex_io(const nir_variable *var, gl_shader_stage stage)
 {
-   gl_shader_stage stage = state->builder.shader->stage;
+   if (var->data.patch || !glsl_type_is_array(var->type))
+      return false;
 
-   return var->data.mode == nir_var_shader_in && !var->data.patch &&
-          (stage == MESA_SHADER_TESS_CTRL ||
-           stage == MESA_SHADER_TESS_EVAL ||
-           stage == MESA_SHADER_GEOMETRY);
-}
+   if (var->data.mode == nir_var_shader_in)
+      return stage == MESA_SHADER_GEOMETRY ||
+             stage == MESA_SHADER_TESS_CTRL ||
+             stage == MESA_SHADER_TESS_EVAL;
 
-static bool
-is_per_vertex_output(struct lower_io_state *state, nir_variable *var)
-{
-   gl_shader_stage stage = state->builder.shader->stage;
-   return var->data.mode == nir_var_shader_out && !var->data.patch &&
-          stage == MESA_SHADER_TESS_CTRL;
+   if (var->data.mode == nir_var_shader_out)
+      return stage == MESA_SHADER_TESS_CTRL;
+
+   return false;
 }
 
 static nir_ssa_def *
 get_io_offset(nir_builder *b, nir_deref_var *deref,
               nir_ssa_def **vertex_index,
-              int (*type_size)(const struct glsl_type *))
+              int (*type_size)(const struct glsl_type *),
+              unsigned *component)
 {
    nir_deref *tail = &deref->deref;
 
@@ -98,7 +97,6 @@ get_io_offset(nir_builder *b, nir_deref_var *deref,
     */
    if (vertex_index != NULL) {
       tail = tail->child;
-      assert(tail->deref_type == nir_deref_type_array);
       nir_deref_array *deref_array = nir_deref_as_array(tail);
 
       nir_ssa_def *vtx = nir_imm_int(b, deref_array->base_offset);
@@ -108,6 +106,19 @@ get_io_offset(nir_builder *b, nir_deref_var *deref,
       *vertex_index = vtx;
    }
 
+   if (deref->var->data.compact) {
+      assert(tail->child->deref_type == nir_deref_type_array);
+      assert(glsl_type_is_scalar(glsl_without_array(deref->var->type)));
+      nir_deref_array *deref_array = nir_deref_as_array(tail->child);
+      /* We always lower indirect dereferences for "compact" array vars. */
+      assert(deref_array->deref_array_type == nir_deref_array_type_direct);
+
+      const unsigned total_offset = *component + deref_array->base_offset;
+      const unsigned slot_offset = total_offset / 4;
+      *component = total_offset % 4;
+      return nir_imm_int(b, type_size(glsl_vec4_type()) * slot_offset);
+   }
+
    /* Just emit code and let constant-folding go to town */
    nir_ssa_def *offset = nir_imm_int(b, 0);
 
@@ -143,19 +154,44 @@ get_io_offset(nir_builder *b, nir_deref_var *deref,
    return offset;
 }
 
-static nir_intrinsic_op
-load_op(struct lower_io_state *state,
-        nir_variable_mode mode, bool per_vertex)
+static nir_intrinsic_instr *
+lower_load(nir_intrinsic_instr *intrin, struct lower_io_state *state,
+           nir_ssa_def *vertex_index, nir_ssa_def *offset,
+           unsigned component)
 {
+   const nir_shader *nir = state->builder.shader;
+   nir_variable *var = intrin->variables[0]->var;
+   nir_variable_mode mode = var->data.mode;
+   nir_ssa_def *barycentric = NULL;
+
    nir_intrinsic_op op;
    switch (mode) {
    case nir_var_shader_in:
-      op = per_vertex ? nir_intrinsic_load_per_vertex_input :
-                        nir_intrinsic_load_input;
+      if (nir->info.stage == MESA_SHADER_FRAGMENT &&
+          nir->options->use_interpolated_input_intrinsics &&
+          var->data.interpolation != INTERP_MODE_FLAT) {
+         assert(vertex_index == NULL);
+
+         nir_intrinsic_op bary_op;
+         if (var->data.sample ||
+             (state->options & nir_lower_io_force_sample_interpolation))
+            bary_op = nir_intrinsic_load_barycentric_sample;
+         else if (var->data.centroid)
+            bary_op = nir_intrinsic_load_barycentric_centroid;
+         else
+            bary_op = nir_intrinsic_load_barycentric_pixel;
+
+         barycentric = nir_load_barycentric(&state->builder, bary_op,
+                                            var->data.interpolation);
+         op = nir_intrinsic_load_interpolated_input;
+      } else {
+         op = vertex_index ? nir_intrinsic_load_per_vertex_input :
+                             nir_intrinsic_load_input;
+      }
       break;
    case nir_var_shader_out:
-      op = per_vertex ? nir_intrinsic_load_per_vertex_output :
-                        nir_intrinsic_load_output;
+      op = vertex_index ? nir_intrinsic_load_per_vertex_output :
+                          nir_intrinsic_load_output;
       break;
    case nir_var_uniform:
       op = nir_intrinsic_load_uniform;
@@ -166,34 +202,80 @@ load_op(struct lower_io_state *state,
    default:
       unreachable("Unknown variable mode");
    }
-   return op;
+
+   nir_intrinsic_instr *load =
+      nir_intrinsic_instr_create(state->builder.shader, op);
+   load->num_components = intrin->num_components;
+
+   nir_intrinsic_set_base(load, var->data.driver_location);
+   if (mode == nir_var_shader_in || mode == nir_var_shader_out)
+      nir_intrinsic_set_component(load, component);
+
+   if (load->intrinsic == nir_intrinsic_load_uniform)
+      nir_intrinsic_set_range(load, state->type_size(var->type));
+
+   if (vertex_index) {
+      load->src[0] = nir_src_for_ssa(vertex_index);
+      load->src[1] = nir_src_for_ssa(offset);
+   } else if (barycentric) {
+      load->src[0] = nir_src_for_ssa(barycentric);
+      load->src[1] = nir_src_for_ssa(offset);
+   } else {
+      load->src[0] = nir_src_for_ssa(offset);
+   }
+
+   return load;
 }
 
-static nir_intrinsic_op
-store_op(struct lower_io_state *state,
-         nir_variable_mode mode, bool per_vertex)
+static nir_intrinsic_instr *
+lower_store(nir_intrinsic_instr *intrin, struct lower_io_state *state,
+            nir_ssa_def *vertex_index, nir_ssa_def *offset,
+            unsigned component)
 {
+   nir_variable *var = intrin->variables[0]->var;
+   nir_variable_mode mode = var->data.mode;
+
    nir_intrinsic_op op;
-   switch (mode) {
-   case nir_var_shader_in:
-   case nir_var_shader_out:
-      op = per_vertex ? nir_intrinsic_store_per_vertex_output :
-                        nir_intrinsic_store_output;
-      break;
-   case nir_var_shared:
+   if (mode == nir_var_shared) {
       op = nir_intrinsic_store_shared;
-      break;
-   default:
-      unreachable("Unknown variable mode");
+   } else {
+      assert(mode == nir_var_shader_out);
+      op = vertex_index ? nir_intrinsic_store_per_vertex_output :
+                          nir_intrinsic_store_output;
    }
-   return op;
+
+   nir_intrinsic_instr *store =
+      nir_intrinsic_instr_create(state->builder.shader, op);
+   store->num_components = intrin->num_components;
+
+   nir_src_copy(&store->src[0], &intrin->src[0], store);
+
+   nir_intrinsic_set_base(store, var->data.driver_location);
+
+   if (mode == nir_var_shader_out)
+      nir_intrinsic_set_component(store, component);
+
+   nir_intrinsic_set_write_mask(store, nir_intrinsic_write_mask(intrin));
+
+   if (vertex_index)
+      store->src[1] = nir_src_for_ssa(vertex_index);
+
+   store->src[vertex_index ? 2 : 1] = nir_src_for_ssa(offset);
+
+   return store;
 }
 
-static nir_intrinsic_op
-atomic_op(nir_intrinsic_op opcode)
+static nir_intrinsic_instr *
+lower_atomic(nir_intrinsic_instr *intrin, struct lower_io_state *state,
+             nir_ssa_def *offset)
 {
-   switch (opcode) {
-#define OP(O) case nir_intrinsic_var_##O: return nir_intrinsic_shared_##O;
+   nir_variable *var = intrin->variables[0]->var;
+
+   assert(var->data.mode == nir_var_shared);
+
+   nir_intrinsic_op op;
+   switch (intrin->intrinsic) {
+#define OP(O) case nir_intrinsic_var_##O: op = nir_intrinsic_shared_##O; break;
    OP(atomic_exchange)
    OP(atomic_comp_swap)
    OP(atomic_add)
@@ -208,16 +290,83 @@ atomic_op(nir_intrinsic_op opcode)
    default:
       unreachable("Invalid atomic");
    }
+
+   nir_intrinsic_instr *atomic =
+      nir_intrinsic_instr_create(state->builder.shader, op);
+
+   nir_intrinsic_set_base(atomic, var->data.driver_location);
+
+   atomic->src[0] = nir_src_for_ssa(offset);
+   for (unsigned i = 0; i < nir_intrinsic_infos[intrin->intrinsic].num_srcs; i++) {
+      nir_src_copy(&atomic->src[i+1], &intrin->src[i], atomic);
+   }
+
+   return atomic;
 }
 
-static bool
-nir_lower_io_block(nir_block *block, void *void_state)
+static nir_intrinsic_instr *
+lower_interpolate_at(nir_intrinsic_instr *intrin, struct lower_io_state *state,
+                     nir_ssa_def *offset, unsigned component)
 {
-   struct lower_io_state *state = void_state;
+   nir_variable *var = intrin->variables[0]->var;
+
+   assert(var->data.mode == nir_var_shader_in);
+
+   /* Ignore interpolateAt() for flat variables - flat is flat. */
+   if (var->data.interpolation == INTERP_MODE_FLAT)
+      return lower_load(intrin, state, NULL, offset, component);
+
+   nir_intrinsic_op bary_op;
+   switch (intrin->intrinsic) {
+   case nir_intrinsic_interp_var_at_centroid:
+      bary_op = (state->options & nir_lower_io_force_sample_interpolation) ?
+                nir_intrinsic_load_barycentric_sample :
+                nir_intrinsic_load_barycentric_centroid;
+      break;
+   case nir_intrinsic_interp_var_at_sample:
+      bary_op = nir_intrinsic_load_barycentric_at_sample;
+      break;
+   case nir_intrinsic_interp_var_at_offset:
+      bary_op = nir_intrinsic_load_barycentric_at_offset;
+      break;
+   default:
+      unreachable("Bogus interpolateAt() intrinsic.");
+   }
+
+   nir_intrinsic_instr *bary_setup =
+      nir_intrinsic_instr_create(state->builder.shader, bary_op);
+
+   nir_ssa_dest_init(&bary_setup->instr, &bary_setup->dest, 2, 32, NULL);
+   nir_intrinsic_set_interp_mode(bary_setup, var->data.interpolation);
+
+   if (intrin->intrinsic != nir_intrinsic_interp_var_at_centroid)
+      nir_src_copy(&bary_setup->src[0], &intrin->src[0], bary_setup);
+
+   nir_builder_instr_insert(&state->builder, &bary_setup->instr);
+
+   nir_intrinsic_instr *load =
+      nir_intrinsic_instr_create(state->builder.shader,
+                                 nir_intrinsic_load_interpolated_input);
+   load->num_components = intrin->num_components;
+
+   nir_intrinsic_set_base(load, var->data.driver_location);
+   nir_intrinsic_set_component(load, component);
 
+   load->src[0] = nir_src_for_ssa(&bary_setup->dest.ssa);
+   load->src[1] = nir_src_for_ssa(offset);
+
+   return load;
+}
+
+static bool
+nir_lower_io_block(nir_block *block,
+                   struct lower_io_state *state)
+{
    nir_builder *b = &state->builder;
+   const nir_shader_compiler_options *options = b->shader->options;
+   bool progress = false;
 
-   nir_foreach_instr_safe(block, instr) {
+   nir_foreach_instr_safe(instr, block) {
       if (instr->type != nir_instr_type_intrinsic)
          continue;
 
@@ -238,14 +387,21 @@ nir_lower_io_block(nir_block *block, void *void_state)
       case nir_intrinsic_var_atomic_comp_swap:
          /* We can lower the io for this nir instrinsic */
          break;
+      case nir_intrinsic_interp_var_at_centroid:
+      case nir_intrinsic_interp_var_at_sample:
+      case nir_intrinsic_interp_var_at_offset:
+         /* We can optionally lower these to load_interpolated_input */
+         if (options->use_interpolated_input_intrinsics)
+            break;
       default:
          /* We can't lower the io for this nir instrinsic, so skip it */
          continue;
       }
 
-      nir_variable_mode mode = intrin->variables[0]->var->data.mode;
+      nir_variable *var = intrin->variables[0]->var;
+      nir_variable_mode mode = var->data.mode;
 
-      if (state->mode != nir_var_all && state->mode != mode)
+      if ((state->modes & mode) == 0)
          continue;
 
       if (mode != nir_var_shader_in &&
@@ -256,80 +412,28 @@ nir_lower_io_block(nir_block *block, void *void_state)
 
       b->cursor = nir_before_instr(instr);
 
-      switch (intrin->intrinsic) {
-      case nir_intrinsic_load_var: {
-         bool per_vertex =
-            is_per_vertex_input(state, intrin->variables[0]->var) ||
-            is_per_vertex_output(state, intrin->variables[0]->var);
-
-         nir_ssa_def *offset;
-         nir_ssa_def *vertex_index;
+      const bool per_vertex = nir_is_per_vertex_io(var, b->shader->info.stage);
 
-         offset = get_io_offset(b, intrin->variables[0],
-                                per_vertex ? &vertex_index : NULL,
-                                state->type_size);
+      nir_ssa_def *offset;
+      nir_ssa_def *vertex_index = NULL;
+      unsigned component_offset = var->data.location_frac;
 
-         nir_intrinsic_instr *load =
-            nir_intrinsic_instr_create(state->mem_ctx,
-                                       load_op(state, mode, per_vertex));
-         load->num_components = intrin->num_components;
+      offset = get_io_offset(b, intrin->variables[0],
+                             per_vertex ? &vertex_index : NULL,
+                             state->type_size, &component_offset);
 
-         nir_intrinsic_set_base(load,
-            intrin->variables[0]->var->data.driver_location);
-
-         if (per_vertex)
-            load->src[0] = nir_src_for_ssa(vertex_index);
-
-         load->src[per_vertex ? 1 : 0] = nir_src_for_ssa(offset);
-
-         if (intrin->dest.is_ssa) {
-            nir_ssa_dest_init(&load->instr, &load->dest,
-                              intrin->num_components,
-                              intrin->dest.ssa.bit_size, NULL);
-            nir_ssa_def_rewrite_uses(&intrin->dest.ssa,
-                                     nir_src_for_ssa(&load->dest.ssa));
-         } else {
-            nir_dest_copy(&load->dest, &intrin->dest, state->mem_ctx);
-         }
+      nir_intrinsic_instr *replacement;
 
-         nir_instr_insert_before(&intrin->instr, &load->instr);
-         nir_instr_remove(&intrin->instr);
+      switch (intrin->intrinsic) {
+      case nir_intrinsic_load_var:
+         replacement = lower_load(intrin, state, vertex_index, offset,
+                                  component_offset);
          break;
-      }
-
-      case nir_intrinsic_store_var: {
-         assert(mode == nir_var_shader_out || mode == nir_var_shared);
-
-         nir_ssa_def *offset;
-         nir_ssa_def *vertex_index;
-
-         bool per_vertex =
-            is_per_vertex_output(state, intrin->variables[0]->var);
-
-         offset = get_io_offset(b, intrin->variables[0],
-                                per_vertex ? &vertex_index : NULL,
-                                state->type_size);
-
-         nir_intrinsic_instr *store =
-            nir_intrinsic_instr_create(state->mem_ctx,
-                                       store_op(state, mode, per_vertex));
-         store->num_components = intrin->num_components;
-
-         nir_src_copy(&store->src[0], &intrin->src[0], store);
-
-         nir_intrinsic_set_base(store,
-            intrin->variables[0]->var->data.driver_location);
-         nir_intrinsic_set_write_mask(store, nir_intrinsic_write_mask(intrin));
-
-         if (per_vertex)
-            store->src[1] = nir_src_for_ssa(vertex_index);
 
-         store->src[per_vertex ? 2 : 1] = nir_src_for_ssa(offset);
-
-         nir_instr_insert_before(&intrin->instr, &store->instr);
-         nir_instr_remove(&intrin->instr);
+      case nir_intrinsic_store_var:
+         replacement = lower_store(intrin, state, vertex_index, offset,
+                                   component_offset);
          break;
-      }
 
       case nir_intrinsic_var_atomic_add:
       case nir_intrinsic_var_atomic_imin:
@@ -340,82 +444,85 @@ nir_lower_io_block(nir_block *block, void *void_state)
       case nir_intrinsic_var_atomic_or:
       case nir_intrinsic_var_atomic_xor:
       case nir_intrinsic_var_atomic_exchange:
-      case nir_intrinsic_var_atomic_comp_swap: {
-         assert(mode == nir_var_shared);
-
-         nir_ssa_def *offset;
-
-         offset = get_io_offset(b, intrin->variables[0],
-                                NULL, state->type_size);
-
-         nir_intrinsic_instr *atomic =
-            nir_intrinsic_instr_create(state->mem_ctx,
-                                       atomic_op(intrin->intrinsic));
-
-         atomic->src[0] = nir_src_for_ssa(offset);
+      case nir_intrinsic_var_atomic_comp_swap:
+         assert(vertex_index == NULL);
+         replacement = lower_atomic(intrin, state, offset);
+         break;
 
-         atomic->const_index[0] =
-            intrin->variables[0]->var->data.driver_location;
+      case nir_intrinsic_interp_var_at_centroid:
+      case nir_intrinsic_interp_var_at_sample:
+      case nir_intrinsic_interp_var_at_offset:
+         assert(vertex_index == NULL);
+         replacement = lower_interpolate_at(intrin, state, offset,
+                                            component_offset);
+         break;
 
-         for (unsigned i = 0;
-              i < nir_op_infos[intrin->intrinsic].num_inputs;
-              i++) {
-            nir_src_copy(&atomic->src[i+1], &intrin->src[i], atomic);
-         }
+      default:
+         continue;
+      }
 
+      if (nir_intrinsic_infos[intrin->intrinsic].has_dest) {
          if (intrin->dest.is_ssa) {
-            nir_ssa_dest_init(&atomic->instr, &atomic->dest,
+            nir_ssa_dest_init(&replacement->instr, &replacement->dest,
                               intrin->dest.ssa.num_components,
                               intrin->dest.ssa.bit_size, NULL);
             nir_ssa_def_rewrite_uses(&intrin->dest.ssa,
-                                     nir_src_for_ssa(&atomic->dest.ssa));
+                                     nir_src_for_ssa(&replacement->dest.ssa));
          } else {
-            nir_dest_copy(&atomic->dest, &intrin->dest, state->mem_ctx);
+            nir_dest_copy(&replacement->dest, &intrin->dest, &intrin->instr);
          }
-
-         nir_instr_insert_before(&intrin->instr, &atomic->instr);
-         nir_instr_remove(&intrin->instr);
-         break;
       }
 
-      default:
-         break;
-      }
+      nir_instr_insert_before(&intrin->instr, &replacement->instr);
+      nir_instr_remove(&intrin->instr);
+      progress = true;
    }
 
-   return true;
+   return progress;
 }
 
-static void
+static bool
 nir_lower_io_impl(nir_function_impl *impl,
-                  nir_variable_mode mode,
-                  int (*type_size)(const struct glsl_type *))
+                  nir_variable_mode modes,
+                  int (*type_size)(const struct glsl_type *),
+                  nir_lower_io_options options)
 {
    struct lower_io_state state;
+   bool progress = false;
 
    nir_builder_init(&state.builder, impl);
-   state.mem_ctx = ralloc_parent(impl);
-   state.mode = mode;
+   state.modes = modes;
    state.type_size = type_size;
+   state.options = options;
 
-   nir_foreach_block(impl, nir_lower_io_block, &state);
+   nir_foreach_block(block, impl) {
+      progress |= nir_lower_io_block(block, &state);
+   }
 
    nir_metadata_preserve(impl, nir_metadata_block_index |
                                nir_metadata_dominance);
+   return progress;
 }
 
-void
-nir_lower_io(nir_shader *shader, nir_variable_mode mode,
-             int (*type_size)(const struct glsl_type *))
+bool
+nir_lower_io(nir_shader *shader, nir_variable_mode modes,
+             int (*type_size)(const struct glsl_type *),
+             nir_lower_io_options options)
 {
-   nir_foreach_function(shader, function) {
-      if (function->impl)
-         nir_lower_io_impl(function->impl, mode, type_size);
+   bool progress = false;
+
+   nir_foreach_function(function, shader) {
+      if (function->impl) {
+         progress |= nir_lower_io_impl(function->impl, modes,
+                                       type_size, options);
+      }
    }
+
+   return progress;
 }
 
 /**
- * Return the offset soruce for a load/store intrinsic.
+ * Return the offset source for a load/store intrinsic.
  */
 nir_src *
 nir_get_io_offset_src(nir_intrinsic_instr *instr)
@@ -425,10 +532,14 @@ nir_get_io_offset_src(nir_intrinsic_instr *instr)
    case nir_intrinsic_load_output:
    case nir_intrinsic_load_uniform:
       return &instr->src[0];
+   case nir_intrinsic_load_ubo:
+   case nir_intrinsic_load_ssbo:
    case nir_intrinsic_load_per_vertex_input:
    case nir_intrinsic_load_per_vertex_output:
+   case nir_intrinsic_load_interpolated_input:
    case nir_intrinsic_store_output:
       return &instr->src[1];
+   case nir_intrinsic_store_ssbo:
    case nir_intrinsic_store_per_vertex_output:
       return &instr->src[2];
    default: