X-Git-Url: https://git.libre-soc.org/?a=blobdiff_plain;f=src%2Fglsl%2Fnir%2Fnir_lower_io.c;h=ec6d09d5b6d7f7eb295bfe92e5800a2629b0d94b;hb=d5c9955d3eaa7311e2b2350b6964bae516c7b7b2;hp=ddbc249371b11eec5d350a2b8c3463627e7eb12d;hpb=27c6e3e4ca665c528b94937b1b0a7c10de515d89;p=mesa.git

diff --git a/src/glsl/nir/nir_lower_io.c b/src/glsl/nir/nir_lower_io.c
index ddbc249371b..ec6d09d5b6d 100644
--- a/src/glsl/nir/nir_lower_io.c
+++ b/src/glsl/nir/nir_lower_io.c
@@ -29,66 +29,31 @@
 /*
  * This lowering pass converts references to input/output variables with
  * loads/stores to actual input/output intrinsics.
- *
- * NOTE: This pass really only works for scalar backends at the moment due
- * to the way it packes the input/output data.
  */
 
 #include "nir.h"
+#include "nir_builder.h"
 
 struct lower_io_state {
+   nir_builder builder;
    void *mem_ctx;
+   int (*type_size)(const struct glsl_type *type);
+   nir_variable_mode mode;
 };
 
-static unsigned
-type_size(const struct glsl_type *type)
-{
-   unsigned int size, i;
-
-   switch (glsl_get_base_type(type)) {
-   case GLSL_TYPE_UINT:
-   case GLSL_TYPE_INT:
-   case GLSL_TYPE_FLOAT:
-   case GLSL_TYPE_BOOL:
-      return glsl_get_components(type);
-   case GLSL_TYPE_ARRAY:
-      return type_size(glsl_get_array_element(type)) * glsl_get_length(type);
-   case GLSL_TYPE_STRUCT:
-      size = 0;
-      for (i = 0; i < glsl_get_length(type); i++) {
-         size += type_size(glsl_get_struct_field(type, i));
-      }
-      return size;
-   case GLSL_TYPE_SAMPLER:
-      return 0;
-   case GLSL_TYPE_ATOMIC_UINT:
-      return 0;
-   case GLSL_TYPE_INTERFACE:
-      return 0;
-   case GLSL_TYPE_IMAGE:
-      return 0;
-   case GLSL_TYPE_VOID:
-   case GLSL_TYPE_ERROR:
-      unreachable("not reached");
-   }
-
-   return 0;
-}
-
-static void
-assign_var_locations(struct hash_table *ht, unsigned *size)
+void
+nir_assign_var_locations(struct exec_list *var_list, unsigned *size,
+                         int (*type_size)(const struct glsl_type *))
 {
    unsigned location = 0;
 
-   struct hash_entry *entry;
-   hash_table_foreach(ht, entry) {
-      nir_variable *var = (nir_variable *) entry->data;
-
+   nir_foreach_variable(var, var_list) {
       /*
        * UBO's have their own address spaces, so don't count them towards the
        * number of global uniforms
        */
-      if (var->data.mode == nir_var_uniform && var->interface_type != NULL)
+      if ((var->data.mode == nir_var_uniform || var->data.mode == nir_var_shader_storage) &&
+          var->interface_type != NULL)
          continue;
 
       var->data.driver_location = location;
@@ -98,36 +63,54 @@ assign_var_locations(struct hash_table *ht, unsigned *size)
    *size = location;
 }
 
-static void
-assign_var_locations_shader(nir_shader *shader)
+/**
+ * Returns true if we're processing a stage whose inputs are arrays indexed
+ * by a vertex number (such as geometry shader inputs).
+ */
+static bool
+is_per_vertex_input(struct lower_io_state *state, nir_variable *var)
 {
-   assign_var_locations(shader->inputs, &shader->num_inputs);
-   assign_var_locations(shader->outputs, &shader->num_outputs);
-   assign_var_locations(shader->uniforms, &shader->num_uniforms);
+   gl_shader_stage stage = state->builder.shader->stage;
+
+   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);
 }
 
 static bool
-deref_has_indirect(nir_deref_var *deref)
+is_per_vertex_output(struct lower_io_state *state, nir_variable *var)
 {
-   for (nir_deref *tail = deref->deref.child; tail; tail = tail->child) {
-      if (tail->deref_type == nir_deref_type_array) {
-         nir_deref_array *arr = nir_deref_as_array(tail);
-         if (arr->deref_array_type == nir_deref_array_type_indirect)
-            return true;
-      }
-   }
-
-   return false;
+   gl_shader_stage stage = state->builder.shader->stage;
+   return var->data.mode == nir_var_shader_out && !var->data.patch &&
+          stage == MESA_SHADER_TESS_CTRL;
 }
 
-static unsigned
-get_io_offset(nir_deref_var *deref, nir_instr *instr, nir_src *indirect,
-              struct lower_io_state *state)
+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 *))
 {
-   bool found_indirect = false;
-   unsigned base_offset = 0;
-
    nir_deref *tail = &deref->deref;
+
+   /* For per-vertex input arrays (i.e. geometry shader inputs), keep the
+    * outermost array index separate.  Process the rest normally.
+    */
+   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);
+      if (deref_array->deref_array_type == nir_deref_array_type_indirect) {
+         vtx = nir_iadd(b, vtx, nir_ssa_for_src(b, deref_array->indirect, 1));
+      }
+      *vertex_index = vtx;
+   }
+
+   /* Just emit code and let constant-folding go to town */
+   nir_ssa_def *offset = nir_imm_int(b, 0);
+
    while (tail->child != NULL) {
       const struct glsl_type *parent_type = tail->type;
       tail = tail->child;
@@ -136,51 +119,51 @@ get_io_offset(nir_deref_var *deref, nir_instr *instr, nir_src *indirect,
          nir_deref_array *deref_array = nir_deref_as_array(tail);
          unsigned size = type_size(tail->type);
 
-         base_offset += size * deref_array->base_offset;
+         offset = nir_iadd(b, offset,
+                           nir_imm_int(b, size * deref_array->base_offset));
 
          if (deref_array->deref_array_type == nir_deref_array_type_indirect) {
-            nir_load_const_instr *load_const =
-               nir_load_const_instr_create(state->mem_ctx, 1);
-            load_const->value.u[0] = size;
-            nir_instr_insert_before(instr, &load_const->instr);
-
-            nir_alu_instr *mul = nir_alu_instr_create(state->mem_ctx,
-                                                      nir_op_imul);
-            mul->src[0].src.is_ssa = true;
-            mul->src[0].src.ssa = &load_const->def;
-            nir_src_copy(&mul->src[1].src, &deref_array->indirect,
-                         state->mem_ctx);
-            mul->dest.write_mask = 1;
-            nir_ssa_dest_init(&mul->instr, &mul->dest.dest, 1, NULL);
-            nir_instr_insert_before(instr, &mul->instr);
-
-            if (found_indirect) {
-               nir_alu_instr *add = nir_alu_instr_create(state->mem_ctx,
-                                                         nir_op_iadd);
-               add->src[0].src = *indirect;
-               add->src[1].src.is_ssa = true;
-               add->src[1].src.ssa = &mul->dest.dest.ssa;
-               add->dest.write_mask = 1;
-               nir_ssa_dest_init(&add->instr, &add->dest.dest, 1, NULL);
-               nir_instr_insert_before(instr, &add->instr);
-
-               indirect->is_ssa = true;
-               indirect->ssa = &add->dest.dest.ssa;
-            } else {
-               indirect->is_ssa = true;
-               indirect->ssa = &mul->dest.dest.ssa;
-               found_indirect = true;
-            }
+            nir_ssa_def *mul =
+               nir_imul(b, nir_imm_int(b, size),
+                        nir_ssa_for_src(b, deref_array->indirect, 1));
+
+            offset = nir_iadd(b, offset, mul);
          }
       } else if (tail->deref_type == nir_deref_type_struct) {
          nir_deref_struct *deref_struct = nir_deref_as_struct(tail);
 
-         for (unsigned i = 0; i < deref_struct->index; i++)
-            base_offset += type_size(glsl_get_struct_field(parent_type, i));
+         unsigned field_offset = 0;
+         for (unsigned i = 0; i < deref_struct->index; i++) {
+            field_offset += type_size(glsl_get_struct_field(parent_type, i));
+         }
+         offset = nir_iadd(b, offset, nir_imm_int(b, field_offset));
       }
    }
 
-   return base_offset;
+   return offset;
+}
+
+static nir_intrinsic_op
+load_op(struct lower_io_state *state,
+        nir_variable_mode mode, bool per_vertex)
+{
+   nir_intrinsic_op op;
+   switch (mode) {
+   case nir_var_shader_in:
+      op = per_vertex ? 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;
+      break;
+   case nir_var_uniform:
+      op = nir_intrinsic_load_uniform;
+      break;
+   default:
+      unreachable("Unknown variable mode");
+   }
+   return op;
 }
 
 static bool
@@ -188,56 +171,61 @@ nir_lower_io_block(nir_block *block, void *void_state)
 {
    struct lower_io_state *state = void_state;
 
+   nir_builder *b = &state->builder;
+
    nir_foreach_instr_safe(block, instr) {
       if (instr->type != nir_instr_type_intrinsic)
          continue;
 
       nir_intrinsic_instr *intrin = nir_instr_as_intrinsic(instr);
 
+      if (intrin->intrinsic != nir_intrinsic_load_var &&
+          intrin->intrinsic != nir_intrinsic_store_var)
+         continue;
+
+      nir_variable_mode mode = intrin->variables[0]->var->data.mode;
+
+      if (state->mode != nir_var_all && state->mode != mode)
+         continue;
+
+      if (mode != nir_var_shader_in &&
+          mode != nir_var_shader_out &&
+          mode != nir_var_uniform)
+         continue;
+
+      b->cursor = nir_before_instr(instr);
+
       switch (intrin->intrinsic) {
       case nir_intrinsic_load_var: {
-         nir_variable_mode mode = intrin->variables[0]->var->data.mode;
-         if (mode != nir_var_shader_in && mode != nir_var_uniform)
-            continue;
-
-         bool has_indirect = deref_has_indirect(intrin->variables[0]);
-
-         /* Figure out the opcode */
-         nir_intrinsic_op load_op;
-         switch (mode) {
-         case nir_var_shader_in:
-            load_op = has_indirect ? nir_intrinsic_load_input_indirect :
-                                     nir_intrinsic_load_input;
-            break;
-         case nir_var_uniform:
-            load_op = has_indirect ? nir_intrinsic_load_uniform_indirect :
-                                     nir_intrinsic_load_uniform;
-            break;
-         default:
-            unreachable("Unknown variable mode");
-         }
+         bool per_vertex =
+            is_per_vertex_input(state, intrin->variables[0]->var) ||
+            is_per_vertex_output(state, intrin->variables[0]->var);
 
-         nir_intrinsic_instr *load = nir_intrinsic_instr_create(state->mem_ctx,
-                                                                load_op);
+         nir_ssa_def *offset;
+         nir_ssa_def *vertex_index;
+
+         offset = get_io_offset(b, intrin->variables[0],
+                                per_vertex ? &vertex_index : NULL,
+                                state->type_size);
+
+         nir_intrinsic_instr *load =
+            nir_intrinsic_instr_create(state->mem_ctx,
+                                       load_op(state, mode, per_vertex));
          load->num_components = intrin->num_components;
 
-         nir_src indirect;
-         unsigned offset = get_io_offset(intrin->variables[0],
-                                         &intrin->instr, &indirect, state);
-         offset += intrin->variables[0]->var->data.driver_location;
+         load->const_index[0] =
+            intrin->variables[0]->var->data.driver_location;
 
-         load->const_index[0] = offset;
-         load->const_index[1] = 1;
+         if (per_vertex)
+            load->src[0] = nir_src_for_ssa(vertex_index);
 
-         if (has_indirect)
-            load->src[0] = indirect;
+         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, NULL);
             nir_ssa_def_rewrite_uses(&intrin->dest.ssa,
-                                     nir_src_for_ssa(&load->dest.ssa),
-                                     state->mem_ctx);
+                                     nir_src_for_ssa(&load->dest.ssa));
          } else {
             nir_dest_copy(&load->dest, &intrin->dest, state->mem_ctx);
          }
@@ -248,34 +236,35 @@ nir_lower_io_block(nir_block *block, void *void_state)
       }
 
       case nir_intrinsic_store_var: {
-         if (intrin->variables[0]->var->data.mode != nir_var_shader_out)
-            continue;
+         assert(mode == nir_var_shader_out);
 
-         bool has_indirect = deref_has_indirect(intrin->variables[0]);
+         nir_ssa_def *offset;
+         nir_ssa_def *vertex_index;
 
-         nir_intrinsic_op store_op;
-         if (has_indirect) {
-            store_op = nir_intrinsic_store_output_indirect;
-         } else {
-            store_op = nir_intrinsic_store_output;
-         }
+         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_op store_op =
+            per_vertex ? nir_intrinsic_store_per_vertex_output :
+                         nir_intrinsic_store_output;
 
          nir_intrinsic_instr *store = nir_intrinsic_instr_create(state->mem_ctx,
                                                                  store_op);
          store->num_components = intrin->num_components;
 
-         nir_src indirect;
-         unsigned offset = get_io_offset(intrin->variables[0],
-                                         &intrin->instr, &indirect, state);
-         offset += intrin->variables[0]->var->data.driver_location;
+         nir_src_copy(&store->src[0], &intrin->src[0], store);
 
-         store->const_index[0] = offset;
-         store->const_index[1] = 1;
+         store->const_index[0] =
+            intrin->variables[0]->var->data.driver_location;
 
-         nir_src_copy(&store->src[0], &intrin->src[0], state->mem_ctx);
+         if (per_vertex)
+            store->src[1] = nir_src_for_ssa(vertex_index);
 
-         if (has_indirect)
-            store->src[1] = indirect;
+         store->src[per_vertex ? 2 : 1] = nir_src_for_ssa(offset);
 
          nir_instr_insert_before(&intrin->instr, &store->instr);
          nir_instr_remove(&intrin->instr);
@@ -291,11 +280,16 @@ nir_lower_io_block(nir_block *block, void *void_state)
 }
 
 static void
-nir_lower_io_impl(nir_function_impl *impl)
+nir_lower_io_impl(nir_function_impl *impl,
+                  nir_variable_mode mode,
+                  int (*type_size)(const struct glsl_type *))
 {
    struct lower_io_state state;
 
+   nir_builder_init(&state.builder, impl);
    state.mem_ctx = ralloc_parent(impl);
+   state.mode = mode;
+   state.type_size = type_size;
 
    nir_foreach_block(impl, nir_lower_io_block, &state);
 
@@ -304,12 +298,53 @@ nir_lower_io_impl(nir_function_impl *impl)
 }
 
 void
-nir_lower_io(nir_shader *shader)
+nir_lower_io(nir_shader *shader, nir_variable_mode mode,
+             int (*type_size)(const struct glsl_type *))
 {
-   assign_var_locations_shader(shader);
-
    nir_foreach_overload(shader, overload) {
       if (overload->impl)
-         nir_lower_io_impl(overload->impl);
+         nir_lower_io_impl(overload->impl, mode, type_size);
+   }
+}
+
+/**
+ * Return the offset soruce for a load/store intrinsic.
+ */
+nir_src *
+nir_get_io_offset_src(nir_intrinsic_instr *instr)
+{
+   switch (instr->intrinsic) {
+   case nir_intrinsic_load_input:
+   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_store_output:
+      return &instr->src[1];
+   case nir_intrinsic_store_ssbo:
+   case nir_intrinsic_store_per_vertex_output:
+      return &instr->src[2];
+   default:
+      return NULL;
+   }
+}
+
+/**
+ * Return the vertex index source for a load/store per_vertex intrinsic.
+ */
+nir_src *
+nir_get_io_vertex_index_src(nir_intrinsic_instr *instr)
+{
+   switch (instr->intrinsic) {
+   case nir_intrinsic_load_per_vertex_input:
+   case nir_intrinsic_load_per_vertex_output:
+      return &instr->src[0];
+   case nir_intrinsic_store_per_vertex_output:
+      return &instr->src[1];
+   default:
+      return NULL;
    }
 }