st/glsl: move nir linking loop to new function st_link_nir()
[mesa.git] / src / mesa / state_tracker / st_glsl_to_nir.cpp
index e6e6bb953f41dc5fb1114babb4fc178647bfde99..8b66f8277a85b09f6f765133853a9f2bbfc4e963 100644 (file)
 #include "main/errors.h"
 #include "main/shaderapi.h"
 #include "main/uniforms.h"
-#include "util/string_to_uint_map.h"
 
 #include "st_context.h"
 #include "st_program.h"
-#include "st_glsl_types.h"
 
 #include "compiler/nir/nir.h"
 #include "compiler/glsl_types.h"
 #include "compiler/glsl/glsl_to_nir.h"
 #include "compiler/glsl/ir.h"
+#include "compiler/glsl/string_to_uint_map.h"
 
 
+static int
+type_size(const struct glsl_type *type)
+{
+   return type->count_attribute_slots(false);
+}
+
 /* Depending on PIPE_CAP_TGSI_TEXCOORD (st->needs_texcoord_semantic) we
  * may need to fix up varying slots so the glsl->nir path is aligned
  * with the anything->tgsi->nir path.
@@ -80,10 +85,10 @@ st_nir_assign_vs_in_locations(struct gl_program *prog, nir_shader *nir)
 
    /* TODO de-duplicate w/ similar code in st_translate_vertex_program()? */
    for (attr = 0; attr < VERT_ATTRIB_MAX; attr++) {
-      if ((prog->InputsRead & BITFIELD64_BIT(attr)) != 0) {
+      if ((prog->info.inputs_read & BITFIELD64_BIT(attr)) != 0) {
          input_to_index[attr] = num_inputs;
          num_inputs++;
-         if ((prog->DoubleInputsRead & BITFIELD64_BIT(attr)) != 0) {
+         if ((prog->info.double_inputs_read & BITFIELD64_BIT(attr)) != 0) {
             /* add placeholder for second part of a double attribute */
             num_inputs++;
          }
@@ -92,6 +97,9 @@ st_nir_assign_vs_in_locations(struct gl_program *prog, nir_shader *nir)
       }
    }
 
+   /* bit of a hack, mirroring st_translate_vertex_program */
+   input_to_index[VERT_ATTRIB_EDGEFLAG] = num_inputs;
+
    nir->num_inputs = 0;
    nir_foreach_variable_safe(var, &nir->inputs) {
       attr = var->data.location;
@@ -113,6 +121,34 @@ st_nir_assign_vs_in_locations(struct gl_program *prog, nir_shader *nir)
    }
 }
 
+static void
+st_nir_assign_var_locations(struct exec_list *var_list, unsigned *size)
+{
+   unsigned location = 0;
+   unsigned assigned_locations[VARYING_SLOT_MAX];
+   uint64_t processed_locs = 0;
+
+   nir_foreach_variable(var, var_list) {
+      /* Because component packing allows varyings to share the same location
+       * we may have already have processed this location.
+       */
+      if (var->data.location >= VARYING_SLOT_VAR0 &&
+          processed_locs & ((uint64_t)1 << var->data.location)) {
+         var->data.driver_location = assigned_locations[var->data.location];
+         *size += type_size(var->type);
+         continue;
+      }
+
+      assigned_locations[var->data.location] = location;
+      var->data.driver_location = location;
+      location += type_size(var->type);
+
+      processed_locs |= ((uint64_t)1 << var->data.location);
+   }
+
+   *size += location;
+}
+
 static int
 st_nir_lookup_parameter_index(const struct gl_program_parameter_list *params,
                               const char *name)
@@ -168,6 +204,7 @@ st_nir_assign_uniform_locations(struct gl_program *prog,
 {
    int max = 0;
    int shaderidx = 0;
+   int imageidx = 0;
 
    nir_foreach_variable(uniform, uniform_list) {
       int loc;
@@ -180,11 +217,15 @@ st_nir_assign_uniform_locations(struct gl_program *prog,
           uniform->interface_type != NULL)
          continue;
 
-      if (uniform->type->is_sampler()) {
-         unsigned val;
+      if (uniform->type->is_sampler() || uniform->type->is_image()) {
+         unsigned val = 0;
          bool found = shader_program->UniformHash->get(val, uniform->name);
-         loc = shaderidx++;
+         if (uniform->type->is_sampler())
+            loc = shaderidx++;
+         else
+            loc = imageidx++;
          assert(found);
+         (void) found; /* silence unused var warning */
          /* this ensure that nir_lower_samplers looks at the correct
           * shader_program->UniformStorage[location]:
           */
@@ -201,7 +242,7 @@ st_nir_assign_uniform_locations(struct gl_program *prog,
 
       uniform->data.driver_location = loc;
 
-      max = MAX2(max, loc + st_glsl_type_size(uniform->type));
+      max = MAX2(max, loc + type_size(uniform->type));
    }
    *size = max;
 }
@@ -218,7 +259,7 @@ st_glsl_to_nir(struct st_context *st, struct gl_program *prog,
                gl_shader_stage stage)
 {
    struct pipe_screen *pscreen = st->pipe->screen;
-   enum pipe_shader_type ptarget = st_shader_stage_to_ptarget(stage);
+   enum pipe_shader_type ptarget = pipe_shader_type_from_mesa(stage);
    const nir_shader_compiler_options *options;
    nir_shader *nir;
 
@@ -232,7 +273,37 @@ st_glsl_to_nir(struct st_context *st, struct gl_program *prog,
       return prog->nir;
 
    nir = glsl_to_nir(shader_program, stage, options);
-   prog->nir = nir;
+
+   /* Make a pass over the IR to add state references for any built-in
+    * uniforms that are used.  This has to be done now (during linking).
+    * Code generation doesn't happen until the first time this shader is
+    * used for rendering.  Waiting until then to generate the parameters is
+    * too late.  At that point, the values for the built-in uniforms won't
+    * get sent to the shader.
+    */
+   nir_foreach_variable(var, &nir->uniforms) {
+      if (strncmp(var->name, "gl_", 3) == 0) {
+         const nir_state_slot *const slots = var->state_slots;
+         assert(var->state_slots != NULL);
+
+         for (unsigned int i = 0; i < var->num_state_slots; i++) {
+            _mesa_add_state_reference(prog->Parameters,
+                                      (gl_state_index *)slots[i].tokens);
+         }
+      }
+   }
+
+   /* Avoid reallocation of the program parameter list, because the uniform
+    * storage is only associated with the original parameter list.
+    * This should be enough for Bitmap and DrawPixels constants.
+    */
+   _mesa_reserve_parameter_storage(prog->Parameters, 8);
+
+   /* This has to be done last.  Any operation the can cause
+    * prog->ParameterValues to get reallocated (e.g., anything that adds a
+    * program constant) has to happen before creating this linkage.
+    */
+   _mesa_associate_uniform_storage(st->ctx, shader_program, prog, true);
 
    NIR_PASS_V(nir, nir_lower_io_to_temporaries,
          nir_shader_get_entrypoint(nir),
@@ -240,14 +311,13 @@ st_glsl_to_nir(struct st_context *st, struct gl_program *prog,
    NIR_PASS_V(nir, nir_lower_global_vars_to_local);
    NIR_PASS_V(nir, nir_split_var_copies);
    NIR_PASS_V(nir, nir_lower_var_copies);
-   NIR_PASS_V(nir, st_nir_lower_builtin);
 
    /* fragment shaders may need : */
    if (stage == MESA_SHADER_FRAGMENT) {
       static const gl_state_index wposTransformState[STATE_LENGTH] = {
          STATE_INTERNAL, STATE_FB_WPOS_Y_TRANSFORM
       };
-      nir_lower_wpos_ytransform_options wpos_options = {0};
+      nir_lower_wpos_ytransform_options wpos_options = { { 0 } };
       struct pipe_screen *pscreen = st->pipe->screen;
 
       memcpy(wpos_options.state_tokens, wposTransformState,
@@ -267,6 +337,16 @@ st_glsl_to_nir(struct st_context *st, struct gl_program *prog,
       }
    }
 
+   NIR_PASS_V(nir, nir_lower_system_values);
+
+   nir_shader_gather_info(nir, nir_shader_get_entrypoint(nir));
+   prog->info = nir->info;
+
+   st_set_prog_affected_state_flags(prog);
+
+   NIR_PASS_V(nir, st_nir_lower_builtin);
+   NIR_PASS_V(nir, nir_lower_atomics, shader_program);
+
    if (st->ctx->_Shader->Flags & GLSL_DUMP) {
       _mesa_log("\n");
       _mesa_log("NIR IR for linked %s program %d:\n",
@@ -276,6 +356,8 @@ st_glsl_to_nir(struct st_context *st, struct gl_program *prog,
       _mesa_log("\n\n");
    }
 
+   prog->nir = nir;
+
    return nir;
 }
 
@@ -309,58 +391,48 @@ sort_varyings(struct exec_list *var_list)
  * variant lowering.
  */
 void
-st_finalize_nir(struct st_context *st, struct gl_program *prog, nir_shader *nir)
+st_finalize_nir(struct st_context *st, struct gl_program *prog,
+                struct gl_shader_program *shader_program, nir_shader *nir)
 {
+   struct pipe_screen *screen = st->pipe->screen;
+
    NIR_PASS_V(nir, nir_split_var_copies);
    NIR_PASS_V(nir, nir_lower_var_copies);
    NIR_PASS_V(nir, nir_lower_io_types);
 
-   if (nir->stage == MESA_SHADER_VERTEX) {
+   if (nir->info.stage == MESA_SHADER_VERTEX) {
       /* Needs special handling so drvloc matches the vbo state: */
       st_nir_assign_vs_in_locations(prog, nir);
       /* Re-lower global vars, to deal with any dead VS inputs. */
       NIR_PASS_V(nir, nir_lower_global_vars_to_local);
 
       sort_varyings(&nir->outputs);
-      nir_assign_var_locations(&nir->outputs,
-                               &nir->num_outputs,
-                               VARYING_SLOT_VAR0,
-                               st_glsl_type_size);
+      st_nir_assign_var_locations(&nir->outputs,
+                                  &nir->num_outputs);
       st_nir_fixup_varying_slots(st, &nir->outputs);
-   } else if (nir->stage == MESA_SHADER_FRAGMENT) {
+   } else if (nir->info.stage == MESA_SHADER_FRAGMENT) {
       sort_varyings(&nir->inputs);
-      nir_assign_var_locations(&nir->inputs,
-                               &nir->num_inputs,
-                               VARYING_SLOT_VAR0,
-                               st_glsl_type_size);
+      st_nir_assign_var_locations(&nir->inputs,
+                                  &nir->num_inputs);
       st_nir_fixup_varying_slots(st, &nir->inputs);
-      nir_assign_var_locations(&nir->outputs,
-                               &nir->num_outputs,
-                               FRAG_RESULT_DATA0,
-                               st_glsl_type_size);
+      st_nir_assign_var_locations(&nir->outputs,
+                                  &nir->num_outputs);
+   } else if (nir->info.stage == MESA_SHADER_COMPUTE) {
+       /* TODO? */
    } else {
       unreachable("invalid shader type for tgsi bypass\n");
    }
 
-   struct gl_shader_program *shader_program;
-   switch (nir->stage) {
-   case MESA_SHADER_VERTEX:
-      shader_program = ((struct st_vertex_program *)prog)->shader_program;
-      break;
-   case MESA_SHADER_FRAGMENT:
-      shader_program = ((struct st_fragment_program *)prog)->shader_program;
-      break;
-   default:
-      assert(!"should not be reached");
-      return;
-   }
+   NIR_PASS_V(nir, nir_lower_atomics_to_ssbo,
+         st->ctx->Const.Program[nir->info.stage].MaxAtomicBuffers);
 
    st_nir_assign_uniform_locations(prog, shader_program,
                                    &nir->uniforms, &nir->num_uniforms);
 
-   NIR_PASS_V(nir, nir_lower_system_values);
-   NIR_PASS_V(nir, nir_lower_io, nir_var_all, st_glsl_type_size);
-   NIR_PASS_V(nir, nir_lower_samplers, shader_program);
+   if (screen->get_param(screen, PIPE_CAP_NIR_SAMPLERS_AS_DEREF))
+      NIR_PASS_V(nir, nir_lower_samplers_as_deref, shader_program);
+   else
+      NIR_PASS_V(nir, nir_lower_samplers, shader_program);
 }
 
 struct gl_program *
@@ -368,44 +440,19 @@ st_nir_get_mesa_program(struct gl_context *ctx,
                         struct gl_shader_program *shader_program,
                         struct gl_linked_shader *shader)
 {
+   struct st_context *st = st_context(ctx);
    struct gl_program *prog;
-   GLenum target = _mesa_shader_stage_to_program(shader->Stage);
 
    validate_ir_tree(shader->ir);
 
-   prog = ctx->Driver.NewProgram(ctx, target, shader_program->Name);
-   if (!prog)
-      return NULL;
+   prog = shader->Program;
 
    prog->Parameters = _mesa_new_parameter_list();
 
-   _mesa_copy_linked_program_data(shader->Stage, shader_program, prog);
-   _mesa_generate_parameters_list_for_uniforms(shader_program, shader,
+   _mesa_copy_linked_program_data(shader_program, shader);
+   _mesa_generate_parameters_list_for_uniforms(ctx, shader_program, shader,
                                                prog->Parameters);
 
-   /* Make a pass over the IR to add state references for any built-in
-    * uniforms that are used.  This has to be done now (during linking).
-    * Code generation doesn't happen until the first time this shader is
-    * used for rendering.  Waiting until then to generate the parameters is
-    * too late.  At that point, the values for the built-in uniforms won't
-    * get sent to the shader.
-    */
-   foreach_in_list(ir_instruction, node, shader->ir) {
-      ir_variable *var = node->as_variable();
-
-      if ((var == NULL) || (var->data.mode != ir_var_uniform) ||
-          (strncmp(var->name, "gl_", 3) != 0))
-         continue;
-
-      const ir_state_slot *const slots = var->get_state_slots();
-      assert(slots != NULL);
-
-      for (unsigned int i = 0; i < var->get_num_state_slots(); i++) {
-         _mesa_add_state_reference(prog->Parameters,
-                                   (gl_state_index *) slots[i].tokens);
-      }
-   }
-
    if (ctx->_Shader->Flags & GLSL_DUMP) {
       _mesa_log("\n");
       _mesa_log("GLSL IR for linked %s program %d:\n",
@@ -415,47 +462,75 @@ st_nir_get_mesa_program(struct gl_context *ctx,
       _mesa_log("\n\n");
    }
 
-   prog->Instructions = NULL;
-   prog->NumInstructions = 0;
-
-   do_set_program_inouts(shader->ir, prog, shader->Stage);
-
-   prog->SamplersUsed = shader->active_samplers;
-   prog->ShadowSamplers = shader->shadow_samplers;
+   prog->ExternalSamplersUsed = gl_external_samplers(prog);
    _mesa_update_shader_textures_used(shader_program, prog);
 
-   _mesa_reference_program(ctx, &shader->Program, prog);
-
-   /* Avoid reallocation of the program parameter list, because the uniform
-    * storage is only associated with the original parameter list.
-    * This should be enough for Bitmap and DrawPixels constants.
-    */
-   _mesa_reserve_parameter_storage(prog->Parameters, 8);
-
-   /* This has to be done last.  Any operation the can cause
-    * prog->ParameterValues to get reallocated (e.g., anything that adds a
-    * program constant) has to happen before creating this linkage.
-    */
-   _mesa_associate_uniform_storage(ctx, shader_program, prog->Parameters);
-
    struct st_vertex_program *stvp;
+   struct st_common_program *stp;
    struct st_fragment_program *stfp;
+   struct st_compute_program *stcp;
+
+   nir_shader *nir = st_glsl_to_nir(st, prog, shader_program, shader->Stage);
 
    switch (shader->Stage) {
    case MESA_SHADER_VERTEX:
       stvp = (struct st_vertex_program *)prog;
       stvp->shader_program = shader_program;
+      stvp->tgsi.type = PIPE_SHADER_IR_NIR;
+      stvp->tgsi.ir.nir = nir;
+      break;
+   case MESA_SHADER_GEOMETRY:
+   case MESA_SHADER_TESS_CTRL:
+   case MESA_SHADER_TESS_EVAL:
+      stp = (struct st_common_program *)prog;
+      stp->shader_program = shader_program;
+      stp->tgsi.type = PIPE_SHADER_IR_NIR;
+      stp->tgsi.ir.nir = nir;
       break;
    case MESA_SHADER_FRAGMENT:
       stfp = (struct st_fragment_program *)prog;
       stfp->shader_program = shader_program;
+      stfp->tgsi.type = PIPE_SHADER_IR_NIR;
+      stfp->tgsi.ir.nir = nir;
+      break;
+   case MESA_SHADER_COMPUTE:
+      stcp = (struct st_compute_program *)prog;
+      stcp->shader_program = shader_program;
+      stcp->tgsi.ir_type = PIPE_SHADER_IR_NIR;
+      stcp->tgsi.prog = nir_shader_clone(NULL, nir);
       break;
    default:
       assert(!"should not be reached");
       return NULL;
    }
 
+
    return prog;
 }
 
+bool
+st_link_nir(struct gl_context *ctx,
+            struct gl_shader_program *shader_program)
+{
+   for (unsigned i = 0; i < MESA_SHADER_STAGES; i++) {
+      struct gl_linked_shader *shader = shader_program->_LinkedShaders[i];
+      if (shader == NULL)
+         continue;
+
+      struct gl_program *linked_prog =
+         st_nir_get_mesa_program(ctx, shader_program, shader);
+
+      if (linked_prog) {
+         if (!ctx->Driver.ProgramStringNotify(ctx,
+                                              _mesa_shader_stage_to_program(i),
+                                              linked_prog)) {
+            _mesa_reference_program(ctx, &shader->Program, NULL);
+            return false;
+         }
+      }
+   }
+
+   return true;
+}
+
 } /* extern "C" */