[RFCv0 4/8] mesa/st: add support for NIR as possible driver IR
Rob Clark
robdclark at gmail.com
Mon Oct 19 12:47:57 PDT 2015
---
src/mesa/state_tracker/st_glsl_to_tgsi.cpp | 223 ++++++++++++++++++++++++++++-
src/mesa/state_tracker/st_program.c | 43 +++++-
2 files changed, 260 insertions(+), 6 deletions(-)
diff --git a/src/mesa/state_tracker/st_glsl_to_tgsi.cpp b/src/mesa/state_tracker/st_glsl_to_tgsi.cpp
index 06f510d..49f496f 100644
--- a/src/mesa/state_tracker/st_glsl_to_tgsi.cpp
+++ b/src/mesa/state_tracker/st_glsl_to_tgsi.cpp
@@ -35,6 +35,9 @@
#include "glsl_parser_extras.h"
#include "ir_optimization.h"
+#include "nir.h"
+#include "glsl_to_nir.h"
+
#include "main/errors.h"
#include "main/shaderobj.h"
#include "main/uniforms.h"
@@ -5486,9 +5489,9 @@ out:
* generating Mesa IR.
*/
static struct gl_program *
-get_mesa_program(struct gl_context *ctx,
- struct gl_shader_program *shader_program,
- struct gl_shader *shader)
+get_mesa_program_tgsi(struct gl_context *ctx,
+ struct gl_shader_program *shader_program,
+ struct gl_shader *shader)
{
glsl_to_tgsi_visitor* v;
struct gl_program *prog;
@@ -5680,6 +5683,220 @@ get_mesa_program(struct gl_context *ctx,
return prog;
}
+/* TODO dup'd from brw_vec4_vistor.cpp.. what should we do? */
+static int
+type_size_vec4(const struct glsl_type *type)
+{
+ unsigned int i;
+ int size;
+
+ switch (type->base_type) {
+ case GLSL_TYPE_UINT:
+ case GLSL_TYPE_INT:
+ case GLSL_TYPE_FLOAT:
+ case GLSL_TYPE_BOOL:
+ if (type->is_matrix()) {
+ return type->matrix_columns;
+ } else {
+ /* Regardless of size of vector, it gets a vec4. This is bad
+ * packing for things like floats, but otherwise arrays become a
+ * mess. Hopefully a later pass over the code can pack scalars
+ * down if appropriate.
+ */
+ return 1;
+ }
+ case GLSL_TYPE_ARRAY:
+ assert(type->length > 0);
+ return type_size_vec4(type->fields.array) * type->length;
+ case GLSL_TYPE_STRUCT:
+ size = 0;
+ for (i = 0; i < type->length; i++) {
+ size += type_size_vec4(type->fields.structure[i].type);
+ }
+ return size;
+ case GLSL_TYPE_SUBROUTINE:
+ return 1;
+
+ case GLSL_TYPE_SAMPLER:
+ /* Samplers take up no register space, since they're baked in at
+ * link time.
+ */
+ return 0;
+ case GLSL_TYPE_ATOMIC_UINT:
+ return 0;
+ case GLSL_TYPE_IMAGE:
+// return DIV_ROUND_UP(BRW_IMAGE_PARAM_SIZE, 4);
+ case GLSL_TYPE_VOID:
+ case GLSL_TYPE_DOUBLE:
+ case GLSL_TYPE_ERROR:
+ case GLSL_TYPE_INTERFACE:
+ unreachable("not reached");
+ }
+
+ return 0;
+}
+
+static struct gl_program *
+get_mesa_program_nir(struct gl_context *ctx,
+ struct gl_shader_program *shader_program,
+ struct gl_shader *shader)
+{
+ struct gl_program *prog;
+ GLenum target = _mesa_shader_stage_to_program(shader->Stage);
+ struct gl_shader_compiler_options *options =
+ &ctx->Const.ShaderCompilerOptions[_mesa_shader_enum_to_shader_stage(shader->Type)];
+ struct pipe_screen *pscreen = ctx->st->pipe->screen;
+ unsigned ptarget = st_shader_stage_to_ptarget(shader->Stage);
+ nir_shader_compiler_options nir_options = {0};
+ nir_shader *nir;
+
+ /* TODO maybe options marked with 'XXX' should be PIPE_SHADER_CAP_x's */
+ nir_options.lower_ffma =
+ !pscreen->get_shader_param(pscreen, ptarget, PIPE_SHADER_CAP_TGSI_FMA_SUPPORTED);
+ nir_options.lower_flrp = true; // XXX
+ nir_options.lower_fpow = options->EmitNoPow;
+ nir_options.lower_fsat = options->EmitNoSat;
+ nir_options.lower_fsqrt =
+ !pscreen->get_shader_param(pscreen, ptarget, PIPE_SHADER_CAP_TGSI_SQRT_SUPPORTED);
+ nir_options.lower_negate = false; // XXX
+ nir_options.lower_sub = false; // XXX
+ nir_options.lower_scmp = true; // XXX
+ nir_options.fdot_replicates = false;// XXX
+ nir_options.lower_ffract = true; // XXX
+ nir_options.native_integers = ctx->Const.NativeIntegers;
+
+ validate_ir_tree(shader->ir);
+
+ prog = ctx->Driver.NewProgram(ctx, target, shader_program->Name);
+ if (!prog)
+ return NULL;
+ 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,
+ prog->Parameters);
+
+// /* Remove reads from output registers. */
+// lower_output_reads(shader->Stage, shader->ir);
+
+ /* 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",
+ _mesa_shader_stage_to_string(shader->Stage),
+ shader_program->Name);
+ _mesa_print_ir(_mesa_get_log_file(), shader->ir, NULL);
+ _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;
+ _mesa_update_shader_textures_used(shader_program, prog);
+
+ _mesa_reference_program(ctx, &shader->Program, prog);
+
+ nir = glsl_to_nir(shader_program, shader->Stage, &nir_options);
+ prog->nir = nir;
+
+ nir_validate_shader(nir);
+
+ if (shader->Stage == MESA_SHADER_GEOMETRY) {
+ nir_lower_gs_intrinsics(nir);
+ nir_validate_shader(nir);
+ }
+
+ nir_print_shader(nir, _mesa_get_log_file());
+
+ nir_lower_global_vars_to_local(nir);
+ nir_validate_shader(nir);
+
+ nir_lower_var_copies(nir);
+ nir_validate_shader(nir);
+
+ nir_assign_var_locations(&nir->inputs,
+ &nir->num_inputs,
+ type_size_vec4);
+ nir_assign_var_locations(&nir->outputs,
+ &nir->num_outputs,
+ type_size_vec4);
+ nir_assign_var_locations(&nir->uniforms,
+ &nir->num_uniforms,
+ type_size_vec4);
+
+ nir_lower_io(nir, nir_var_all, type_size_vec4);
+ nir_validate_shader(nir);
+
+
+ // XXX do we need anything else from brw_create_nir().. and what
+ // is best way to split up which things should be here vs driver?
+ // currently just trying to make the result here similar to what
+ // we get from tgsi_to_nir().. so lower_io, etc..
+
+ // XXX new flag, probably?
+ if (ctx->_Shader->Flags & GLSL_DUMP) {
+ _mesa_log("\n");
+ _mesa_log("NIR IR for linked %s program %d:\n",
+ _mesa_shader_stage_to_string(shader->Stage),
+ shader_program->Name);
+ nir_print_shader(nir, _mesa_get_log_file());
+ _mesa_log("\n\n");
+ }
+
+ /* 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);
+
+ return prog;
+}
+
+static struct gl_program *
+get_mesa_program(struct gl_context *ctx,
+ struct gl_shader_program *shader_program,
+ struct gl_shader *shader)
+{
+ struct pipe_screen *pscreen = ctx->st->pipe->screen;
+ unsigned ptarget = st_shader_stage_to_ptarget(shader->Stage);
+ enum pipe_shader_ir preferred_ir = (enum pipe_shader_ir)
+ pscreen->get_shader_param(pscreen, ptarget, PIPE_SHADER_CAP_PREFERRED_IR);
+ /* TODO: only do this for VERTEX stage for now: */
+ if ((preferred_ir == PIPE_SHADER_IR_NIR) &&
+ (shader->Type == GL_VERTEX_SHADER)) {
+ return get_mesa_program_nir(ctx, shader_program, shader);
+ } else {
+ return get_mesa_program_tgsi(ctx, shader_program, shader);
+ }
+}
+
+
extern "C" {
static void
diff --git a/src/mesa/state_tracker/st_program.c b/src/mesa/state_tracker/st_program.c
index 6a69ba7..3617160 100644
--- a/src/mesa/state_tracker/st_program.c
+++ b/src/mesa/state_tracker/st_program.c
@@ -38,6 +38,8 @@
#include "program/prog_print.h"
#include "program/programopt.h"
+#include "nir.h"
+
#include "pipe/p_context.h"
#include "pipe/p_defines.h"
#include "pipe/p_shader_tokens.h"
@@ -70,8 +72,13 @@ delete_vp_variant(struct st_context *st, struct st_vp_variant *vpv)
if (vpv->draw_shader)
draw_delete_vertex_shader( st->draw, vpv->draw_shader );
- if (vpv->tgsi.tokens)
- ureg_free_tokens(vpv->tgsi.tokens);
+ if (vpv->tgsi.ir == PIPE_SHADER_IR_NIR) {
+ if (vpv->tgsi.nir)
+ ralloc_free(vpv->tgsi.nir);
+ } else {
+ if (vpv->tgsi.tokens)
+ ureg_free_tokens(vpv->tgsi.tokens);
+ }
free( vpv );
}
@@ -270,6 +277,20 @@ st_translate_vertex_program(struct st_context *st,
stvp->num_inputs = 0;
+ if (stvp->Base.Base.nir) {
+ /* TODO should we be doing glsl_to_nir() here instead of in
+ * get_mesa_program_nir()?
+ *
+ * NOTE glsl_to_nir() path doesn't apply for GL_*_PROGRAM_ARB
+ * so I *think* that simplifies some things..
+ */
+ nir_shader *nir = stvp->Base.Base.nir;
+ stvp->tgsi.ir = PIPE_SHADER_IR_NIR;
+ stvp->tgsi.nir = nir;
+ stvp->num_inputs = nir->num_inputs;
+ return true;
+ }
+
if (stvp->Base.IsPositionInvariant)
_mesa_insert_mvp_code(st->ctx, &stvp->Base);
@@ -466,10 +487,26 @@ st_create_vp_variant(struct st_context *st,
struct pipe_context *pipe = st->pipe;
vpv->key = *key;
- vpv->tgsi.tokens = tgsi_dup_tokens(stvp->tgsi.tokens);
vpv->tgsi.stream_output = stvp->tgsi.stream_output;
vpv->num_inputs = stvp->num_inputs;
+ if (stvp->tgsi.ir == PIPE_SHADER_IR_NIR) {
+ vpv->tgsi.ir = PIPE_SHADER_IR_NIR;
+ /* TODO, we need nir_clone() here for dealing w/ variants
+ * that are maintained at the mesa-st layer. (Otoh, the
+ * pipe-drivers already maintain their own variants, it
+ * might be nicer to just expose the variant-key to the
+ * driver and let it decide what to lower, since it is
+ * doing that already for other things..)
+ */
+ vpv->tgsi.nir = stvp->tgsi.nir;
+
+ vpv->driver_shader = pipe->create_vs_state(pipe, &vpv->tgsi);
+ return vpv;
+ }
+
+ vpv->tgsi.tokens = tgsi_dup_tokens(stvp->tgsi.tokens);
+
/* Emulate features. */
if (key->clamp_color || key->passthrough_edgeflags) {
const struct tgsi_token *tokens;
--
2.5.0
More information about the dri-devel
mailing list