#include "main/macros.h"
#include "main/shaderobj.h"
+#include "main/uniforms.h"
#include "program/prog_parameter.h"
#include "program/prog_print.h"
#include "program/prog_optimize.h"
+#include "program/register_allocate.h"
+#include "program/sampler.h"
#include "program/hash_table.h"
#include "brw_context.h"
#include "brw_eu.h"
};
static int using_new_fs = -1;
+static struct brw_reg brw_reg_from_fs_reg(class fs_reg *reg);
struct gl_shader *
brw_new_shader(GLcontext *ctx, GLuint name, GLuint type)
GLboolean
brw_link_shader(GLcontext *ctx, struct gl_shader_program *prog)
{
+ struct intel_context *intel = intel_context(ctx);
if (using_new_fs == -1)
using_new_fs = getenv("INTEL_NEW_FS") != NULL;
do_div_to_mul_rcp(shader->ir);
do_sub_to_add_neg(shader->ir);
do_explog_to_explog2(shader->ir);
-
- brw_do_channel_expressions(shader->ir);
- brw_do_vector_splitting(shader->ir);
+ do_lower_texture_projection(shader->ir);
+ brw_do_cubemap_normalize(shader->ir);
do {
progress = false;
+ brw_do_channel_expressions(shader->ir);
+ brw_do_vector_splitting(shader->ir);
+
progress = do_lower_jumps(shader->ir, true, true,
true, /* main return */
false, /* continue */
GL_TRUE, /* temp */
GL_TRUE /* uniform */
) || progress;
+ if (intel->gen == 6) {
+ progress = do_if_to_cond_assign(shader->ir) || progress;
+ }
} while (progress);
validate_ir_tree(shader->ir);
case GLSL_TYPE_BOOL:
return type->components();
case GLSL_TYPE_ARRAY:
- /* FINISHME: uniform/varying arrays. */
return type_size(type->fields.array) * type->length;
case GLSL_TYPE_STRUCT:
size = 0;
/** Register file: ARF, GRF, MRF, IMM. */
enum register_file file;
- /** Abstract register number. 0 = fixed hw reg */
+ /** virtual register number. 0 = fixed hw reg */
int reg;
- /** Offset within the abstract register. */
+ /** Offset within the virtual register. */
int reg_offset;
/** HW register number. Generally unset until register allocation. */
int hw_reg;
this->conditional_mod = BRW_CONDITIONAL_NONE;
this->predicated = false;
this->sampler = 0;
+ this->target = 0;
+ this->eot = false;
+ this->header_present = false;
this->shadow_compare = false;
}
bool predicated;
int conditional_mod; /**< BRW_CONDITIONAL_* */
- int mlen; /** SEND message length */
+ int mlen; /**< SEND message length */
int sampler;
+ int target; /**< MRT target. */
+ bool eot;
+ bool header_present;
bool shadow_compare;
/** @{
this->c = c;
this->p = &c->func;
this->brw = p->brw;
+ this->fp = brw->fragment_program;
this->intel = &brw->intel;
this->ctx = &intel->ctx;
this->mem_ctx = talloc_new(NULL);
this->shader = shader;
this->fail = false;
- this->next_abstract_grf = 1;
this->variable_ht = hash_table_ctor(0,
hash_table_pointer_hash,
hash_table_pointer_compare);
this->current_annotation = NULL;
this->annotation_string = NULL;
this->annotation_ir = NULL;
+ this->base_ir = NULL;
+
+ this->virtual_grf_sizes = NULL;
+ this->virtual_grf_next = 1;
+ this->virtual_grf_array_size = 0;
+ this->virtual_grf_def = NULL;
+ this->virtual_grf_use = NULL;
+
+ this->kill_emitted = false;
}
+
~fs_visitor()
{
talloc_free(this->mem_ctx);
}
fs_reg *variable_storage(ir_variable *var);
+ int virtual_grf_alloc(int size);
void visit(ir_variable *ir);
void visit(ir_assignment *ir);
fs_inst *emit(fs_inst inst);
void assign_curb_setup();
+ void calculate_urb_setup();
void assign_urb_setup();
void assign_regs();
+ void assign_regs_trivial();
+ void calculate_live_intervals();
+ bool propagate_constants();
+ bool dead_code_eliminate();
+ bool virtual_grf_interferes(int a, int b);
void generate_code();
void generate_fb_write(fs_inst *inst);
void generate_linterp(fs_inst *inst, struct brw_reg dst,
struct brw_reg *src);
void generate_tex(fs_inst *inst, struct brw_reg dst, struct brw_reg src);
void generate_math(fs_inst *inst, struct brw_reg dst, struct brw_reg *src);
- void generate_discard(fs_inst *inst);
+ void generate_discard(fs_inst *inst, struct brw_reg temp);
void generate_ddx(fs_inst *inst, struct brw_reg dst, struct brw_reg src);
void generate_ddy(fs_inst *inst, struct brw_reg dst, struct brw_reg src);
void emit_dummy_fs();
- void emit_interpolation();
- void emit_pinterp(int location);
+ fs_reg *emit_fragcoord_interpolation(ir_variable *ir);
+ fs_reg *emit_frontfacing_interpolation(ir_variable *ir);
+ fs_reg *emit_general_interpolation(ir_variable *ir);
+ void emit_interpolation_setup_gen4();
+ void emit_interpolation_setup_gen6();
+ fs_inst *emit_texture_gen4(ir_texture *ir, fs_reg dst, fs_reg coordinate);
+ fs_inst *emit_texture_gen5(ir_texture *ir, fs_reg dst, fs_reg coordinate);
void emit_fb_writes();
+ void emit_assignment_writes(fs_reg &l, fs_reg &r,
+ const glsl_type *type, bool predicated);
struct brw_reg interp_reg(int location, int channel);
+ int setup_uniform_values(int loc, const glsl_type *type);
+ void setup_builtin_uniform_values(ir_variable *ir);
struct brw_context *brw;
+ const struct gl_fragment_program *fp;
struct intel_context *intel;
GLcontext *ctx;
struct brw_wm_compile *c;
struct brw_shader *shader;
void *mem_ctx;
exec_list instructions;
- int next_abstract_grf;
+
+ int *virtual_grf_sizes;
+ int virtual_grf_next;
+ int virtual_grf_array_size;
+ int *virtual_grf_def;
+ int *virtual_grf_use;
+
struct hash_table *variable_ht;
ir_variable *frag_color, *frag_data, *frag_depth;
int first_non_payload_grf;
+ int urb_setup[FRAG_ATTRIB_MAX];
+ bool kill_emitted;
/** @{ debug annotation info */
const char *current_annotation;
fs_reg pixel_x;
fs_reg pixel_y;
+ fs_reg wpos_w;
fs_reg pixel_w;
fs_reg delta_x;
fs_reg delta_y;
- fs_reg interp_attrs[64];
int grf_used;
};
+int
+fs_visitor::virtual_grf_alloc(int size)
+{
+ if (virtual_grf_array_size <= virtual_grf_next) {
+ if (virtual_grf_array_size == 0)
+ virtual_grf_array_size = 16;
+ else
+ virtual_grf_array_size *= 2;
+ virtual_grf_sizes = talloc_realloc(mem_ctx, virtual_grf_sizes,
+ int, virtual_grf_array_size);
+
+ /* This slot is always unused. */
+ virtual_grf_sizes[0] = 0;
+ }
+ virtual_grf_sizes[virtual_grf_next] = size;
+ return virtual_grf_next++;
+}
+
/** Fixed HW reg constructor. */
fs_reg::fs_reg(enum register_file file, int hw_reg)
{
this->type = BRW_REGISTER_TYPE_F;
}
+int
+brw_type_for_base_type(const struct glsl_type *type)
+{
+ switch (type->base_type) {
+ case GLSL_TYPE_FLOAT:
+ return BRW_REGISTER_TYPE_F;
+ case GLSL_TYPE_INT:
+ case GLSL_TYPE_BOOL:
+ return BRW_REGISTER_TYPE_D;
+ case GLSL_TYPE_UINT:
+ return BRW_REGISTER_TYPE_UD;
+ case GLSL_TYPE_ARRAY:
+ case GLSL_TYPE_STRUCT:
+ /* These should be overridden with the type of the member when
+ * dereferenced into. BRW_REGISTER_TYPE_UD seems like a likely
+ * way to trip up if we don't.
+ */
+ return BRW_REGISTER_TYPE_UD;
+ default:
+ assert(!"not reached");
+ return BRW_REGISTER_TYPE_F;
+ }
+}
+
/** Automatic reg constructor. */
fs_reg::fs_reg(class fs_visitor *v, const struct glsl_type *type)
{
init();
this->file = GRF;
- this->reg = v->next_abstract_grf;
+ this->reg = v->virtual_grf_alloc(type_size(type));
this->reg_offset = 0;
- v->next_abstract_grf += type_size(type);
+ this->type = brw_type_for_base_type(type);
+}
+
+fs_reg *
+fs_visitor::variable_storage(ir_variable *var)
+{
+ return (fs_reg *)hash_table_find(this->variable_ht, var);
+}
+
+/* Our support for uniforms is piggy-backed on the struct
+ * gl_fragment_program, because that's where the values actually
+ * get stored, rather than in some global gl_shader_program uniform
+ * store.
+ */
+int
+fs_visitor::setup_uniform_values(int loc, const glsl_type *type)
+{
+ unsigned int offset = 0;
+ float *vec_values;
+
+ if (type->is_matrix()) {
+ const glsl_type *column = glsl_type::get_instance(GLSL_TYPE_FLOAT,
+ type->vector_elements,
+ 1);
+
+ for (unsigned int i = 0; i < type->matrix_columns; i++) {
+ offset += setup_uniform_values(loc + offset, column);
+ }
+
+ return offset;
+ }
switch (type->base_type) {
case GLSL_TYPE_FLOAT:
- this->type = BRW_REGISTER_TYPE_F;
- break;
+ case GLSL_TYPE_UINT:
case GLSL_TYPE_INT:
case GLSL_TYPE_BOOL:
- this->type = BRW_REGISTER_TYPE_D;
- break;
- case GLSL_TYPE_UINT:
- this->type = BRW_REGISTER_TYPE_UD;
- break;
+ vec_values = fp->Base.Parameters->ParameterValues[loc];
+ for (unsigned int i = 0; i < type->vector_elements; i++) {
+ c->prog_data.param[c->prog_data.nr_params++] = &vec_values[i];
+ }
+ return 1;
+
+ case GLSL_TYPE_STRUCT:
+ for (unsigned int i = 0; i < type->length; i++) {
+ offset += setup_uniform_values(loc + offset,
+ type->fields.structure[i].type);
+ }
+ return offset;
+
+ case GLSL_TYPE_ARRAY:
+ for (unsigned int i = 0; i < type->length; i++) {
+ offset += setup_uniform_values(loc + offset, type->fields.array);
+ }
+ return offset;
+
+ case GLSL_TYPE_SAMPLER:
+ /* The sampler takes up a slot, but we don't use any values from it. */
+ return 1;
+
default:
assert(!"not reached");
- this->type = BRW_REGISTER_TYPE_F;
- break;
+ return 0;
+ }
+}
+
+
+/* Our support for builtin uniforms is even scarier than non-builtin.
+ * It sits on top of the PROG_STATE_VAR parameters that are
+ * automatically updated from GL context state.
+ */
+void
+fs_visitor::setup_builtin_uniform_values(ir_variable *ir)
+{
+ const struct gl_builtin_uniform_desc *statevar = NULL;
+
+ for (unsigned int i = 0; _mesa_builtin_uniform_desc[i].name; i++) {
+ statevar = &_mesa_builtin_uniform_desc[i];
+ if (strcmp(ir->name, _mesa_builtin_uniform_desc[i].name) == 0)
+ break;
+ }
+
+ if (!statevar->name) {
+ this->fail = true;
+ printf("Failed to find builtin uniform `%s'\n", ir->name);
+ return;
+ }
+
+ int array_count;
+ if (ir->type->is_array()) {
+ array_count = ir->type->length;
+ } else {
+ array_count = 1;
+ }
+
+ for (int a = 0; a < array_count; a++) {
+ for (unsigned int i = 0; i < statevar->num_elements; i++) {
+ struct gl_builtin_uniform_element *element = &statevar->elements[i];
+ int tokens[STATE_LENGTH];
+
+ memcpy(tokens, element->tokens, sizeof(element->tokens));
+ if (ir->type->is_array()) {
+ tokens[1] = a;
+ }
+
+ /* This state reference has already been setup by ir_to_mesa,
+ * but we'll get the same index back here.
+ */
+ int index = _mesa_add_state_reference(this->fp->Base.Parameters,
+ (gl_state_index *)tokens);
+ float *vec_values = this->fp->Base.Parameters->ParameterValues[index];
+
+ /* Add each of the unique swizzles of the element as a
+ * parameter. This'll end up matching the expected layout of
+ * the array/matrix/structure we're trying to fill in.
+ */
+ int last_swiz = -1;
+ for (unsigned int i = 0; i < 4; i++) {
+ int swiz = GET_SWZ(element->swizzle, i);
+ if (swiz == last_swiz)
+ break;
+ last_swiz = swiz;
+
+ c->prog_data.param[c->prog_data.nr_params++] = &vec_values[swiz];
+ }
+ }
}
}
fs_reg *
-fs_visitor::variable_storage(ir_variable *var)
+fs_visitor::emit_fragcoord_interpolation(ir_variable *ir)
{
- return (fs_reg *)hash_table_find(this->variable_ht, var);
+ fs_reg *reg = new(this->mem_ctx) fs_reg(this, ir->type);
+ fs_reg wpos = *reg;
+ fs_reg neg_y = this->pixel_y;
+ neg_y.negate = true;
+
+ /* gl_FragCoord.x */
+ if (ir->pixel_center_integer) {
+ emit(fs_inst(BRW_OPCODE_MOV, wpos, this->pixel_x));
+ } else {
+ emit(fs_inst(BRW_OPCODE_ADD, wpos, this->pixel_x, fs_reg(0.5f)));
+ }
+ wpos.reg_offset++;
+
+ /* gl_FragCoord.y */
+ if (ir->origin_upper_left && ir->pixel_center_integer) {
+ emit(fs_inst(BRW_OPCODE_MOV, wpos, this->pixel_y));
+ } else {
+ fs_reg pixel_y = this->pixel_y;
+ float offset = (ir->pixel_center_integer ? 0.0 : 0.5);
+
+ if (!ir->origin_upper_left) {
+ pixel_y.negate = true;
+ offset += c->key.drawable_height - 1.0;
+ }
+
+ emit(fs_inst(BRW_OPCODE_ADD, wpos, pixel_y, fs_reg(offset)));
+ }
+ wpos.reg_offset++;
+
+ /* gl_FragCoord.z */
+ emit(fs_inst(FS_OPCODE_LINTERP, wpos, this->delta_x, this->delta_y,
+ interp_reg(FRAG_ATTRIB_WPOS, 2)));
+ wpos.reg_offset++;
+
+ /* gl_FragCoord.w: Already set up in emit_interpolation */
+ emit(fs_inst(BRW_OPCODE_MOV, wpos, this->wpos_w));
+
+ return reg;
+}
+
+fs_reg *
+fs_visitor::emit_general_interpolation(ir_variable *ir)
+{
+ fs_reg *reg = new(this->mem_ctx) fs_reg(this, ir->type);
+ /* Interpolation is always in floating point regs. */
+ reg->type = BRW_REGISTER_TYPE_F;
+ fs_reg attr = *reg;
+
+ unsigned int array_elements;
+ const glsl_type *type;
+
+ if (ir->type->is_array()) {
+ array_elements = ir->type->length;
+ if (array_elements == 0) {
+ this->fail = true;
+ }
+ type = ir->type->fields.array;
+ } else {
+ array_elements = 1;
+ type = ir->type;
+ }
+
+ int location = ir->location;
+ for (unsigned int i = 0; i < array_elements; i++) {
+ for (unsigned int j = 0; j < type->matrix_columns; j++) {
+ if (urb_setup[location] == -1) {
+ /* If there's no incoming setup data for this slot, don't
+ * emit interpolation for it.
+ */
+ attr.reg_offset += type->vector_elements;
+ location++;
+ continue;
+ }
+
+ for (unsigned int c = 0; c < type->vector_elements; c++) {
+ struct brw_reg interp = interp_reg(location, c);
+ emit(fs_inst(FS_OPCODE_LINTERP,
+ attr,
+ this->delta_x,
+ this->delta_y,
+ fs_reg(interp)));
+ attr.reg_offset++;
+ }
+ attr.reg_offset -= type->vector_elements;
+
+ if (intel->gen < 6) {
+ for (unsigned int c = 0; c < type->vector_elements; c++) {
+ emit(fs_inst(BRW_OPCODE_MUL,
+ attr,
+ attr,
+ this->pixel_w));
+ attr.reg_offset++;
+ }
+ }
+ location++;
+ }
+ }
+
+ return reg;
+}
+
+fs_reg *
+fs_visitor::emit_frontfacing_interpolation(ir_variable *ir)
+{
+ fs_reg *reg = new(this->mem_ctx) fs_reg(this, ir->type);
+
+ /* The frontfacing comes in as a bit in the thread payload. */
+ if (intel->gen >= 6) {
+ emit(fs_inst(BRW_OPCODE_ASR,
+ *reg,
+ fs_reg(retype(brw_vec1_grf(0, 0), BRW_REGISTER_TYPE_D)),
+ fs_reg(15)));
+ emit(fs_inst(BRW_OPCODE_NOT,
+ *reg,
+ *reg));
+ emit(fs_inst(BRW_OPCODE_AND,
+ *reg,
+ *reg,
+ fs_reg(1)));
+ } else {
+ fs_reg *reg = new(this->mem_ctx) fs_reg(this, ir->type);
+ struct brw_reg r1_6ud = retype(brw_vec1_grf(1, 6), BRW_REGISTER_TYPE_UD);
+ /* bit 31 is "primitive is back face", so checking < (1 << 31) gives
+ * us front face
+ */
+ fs_inst *inst = emit(fs_inst(BRW_OPCODE_CMP,
+ *reg,
+ fs_reg(r1_6ud),
+ fs_reg(1u << 31)));
+ inst->conditional_mod = BRW_CONDITIONAL_L;
+ emit(fs_inst(BRW_OPCODE_AND, *reg, *reg, fs_reg(1u)));
+ }
+
+ return reg;
}
void
{
fs_reg *reg = NULL;
+ if (variable_storage(ir))
+ return;
+
if (strcmp(ir->name, "gl_FragColor") == 0) {
this->frag_color = ir;
} else if (strcmp(ir->name, "gl_FragData") == 0) {
this->frag_data = ir;
} else if (strcmp(ir->name, "gl_FragDepth") == 0) {
this->frag_depth = ir;
- assert(!"FINISHME: this hangs currently.");
}
if (ir->mode == ir_var_in) {
- reg = &this->interp_attrs[ir->location];
+ if (!strcmp(ir->name, "gl_FragCoord")) {
+ reg = emit_fragcoord_interpolation(ir);
+ } else if (!strcmp(ir->name, "gl_FrontFacing")) {
+ reg = emit_frontfacing_interpolation(ir);
+ } else {
+ reg = emit_general_interpolation(ir);
+ }
+ assert(reg);
+ hash_table_insert(this->variable_ht, reg, ir);
+ return;
}
if (ir->mode == ir_var_uniform) {
- const float *vec_values;
int param_index = c->prog_data.nr_params;
- /* FINISHME: This is wildly incomplete. */
- assert(ir->type->is_scalar() || ir->type->is_vector() ||
- ir->type->is_sampler());
-
- const struct gl_program *fp = &this->brw->fragment_program->Base;
- /* Our support for uniforms is piggy-backed on the struct
- * gl_fragment_program, because that's where the values actually
- * get stored, rather than in some global gl_shader_program uniform
- * store.
- */
- vec_values = fp->Parameters->ParameterValues[ir->location];
- for (unsigned int i = 0; i < ir->type->vector_elements; i++) {
- c->prog_data.param[c->prog_data.nr_params++] = &vec_values[i];
+ if (!strncmp(ir->name, "gl_", 3)) {
+ setup_builtin_uniform_values(ir);
+ } else {
+ setup_uniform_values(ir->location, ir->type);
}
reg = new(this->mem_ctx) fs_reg(UNIFORM, param_index);
void
fs_visitor::visit(ir_dereference_record *ir)
{
- assert(!"FINISHME");
+ const glsl_type *struct_type = ir->record->type;
+
+ ir->record->accept(this);
+
+ unsigned int offset = 0;
+ for (unsigned int i = 0; i < struct_type->length; i++) {
+ if (strcmp(struct_type->fields.structure[i].name, ir->field) == 0)
+ break;
+ offset += type_size(struct_type->fields.structure[i].type);
+ }
+ this->result.reg_offset += offset;
+ this->result.type = brw_type_for_base_type(ir->type);
}
void
ir->array->accept(this);
index = ir->array_index->as_constant();
- if (ir->type->is_matrix()) {
- element_size = ir->type->vector_elements;
- } else {
- element_size = type_size(ir->type);
- }
+ element_size = type_size(ir->type);
+ this->result.type = brw_type_for_base_type(ir->type);
if (index) {
assert(this->result.file == UNIFORM ||
this->result.reg != 0));
this->result.reg_offset += index->value.i[0] * element_size;
} else {
- assert(!"FINISHME: non-constant matrix column");
+ assert(!"FINISHME: non-constant array element");
}
}
}
}
+void
+fs_visitor::emit_assignment_writes(fs_reg &l, fs_reg &r,
+ const glsl_type *type, bool predicated)
+{
+ switch (type->base_type) {
+ case GLSL_TYPE_FLOAT:
+ case GLSL_TYPE_UINT:
+ case GLSL_TYPE_INT:
+ case GLSL_TYPE_BOOL:
+ for (unsigned int i = 0; i < type->components(); i++) {
+ l.type = brw_type_for_base_type(type);
+ r.type = brw_type_for_base_type(type);
+
+ fs_inst *inst = emit(fs_inst(BRW_OPCODE_MOV, l, r));
+ inst->predicated = predicated;
+
+ l.reg_offset++;
+ r.reg_offset++;
+ }
+ break;
+ case GLSL_TYPE_ARRAY:
+ for (unsigned int i = 0; i < type->length; i++) {
+ emit_assignment_writes(l, r, type->fields.array, predicated);
+ }
+
+ case GLSL_TYPE_STRUCT:
+ for (unsigned int i = 0; i < type->length; i++) {
+ emit_assignment_writes(l, r, type->fields.structure[i].type,
+ predicated);
+ }
+ break;
+
+ case GLSL_TYPE_SAMPLER:
+ break;
+
+ default:
+ assert(!"not reached");
+ break;
+ }
+}
+
void
fs_visitor::visit(ir_assignment *ir)
{
struct fs_reg l, r;
- int i;
- int write_mask;
fs_inst *inst;
/* FINISHME: arrays on the lhs */
ir->rhs->accept(this);
r = this->result;
- /* FINISHME: This should really set to the correct maximal writemask for each
- * FINISHME: component written (in the loops below). This case can only
- * FINISHME: occur for matrices, arrays, and structures.
- */
- if (ir->write_mask == 0) {
- assert(!ir->lhs->type->is_scalar() && !ir->lhs->type->is_vector());
- write_mask = WRITEMASK_XYZW;
- } else {
- assert(ir->lhs->type->is_vector() || ir->lhs->type->is_scalar());
- write_mask = ir->write_mask;
- }
-
assert(l.file != BAD_FILE);
assert(r.file != BAD_FILE);
if (ir->condition) {
/* Get the condition bool into the predicate. */
ir->condition->accept(this);
- inst = emit(fs_inst(BRW_OPCODE_CMP, this->result, fs_reg(0)));
+ inst = emit(fs_inst(BRW_OPCODE_CMP, reg_null, this->result, fs_reg(0)));
inst->conditional_mod = BRW_CONDITIONAL_NZ;
}
- for (i = 0; i < type_size(ir->lhs->type); i++) {
- if (i >= 4 || (write_mask & (1 << i))) {
- inst = emit(fs_inst(BRW_OPCODE_MOV, l, r));
- if (ir->condition)
- inst->predicated = true;
- r.reg_offset++;
+ if (ir->lhs->type->is_scalar() ||
+ ir->lhs->type->is_vector()) {
+ for (int i = 0; i < ir->lhs->type->vector_elements; i++) {
+ if (ir->write_mask & (1 << i)) {
+ inst = emit(fs_inst(BRW_OPCODE_MOV, l, r));
+ if (ir->condition)
+ inst->predicated = true;
+ r.reg_offset++;
+ }
+ l.reg_offset++;
}
- l.reg_offset++;
+ } else {
+ emit_assignment_writes(l, r, ir->lhs->type, ir->condition != NULL);
}
}
-void
-fs_visitor::visit(ir_texture *ir)
+fs_inst *
+fs_visitor::emit_texture_gen4(ir_texture *ir, fs_reg dst, fs_reg coordinate)
{
+ int mlen;
int base_mrf = 2;
- fs_inst *inst = NULL;
- unsigned int mlen = 0;
+ bool simd16 = false;
+ fs_reg orig_dst;
- ir->coordinate->accept(this);
- fs_reg coordinate = this->result;
+ if (ir->shadow_comparitor) {
+ for (mlen = 0; mlen < ir->coordinate->type->vector_elements; mlen++) {
+ emit(fs_inst(BRW_OPCODE_MOV, fs_reg(MRF, base_mrf + mlen),
+ coordinate));
+ coordinate.reg_offset++;
+ }
+ /* gen4's SIMD8 sampler always has the slots for u,v,r present. */
+ mlen = 3;
- if (ir->projector) {
- fs_reg inv_proj = fs_reg(this, glsl_type::float_type);
+ if (ir->op == ir_tex) {
+ /* There's no plain shadow compare message, so we use shadow
+ * compare with a bias of 0.0.
+ */
+ emit(fs_inst(BRW_OPCODE_MOV, fs_reg(MRF, base_mrf + mlen),
+ fs_reg(0.0f)));
+ mlen++;
+ } else if (ir->op == ir_txb) {
+ ir->lod_info.bias->accept(this);
+ emit(fs_inst(BRW_OPCODE_MOV, fs_reg(MRF, base_mrf + mlen),
+ this->result));
+ mlen++;
+ } else {
+ assert(ir->op == ir_txl);
+ ir->lod_info.lod->accept(this);
+ emit(fs_inst(BRW_OPCODE_MOV, fs_reg(MRF, base_mrf + mlen),
+ this->result));
+ mlen++;
+ }
- ir->projector->accept(this);
- emit(fs_inst(FS_OPCODE_RCP, inv_proj, this->result));
+ ir->shadow_comparitor->accept(this);
+ emit(fs_inst(BRW_OPCODE_MOV, fs_reg(MRF, base_mrf + mlen), this->result));
+ mlen++;
+ } else if (ir->op == ir_tex) {
+ for (mlen = 0; mlen < ir->coordinate->type->vector_elements; mlen++) {
+ emit(fs_inst(BRW_OPCODE_MOV, fs_reg(MRF, base_mrf + mlen),
+ coordinate));
+ coordinate.reg_offset++;
+ }
+ /* gen4's SIMD8 sampler always has the slots for u,v,r present. */
+ mlen = 3;
+ } else {
+ /* Oh joy. gen4 doesn't have SIMD8 non-shadow-compare bias/lod
+ * instructions. We'll need to do SIMD16 here.
+ */
+ assert(ir->op == ir_txb || ir->op == ir_txl);
- fs_reg proj_coordinate = fs_reg(this, ir->coordinate->type);
- for (unsigned int i = 0; i < ir->coordinate->type->vector_elements; i++) {
- emit(fs_inst(BRW_OPCODE_MUL, proj_coordinate, coordinate, inv_proj));
+ for (mlen = 0; mlen < ir->coordinate->type->vector_elements * 2;) {
+ emit(fs_inst(BRW_OPCODE_MOV, fs_reg(MRF, base_mrf + mlen),
+ coordinate));
coordinate.reg_offset++;
- proj_coordinate.reg_offset++;
+ mlen++;
+
+ /* The unused upper half. */
+ mlen++;
+ }
+
+ /* lod/bias appears after u/v/r. */
+ mlen = 6;
+
+ if (ir->op == ir_txb) {
+ ir->lod_info.bias->accept(this);
+ emit(fs_inst(BRW_OPCODE_MOV, fs_reg(MRF, base_mrf + mlen),
+ this->result));
+ mlen++;
+ } else {
+ ir->lod_info.lod->accept(this);
+ emit(fs_inst(BRW_OPCODE_MOV, fs_reg(MRF, base_mrf + mlen),
+ this->result));
+ mlen++;
}
- proj_coordinate.reg_offset = 0;
- coordinate = proj_coordinate;
+ /* The unused upper half. */
+ mlen++;
+
+ /* Now, since we're doing simd16, the return is 2 interleaved
+ * vec4s where the odd-indexed ones are junk. We'll need to move
+ * this weirdness around to the expected layout.
+ */
+ simd16 = true;
+ orig_dst = dst;
+ dst = fs_reg(this, glsl_type::get_array_instance(glsl_type::vec4_type,
+ 2));
+ dst.type = BRW_REGISTER_TYPE_F;
+ }
+
+ fs_inst *inst = NULL;
+ switch (ir->op) {
+ case ir_tex:
+ inst = emit(fs_inst(FS_OPCODE_TEX, dst, fs_reg(MRF, base_mrf)));
+ break;
+ case ir_txb:
+ inst = emit(fs_inst(FS_OPCODE_TXB, dst, fs_reg(MRF, base_mrf)));
+ break;
+ case ir_txl:
+ inst = emit(fs_inst(FS_OPCODE_TXL, dst, fs_reg(MRF, base_mrf)));
+ break;
+ case ir_txd:
+ case ir_txf:
+ assert(!"GLSL 1.30 features unsupported");
+ break;
+ }
+ inst->mlen = mlen;
+
+ if (simd16) {
+ for (int i = 0; i < 4; i++) {
+ emit(fs_inst(BRW_OPCODE_MOV, orig_dst, dst));
+ orig_dst.reg_offset++;
+ dst.reg_offset += 2;
+ }
}
+ return inst;
+}
+
+fs_inst *
+fs_visitor::emit_texture_gen5(ir_texture *ir, fs_reg dst, fs_reg coordinate)
+{
+ /* gen5's SIMD8 sampler has slots for u, v, r, array index, then
+ * optional parameters like shadow comparitor or LOD bias. If
+ * optional parameters aren't present, those base slots are
+ * optional and don't need to be included in the message.
+ *
+ * We don't fill in the unnecessary slots regardless, which may
+ * look surprising in the disassembly.
+ */
+ int mlen;
+ int base_mrf = 2;
+
for (mlen = 0; mlen < ir->coordinate->type->vector_elements; mlen++) {
emit(fs_inst(BRW_OPCODE_MOV, fs_reg(MRF, base_mrf + mlen), coordinate));
coordinate.reg_offset++;
}
- /* Pre-Ironlake, the 8-wide sampler always took u,v,r. */
- if (intel->gen < 5)
- mlen = 3;
-
if (ir->shadow_comparitor) {
- /* For shadow comparisons, we have to supply u,v,r. */
- mlen = 3;
+ mlen = MAX2(mlen, 4);
ir->shadow_comparitor->accept(this);
emit(fs_inst(BRW_OPCODE_MOV, fs_reg(MRF, base_mrf + mlen), this->result));
mlen++;
}
- /* Do we ever want to handle writemasking on texture samples? Is it
- * performance relevant?
- */
- fs_reg dst = fs_reg(this, glsl_type::vec4_type);
-
+ fs_inst *inst = NULL;
switch (ir->op) {
case ir_tex:
inst = emit(fs_inst(FS_OPCODE_TEX, dst, fs_reg(MRF, base_mrf)));
break;
case ir_txb:
ir->lod_info.bias->accept(this);
+ mlen = MAX2(mlen, 4);
emit(fs_inst(BRW_OPCODE_MOV, fs_reg(MRF, base_mrf + mlen), this->result));
mlen++;
break;
case ir_txl:
ir->lod_info.lod->accept(this);
+ mlen = MAX2(mlen, 4);
emit(fs_inst(BRW_OPCODE_MOV, fs_reg(MRF, base_mrf + mlen), this->result));
mlen++;
assert(!"GLSL 1.30 features unsupported");
break;
}
-
- this->result = dst;
-
- if (ir->shadow_comparitor)
- inst->shadow_compare = true;
inst->mlen = mlen;
+
+ return inst;
}
void
-fs_visitor::visit(ir_swizzle *ir)
+fs_visitor::visit(ir_texture *ir)
{
- ir->val->accept(this);
- fs_reg val = this->result;
+ fs_inst *inst = NULL;
- fs_reg result = fs_reg(this, ir->type);
- this->result = result;
+ ir->coordinate->accept(this);
+ fs_reg coordinate = this->result;
+
+ /* Should be lowered by do_lower_texture_projection */
+ assert(!ir->projector);
+
+ /* Writemasking doesn't eliminate channels on SIMD8 texture
+ * samples, so don't worry about them.
+ */
+ fs_reg dst = fs_reg(this, glsl_type::vec4_type);
+
+ if (intel->gen < 5) {
+ inst = emit_texture_gen4(ir, dst, coordinate);
+ } else {
+ inst = emit_texture_gen5(ir, dst, coordinate);
+ }
+
+ inst->sampler =
+ _mesa_get_sampler_uniform_value(ir->sampler,
+ ctx->Shader.CurrentProgram,
+ &brw->fragment_program->Base);
+ inst->sampler = c->fp->program.Base.SamplerUnits[inst->sampler];
+
+ this->result = dst;
+
+ if (ir->shadow_comparitor)
+ inst->shadow_compare = true;
+
+ if (c->key.tex_swizzles[inst->sampler] != SWIZZLE_NOOP) {
+ fs_reg swizzle_dst = fs_reg(this, glsl_type::vec4_type);
+
+ for (int i = 0; i < 4; i++) {
+ int swiz = GET_SWZ(c->key.tex_swizzles[inst->sampler], i);
+ fs_reg l = swizzle_dst;
+ l.reg_offset += i;
+
+ if (swiz == SWIZZLE_ZERO) {
+ emit(fs_inst(BRW_OPCODE_MOV, l, fs_reg(0.0f)));
+ } else if (swiz == SWIZZLE_ONE) {
+ emit(fs_inst(BRW_OPCODE_MOV, l, fs_reg(1.0f)));
+ } else {
+ fs_reg r = dst;
+ r.reg_offset += GET_SWZ(c->key.tex_swizzles[inst->sampler], i);
+ emit(fs_inst(BRW_OPCODE_MOV, l, r));
+ }
+ }
+ this->result = swizzle_dst;
+ }
+}
+
+void
+fs_visitor::visit(ir_swizzle *ir)
+{
+ ir->val->accept(this);
+ fs_reg val = this->result;
+
+ if (ir->type->vector_elements == 1) {
+ this->result.reg_offset += ir->mask.x;
+ return;
+ }
+
+ fs_reg result = fs_reg(this, ir->type);
+ this->result = result;
for (unsigned int i = 0; i < ir->type->vector_elements; i++) {
fs_reg channel = val;
void
fs_visitor::visit(ir_discard *ir)
{
+ fs_reg temp = fs_reg(this, glsl_type::uint_type);
+
assert(ir->condition == NULL); /* FINISHME */
- emit(fs_inst(FS_OPCODE_DISCARD));
+ emit(fs_inst(FS_OPCODE_DISCARD, temp, temp));
+ kill_emitted = true;
}
void
void
fs_visitor::visit(ir_loop *ir)
{
- assert(!ir->from);
- assert(!ir->to);
- assert(!ir->increment);
- assert(!ir->counter);
+ fs_reg counter = reg_undef;
+
+ if (ir->counter) {
+ this->base_ir = ir->counter;
+ ir->counter->accept(this);
+ counter = *(variable_storage(ir->counter));
+
+ if (ir->from) {
+ this->base_ir = ir->from;
+ ir->from->accept(this);
+
+ emit(fs_inst(BRW_OPCODE_MOV, counter, this->result));
+ }
+ }
emit(fs_inst(BRW_OPCODE_DO));
- /* Start a safety counter. If the user messed up their loop
- * counting, we don't want to hang the GPU.
- */
- fs_reg max_iter = fs_reg(this, glsl_type::int_type);
- emit(fs_inst(BRW_OPCODE_MOV, max_iter, fs_reg(10000)));
+ if (ir->to) {
+ this->base_ir = ir->to;
+ ir->to->accept(this);
+
+ fs_inst *inst = emit(fs_inst(BRW_OPCODE_CMP, reg_null,
+ counter, this->result));
+ switch (ir->cmp) {
+ case ir_binop_equal:
+ inst->conditional_mod = BRW_CONDITIONAL_Z;
+ break;
+ case ir_binop_nequal:
+ inst->conditional_mod = BRW_CONDITIONAL_NZ;
+ break;
+ case ir_binop_gequal:
+ inst->conditional_mod = BRW_CONDITIONAL_GE;
+ break;
+ case ir_binop_lequal:
+ inst->conditional_mod = BRW_CONDITIONAL_LE;
+ break;
+ case ir_binop_greater:
+ inst->conditional_mod = BRW_CONDITIONAL_G;
+ break;
+ case ir_binop_less:
+ inst->conditional_mod = BRW_CONDITIONAL_L;
+ break;
+ default:
+ assert(!"not reached: unknown loop condition");
+ this->fail = true;
+ break;
+ }
+
+ inst = emit(fs_inst(BRW_OPCODE_BREAK));
+ inst->predicated = true;
+ }
foreach_iter(exec_list_iterator, iter, ir->body_instructions) {
ir_instruction *ir = (ir_instruction *)iter.get();
- fs_inst *inst;
this->base_ir = ir;
ir->accept(this);
+ }
- /* Check the maximum loop iters counter. */
- inst = emit(fs_inst(BRW_OPCODE_ADD, max_iter, max_iter, fs_reg(-1)));
- inst->conditional_mod = BRW_CONDITIONAL_Z;
-
- inst = emit(fs_inst(BRW_OPCODE_BREAK));
- inst->predicated = true;
+ if (ir->increment) {
+ this->base_ir = ir->increment;
+ ir->increment->accept(this);
+ emit(fs_inst(BRW_OPCODE_ADD, counter, counter, this->result));
}
emit(fs_inst(BRW_OPCODE_WHILE));
struct brw_reg
fs_visitor::interp_reg(int location, int channel)
{
- int regnr = location * 2 + channel / 2;
+ int regnr = urb_setup[location] * 2 + channel / 2;
int stride = (channel & 1) * 4;
+ assert(urb_setup[location] != -1);
+
return brw_vec1_grf(regnr, stride);
}
/** Emits the interpolation for the varying inputs. */
void
-fs_visitor::emit_interpolation()
+fs_visitor::emit_interpolation_setup_gen4()
{
struct brw_reg g1_uw = retype(brw_vec1_grf(1, 0), BRW_REGISTER_TYPE_UW);
- /* For now, the source regs for the setup URB data will be unset,
- * since we don't know until codegen how many push constants we'll
- * use, and therefore what the setup URB offset is.
- */
- fs_reg src_reg = reg_undef;
this->current_annotation = "compute pixel centers";
this->pixel_x = fs_reg(this, glsl_type::uint_type);
fs_reg(brw_imm_v(0x11001100))));
this->current_annotation = "compute pixel deltas from v0";
- this->delta_x = fs_reg(this, glsl_type::float_type);
- this->delta_y = fs_reg(this, glsl_type::float_type);
+ if (brw->has_pln) {
+ this->delta_x = fs_reg(this, glsl_type::vec2_type);
+ this->delta_y = this->delta_x;
+ this->delta_y.reg_offset++;
+ } else {
+ this->delta_x = fs_reg(this, glsl_type::float_type);
+ this->delta_y = fs_reg(this, glsl_type::float_type);
+ }
emit(fs_inst(BRW_OPCODE_ADD,
this->delta_x,
this->pixel_x,
emit(fs_inst(BRW_OPCODE_ADD,
this->delta_y,
this->pixel_y,
- fs_reg(brw_vec1_grf(1, 1))));
+ fs_reg(negate(brw_vec1_grf(1, 1)))));
this->current_annotation = "compute pos.w and 1/pos.w";
- /* Compute wpos. Unlike many other varying inputs, we usually need it
- * to produce 1/w, and the varying variable wouldn't show up.
+ /* Compute wpos.w. It's always in our setup, since it's needed to
+ * interpolate the other attributes.
*/
- fs_reg wpos = fs_reg(this, glsl_type::vec4_type);
- this->interp_attrs[FRAG_ATTRIB_WPOS] = wpos;
- emit(fs_inst(BRW_OPCODE_MOV, wpos, this->pixel_x)); /* FINISHME: ARB_fcc */
- wpos.reg_offset++;
- emit(fs_inst(BRW_OPCODE_MOV, wpos, this->pixel_y)); /* FINISHME: ARB_fcc */
- wpos.reg_offset++;
- emit(fs_inst(FS_OPCODE_LINTERP, wpos, this->delta_x, this->delta_y,
- interp_reg(FRAG_ATTRIB_WPOS, 2)));
- wpos.reg_offset++;
- emit(fs_inst(FS_OPCODE_LINTERP, wpos, this->delta_x, this->delta_y,
+ this->wpos_w = fs_reg(this, glsl_type::float_type);
+ emit(fs_inst(FS_OPCODE_LINTERP, wpos_w, this->delta_x, this->delta_y,
interp_reg(FRAG_ATTRIB_WPOS, 3)));
- /* Compute the pixel W value from wpos.w. */
+ /* Compute the pixel 1/W value from wpos.w. */
this->pixel_w = fs_reg(this, glsl_type::float_type);
- emit(fs_inst(FS_OPCODE_RCP, this->pixel_w, wpos));
-
- /* FINISHME: gl_FrontFacing */
+ emit(fs_inst(FS_OPCODE_RCP, this->pixel_w, wpos_w));
+ this->current_annotation = NULL;
+}
- foreach_iter(exec_list_iterator, iter, *this->shader->ir) {
- ir_instruction *ir = (ir_instruction *)iter.get();
- ir_variable *var = ir->as_variable();
+/** Emits the interpolation for the varying inputs. */
+void
+fs_visitor::emit_interpolation_setup_gen6()
+{
+ struct brw_reg g1_uw = retype(brw_vec1_grf(1, 0), BRW_REGISTER_TYPE_UW);
- if (!var)
- continue;
+ /* If the pixel centers end up used, the setup is the same as for gen4. */
+ this->current_annotation = "compute pixel centers";
+ this->pixel_x = fs_reg(this, glsl_type::uint_type);
+ this->pixel_y = fs_reg(this, glsl_type::uint_type);
+ this->pixel_x.type = BRW_REGISTER_TYPE_UW;
+ this->pixel_y.type = BRW_REGISTER_TYPE_UW;
+ emit(fs_inst(BRW_OPCODE_ADD,
+ this->pixel_x,
+ fs_reg(stride(suboffset(g1_uw, 4), 2, 4, 0)),
+ fs_reg(brw_imm_v(0x10101010))));
+ emit(fs_inst(BRW_OPCODE_ADD,
+ this->pixel_y,
+ fs_reg(stride(suboffset(g1_uw, 5), 2, 4, 0)),
+ fs_reg(brw_imm_v(0x11001100))));
- if (var->mode != ir_var_in)
- continue;
+ this->current_annotation = "compute 1/pos.w";
+ this->wpos_w = fs_reg(brw_vec8_grf(c->key.source_w_reg, 0));
+ this->pixel_w = fs_reg(this, glsl_type::float_type);
+ emit(fs_inst(FS_OPCODE_RCP, this->pixel_w, wpos_w));
- /* If it's already set up (WPOS), skip. */
- if (var->location == 0)
- continue;
+ this->delta_x = fs_reg(brw_vec8_grf(2, 0));
+ this->delta_y = fs_reg(brw_vec8_grf(3, 0));
- this->current_annotation = talloc_asprintf(this->mem_ctx,
- "interpolate %s "
- "(FRAG_ATTRIB[%d])",
- var->name,
- var->location);
- emit_pinterp(var->location);
- }
this->current_annotation = NULL;
}
void
-fs_visitor::emit_pinterp(int location)
+fs_visitor::emit_fb_writes()
{
- fs_reg interp_attr = fs_reg(this, glsl_type::vec4_type);
- this->interp_attrs[location] = interp_attr;
-
- for (unsigned int i = 0; i < 4; i++) {
- struct brw_reg interp = interp_reg(location, i);
- emit(fs_inst(FS_OPCODE_LINTERP,
- interp_attr,
- this->delta_x,
- this->delta_y,
- fs_reg(interp)));
- interp_attr.reg_offset++;
- }
- interp_attr.reg_offset -= 4;
-
- for (unsigned int i = 0; i < 4; i++) {
- emit(fs_inst(BRW_OPCODE_MUL,
- interp_attr,
- interp_attr,
- this->pixel_w));
- interp_attr.reg_offset++;
+ this->current_annotation = "FB write header";
+ GLboolean header_present = GL_TRUE;
+ int nr = 0;
+
+ if (intel->gen >= 6 &&
+ !this->kill_emitted &&
+ c->key.nr_color_regions == 1) {
+ header_present = false;
}
-}
-void
-fs_visitor::emit_fb_writes()
-{
- this->current_annotation = "FB write";
+ if (header_present) {
+ /* m0, m1 header */
+ nr += 2;
+ }
+
+ if (c->key.aa_dest_stencil_reg) {
+ emit(fs_inst(BRW_OPCODE_MOV, fs_reg(MRF, nr++),
+ fs_reg(brw_vec8_grf(c->key.aa_dest_stencil_reg, 0))));
+ }
+
+ /* Reserve space for color. It'll be filled in per MRT below. */
+ int color_mrf = nr;
+ nr += 4;
+
+ if (c->key.source_depth_to_render_target) {
+ if (c->key.computes_depth) {
+ /* Hand over gl_FragDepth. */
+ assert(this->frag_depth);
+ fs_reg depth = *(variable_storage(this->frag_depth));
+
+ emit(fs_inst(BRW_OPCODE_MOV, fs_reg(MRF, nr++), depth));
+ } else {
+ /* Pass through the payload depth. */
+ emit(fs_inst(BRW_OPCODE_MOV, fs_reg(MRF, nr++),
+ fs_reg(brw_vec8_grf(c->key.source_depth_reg, 0))));
+ }
+ }
+
+ if (c->key.dest_depth_reg) {
+ emit(fs_inst(BRW_OPCODE_MOV, fs_reg(MRF, nr++),
+ fs_reg(brw_vec8_grf(c->key.dest_depth_reg, 0))));
+ }
+
+ fs_reg color = reg_undef;
+ if (this->frag_color)
+ color = *(variable_storage(this->frag_color));
+ else if (this->frag_data)
+ color = *(variable_storage(this->frag_data));
+
+ for (int target = 0; target < c->key.nr_color_regions; target++) {
+ this->current_annotation = talloc_asprintf(this->mem_ctx,
+ "FB write target %d",
+ target);
+ if (this->frag_color || this->frag_data) {
+ for (int i = 0; i < 4; i++) {
+ emit(fs_inst(BRW_OPCODE_MOV,
+ fs_reg(MRF, color_mrf + i),
+ color));
+ color.reg_offset++;
+ }
+ }
- assert(this->frag_color || !"FINISHME: MRT");
- fs_reg color = *(variable_storage(this->frag_color));
+ if (this->frag_color)
+ color.reg_offset -= 4;
- for (int i = 0; i < 4; i++) {
- emit(fs_inst(BRW_OPCODE_MOV,
- fs_reg(MRF, 2 + i),
- color));
- color.reg_offset++;
+ fs_inst *inst = emit(fs_inst(FS_OPCODE_FB_WRITE,
+ reg_undef, reg_undef));
+ inst->target = target;
+ inst->mlen = nr;
+ if (target == c->key.nr_color_regions - 1)
+ inst->eot = true;
+ inst->header_present = header_present;
}
- emit(fs_inst(FS_OPCODE_FB_WRITE,
- fs_reg(0),
- fs_reg(0)));
+ if (c->key.nr_color_regions == 0) {
+ fs_inst *inst = emit(fs_inst(FS_OPCODE_FB_WRITE,
+ reg_undef, reg_undef));
+ inst->mlen = nr;
+ inst->eot = true;
+ inst->header_present = header_present;
+ }
this->current_annotation = NULL;
}
void
fs_visitor::generate_fb_write(fs_inst *inst)
{
- GLboolean eot = 1; /* FINISHME: MRT */
- /* FINISHME: AADS */
+ GLboolean eot = inst->eot;
+ struct brw_reg implied_header;
/* Header is 2 regs, g0 and g1 are the contents. g0 will be implied
* move, here's g1.
brw_push_insn_state(p);
brw_set_mask_control(p, BRW_MASK_DISABLE);
brw_set_compression_control(p, BRW_COMPRESSION_NONE);
- brw_MOV(p,
- brw_message_reg(1),
- brw_vec8_grf(1, 0));
- brw_pop_insn_state(p);
- int nr = 2 + 4;
+ if (inst->header_present) {
+ if (intel->gen >= 6) {
+ brw_MOV(p,
+ brw_message_reg(0),
+ brw_vec8_grf(0, 0));
+ implied_header = brw_null_reg();
+ } else {
+ implied_header = retype(brw_vec8_grf(0, 0), BRW_REGISTER_TYPE_UW);
+ }
+
+ brw_MOV(p,
+ brw_message_reg(1),
+ brw_vec8_grf(1, 0));
+ } else {
+ implied_header = brw_null_reg();
+ }
+
+ brw_pop_insn_state(p);
brw_fb_WRITE(p,
8, /* dispatch_width */
retype(vec8(brw_null_reg()), BRW_REGISTER_TYPE_UW),
0, /* base MRF */
- retype(brw_vec8_grf(0, 0), BRW_REGISTER_TYPE_UW),
- 0, /* FINISHME: MRT target */
- nr,
+ implied_header,
+ inst->target,
+ inst->mlen,
0,
eot);
}
{
int msg_type = -1;
int rlen = 4;
+ uint32_t simd_mode = BRW_SAMPLER_SIMD_MODE_SIMD8;
- if (intel->gen == 5) {
+ if (intel->gen >= 5) {
switch (inst->opcode) {
case FS_OPCODE_TEX:
if (inst->shadow_compare) {
/* Note that G45 and older determines shadow compare and dispatch width
* from message length for most messages.
*/
+ msg_type = BRW_SAMPLER_MESSAGE_SIMD8_SAMPLE;
if (inst->shadow_compare) {
- msg_type = BRW_SAMPLER_MESSAGE_SIMD16_SAMPLE_COMPARE;
+ assert(inst->mlen == 5);
} else {
- msg_type = BRW_SAMPLER_MESSAGE_SIMD16_SAMPLE;
+ assert(inst->mlen <= 6);
}
+ break;
case FS_OPCODE_TXB:
if (inst->shadow_compare) {
- assert(!"FINISHME: shadow compare with bias.");
- msg_type = BRW_SAMPLER_MESSAGE_SIMD16_SAMPLE_BIAS;
+ assert(inst->mlen == 5);
+ msg_type = BRW_SAMPLER_MESSAGE_SIMD8_SAMPLE;
} else {
+ assert(inst->mlen == 8);
msg_type = BRW_SAMPLER_MESSAGE_SIMD16_SAMPLE_BIAS;
- rlen = 8;
+ simd_mode = BRW_SAMPLER_SIMD_MODE_SIMD16;
}
break;
}
}
assert(msg_type != -1);
+ if (simd_mode == BRW_SAMPLER_SIMD_MODE_SIMD16) {
+ rlen = 8;
+ dst = vec16(dst);
+ }
+
/* g0 header. */
src.nr--;
inst->mlen + 1,
0,
1,
- BRW_SAMPLER_SIMD_MODE_SIMD8);
+ simd_mode);
}
}
void
-fs_visitor::generate_discard(fs_inst *inst)
+fs_visitor::generate_discard(fs_inst *inst, struct brw_reg temp)
{
struct brw_reg g0 = retype(brw_vec1_grf(0, 0), BRW_REGISTER_TYPE_UW);
+ temp = brw_uw1_reg(temp.file, temp.nr, 0);
+
brw_push_insn_state(p);
brw_set_mask_control(p, BRW_MASK_DISABLE);
- brw_NOT(p, c->emit_mask_reg, brw_mask_reg(1)); /* IMASK */
- brw_AND(p, g0, c->emit_mask_reg, g0);
+ brw_NOT(p, temp, brw_mask_reg(1)); /* IMASK */
+ brw_AND(p, g0, temp, g0);
brw_pop_insn_state(p);
}
-static void
-trivial_assign_reg(int header_size, fs_reg *reg)
-{
- if (reg->file == GRF && reg->reg != 0) {
- reg->hw_reg = header_size + reg->reg - 1 + reg->reg_offset;
- reg->reg = 0;
- }
-}
-
void
fs_visitor::assign_curb_setup()
{
c->prog_data.first_curbe_grf = c->key.nr_payload_regs;
c->prog_data.curb_read_length = ALIGN(c->prog_data.nr_params, 8) / 8;
- if (intel->gen == 5 && (c->prog_data.first_curbe_grf +
- c->prog_data.curb_read_length) & 1) {
- /* Align the start of the interpolation coefficients so that we can use
- * the PLN instruction.
- */
- c->prog_data.first_curbe_grf++;
- }
-
/* Map the offsets in the UNIFORM file to fixed HW regs. */
foreach_iter(exec_list_iterator, iter, this->instructions) {
fs_inst *inst = (fs_inst *)iter.get();
}
void
-fs_visitor::assign_urb_setup()
+fs_visitor::calculate_urb_setup()
{
- int urb_start = c->prog_data.first_curbe_grf + c->prog_data.curb_read_length;
- int interp_reg_nr[FRAG_ATTRIB_MAX];
-
- c->prog_data.urb_read_length = 0;
+ for (unsigned int i = 0; i < FRAG_ATTRIB_MAX; i++) {
+ urb_setup[i] = -1;
+ }
+ int urb_next = 0;
/* Figure out where each of the incoming setup attributes lands. */
- for (unsigned int i = 0; i < FRAG_ATTRIB_MAX; i++) {
- interp_reg_nr[i] = -1;
+ if (intel->gen >= 6) {
+ for (unsigned int i = 0; i < FRAG_ATTRIB_MAX; i++) {
+ if (brw->fragment_program->Base.InputsRead & BITFIELD64_BIT(i)) {
+ urb_setup[i] = urb_next++;
+ }
+ }
+ } else {
+ /* FINISHME: The sf doesn't map VS->FS inputs for us very well. */
+ for (unsigned int i = 0; i < VERT_RESULT_MAX; i++) {
+ if (c->key.vp_outputs_written & BITFIELD64_BIT(i)) {
+ int fp_index;
+
+ if (i >= VERT_RESULT_VAR0)
+ fp_index = i - (VERT_RESULT_VAR0 - FRAG_ATTRIB_VAR0);
+ else if (i <= VERT_RESULT_TEX7)
+ fp_index = i;
+ else
+ fp_index = -1;
+
+ if (fp_index >= 0)
+ urb_setup[fp_index] = urb_next++;
+ }
+ }
+ }
- if (i != FRAG_ATTRIB_WPOS &&
- !(brw->fragment_program->Base.InputsRead & BITFIELD64_BIT(i)))
- continue;
+ /* Each attribute is 4 setup channels, each of which is half a reg. */
+ c->prog_data.urb_read_length = urb_next * 2;
+}
- /* Each attribute is 4 setup channels, each of which is half a reg. */
- interp_reg_nr[i] = urb_start + c->prog_data.urb_read_length;
- c->prog_data.urb_read_length += 2;
- }
+void
+fs_visitor::assign_urb_setup()
+{
+ int urb_start = c->prog_data.first_curbe_grf + c->prog_data.curb_read_length;
- /* Map the register numbers for FS_OPCODE_LINTERP so that it uses
- * the correct setup input.
+ /* Offset all the urb_setup[] index by the actual position of the
+ * setup regs, now that the location of the constants has been chosen.
*/
foreach_iter(exec_list_iterator, iter, this->instructions) {
fs_inst *inst = (fs_inst *)iter.get();
assert(inst->src[2].file == FIXED_HW_REG);
- int location = inst->src[2].fixed_hw_reg.nr / 2;
- assert(interp_reg_nr[location] != -1);
- inst->src[2].fixed_hw_reg.nr = (interp_reg_nr[location] +
- (inst->src[2].fixed_hw_reg.nr & 1));
+ inst->src[2].fixed_hw_reg.nr += urb_start;
}
this->first_non_payload_grf = urb_start + c->prog_data.urb_read_length;
}
+static void
+assign_reg(int *reg_hw_locations, fs_reg *reg)
+{
+ if (reg->file == GRF && reg->reg != 0) {
+ reg->hw_reg = reg_hw_locations[reg->reg] + reg->reg_offset;
+ reg->reg = 0;
+ }
+}
+
void
-fs_visitor::assign_regs()
+fs_visitor::assign_regs_trivial()
{
- int header_size = this->first_non_payload_grf;
int last_grf = 0;
+ int hw_reg_mapping[this->virtual_grf_next];
+ int i;
+
+ hw_reg_mapping[0] = 0;
+ hw_reg_mapping[1] = this->first_non_payload_grf;
+ for (i = 2; i < this->virtual_grf_next; i++) {
+ hw_reg_mapping[i] = (hw_reg_mapping[i - 1] +
+ this->virtual_grf_sizes[i - 1]);
+ }
+ last_grf = hw_reg_mapping[i - 1] + this->virtual_grf_sizes[i - 1];
- /* FINISHME: trivial assignment of register numbers */
foreach_iter(exec_list_iterator, iter, this->instructions) {
fs_inst *inst = (fs_inst *)iter.get();
- trivial_assign_reg(header_size, &inst->dst);
- trivial_assign_reg(header_size, &inst->src[0]);
- trivial_assign_reg(header_size, &inst->src[1]);
+ assign_reg(hw_reg_mapping, &inst->dst);
+ assign_reg(hw_reg_mapping, &inst->src[0]);
+ assign_reg(hw_reg_mapping, &inst->src[1]);
+ }
+
+ this->grf_used = last_grf + 1;
+}
+
+void
+fs_visitor::assign_regs()
+{
+ int last_grf = 0;
+ int hw_reg_mapping[this->virtual_grf_next + 1];
+ int base_reg_count = BRW_MAX_GRF - this->first_non_payload_grf;
+ int class_sizes[base_reg_count];
+ int class_count = 0;
+ int aligned_pair_class = -1;
+
+ /* Set up the register classes.
+ *
+ * The base registers store a scalar value. For texture samples,
+ * we get virtual GRFs composed of 4 contiguous hw register. For
+ * structures and arrays, we store them as contiguous larger things
+ * than that, though we should be able to do better most of the
+ * time.
+ */
+ class_sizes[class_count++] = 1;
+ if (brw->has_pln && intel->gen < 6) {
+ /* Always set up the (unaligned) pairs for gen5, so we can find
+ * them for making the aligned pair class.
+ */
+ class_sizes[class_count++] = 2;
+ }
+ for (int r = 1; r < this->virtual_grf_next; r++) {
+ int i;
+
+ for (i = 0; i < class_count; i++) {
+ if (class_sizes[i] == this->virtual_grf_sizes[r])
+ break;
+ }
+ if (i == class_count) {
+ if (this->virtual_grf_sizes[r] >= base_reg_count) {
+ fprintf(stderr, "Object too large to register allocate.\n");
+ this->fail = true;
+ }
+
+ class_sizes[class_count++] = this->virtual_grf_sizes[r];
+ }
+ }
+
+ int ra_reg_count = 0;
+ int class_base_reg[class_count];
+ int class_reg_count[class_count];
+ int classes[class_count + 1];
- last_grf = MAX2(last_grf, inst->dst.hw_reg);
- last_grf = MAX2(last_grf, inst->src[0].hw_reg);
- last_grf = MAX2(last_grf, inst->src[1].hw_reg);
+ for (int i = 0; i < class_count; i++) {
+ class_base_reg[i] = ra_reg_count;
+ class_reg_count[i] = base_reg_count - (class_sizes[i] - 1);
+ ra_reg_count += class_reg_count[i];
+ }
+
+ struct ra_regs *regs = ra_alloc_reg_set(ra_reg_count);
+ for (int i = 0; i < class_count; i++) {
+ classes[i] = ra_alloc_reg_class(regs);
+
+ for (int i_r = 0; i_r < class_reg_count[i]; i_r++) {
+ ra_class_add_reg(regs, classes[i], class_base_reg[i] + i_r);
+ }
+
+ /* Add conflicts between our contiguous registers aliasing
+ * base regs and other register classes' contiguous registers
+ * that alias base regs, or the base regs themselves for classes[0].
+ */
+ for (int c = 0; c <= i; c++) {
+ for (int i_r = 0; i_r < class_reg_count[i]; i_r++) {
+ for (int c_r = MAX2(0, i_r - (class_sizes[c] - 1));
+ c_r < MIN2(class_reg_count[c], i_r + class_sizes[i]);
+ c_r++) {
+
+ if (0) {
+ printf("%d/%d conflicts %d/%d\n",
+ class_sizes[i], this->first_non_payload_grf + i_r,
+ class_sizes[c], this->first_non_payload_grf + c_r);
+ }
+
+ ra_add_reg_conflict(regs,
+ class_base_reg[i] + i_r,
+ class_base_reg[c] + c_r);
+ }
+ }
+ }
+ }
+
+ /* Add a special class for aligned pairs, which we'll put delta_x/y
+ * in on gen5 so that we can do PLN.
+ */
+ if (brw->has_pln && intel->gen < 6) {
+ int reg_count = (base_reg_count - 1) / 2;
+ int unaligned_pair_class = 1;
+ assert(class_sizes[unaligned_pair_class] == 2);
+
+ aligned_pair_class = class_count;
+ classes[aligned_pair_class] = ra_alloc_reg_class(regs);
+ class_base_reg[aligned_pair_class] = 0;
+ class_reg_count[aligned_pair_class] = 0;
+ int start = (this->first_non_payload_grf & 1) ? 1 : 0;
+
+ for (int i = 0; i < reg_count; i++) {
+ ra_class_add_reg(regs, classes[aligned_pair_class],
+ class_base_reg[unaligned_pair_class] + i * 2 + start);
+ }
+ class_count++;
+ }
+
+ ra_set_finalize(regs);
+
+ struct ra_graph *g = ra_alloc_interference_graph(regs,
+ this->virtual_grf_next);
+ /* Node 0 is just a placeholder to keep virtual_grf[] mapping 1:1
+ * with nodes.
+ */
+ ra_set_node_class(g, 0, classes[0]);
+
+ for (int i = 1; i < this->virtual_grf_next; i++) {
+ for (int c = 0; c < class_count; c++) {
+ if (class_sizes[c] == this->virtual_grf_sizes[i]) {
+ if (aligned_pair_class >= 0 &&
+ this->delta_x.reg == i) {
+ ra_set_node_class(g, i, classes[aligned_pair_class]);
+ } else {
+ ra_set_node_class(g, i, classes[c]);
+ }
+ break;
+ }
+ }
+
+ for (int j = 1; j < i; j++) {
+ if (virtual_grf_interferes(i, j)) {
+ ra_add_node_interference(g, i, j);
+ }
+ }
+ }
+
+ /* FINISHME: Handle spilling */
+ if (!ra_allocate_no_spills(g)) {
+ fprintf(stderr, "Failed to allocate registers.\n");
+ this->fail = true;
+ return;
+ }
+
+ /* Get the chosen virtual registers for each node, and map virtual
+ * regs in the register classes back down to real hardware reg
+ * numbers.
+ */
+ hw_reg_mapping[0] = 0; /* unused */
+ for (int i = 1; i < this->virtual_grf_next; i++) {
+ int reg = ra_get_node_reg(g, i);
+ int hw_reg = -1;
+
+ for (int c = 0; c < class_count; c++) {
+ if (reg >= class_base_reg[c] &&
+ reg < class_base_reg[c] + class_reg_count[c]) {
+ hw_reg = reg - class_base_reg[c];
+ break;
+ }
+ }
+
+ assert(hw_reg != -1);
+ hw_reg_mapping[i] = this->first_non_payload_grf + hw_reg;
+ last_grf = MAX2(last_grf,
+ hw_reg_mapping[i] + this->virtual_grf_sizes[i] - 1);
+ }
+
+ foreach_iter(exec_list_iterator, iter, this->instructions) {
+ fs_inst *inst = (fs_inst *)iter.get();
+
+ assign_reg(hw_reg_mapping, &inst->dst);
+ assign_reg(hw_reg_mapping, &inst->src[0]);
+ assign_reg(hw_reg_mapping, &inst->src[1]);
}
this->grf_used = last_grf + 1;
+
+ talloc_free(g);
+ talloc_free(regs);
+}
+
+void
+fs_visitor::calculate_live_intervals()
+{
+ int num_vars = this->virtual_grf_next;
+ int *def = talloc_array(mem_ctx, int, num_vars);
+ int *use = talloc_array(mem_ctx, int, num_vars);
+ int loop_depth = 0;
+ int loop_start = 0;
+
+ for (int i = 0; i < num_vars; i++) {
+ def[i] = 1 << 30;
+ use[i] = -1;
+ }
+
+ int ip = 0;
+ foreach_iter(exec_list_iterator, iter, this->instructions) {
+ fs_inst *inst = (fs_inst *)iter.get();
+
+ if (inst->opcode == BRW_OPCODE_DO) {
+ if (loop_depth++ == 0)
+ loop_start = ip;
+ } else if (inst->opcode == BRW_OPCODE_WHILE) {
+ loop_depth--;
+
+ if (loop_depth == 0) {
+ /* FINISHME:
+ *
+ * Patches up any vars marked for use within the loop as
+ * live until the end. This is conservative, as there
+ * will often be variables defined and used inside the
+ * loop but dead at the end of the loop body.
+ */
+ for (int i = 0; i < num_vars; i++) {
+ if (use[i] == loop_start) {
+ use[i] = ip;
+ }
+ }
+ }
+ } else {
+ int eip = ip;
+
+ if (loop_depth)
+ eip = loop_start;
+
+ for (unsigned int i = 0; i < 3; i++) {
+ if (inst->src[i].file == GRF && inst->src[i].reg != 0) {
+ use[inst->src[i].reg] = MAX2(use[inst->src[i].reg], eip);
+ }
+ }
+ if (inst->dst.file == GRF && inst->dst.reg != 0) {
+ def[inst->dst.reg] = MIN2(def[inst->dst.reg], eip);
+ }
+ }
+
+ ip++;
+ }
+
+ talloc_free(this->virtual_grf_def);
+ talloc_free(this->virtual_grf_use);
+ this->virtual_grf_def = def;
+ this->virtual_grf_use = use;
+}
+
+/**
+ * Attempts to move immediate constants into the immediate
+ * constant slot of following instructions.
+ *
+ * Immediate constants are a bit tricky -- they have to be in the last
+ * operand slot, you can't do abs/negate on them,
+ */
+
+bool
+fs_visitor::propagate_constants()
+{
+ bool progress = false;
+
+ foreach_iter(exec_list_iterator, iter, this->instructions) {
+ fs_inst *inst = (fs_inst *)iter.get();
+
+ if (inst->opcode != BRW_OPCODE_MOV ||
+ inst->predicated ||
+ inst->dst.file != GRF || inst->src[0].file != IMM ||
+ inst->dst.type != inst->src[0].type)
+ continue;
+
+ /* Don't bother with cases where we should have had the
+ * operation on the constant folded in GLSL already.
+ */
+ if (inst->saturate)
+ continue;
+
+ /* Found a move of a constant to a GRF. Find anything else using the GRF
+ * before it's written, and replace it with the constant if we can.
+ */
+ exec_list_iterator scan_iter = iter;
+ scan_iter.next();
+ for (; scan_iter.has_next(); scan_iter.next()) {
+ fs_inst *scan_inst = (fs_inst *)scan_iter.get();
+
+ if (scan_inst->opcode == BRW_OPCODE_DO ||
+ scan_inst->opcode == BRW_OPCODE_WHILE ||
+ scan_inst->opcode == BRW_OPCODE_ELSE ||
+ scan_inst->opcode == BRW_OPCODE_ENDIF) {
+ break;
+ }
+
+ for (int i = 2; i >= 0; i--) {
+ if (scan_inst->src[i].file != GRF ||
+ scan_inst->src[i].reg != inst->dst.reg ||
+ scan_inst->src[i].reg_offset != inst->dst.reg_offset)
+ continue;
+
+ /* Don't bother with cases where we should have had the
+ * operation on the constant folded in GLSL already.
+ */
+ if (scan_inst->src[i].negate || scan_inst->src[i].abs)
+ continue;
+
+ switch (scan_inst->opcode) {
+ case BRW_OPCODE_MOV:
+ scan_inst->src[i] = inst->src[0];
+ progress = true;
+ break;
+
+ case BRW_OPCODE_MUL:
+ case BRW_OPCODE_ADD:
+ if (i == 1) {
+ scan_inst->src[i] = inst->src[0];
+ progress = true;
+ } else if (i == 0 && scan_inst->src[1].file != IMM) {
+ /* Fit this constant in by commuting the operands */
+ scan_inst->src[0] = scan_inst->src[1];
+ scan_inst->src[1] = inst->src[0];
+ }
+ break;
+ case BRW_OPCODE_CMP:
+ if (i == 1) {
+ scan_inst->src[i] = inst->src[0];
+ progress = true;
+ }
+ }
+ }
+
+ if (scan_inst->dst.file == GRF &&
+ scan_inst->dst.reg == inst->dst.reg &&
+ (scan_inst->dst.reg_offset == inst->dst.reg_offset ||
+ scan_inst->opcode == FS_OPCODE_TEX)) {
+ break;
+ }
+ }
+ }
+
+ return progress;
+}
+/**
+ * Must be called after calculate_live_intervales() to remove unused
+ * writes to registers -- register allocation will fail otherwise
+ * because something deffed but not used won't be considered to
+ * interfere with other regs.
+ */
+bool
+fs_visitor::dead_code_eliminate()
+{
+ bool progress = false;
+ int num_vars = this->virtual_grf_next;
+ bool dead[num_vars];
+
+ for (int i = 0; i < num_vars; i++) {
+ /* This would be ">=", but FS_OPCODE_DISCARD has a src == dst where
+ * it writes dst then reads it as src.
+ */
+ dead[i] = this->virtual_grf_def[i] > this->virtual_grf_use[i];
+
+ if (dead[i]) {
+ /* Mark off its interval so it won't interfere with anything. */
+ this->virtual_grf_def[i] = -1;
+ this->virtual_grf_use[i] = -1;
+ }
+ }
+
+ foreach_iter(exec_list_iterator, iter, this->instructions) {
+ fs_inst *inst = (fs_inst *)iter.get();
+
+ if (inst->dst.file == GRF && dead[inst->dst.reg]) {
+ inst->remove();
+ progress = true;
+ }
+ }
+
+ return progress;
+}
+
+bool
+fs_visitor::virtual_grf_interferes(int a, int b)
+{
+ int start = MAX2(this->virtual_grf_def[a], this->virtual_grf_def[b]);
+ int end = MIN2(this->virtual_grf_use[a], this->virtual_grf_use[b]);
+
+ /* For dead code, just check if the def interferes with the other range. */
+ if (this->virtual_grf_use[a] == -1) {
+ return (this->virtual_grf_def[a] >= this->virtual_grf_def[b] &&
+ this->virtual_grf_def[a] < this->virtual_grf_use[b]);
+ }
+ if (this->virtual_grf_use[b] == -1) {
+ return (this->virtual_grf_def[b] >= this->virtual_grf_def[a] &&
+ this->virtual_grf_def[b] < this->virtual_grf_use[a]);
+ }
+
+ return start <= end;
}
static struct brw_reg brw_reg_from_fs_reg(fs_reg *reg)
case BRW_OPCODE_XOR:
brw_XOR(p, dst, src[0], src[1]);
break;
+ case BRW_OPCODE_NOT:
+ brw_NOT(p, dst, src[0]);
+ break;
+ case BRW_OPCODE_ASR:
+ brw_ASR(p, dst, src[0], src[1]);
+ break;
+ case BRW_OPCODE_SHR:
+ brw_SHR(p, dst, src[0], src[1]);
+ break;
+ case BRW_OPCODE_SHL:
+ brw_SHL(p, dst, src[0], src[1]);
+ break;
case BRW_OPCODE_CMP:
brw_CMP(p, dst, inst->conditional_mod, src[0], src[1]);
struct brw_instruction *inst0, *inst1;
GLuint br = 1;
- if (intel->gen == 5)
+ if (intel->gen >= 5)
br = 2;
assert(loop_stack_depth > 0);
generate_tex(inst, dst, src[0]);
break;
case FS_OPCODE_DISCARD:
- generate_discard(inst);
+ generate_discard(inst, dst /* src0 == dst */);
break;
case FS_OPCODE_DDX:
generate_ddx(inst, dst, src[0]);
if (0) {
v.emit_dummy_fs();
} else {
- v.emit_interpolation();
+ v.calculate_urb_setup();
+ if (intel->gen < 6)
+ v.emit_interpolation_setup_gen4();
+ else
+ v.emit_interpolation_setup_gen6();
/* Generate FS IR for main(). (the visitor only descends into
* functions called "main").
v.emit_fb_writes();
v.assign_curb_setup();
v.assign_urb_setup();
- v.assign_regs();
+
+ bool progress;
+ do {
+ progress = false;
+
+ v.calculate_live_intervals();
+ progress = v.propagate_constants() || progress;
+ progress = v.dead_code_eliminate() || progress;
+ } while (progress);
+
+ if (0)
+ v.assign_regs_trivial();
+ else
+ v.assign_regs();
}
- v.generate_code();
+ if (!v.fail)
+ v.generate_code();
assert(!v.fail); /* FINISHME: Cleanly fail, tested at link time, etc. */