#include "brw_wm.h"
#include "talloc.h"
}
+#include "brw_fs.h"
#include "../glsl/glsl_types.h"
#include "../glsl/ir_optimization.h"
#include "../glsl/ir_print_visitor.h"
-enum register_file {
- ARF = BRW_ARCHITECTURE_REGISTER_FILE,
- GRF = BRW_GENERAL_REGISTER_FILE,
- MRF = BRW_MESSAGE_REGISTER_FILE,
- IMM = BRW_IMMEDIATE_VALUE,
- FIXED_HW_REG, /* a struct brw_reg */
- UNIFORM, /* prog_data->params[hw_reg] */
- BAD_FILE
-};
-
-enum fs_opcodes {
- FS_OPCODE_FB_WRITE = 256,
- FS_OPCODE_RCP,
- FS_OPCODE_RSQ,
- FS_OPCODE_SQRT,
- FS_OPCODE_EXP2,
- FS_OPCODE_LOG2,
- FS_OPCODE_POW,
- FS_OPCODE_SIN,
- FS_OPCODE_COS,
- FS_OPCODE_DDX,
- FS_OPCODE_DDY,
- FS_OPCODE_LINTERP,
- FS_OPCODE_TEX,
- FS_OPCODE_TXB,
- FS_OPCODE_TXL,
- FS_OPCODE_DISCARD,
-};
-
-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)
+brw_new_shader(struct gl_context *ctx, GLuint name, GLuint type)
{
struct brw_shader *shader;
}
struct gl_shader_program *
-brw_new_shader_program(GLcontext *ctx, GLuint name)
+brw_new_shader_program(struct gl_context *ctx, GLuint name)
{
struct brw_shader_program *prog;
prog = talloc_zero(NULL, struct brw_shader_program);
}
GLboolean
-brw_compile_shader(GLcontext *ctx, struct gl_shader *shader)
+brw_compile_shader(struct gl_context *ctx, struct gl_shader *shader)
{
if (!_mesa_ir_compile_shader(ctx, shader))
return GL_FALSE;
}
GLboolean
-brw_link_shader(GLcontext *ctx, struct gl_shader_program *prog)
+brw_link_shader(struct gl_context *ctx, struct gl_shader_program *prog)
{
- if (using_new_fs == -1)
- using_new_fs = getenv("INTEL_NEW_FS") != NULL;
-
- for (unsigned i = 0; i < prog->_NumLinkedShaders; i++) {
- struct brw_shader *shader = (struct brw_shader *)prog->_LinkedShaders[i];
-
- if (using_new_fs && shader->base.Type == GL_FRAGMENT_SHADER) {
- void *mem_ctx = talloc_new(NULL);
- bool progress;
-
- if (shader->ir)
- talloc_free(shader->ir);
- shader->ir = new(shader) exec_list;
- clone_ir_list(mem_ctx, shader->ir, shader->base.ir);
-
- do_mat_op_to_vec(shader->ir);
- do_mod_to_fract(shader->ir);
- do_div_to_mul_rcp(shader->ir);
- do_sub_to_add_neg(shader->ir);
- do_explog_to_explog2(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 */
- false /* loops */
- ) || progress;
-
- progress = do_common_optimization(shader->ir, true, 32) || progress;
-
- progress = lower_noise(shader->ir) || progress;
- progress =
- lower_variable_index_to_cond_assign(shader->ir,
- GL_TRUE, /* input */
- GL_TRUE, /* output */
- GL_TRUE, /* temp */
- GL_TRUE /* uniform */
- ) || progress;
- } while (progress);
-
- validate_ir_tree(shader->ir);
-
- reparent_ir(shader->ir, shader->ir);
- talloc_free(mem_ctx);
- }
+ struct brw_shader *shader =
+ (struct brw_shader *)prog->_LinkedShaders[MESA_SHADER_FRAGMENT];
+ if (shader != NULL) {
+ void *mem_ctx = talloc_new(NULL);
+ bool progress;
+
+ if (shader->ir)
+ talloc_free(shader->ir);
+ shader->ir = new(shader) exec_list;
+ clone_ir_list(mem_ctx, shader->ir, shader->base.ir);
+
+ do_mat_op_to_vec(shader->ir);
+ do_mod_to_fract(shader->ir);
+ do_div_to_mul_rcp(shader->ir);
+ do_sub_to_add_neg(shader->ir);
+ do_explog_to_explog2(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 */
+ false /* loops */
+ ) || progress;
+
+ progress = do_common_optimization(shader->ir, true, 32) || progress;
+
+ progress = lower_noise(shader->ir) || progress;
+ progress =
+ lower_variable_index_to_cond_assign(shader->ir,
+ GL_TRUE, /* input */
+ GL_TRUE, /* output */
+ GL_TRUE, /* temp */
+ GL_TRUE /* uniform */
+ ) || progress;
+ } while (progress);
+
+ validate_ir_tree(shader->ir);
+
+ reparent_ir(shader->ir, shader->ir);
+ talloc_free(mem_ctx);
}
if (!_mesa_ir_link_shader(ctx, prog))
}
}
-class fs_reg {
-public:
- /* Callers of this talloc-based new need not call delete. It's
- * easier to just talloc_free 'ctx' (or any of its ancestors). */
- static void* operator new(size_t size, void *ctx)
- {
- void *node;
-
- node = talloc_size(ctx, size);
- assert(node != NULL);
-
- return node;
- }
-
- void init()
- {
- this->reg = 0;
- this->reg_offset = 0;
- this->negate = 0;
- this->abs = 0;
- this->hw_reg = -1;
- }
-
- /** Generic unset register constructor. */
- fs_reg()
- {
- init();
- this->file = BAD_FILE;
- }
-
- /** Immediate value constructor. */
- fs_reg(float f)
- {
- init();
- this->file = IMM;
- this->type = BRW_REGISTER_TYPE_F;
- this->imm.f = f;
- }
-
- /** Immediate value constructor. */
- fs_reg(int32_t i)
- {
- init();
- this->file = IMM;
- this->type = BRW_REGISTER_TYPE_D;
- this->imm.i = i;
- }
-
- /** Immediate value constructor. */
- fs_reg(uint32_t u)
- {
- init();
- this->file = IMM;
- this->type = BRW_REGISTER_TYPE_UD;
- this->imm.u = u;
- }
-
- /** Fixed brw_reg Immediate value constructor. */
- fs_reg(struct brw_reg fixed_hw_reg)
- {
- init();
- this->file = FIXED_HW_REG;
- this->fixed_hw_reg = fixed_hw_reg;
- this->type = fixed_hw_reg.type;
- }
-
- fs_reg(enum register_file file, int hw_reg);
- fs_reg(class fs_visitor *v, const struct glsl_type *type);
-
- /** Register file: ARF, GRF, MRF, IMM. */
- enum register_file file;
- /** virtual register number. 0 = fixed hw reg */
- int reg;
- /** Offset within the virtual register. */
- int reg_offset;
- /** HW register number. Generally unset until register allocation. */
- int hw_reg;
- /** Register type. BRW_REGISTER_TYPE_* */
- int type;
- bool negate;
- bool abs;
- struct brw_reg fixed_hw_reg;
-
- /** Value for file == BRW_IMMMEDIATE_FILE */
- union {
- int32_t i;
- uint32_t u;
- float f;
- } imm;
-};
-
-static const fs_reg reg_undef;
-static const fs_reg reg_null(ARF, BRW_ARF_NULL);
-
-class fs_inst : public exec_node {
-public:
- /* Callers of this talloc-based new need not call delete. It's
- * easier to just talloc_free 'ctx' (or any of its ancestors). */
- static void* operator new(size_t size, void *ctx)
- {
- void *node;
-
- node = talloc_zero_size(ctx, size);
- assert(node != NULL);
-
- return node;
- }
-
- void init()
- {
- this->opcode = BRW_OPCODE_NOP;
- this->saturate = false;
- this->conditional_mod = BRW_CONDITIONAL_NONE;
- this->predicated = false;
- this->sampler = 0;
- this->target = 0;
- this->eot = false;
- this->shadow_compare = false;
- }
-
- fs_inst()
- {
- init();
- }
-
- fs_inst(int opcode)
- {
- init();
- this->opcode = opcode;
- }
-
- fs_inst(int opcode, fs_reg dst, fs_reg src0)
- {
- init();
- this->opcode = opcode;
- this->dst = dst;
- this->src[0] = src0;
- }
-
- fs_inst(int opcode, fs_reg dst, fs_reg src0, fs_reg src1)
- {
- init();
- this->opcode = opcode;
- this->dst = dst;
- this->src[0] = src0;
- this->src[1] = src1;
- }
-
- fs_inst(int opcode, fs_reg dst, fs_reg src0, fs_reg src1, fs_reg src2)
- {
- init();
- this->opcode = opcode;
- this->dst = dst;
- this->src[0] = src0;
- this->src[1] = src1;
- this->src[2] = src2;
- }
-
- int opcode; /* BRW_OPCODE_* or FS_OPCODE_* */
- fs_reg dst;
- fs_reg src[3];
- bool saturate;
- bool predicated;
- int conditional_mod; /**< BRW_CONDITIONAL_* */
-
- int mlen; /**< SEND message length */
- int sampler;
- int target; /**< MRT target. */
- bool eot;
- bool shadow_compare;
-
- /** @{
- * Annotation for the generated IR. One of the two can be set.
- */
- ir_instruction *ir;
- const char *annotation;
- /** @} */
-};
-
-class fs_visitor : public ir_visitor
-{
-public:
-
- fs_visitor(struct brw_wm_compile *c, struct brw_shader *shader)
- {
- 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->variable_ht = hash_table_ctor(0,
- hash_table_pointer_hash,
- hash_table_pointer_compare);
-
- this->frag_color = NULL;
- this->frag_data = NULL;
- this->frag_depth = NULL;
- this->first_non_payload_grf = 0;
-
- 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;
- }
- ~fs_visitor()
- {
- talloc_free(this->mem_ctx);
- hash_table_dtor(this->variable_ht);
- }
-
- fs_reg *variable_storage(ir_variable *var);
- int virtual_grf_alloc(int size);
-
- void visit(ir_variable *ir);
- void visit(ir_assignment *ir);
- void visit(ir_dereference_variable *ir);
- void visit(ir_dereference_record *ir);
- void visit(ir_dereference_array *ir);
- void visit(ir_expression *ir);
- void visit(ir_texture *ir);
- void visit(ir_if *ir);
- void visit(ir_constant *ir);
- void visit(ir_swizzle *ir);
- void visit(ir_return *ir);
- void visit(ir_loop *ir);
- void visit(ir_loop_jump *ir);
- void visit(ir_discard *ir);
- void visit(ir_call *ir);
- void visit(ir_function *ir);
- void visit(ir_function_signature *ir);
-
- fs_inst *emit(fs_inst inst);
- void assign_curb_setup();
- void assign_urb_setup();
- void assign_regs();
- void assign_regs_trivial();
- void calculate_live_intervals();
- 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, 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_fragcoord_interpolation(ir_variable *ir);
- void emit_general_interpolation(ir_variable *ir);
- void emit_interpolation_setup();
- void emit_fb_writes();
-
- 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_compile *p;
- struct brw_shader *shader;
- void *mem_ctx;
- exec_list instructions;
-
- 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;
-
- /** @{ debug annotation info */
- const char *current_annotation;
- ir_instruction *base_ir;
- const char **annotation_string;
- ir_instruction **annotation_ir;
- /** @} */
-
- bool fail;
-
- /* Result of last visit() method. */
- fs_reg result;
-
- fs_reg pixel_x;
- fs_reg pixel_y;
- fs_reg wpos_w;
- fs_reg pixel_w;
- fs_reg delta_x;
- fs_reg delta_y;
-
- int grf_used;
-
-};
-
int
fs_visitor::virtual_grf_alloc(int size)
{
this->type = BRW_REGISTER_TYPE_F;
}
+/** Fixed HW reg constructor. */
+fs_reg::fs_reg(enum register_file file, int hw_reg, uint32_t type)
+{
+ init();
+ this->file = file;
+ this->hw_reg = hw_reg;
+ this->type = type;
+}
+
int
brw_type_for_base_type(const struct glsl_type *type)
{
return BRW_REGISTER_TYPE_UD;
case GLSL_TYPE_ARRAY:
case GLSL_TYPE_STRUCT:
+ case GLSL_TYPE_SAMPLER:
/* 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.
case GLSL_TYPE_BOOL:
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];
+ unsigned int param = c->prog_data.nr_params++;
+
+ assert(param < ARRAY_SIZE(c->prog_data.param));
+
+ switch (type->base_type) {
+ case GLSL_TYPE_FLOAT:
+ c->prog_data.param_convert[param] = PARAM_NO_CONVERT;
+ break;
+ case GLSL_TYPE_UINT:
+ c->prog_data.param_convert[param] = PARAM_CONVERT_F2U;
+ break;
+ case GLSL_TYPE_INT:
+ c->prog_data.param_convert[param] = PARAM_CONVERT_F2I;
+ break;
+ case GLSL_TYPE_BOOL:
+ c->prog_data.param_convert[param] = PARAM_CONVERT_F2B;
+ break;
+ }
+
+ c->prog_data.param[param] = &vec_values[i];
}
return 1;
*/
int last_swiz = -1;
for (unsigned int i = 0; i < 4; i++) {
- int this_swiz = GET_SWZ(element->swizzle, i);
- if (this_swiz == last_swiz)
+ int swiz = GET_SWZ(element->swizzle, i);
+ if (swiz == last_swiz)
break;
- last_swiz = this_swiz;
+ last_swiz = swiz;
- c->prog_data.param[c->prog_data.nr_params++] = &vec_values[i];
+ c->prog_data.param_convert[c->prog_data.nr_params] =
+ PARAM_NO_CONVERT;
+ c->prog_data.param[c->prog_data.nr_params++] = &vec_values[swiz];
}
}
}
}
-void
+fs_reg *
fs_visitor::emit_fragcoord_interpolation(ir_variable *ir)
{
fs_reg *reg = new(this->mem_ctx) fs_reg(this, ir->type);
/* gl_FragCoord.w: Already set up in emit_interpolation */
emit(fs_inst(BRW_OPCODE_MOV, wpos, this->wpos_w));
- hash_table_insert(this->variable_ht, reg, ir);
+ return reg;
}
-
-void
+fs_reg *
fs_visitor::emit_general_interpolation(ir_variable *ir)
{
fs_reg *reg = new(this->mem_ctx) fs_reg(this, 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 (!(fp->Base.InputsRead & BITFIELD64_BIT(location))) {
+ if (urb_setup[location] == -1) {
/* If there's no incoming setup data for this slot, don't
- * emit interpolation for it (since it's not used, and
- * we'd fall over later trying to find the setup data.
+ * emit interpolation for it.
*/
attr.reg_offset += type->vector_elements;
+ location++;
continue;
}
fs_reg(interp)));
attr.reg_offset++;
}
- attr.reg_offset -= type->vector_elements;
- for (unsigned int c = 0; c < type->vector_elements; c++) {
- emit(fs_inst(BRW_OPCODE_MUL,
- attr,
- attr,
- this->pixel_w));
- attr.reg_offset++;
+ if (intel->gen < 6) {
+ attr.reg_offset -= type->vector_elements;
+ 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++;
}
}
- hash_table_insert(this->variable_ht, reg, ir);
+ 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 {
+ 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;
+}
+
+fs_inst *
+fs_visitor::emit_math(fs_opcodes opcode, fs_reg dst, fs_reg src)
+{
+ switch (opcode) {
+ case FS_OPCODE_RCP:
+ case FS_OPCODE_RSQ:
+ case FS_OPCODE_SQRT:
+ case FS_OPCODE_EXP2:
+ case FS_OPCODE_LOG2:
+ case FS_OPCODE_SIN:
+ case FS_OPCODE_COS:
+ break;
+ default:
+ assert(!"not reached: bad math opcode");
+ return NULL;
+ }
+
+ /* Can't do hstride == 0 args to gen6 math, so expand it out. We
+ * might be able to do better by doing execsize = 1 math and then
+ * expanding that result out, but we would need to be careful with
+ * masking.
+ */
+ if (intel->gen >= 6 && src.file == UNIFORM) {
+ fs_reg expanded = fs_reg(this, glsl_type::float_type);
+ emit(fs_inst(BRW_OPCODE_MOV, expanded, src));
+ src = expanded;
+ }
+
+ fs_inst *inst = emit(fs_inst(opcode, dst, src));
+
+ if (intel->gen < 6) {
+ inst->base_mrf = 2;
+ inst->mlen = 1;
+ }
+
+ return inst;
+}
+
+fs_inst *
+fs_visitor::emit_math(fs_opcodes opcode, fs_reg dst, fs_reg src0, fs_reg src1)
+{
+ int base_mrf = 2;
+ fs_inst *inst;
+
+ assert(opcode == FS_OPCODE_POW);
+
+ if (intel->gen >= 6) {
+ /* Can't do hstride == 0 args to gen6 math, so expand it out. */
+ if (src0.file == UNIFORM) {
+ fs_reg expanded = fs_reg(this, glsl_type::float_type);
+ emit(fs_inst(BRW_OPCODE_MOV, expanded, src0));
+ src0 = expanded;
+ }
+
+ if (src1.file == UNIFORM) {
+ fs_reg expanded = fs_reg(this, glsl_type::float_type);
+ emit(fs_inst(BRW_OPCODE_MOV, expanded, src1));
+ src1 = expanded;
+ }
+
+ inst = emit(fs_inst(opcode, dst, src0, src1));
+ } else {
+ emit(fs_inst(BRW_OPCODE_MOV, fs_reg(MRF, base_mrf + 1), src1));
+ inst = emit(fs_inst(opcode, dst, src0, reg_null_f));
+
+ inst->base_mrf = base_mrf;
+ inst->mlen = 2;
+ }
+ return inst;
}
void
if (ir->mode == ir_var_in) {
if (!strcmp(ir->name, "gl_FragCoord")) {
- emit_fragcoord_interpolation(ir);
- return;
+ reg = emit_fragcoord_interpolation(ir);
} else if (!strcmp(ir->name, "gl_FrontFacing")) {
- 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)));
+ reg = emit_frontfacing_interpolation(ir);
} else {
- emit_general_interpolation(ir);
- return;
+ reg = emit_general_interpolation(ir);
}
+ assert(reg);
+ hash_table_insert(this->variable_ht, reg, ir);
+ return;
}
if (ir->mode == ir_var_uniform) {
}
reg = new(this->mem_ctx) fs_reg(UNIFORM, param_index);
+ reg->type = brw_type_for_base_type(ir->type);
}
if (!reg)
{
unsigned int operand;
fs_reg op[2], temp;
- fs_reg result;
fs_inst *inst;
for (operand = 0; operand < ir->get_num_operands(); operand++) {
switch (ir->operation) {
case ir_unop_logic_not:
- emit(fs_inst(BRW_OPCODE_ADD, this->result, op[0], fs_reg(-1)));
+ /* Note that BRW_OPCODE_NOT is not appropriate here, since it is
+ * ones complement of the whole register, not just bit 0.
+ */
+ emit(fs_inst(BRW_OPCODE_XOR, this->result, op[0], fs_reg(1)));
break;
case ir_unop_neg:
op[0].negate = !op[0].negate;
emit(fs_inst(BRW_OPCODE_MOV, this->result, fs_reg(0.0f)));
- inst = emit(fs_inst(BRW_OPCODE_CMP, reg_null, op[0], fs_reg(0.0f)));
+ inst = emit(fs_inst(BRW_OPCODE_CMP, reg_null_f, op[0], fs_reg(0.0f)));
inst->conditional_mod = BRW_CONDITIONAL_G;
inst = emit(fs_inst(BRW_OPCODE_MOV, this->result, fs_reg(1.0f)));
inst->predicated = true;
- inst = emit(fs_inst(BRW_OPCODE_CMP, reg_null, op[0], fs_reg(0.0f)));
+ inst = emit(fs_inst(BRW_OPCODE_CMP, reg_null_f, op[0], fs_reg(0.0f)));
inst->conditional_mod = BRW_CONDITIONAL_L;
inst = emit(fs_inst(BRW_OPCODE_MOV, this->result, fs_reg(-1.0f)));
inst->predicated = true;
break;
case ir_unop_rcp:
- emit(fs_inst(FS_OPCODE_RCP, this->result, op[0]));
+ emit_math(FS_OPCODE_RCP, this->result, op[0]);
break;
case ir_unop_exp2:
- emit(fs_inst(FS_OPCODE_EXP2, this->result, op[0]));
+ emit_math(FS_OPCODE_EXP2, this->result, op[0]);
break;
case ir_unop_log2:
- emit(fs_inst(FS_OPCODE_LOG2, this->result, op[0]));
+ emit_math(FS_OPCODE_LOG2, this->result, op[0]);
break;
case ir_unop_exp:
case ir_unop_log:
assert(!"not reached: should be handled by ir_explog_to_explog2");
break;
case ir_unop_sin:
- emit(fs_inst(FS_OPCODE_SIN, this->result, op[0]));
+ emit_math(FS_OPCODE_SIN, this->result, op[0]);
break;
case ir_unop_cos:
- emit(fs_inst(FS_OPCODE_COS, this->result, op[0]));
+ emit_math(FS_OPCODE_COS, this->result, op[0]);
break;
case ir_unop_dFdx:
break;
case ir_unop_sqrt:
- emit(fs_inst(FS_OPCODE_SQRT, this->result, op[0]));
+ emit_math(FS_OPCODE_SQRT, this->result, op[0]);
break;
case ir_unop_rsq:
- emit(fs_inst(FS_OPCODE_RSQ, this->result, op[0]));
+ emit_math(FS_OPCODE_RSQ, this->result, op[0]);
break;
case ir_unop_i2f:
case ir_unop_b2f:
case ir_unop_b2i:
- emit(fs_inst(BRW_OPCODE_MOV, this->result, op[0]));
- break;
case ir_unop_f2i:
emit(fs_inst(BRW_OPCODE_MOV, this->result, op[0]));
break;
case ir_unop_i2b:
inst = emit(fs_inst(BRW_OPCODE_CMP, this->result, op[0], fs_reg(0.0f)));
inst->conditional_mod = BRW_CONDITIONAL_NZ;
+ inst = emit(fs_inst(BRW_OPCODE_AND, this->result,
+ this->result, fs_reg(1)));
+ break;
case ir_unop_trunc:
- emit(fs_inst(BRW_OPCODE_RNDD, this->result, op[0]));
+ emit(fs_inst(BRW_OPCODE_RNDZ, this->result, op[0]));
break;
case ir_unop_ceil:
- op[0].negate = ~op[0].negate;
+ op[0].negate = !op[0].negate;
inst = emit(fs_inst(BRW_OPCODE_RNDD, this->result, op[0]));
this->result.negate = true;
break;
case ir_unop_fract:
inst = emit(fs_inst(BRW_OPCODE_FRC, this->result, op[0]));
break;
+ case ir_unop_round_even:
+ emit(fs_inst(BRW_OPCODE_RNDE, this->result, op[0]));
+ break;
case ir_binop_min:
inst = emit(fs_inst(BRW_OPCODE_CMP, this->result, op[0], op[1]));
break;
case ir_binop_pow:
- inst = emit(fs_inst(FS_OPCODE_POW, this->result, op[0], op[1]));
+ emit_math(FS_OPCODE_POW, this->result, op[0], op[1]);
break;
case ir_unop_bit_not:
- case ir_unop_u2f:
- case ir_binop_lshift:
- case ir_binop_rshift:
+ inst = emit(fs_inst(BRW_OPCODE_NOT, this->result, op[0]));
+ break;
case ir_binop_bit_and:
+ inst = emit(fs_inst(BRW_OPCODE_AND, this->result, op[0], op[1]));
+ break;
case ir_binop_bit_xor:
+ inst = emit(fs_inst(BRW_OPCODE_XOR, this->result, op[0], op[1]));
+ break;
case ir_binop_bit_or:
+ inst = emit(fs_inst(BRW_OPCODE_OR, this->result, op[0], op[1]));
+ break;
+
+ case ir_unop_u2f:
+ case ir_binop_lshift:
+ case ir_binop_rshift:
assert(!"GLSL 1.30 features unsupported");
break;
}
}
+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);
+ }
+ break;
+
+ 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, reg_null, this->result, fs_reg(0)));
- inst->conditional_mod = BRW_CONDITIONAL_NZ;
+ emit_bool_to_cond_code(ir->condition);
}
- 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 base_mrf = 2;
- fs_inst *inst = NULL;
- unsigned int mlen = 0;
+ int mlen;
+ int base_mrf = 1;
+ bool simd16 = false;
+ fs_reg orig_dst;
- ir->coordinate->accept(this);
- fs_reg coordinate = this->result;
+ /* g0 header. */
+ mlen = 1;
+
+ if (ir->shadow_comparitor) {
+ for (int i = 0; i < ir->coordinate->type->vector_elements; i++) {
+ emit(fs_inst(BRW_OPCODE_MOV, fs_reg(MRF, base_mrf + mlen + i),
+ 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 (int i = 0; i < ir->coordinate->type->vector_elements; i++) {
+ emit(fs_inst(BRW_OPCODE_MOV, fs_reg(MRF, base_mrf + mlen + i),
+ 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 (int i = 0; i < ir->coordinate->type->vector_elements; i++) {
+ emit(fs_inst(BRW_OPCODE_MOV, fs_reg(MRF, base_mrf + mlen + i * 2),
+ coordinate));
coordinate.reg_offset++;
- proj_coordinate.reg_offset++;
}
- proj_coordinate.reg_offset = 0;
- coordinate = proj_coordinate;
+ /* 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++;
+ }
+
+ /* 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;
}
- 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++;
+ fs_inst *inst = NULL;
+ switch (ir->op) {
+ case ir_tex:
+ inst = emit(fs_inst(FS_OPCODE_TEX, dst));
+ break;
+ case ir_txb:
+ inst = emit(fs_inst(FS_OPCODE_TXB, dst));
+ break;
+ case ir_txl:
+ inst = emit(fs_inst(FS_OPCODE_TXL, dst));
+ break;
+ case ir_txd:
+ case ir_txf:
+ assert(!"GLSL 1.30 features unsupported");
+ break;
+ }
+ inst->base_mrf = base_mrf;
+ 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;
+ }
}
- /* Pre-Ironlake, the 8-wide sampler always took u,v,r. */
- if (intel->gen < 5)
- mlen = 3;
+ 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 = 1; /* g0 header always present. */
+ int base_mrf = 1;
+
+ for (int i = 0; i < ir->coordinate->type->vector_elements; i++) {
+ emit(fs_inst(BRW_OPCODE_MOV, fs_reg(MRF, base_mrf + mlen + i),
+ coordinate));
+ coordinate.reg_offset++;
+ }
+ mlen += ir->coordinate->type->vector_elements;
if (ir->shadow_comparitor) {
- /* For shadow comparisons, we have to supply u,v,r. */
- mlen = 3;
+ mlen = MAX2(mlen, 5);
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)));
+ inst = emit(fs_inst(FS_OPCODE_TEX, dst));
break;
case ir_txb:
ir->lod_info.bias->accept(this);
+ mlen = MAX2(mlen, 5);
emit(fs_inst(BRW_OPCODE_MOV, fs_reg(MRF, base_mrf + mlen), this->result));
mlen++;
- inst = emit(fs_inst(FS_OPCODE_TXB, dst, fs_reg(MRF, base_mrf)));
+ inst = emit(fs_inst(FS_OPCODE_TXB, dst));
break;
case ir_txl:
ir->lod_info.lod->accept(this);
+ mlen = MAX2(mlen, 5);
emit(fs_inst(BRW_OPCODE_MOV, fs_reg(MRF, base_mrf + mlen), this->result));
mlen++;
- inst = emit(fs_inst(FS_OPCODE_TXL, dst, fs_reg(MRF, base_mrf)));
+ inst = emit(fs_inst(FS_OPCODE_TXL, dst));
break;
case ir_txd:
case ir_txf:
assert(!"GLSL 1.30 features unsupported");
break;
}
-
- 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;
+ inst->base_mrf = base_mrf;
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;
+ int sampler;
+ fs_inst *inst = NULL;
- fs_reg result = fs_reg(this, ir->type);
- this->result = result;
+ ir->coordinate->accept(this);
+ fs_reg coordinate = this->result;
- for (unsigned int i = 0; i < ir->type->vector_elements; i++) {
- fs_reg channel = val;
- int swiz = 0;
+ /* Should be lowered by do_lower_texture_projection */
+ assert(!ir->projector);
- switch (i) {
- case 0:
- swiz = ir->mask.x;
- break;
- case 1:
- swiz = ir->mask.y;
- break;
- case 2:
- swiz = ir->mask.z;
- break;
- case 3:
- swiz = ir->mask.w;
- break;
+ sampler = _mesa_get_sampler_uniform_value(ir->sampler,
+ ctx->Shader.CurrentFragmentProgram,
+ &brw->fragment_program->Base);
+ sampler = c->fp->program.Base.SamplerUnits[sampler];
+
+ /* The 965 requires the EU to do the normalization of GL rectangle
+ * texture coordinates. We use the program parameter state
+ * tracking to get the scaling factor.
+ */
+ if (ir->sampler->type->sampler_dimensionality == GLSL_SAMPLER_DIM_RECT) {
+ struct gl_program_parameter_list *params = c->fp->program.Base.Parameters;
+ int tokens[STATE_LENGTH] = {
+ STATE_INTERNAL,
+ STATE_TEXRECT_SCALE,
+ sampler,
+ 0,
+ 0
+ };
+
+ c->prog_data.param_convert[c->prog_data.nr_params] =
+ PARAM_NO_CONVERT;
+ c->prog_data.param_convert[c->prog_data.nr_params + 1] =
+ PARAM_NO_CONVERT;
+
+ fs_reg scale_x = fs_reg(UNIFORM, c->prog_data.nr_params);
+ fs_reg scale_y = fs_reg(UNIFORM, c->prog_data.nr_params + 1);
+ GLuint index = _mesa_add_state_reference(params,
+ (gl_state_index *)tokens);
+ float *vec_values = this->fp->Base.Parameters->ParameterValues[index];
+
+ c->prog_data.param[c->prog_data.nr_params++] = &vec_values[0];
+ c->prog_data.param[c->prog_data.nr_params++] = &vec_values[1];
+
+ fs_reg dst = fs_reg(this, ir->coordinate->type);
+ fs_reg src = coordinate;
+ coordinate = dst;
+
+ emit(fs_inst(BRW_OPCODE_MUL, dst, src, scale_x));
+ dst.reg_offset++;
+ src.reg_offset++;
+ emit(fs_inst(BRW_OPCODE_MUL, dst, src, scale_y));
+ }
+
+ /* 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 = 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;
+ int swiz = 0;
+
+ switch (i) {
+ case 0:
+ swiz = ir->mask.x;
+ break;
+ case 1:
+ swiz = ir->mask.y;
+ break;
+ case 2:
+ swiz = ir->mask.z;
+ break;
+ case 3:
+ swiz = ir->mask.w;
+ break;
}
channel.reg_offset += swiz;
assert(ir->condition == NULL); /* FINISHME */
- emit(fs_inst(FS_OPCODE_DISCARD, temp, temp));
+ emit(fs_inst(FS_OPCODE_DISCARD_NOT, temp, reg_null_d));
+ emit(fs_inst(FS_OPCODE_DISCARD_AND, reg_null_d, temp));
+ kill_emitted = true;
}
void
}
}
+void
+fs_visitor::emit_bool_to_cond_code(ir_rvalue *ir)
+{
+ ir_expression *expr = ir->as_expression();
+
+ if (expr) {
+ fs_reg op[2];
+ fs_inst *inst;
+
+ for (unsigned int i = 0; i < expr->get_num_operands(); i++) {
+ assert(expr->operands[i]->type->is_scalar());
+
+ expr->operands[i]->accept(this);
+ op[i] = this->result;
+ }
+
+ switch (expr->operation) {
+ case ir_unop_logic_not:
+ inst = emit(fs_inst(BRW_OPCODE_AND, reg_null_d, op[0], fs_reg(1)));
+ inst->conditional_mod = BRW_CONDITIONAL_Z;
+ break;
+
+ case ir_binop_logic_xor:
+ inst = emit(fs_inst(BRW_OPCODE_XOR, reg_null_d, op[0], op[1]));
+ inst->conditional_mod = BRW_CONDITIONAL_NZ;
+ break;
+
+ case ir_binop_logic_or:
+ inst = emit(fs_inst(BRW_OPCODE_OR, reg_null_d, op[0], op[1]));
+ inst->conditional_mod = BRW_CONDITIONAL_NZ;
+ break;
+
+ case ir_binop_logic_and:
+ inst = emit(fs_inst(BRW_OPCODE_AND, reg_null_d, op[0], op[1]));
+ inst->conditional_mod = BRW_CONDITIONAL_NZ;
+ break;
+
+ case ir_unop_f2b:
+ if (intel->gen >= 6) {
+ inst = emit(fs_inst(BRW_OPCODE_CMP, reg_null_d,
+ op[0], fs_reg(0.0f)));
+ } else {
+ inst = emit(fs_inst(BRW_OPCODE_MOV, reg_null_d, op[0]));
+ }
+ inst->conditional_mod = BRW_CONDITIONAL_NZ;
+ break;
+
+ case ir_unop_i2b:
+ if (intel->gen >= 6) {
+ inst = emit(fs_inst(BRW_OPCODE_CMP, reg_null_d, op[0], fs_reg(0)));
+ } else {
+ inst = emit(fs_inst(BRW_OPCODE_MOV, reg_null_d, op[0]));
+ }
+ inst->conditional_mod = BRW_CONDITIONAL_NZ;
+ break;
+
+ case ir_binop_greater:
+ inst = emit(fs_inst(BRW_OPCODE_CMP, reg_null_d, op[0], op[1]));
+ inst->conditional_mod = BRW_CONDITIONAL_G;
+ break;
+ case ir_binop_gequal:
+ inst = emit(fs_inst(BRW_OPCODE_CMP, reg_null_d, op[0], op[1]));
+ inst->conditional_mod = BRW_CONDITIONAL_GE;
+ break;
+ case ir_binop_less:
+ inst = emit(fs_inst(BRW_OPCODE_CMP, reg_null_d, op[0], op[1]));
+ inst->conditional_mod = BRW_CONDITIONAL_L;
+ break;
+ case ir_binop_lequal:
+ inst = emit(fs_inst(BRW_OPCODE_CMP, reg_null_d, op[0], op[1]));
+ inst->conditional_mod = BRW_CONDITIONAL_LE;
+ break;
+ case ir_binop_equal:
+ case ir_binop_all_equal:
+ inst = emit(fs_inst(BRW_OPCODE_CMP, reg_null_d, op[0], op[1]));
+ inst->conditional_mod = BRW_CONDITIONAL_Z;
+ break;
+ case ir_binop_nequal:
+ case ir_binop_any_nequal:
+ inst = emit(fs_inst(BRW_OPCODE_CMP, reg_null_d, op[0], op[1]));
+ inst->conditional_mod = BRW_CONDITIONAL_NZ;
+ break;
+ default:
+ assert(!"not reached");
+ this->fail = true;
+ break;
+ }
+ return;
+ }
+
+ ir->accept(this);
+
+ if (intel->gen >= 6) {
+ fs_inst *inst = emit(fs_inst(BRW_OPCODE_AND, reg_null_d,
+ this->result, fs_reg(1)));
+ inst->conditional_mod = BRW_CONDITIONAL_NZ;
+ } else {
+ fs_inst *inst = emit(fs_inst(BRW_OPCODE_MOV, reg_null_d, this->result));
+ inst->conditional_mod = BRW_CONDITIONAL_NZ;
+ }
+}
+
+/**
+ * Emit a gen6 IF statement with the comparison folded into the IF
+ * instruction.
+ */
+void
+fs_visitor::emit_if_gen6(ir_if *ir)
+{
+ ir_expression *expr = ir->condition->as_expression();
+
+ if (expr) {
+ fs_reg op[2];
+ fs_inst *inst;
+ fs_reg temp;
+
+ for (unsigned int i = 0; i < expr->get_num_operands(); i++) {
+ assert(expr->operands[i]->type->is_scalar());
+
+ expr->operands[i]->accept(this);
+ op[i] = this->result;
+ }
+
+ switch (expr->operation) {
+ case ir_unop_logic_not:
+ inst = emit(fs_inst(BRW_OPCODE_IF, temp, op[0], fs_reg(1)));
+ inst->conditional_mod = BRW_CONDITIONAL_Z;
+ return;
+
+ case ir_binop_logic_xor:
+ inst = emit(fs_inst(BRW_OPCODE_IF, reg_null_d, op[0], op[1]));
+ inst->conditional_mod = BRW_CONDITIONAL_NZ;
+ return;
+
+ case ir_binop_logic_or:
+ temp = fs_reg(this, glsl_type::bool_type);
+ emit(fs_inst(BRW_OPCODE_OR, temp, op[0], op[1]));
+ inst = emit(fs_inst(BRW_OPCODE_IF, reg_null_d, temp, fs_reg(0)));
+ inst->conditional_mod = BRW_CONDITIONAL_NZ;
+ return;
+
+ case ir_binop_logic_and:
+ temp = fs_reg(this, glsl_type::bool_type);
+ emit(fs_inst(BRW_OPCODE_AND, temp, op[0], op[1]));
+ inst = emit(fs_inst(BRW_OPCODE_IF, reg_null_d, temp, fs_reg(0)));
+ inst->conditional_mod = BRW_CONDITIONAL_NZ;
+ return;
+
+ case ir_unop_f2b:
+ inst = emit(fs_inst(BRW_OPCODE_IF, reg_null_f, op[0], fs_reg(0)));
+ inst->conditional_mod = BRW_CONDITIONAL_NZ;
+ return;
+
+ case ir_unop_i2b:
+ inst = emit(fs_inst(BRW_OPCODE_IF, reg_null_d, op[0], fs_reg(0)));
+ inst->conditional_mod = BRW_CONDITIONAL_NZ;
+ return;
+
+ case ir_binop_greater:
+ inst = emit(fs_inst(BRW_OPCODE_IF, reg_null_d, op[0], op[1]));
+ inst->conditional_mod = BRW_CONDITIONAL_G;
+ return;
+ case ir_binop_gequal:
+ inst = emit(fs_inst(BRW_OPCODE_IF, reg_null_d, op[0], op[1]));
+ inst->conditional_mod = BRW_CONDITIONAL_GE;
+ return;
+ case ir_binop_less:
+ inst = emit(fs_inst(BRW_OPCODE_IF, reg_null_d, op[0], op[1]));
+ inst->conditional_mod = BRW_CONDITIONAL_L;
+ return;
+ case ir_binop_lequal:
+ inst = emit(fs_inst(BRW_OPCODE_IF, reg_null_d, op[0], op[1]));
+ inst->conditional_mod = BRW_CONDITIONAL_LE;
+ return;
+ case ir_binop_equal:
+ case ir_binop_all_equal:
+ inst = emit(fs_inst(BRW_OPCODE_IF, reg_null_d, op[0], op[1]));
+ inst->conditional_mod = BRW_CONDITIONAL_Z;
+ return;
+ case ir_binop_nequal:
+ case ir_binop_any_nequal:
+ inst = emit(fs_inst(BRW_OPCODE_IF, reg_null_d, op[0], op[1]));
+ inst->conditional_mod = BRW_CONDITIONAL_NZ;
+ return;
+ default:
+ assert(!"not reached");
+ inst = emit(fs_inst(BRW_OPCODE_IF, reg_null_d, op[0], fs_reg(0)));
+ inst->conditional_mod = BRW_CONDITIONAL_NZ;
+ this->fail = true;
+ return;
+ }
+ return;
+ }
+
+ ir->condition->accept(this);
+
+ fs_inst *inst = emit(fs_inst(BRW_OPCODE_IF, reg_null_d, this->result, fs_reg(0)));
+ inst->conditional_mod = BRW_CONDITIONAL_NZ;
+}
+
void
fs_visitor::visit(ir_if *ir)
{
*/
this->base_ir = ir->condition;
- /* Generate the condition into the condition code. */
- ir->condition->accept(this);
- inst = emit(fs_inst(BRW_OPCODE_MOV, fs_reg(brw_null_reg()), this->result));
- inst->conditional_mod = BRW_CONDITIONAL_NZ;
+ if (intel->gen >= 6) {
+ emit_if_gen6(ir);
+ } else {
+ emit_bool_to_cond_code(ir->condition);
- inst = emit(fs_inst(BRW_OPCODE_IF));
- inst->predicated = true;
+ inst = emit(fs_inst(BRW_OPCODE_IF));
+ inst->predicated = true;
+ }
foreach_iter(exec_list_iterator, iter, ir->then_instructions) {
ir_instruction *ir = (ir_instruction *)iter.get();
this->base_ir = ir->to;
ir->to->accept(this);
- fs_inst *inst = emit(fs_inst(BRW_OPCODE_CMP, reg_null,
+ fs_inst *inst = emit(fs_inst(BRW_OPCODE_CMP, reg_null_d,
counter, this->result));
switch (ir->cmp) {
case ir_binop_equal:
write = emit(fs_inst(FS_OPCODE_FB_WRITE,
fs_reg(0),
fs_reg(0)));
+ write->base_mrf = 0;
}
/* The register location here is relative to the start of the URB
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_setup()
+fs_visitor::emit_interpolation_setup_gen4()
{
struct brw_reg g1_uw = retype(brw_vec1_grf(1, 0), BRW_REGISTER_TYPE_UW);
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,
interp_reg(FRAG_ATTRIB_WPOS, 3)));
/* 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_w));
+ emit_math(FS_OPCODE_RCP, this->pixel_w, wpos_w);
+ this->current_annotation = NULL;
+}
+
+/** 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 the pixel centers end up used, the setup is the same as for gen4. */
+ this->current_annotation = "compute pixel centers";
+ fs_reg int_pixel_x = fs_reg(this, glsl_type::uint_type);
+ fs_reg int_pixel_y = fs_reg(this, glsl_type::uint_type);
+ int_pixel_x.type = BRW_REGISTER_TYPE_UW;
+ int_pixel_y.type = BRW_REGISTER_TYPE_UW;
+ emit(fs_inst(BRW_OPCODE_ADD,
+ int_pixel_x,
+ fs_reg(stride(suboffset(g1_uw, 4), 2, 4, 0)),
+ fs_reg(brw_imm_v(0x10101010))));
+ emit(fs_inst(BRW_OPCODE_ADD,
+ int_pixel_y,
+ fs_reg(stride(suboffset(g1_uw, 5), 2, 4, 0)),
+ fs_reg(brw_imm_v(0x11001100))));
+
+ /* As of gen6, we can no longer mix float and int sources. We have
+ * to turn the integer pixel centers into floats for their actual
+ * use.
+ */
+ this->pixel_x = fs_reg(this, glsl_type::float_type);
+ this->pixel_y = fs_reg(this, glsl_type::float_type);
+ emit(fs_inst(BRW_OPCODE_MOV, this->pixel_x, int_pixel_x));
+ emit(fs_inst(BRW_OPCODE_MOV, this->pixel_y, int_pixel_y));
+
+ 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_math(FS_OPCODE_RCP, this->pixel_w, wpos_w);
+
+ this->delta_x = fs_reg(brw_vec8_grf(2, 0));
+ this->delta_y = fs_reg(brw_vec8_grf(3, 0));
+
this->current_annotation = NULL;
}
fs_visitor::emit_fb_writes()
{
this->current_annotation = "FB write header";
+ GLboolean header_present = GL_TRUE;
int nr = 0;
- /* m0, m1 header */
- nr += 2;
+ if (intel->gen >= 6 &&
+ !this->kill_emitted &&
+ c->key.nr_color_regions == 1) {
+ header_present = false;
+ }
+
+ 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_inst *inst = emit(fs_inst(FS_OPCODE_FB_WRITE,
reg_undef, reg_undef));
inst->target = target;
+ inst->base_mrf = 0;
inst->mlen = nr;
if (target == c->key.nr_color_regions - 1)
inst->eot = true;
+ inst->header_present = header_present;
}
if (c->key.nr_color_regions == 0) {
fs_inst *inst = emit(fs_inst(FS_OPCODE_FB_WRITE,
reg_undef, reg_undef));
+ inst->base_mrf = 0;
inst->mlen = nr;
inst->eot = true;
+ inst->header_present = header_present;
}
this->current_annotation = NULL;
fs_visitor::generate_fb_write(fs_inst *inst)
{
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));
+
+ if (inst->header_present) {
+ if (intel->gen >= 6) {
+ brw_MOV(p,
+ brw_message_reg(inst->base_mrf),
+ brw_vec8_grf(0, 0));
+
+ if (inst->target > 0) {
+ /* Set the render target index for choosing BLEND_STATE. */
+ brw_MOV(p, retype(brw_vec1_reg(BRW_MESSAGE_REGISTER_FILE, 0, 2),
+ BRW_REGISTER_TYPE_UD),
+ brw_imm_ud(inst->target));
+ }
+
+ /* Clear viewport index, render target array index. */
+ brw_AND(p, retype(brw_vec1_reg(BRW_MESSAGE_REGISTER_FILE, 0, 0),
+ BRW_REGISTER_TYPE_UD),
+ retype(brw_vec1_grf(0, 0), BRW_REGISTER_TYPE_UD),
+ brw_imm_ud(0xf7ff));
+
+ implied_header = brw_null_reg();
+ } else {
+ implied_header = retype(brw_vec8_grf(0, 0), BRW_REGISTER_TYPE_UW);
+ }
+
+ brw_MOV(p,
+ brw_message_reg(inst->base_mrf + 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),
+ inst->base_mrf,
+ implied_header,
inst->target,
inst->mlen,
0,
break;
}
- if (inst->opcode == FS_OPCODE_POW) {
- brw_MOV(p, brw_message_reg(3), src[1]);
- }
+ if (intel->gen >= 6) {
+ assert(inst->mlen == 0);
+
+ if (inst->opcode == FS_OPCODE_POW) {
+ brw_math2(p, dst, op, src[0], src[1]);
+ } else {
+ brw_math(p, dst,
+ op,
+ inst->saturate ? BRW_MATH_SATURATE_SATURATE :
+ BRW_MATH_SATURATE_NONE,
+ 0, src[0],
+ BRW_MATH_DATA_VECTOR,
+ BRW_MATH_PRECISION_FULL);
+ }
+ } else {
+ assert(inst->mlen >= 1);
- brw_math(p, dst,
- op,
- inst->saturate ? BRW_MATH_SATURATE_SATURATE :
- BRW_MATH_SATURATE_NONE,
- 2, src[0],
- BRW_MATH_DATA_VECTOR,
- BRW_MATH_PRECISION_FULL);
+ brw_math(p, dst,
+ op,
+ inst->saturate ? BRW_MATH_SATURATE_SATURATE :
+ BRW_MATH_SATURATE_NONE,
+ inst->base_mrf, src[0],
+ BRW_MATH_DATA_VECTOR,
+ BRW_MATH_PRECISION_FULL);
+ }
}
void
-fs_visitor::generate_tex(fs_inst *inst, struct brw_reg dst, struct brw_reg src)
+fs_visitor::generate_tex(fs_inst *inst, struct brw_reg dst)
{
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 == 6);
} else {
- msg_type = BRW_SAMPLER_MESSAGE_SIMD16_SAMPLE;
+ assert(inst->mlen <= 4);
}
+ 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 == 6);
+ msg_type = BRW_SAMPLER_MESSAGE_SIMD8_SAMPLE;
} else {
+ assert(inst->mlen == 9);
msg_type = BRW_SAMPLER_MESSAGE_SIMD16_SAMPLE_BIAS;
- rlen = 8;
+ simd_mode = BRW_SAMPLER_SIMD_MODE_SIMD16;
}
break;
}
}
assert(msg_type != -1);
- /* g0 header. */
- src.nr--;
+ if (simd_mode == BRW_SAMPLER_SIMD_MODE_SIMD16) {
+ rlen = 8;
+ dst = vec16(dst);
+ }
brw_SAMPLE(p,
retype(dst, BRW_REGISTER_TYPE_UW),
- src.nr,
+ inst->base_mrf,
retype(brw_vec8_grf(0, 0), BRW_REGISTER_TYPE_UW),
SURF_INDEX_TEXTURE(inst->sampler),
inst->sampler,
WRITEMASK_XYZW,
msg_type,
rlen,
- inst->mlen + 1,
+ inst->mlen,
0,
1,
- BRW_SAMPLER_SIMD_MODE_SIMD8);
+ simd_mode);
}
}
void
-fs_visitor::generate_discard(fs_inst *inst, struct brw_reg temp)
+fs_visitor::generate_discard_not(fs_inst *inst, struct brw_reg mask)
{
- 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, temp, brw_mask_reg(1)); /* IMASK */
- brw_AND(p, g0, temp, g0);
- brw_pop_insn_state(p);
+ if (intel->gen >= 6) {
+ /* Gen6 no longer has the mask reg for us to just read the
+ * active channels from. However, cmp updates just the channels
+ * of the flag reg that are enabled, so we can get at the
+ * channel enables that way. In this step, make a reg of ones
+ * we'll compare to.
+ */
+ brw_MOV(p, mask, brw_imm_ud(1));
+ } else {
+ brw_push_insn_state(p);
+ brw_set_mask_control(p, BRW_MASK_DISABLE);
+ brw_NOT(p, mask, brw_mask_reg(1)); /* IMASK */
+ brw_pop_insn_state(p);
+ }
}
void
-fs_visitor::assign_curb_setup()
+fs_visitor::generate_discard_and(fs_inst *inst, struct brw_reg mask)
{
- 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 >= 6) {
+ struct brw_reg f0 = brw_flag_reg();
+ struct brw_reg g1 = retype(brw_vec1_grf(1, 7), BRW_REGISTER_TYPE_UW);
+
+ brw_push_insn_state(p);
+ brw_set_mask_control(p, BRW_MASK_DISABLE);
+ brw_MOV(p, f0, brw_imm_uw(0xffff)); /* inactive channels undiscarded */
+ brw_pop_insn_state(p);
+
+ brw_CMP(p, retype(brw_null_reg(), BRW_REGISTER_TYPE_UD),
+ BRW_CONDITIONAL_Z, mask, brw_imm_ud(0)); /* active channels fail test */
+ /* Undo CMP's whacking of predication*/
+ brw_set_predicate_control(p, BRW_PREDICATE_NONE);
+
+ brw_push_insn_state(p);
+ brw_set_mask_control(p, BRW_MASK_DISABLE);
+ brw_AND(p, g1, f0, g1);
+ brw_pop_insn_state(p);
+ } else {
+ struct brw_reg g0 = retype(brw_vec1_grf(0, 0), BRW_REGISTER_TYPE_UW);
- 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++;
- }
+ mask = brw_uw1_reg(mask.file, mask.nr, 0);
- /* 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();
+ brw_push_insn_state(p);
+ brw_set_mask_control(p, BRW_MASK_DISABLE);
+ brw_AND(p, g0, mask, g0);
+ brw_pop_insn_state(p);
+ }
+}
- for (unsigned int i = 0; i < 3; i++) {
- if (inst->src[i].file == UNIFORM) {
- int constant_nr = inst->src[i].hw_reg + inst->src[i].reg_offset;
- struct brw_reg brw_reg = brw_vec1_grf(c->prog_data.first_curbe_grf +
- constant_nr / 8,
- constant_nr % 8);
+void
+fs_visitor::generate_spill(fs_inst *inst, struct brw_reg src)
+{
+ assert(inst->mlen != 0);
- inst->src[i].file = FIXED_HW_REG;
- inst->src[i].fixed_hw_reg = brw_reg;
- }
- }
- }
+ brw_MOV(p,
+ retype(brw_message_reg(inst->base_mrf + 1), BRW_REGISTER_TYPE_UD),
+ retype(src, BRW_REGISTER_TYPE_UD));
+ brw_oword_block_write_scratch(p, brw_message_reg(inst->base_mrf), 1,
+ inst->offset);
}
void
-fs_visitor::assign_urb_setup()
+fs_visitor::generate_unspill(fs_inst *inst, struct brw_reg dst)
{
- 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;
+ assert(inst->mlen != 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;
+ /* Clear any post destination dependencies that would be ignored by
+ * the block read. See the B-Spec for pre-gen5 send instruction.
+ *
+ * This could use a better solution, since texture sampling and
+ * math reads could potentially run into it as well -- anywhere
+ * that we have a SEND with a destination that is a register that
+ * was written but not read within the last N instructions (what's
+ * N? unsure). This is rare because of dead code elimination, but
+ * not impossible.
+ */
+ if (intel->gen == 4 && !intel->is_g4x)
+ brw_MOV(p, brw_null_reg(), dst);
- if (i != FRAG_ATTRIB_WPOS &&
- !(brw->fragment_program->Base.InputsRead & BITFIELD64_BIT(i)))
- continue;
+ brw_oword_block_read_scratch(p, dst, brw_message_reg(inst->base_mrf), 1,
+ inst->offset);
- /* 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;
+ if (intel->gen == 4 && !intel->is_g4x) {
+ /* gen4 errata: destination from a send can't be used as a
+ * destination until it's been read. Just read it so we don't
+ * have to worry.
+ */
+ brw_MOV(p, brw_null_reg(), dst);
}
+}
- /* Map the register numbers for FS_OPCODE_LINTERP so that it uses
- * the correct setup input.
- */
- foreach_iter(exec_list_iterator, iter, this->instructions) {
- fs_inst *inst = (fs_inst *)iter.get();
- if (inst->opcode != FS_OPCODE_LINTERP)
- continue;
+void
+fs_visitor::generate_pull_constant_load(fs_inst *inst, struct brw_reg dst)
+{
+ assert(inst->mlen != 0);
- assert(inst->src[2].file == FIXED_HW_REG);
+ /* Clear any post destination dependencies that would be ignored by
+ * the block read. See the B-Spec for pre-gen5 send instruction.
+ *
+ * This could use a better solution, since texture sampling and
+ * math reads could potentially run into it as well -- anywhere
+ * that we have a SEND with a destination that is a register that
+ * was written but not read within the last N instructions (what's
+ * N? unsure). This is rare because of dead code elimination, but
+ * not impossible.
+ */
+ if (intel->gen == 4 && !intel->is_g4x)
+ brw_MOV(p, brw_null_reg(), dst);
- 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));
- }
+ brw_oword_block_read(p, dst, brw_message_reg(inst->base_mrf),
+ inst->offset, SURF_INDEX_FRAG_CONST_BUFFER);
- this->first_non_payload_grf = urb_start + c->prog_data.urb_read_length;
+ if (intel->gen == 4 && !intel->is_g4x) {
+ /* gen4 errata: destination from a send can't be used as a
+ * destination until it's been read. Just read it so we don't
+ * have to worry.
+ */
+ brw_MOV(p, brw_null_reg(), dst);
+ }
}
-static void
-assign_reg(int *reg_hw_locations, fs_reg *reg)
+void
+fs_visitor::assign_curb_setup()
{
- if (reg->file == GRF && reg->reg != 0) {
- reg->hw_reg = reg_hw_locations[reg->reg] + reg->reg_offset;
- reg->reg = 0;
+ 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;
+
+ /* 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();
+
+ for (unsigned int i = 0; i < 3; i++) {
+ if (inst->src[i].file == UNIFORM) {
+ int constant_nr = inst->src[i].hw_reg + inst->src[i].reg_offset;
+ struct brw_reg brw_reg = brw_vec1_grf(c->prog_data.first_curbe_grf +
+ constant_nr / 8,
+ constant_nr % 8);
+
+ inst->src[i].file = FIXED_HW_REG;
+ inst->src[i].fixed_hw_reg = retype(brw_reg, inst->src[i].type);
+ }
+ }
}
}
void
-fs_visitor::assign_regs_trivial()
+fs_visitor::calculate_urb_setup()
{
- int last_grf = 0;
- int hw_reg_mapping[this->virtual_grf_next];
- int i;
+ for (unsigned int i = 0; i < FRAG_ATTRIB_MAX; i++) {
+ urb_setup[i] = -1;
+ }
- 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]);
+ int urb_next = 0;
+ /* Figure out where each of the incoming setup attributes lands. */
+ 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++;
+ }
+ }
}
- last_grf = hw_reg_mapping[i - 1] + this->virtual_grf_sizes[i - 1];
+ /* Each attribute is 4 setup channels, each of which is half a reg. */
+ c->prog_data.urb_read_length = urb_next * 2;
+}
+
+void
+fs_visitor::assign_urb_setup()
+{
+ int urb_start = c->prog_data.first_curbe_grf + c->prog_data.curb_read_length;
+
+ /* 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();
- assign_reg(hw_reg_mapping, &inst->dst);
- assign_reg(hw_reg_mapping, &inst->src[0]);
- assign_reg(hw_reg_mapping, &inst->src[1]);
+ if (inst->opcode != FS_OPCODE_LINTERP)
+ continue;
+
+ assert(inst->src[2].file == FIXED_HW_REG);
+
+ inst->src[2].fixed_hw_reg.nr += urb_start;
}
- this->grf_used = last_grf + 1;
+ this->first_non_payload_grf = urb_start + c->prog_data.urb_read_length;
}
+/**
+ * Split large virtual GRFs into separate components if we can.
+ *
+ * This is mostly duplicated with what brw_fs_vector_splitting does,
+ * but that's really conservative because it's afraid of doing
+ * splitting that doesn't result in real progress after the rest of
+ * the optimization phases, which would cause infinite looping in
+ * optimization. We can do it once here, safely. This also has the
+ * opportunity to split interpolated values, or maybe even uniforms,
+ * which we don't have at the IR level.
+ *
+ * We want to split, because virtual GRFs are what we register
+ * allocate and spill (due to contiguousness requirements for some
+ * instructions), and they're what we naturally generate in the
+ * codegen process, but most virtual GRFs don't actually need to be
+ * contiguous sets of GRFs. If we split, we'll end up with reduced
+ * live intervals and better dead code elimination and coalescing.
+ */
void
-fs_visitor::assign_regs()
+fs_visitor::split_virtual_grfs()
{
- 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;
-
- calculate_live_intervals();
-
- /* 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;
- for (int r = 1; r < this->virtual_grf_next; r++) {
- int i;
+ int num_vars = this->virtual_grf_next;
+ bool split_grf[num_vars];
+ int new_virtual_grf[num_vars];
- for (i = 0; i < class_count; i++) {
- if (class_sizes[i] == this->virtual_grf_sizes[r])
- break;
- }
- if (i == class_count) {
- class_sizes[class_count++] = this->virtual_grf_sizes[r];
- }
+ /* Try to split anything > 0 sized. */
+ for (int i = 0; i < num_vars; i++) {
+ if (this->virtual_grf_sizes[i] != 1)
+ split_grf[i] = true;
+ else
+ split_grf[i] = false;
}
- int ra_reg_count = 0;
- int class_base_reg[class_count];
- int class_reg_count[class_count];
- int classes[class_count];
-
- 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];
+ if (brw->has_pln) {
+ /* PLN opcodes rely on the delta_xy being contiguous. */
+ split_grf[this->delta_x.reg] = false;
}
- 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);
+ foreach_iter(exec_list_iterator, iter, this->instructions) {
+ fs_inst *inst = (fs_inst *)iter.get();
- 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);
+ /* Texturing produces 4 contiguous registers, so no splitting. */
+ if ((inst->opcode == FS_OPCODE_TEX ||
+ inst->opcode == FS_OPCODE_TXB ||
+ inst->opcode == FS_OPCODE_TXL) &&
+ inst->dst.file == GRF) {
+ split_grf[inst->dst.reg] = false;
}
+ }
- /* 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] - 1; i_r++) {
- for (int c_r = MAX2(0, i_r - (class_sizes[c] - 1));
- c_r <= MIN2(class_reg_count[c] - 1, i_r + class_sizes[i] - 1);
- c_r++) {
-
- if (0) {
- printf("%d/%d conflicts %d/%d\n",
- class_sizes[i], i_r,
- class_sizes[c], c_r);
- }
-
- ra_add_reg_conflict(regs,
- class_base_reg[i] + i_r,
- class_base_reg[c] + c_r);
- }
+ /* Allocate new space for split regs. Note that the virtual
+ * numbers will be contiguous.
+ */
+ for (int i = 0; i < num_vars; i++) {
+ if (split_grf[i]) {
+ new_virtual_grf[i] = virtual_grf_alloc(1);
+ for (int j = 2; j < this->virtual_grf_sizes[i]; j++) {
+ int reg = virtual_grf_alloc(1);
+ assert(reg == new_virtual_grf[i] + j - 1);
}
+ this->virtual_grf_sizes[i] = 1;
}
}
- 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]);
+ foreach_iter(exec_list_iterator, iter, this->instructions) {
+ fs_inst *inst = (fs_inst *)iter.get();
- 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]) {
- ra_set_node_class(g, i, classes[c]);
- break;
- }
+ if (inst->dst.file == GRF &&
+ split_grf[inst->dst.reg] &&
+ inst->dst.reg_offset != 0) {
+ inst->dst.reg = (new_virtual_grf[inst->dst.reg] +
+ inst->dst.reg_offset - 1);
+ inst->dst.reg_offset = 0;
}
-
- for (int j = 1; j < i; j++) {
- if (virtual_grf_interferes(i, j)) {
- ra_add_node_interference(g, i, j);
+ for (int i = 0; i < 3; i++) {
+ if (inst->src[i].file == GRF &&
+ split_grf[inst->src[i].reg] &&
+ inst->src[i].reg_offset != 0) {
+ inst->src[i].reg = (new_virtual_grf[inst->src[i].reg] +
+ inst->src[i].reg_offset - 1);
+ inst->src[i].reg_offset = 0;
}
}
}
+}
- /* FINISHME: Handle spilling */
- if (!ra_allocate_no_spills(g)) {
- fprintf(stderr, "Failed to allocate registers.\n");
- this->fail = true;
+/**
+ * Choose accesses from the UNIFORM file to demote to using the pull
+ * constant buffer.
+ *
+ * We allow a fragment shader to have more than the specified minimum
+ * maximum number of fragment shader uniform components (64). If
+ * there are too many of these, they'd fill up all of register space.
+ * So, this will push some of them out to the pull constant buffer and
+ * update the program to load them.
+ */
+void
+fs_visitor::setup_pull_constants()
+{
+ /* Only allow 16 registers (128 uniform components) as push constants. */
+ unsigned int max_uniform_components = 16 * 8;
+ if (c->prog_data.nr_params <= max_uniform_components)
return;
- }
- /* Get the chosen virtual registers for each node, and map virtual
- * regs in the register classes back down to real hardware reg
- * numbers.
+ /* Just demote the end of the list. We could probably do better
+ * here, demoting things that are rarely used in the program first.
*/
- 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] - 1) {
- 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);
- }
+ int pull_uniform_base = max_uniform_components;
+ int pull_uniform_count = c->prog_data.nr_params - pull_uniform_base;
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]);
- }
+ for (int i = 0; i < 3; i++) {
+ if (inst->src[i].file != UNIFORM)
+ continue;
- this->grf_used = last_grf + 1;
+ int uniform_nr = inst->src[i].hw_reg + inst->src[i].reg_offset;
+ if (uniform_nr < pull_uniform_base)
+ continue;
- talloc_free(g);
- talloc_free(regs);
+ fs_reg dst = fs_reg(this, glsl_type::float_type);
+ fs_inst *pull = new(mem_ctx) fs_inst(FS_OPCODE_PULL_CONSTANT_LOAD,
+ dst);
+ pull->offset = ((uniform_nr - pull_uniform_base) * 4) & ~15;
+ pull->ir = inst->ir;
+ pull->annotation = inst->annotation;
+ pull->base_mrf = 14;
+ pull->mlen = 1;
+
+ inst->insert_before(pull);
+
+ inst->src[i].file = GRF;
+ inst->src[i].reg = dst.reg;
+ inst->src[i].reg_offset = 0;
+ inst->src[i].smear = (uniform_nr - pull_uniform_base) & 3;
+ }
+ }
+
+ for (int i = 0; i < pull_uniform_count; i++) {
+ c->prog_data.pull_param[i] = c->prog_data.param[pull_uniform_base + i];
+ c->prog_data.pull_param_convert[i] =
+ c->prog_data.param_convert[pull_uniform_base + i];
+ }
+ c->prog_data.nr_params -= pull_uniform_count;
+ c->prog_data.nr_pull_params = pull_uniform_count;
}
void
int *use = talloc_array(mem_ctx, int, num_vars);
int loop_depth = 0;
int loop_start = 0;
+ int bb_header_ip = 0;
for (int i = 0; i < num_vars; i++) {
def[i] = 1 << 30;
- use[i] = 0;
+ use[i] = -1;
}
int ip = 0;
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.
+ /* Patches up the use of vars marked for being live across
+ * the whole loop.
*/
for (int i = 0; i < num_vars; i++) {
if (use[i] == loop_start) {
}
}
} 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) {
- def[inst->src[i].reg] = MIN2(def[inst->src[i].reg], eip);
- use[inst->src[i].reg] = MAX2(use[inst->src[i].reg], eip);
+ int reg = inst->src[i].reg;
+
+ if (!loop_depth || (this->virtual_grf_sizes[reg] == 1 &&
+ def[reg] >= bb_header_ip)) {
+ use[reg] = ip;
+ } else {
+ def[reg] = MIN2(loop_start, def[reg]);
+ use[reg] = loop_start;
+
+ /* Nobody else is going to go smash our start to
+ * later in the loop now, because def[reg] now
+ * points before the bb header.
+ */
+ }
}
}
if (inst->dst.file == GRF && inst->dst.reg != 0) {
- def[inst->dst.reg] = MIN2(def[inst->dst.reg], eip);
- use[inst->dst.reg] = MAX2(use[inst->dst.reg], eip);
+ int reg = inst->dst.reg;
+
+ if (!loop_depth || (this->virtual_grf_sizes[reg] == 1 &&
+ !inst->predicated)) {
+ def[reg] = MIN2(def[reg], ip);
+ } else {
+ def[reg] = MIN2(def[reg], loop_start);
+ }
}
}
ip++;
+
+ /* Set the basic block header IP. This is used for determining
+ * if a complete def of single-register virtual GRF in a loop
+ * dominates a use in the same basic block. It's a quick way to
+ * reduce the live interval range of most register used in a
+ * loop.
+ */
+ if (inst->opcode == BRW_OPCODE_IF ||
+ inst->opcode == BRW_OPCODE_ELSE ||
+ inst->opcode == BRW_OPCODE_ENDIF ||
+ inst->opcode == BRW_OPCODE_DO ||
+ inst->opcode == BRW_OPCODE_WHILE ||
+ inst->opcode == BRW_OPCODE_BREAK ||
+ inst->opcode == BRW_OPCODE_CONTINUE) {
+ bb_header_ip = 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++) {
+ 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::register_coalesce()
+{
+ 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->saturate ||
+ inst->dst.file != GRF || inst->src[0].file != GRF ||
+ inst->dst.type != inst->src[0].type)
+ continue;
+
+ /* Found a move of a GRF to a GRF. Let's see if we can coalesce
+ * them: check for no writes to either one until the exit of the
+ * program.
+ */
+ bool interfered = false;
+ 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_ENDIF) {
+ interfered = true;
+ iter = scan_iter;
+ break;
+ }
+
+ if (scan_inst->dst.file == GRF) {
+ if (scan_inst->dst.reg == inst->dst.reg &&
+ (scan_inst->dst.reg_offset == inst->dst.reg_offset ||
+ scan_inst->opcode == FS_OPCODE_TEX)) {
+ interfered = true;
+ break;
+ }
+ if (scan_inst->dst.reg == inst->src[0].reg &&
+ (scan_inst->dst.reg_offset == inst->src[0].reg_offset ||
+ scan_inst->opcode == FS_OPCODE_TEX)) {
+ interfered = true;
+ break;
+ }
+ }
+ }
+ if (interfered) {
+ continue;
+ }
+
+ /* Update live interval so we don't have to recalculate. */
+ this->virtual_grf_use[inst->src[0].reg] = MAX2(virtual_grf_use[inst->src[0].reg],
+ virtual_grf_use[inst->dst.reg]);
+
+ /* Rewrite the later usage to point at the source of the move to
+ * be removed.
+ */
+ for (exec_list_iterator scan_iter = iter; scan_iter.has_next();
+ scan_iter.next()) {
+ fs_inst *scan_inst = (fs_inst *)scan_iter.get();
+
+ for (int i = 0; i < 3; 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) {
+ scan_inst->src[i].reg = inst->src[0].reg;
+ scan_inst->src[i].reg_offset = inst->src[0].reg_offset;
+ scan_inst->src[i].abs |= inst->src[0].abs;
+ scan_inst->src[i].negate ^= inst->src[0].negate;
+ scan_inst->src[i].smear = inst->src[0].smear;
+ }
+ }
+ }
+
+ inst->remove();
+ progress = true;
+ }
+
+ return progress;
+}
+
+
+bool
+fs_visitor::compute_to_mrf()
+{
+ bool progress = false;
+ int next_ip = 0;
+
+ foreach_iter(exec_list_iterator, iter, this->instructions) {
+ fs_inst *inst = (fs_inst *)iter.get();
+
+ int ip = next_ip;
+ next_ip++;
+
+ if (inst->opcode != BRW_OPCODE_MOV ||
+ inst->predicated ||
+ inst->dst.file != MRF || inst->src[0].file != GRF ||
+ inst->dst.type != inst->src[0].type ||
+ inst->src[0].abs || inst->src[0].negate || inst->src[0].smear != -1)
+ continue;
+
+ /* Can't compute-to-MRF this GRF if someone else was going to
+ * read it later.
+ */
+ if (this->virtual_grf_use[inst->src[0].reg] > ip)
+ continue;
+
+ /* Found a move of a GRF to a MRF. Let's see if we can go
+ * rewrite the thing that made this GRF to write into the MRF.
+ */
+ bool found = false;
+ fs_inst *scan_inst;
+ for (scan_inst = (fs_inst *)inst->prev;
+ scan_inst->prev != NULL;
+ scan_inst = (fs_inst *)scan_inst->prev) {
+ /* We don't handle flow control here. Most computation of
+ * values that end up in MRFs are shortly before the MRF
+ * write anyway.
+ */
+ if (scan_inst->opcode == BRW_OPCODE_DO ||
+ scan_inst->opcode == BRW_OPCODE_WHILE ||
+ scan_inst->opcode == BRW_OPCODE_ENDIF) {
+ break;
+ }
+
+ /* You can't read from an MRF, so if someone else reads our
+ * MRF's source GRF that we wanted to rewrite, that stops us.
+ */
+ bool interfered = false;
+ for (int i = 0; i < 3; i++) {
+ if (scan_inst->src[i].file == GRF &&
+ scan_inst->src[i].reg == inst->src[0].reg &&
+ scan_inst->src[i].reg_offset == inst->src[0].reg_offset) {
+ interfered = true;
+ }
+ }
+ if (interfered)
+ break;
+
+ if (scan_inst->dst.file == MRF &&
+ scan_inst->dst.hw_reg == inst->dst.hw_reg) {
+ /* Somebody else wrote our MRF here, so we can't can't
+ * compute-to-MRF before that.
+ */
+ break;
+ }
+
+ if (scan_inst->mlen > 0) {
+ /* Found a SEND instruction, which will do some amount of
+ * implied write that may overwrite our MRF that we were
+ * hoping to compute-to-MRF somewhere above it. Nothing
+ * we have implied-writes more than 2 MRFs from base_mrf,
+ * though.
+ */
+ int implied_write_len = MIN2(scan_inst->mlen, 2);
+ if (inst->dst.hw_reg >= scan_inst->base_mrf &&
+ inst->dst.hw_reg < scan_inst->base_mrf + implied_write_len) {
+ break;
+ }
+ }
+
+ if (scan_inst->dst.file == GRF &&
+ scan_inst->dst.reg == inst->src[0].reg) {
+ /* Found the last thing to write our reg we want to turn
+ * into a compute-to-MRF.
+ */
+
+ if (scan_inst->opcode == FS_OPCODE_TEX) {
+ /* texturing writes several continuous regs, so we can't
+ * compute-to-mrf that.
+ */
+ break;
+ }
+
+ /* If it's predicated, it (probably) didn't populate all
+ * the channels.
+ */
+ if (scan_inst->predicated)
+ break;
+
+ /* SEND instructions can't have MRF as a destination. */
+ if (scan_inst->mlen)
+ break;
+
+ if (intel->gen >= 6) {
+ /* gen6 math instructions must have the destination be
+ * GRF, so no compute-to-MRF for them.
+ */
+ if (scan_inst->opcode == FS_OPCODE_RCP ||
+ scan_inst->opcode == FS_OPCODE_RSQ ||
+ scan_inst->opcode == FS_OPCODE_SQRT ||
+ scan_inst->opcode == FS_OPCODE_EXP2 ||
+ scan_inst->opcode == FS_OPCODE_LOG2 ||
+ scan_inst->opcode == FS_OPCODE_SIN ||
+ scan_inst->opcode == FS_OPCODE_COS ||
+ scan_inst->opcode == FS_OPCODE_POW) {
+ break;
+ }
+ }
+
+ if (scan_inst->dst.reg_offset == inst->src[0].reg_offset) {
+ /* Found the creator of our MRF's source value. */
+ found = true;
+ break;
+ }
+ }
+ }
+ if (found) {
+ scan_inst->dst.file = MRF;
+ scan_inst->dst.hw_reg = inst->dst.hw_reg;
+ scan_inst->saturate |= inst->saturate;
+ 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]);
- return start <= end;
+ /* 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 GRF:
case ARF:
case MRF:
- brw_reg = brw_vec8_reg(reg->file,
- reg->hw_reg, 0);
+ if (reg->smear == -1) {
+ brw_reg = brw_vec8_reg(reg->file,
+ reg->hw_reg, 0);
+ } else {
+ brw_reg = brw_vec1_reg(reg->file,
+ reg->hw_reg, reg->smear);
+ }
brw_reg = retype(brw_reg, reg->type);
break;
case IMM:
void
fs_visitor::generate_code()
{
- unsigned int annotation_len = 0;
int last_native_inst = 0;
struct brw_instruction *if_stack[16], *loop_stack[16];
int if_stack_depth = 0, loop_stack_depth = 0;
int if_depth_in_loop[16];
+ const char *last_annotation_string = NULL;
+ ir_instruction *last_annotation_ir = NULL;
+
+ if (unlikely(INTEL_DEBUG & DEBUG_WM)) {
+ printf("Native code for fragment shader %d:\n",
+ ctx->Shader.CurrentFragmentProgram->Name);
+ }
if_depth_in_loop[loop_stack_depth] = 0;
fs_inst *inst = (fs_inst *)iter.get();
struct brw_reg src[3], dst;
+ if (unlikely(INTEL_DEBUG & DEBUG_WM)) {
+ if (last_annotation_ir != inst->ir) {
+ last_annotation_ir = inst->ir;
+ if (last_annotation_ir) {
+ printf(" ");
+ last_annotation_ir->print();
+ printf("\n");
+ }
+ }
+ if (last_annotation_string != inst->annotation) {
+ last_annotation_string = inst->annotation;
+ if (last_annotation_string)
+ printf(" %s\n", last_annotation_string);
+ }
+ }
+
for (unsigned int i = 0; i < 3; i++) {
src[i] = brw_reg_from_fs_reg(&inst->src[i]);
}
case BRW_OPCODE_RNDD:
brw_RNDD(p, dst, src[0]);
break;
+ case BRW_OPCODE_RNDE:
+ brw_RNDE(p, dst, src[0]);
+ break;
case BRW_OPCODE_RNDZ:
brw_RNDZ(p, dst, src[0]);
break;
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]);
case BRW_OPCODE_IF:
assert(if_stack_depth < 16);
- if_stack[if_stack_depth] = brw_IF(p, BRW_EXECUTE_8);
+ if (inst->src[0].file != BAD_FILE) {
+ assert(intel->gen >= 6);
+ if_stack[if_stack_depth] = brw_IF_gen6(p, inst->conditional_mod, src[0], src[1]);
+ } else {
+ if_stack[if_stack_depth] = brw_IF(p, BRW_EXECUTE_8);
+ }
if_depth_in_loop[loop_stack_depth]++;
if_stack_depth++;
break;
+
case BRW_OPCODE_ELSE:
if_stack[if_stack_depth - 1] =
brw_ELSE(p, if_stack[if_stack_depth - 1]);
struct brw_instruction *inst0, *inst1;
GLuint br = 1;
- if (intel->gen == 5)
+ if (intel->gen >= 5)
br = 2;
assert(loop_stack_depth > 0);
case FS_OPCODE_TEX:
case FS_OPCODE_TXB:
case FS_OPCODE_TXL:
- generate_tex(inst, dst, src[0]);
+ generate_tex(inst, dst);
break;
- case FS_OPCODE_DISCARD:
- generate_discard(inst, dst /* src0 == dst */);
+ case FS_OPCODE_DISCARD_NOT:
+ generate_discard_not(inst, dst);
+ break;
+ case FS_OPCODE_DISCARD_AND:
+ generate_discard_and(inst, src[0]);
break;
case FS_OPCODE_DDX:
generate_ddx(inst, dst, src[0]);
case FS_OPCODE_DDY:
generate_ddy(inst, dst, src[0]);
break;
+
+ case FS_OPCODE_SPILL:
+ generate_spill(inst, src[0]);
+ break;
+
+ case FS_OPCODE_UNSPILL:
+ generate_unspill(inst, dst);
+ break;
+
+ case FS_OPCODE_PULL_CONSTANT_LOAD:
+ generate_pull_constant_load(inst, dst);
+ break;
+
case FS_OPCODE_FB_WRITE:
generate_fb_write(inst);
break;
this->fail = true;
}
- if (annotation_len < p->nr_insn) {
- annotation_len *= 2;
- if (annotation_len < 16)
- annotation_len = 16;
-
- this->annotation_string = talloc_realloc(this->mem_ctx,
- annotation_string,
- const char *,
- annotation_len);
- this->annotation_ir = talloc_realloc(this->mem_ctx,
- annotation_ir,
- ir_instruction *,
- annotation_len);
+ if (unlikely(INTEL_DEBUG & DEBUG_WM)) {
+ for (unsigned int i = last_native_inst; i < p->nr_insn; i++) {
+ if (0) {
+ printf("0x%08x 0x%08x 0x%08x 0x%08x ",
+ ((uint32_t *)&p->store[i])[3],
+ ((uint32_t *)&p->store[i])[2],
+ ((uint32_t *)&p->store[i])[1],
+ ((uint32_t *)&p->store[i])[0]);
+ }
+ brw_disasm(stdout, &p->store[i], intel->gen);
+ printf("\n");
+ }
}
- for (unsigned int i = last_native_inst; i < p->nr_insn; i++) {
- this->annotation_string[i] = inst->annotation;
- this->annotation_ir[i] = inst->ir;
- }
last_native_inst = p->nr_insn;
}
}
GLboolean
brw_wm_fs_emit(struct brw_context *brw, struct brw_wm_compile *c)
{
- struct brw_compile *p = &c->func;
struct intel_context *intel = &brw->intel;
- GLcontext *ctx = &intel->ctx;
- struct brw_shader *shader = NULL;
- struct gl_shader_program *prog = ctx->Shader.CurrentProgram;
+ struct gl_context *ctx = &intel->ctx;
+ struct gl_shader_program *prog = ctx->Shader.CurrentFragmentProgram;
if (!prog)
return GL_FALSE;
- if (!using_new_fs)
- return GL_FALSE;
-
- for (unsigned int i = 0; i < prog->_NumLinkedShaders; i++) {
- if (prog->_LinkedShaders[i]->Type == GL_FRAGMENT_SHADER) {
- shader = (struct brw_shader *)prog->_LinkedShaders[i];
- break;
- }
- }
+ struct brw_shader *shader =
+ (brw_shader *) prog->_LinkedShaders[MESA_SHADER_FRAGMENT];
if (!shader)
return GL_FALSE;
*/
c->dispatch_width = 8;
- if (INTEL_DEBUG & DEBUG_WM) {
+ if (unlikely(INTEL_DEBUG & DEBUG_WM)) {
printf("GLSL IR for native fragment shader %d:\n", prog->Name);
_mesa_print_ir(shader->ir, NULL);
printf("\n");
if (0) {
v.emit_dummy_fs();
} else {
- v.emit_interpolation_setup();
+ 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.split_virtual_grfs();
+ v.setup_pull_constants();
+
v.assign_curb_setup();
v.assign_urb_setup();
+
+ bool progress;
+ do {
+ progress = false;
+ v.calculate_live_intervals();
+ progress = v.propagate_constants() || progress;
+ progress = v.register_coalesce() || progress;
+ progress = v.compute_to_mrf() || progress;
+ progress = v.dead_code_eliminate() || progress;
+ } while (progress);
+
+ if (0) {
+ /* Debug of register spilling: Go spill everything. */
+ int virtual_grf_count = v.virtual_grf_next;
+ for (int i = 1; i < virtual_grf_count; i++) {
+ v.spill_reg(i);
+ }
+ v.calculate_live_intervals();
+ }
+
if (0)
v.assign_regs_trivial();
- else
- v.assign_regs();
+ else {
+ while (!v.assign_regs()) {
+ if (v.fail)
+ break;
+
+ v.calculate_live_intervals();
+ }
+ }
}
- v.generate_code();
+ if (!v.fail)
+ v.generate_code();
assert(!v.fail); /* FINISHME: Cleanly fail, tested at link time, etc. */
if (v.fail)
return GL_FALSE;
- if (INTEL_DEBUG & DEBUG_WM) {
- const char *last_annotation_string = NULL;
- ir_instruction *last_annotation_ir = NULL;
-
- printf("Native code for fragment shader %d:\n", prog->Name);
- for (unsigned int i = 0; i < p->nr_insn; i++) {
- if (last_annotation_ir != v.annotation_ir[i]) {
- last_annotation_ir = v.annotation_ir[i];
- if (last_annotation_ir) {
- printf(" ");
- last_annotation_ir->print();
- printf("\n");
- }
- }
- if (last_annotation_string != v.annotation_string[i]) {
- last_annotation_string = v.annotation_string[i];
- if (last_annotation_string)
- printf(" %s\n", last_annotation_string);
- }
- brw_disasm(stdout, &p->store[i], intel->gen);
- }
- printf("\n");
- }
-
c->prog_data.total_grf = v.grf_used;
- c->prog_data.total_scratch = 0;
return GL_TRUE;
}
projects / mesa.git / blobdiff