}
#include "brw_fs.h"
#include "glsl/glsl_types.h"
-#include "glsl/ir_print_visitor.h"
void
fs_inst::init()
ALU2(SHR)
ALU2(ASR)
ALU3(LRP)
+ALU1(BFREV)
+ALU3(BFE)
+ALU2(BFI1)
+ALU3(BFI2)
+ALU1(FBH)
+ALU1(FBL)
+ALU1(CBIT)
+ALU3(MAD)
/** Gen4 predicated IF. */
fs_inst *
fs_inst *
fs_visitor::IF(fs_reg src0, fs_reg src1, uint32_t condition)
{
- assert(intel->gen >= 6);
+ assert(brw->gen >= 6);
fs_inst *inst = new(mem_ctx) fs_inst(BRW_OPCODE_IF,
reg_null_d, src0, src1);
inst->conditional_mod = condition;
* mostly work out for float-interpreted-as-int since our comparisons are
* for >0, =0, <0.
*/
- if (intel->gen == 4) {
+ if (brw->gen == 4) {
dst.type = src0.type;
- if (dst.file == FIXED_HW_REG)
+ if (dst.file == HW_REG)
dst.fixed_hw_reg.type = dst.type;
}
varying_offset, const_offset & ~3));
int scale = 1;
- if (intel->gen == 4 && dispatch_width == 8) {
+ if (brw->gen == 4 && dispatch_width == 8) {
/* Pre-gen5, we can either use a SIMD8 message that requires (header,
* u, v, r) as parameters, or we can just use the SIMD16 message
* consisting of (header, u). We choose the second, at the cost of a
}
enum opcode op;
- if (intel->gen >= 7)
+ if (brw->gen >= 7)
op = FS_OPCODE_VARYING_PULL_CONSTANT_LOAD_GEN7;
else
op = FS_OPCODE_VARYING_PULL_CONSTANT_LOAD;
inst->regs_written = 4 * scale;
instructions.push_tail(inst);
- if (intel->gen < 7) {
+ if (brw->gen < 7) {
inst->base_mrf = 13;
inst->header_present = true;
- if (intel->gen == 4)
+ if (brw->gen == 4)
inst->mlen = 3;
else
inst->mlen = 1 + dispatch_width / 8;
bool
fs_visitor::can_do_source_mods(fs_inst *inst)
{
- if (intel->gen == 6 && inst->is_math())
+ if (brw->gen == 6 && inst->is_math())
return false;
if (inst->is_send_from_grf())
fs_reg::fs_reg(struct brw_reg fixed_hw_reg)
{
init();
- this->file = FIXED_HW_REG;
+ this->file = HW_REG;
this->fixed_hw_reg = fixed_hw_reg;
this->type = fixed_hw_reg.type;
}
fs_reg
fs_visitor::get_timestamp()
{
- assert(intel->gen >= 7);
+ assert(brw->gen >= 7);
fs_reg ts = fs_reg(retype(brw_vec1_reg(BRW_ARCHITECTURE_REGISTER_FILE,
BRW_ARF_TIMESTAMP,
bool
fs_inst::is_partial_write()
{
- return (this->predicate ||
+ return ((this->predicate && this->opcode != BRW_OPCODE_SEL) ||
this->force_uncompressed ||
this->force_sechalf);
}
import_uniforms_callback,
variable_ht);
this->params_remap = v->params_remap;
+ this->nr_params_remap = v->nr_params_remap;
}
/* Our support for uniforms is piggy-backed on the struct
wpos.reg_offset++;
/* gl_FragCoord.z */
- if (intel->gen >= 6) {
+ if (brw->gen >= 6) {
emit(MOV(wpos, fs_reg(brw_vec8_grf(c->source_depth_reg, 0))));
} else {
emit(FS_OPCODE_LINTERP, wpos,
bool is_centroid)
{
brw_wm_barycentric_interp_mode barycoord_mode;
- if (intel->gen >= 6) {
+ if (brw->gen >= 6) {
if (is_centroid) {
if (interpolation_mode == INTERP_QUALIFIER_SMOOTH)
barycoord_mode = BRW_WM_PERSPECTIVE_CENTROID_BARYCENTRIC;
inst->predicate = BRW_PREDICATE_NORMAL;
inst->predicate_inverse = true;
}
- if (intel->gen < 6) {
+ if (brw->gen < 6 && interpolation_mode == INTERP_QUALIFIER_SMOOTH) {
emit(BRW_OPCODE_MUL, attr, attr, this->pixel_w);
}
attr.reg_offset++;
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) {
+ if (brw->gen >= 6) {
emit(BRW_OPCODE_ASR, *reg,
fs_reg(retype(brw_vec1_grf(0, 0), BRW_REGISTER_TYPE_D)),
fs_reg(15));
* The hardware ignores source modifiers (negate and abs) on math
* instructions, so we also move to a temp to set those up.
*/
- if (intel->gen == 6 && src.file != UNIFORM && src.file != IMM &&
+ if (brw->gen == 6 && src.file != UNIFORM && src.file != IMM &&
!src.abs && !src.negate)
return src;
/* Gen7 relaxes most of the above restrictions, but still can't use IMM
* operands to math
*/
- if (intel->gen >= 7 && src.file != IMM)
+ if (brw->gen >= 7 && src.file != IMM)
return src;
fs_reg expanded = fs_reg(this, glsl_type::float_type);
* Gen 6 hardware ignores source modifiers (negate and abs) on math
* instructions, so we also move to a temp to set those up.
*/
- if (intel->gen >= 6)
+ if (brw->gen >= 6)
src = fix_math_operand(src);
fs_inst *inst = emit(opcode, dst, src);
- if (intel->gen < 6) {
+ if (brw->gen < 6) {
inst->base_mrf = 2;
inst->mlen = dispatch_width / 8;
}
switch (opcode) {
case SHADER_OPCODE_INT_QUOTIENT:
case SHADER_OPCODE_INT_REMAINDER:
- if (intel->gen >= 7 && dispatch_width == 16)
+ if (brw->gen >= 7 && dispatch_width == 16)
fail("16-wide INTDIV unsupported\n");
break;
case SHADER_OPCODE_POW:
return NULL;
}
- if (intel->gen >= 6) {
+ if (brw->gen >= 6) {
src0 = fix_math_operand(src0);
src1 = fix_math_operand(src1);
constant_nr / 8,
constant_nr % 8);
- inst->src[i].file = FIXED_HW_REG;
+ inst->src[i].file = HW_REG;
inst->src[i].fixed_hw_reg = retype(brw_reg, inst->src[i].type);
}
}
int urb_next = 0;
/* Figure out where each of the incoming setup attributes lands. */
- if (intel->gen >= 6) {
+ if (brw->gen >= 6) {
for (unsigned int i = 0; i < VARYING_SLOT_MAX; i++) {
if (fp->Base.InputsRead & BITFIELD64_BIT(i)) {
urb_setup[i] = urb_next++;
fs_inst *inst = (fs_inst *)node;
if (inst->opcode == FS_OPCODE_LINTERP) {
- assert(inst->src[2].file == FIXED_HW_REG);
+ assert(inst->src[2].file == HW_REG);
inst->src[2].fixed_hw_reg.nr += urb_start;
}
if (inst->opcode == FS_OPCODE_CINTERP) {
- assert(inst->src[0].file == FIXED_HW_REG);
+ assert(inst->src[0].file == HW_REG);
inst->src[0].fixed_hw_reg.nr += urb_start;
}
}
* the send is reading the whole thing.
*/
if (inst->is_send_from_grf()) {
- split_grf[inst->src[0].reg] = false;
+ for (int i = 0; i < 3; i++) {
+ if (inst->src[i].file == GRF) {
+ split_grf[inst->src[i].reg] = false;
+ }
+ }
}
}
remap_table[i] = new_index;
virtual_grf_sizes[new_index] = virtual_grf_sizes[i];
if (live_intervals_valid) {
- virtual_grf_use[new_index] = virtual_grf_use[i];
- virtual_grf_def[new_index] = virtual_grf_def[i];
+ virtual_grf_start[new_index] = virtual_grf_start[i];
+ virtual_grf_end[new_index] = virtual_grf_end[i];
}
++new_index;
}
{
if (dispatch_width == 8) {
this->params_remap = ralloc_array(mem_ctx, int, c->prog_data.nr_params);
+ this->nr_params_remap = c->prog_data.nr_params;
for (unsigned int i = 0; i < c->prog_data.nr_params; i++)
this->params_remap[i] = -1;
if (inst->src[i].file != UNIFORM)
continue;
- assert(constant_nr < (int)c->prog_data.nr_params);
+ /* Section 5.11 of the OpenGL 4.3 spec says:
+ *
+ * "Out-of-bounds reads return undefined values, which include
+ * values from other variables of the active program or zero."
+ */
+ if (constant_nr < 0 || constant_nr >= (int)c->prog_data.nr_params) {
+ constant_nr = 0;
+ }
/* For now, set this to non-negative. We'll give it the
* actual new number in a moment, in order to keep the
if (inst->src[i].file != UNIFORM)
continue;
+ /* as above alias to 0 */
+ if (constant_nr < 0 || constant_nr >= (int)this->nr_params_remap) {
+ constant_nr = 0;
+ }
assert(this->params_remap[constant_nr] != -1);
inst->src[i].reg = this->params_remap[constant_nr];
inst->src[i].reg_offset = 0;
}
/**
- * 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.
+ * Removes any instructions writing a VGRF where that VGRF is not used by any
+ * later instruction.
*/
bool
fs_visitor::dead_code_eliminate()
foreach_list_safe(node, &this->instructions) {
fs_inst *inst = (fs_inst *)node;
- if (inst->dst.file == GRF && this->virtual_grf_use[inst->dst.reg] <= pc) {
- inst->remove();
- progress = true;
+ if (inst->dst.file == GRF) {
+ assert(this->virtual_grf_end[inst->dst.reg] >= pc);
+ if (this->virtual_grf_end[inst->dst.reg] == pc) {
+ inst->remove();
+ progress = true;
+ }
}
pc++;
}
}
+ if (has_source_modifiers) {
+ 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 &&
+ inst->dst.type != scan_inst->src[i].type)
+ {
+ interfered = true;
+ break;
+ }
+ }
+ }
+
+
/* The gen6 MATH instruction can't handle source modifiers or
* unusual register regions, so avoid coalescing those for
* now. We should do something more specific.
/* 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)
+ if (this->virtual_grf_end[inst->src[0].reg] > ip)
continue;
/* Found a move of a GRF to a MRF. Let's see if we can go
if (scan_inst->mlen)
break;
- if (intel->gen == 6) {
+ if (brw->gen == 6) {
/* gen6 math instructions must have the destination be
* GRF, so no compute-to-MRF for them.
*/
int grf;
if (inst->src[i].file == GRF) {
grf = inst->src[i].reg;
- } else if (inst->src[i].file == FIXED_HW_REG &&
+ } else if (inst->src[i].file == HW_REG &&
inst->src[i].fixed_hw_reg.file == BRW_GENERAL_REGISTER_FILE) {
grf = inst->src[i].fixed_hw_reg.nr;
} else {
void
fs_visitor::insert_gen4_send_dependency_workarounds()
{
- if (intel->gen != 4 || intel->is_g4x)
+ if (brw->gen != 4 || brw->is_g4x)
return;
/* Note that we're done with register allocation, so GRF fs_regs always
if (inst->opcode != FS_OPCODE_UNIFORM_PULL_CONSTANT_LOAD)
continue;
- if (intel->gen >= 7) {
+ if (brw->gen >= 7) {
/* The offset arg before was a vec4-aligned byte offset. We need to
* turn it into a dword offset.
*/
}
void
-fs_visitor::dump_instruction(fs_inst *inst)
+fs_visitor::dump_instruction(backend_instruction *be_inst)
{
+ fs_inst *inst = (fs_inst *)be_inst;
+
if (inst->predicate) {
printf("(%cf0.%d) ",
inst->predicate_inverse ? '-' : '+',
if (inst->conditional_mod) {
printf(".cmod");
if (!inst->predicate &&
- (intel->gen < 5 || (inst->opcode != BRW_OPCODE_SEL &&
+ (brw->gen < 5 || (inst->opcode != BRW_OPCODE_SEL &&
inst->opcode != BRW_OPCODE_IF &&
inst->opcode != BRW_OPCODE_WHILE))) {
- printf(".f0.%d\n", inst->flag_subreg);
+ printf(".f0.%d", inst->flag_subreg);
}
}
printf(" ");
case UNIFORM:
printf("***u%d***", inst->dst.reg);
break;
+ case ARF:
+ if (inst->dst.reg == BRW_ARF_NULL)
+ printf("(null)");
+ else
+ printf("arf%d", inst->dst.reg);
+ break;
default:
printf("???");
break;
printf("\n");
}
-void
-fs_visitor::dump_instructions()
-{
- int ip = 0;
- foreach_list(node, &this->instructions) {
- fs_inst *inst = (fs_inst *)node;
- printf("%d: ", ip++);
- dump_instruction(inst);
- }
-}
-
/**
* Possibly returns an instruction that set up @param reg.
*
(fp->Base.InputsRead & (1 << VARYING_SLOT_POS)) != 0;
unsigned barycentric_interp_modes = c->prog_data.barycentric_interp_modes;
- assert(intel->gen >= 6);
+ assert(brw->gen >= 6);
/* R0-1: masks, pixel X/Y coordinates. */
c->nr_payload_regs = 2;
sanity_param_count = fp->Base.Parameters->NumParameters;
uint32_t orig_nr_params = c->prog_data.nr_params;
- if (intel->gen >= 6)
+ if (brw->gen >= 6)
setup_payload_gen6();
else
setup_payload_gen4();
emit_shader_time_begin();
calculate_urb_setup();
- if (intel->gen < 6)
+ if (brw->gen < 6)
emit_interpolation_setup_gen4();
else
emit_interpolation_setup_gen6();
struct gl_shader_program *prog,
unsigned *final_assembly_size)
{
- struct intel_context *intel = &brw->intel;
bool start_busy = false;
float start_time = 0;
- if (unlikely(intel->perf_debug)) {
- start_busy = (intel->batch.last_bo &&
- drm_intel_bo_busy(intel->batch.last_bo));
+ if (unlikely(brw->perf_debug)) {
+ start_busy = (brw->batch.last_bo &&
+ drm_intel_bo_busy(brw->batch.last_bo));
start_time = get_time();
}
exec_list *simd16_instructions = NULL;
fs_visitor v2(brw, c, prog, fp, 16);
- bool no16 = INTEL_DEBUG & DEBUG_NO16;
- if (intel->gen >= 5 && c->prog_data.nr_pull_params == 0 && likely(!no16)) {
- v2.import_uniforms(&v);
- if (!v2.run()) {
- perf_debug("16-wide shader failed to compile, falling back to "
- "8-wide at a 10-20%% performance cost: %s", v2.fail_msg);
+ if (brw->gen >= 5 && likely(!(INTEL_DEBUG & DEBUG_NO16))) {
+ if (c->prog_data.nr_pull_params == 0) {
+ /* Try a 16-wide compile */
+ v2.import_uniforms(&v);
+ if (!v2.run()) {
+ perf_debug("16-wide shader failed to compile, falling back to "
+ "8-wide at a 10-20%% performance cost: %s", v2.fail_msg);
+ } else {
+ simd16_instructions = &v2.instructions;
+ }
} else {
- simd16_instructions = &v2.instructions;
+ perf_debug("Skipping 16-wide due to pull parameters.\n");
}
}
simd16_instructions,
final_assembly_size);
- if (unlikely(intel->perf_debug) && shader) {
+ if (unlikely(brw->perf_debug) && shader) {
if (shader->compiled_once)
brw_wm_debug_recompile(brw, prog, &c->key);
shader->compiled_once = true;
- if (start_busy && !drm_intel_bo_busy(intel->batch.last_bo)) {
+ if (start_busy && !drm_intel_bo_busy(brw->batch.last_bo)) {
perf_debug("FS compile took %.03f ms and stalled the GPU\n",
(get_time() - start_time) * 1000);
}
brw_fs_precompile(struct gl_context *ctx, struct gl_shader_program *prog)
{
struct brw_context *brw = brw_context(ctx);
- struct intel_context *intel = &brw->intel;
struct brw_wm_prog_key key;
if (!prog->_LinkedShaders[MESA_SHADER_FRAGMENT])
memset(&key, 0, sizeof(key));
- if (intel->gen < 6) {
+ if (brw->gen < 6) {
if (fp->UsesKill)
key.iz_lookup |= IZ_PS_KILL_ALPHATEST_BIT;
key.iz_lookup |= IZ_DEPTH_WRITE_ENABLE_BIT;
}
- if (intel->gen < 6)
+ if (brw->gen < 6)
key.input_slots_valid |= BITFIELD64_BIT(VARYING_SLOT_POS);
for (int i = 0; i < VARYING_SLOT_MAX; i++) {
if (!(fp->Base.InputsRead & BITFIELD64_BIT(i)))
continue;
- if (intel->gen < 6) {
+ if (brw->gen < 6) {
if (_mesa_varying_slot_in_fs((gl_varying_slot) i))
key.input_slots_valid |= BITFIELD64_BIT(i);
}
key.clamp_fragment_color = ctx->API == API_OPENGL_COMPAT;
- for (int i = 0; i < MAX_SAMPLERS; i++) {
+ unsigned sampler_count = _mesa_fls(fp->Base.SamplersUsed);
+ for (unsigned i = 0; i < sampler_count; i++) {
if (fp->Base.ShadowSamplers & (1 << i)) {
/* Assume DEPTH_TEXTURE_MODE is the default: X, X, X, 1 */
key.tex.swizzles[i] =