#include "brw_fs.h"
#include "brw_cfg.h"
-#include "glsl/glsl_types.h"
+#include "glsl/nir/glsl_types.h"
#include "glsl/ir_optimization.h"
+using namespace brw;
+
static void
assign_reg(unsigned *reg_hw_locations, fs_reg *reg)
{
- if (reg->file == GRF) {
- reg->reg = reg_hw_locations[reg->reg] + reg->reg_offset;
+ if (reg->file == VGRF) {
+ reg->nr = reg_hw_locations[reg->nr] + reg->reg_offset;
reg->reg_offset = 0;
}
}
}
static void
-brw_alloc_reg_set(struct brw_compiler *compiler, int reg_width)
+brw_alloc_reg_set(struct brw_compiler *compiler, int dispatch_width)
{
const struct brw_device_info *devinfo = compiler->devinfo;
int base_reg_count = BRW_MAX_GRF;
- int index = reg_width - 1;
+ int index = (dispatch_width / 8) - 1;
+
+ if (dispatch_width > 8 && devinfo->gen >= 7) {
+ /* For IVB+, we don't need the PLN hacks or the even-reg alignment in
+ * SIMD16. Therefore, we can use the exact same register sets for
+ * SIMD16 as we do for SIMD8 and we don't need to recalculate them.
+ */
+ compiler->fs_reg_sets[index] = compiler->fs_reg_sets[0];
+ return;
+ }
/* The registers used to make up almost all values handled in the compiler
* are a scalar value occupying a single register (or 2 registers in the
/* Compute the total number of registers across all classes. */
int ra_reg_count = 0;
for (int i = 0; i < class_count; i++) {
- if (devinfo->gen <= 5 && reg_width == 2) {
+ if (devinfo->gen <= 5 && dispatch_width == 16) {
/* From the G45 PRM:
*
* In order to reduce the hardware complexity, the following
}
uint8_t *ra_reg_to_grf = ralloc_array(compiler, uint8_t, ra_reg_count);
- struct ra_regs *regs = ra_alloc_reg_set(compiler, ra_reg_count);
+ struct ra_regs *regs = ra_alloc_reg_set(compiler, ra_reg_count, false);
if (devinfo->gen >= 6)
ra_set_allocate_round_robin(regs);
int *classes = ralloc_array(compiler, int, class_count);
int pairs_reg_count = 0;
for (int i = 0; i < class_count; i++) {
int class_reg_count;
- if (devinfo->gen <= 5 && reg_width == 2) {
+ if (devinfo->gen <= 5 && dispatch_width == 16) {
class_reg_count = (base_reg_count - (class_sizes[i] - 1)) / 2;
/* See comment below. The only difference here is that we are
pairs_reg_count = class_reg_count;
}
- if (devinfo->gen <= 5 && reg_width == 2) {
+ if (devinfo->gen <= 5 && dispatch_width == 16) {
for (int j = 0; j < class_reg_count; j++) {
ra_class_add_reg(regs, classes[i], reg);
for (int base_reg = j;
base_reg < j + (class_sizes[i] + 1) / 2;
base_reg++) {
- ra_add_transitive_reg_conflict(regs, base_reg, reg);
+ ra_add_reg_conflict(regs, base_reg, reg);
}
reg++;
for (int base_reg = j;
base_reg < j + class_sizes[i];
base_reg++) {
- ra_add_transitive_reg_conflict(regs, base_reg, reg);
+ ra_add_reg_conflict(regs, base_reg, reg);
}
reg++;
}
assert(reg == ra_reg_count);
+ /* Applying transitivity to all of the base registers gives us the
+ * appropreate register conflict relationships everywhere.
+ */
+ for (int reg = 0; reg < base_reg_count; reg++)
+ ra_make_reg_conflicts_transitive(regs, reg);
+
/* Add a special class for aligned pairs, which we'll put delta_xy
* in on Gen <= 6 so that we can do PLN.
*/
- if (devinfo->has_pln && reg_width == 1 && devinfo->gen <= 6) {
+ if (devinfo->has_pln && dispatch_width == 8 && devinfo->gen <= 6) {
aligned_pairs_class = ra_alloc_reg_class(regs);
for (int i = 0; i < pairs_reg_count; i++) {
void
brw_fs_alloc_reg_sets(struct brw_compiler *compiler)
{
- brw_alloc_reg_set(compiler, 1);
- brw_alloc_reg_set(compiler, 2);
+ brw_alloc_reg_set(compiler, 8);
+ brw_alloc_reg_set(compiler, 16);
}
static int
unreachable("not reached");
}
-/**
- * Sets up interference between thread payload registers and the virtual GRFs
- * to be allocated for program temporaries.
- *
- * We want to be able to reallocate the payload for our virtual GRFs, notably
- * because the setup coefficients for a full set of 16 FS inputs takes up 8 of
- * our 128 registers.
- *
- * The layout of the payload registers is:
- *
- * 0..payload.num_regs-1: fixed function setup (including bary coordinates).
- * payload.num_regs..payload.num_regs+curb_read_lengh-1: uniform data
- * payload.num_regs+curb_read_lengh..first_non_payload_grf-1: setup coefficients.
- *
- * And we have payload_node_count nodes covering these registers in order
- * (note that in SIMD16, a node is two registers).
- */
-void
-fs_visitor::setup_payload_interference(struct ra_graph *g,
- int payload_node_count,
- int first_payload_node)
+void fs_visitor::calculate_payload_ranges(int payload_node_count,
+ int *payload_last_use_ip)
{
int loop_depth = 0;
int loop_end_ip = 0;
- int payload_last_use_ip[payload_node_count];
- memset(payload_last_use_ip, 0, sizeof(payload_last_use_ip));
+ for (int i = 0; i < payload_node_count; i++)
+ payload_last_use_ip[i] = -1;
+
int ip = 0;
foreach_block_and_inst(block, fs_inst, inst, cfg) {
switch (inst->opcode) {
else
use_ip = ip;
- /* Note that UNIFORM args have been turned into FIXED_HW_REG by
+ /* Note that UNIFORM args have been turned into FIXED_GRF by
* assign_curbe_setup(), and interpolation uses fixed hardware regs from
* the start (see interp_reg()).
*/
for (int i = 0; i < inst->sources; i++) {
- if (inst->src[i].file == HW_REG &&
- inst->src[i].fixed_hw_reg.file == BRW_GENERAL_REGISTER_FILE) {
- int node_nr = inst->src[i].fixed_hw_reg.nr;
+ if (inst->src[i].file == FIXED_GRF) {
+ int node_nr = inst->src[i].nr;
if (node_nr >= payload_node_count)
continue;
- payload_last_use_ip[node_nr] = use_ip;
+ for (int j = 0; j < inst->regs_read(i); j++) {
+ payload_last_use_ip[node_nr + j] = use_ip;
+ assert(node_nr + j < payload_node_count);
+ }
}
}
/* Special case instructions which have extra implied registers used. */
switch (inst->opcode) {
- case FS_OPCODE_LINTERP:
- /* On gen6+ in SIMD16, there are 4 adjacent registers used by
- * PLN's sourcing of the deltas, while we list only the first one
- * in the arguments. Pre-gen6, the deltas are computed in normal
- * VGRFs.
- */
- if (devinfo->gen >= 6) {
- int delta_x_arg = 0;
- if (inst->src[delta_x_arg].file == HW_REG &&
- inst->src[delta_x_arg].fixed_hw_reg.file ==
- BRW_GENERAL_REGISTER_FILE) {
- for (int i = 1; i < 4; ++i) {
- int node = inst->src[delta_x_arg].fixed_hw_reg.nr + i;
- assert(node < payload_node_count);
- payload_last_use_ip[node] = use_ip;
- }
- }
- }
+ case CS_OPCODE_CS_TERMINATE:
+ payload_last_use_ip[0] = use_ip;
break;
default:
ip++;
}
+}
+
+
+/**
+ * Sets up interference between thread payload registers and the virtual GRFs
+ * to be allocated for program temporaries.
+ *
+ * We want to be able to reallocate the payload for our virtual GRFs, notably
+ * because the setup coefficients for a full set of 16 FS inputs takes up 8 of
+ * our 128 registers.
+ *
+ * The layout of the payload registers is:
+ *
+ * 0..payload.num_regs-1: fixed function setup (including bary coordinates).
+ * payload.num_regs..payload.num_regs+curb_read_lengh-1: uniform data
+ * payload.num_regs+curb_read_lengh..first_non_payload_grf-1: setup coefficients.
+ *
+ * And we have payload_node_count nodes covering these registers in order
+ * (note that in SIMD16, a node is two registers).
+ */
+void
+fs_visitor::setup_payload_interference(struct ra_graph *g,
+ int payload_node_count,
+ int first_payload_node)
+{
+ int payload_last_use_ip[payload_node_count];
+ calculate_payload_ranges(payload_node_count, payload_last_use_ip);
for (int i = 0; i < payload_node_count; i++) {
+ if (payload_last_use_ip[i] == -1)
+ continue;
+
/* Mark the payload node as interfering with any virtual grf that is
* live between the start of the program and our last use of the payload
* node.
* see if we can actually use MRFs to do spills without overwriting normal MRF
* contents.
*/
-void
-fs_visitor::get_used_mrfs(bool *mrf_used)
+static void
+get_used_mrfs(fs_visitor *v, bool *mrf_used)
{
- int reg_width = dispatch_width / 8;
+ int reg_width = v->dispatch_width / 8;
- memset(mrf_used, 0, BRW_MAX_MRF * sizeof(bool));
+ memset(mrf_used, 0, BRW_MAX_MRF(v->devinfo->gen) * sizeof(bool));
- foreach_block_and_inst(block, fs_inst, inst, cfg) {
+ foreach_block_and_inst(block, fs_inst, inst, v->cfg) {
if (inst->dst.file == MRF) {
- int reg = inst->dst.reg & ~BRW_MRF_COMPR4;
+ int reg = inst->dst.nr & ~BRW_MRF_COMPR4;
mrf_used[reg] = true;
if (reg_width == 2) {
- if (inst->dst.reg & BRW_MRF_COMPR4) {
+ if (inst->dst.nr & BRW_MRF_COMPR4) {
mrf_used[reg + 4] = true;
} else {
mrf_used[reg + 1] = true;
}
if (inst->mlen > 0) {
- for (int i = 0; i < implied_mrf_writes(inst); i++) {
+ for (int i = 0; i < v->implied_mrf_writes(inst); i++) {
mrf_used[inst->base_mrf + i] = true;
}
}
* Sets interference between virtual GRFs and usage of the high GRFs for SEND
* messages (treated as MRFs in code generation).
*/
-void
-fs_visitor::setup_mrf_hack_interference(struct ra_graph *g, int first_mrf_node)
+static void
+setup_mrf_hack_interference(fs_visitor *v, struct ra_graph *g,
+ int first_mrf_node, int *first_used_mrf)
{
- bool mrf_used[BRW_MAX_MRF];
- get_used_mrfs(mrf_used);
+ bool mrf_used[BRW_MAX_MRF(v->devinfo->gen)];
+ get_used_mrfs(v, mrf_used);
- for (int i = 0; i < BRW_MAX_MRF; i++) {
+ *first_used_mrf = BRW_MAX_MRF(v->devinfo->gen);
+ for (int i = 0; i < BRW_MAX_MRF(v->devinfo->gen); i++) {
/* Mark each MRF reg node as being allocated to its physical register.
*
* The alternative would be to have per-physical-register classes, which
* that are used as conflicting with all virtual GRFs.
*/
if (mrf_used[i]) {
- for (unsigned j = 0; j < this->alloc.count; j++) {
+ if (i < *first_used_mrf)
+ *first_used_mrf = i;
+
+ for (unsigned j = 0; j < v->alloc.count; j++) {
ra_add_node_interference(g, first_mrf_node + i, j);
}
}
bool
fs_visitor::assign_regs(bool allow_spilling)
{
- struct brw_compiler *compiler = brw->intelScreen->compiler;
/* Most of this allocation was written for a reg_width of 1
* (dispatch_width == 8). In extending to SIMD16, the code was
* left in place and it was converted to have the hardware
* that register and set it to the appropriate class.
*/
if (compiler->fs_reg_sets[rsi].aligned_pairs_class >= 0 &&
- this->delta_xy[BRW_WM_PERSPECTIVE_PIXEL_BARYCENTRIC].file == GRF &&
- this->delta_xy[BRW_WM_PERSPECTIVE_PIXEL_BARYCENTRIC].reg == i) {
+ this->delta_xy[BRW_WM_PERSPECTIVE_PIXEL_BARYCENTRIC].file == VGRF &&
+ this->delta_xy[BRW_WM_PERSPECTIVE_PIXEL_BARYCENTRIC].nr == i) {
c = compiler->fs_reg_sets[rsi].aligned_pairs_class;
}
setup_payload_interference(g, payload_node_count, first_payload_node);
if (devinfo->gen >= 7) {
- setup_mrf_hack_interference(g, first_mrf_hack_node);
+ int first_used_mrf = BRW_MAX_MRF(devinfo->gen);
+ setup_mrf_hack_interference(this, g, first_mrf_hack_node,
+ &first_used_mrf);
foreach_block_and_inst(block, fs_inst, inst, cfg) {
/* When we do send-from-GRF for FB writes, we need to ensure that
* highest register that works.
*/
if (inst->eot) {
- int size = alloc.sizes[inst->src[0].reg];
+ int size = alloc.sizes[inst->src[0].nr];
int reg = compiler->fs_reg_sets[rsi].class_to_ra_reg_range[size] - 1;
- ra_set_node_reg(g, inst->src[0].reg, reg);
+
+ /* If something happened to spill, we want to push the EOT send
+ * register early enough in the register file that we don't
+ * conflict with any used MRF hack registers.
+ */
+ reg -= BRW_MAX_MRF(devinfo->gen) - first_used_mrf;
+
+ ra_set_node_reg(g, inst->src[0].nr, reg);
break;
}
}
* destination interfere.
*/
foreach_block_and_inst(block, fs_inst, inst, cfg) {
- if (inst->dst.file != GRF)
+ if (inst->dst.file != VGRF)
continue;
for (int i = 0; i < inst->sources; ++i) {
- if (inst->src[i].file == GRF) {
- ra_add_node_interference(g, inst->dst.reg, inst->src[i].reg);
+ if (inst->src[i].file == VGRF) {
+ ra_add_node_interference(g, inst->dst.nr, inst->src[i].nr);
}
}
}
}
/* Debug of register spilling: Go spill everything. */
- if (unlikely(INTEL_DEBUG & DEBUG_SPILL)) {
+ if (unlikely(INTEL_DEBUG & DEBUG_SPILL_FS)) {
int reg = choose_spill_reg(g);
if (reg != -1) {
uint32_t spill_offset, int count)
{
int reg_size = 1;
- if (dispatch_width == 16 && count % 2 == 0) {
+ if (dispatch_width == 16 && count % 2 == 0)
reg_size = 2;
- dst.width = 16;
- }
+
+ const fs_builder ibld = bld.annotate(inst->annotation, inst->ir)
+ .group(reg_size * 8, 0)
+ .at(block, inst);
for (int i = 0; i < count / reg_size; i++) {
/* The gen7 descriptor-based offset is 12 bits of HWORD units. */
bool gen7_read = devinfo->gen >= 7 && spill_offset < (1 << 12) * REG_SIZE;
-
- fs_inst *unspill_inst =
- new(mem_ctx) fs_inst(gen7_read ?
- SHADER_OPCODE_GEN7_SCRATCH_READ :
- SHADER_OPCODE_GEN4_SCRATCH_READ,
- dst);
+ fs_inst *unspill_inst = ibld.emit(gen7_read ?
+ SHADER_OPCODE_GEN7_SCRATCH_READ :
+ SHADER_OPCODE_GEN4_SCRATCH_READ,
+ dst);
unspill_inst->offset = spill_offset;
- unspill_inst->ir = inst->ir;
- unspill_inst->annotation = inst->annotation;
unspill_inst->regs_written = reg_size;
if (!gen7_read) {
- unspill_inst->base_mrf = 14;
+ unspill_inst->base_mrf = FIRST_SPILL_MRF(devinfo->gen) + 1;
unspill_inst->mlen = 1; /* header contains offset */
}
- inst->insert_before(block, unspill_inst);
dst.reg_offset += reg_size;
spill_offset += reg_size * REG_SIZE;
uint32_t spill_offset, int count)
{
int reg_size = 1;
- int spill_base_mrf = 14;
+ int spill_base_mrf = FIRST_SPILL_MRF(devinfo->gen) + 1;
if (dispatch_width == 16 && count % 2 == 0) {
- spill_base_mrf = 13;
+ spill_base_mrf = FIRST_SPILL_MRF(devinfo->gen);
reg_size = 2;
}
+ const fs_builder ibld = bld.annotate(inst->annotation, inst->ir)
+ .group(reg_size * 8, 0)
+ .at(block, inst->next);
+
for (int i = 0; i < count / reg_size; i++) {
fs_inst *spill_inst =
- new(mem_ctx) fs_inst(SHADER_OPCODE_GEN4_SCRATCH_WRITE,
- reg_size * 8, reg_null_f, src);
+ ibld.emit(SHADER_OPCODE_GEN4_SCRATCH_WRITE, ibld.null_reg_f(), src);
src.reg_offset += reg_size;
spill_inst->offset = spill_offset + i * reg_size * REG_SIZE;
- spill_inst->ir = inst->ir;
- spill_inst->annotation = inst->annotation;
spill_inst->mlen = 1 + reg_size; /* header, value */
spill_inst->base_mrf = spill_base_mrf;
- inst->insert_after(block, spill_inst);
}
}
*/
foreach_block_and_inst(block, fs_inst, inst, cfg) {
for (unsigned int i = 0; i < inst->sources; i++) {
- if (inst->src[i].file == GRF) {
- spill_costs[inst->src[i].reg] += loop_scale;
+ if (inst->src[i].file == VGRF) {
+ spill_costs[inst->src[i].nr] += loop_scale;
/* Register spilling logic assumes full-width registers; smeared
* registers have a width of 1 so if we try to spill them we'll
* register pressure anyhow.
*/
if (!inst->src[i].is_contiguous()) {
- no_spill[inst->src[i].reg] = true;
+ no_spill[inst->src[i].nr] = true;
}
}
}
- if (inst->dst.file == GRF) {
- spill_costs[inst->dst.reg] += inst->regs_written * loop_scale;
+ if (inst->dst.file == VGRF) {
+ spill_costs[inst->dst.nr] += inst->regs_written * loop_scale;
if (!inst->dst.is_contiguous()) {
- no_spill[inst->dst.reg] = true;
+ no_spill[inst->dst.nr] = true;
}
}
break;
case SHADER_OPCODE_GEN4_SCRATCH_WRITE:
- if (inst->src[0].file == GRF)
- no_spill[inst->src[0].reg] = true;
+ if (inst->src[0].file == VGRF)
+ no_spill[inst->src[0].nr] = true;
break;
case SHADER_OPCODE_GEN4_SCRATCH_READ:
case SHADER_OPCODE_GEN7_SCRATCH_READ:
- if (inst->dst.file == GRF)
- no_spill[inst->dst.reg] = true;
+ if (inst->dst.file == VGRF)
+ no_spill[inst->dst.nr] = true;
break;
default:
int size = alloc.sizes[spill_reg];
unsigned int spill_offset = last_scratch;
assert(ALIGN(spill_offset, 16) == spill_offset); /* oword read/write req. */
- int spill_base_mrf = dispatch_width > 8 ? 13 : 14;
+ int spill_base_mrf = dispatch_width > 8 ? FIRST_SPILL_MRF(devinfo->gen) :
+ FIRST_SPILL_MRF(devinfo->gen) + 1;
/* Spills may use MRFs 13-15 in the SIMD16 case. Our texturing is done
* using up to 11 MRFs starting from either m1 or m2, and fb writes can use
* SIMD16 mode, because we'd stomp the FB writes.
*/
if (!spilled_any_registers) {
- bool mrf_used[BRW_MAX_MRF];
- get_used_mrfs(mrf_used);
+ bool mrf_used[BRW_MAX_MRF(devinfo->gen)];
+ get_used_mrfs(this, mrf_used);
- for (int i = spill_base_mrf; i < BRW_MAX_MRF; i++) {
+ for (int i = spill_base_mrf; i < BRW_MAX_MRF(devinfo->gen); i++) {
if (mrf_used[i]) {
fail("Register spilling not supported with m%d used", i);
return;
*/
foreach_block_and_inst (block, fs_inst, inst, cfg) {
for (unsigned int i = 0; i < inst->sources; i++) {
- if (inst->src[i].file == GRF &&
- inst->src[i].reg == spill_reg) {
+ if (inst->src[i].file == VGRF &&
+ inst->src[i].nr == spill_reg) {
int regs_read = inst->regs_read(i);
int subset_spill_offset = (spill_offset +
REG_SIZE * inst->src[i].reg_offset);
- fs_reg unspill_dst(GRF, alloc.allocate(regs_read));
+ fs_reg unspill_dst(VGRF, alloc.allocate(regs_read));
- inst->src[i].reg = unspill_dst.reg;
+ inst->src[i].nr = unspill_dst.nr;
inst->src[i].reg_offset = 0;
emit_unspill(block, inst, unspill_dst, subset_spill_offset,
}
}
- if (inst->dst.file == GRF &&
- inst->dst.reg == spill_reg) {
+ if (inst->dst.file == VGRF &&
+ inst->dst.nr == spill_reg) {
int subset_spill_offset = (spill_offset +
REG_SIZE * inst->dst.reg_offset);
- fs_reg spill_src(GRF, alloc.allocate(inst->regs_written));
+ fs_reg spill_src(VGRF, alloc.allocate(inst->regs_written));
- inst->dst.reg = spill_src.reg;
+ inst->dst.nr = spill_src.nr;
inst->dst.reg_offset = 0;
/* If we're immediately spilling the register, we should not use