/* 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
- * case of 16-wide, which is handled by dividing base_reg_count by 2 and
+ * case of SIMD16, which is handled by dividing base_reg_count by 2 and
* multiplying allocated register numbers by 2). Things that were
* aggregates of scalar values at the GLSL level were split to scalar
* values by split_virtual_grfs().
int class_sizes[BRW_MAX_MRF];
if (brw->gen >= 7) {
- for (class_count = 0; class_count < 11; class_count++)
+ for (class_count = 0; class_count < MAX_SAMPLER_MESSAGE_SIZE;
+ class_count++)
class_sizes[class_count] = class_count + 1;
} else {
for (class_count = 0; class_count < 4; class_count++)
* nr_payload_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 16-wide, a node is two registers).
+ * (note that in SIMD16, a node is two registers).
*/
void
fs_visitor::setup_payload_interference(struct ra_graph *g,
break;
case FS_OPCODE_LINTERP:
- /* On gen6+ in 16-wide, there are 4 adjacent registers (so 2 nodes)
+ /* On gen6+ in SIMD16, there are 4 adjacent registers (so 2 nodes)
* used by PLN's sourcing of the deltas, while we list only the first
* two in the arguments (1 node). Pre-gen6, the deltas are computed
* in normal VGRFs.
}
/**
- * Sets interference between virtual GRFs and usage of the high GRFs for SEND
- * messages (treated as MRFs in code generation).
+ * Sets the mrf_used array to indicate which MRFs are used by the shader IR
+ *
+ * This is used in assign_regs() to decide which of the GRFs that we use as
+ * MRFs on gen7 get normally register allocated, and in register spilling to
+ * see if we can actually use MRFs to do spills without overwriting normal MRF
+ * contents.
*/
void
-fs_visitor::setup_mrf_hack_interference(struct ra_graph *g, int first_mrf_node)
+fs_visitor::get_used_mrfs(bool *mrf_used)
{
- int mrf_count = BRW_MAX_GRF - GEN7_MRF_HACK_START;
int reg_width = dispatch_width / 8;
- /* Identify all the MRFs used in the program. */
- bool mrf_used[mrf_count];
- memset(mrf_used, 0, sizeof(mrf_used));
+ memset(mrf_used, 0, BRW_MAX_MRF * sizeof(bool));
+
foreach_list(node, &this->instructions) {
fs_inst *inst = (fs_inst *)node;
}
}
}
+}
+
+/**
+ * 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)
+{
+ int reg_width = dispatch_width / 8;
- for (int i = 0; i < mrf_count; i++) {
- /* Mark each payload reg node as being allocated to its physical register.
+ bool mrf_used[BRW_MAX_MRF];
+ get_used_mrfs(mrf_used);
+
+ for (int i = 0; i < BRW_MAX_MRF; i++) {
+ /* Mark each MRF reg node as being allocated to its physical register.
*
* The alternative would be to have per-physical-register classes, which
* would just be silly.
}
bool
-fs_visitor::assign_regs()
+fs_visitor::assign_regs(bool allow_spilling)
{
/* Most of this allocation was written for a reg_width of 1
- * (dispatch_width == 8). In extending to 16-wide, the code was
+ * (dispatch_width == 8). In extending to SIMD16, the code was
* left in place and it was converted to have the hardware
* registers it's allocating be contiguous physical pairs of regs
* for reg_width == 2.
if (brw->gen >= 7)
setup_mrf_hack_interference(g, first_mrf_hack_node);
+ /* Debug of register spilling: Go spill everything. */
+ if (0) {
+ int reg = choose_spill_reg(g);
+
+ if (reg != -1) {
+ spill_reg(reg);
+ ralloc_free(g);
+ return false;
+ }
+ }
+
if (!ra_allocate_no_spills(g)) {
/* Failed to allocate registers. Spill a reg, and the caller will
* loop back into here to try again.
if (reg == -1) {
fail("no register to spill:\n");
dump_instructions();
- } else if (dispatch_width == 16) {
- fail("Failure to register allocate. Reduce number of live scalar "
- "values to avoid this.");
- } else {
- spill_reg(reg);
+ } else if (allow_spilling) {
+ spill_reg(reg);
}
-
ralloc_free(g);
return false;
int count)
{
for (int i = 0; i < count; i++) {
- fs_inst *unspill_inst = new(mem_ctx) fs_inst(FS_OPCODE_UNSPILL, dst);
+ /* The gen7 descriptor-based offset is 12 bits of HWORD units. */
+ bool gen7_read = brw->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);
unspill_inst->offset = spill_offset;
unspill_inst->ir = inst->ir;
unspill_inst->annotation = inst->annotation;
- /* Choose a MRF that won't conflict with an MRF that's live across the
- * spill. Nothing else will make it up to MRF 14/15.
- */
- unspill_inst->base_mrf = 14;
- unspill_inst->mlen = 1; /* header contains offset */
+ if (!gen7_read) {
+ unspill_inst->base_mrf = 14;
+ unspill_inst->mlen = 1; /* header contains offset */
+ }
inst->insert_before(unspill_inst);
dst.reg_offset++;
- spill_offset += REG_SIZE;
+ spill_offset += dispatch_width * sizeof(float);
}
}
loop_scale /= 10;
break;
- case FS_OPCODE_SPILL:
+ case SHADER_OPCODE_GEN4_SCRATCH_WRITE:
if (inst->src[0].file == GRF)
no_spill[inst->src[0].reg] = true;
break;
- case FS_OPCODE_UNSPILL:
+ case SHADER_OPCODE_GEN4_SCRATCH_READ:
+ case SHADER_OPCODE_GEN7_SCRATCH_READ:
if (inst->dst.file == GRF)
no_spill[inst->dst.reg] = true;
break;
void
fs_visitor::spill_reg(int spill_reg)
{
+ int reg_size = dispatch_width * sizeof(float);
int size = virtual_grf_sizes[spill_reg];
unsigned int spill_offset = c->last_scratch;
assert(ALIGN(spill_offset, 16) == spill_offset); /* oword read/write req. */
- c->last_scratch += size * REG_SIZE;
+ int spill_base_mrf = dispatch_width > 8 ? 13 : 14;
+
+ /* 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
+ * up to m13 (gen6+ simd16: 2 header + 8 color + 2 src0alpha + 2 omask) or
+ * m15 (gen4-5 simd16: 2 header + 8 color + 1 aads + 2 src depth + 2 dst
+ * depth), starting from m1. In summary: We may not be able to spill in
+ * SIMD16 mode, because we'd stomp the FB writes.
+ */
+ if (!spilled_any_registers) {
+ bool mrf_used[BRW_MAX_MRF];
+ get_used_mrfs(mrf_used);
+
+ for (int i = spill_base_mrf; i < BRW_MAX_MRF; i++) {
+ if (mrf_used[i]) {
+ fail("Register spilling not supported with m%d used", i);
+ return;
+ }
+ }
+
+ spilled_any_registers = true;
+ }
+
+ c->last_scratch += size * reg_size;
/* Generate spill/unspill instructions for the objects being
* spilled. Right now, we spill or unspill the whole thing to a
if (inst->src[i].file == GRF &&
inst->src[i].reg == spill_reg) {
int regs_read = inst->regs_read(this, i);
+ int subset_spill_offset = (spill_offset +
+ reg_size * inst->src[i].reg_offset);
inst->src[i].reg = virtual_grf_alloc(regs_read);
inst->src[i].reg_offset = 0;
- emit_unspill(inst, inst->src[i],
- spill_offset + REG_SIZE * inst->src[i].reg_offset,
- regs_read);
+ emit_unspill(inst, inst->src[i], subset_spill_offset, regs_read);
}
}
if (inst->dst.file == GRF &&
inst->dst.reg == spill_reg) {
int subset_spill_offset = (spill_offset +
- REG_SIZE * inst->dst.reg_offset);
+ reg_size * inst->dst.reg_offset);
inst->dst.reg = virtual_grf_alloc(inst->regs_written);
inst->dst.reg_offset = 0;
spill_src.smear = -1;
for (int chan = 0; chan < inst->regs_written; chan++) {
- fs_inst *spill_inst = new(mem_ctx) fs_inst(FS_OPCODE_SPILL,
- reg_null_f, spill_src);
+ fs_inst *spill_inst =
+ new(mem_ctx) fs_inst(SHADER_OPCODE_GEN4_SCRATCH_WRITE,
+ reg_null_f, spill_src);
spill_src.reg_offset++;
- spill_inst->offset = subset_spill_offset + chan * REG_SIZE;
+ spill_inst->offset = subset_spill_offset + chan * reg_size;
spill_inst->ir = inst->ir;
spill_inst->annotation = inst->annotation;
- spill_inst->base_mrf = 14;
- spill_inst->mlen = 2; /* header, value */
+ spill_inst->mlen = 1 + dispatch_width / 8; /* header, value */
+ spill_inst->base_mrf = spill_base_mrf;
inst->insert_after(spill_inst);
}
}