[Mesa-dev] [PATCH 06/10] i965/vs: Add support for compute-to-MRF.
Eric Anholt
eric at anholt.net
Thu Sep 8 23:32:23 PDT 2011
Removes 1.8% of the instructions from 97% of the vertex shaders in
shader-db.
---
src/mesa/drivers/dri/i965/brw_vec4.cpp | 177 +++++++++++++++++++++++++++
src/mesa/drivers/dri/i965/brw_vec4.h | 1 +
src/mesa/drivers/dri/i965/brw_vec4_emit.cpp | 1 +
3 files changed, 179 insertions(+), 0 deletions(-)
diff --git a/src/mesa/drivers/dri/i965/brw_vec4.cpp b/src/mesa/drivers/dri/i965/brw_vec4.cpp
index 2a8d63a..4a191af 100644
--- a/src/mesa/drivers/dri/i965/brw_vec4.cpp
+++ b/src/mesa/drivers/dri/i965/brw_vec4.cpp
@@ -519,4 +519,181 @@ vec4_visitor::move_push_constants_to_pull_constants()
pack_uniform_registers();
}
+/*
+ * Tries to reduce extra MOV instructions by taking GRFs that get just
+ * written and then MOVed into an MRF and making the original write of
+ * the GRF write directly to the MRF instead.
+ */
+bool
+vec4_visitor::opt_compute_to_mrf()
+{
+ bool progress = false;
+ int next_ip = 0;
+
+ calculate_live_intervals();
+
+ foreach_list_safe(node, &this->instructions) {
+ vec4_instruction *inst = (vec4_instruction *)node;
+
+ int ip = next_ip;
+ next_ip++;
+
+ if (inst->opcode != BRW_OPCODE_MOV ||
+ inst->predicate ||
+ 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].reladdr)
+ continue;
+
+ int mrf = inst->dst.reg;
+
+ /* 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;
+
+ /* We need to check interference with the MRF between this
+ * instruction and the earliest instruction involved in writing
+ * the GRF we're eliminating. To do that, keep track of which
+ * of our source channels we've seen initialized.
+ */
+ bool chans_needed[4] = {false, false, false, false};
+ int chans_remaining = 0;
+ for (int i = 0; i < 4; i++) {
+ int chan = BRW_GET_SWZ(inst->src[0].swizzle, i);
+
+ if (!(inst->dst.writemask & (1 << i)))
+ continue;
+
+ /* We don't handle compute-to-MRF across a swizzle. We would
+ * need to be able to rewrite instructions above to output
+ * results to different channels.
+ */
+ if (chan != i)
+ chans_remaining = 5;
+
+ if (!chans_needed[chan]) {
+ chans_needed[chan] = true;
+ chans_remaining++;
+ }
+ }
+ if (chans_remaining > 4)
+ continue;
+
+ /* Now walk up the instruction stream trying to see if we can
+ * rewrite everything writing to the GRF into the MRF instead.
+ */
+ vec4_instruction *scan_inst;
+ for (scan_inst = (vec4_instruction *)inst->prev;
+ scan_inst->prev != NULL;
+ scan_inst = (vec4_instruction *)scan_inst->prev) {
+ if (scan_inst->dst.file == GRF &&
+ scan_inst->dst.reg == inst->src[0].reg &&
+ scan_inst->dst.reg_offset == inst->src[0].reg_offset) {
+ /* Found something writing to the reg we want to turn into
+ * a compute-to-MRF.
+ */
+
+ /* 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->is_math()) {
+ break;
+ }
+ }
+
+ /* Mark which channels we found unconditional writes for. */
+ if (!scan_inst->predicate) {
+ for (int i = 0; i < 4; i++) {
+ if (scan_inst->dst.writemask & (1 << i) &&
+ chans_needed[i]) {
+ chans_needed[i] = false;
+ chans_remaining--;
+ }
+ }
+ }
+
+ if (chans_remaining == 0)
+ break;
+ }
+
+ /* 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_ELSE ||
+ 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 somebody else writes our MRF here, we can't
+ * compute-to-MRF before that.
+ */
+ if (scan_inst->dst.file == MRF && mrf == scan_inst->dst.reg)
+ break;
+
+ if (scan_inst->mlen > 0) {
+ /* Found a SEND instruction, which means that there are
+ * live values in MRFs from base_mrf to base_mrf +
+ * scan_inst->mlen - 1. Don't go pushing our MRF write up
+ * above it.
+ */
+ if (mrf >= scan_inst->base_mrf &&
+ mrf < scan_inst->base_mrf + scan_inst->mlen) {
+ break;
+ }
+ }
+ }
+
+ if (chans_remaining == 0) {
+ /* If we've made it here, we have an inst we want to
+ * compute-to-MRF, and a scan_inst pointing to the earliest
+ * instruction involved in computing the value. Now go
+ * rewrite the instruction stream between the two.
+ */
+
+ while (scan_inst != inst) {
+ if (scan_inst->dst.file == GRF &&
+ scan_inst->dst.reg == inst->src[0].reg &&
+ scan_inst->dst.reg_offset == inst->src[0].reg_offset) {
+ scan_inst->dst.file = MRF;
+ scan_inst->dst.reg = mrf;
+ scan_inst->dst.reg_offset = 0;
+ scan_inst->saturate |= inst->saturate;
+ }
+ scan_inst = (vec4_instruction *)scan_inst->next;
+ }
+ inst->remove();
+ progress = true;
+ }
+ }
+
+ if (progress)
+ live_intervals_valid = false;
+
+ return progress;
+}
+
} /* namespace brw */
diff --git a/src/mesa/drivers/dri/i965/brw_vec4.h b/src/mesa/drivers/dri/i965/brw_vec4.h
index e305e6f..90c1cbc 100644
--- a/src/mesa/drivers/dri/i965/brw_vec4.h
+++ b/src/mesa/drivers/dri/i965/brw_vec4.h
@@ -400,6 +400,7 @@ public:
bool virtual_grf_interferes(int a, int b);
bool opt_copy_propagation();
bool opt_algebraic();
+ bool opt_compute_to_mrf();
vec4_instruction *emit(vec4_instruction *inst);
diff --git a/src/mesa/drivers/dri/i965/brw_vec4_emit.cpp b/src/mesa/drivers/dri/i965/brw_vec4_emit.cpp
index 15f2458..ecaf8b4 100644
--- a/src/mesa/drivers/dri/i965/brw_vec4_emit.cpp
+++ b/src/mesa/drivers/dri/i965/brw_vec4_emit.cpp
@@ -622,6 +622,7 @@ vec4_visitor::run()
progress = dead_code_eliminate() || progress;
progress = opt_copy_propagation() || progress;
progress = opt_algebraic() || progress;
+ progress = opt_compute_to_mrf() || progress;
} while (progress);
--
1.7.5.4
More information about the mesa-dev
mailing list