Mesa (master): vc4: Move register allocation to a separate file.
Eric Anholt
anholt at kemper.freedesktop.org
Tue Sep 16 22:16:57 UTC 2014
Module: Mesa
Branch: master
Commit: ad02ba42f0931abfdb79af29678267bd063e92f3
URL: http://cgit.freedesktop.org/mesa/mesa/commit/?id=ad02ba42f0931abfdb79af29678267bd063e92f3
Author: Eric Anholt <eric at anholt.net>
Date: Mon Sep 15 11:45:56 2014 -0700
vc4: Move register allocation to a separate file.
I'm going to be rewriting it all, and having it mixed up with the
QIR-to-QPU opcode translation was messy.
---
src/gallium/drivers/vc4/Makefile.sources | 1 +
src/gallium/drivers/vc4/vc4_context.h | 1 +
src/gallium/drivers/vc4/vc4_qpu_emit.c | 106 +--------------
src/gallium/drivers/vc4/vc4_register_allocate.c | 157 +++++++++++++++++++++++
4 files changed, 165 insertions(+), 100 deletions(-)
diff --git a/src/gallium/drivers/vc4/Makefile.sources b/src/gallium/drivers/vc4/Makefile.sources
index f64848f..bcb4209 100644
--- a/src/gallium/drivers/vc4/Makefile.sources
+++ b/src/gallium/drivers/vc4/Makefile.sources
@@ -23,6 +23,7 @@ C_SOURCES := \
vc4_qpu_emit.c \
vc4_qpu.h \
vc4_qpu_validate.c \
+ vc4_register_allocate.c \
vc4_resource.c \
vc4_resource.h \
vc4_screen.c \
diff --git a/src/gallium/drivers/vc4/vc4_context.h b/src/gallium/drivers/vc4/vc4_context.h
index b82c08c..e586433 100644
--- a/src/gallium/drivers/vc4/vc4_context.h
+++ b/src/gallium/drivers/vc4/vc4_context.h
@@ -230,6 +230,7 @@ void vc4_flush(struct pipe_context *pctx);
void vc4_flush_for_bo(struct pipe_context *pctx, struct vc4_bo *bo);
void vc4_emit_state(struct pipe_context *pctx);
void vc4_generate_code(struct vc4_compile *c);
+struct qpu_reg *vc4_register_allocate(struct vc4_compile *c);
void vc4_update_compiled_shaders(struct vc4_context *vc4, uint8_t prim_mode);
bool vc4_rt_format_supported(enum pipe_format f);
diff --git a/src/gallium/drivers/vc4/vc4_qpu_emit.c b/src/gallium/drivers/vc4/vc4_qpu_emit.c
index 26520fe..6c4c76d 100644
--- a/src/gallium/drivers/vc4/vc4_qpu_emit.c
+++ b/src/gallium/drivers/vc4/vc4_qpu_emit.c
@@ -210,43 +210,11 @@ serialize_insts(struct vc4_compile *c)
void
vc4_generate_code(struct vc4_compile *c)
{
- struct qpu_reg allocate_to_qpu_reg[4 + 32 + 32];
- bool reg_in_use[ARRAY_SIZE(allocate_to_qpu_reg)];
- int *reg_allocated = calloc(c->num_temps, sizeof(*reg_allocated));
- int *reg_uses_remaining =
- calloc(c->num_temps, sizeof(*reg_uses_remaining));
+ struct qpu_reg *temp_registers = vc4_register_allocate(c);
bool discard = false;
- for (int i = 0; i < ARRAY_SIZE(reg_in_use); i++)
- reg_in_use[i] = false;
- for (int i = 0; i < c->num_temps; i++)
- reg_allocated[i] = -1;
-
- uint32_t next_reg = 0;
- for (int i = 0; i < 4; i++)
- allocate_to_qpu_reg[next_reg++] = qpu_rn(i == 3 ? 4 : i);
- for (int i = 0; i < 32; i++)
- allocate_to_qpu_reg[next_reg++] = qpu_ra(i);
- for (int i = 0; i < 32; i++)
- allocate_to_qpu_reg[next_reg++] = qpu_rb(i);
- assert(next_reg == ARRAY_SIZE(allocate_to_qpu_reg));
-
make_empty_list(&c->qpu_inst_list);
- struct simple_node *node;
- foreach(node, &c->instructions) {
- struct qinst *qinst = (struct qinst *)node;
-
- if (qinst->dst.file == QFILE_TEMP)
- reg_uses_remaining[qinst->dst.index]++;
- for (int i = 0; i < qir_get_op_nsrc(qinst->op); i++) {
- if (qinst->src[i].file == QFILE_TEMP)
- reg_uses_remaining[qinst->src[i].index]++;
- }
- if (qinst->op == QOP_FRAG_Z)
- reg_in_use[3 + 32 + QPU_R_FRAG_PAYLOAD_ZW] = true;
- }
-
switch (c->stage) {
case QSTAGE_VERT:
case QSTAGE_COORD:
@@ -259,6 +227,7 @@ vc4_generate_code(struct vc4_compile *c)
break;
}
+ struct simple_node *node;
foreach(node, &c->instructions) {
struct qinst *qinst = (struct qinst *)node;
@@ -306,18 +275,7 @@ vc4_generate_code(struct vc4_compile *c)
src[i] = qpu_rn(0);
break;
case QFILE_TEMP:
- if (reg_allocated[index] == -1) {
- fprintf(stderr, "undefined reg use: ");
- qir_dump_inst(qinst);
- fprintf(stderr, "\n");
-
- src[i] = qpu_rn(0);
- } else {
- src[i] = allocate_to_qpu_reg[reg_allocated[index]];
- reg_uses_remaining[index]--;
- if (reg_uses_remaining[index] == 0)
- reg_in_use[reg_allocated[index]] = false;
- }
+ src[i] = temp_registers[index];
break;
case QFILE_UNIF:
src[i] = qpu_unif();
@@ -333,63 +291,9 @@ vc4_generate_code(struct vc4_compile *c)
case QFILE_NULL:
dst = qpu_ra(QPU_W_NOP);
break;
-
case QFILE_TEMP:
- if (reg_allocated[qinst->dst.index] == -1) {
- int alloc;
- for (alloc = 0;
- alloc < ARRAY_SIZE(reg_in_use);
- alloc++) {
- struct qpu_reg reg = allocate_to_qpu_reg[alloc];
-
- switch (qinst->op) {
- case QOP_PACK_SCALED:
- /* The pack flags require an
- * A-file register.
- */
- if (reg.mux != QPU_MUX_A)
- continue;
- break;
- case QOP_TEX_RESULT:
- case QOP_TLB_COLOR_READ:
- /* Only R4-generating
- * instructions get to store
- * values in R4 for now, until
- * we figure out how to do
- * interference.
- */
- if (reg.mux != QPU_MUX_R4)
- continue;
- break;
- case QOP_FRAG_Z:
- if (reg.mux != QPU_MUX_B ||
- reg.addr != QPU_R_FRAG_PAYLOAD_ZW) {
- continue;
- }
- break;
- default:
- if (reg.mux == QPU_MUX_R4)
- continue;
- break;
- }
-
- if (!reg_in_use[alloc])
- break;
- }
- assert(alloc != ARRAY_SIZE(reg_in_use) && "need better reg alloc");
- reg_in_use[alloc] = true;
- reg_allocated[qinst->dst.index] = alloc;
- }
-
- dst = allocate_to_qpu_reg[reg_allocated[qinst->dst.index]];
-
- reg_uses_remaining[qinst->dst.index]--;
- if (reg_uses_remaining[qinst->dst.index] == 0) {
- reg_in_use[reg_allocated[qinst->dst.index]] =
- false;
- }
+ dst = temp_registers[qinst->dst.index];
break;
-
case QFILE_VARY:
case QFILE_UNIF:
assert(!"not reached");
@@ -645,4 +549,6 @@ vc4_generate_code(struct vc4_compile *c)
vc4_dump_program(c);
vc4_qpu_validate(c->qpu_insts, c->qpu_inst_count);
+
+ free(temp_registers);
}
diff --git a/src/gallium/drivers/vc4/vc4_register_allocate.c b/src/gallium/drivers/vc4/vc4_register_allocate.c
new file mode 100644
index 0000000..97d4fb3
--- /dev/null
+++ b/src/gallium/drivers/vc4/vc4_register_allocate.c
@@ -0,0 +1,157 @@
+/*
+ * Copyright © 2014 Broadcom
+ *
+ * Permission is hereby granted, free of charge, to any person obtaining a
+ * copy of this software and associated documentation files (the "Software"),
+ * to deal in the Software without restriction, including without limitation
+ * the rights to use, copy, modify, merge, publish, distribute, sublicense,
+ * and/or sell copies of the Software, and to permit persons to whom the
+ * Software is furnished to do so, subject to the following conditions:
+ *
+ * The above copyright notice and this permission notice (including the next
+ * paragraph) shall be included in all copies or substantial portions of the
+ * Software.
+ *
+ * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
+ * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
+ * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL
+ * THE AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
+ * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING
+ * FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS
+ * IN THE SOFTWARE.
+ */
+
+#include <inttypes.h>
+
+#include "vc4_context.h"
+#include "vc4_qir.h"
+#include "vc4_qpu.h"
+
+/**
+ * Returns a mapping from QFILE_TEMP indices to struct qpu_regs.
+ *
+ * The return value should be freed by the caller.
+ */
+struct qpu_reg *
+vc4_register_allocate(struct vc4_compile *c)
+{
+ struct simple_node *node;
+ struct qpu_reg allocate_to_qpu_reg[4 + 32 + 32];
+ bool reg_in_use[ARRAY_SIZE(allocate_to_qpu_reg)];
+ int *reg_allocated = calloc(c->num_temps, sizeof(*reg_allocated));
+ int *reg_uses_remaining =
+ calloc(c->num_temps, sizeof(*reg_uses_remaining));
+ struct qpu_reg *temp_registers = calloc(c->num_temps,
+ sizeof(*temp_registers));
+
+ for (int i = 0; i < ARRAY_SIZE(reg_in_use); i++)
+ reg_in_use[i] = false;
+ for (int i = 0; i < c->num_temps; i++)
+ reg_allocated[i] = -1;
+
+ /* If things aren't ever written (undefined values), just read from
+ * r0.
+ */
+ for (int i = 0; i < c->num_temps; i++)
+ temp_registers[i] = qpu_rn(0);
+
+ uint32_t next_reg = 0;
+ for (int i = 0; i < 4; i++)
+ allocate_to_qpu_reg[next_reg++] = qpu_rn(i == 3 ? 4 : i);
+ for (int i = 0; i < 32; i++)
+ allocate_to_qpu_reg[next_reg++] = qpu_ra(i);
+ for (int i = 0; i < 32; i++)
+ allocate_to_qpu_reg[next_reg++] = qpu_rb(i);
+ assert(next_reg == ARRAY_SIZE(allocate_to_qpu_reg));
+
+ foreach(node, &c->instructions) {
+ struct qinst *qinst = (struct qinst *)node;
+
+ if (qinst->dst.file == QFILE_TEMP)
+ reg_uses_remaining[qinst->dst.index]++;
+ for (int i = 0; i < qir_get_op_nsrc(qinst->op); i++) {
+ if (qinst->src[i].file == QFILE_TEMP)
+ reg_uses_remaining[qinst->src[i].index]++;
+ }
+ if (qinst->op == QOP_FRAG_Z)
+ reg_in_use[3 + 32 + QPU_R_FRAG_PAYLOAD_ZW] = true;
+ }
+
+ foreach(node, &c->instructions) {
+ struct qinst *qinst = (struct qinst *)node;
+
+ for (int i = 0; i < qir_get_op_nsrc(qinst->op); i++) {
+ int index = qinst->src[i].index;
+
+ if (qinst->src[i].file != QFILE_TEMP)
+ continue;
+
+ if (reg_allocated[index] == -1) {
+ fprintf(stderr, "undefined reg use: ");
+ qir_dump_inst(qinst);
+ fprintf(stderr, "\n");
+ } else {
+ reg_uses_remaining[index]--;
+ if (reg_uses_remaining[index] == 0)
+ reg_in_use[reg_allocated[index]] = false;
+ }
+ }
+
+ if (qinst->dst.file == QFILE_TEMP) {
+ if (reg_allocated[qinst->dst.index] == -1) {
+ int alloc;
+ for (alloc = 0;
+ alloc < ARRAY_SIZE(reg_in_use);
+ alloc++) {
+ struct qpu_reg reg = allocate_to_qpu_reg[alloc];
+
+ switch (qinst->op) {
+ case QOP_PACK_SCALED:
+ /* The pack flags require an
+ * A-file register.
+ */
+ if (reg.mux != QPU_MUX_A)
+ continue;
+ break;
+ case QOP_TEX_RESULT:
+ case QOP_TLB_COLOR_READ:
+ /* Only R4-generating
+ * instructions get to store
+ * values in R4 for now, until
+ * we figure out how to do
+ * interference.
+ */
+ if (reg.mux != QPU_MUX_R4)
+ continue;
+ break;
+ case QOP_FRAG_Z:
+ if (reg.mux != QPU_MUX_B ||
+ reg.addr != QPU_R_FRAG_PAYLOAD_ZW) {
+ continue;
+ }
+ break;
+ default:
+ if (reg.mux == QPU_MUX_R4)
+ continue;
+ break;
+ }
+
+ if (!reg_in_use[alloc])
+ break;
+ }
+ assert(alloc != ARRAY_SIZE(reg_in_use) && "need better reg alloc");
+ reg_in_use[alloc] = true;
+ reg_allocated[qinst->dst.index] = alloc;
+ temp_registers[qinst->dst.index] = allocate_to_qpu_reg[alloc];
+ }
+
+ reg_uses_remaining[qinst->dst.index]--;
+ if (reg_uses_remaining[qinst->dst.index] == 0) {
+ reg_in_use[reg_allocated[qinst->dst.index]] =
+ false;
+ }
+ }
+ }
+
+ return temp_registers;
+}
More information about the mesa-commit
mailing list