meson: Add support for the vc4 driver.

[mesa.git] / src / gallium / drivers / vc4 / vc4_qir_schedule.c

diff --git a/src/gallium/drivers/vc4/vc4_qir_schedule.c b/src/gallium/drivers/vc4/vc4_qir_schedule.c
index ea48a8583377996a55322ba4eef680134355505f..5118caf317c02f0d9a2ecb51d75712b4aa14dcb8 100644 (file)
--- a/src/gallium/drivers/vc4/vc4_qir_schedule.c
+++ b/src/gallium/drivers/vc4/vc4_qir_schedule.c
@@ -434,10 +434,20 @@ get_register_pressure_cost(struct schedule_state *state, struct qinst *inst)
                 cost--;
 
         for (int i = 0; i < qir_get_nsrc(inst); i++) {
-                if (inst->src[i].file == QFILE_TEMP &&
-                    !BITSET_TEST(state->temp_live, inst->src[i].index)) {
-                        cost++;
+                if (inst->src[i].file != QFILE_TEMP ||
+                    BITSET_TEST(state->temp_live, inst->src[i].index)) {
+                        continue;
                 }
+
+                bool already_counted = false;
+                for (int j = 0; j < i; j++) {
+                        if (inst->src[i].file == inst->src[j].file &&
+                            inst->src[i].index == inst->src[j].index) {
+                                already_counted = true;
+                        }
+                }
+                if (!already_counted)
+                        cost++;
         }
 
         return cost;
@@ -569,6 +579,28 @@ latency_between(struct schedule_node *before, struct schedule_node *after)
             after->inst->op == QOP_TEX_RESULT)
                 return 100;
 
+        switch (before->inst->op) {
+        case QOP_RCP:
+        case QOP_RSQ:
+        case QOP_EXP2:
+        case QOP_LOG2:
+                for (int i = 0; i < qir_get_nsrc(after->inst); i++) {
+                        if (after->inst->src[i].file ==
+                            before->inst->dst.file &&
+                            after->inst->src[i].index ==
+                            before->inst->dst.index) {
+                                /* There are two QPU delay slots before we can
+                                 * read a math result, which could be up to 4
+                                 * QIR instructions if they packed well.
+                                 */
+                                return 4;
+                        }
+                }
+                break;
+        default:
+                break;
+        }
+
         return 1;
 }