vc4: Add another check for invalid TLB scoreboard handling.

This was caught by an assertion in the simulator.

diff --git a/src/gallium/drivers/vc4/vc4_qpu.c b/src/gallium/drivers/vc4/vc4_qpu.c
index a551a0f..093ca07 100644 (file)
--- a/src/gallium/drivers/vc4/vc4_qpu.c
+++ b/src/gallium/drivers/vc4/vc4_qpu.c
@@ -244,3 +244,26 @@ qpu_set_cond_mul(uint64_t inst, uint32_t sig)
         return (inst & ~QPU_COND_MUL_MASK) | QPU_SET_FIELD(sig, QPU_COND_MUL);
 }
 
+bool
+qpu_waddr_is_tlb(uint32_t waddr)
+{
+        switch (waddr) {
+        case QPU_W_TLB_COLOR_ALL:
+        case QPU_W_TLB_COLOR_MS:
+        case QPU_W_TLB_Z:
+                return true;
+        default:
+                return false;
+        }
+}
+
+bool
+qpu_inst_is_tlb(uint64_t inst)
+{
+        uint32_t sig = QPU_GET_FIELD(inst, QPU_SIG);
+
+        return (qpu_waddr_is_tlb(QPU_GET_FIELD(inst, QPU_WADDR_ADD)) ||
+                qpu_waddr_is_tlb(QPU_GET_FIELD(inst, QPU_WADDR_MUL)) ||
+                sig == QPU_SIG_COLOR_LOAD ||
+                sig == QPU_SIG_WAIT_FOR_SCOREBOARD);
+}

diff --git a/src/gallium/drivers/vc4/vc4_qpu.h b/src/gallium/drivers/vc4/vc4_qpu.h
index 2f37108..5f4caab 100644 (file)
--- a/src/gallium/drivers/vc4/vc4_qpu.h
+++ b/src/gallium/drivers/vc4/vc4_qpu.h
@@ -133,6 +133,9 @@ uint64_t qpu_set_sig(uint64_t inst, uint32_t sig);
 uint64_t qpu_set_cond_add(uint64_t inst, uint32_t cond);
 uint64_t qpu_set_cond_mul(uint64_t inst, uint32_t cond);
 
+bool qpu_waddr_is_tlb(uint32_t waddr);
+bool qpu_inst_is_tlb(uint64_t inst);
+
 static inline uint64_t
 qpu_load_imm_f(struct qpu_reg dst, float val)
 {

diff --git a/src/gallium/drivers/vc4/vc4_qpu_validate.c b/src/gallium/drivers/vc4/vc4_qpu_validate.c
index d043733..8fe5f41 100644 (file)
--- a/src/gallium/drivers/vc4/vc4_qpu_validate.c
+++ b/src/gallium/drivers/vc4/vc4_qpu_validate.c
@@ -91,11 +91,17 @@ writes_sfu(uint64_t inst)
 void
 vc4_qpu_validate(uint64_t *insts, uint32_t num_inst)
 {
+        bool scoreboard_locked = false;
+
         for (int i = 0; i < num_inst; i++) {
                 uint64_t inst = insts[i];
 
-                if (QPU_GET_FIELD(inst, QPU_SIG) != QPU_SIG_PROG_END)
+                if (QPU_GET_FIELD(inst, QPU_SIG) != QPU_SIG_PROG_END) {
+                        if (qpu_inst_is_tlb(inst))
+                                scoreboard_locked = true;
+
                         continue;
+                }
 
                 /* "The Thread End instruction must not write to either physical
                  *  regfile A or B."
@@ -103,6 +109,11 @@ vc4_qpu_validate(uint64_t *insts, uint32_t num_inst)
                 assert(QPU_GET_FIELD(inst, QPU_WADDR_ADD) >= 32);
                 assert(QPU_GET_FIELD(inst, QPU_WADDR_MUL) >= 32);
 
+                /* Can't trigger an implicit wait on scoreboard in the program
+                 * end instruction.
+                 */
+                assert(!qpu_inst_is_tlb(inst) || scoreboard_locked);
+
                 /* Two delay slots will be executed. */
                 assert(i + 2 <= num_inst);
 
@@ -141,13 +152,7 @@ vc4_qpu_validate(uint64_t *insts, uint32_t num_inst)
         for (int i = 0; i < 2; i++) {
                 uint64_t inst = insts[i];
 
-                assert(QPU_GET_FIELD(inst, QPU_SIG) != QPU_SIG_COLOR_LOAD);
-                assert(QPU_GET_FIELD(inst, QPU_SIG) !=
-                       QPU_SIG_WAIT_FOR_SCOREBOARD);
-                assert(!writes_reg(inst, QPU_W_TLB_COLOR_MS));
-                assert(!writes_reg(inst, QPU_W_TLB_COLOR_ALL));
-                assert(!writes_reg(inst, QPU_W_TLB_Z));
-
+                assert(!qpu_inst_is_tlb(inst));
         }
 
         /* "If TMU_NOSWAP is written, the write must be three instructions