+/* TODO we probably have more than 8 scratch regs.. although the first
+ * 8 is what kernel dumps, and it is kinda useful to be able to see
+ * the value in kernel traces
+ */
+#define OVERFLOW_FLAG_REG REG_A6XX_CP_SCRATCH_REG(0)
+
+/*
+ * If overflow is detected, either 0x1 (VSC_DATA overflow) or 0x3
+ * (VSC_DATA2 overflow) plus the size of the overflowed buffer is
+ * written to control->vsc_overflow. This allows the CPU to
+ * detect which buffer overflowed (and, since the current size is
+ * encoded as well, this protects against already-submitted but
+ * not executed batches from fooling the CPU into increasing the
+ * size again unnecessarily).
+ *
+ * To conditionally use VSC data in draw pass only if there is no
+ * overflow, we use a scratch reg (OVERFLOW_FLAG_REG) to hold 1
+ * if no overflow, or 0 in case of overflow. The value is inverted
+ * to make the CP_COND_REG_EXEC stuff easier.
+ */
+static void
+emit_vsc_overflow_test(struct fd_batch *batch)
+{
+ struct fd_ringbuffer *ring = batch->gmem;
+ struct fd_gmem_stateobj *gmem = &batch->ctx->gmem;
+ struct fd6_context *fd6_ctx = fd6_context(batch->ctx);
+
+ debug_assert((fd6_ctx->vsc_data_pitch & 0x3) == 0);
+ debug_assert((fd6_ctx->vsc_data2_pitch & 0x3) == 0);
+
+ /* Clear vsc_scratch: */
+ OUT_PKT7(ring, CP_MEM_WRITE, 3);
+ OUT_RELOCW(ring, control_ptr(fd6_ctx, vsc_scratch));
+ OUT_RING(ring, 0x0);
+
+ /* Check for overflow, write vsc_scratch if detected: */
+ for (int i = 0; i < gmem->num_vsc_pipes; i++) {
+ OUT_PKT7(ring, CP_COND_WRITE5, 8);
+ OUT_RING(ring, CP_COND_WRITE5_0_FUNCTION(WRITE_GE) |
+ CP_COND_WRITE5_0_WRITE_MEMORY);
+ OUT_RING(ring, CP_COND_WRITE5_1_POLL_ADDR_LO(REG_A6XX_VSC_SIZE_REG(i)));
+ OUT_RING(ring, CP_COND_WRITE5_2_POLL_ADDR_HI(0));
+ OUT_RING(ring, CP_COND_WRITE5_3_REF(fd6_ctx->vsc_data_pitch));
+ OUT_RING(ring, CP_COND_WRITE5_4_MASK(~0));
+ OUT_RELOCW(ring, control_ptr(fd6_ctx, vsc_scratch)); /* WRITE_ADDR_LO/HI */
+ OUT_RING(ring, CP_COND_WRITE5_7_WRITE_DATA(1 + fd6_ctx->vsc_data_pitch));
+
+ OUT_PKT7(ring, CP_COND_WRITE5, 8);
+ OUT_RING(ring, CP_COND_WRITE5_0_FUNCTION(WRITE_GE) |
+ CP_COND_WRITE5_0_WRITE_MEMORY);
+ OUT_RING(ring, CP_COND_WRITE5_1_POLL_ADDR_LO(REG_A6XX_VSC_SIZE2_REG(i)));
+ OUT_RING(ring, CP_COND_WRITE5_2_POLL_ADDR_HI(0));
+ OUT_RING(ring, CP_COND_WRITE5_3_REF(fd6_ctx->vsc_data2_pitch));
+ OUT_RING(ring, CP_COND_WRITE5_4_MASK(~0));
+ OUT_RELOCW(ring, control_ptr(fd6_ctx, vsc_scratch)); /* WRITE_ADDR_LO/HI */
+ OUT_RING(ring, CP_COND_WRITE5_7_WRITE_DATA(3 + fd6_ctx->vsc_data2_pitch));
+ }
+
+ OUT_PKT7(ring, CP_WAIT_MEM_WRITES, 0);
+
+ OUT_PKT7(ring, CP_WAIT_FOR_ME, 0);
+
+ OUT_PKT7(ring, CP_MEM_TO_REG, 3);
+ OUT_RING(ring, CP_MEM_TO_REG_0_REG(OVERFLOW_FLAG_REG) |
+ CP_MEM_TO_REG_0_CNT(1 - 1));
+ OUT_RELOC(ring, control_ptr(fd6_ctx, vsc_scratch)); /* SRC_LO/HI */
+
+ /*
+ * This is a bit awkward, we really want a way to invert the
+ * CP_REG_TEST/CP_COND_REG_EXEC logic, so that we can conditionally
+ * execute cmds to use hwbinning when a bit is *not* set. This
+ * dance is to invert OVERFLOW_FLAG_REG
+ *
+ * A CP_NOP packet is used to skip executing the 'else' clause
+ * if (b0 set)..
+ */
+
+ BEGIN_RING(ring, 10); /* ensure if/else doesn't get split */
+
+ /* b0 will be set if VSC_DATA or VSC_DATA2 overflow: */
+ OUT_PKT7(ring, CP_REG_TEST, 1);
+ OUT_RING(ring, A6XX_CP_REG_TEST_0_REG(OVERFLOW_FLAG_REG) |
+ A6XX_CP_REG_TEST_0_BIT(0) |
+ A6XX_CP_REG_TEST_0_UNK25);
+
+ OUT_PKT7(ring, CP_COND_REG_EXEC, 2);
+ OUT_RING(ring, 0x10000000);
+ OUT_RING(ring, 7); /* conditionally execute next 7 dwords */
+
+ /* if (b0 set) */ {
+ /*
+ * On overflow, mirror the value to control->vsc_overflow
+ * which CPU is checking to detect overflow (see
+ * check_vsc_overflow())
+ */
+ OUT_PKT7(ring, CP_REG_TO_MEM, 3);
+ OUT_RING(ring, CP_REG_TO_MEM_0_REG(OVERFLOW_FLAG_REG) |
+ CP_REG_TO_MEM_0_CNT(1 - 1));
+ OUT_RELOCW(ring, control_ptr(fd6_ctx, vsc_overflow));
+
+ OUT_PKT4(ring, OVERFLOW_FLAG_REG, 1);
+ OUT_RING(ring, 0x0);
+
+ OUT_PKT7(ring, CP_NOP, 2); /* skip 'else' when 'if' is taken */
+ } /* else */ {
+ OUT_PKT4(ring, OVERFLOW_FLAG_REG, 1);
+ OUT_RING(ring, 0x1);
+ }
+}
+
+static void
+check_vsc_overflow(struct fd_context *ctx)
+{
+ struct fd6_context *fd6_ctx = fd6_context(ctx);
+ struct fd6_control *control = fd_bo_map(fd6_ctx->control_mem);
+ uint32_t vsc_overflow = control->vsc_overflow;
+
+ if (!vsc_overflow)
+ return;
+
+ /* clear overflow flag: */
+ control->vsc_overflow = 0;
+
+ unsigned buffer = vsc_overflow & 0x3;
+ unsigned size = vsc_overflow & ~0x3;
+
+ if (buffer == 0x1) {
+ /* VSC_PIPE_DATA overflow: */
+
+ if (size < fd6_ctx->vsc_data_pitch) {
+ /* we've already increased the size, this overflow is
+ * from a batch submitted before resize, but executed
+ * after
+ */
+ return;
+ }
+
+ fd_bo_del(fd6_ctx->vsc_data);
+ fd6_ctx->vsc_data = NULL;
+ fd6_ctx->vsc_data_pitch *= 2;
+
+ debug_printf("resized VSC_DATA_PITCH to: 0x%x\n", fd6_ctx->vsc_data_pitch);
+
+ } else if (buffer == 0x3) {
+ /* VSC_PIPE_DATA2 overflow: */
+
+ if (size < fd6_ctx->vsc_data2_pitch) {
+ /* we've already increased the size */
+ return;
+ }
+
+ fd_bo_del(fd6_ctx->vsc_data2);
+ fd6_ctx->vsc_data2 = NULL;
+ fd6_ctx->vsc_data2_pitch *= 2;
+
+ debug_printf("resized VSC_DATA2_PITCH to: 0x%x\n", fd6_ctx->vsc_data2_pitch);
+
+ } else {
+ /* NOTE: it's possible, for example, for overflow to corrupt the
+ * control page. I mostly just see this hit if I set initial VSC
+ * buffer size extremely small. Things still seem to recover,
+ * but maybe we should pre-emptively realloc vsc_data/vsc_data2
+ * and hope for different memory placement?
+ */
+ DBG("invalid vsc_overflow value: 0x%08x", vsc_overflow);
+ }
+}
+
+/*
+ * Emit conditional CP_INDIRECT_BRANCH based on VSC_STATE[p], ie. the IB
+ * is skipped for tiles that have no visible geometry.
+ */
+static void
+emit_conditional_ib(struct fd_batch *batch, struct fd_tile *tile,
+ struct fd_ringbuffer *target)
+{
+ struct fd_ringbuffer *ring = batch->gmem;
+
+ if (target->cur == target->start)
+ return;
+
+ emit_marker6(ring, 6);
+
+ unsigned count = fd_ringbuffer_cmd_count(target);
+
+ BEGIN_RING(ring, 5 + 4 * count); /* ensure conditional doesn't get split */
+
+ OUT_PKT7(ring, CP_REG_TEST, 1);
+ OUT_RING(ring, A6XX_CP_REG_TEST_0_REG(REG_A6XX_VSC_STATE_REG(tile->p)) |
+ A6XX_CP_REG_TEST_0_BIT(tile->n) |
+ A6XX_CP_REG_TEST_0_UNK25);
+
+ OUT_PKT7(ring, CP_COND_REG_EXEC, 2);
+ OUT_RING(ring, 0x10000000);
+ OUT_RING(ring, 4 * count); /* conditionally execute next 4*count dwords */
+
+ for (unsigned i = 0; i < count; i++) {
+ uint32_t dwords;
+ OUT_PKT7(ring, CP_INDIRECT_BUFFER, 3);
+ dwords = fd_ringbuffer_emit_reloc_ring_full(ring, target, i) / 4;
+ assert(dwords > 0);
+ OUT_RING(ring, dwords);
+ }
+
+ emit_marker6(ring, 6);
+}
+