+brw_typed_atomic(struct brw_codegen *p,
+ struct brw_reg dst,
+ struct brw_reg payload,
+ struct brw_reg surface,
+ unsigned atomic_op,
+ unsigned msg_length,
+ bool response_expected) {
+ const struct gen_device_info *devinfo = p->devinfo;
+ const unsigned sfid = (devinfo->gen >= 8 || devinfo->is_haswell ?
+ HSW_SFID_DATAPORT_DATA_CACHE_1 :
+ GEN6_SFID_DATAPORT_RENDER_CACHE);
+ const bool align1 = (brw_inst_access_mode(devinfo, p->current) == BRW_ALIGN_1);
+ /* Mask out unused components -- See comment in brw_untyped_atomic(). */
+ const unsigned mask = align1 ? WRITEMASK_XYZW : WRITEMASK_X;
+ struct brw_inst *insn = brw_send_indirect_surface_message(
+ p, sfid, brw_writemask(dst, mask), payload, surface, msg_length,
+ brw_surface_payload_size(p, response_expected,
+ devinfo->gen >= 8 || devinfo->is_haswell, false),
+ true);
+
+ brw_set_dp_typed_atomic_message(
+ p, insn, atomic_op, response_expected);
+}
+
+static void
+brw_set_dp_typed_surface_read_message(struct brw_codegen *p,
+ struct brw_inst *insn,
+ unsigned num_channels)
+{
+ const struct gen_device_info *devinfo = p->devinfo;
+ /* Set mask of unused channels. */
+ unsigned msg_control = 0xf & (0xf << num_channels);
+
+ if (devinfo->gen >= 8 || devinfo->is_haswell) {
+ if (brw_inst_access_mode(devinfo, p->current) == BRW_ALIGN_1) {
+ if (brw_inst_qtr_control(devinfo, p->current) % 2 == 1)
+ msg_control |= 2 << 4; /* Use high 8 slots of the sample mask */
+ else
+ msg_control |= 1 << 4; /* Use low 8 slots of the sample mask */
+ }
+
+ brw_inst_set_dp_msg_type(devinfo, insn,
+ HSW_DATAPORT_DC_PORT1_TYPED_SURFACE_READ);
+ } else {
+ if (brw_inst_access_mode(devinfo, p->current) == BRW_ALIGN_1) {
+ if (brw_inst_qtr_control(devinfo, p->current) % 2 == 1)
+ msg_control |= 1 << 5; /* Use high 8 slots of the sample mask */
+ }
+
+ brw_inst_set_dp_msg_type(devinfo, insn,
+ GEN7_DATAPORT_RC_TYPED_SURFACE_READ);
+ }
+
+ brw_inst_set_dp_msg_control(devinfo, insn, msg_control);
+}
+
+void
+brw_typed_surface_read(struct brw_codegen *p,
+ struct brw_reg dst,
+ struct brw_reg payload,
+ struct brw_reg surface,
+ unsigned msg_length,
+ unsigned num_channels)
+{
+ const struct gen_device_info *devinfo = p->devinfo;
+ const unsigned sfid = (devinfo->gen >= 8 || devinfo->is_haswell ?
+ HSW_SFID_DATAPORT_DATA_CACHE_1 :
+ GEN6_SFID_DATAPORT_RENDER_CACHE);
+ struct brw_inst *insn = brw_send_indirect_surface_message(
+ p, sfid, dst, payload, surface, msg_length,
+ brw_surface_payload_size(p, num_channels,
+ devinfo->gen >= 8 || devinfo->is_haswell, false),
+ true);
+
+ brw_set_dp_typed_surface_read_message(
+ p, insn, num_channels);
+}
+
+static void
+brw_set_dp_typed_surface_write_message(struct brw_codegen *p,
+ struct brw_inst *insn,
+ unsigned num_channels)
+{
+ const struct gen_device_info *devinfo = p->devinfo;
+ /* Set mask of unused channels. */
+ unsigned msg_control = 0xf & (0xf << num_channels);
+
+ if (devinfo->gen >= 8 || devinfo->is_haswell) {
+ if (brw_inst_access_mode(devinfo, p->current) == BRW_ALIGN_1) {
+ if (brw_inst_qtr_control(devinfo, p->current) % 2 == 1)
+ msg_control |= 2 << 4; /* Use high 8 slots of the sample mask */
+ else
+ msg_control |= 1 << 4; /* Use low 8 slots of the sample mask */
+ }
+
+ brw_inst_set_dp_msg_type(devinfo, insn,
+ HSW_DATAPORT_DC_PORT1_TYPED_SURFACE_WRITE);
+
+ } else {
+ if (brw_inst_access_mode(devinfo, p->current) == BRW_ALIGN_1) {
+ if (brw_inst_qtr_control(devinfo, p->current) % 2 == 1)
+ msg_control |= 1 << 5; /* Use high 8 slots of the sample mask */
+ }
+
+ brw_inst_set_dp_msg_type(devinfo, insn,
+ GEN7_DATAPORT_RC_TYPED_SURFACE_WRITE);
+ }
+
+ brw_inst_set_dp_msg_control(devinfo, insn, msg_control);
+}
+
+void
+brw_typed_surface_write(struct brw_codegen *p,
+ struct brw_reg payload,
+ struct brw_reg surface,
+ unsigned msg_length,
+ unsigned num_channels)
+{
+ const struct gen_device_info *devinfo = p->devinfo;
+ const unsigned sfid = (devinfo->gen >= 8 || devinfo->is_haswell ?
+ HSW_SFID_DATAPORT_DATA_CACHE_1 :
+ GEN6_SFID_DATAPORT_RENDER_CACHE);
+ const bool align1 = (brw_inst_access_mode(devinfo, p->current) == BRW_ALIGN_1);
+ /* Mask out unused components -- See comment in brw_untyped_atomic(). */
+ const unsigned mask = (devinfo->gen == 7 && !devinfo->is_haswell && !align1 ?
+ WRITEMASK_X : WRITEMASK_XYZW);
+ struct brw_inst *insn = brw_send_indirect_surface_message(
+ p, sfid, brw_writemask(brw_null_reg(), mask),
+ payload, surface, msg_length, 0, true);
+
+ brw_set_dp_typed_surface_write_message(
+ p, insn, num_channels);
+}
+
+static void
+brw_set_memory_fence_message(struct brw_codegen *p,
+ struct brw_inst *insn,
+ enum brw_message_target sfid,
+ bool commit_enable)
+{
+ const struct gen_device_info *devinfo = p->devinfo;
+
+ brw_set_message_descriptor(p, insn, sfid,
+ 1 /* message length */,
+ (commit_enable ? 1 : 0) /* response length */,
+ true /* header present */,
+ false);
+
+ switch (sfid) {
+ case GEN6_SFID_DATAPORT_RENDER_CACHE:
+ brw_inst_set_dp_msg_type(devinfo, insn, GEN7_DATAPORT_RC_MEMORY_FENCE);
+ break;
+ case GEN7_SFID_DATAPORT_DATA_CACHE:
+ brw_inst_set_dp_msg_type(devinfo, insn, GEN7_DATAPORT_DC_MEMORY_FENCE);
+ break;
+ default:
+ unreachable("Not reached");
+ }
+
+ if (commit_enable)
+ brw_inst_set_dp_msg_control(devinfo, insn, 1 << 5);
+}
+
+void
+brw_memory_fence(struct brw_codegen *p,
+ struct brw_reg dst)
+{
+ const struct gen_device_info *devinfo = p->devinfo;
+ const bool commit_enable = devinfo->gen == 7 && !devinfo->is_haswell;
+ struct brw_inst *insn;
+
+ brw_push_insn_state(p);
+ brw_set_default_mask_control(p, BRW_MASK_DISABLE);
+ brw_set_default_exec_size(p, BRW_EXECUTE_1);
+ dst = vec1(dst);
+
+ /* Set dst as destination for dependency tracking, the MEMORY_FENCE
+ * message doesn't write anything back.
+ */
+ insn = next_insn(p, BRW_OPCODE_SEND);
+ dst = retype(dst, BRW_REGISTER_TYPE_UW);
+ brw_set_dest(p, insn, dst);
+ brw_set_src0(p, insn, dst);
+ brw_set_memory_fence_message(p, insn, GEN7_SFID_DATAPORT_DATA_CACHE,
+ commit_enable);
+
+ if (devinfo->gen == 7 && !devinfo->is_haswell) {
+ /* IVB does typed surface access through the render cache, so we need to
+ * flush it too. Use a different register so both flushes can be
+ * pipelined by the hardware.
+ */
+ insn = next_insn(p, BRW_OPCODE_SEND);
+ brw_set_dest(p, insn, offset(dst, 1));
+ brw_set_src0(p, insn, offset(dst, 1));
+ brw_set_memory_fence_message(p, insn, GEN6_SFID_DATAPORT_RENDER_CACHE,
+ commit_enable);
+
+ /* Now write the response of the second message into the response of the
+ * first to trigger a pipeline stall -- This way future render and data
+ * cache messages will be properly ordered with respect to past data and
+ * render cache messages.
+ */
+ brw_MOV(p, dst, offset(dst, 1));
+ }
+
+ brw_pop_insn_state(p);
+}
+
+void
+brw_pixel_interpolator_query(struct brw_codegen *p,