+namespace brw {
+ inline fs_reg
+ fetch_payload_reg(const brw::fs_builder &bld, uint8_t regs[2],
+ brw_reg_type type = BRW_REGISTER_TYPE_F, unsigned n = 1)
+ {
+ if (!regs[0])
+ return fs_reg();
+
+ if (bld.dispatch_width() > 16) {
+ const fs_reg tmp = bld.vgrf(type, n);
+ const brw::fs_builder hbld = bld.exec_all().group(16, 0);
+ const unsigned m = bld.dispatch_width() / hbld.dispatch_width();
+ fs_reg *const components = new fs_reg[n * m];
+
+ for (unsigned c = 0; c < n; c++) {
+ for (unsigned g = 0; g < m; g++) {
+ components[c * m + g] =
+ offset(retype(brw_vec8_grf(regs[g], 0), type), hbld, c);
+ }
+ }
+
+ hbld.LOAD_PAYLOAD(tmp, components, n * m, 0);
+
+ delete[] components;
+ return tmp;
+
+ } else {
+ return fs_reg(retype(brw_vec8_grf(regs[0], 0), type));
+ }
+ }
+}
+
+void shuffle_from_32bit_read(const brw::fs_builder &bld,
+ const fs_reg &dst,
+ const fs_reg &src,
+ uint32_t first_component,
+ uint32_t components);
+
+fs_reg shuffle_for_32bit_write(const brw::fs_builder &bld,
+ const fs_reg &src,
+ uint32_t first_component,
+ uint32_t components);