xe: jit: reorder: allow emulated qw swizzle

Author: atkassen

src/gpu/intel/jit/codegen/reorder.hpp

@@ -244,15 +244,9 @@ void emit_reorder_1d_tile(ngen::HW hw, GeneratorT *host,
     // Replace (float -> float) by (int -> int) as word/dword moves have less
     // restrictions.
     if (src_type == dst_type && to_ir(src_type).is_fp()) {
-        int factor = (src_type == ngen::DataType::df ? 2 : 1);
-        if (factor == 1 || (src_stride == 1 && dst_stride == 1)) {
-            src_type
-                    = to_ngen(type_t::u(ngen::getBytes(src_type) / factor * 8));
-            dst_type = src_type;
-            width *= factor;
-            src = src.reinterpret(src_type);
-            dst = dst.reinterpret(dst_type);
-        }
+        src_type = dst_type = to_ngen(type_t::u(ngen::getBits(src_type)));
+        src = src.reinterpret(src_type);
+        dst = dst.reinterpret(dst_type);
     }
 
     const int grf_size = ngen::GRF::bytes(hw);
@@ -297,7 +291,7 @@ void emit_reorder_1d_tile(ngen::HW hw, GeneratorT *host,
                         || !dst.check_bounds(0, 64, /*is_dense=*/true)))
             step = 8;
 
-        if (src_df || dst_df) step = 8;
+        if (src_df || src_q || dst_df || dst_q) step = 8;
 
         // Max supported stride is 4.
         if (src_stride > 4 || dst_stride > 4) step = 1;
@@ -310,9 +304,6 @@ void emit_reorder_1d_tile(ngen::HW hw, GeneratorT *host,
         // Non-power-of-2 strides must be handled element-by-element
         if (!math::is_pow2(src_stride) || !math::is_pow2(dst_stride)) step = 1;
 
-        // Qword does not appear to support swizzling.
-        if (src_q && dst_q && src_stride != dst_stride) step = 1;
-
         return step;
     };
 
@@ -1893,8 +1884,6 @@ class reorder_impl_t {
         const int grf_size = ngen::GRF::bytes(hw_);
 
         if (src_layout_.type() != dst_layout_.type()) return false;
-        // long / f64 swizzle emits scalar instructions
-        if (src_layout_.type().scalar().size() >= 8) return false;
         if (!src_layout_.is_dense()) return false;
         if (!dst_layout_.is_dense()) return false;