4 files changed, 30 insertions, 3 deletions

diff --git a/src/intel/compiler/brw_eu_defines.h b/src/intel/compiler/brw_eu_defines.h
index 3ce7cca0eb1..da723307b73 100644
--- a/src/intel/compiler/brw_eu_defines.h
+++ b/src/intel/compiler/brw_eu_defines.h
@@ -423,6 +423,7 @@ enum opcode {
    SHADER_OPCODE_A64_BYTE_SCATTERED_READ_LOGICAL,
    SHADER_OPCODE_A64_BYTE_SCATTERED_WRITE_LOGICAL,
    SHADER_OPCODE_A64_UNTYPED_ATOMIC_LOGICAL,
+   SHADER_OPCODE_A64_UNTYPED_ATOMIC_INT64_LOGICAL,
    SHADER_OPCODE_A64_UNTYPED_ATOMIC_FLOAT_LOGICAL,
 
    SHADER_OPCODE_TYPED_ATOMIC_LOGICAL,
diff --git a/src/intel/compiler/brw_fs.cpp b/src/intel/compiler/brw_fs.cpp
index 0a390e54354..9f82946f078 100644
--- a/src/intel/compiler/brw_fs.cpp
+++ b/src/intel/compiler/brw_fs.cpp
@@ -855,6 +855,7 @@ fs_inst::components_read(unsigned i) const
       return i == 1 ? src[2].ud : 1;
 
    case SHADER_OPCODE_A64_UNTYPED_ATOMIC_LOGICAL:
+   case SHADER_OPCODE_A64_UNTYPED_ATOMIC_INT64_LOGICAL:
       assert(src[2].file == IMM);
       if (i == 1) {
          /* Data source */
@@ -5298,7 +5299,7 @@ lower_a64_logical_send(const fs_builder &bld, fs_inst *inst)
    if (devinfo->gen >= 9) {
       /* On Skylake and above, we have SENDS */
       mlen = 2 * (inst->exec_size / 8);
-      ex_mlen = src_comps * (inst->exec_size / 8);
+      ex_mlen = src_comps * type_sz(src.type) * inst->exec_size / REG_SIZE;
       payload = retype(bld.move_to_vgrf(addr, 1), BRW_REGISTER_TYPE_UD);
       payload2 = retype(bld.move_to_vgrf(src, src_comps),
                         BRW_REGISTER_TYPE_UD);
@@ -5350,6 +5351,13 @@ lower_a64_logical_send(const fs_builder &bld, fs_inst *inst)
                                             !inst->dst.is_null());
       break;
 
+   case SHADER_OPCODE_A64_UNTYPED_ATOMIC_INT64_LOGICAL:
+      desc = brw_dp_a64_untyped_atomic_desc(devinfo, inst->exec_size, 64,
+                                            arg,   /* atomic_op */
+                                            !inst->dst.is_null());
+      break;
+
+
    case SHADER_OPCODE_A64_UNTYPED_ATOMIC_FLOAT_LOGICAL:
       desc = brw_dp_a64_untyped_atomic_float_desc(devinfo, inst->exec_size,
                                                   arg,   /* atomic_op */
@@ -5558,6 +5566,7 @@ fs_visitor::lower_logical_sends()
       case SHADER_OPCODE_A64_BYTE_SCATTERED_WRITE_LOGICAL:
       case SHADER_OPCODE_A64_BYTE_SCATTERED_READ_LOGICAL:
       case SHADER_OPCODE_A64_UNTYPED_ATOMIC_LOGICAL:
+      case SHADER_OPCODE_A64_UNTYPED_ATOMIC_INT64_LOGICAL:
       case SHADER_OPCODE_A64_UNTYPED_ATOMIC_FLOAT_LOGICAL:
          lower_a64_logical_send(ibld, inst);
          break;
@@ -6147,6 +6156,7 @@ get_lowered_simd_width(const struct gen_device_info *devinfo,
       return devinfo->gen <= 8 ? 8 : MIN2(16, inst->exec_size);
 
    case SHADER_OPCODE_A64_UNTYPED_ATOMIC_LOGICAL:
+   case SHADER_OPCODE_A64_UNTYPED_ATOMIC_INT64_LOGICAL:
    case SHADER_OPCODE_A64_UNTYPED_ATOMIC_FLOAT_LOGICAL:
       return 8;
 
diff --git a/src/intel/compiler/brw_fs_nir.cpp b/src/intel/compiler/brw_fs_nir.cpp
index cf044a3f613..4bf85bfb7e8 100644
--- a/src/intel/compiler/brw_fs_nir.cpp
+++ b/src/intel/compiler/brw_fs_nir.cpp
@@ -4928,6 +4928,13 @@ fs_visitor::nir_emit_ssbo_atomic(const fs_builder &bld,
    if (stage == MESA_SHADER_FRAGMENT)
       brw_wm_prog_data(prog_data)->has_side_effects = true;
 
+   /* The BTI untyped atomic messages only support 32-bit atomics.  If you
+    * just look at the big table of messages in the Vol 7 of the SKL PRM, they
+    * appear to exist.  However, if you look at Vol 2a, there are no message
+    * descriptors provided for Qword atomic ops except for A64 messages.
+    */
+   assert(nir_dest_bit_size(instr->dest) == 32);
+
    fs_reg dest;
    if (nir_intrinsic_infos[instr->intrinsic].has_dest)
       dest = get_nir_dest(instr->dest);
@@ -5092,8 +5099,14 @@ fs_visitor::nir_emit_global_atomic(const fs_builder &bld,
       data = tmp;
    }
 
-   bld.emit(SHADER_OPCODE_A64_UNTYPED_ATOMIC_LOGICAL,
-            dest, addr, data, brw_imm_ud(op));
+   if (nir_dest_bit_size(instr->dest) == 64) {
+      bld.emit(SHADER_OPCODE_A64_UNTYPED_ATOMIC_INT64_LOGICAL,
+               dest, addr, data, brw_imm_ud(op));
+   } else {
+      assert(nir_dest_bit_size(instr->dest) == 32);
+      bld.emit(SHADER_OPCODE_A64_UNTYPED_ATOMIC_LOGICAL,
+               dest, addr, data, brw_imm_ud(op));
+   }
 }
 
 void
diff --git a/src/intel/compiler/brw_shader.cpp b/src/intel/compiler/brw_shader.cpp
index f7ed108bd07..1f98bd08224 100644
--- a/src/intel/compiler/brw_shader.cpp
+++ b/src/intel/compiler/brw_shader.cpp
@@ -308,6 +308,8 @@ brw_instruction_name(const struct gen_device_info *devinfo, enum opcode op)
       return "a64_byte_scattered_write_logical";
    case SHADER_OPCODE_A64_UNTYPED_ATOMIC_LOGICAL:
       return "a64_untyped_atomic_logical";
+   case SHADER_OPCODE_A64_UNTYPED_ATOMIC_INT64_LOGICAL:
+      return "a64_untyped_atomic_int64_logical";
    case SHADER_OPCODE_A64_UNTYPED_ATOMIC_FLOAT_LOGICAL:
       return "a64_untyped_atomic_float_logical";
    case SHADER_OPCODE_TYPED_ATOMIC_LOGICAL:
@@ -1044,6 +1046,7 @@ backend_instruction::has_side_effects() const
    case SHADER_OPCODE_A64_UNTYPED_WRITE_LOGICAL:
    case SHADER_OPCODE_A64_BYTE_SCATTERED_WRITE_LOGICAL:
    case SHADER_OPCODE_A64_UNTYPED_ATOMIC_LOGICAL:
+   case SHADER_OPCODE_A64_UNTYPED_ATOMIC_INT64_LOGICAL:
    case SHADER_OPCODE_A64_UNTYPED_ATOMIC_FLOAT_LOGICAL:
    case SHADER_OPCODE_BYTE_SCATTERED_WRITE_LOGICAL:
    case SHADER_OPCODE_TYPED_ATOMIC_LOGICAL: