intel/fs: Do 8-bit subgroup scan operations in 16 bits

Reviewed-by: Paulo Zanoni <[email protected]>

1 files changed, 39 insertions, 3 deletions

diff --git a/src/intel/compiler/brw_fs_nir.cpp b/src/intel/compiler/brw_fs_nir.cpp
index abce29c92f5..f9e81e14f3a 100644
--- a/src/intel/compiler/brw_fs_nir.cpp
+++ b/src/intel/compiler/brw_fs_nir.cpp
@@ -4925,10 +4925,28 @@ fs_visitor::nir_emit_intrinsic(const fs_builder &bld, nir_intrinsic_instr *instr
       opcode brw_op = brw_op_for_nir_reduction_op(redop);
       brw_conditional_mod cond_mod = brw_cond_mod_for_nir_reduction_op(redop);
 
+      /* There are a couple of register region issues that make things
+       * complicated for 8-bit types:
+       *
+       *    1. Only raw moves are allowed to write to a packed 8-bit
+       *       destination.
+       *    2. If we use a strided destination, the efficient way to do scan
+       *       operations ends up using strides that are too big to encode in
+       *       an instruction.
+       *
+       * To get around these issues, we just do all 8-bit scan operations in
+       * 16 bits.  It's actually fewer instructions than what we'd have to do
+       * if we were trying to do it in native 8-bit types and the results are
+       * the same once we truncate to 8 bits at the end.
+       */
+      brw_reg_type scan_type = src.type;
+      if (type_sz(scan_type) == 1)
+         scan_type = brw_reg_type_from_bit_size(16, src.type);
+
       /* Set up a register for all of our scratching around and initialize it
        * to reduction operation's identity value.
        */
-      fs_reg scan = bld.vgrf(src.type);
+      fs_reg scan = bld.vgrf(scan_type);
       bld.exec_all().emit(SHADER_OPCODE_SEL_EXEC, scan, src, identity);
 
       bld.emit_scan(brw_op, scan, cluster_size, cond_mod);
@@ -4971,10 +4989,28 @@ fs_visitor::nir_emit_intrinsic(const fs_builder &bld, nir_intrinsic_instr *instr
       opcode brw_op = brw_op_for_nir_reduction_op(redop);
       brw_conditional_mod cond_mod = brw_cond_mod_for_nir_reduction_op(redop);
 
+      /* There are a couple of register region issues that make things
+       * complicated for 8-bit types:
+       *
+       *    1. Only raw moves are allowed to write to a packed 8-bit
+       *       destination.
+       *    2. If we use a strided destination, the efficient way to do scan
+       *       operations ends up using strides that are too big to encode in
+       *       an instruction.
+       *
+       * To get around these issues, we just do all 8-bit scan operations in
+       * 16 bits.  It's actually fewer instructions than what we'd have to do
+       * if we were trying to do it in native 8-bit types and the results are
+       * the same once we truncate to 8 bits at the end.
+       */
+      brw_reg_type scan_type = src.type;
+      if (type_sz(scan_type) == 1)
+         scan_type = brw_reg_type_from_bit_size(16, src.type);
+
       /* Set up a register for all of our scratching around and initialize it
        * to reduction operation's identity value.
        */
-      fs_reg scan = bld.vgrf(src.type);
+      fs_reg scan = bld.vgrf(scan_type);
       const fs_builder allbld = bld.exec_all();
       allbld.emit(SHADER_OPCODE_SEL_EXEC, scan, src, identity);
 
@@ -4983,7 +5019,7 @@ fs_visitor::nir_emit_intrinsic(const fs_builder &bld, nir_intrinsic_instr *instr
           * shift of the contents before we can begin.  To make things worse,
           * we can't do this with a normal stride; we have to use indirects.
           */
-         fs_reg shifted = bld.vgrf(src.type);
+         fs_reg shifted = bld.vgrf(scan_type);
          fs_reg idx = bld.vgrf(BRW_REGISTER_TYPE_W);
          allbld.ADD(idx, nir_system_values[SYSTEM_VALUE_SUBGROUP_INVOCATION],
                          brw_imm_w(-1));