Mesa (main): aco: use v_perm_b32 to copy 0xff00/0x00ff/0xff/0x00

[GitLab Mirror]

Module: Mesa
Branch: main
Commit: e68e6c75ca14c768b50d9ad85c37b2443699dcbb
URL:    http://cgit.freedesktop.org/mesa/mesa/commit/?id=e68e6c75ca14c768b50d9ad85c37b2443699dcbb

Author: Rhys Perry <pendingchaos02 at gmail.com>
Date:   Fri May 13 15:11:48 2022 +0100

aco: use v_perm_b32 to copy 0xff00/0x00ff/0xff/0x00

Signed-off-by: Rhys Perry <pendingchaos02 at gmail.com>
Reviewed-by: Timur Kristóf <timur.kristof at gmail.com>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/16595>

---

 src/amd/compiler/aco_lower_to_hw_instr.cpp  | 10 +++++++
 src/amd/compiler/tests/test_to_hw_instr.cpp | 44 +++++++++++++++++++++++++++++
 2 files changed, 54 insertions(+)

diff --git a/src/amd/compiler/aco_lower_to_hw_instr.cpp b/src/amd/compiler/aco_lower_to_hw_instr.cpp
index 2dc32f65ec1..40798da9597 100644
--- a/src/amd/compiler/aco_lower_to_hw_instr.cpp
+++ b/src/amd/compiler/aco_lower_to_hw_instr.cpp
@@ -1085,6 +1085,10 @@ copy_constant(lower_context* ctx, Builder& bld, Definition dst, Operand op)
       assert(dst.regClass() == v1b || dst.regClass() == v2b);
 
       bool use_sdwa = ctx->program->gfx_level >= GFX9 && ctx->program->gfx_level < GFX11;
+      /* We need the v_perm_b32 (VOP3) to be able to take literals, and that's a GFX10+ feature. */
+      bool can_use_perm = ctx->program->gfx_level >= GFX10 &&
+                          (op.constantEquals(0) || op.constantEquals(0xff) ||
+                           op.constantEquals(0xffff) || op.constantEquals(0xff00));
       if (dst.regClass() == v1b && use_sdwa) {
          uint8_t val = op.constantValue();
          Operand op32 = Operand::c32((uint32_t)val | (val & 0x80u ? 0xffffff00u : 0u));
@@ -1118,6 +1122,12 @@ copy_constant(lower_context* ctx, Builder& bld, Definition dst, Operand op)
             Instruction* instr = bld.vop3(aco_opcode::v_pack_b32_f16, dst, op, def_hi);
             instr->vop3().opsel = 2;
          }
+      } else if (can_use_perm) {
+         uint8_t swiz[] = {4, 5, 6, 7};
+         swiz[dst.physReg().byte()] = op.constantValue() & 0xff ? bperm_255 : bperm_0;
+         if (dst.bytes() == 2)
+            swiz[dst.physReg().byte() + 1] = op.constantValue() >> 8 ? bperm_255 : bperm_0;
+         create_bperm(bld, swiz, dst, Operand::zero());
       } else {
          uint32_t offset = dst.physReg().byte() * 8u;
          uint32_t mask = ((1u << (dst.bytes() * 8)) - 1) << offset;
diff --git a/src/amd/compiler/tests/test_to_hw_instr.cpp b/src/amd/compiler/tests/test_to_hw_instr.cpp
index 22587d4edea..67c6bf704bc 100644
--- a/src/amd/compiler/tests/test_to_hw_instr.cpp
+++ b/src/amd/compiler/tests/test_to_hw_instr.cpp
@@ -464,6 +464,50 @@ BEGIN_TEST(to_hw_instr.subdword_constant)
       bld.pseudo(aco_opcode::p_parallelcopy, Definition(v0_lo, v1), Definition(v1_lo, v1b),
                  Operand::zero(), Operand::zero(1));
 
+      bld.reset(program->create_and_insert_block());
+      program->blocks[0].linear_succs.push_back(1);
+      program->blocks[1].linear_preds.push_back(0);
+
+      /* Prevent usage of v_pack_b32_f16, so we use v_perm_b32 instead. */
+      program->blocks[1].fp_mode.denorm16_64 = fp_denorm_flush;
+
+      //>> p_unit_test 13
+      //~gfx9! v1: %_:v[0] = v_and_b32 0xffff0000, %_:v[0]
+      //~gfx9! v1: %_:v[0] = v_or_b32 0xff, %_:v[0]
+      //~gfx(10|11)! v1: %_:v[0] = v_perm_b32 0, %_:v[0], 0x7060c0d
+      bld.pseudo(aco_opcode::p_unit_test, Operand::c32(13u));
+      bld.pseudo(aco_opcode::p_parallelcopy, Definition(v0_lo, v2b),
+                 Operand::c16(0x00ff));
+
+      //! p_unit_test 14
+      //~gfx9! v1: %_:v[0] = v_and_b32 0xffff, %_:v[0]
+      //~gfx9! v1: %_:v[0] = v_or_b32 0xff000000, %_:v[0]
+      //~gfx(10|11)! v1: %_:v[0] = v_perm_b32 0, %_:v[0], 0xd0c0504
+      bld.pseudo(aco_opcode::p_unit_test, Operand::c32(14u));
+      bld.pseudo(aco_opcode::p_parallelcopy, Definition(v0_hi, v2b),
+                 Operand::c16(0xff00));
+
+      //! p_unit_test 15
+      //~gfx(9|10)! v2b: %_:v[0][0:16] = v_mov_b32 0 dst_sel:uword0 dst_preserve src0_sel:dword
+      //~gfx11! v1: %_:v[0] = v_perm_b32 0, %_:v[0], 0x7060c0c
+      bld.pseudo(aco_opcode::p_unit_test, Operand::c32(15u));
+      bld.pseudo(aco_opcode::p_parallelcopy, Definition(v0_lo, v2b),
+                 Operand::zero(2));
+
+      //! p_unit_test 16
+      //~gfx(9|10)! v1b: %_:v[0][0:8] = v_mov_b32 -1 dst_sel:ubyte0 dst_preserve src0_sel:dword
+      //~gfx11! v1: %_:v[0] = v_perm_b32 0, %_:v[0], 0x706050d
+      bld.pseudo(aco_opcode::p_unit_test, Operand::c32(16u));
+      bld.pseudo(aco_opcode::p_parallelcopy, Definition(v0_lo, v1b),
+                 Operand::c8(0xff));
+
+      //! p_unit_test 17
+      //~gfx(9|10)! v1b: %_:v[0][0:8] = v_mov_b32 0 dst_sel:ubyte0 dst_preserve src0_sel:dword
+      //~gfx11! v1: %_:v[0] = v_perm_b32 0, %_:v[0], 0x706050c
+      bld.pseudo(aco_opcode::p_unit_test, Operand::c32(17u));
+      bld.pseudo(aco_opcode::p_parallelcopy, Definition(v0_lo, v1b),
+                 Operand::zero(1));
+
       //! s_endpgm
 
       finish_to_hw_instr_test();