mesa/src/amd/compiler/tests/test_sdwa.cpp

/*
 * Copyright © 2020 Valve Corporation
 *
 * Permission is hereby granted, free of charge, to any person obtaining a
 * copy of this software and associated documentation files (the "Software"),
 * to deal in the Software without restriction, including without limitation
 * the rights to use, copy, modify, merge, publish, distribute, sublicense,
 * and/or sell copies of the Software, and to permit persons to whom the
 * Software is furnished to do so, subject to the following conditions:
 *
 * The above copyright notice and this permission notice (including the next
 * paragraph) shall be included in all copies or substantial portions of the
 * Software.
 *
 * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
 * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
 * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT.  IN NO EVENT SHALL
 * THE AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
 * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING
 * FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS
 * IN THE SOFTWARE.
 *
 */
#include "helpers.h"
#include <stdarg.h>

using namespace aco;

BEGIN_TEST(validate.sdwa.allow)
   for (unsigned i = GFX8; i <= GFX10; i++) {
      //>> v1: %a, v1: %b, s1: %c, s1: %d = p_startpgm
      if (!setup_cs("v1 v1 s1 s1", (amd_gfx_level)i))
         continue;
      //>> Validation results:
      //! Validation passed

      SDWA_instruction *sdwa = &bld.vop2_sdwa(aco_opcode::v_mul_f32, bld.def(v1), inputs[0], inputs[1])->sdwa();
      sdwa->neg[0] = sdwa->neg[1] = sdwa->abs[0] = sdwa->abs[1] = true;

      bld.vop2_sdwa(aco_opcode::v_mul_f32, bld.def(v1b), inputs[0], inputs[1]);

      sdwa = &bld.vop2_sdwa(aco_opcode::v_mul_f32, bld.def(v1), inputs[0], inputs[1])->sdwa();
      sdwa->sel[0] = SubdwordSel::sbyte2;
      sdwa->sel[1] = SubdwordSel::uword1;

      finish_validator_test();
   }
END_TEST

BEGIN_TEST(validate.sdwa.support)
   for (unsigned i = GFX7; i <= GFX11; i++) {
      //>> v1: %a, v1: %b, s1: %c, s1: %d = p_startpgm
      if (!setup_cs("v1 v1 s1 s1", (amd_gfx_level)i))
         continue;
      //>> Validation results:

      //~gfx(7|11)! SDWA is GFX8 to GFX10.3 only: v1: %t0 = v_mul_f32 %a, %b dst_sel:dword src0_sel:dword src1_sel:dword
      //~gfx(7|11)! Validation failed
      //~gfx([89]|10)! Validation passed
      bld.vop2_sdwa(aco_opcode::v_mul_f32, bld.def(v1), inputs[0], inputs[1]);

      finish_validator_test();
   }
END_TEST

BEGIN_TEST(validate.sdwa.operands)
   for (unsigned i = GFX8; i <= GFX10; i++) {
      //>> v1: %vgpr0, v1: %vgp1, s1: %sgpr0, s1: %sgpr1 = p_startpgm
      if (!setup_cs("v1 v1 s1 s1", (amd_gfx_level)i))
         continue;
      //>> Validation results:

      //~gfx8! Wrong source position for SGPR argument: v1: %_ = v_mul_f32 %sgpr0, %vgpr1 dst_sel:dword src0_sel:dword src1_sel:dword
      //~gfx8! Wrong source position for SGPR argument: v1: %_ = v_mul_f32 %vgpr0, %sgpr1 dst_sel:dword src0_sel:dword src1_sel:dword
      bld.vop2_sdwa(aco_opcode::v_mul_f32, bld.def(v1), inputs[2], inputs[1]);
      bld.vop2_sdwa(aco_opcode::v_mul_f32, bld.def(v1), inputs[0], inputs[3]);

      //~gfx8! Wrong source position for constant argument: v1: %_ = v_mul_f32 4, %vgpr1 dst_sel:dword src0_sel:dword src1_sel:dword
      //~gfx8! Wrong source position for constant argument: v1: %_ = v_mul_f32 %vgpr0, 4 dst_sel:dword src0_sel:dword src1_sel:dword
      bld.vop2_sdwa(aco_opcode::v_mul_f32, bld.def(v1), Operand::c32(4u), inputs[1]);
      bld.vop2_sdwa(aco_opcode::v_mul_f32, bld.def(v1), inputs[0], Operand::c32(4u));

      //! Literal applied on wrong instruction format: v1: %_ = v_mul_f32 0x1234, %vgpr1 dst_sel:dword src0_sel:dword src1_sel:dword
      //! Literal applied on wrong instruction format: v1: %_ = v_mul_f32 %vgpr0, 0x1234 dst_sel:dword src0_sel:dword src1_sel:dword
      //! Wrong source position for Literal argument: v1: %_ = v_mul_f32 %vgpr0, 0x1234 dst_sel:dword src0_sel:dword src1_sel:dword
      bld.vop2_sdwa(aco_opcode::v_mul_f32, bld.def(v1), Operand::c32(0x1234u), inputs[1]);
      bld.vop2_sdwa(aco_opcode::v_mul_f32, bld.def(v1), inputs[0], Operand::c32(0x1234u));

      //! Validation failed

      finish_validator_test();
   }
END_TEST

BEGIN_TEST(validate.sdwa.vopc)
   for (unsigned i = GFX8; i <= GFX10; i++) {
      //>> v1: %vgpr0, v1: %vgp1, s1: %sgpr0, s1: %sgpr1 = p_startpgm
      if (!setup_cs("v1 v1 s1 s1", (amd_gfx_level)i))
         continue;
      //>> Validation results:

      bld.vopc_sdwa(aco_opcode::v_cmp_gt_f32, bld.def(bld.lm, vcc), inputs[0], inputs[1]);

      //~gfx8! SDWA+VOPC definition must be fixed to vcc on GFX8: s2: %_ = v_cmp_lt_f32 %vgpr0, %vgpr1 src0_sel:dword src1_sel:dword
      bld.vopc_sdwa(aco_opcode::v_cmp_lt_f32, bld.def(bld.lm), inputs[0], inputs[1]);

      //~gfx(9|10)! SDWA VOPC clamp only supported on GFX8: s2: %_:vcc = v_cmp_eq_f32 %vgpr0, %vgpr1 clamp src0_sel:dword src1_sel:dword
      bld.vopc_sdwa(aco_opcode::v_cmp_eq_f32, bld.def(bld.lm, vcc), inputs[0], inputs[1])->sdwa().clamp = true;

      //! Validation failed

      finish_validator_test();
   }
END_TEST

BEGIN_TEST(validate.sdwa.omod)
   for (unsigned i = GFX8; i <= GFX10; i++) {
      //>> v1: %vgpr0, v1: %vgp1, s1: %sgpr0, s1: %sgpr1 = p_startpgm
      if (!setup_cs("v1 v1 s1 s1", (amd_gfx_level)i))
         continue;
      //>> Validation results:

      //~gfx8! SDWA omod only supported on GFX9+: v1: %_ = v_mul_f32 %vgpr0, %vgpr1 *2 dst_sel:dword src0_sel:dword src1_sel:dword
      //~gfx8! Validation failed
      //~gfx(9|10)! Validation passed
      bld.vop2_sdwa(aco_opcode::v_mul_f32, bld.def(v1), inputs[0], inputs[1])->sdwa().omod = 1;

      finish_validator_test();
   }
END_TEST

BEGIN_TEST(validate.sdwa.vcc)
   for (unsigned i = GFX8; i <= GFX10; i++) {
      //>> v1: %vgpr0, v1: %vgpr1, s2: %sgpr0 = p_startpgm
      if (!setup_cs("v1 v1 s2", (amd_gfx_level)i))
         continue;
      //>> Validation results:

      //! 3rd operand must be fixed to vcc with SDWA: v1: %_ = v_cndmask_b32 %vgpr0, %vgpr1, %_ dst_sel:dword src0_sel:dword src1_sel:dword
      bld.vop2_sdwa(aco_opcode::v_cndmask_b32, bld.def(v1), inputs[0], inputs[1], inputs[2]);
      bld.vop2_sdwa(aco_opcode::v_cndmask_b32, bld.def(v1), inputs[0], inputs[1], bld.vcc(inputs[2]));

      //! 2nd definition must be fixed to vcc with SDWA: v1: %_, s2: %_ = v_add_co_u32 %vgpr0, %vgpr1 dst_sel:dword src0_sel:dword src1_sel:dword
      bld.vop2_sdwa(aco_opcode::v_add_co_u32, bld.def(v1), bld.def(bld.lm), inputs[0], inputs[1]);
      bld.vop2_sdwa(aco_opcode::v_add_co_u32, bld.def(v1), bld.def(bld.lm, vcc), inputs[0], inputs[1]);

      //! Validation failed

      finish_validator_test();
   }
END_TEST

BEGIN_TEST(optimize.sdwa.extract)
   for (unsigned i = GFX7; i <= GFX10; i++) {
   for (unsigned is_signed = 0; is_signed <= 1; is_signed++) {
      //>> v1: %a, v1: %b, s1: %c, s1: %d = p_startpgm
      if (!setup_cs("v1 v1 s1 s1", (amd_gfx_level)i, CHIP_UNKNOWN, is_signed ? "_signed" : "_unsigned"))
         continue;

      //; def standard_test(index, sel):
      //;    res = 'v1: %%res%s = v_mul_f32 %%a, %%b dst_sel:dword src0_sel:dword src1_sel:%c%s\n' % (index, 's' if variant.endswith('_signed') else 'u', sel)
      //;    res += 'p_unit_test %s, %%res%s' % (index, index)
      //;    return res
      //; funcs['standard_test'] = lambda a: standard_test(*(v for v in a.split(',')))

      aco_opcode ext = aco_opcode::p_extract;
      aco_opcode ins = aco_opcode::p_insert;

      {
      //~gfx[^7].*! @standard_test(0,byte0)
      Temp bfe_byte0_b = bld.pseudo(ext, bld.def(v1), inputs[1], Operand::zero(), Operand::c32(8u),
                                    Operand::c32(is_signed));
      writeout(0, bld.vop2(aco_opcode::v_mul_f32, bld.def(v1), inputs[0], bfe_byte0_b));

      //~gfx[^7].*! @standard_test(1,byte1)
      Temp bfe_byte1_b = bld.pseudo(ext, bld.def(v1), inputs[1], Operand::c32(1u), Operand::c32(8u),
                                    Operand::c32(is_signed));
      writeout(1, bld.vop2(aco_opcode::v_mul_f32, bld.def(v1), inputs[0], bfe_byte1_b));

      //~gfx[^7].*! @standard_test(2,byte2)
      Temp bfe_byte2_b = bld.pseudo(ext, bld.def(v1), inputs[1], Operand::c32(2u), Operand::c32(8u),
                                    Operand::c32(is_signed));
      writeout(2, bld.vop2(aco_opcode::v_mul_f32, bld.def(v1), inputs[0], bfe_byte2_b));

      //~gfx[^7].*! @standard_test(3,byte3)
      Temp bfe_byte3_b = bld.pseudo(ext, bld.def(v1), inputs[1], Operand::c32(3u), Operand::c32(8u),
                                    Operand::c32(is_signed));
      writeout(3, bld.vop2(aco_opcode::v_mul_f32, bld.def(v1), inputs[0], bfe_byte3_b));

      //~gfx[^7].*! @standard_test(4,word0)
      Temp bfe_word0_b = bld.pseudo(ext, bld.def(v1), inputs[1], Operand::zero(), Operand::c32(16u),
                                    Operand::c32(is_signed));
      writeout(4, bld.vop2(aco_opcode::v_mul_f32, bld.def(v1), inputs[0], bfe_word0_b));

      //~gfx[^7].*! @standard_test(5,word1)
      Temp bfe_word1_b = bld.pseudo(ext, bld.def(v1), inputs[1], Operand::c32(1u),
                                    Operand::c32(16u), Operand::c32(is_signed));
      writeout(5, bld.vop2(aco_opcode::v_mul_f32, bld.def(v1), inputs[0], bfe_word1_b));

      //~gfx[^7]_unsigned! @standard_test(6,byte0)
      Temp bfi_byte0_b = bld.pseudo(ins, bld.def(v1), inputs[1], Operand::zero(), Operand::c32(8u));
      writeout(6, bld.vop2(aco_opcode::v_mul_f32, bld.def(v1), inputs[0], bfi_byte0_b));

      //~gfx[^7]_unsigned! @standard_test(7,word0)
      Temp bfi_word0_b =
         bld.pseudo(ins, bld.def(v1), inputs[1], Operand::zero(), Operand::c32(16u));
      writeout(7, bld.vop2(aco_opcode::v_mul_f32, bld.def(v1), inputs[0], bfi_word0_b));
      }

      //>> p_unit_test 63
      writeout(63);

      {
      //! v1: %tmp8 = p_insert %b, 1, 8
      //! v1: %res8 = v_mul_f32 %a, %tmp8
      //! p_unit_test 8, %res8
      Temp bfi_byte1_b =
         bld.pseudo(ins, bld.def(v1), inputs[1], Operand::c32(1u), Operand::c32(8u));
      writeout(8, bld.vop2(aco_opcode::v_mul_f32, bld.def(v1), inputs[0], bfi_byte1_b));

      /* v_cvt_f32_ubyte[0-3] can be used instead of v_cvt_f32_u32+sdwa */
      //~gfx7_signed! v1: %bfe_byte0_b = p_extract %b, 0, 8, 1
      //~gfx7_signed! v1: %res9 = v_cvt_f32_u32 %bfe_byte0_b
      //~gfx[^7]+_signed! v1: %res9 = v_cvt_f32_u32 %b dst_sel:dword src0_sel:sbyte0
      //~gfx\d+_unsigned! v1: %res9 = v_cvt_f32_ubyte0 %b
      //! p_unit_test 9, %res9
      Temp bfe_byte0_b = bld.pseudo(ext, bld.def(v1), inputs[1], Operand::zero(), Operand::c32(8u),
                                    Operand::c32(is_signed));
      writeout(9, bld.vop1(aco_opcode::v_cvt_f32_u32, bld.def(v1), bfe_byte0_b));

      //~gfx7_signed! v1: %bfe_byte1_b = p_extract %b, 1, 8, 1
      //~gfx7_signed! v1: %res10 = v_cvt_f32_u32 %bfe_byte1_b
      //~gfx[^7]+_signed! v1: %res10 = v_cvt_f32_u32 %b dst_sel:dword src0_sel:sbyte1
      //~gfx\d+_unsigned! v1: %res10 = v_cvt_f32_ubyte1 %b
      //! p_unit_test 10, %res10
      Temp bfe_byte1_b = bld.pseudo(ext, bld.def(v1), inputs[1], Operand::c32(1u), Operand::c32(8u),
                                    Operand::c32(is_signed));
      writeout(10, bld.vop1(aco_opcode::v_cvt_f32_u32, bld.def(v1), bfe_byte1_b));

      //~gfx7_signed! v1: %bfe_byte2_b = p_extract %b, 2, 8, 1
      //~gfx7_signed! v1: %res11 = v_cvt_f32_u32 %bfe_byte2_b
      //~gfx[^7]+_signed! v1: %res11 = v_cvt_f32_u32 %b dst_sel:dword src0_sel:sbyte2
      //~gfx\d+_unsigned! v1: %res11 = v_cvt_f32_ubyte2 %b
      //! p_unit_test 11, %res11
      Temp bfe_byte2_b = bld.pseudo(ext, bld.def(v1), inputs[1], Operand::c32(2u), Operand::c32(8u),
                                    Operand::c32(is_signed));
      writeout(11, bld.vop1(aco_opcode::v_cvt_f32_u32, bld.def(v1), bfe_byte2_b));

      //~gfx7_signed! v1: %bfe_byte3_b = p_extract %b, 3, 8, 1
      //~gfx7_signed! v1: %res12 = v_cvt_f32_u32 %bfe_byte3_b
      //~gfx[^7]+_signed! v1: %res12 = v_cvt_f32_u32 %b dst_sel:dword src0_sel:sbyte3
      //~gfx\d+_unsigned! v1: %res12 = v_cvt_f32_ubyte3 %b
      //! p_unit_test 12, %res12
      Temp bfe_byte3_b = bld.pseudo(ext, bld.def(v1), inputs[1], Operand::c32(3u), Operand::c32(8u),
                                    Operand::c32(is_signed));
      writeout(12, bld.vop1(aco_opcode::v_cvt_f32_u32, bld.def(v1), bfe_byte3_b));

      /* VOP3-only instructions can't use SDWA but they can use opsel on GFX9+ instead */
      //~gfx(9|10).*! v1: %res13 = v_add_i16 %a, %b
      //~gfx(9|10).*! p_unit_test 13, %res13
      Temp bfe_word0_b = bld.pseudo(ext, bld.def(v1), inputs[1], Operand::zero(), Operand::c32(16u),
                                    Operand::c32(is_signed));
      writeout(13, bld.vop3(aco_opcode::v_add_i16, bld.def(v1), inputs[0], bfe_word0_b));

      //~gfx(9|10).*! v1: %res14 = v_add_i16 %a, hi(%b)
      //~gfx(9|10).*! p_unit_test 14, %res14
      Temp bfe_word1_b = bld.pseudo(ext, bld.def(v1), inputs[1], Operand::c32(1u),
                                    Operand::c32(16u), Operand::c32(is_signed));
      writeout(14, bld.vop3(aco_opcode::v_add_i16, bld.def(v1), inputs[0], bfe_word1_b));
      }

      finish_opt_test();
   }
   }
END_TEST

BEGIN_TEST(optimize.sdwa.extract_modifiers)
   for (unsigned i = GFX8; i <= GFX10; i++) {
      //>> v1: %a, v1: %b, s1: %c, s1: %d = p_startpgm
      if (!setup_cs("v1 v1 s1 s1", (amd_gfx_level)i))
         continue;

      aco_opcode ext = aco_opcode::p_extract;

      //! v1: %res0 = v_mul_f32 %a, -%b dst_sel:dword src0_sel:dword src1_sel:ubyte0
      //! p_unit_test 0, %res0
      Temp byte0 = bld.pseudo(ext, bld.def(v1), inputs[1], Operand::zero(), Operand::c32(8u),
                              Operand::zero());
      Temp neg_byte0 = fneg(byte0);
      writeout(0, bld.vop2(aco_opcode::v_mul_f32, bld.def(v1), inputs[0], neg_byte0));

      //~gfx8! v1: %neg = v_mul_f32 -1.0, %b
      //~gfx8! v1: %res1 = v_mul_f32 %a, %neg dst_sel:dword src0_sel:dword src1_sel:ubyte0
      //~gfx(9|10)! v1: %neg_byte0 = v_mul_f32 -1.0, %b dst_sel:ubyte0 src0_sel:dword src1_sel:dword
      //~gfx(9|10)! v1: %res1 = v_mul_f32 %a, %neg_byte0
      //! p_unit_test 1, %res1
      Temp neg = fneg(inputs[1]);
      Temp byte0_neg =
         bld.pseudo(ext, bld.def(v1), neg, Operand::zero(), Operand::c32(8u), Operand::zero());
      writeout(1, bld.vop2(aco_opcode::v_mul_f32, bld.def(v1), inputs[0], byte0_neg));

      //! v1: %res2 = v_mul_f32 %a, |%b| dst_sel:dword src0_sel:dword src1_sel:ubyte0
      //! p_unit_test 2, %res2
      Temp abs_byte0 = fabs(byte0);
      writeout(2, bld.vop2(aco_opcode::v_mul_f32, bld.def(v1), inputs[0], abs_byte0));

      //! v1: %abs = v_mul_f32 1.0, |%b|
      //! v1: %res3 = v_mul_f32 %a, %abs dst_sel:dword src0_sel:dword src1_sel:ubyte0
      //! p_unit_test 3, %res3
      Temp abs = fabs(inputs[1]);
      Temp byte0_abs =
         bld.pseudo(ext, bld.def(v1), abs, Operand::zero(), Operand::c32(8u), Operand::zero());
      writeout(3, bld.vop2(aco_opcode::v_mul_f32, bld.def(v1), inputs[0], byte0_abs));

      //! v1: %res4 = v_mul_f32 %1, -|%2| dst_sel:dword src0_sel:dword src1_sel:ubyte0
      //! p_unit_test 4, %res4
      Temp neg_abs_byte0 = fneg(abs_byte0);
      writeout(4, bld.vop2(aco_opcode::v_mul_f32, bld.def(v1), inputs[0], neg_abs_byte0));

      //~gfx8! v1: %neg_abs = v_mul_f32 -1.0, %abs
      //~gfx8! v1: %res5 = v_mul_f32 %a, %neg_abs dst_sel:dword src0_sel:dword src1_sel:ubyte0
      //~gfx(9|10)! v1: %neg_abs_byte0 = v_mul_f32 -1.0, %abs dst_sel:ubyte0 src0_sel:dword src1_sel:dword
      //~gfx(9|10)! v1: %res5 = v_mul_f32 %a, %neg_abs_byte0
      //! p_unit_test 5, %res5
      Temp neg_abs = fneg(abs);
      Temp byte0_neg_abs =
         bld.pseudo(ext, bld.def(v1), neg_abs, Operand::zero(), Operand::c32(8u), Operand::zero());
      writeout(5, bld.vop2(aco_opcode::v_mul_f32, bld.def(v1), inputs[0], byte0_neg_abs));

      finish_opt_test();
   }
END_TEST

BEGIN_TEST(optimize.sdwa.extract.sgpr)
   for (unsigned i = GFX8; i <= GFX10; i++) {
      //>> v1: %a, v1: %b, s1: %c, s1: %d = p_startpgm
      if (!setup_cs("v1 v1 s1 s1", (amd_gfx_level)i))
         continue;

      aco_opcode ext = aco_opcode::p_extract;

      //~gfx8! v1: %byte0_b = p_extract %b, 0, 8, 0
      //~gfx8! v1: %res1 = v_mul_f32 %c, %byte0_b
      //~gfx(9|10)! v1: %res1 = v_mul_f32 %c, %b dst_sel:dword src0_sel:dword src1_sel:ubyte0
      //! p_unit_test 1, %res1
      Temp byte0_b = bld.pseudo(ext, bld.def(v1), inputs[1], Operand::zero(), Operand::c32(8u),
                                Operand::zero());
      writeout(1, bld.vop2(aco_opcode::v_mul_f32, bld.def(v1), inputs[2], byte0_b));

      //~gfx8! v1: %byte0_c = p_extract %c, 0, 8, 0
      //~gfx8! v1: %res2 = v_mul_f32 %a, %byte0_c
      //~gfx(9|10)! v1: %res2 = v_mul_f32 %a, %c dst_sel:dword src0_sel:dword src1_sel:ubyte0
      //! p_unit_test 2, %res2
      Temp byte0_c = bld.pseudo(ext, bld.def(v1), inputs[2], Operand::zero(), Operand::c32(8u),
                                Operand::zero());
      writeout(2, bld.vop2(aco_opcode::v_mul_f32, bld.def(v1), inputs[0], byte0_c));

      //~gfx8! v1: %byte0_c_2 = p_extract %c, 0, 8, 0
      //~gfx8! v1: %res3 = v_mul_f32 %c, %byte0_c_2
      //~gfx(9|10)! v1: %res3 = v_mul_f32 %c, %c dst_sel:dword src0_sel:dword src1_sel:ubyte0
      //! p_unit_test 3, %res3
      byte0_c = bld.pseudo(ext, bld.def(v1), inputs[2], Operand::zero(), Operand::c32(8u),
                           Operand::zero());
      writeout(3, bld.vop2(aco_opcode::v_mul_f32, bld.def(v1), inputs[2], byte0_c));

      //~gfx(8|9)! v1: %byte0_c_3 = p_extract %c, 0, 8, 0
      //~gfx(8|9)! v1: %res4 = v_mul_f32 %d, %byte0_c_3
      //~gfx10! v1: %res4 = v_mul_f32 %d, %c dst_sel:dword src0_sel:dword src1_sel:ubyte0
      //! p_unit_test 4, %res4
      byte0_c = bld.pseudo(ext, bld.def(v1), inputs[2], Operand::zero(), Operand::c32(8u),
                           Operand::zero());
      writeout(4, bld.vop2(aco_opcode::v_mul_f32, bld.def(v1), inputs[3], byte0_c));

      finish_opt_test();
   }
END_TEST

BEGIN_TEST(optimize.sdwa.from_vop3)
   for (unsigned i = GFX8; i <= GFX10; i++) {
      //>> v1: %a, v1: %b, s1: %c, s1: %d = p_startpgm
      if (!setup_cs("v1 v1 s1 s1", (amd_gfx_level)i))
         continue;

      //! v1: %res0 = v_mul_f32 -|%a|, %b dst_sel:dword src0_sel:dword src1_sel:ubyte0
      //! p_unit_test 0, %res0
      Temp byte0_b = bld.pseudo(aco_opcode::p_extract, bld.def(v1), inputs[1], Operand::zero(),
                                Operand::c32(8u), Operand::zero());
      VALU_instruction* mul =
         &bld.vop2_e64(aco_opcode::v_mul_f32, bld.def(v1), inputs[0], byte0_b)->valu();
      mul->neg[0] = true;
      mul->abs[0] = true;
      writeout(0, mul->definitions[0].getTemp());

      //~gfx8! v1: %byte0_b_0 = p_extract %b, 0, 8, 0
      //~gfx8! v1: %res1 = v_mul_f32 %a, %byte0_b_0 *4
      //~gfx(9|10)! v1: %res1 = v_mul_f32 %a, %b *4 dst_sel:dword src0_sel:dword src1_sel:ubyte0
      //! p_unit_test 1, %res1
      byte0_b = bld.pseudo(aco_opcode::p_extract, bld.def(v1), inputs[1], Operand::zero(),
                           Operand::c32(8u), Operand::zero());
      mul = &bld.vop2_e64(aco_opcode::v_mul_f32, bld.def(v1), inputs[0], byte0_b)->valu();
      mul->omod = 2;
      writeout(1, mul->definitions[0].getTemp());

      //~gfx8! v1: %byte0_b_1 = p_extract %b, 0, 8, 0
      //~gfx8! v1: %res2 = v_mul_f32 %byte0_b_1, %c
      //~gfx(9|10)! v1: %res2 = v_mul_f32 %b, %c dst_sel:dword src0_sel:ubyte0 src1_sel:dword
      //! p_unit_test 2, %res2
      byte0_b = bld.pseudo(aco_opcode::p_extract, bld.def(v1), inputs[1], Operand::zero(),
                           Operand::c32(8u), Operand::zero());
      writeout(2, bld.vop2_e64(aco_opcode::v_mul_f32, bld.def(v1), byte0_b, inputs[2]));

      if (i >= GFX10) {
         //~gfx10! v1: %byte0_b_2 = p_extract %b, 0, 8, 0
         //~gfx10! v1: %res3 = v_mul_f32 %byte0_b_2, 0x1234
         //~gfx10! p_unit_test 3, %res3
         byte0_b = bld.pseudo(aco_opcode::p_extract, bld.def(v1), inputs[1], Operand::zero(),
                              Operand::c32(8u), Operand::zero());
         writeout(3,
                  bld.vop2_e64(aco_opcode::v_mul_f32, bld.def(v1), byte0_b, Operand::c32(0x1234u)));
      }

      finish_opt_test();
   }
END_TEST

BEGIN_TEST(optimize.sdwa.insert)
   for (unsigned i = GFX7; i <= GFX10; i++) {
      //>> v1: %a, v1: %b = p_startpgm
      if (!setup_cs("v1 v1", (amd_gfx_level)i))
         continue;

      aco_opcode ext = aco_opcode::p_extract;
      aco_opcode ins = aco_opcode::p_insert;

      //~gfx[^7]! v1: %res0 = v_mul_f32 %a, %b dst_sel:ubyte0 src0_sel:dword src1_sel:dword
      //~gfx[^7]! p_unit_test 0, %res0
      Temp val = bld.vop2(aco_opcode::v_mul_f32, bld.def(v1), inputs[0], inputs[1]);
      writeout(0, bld.pseudo(ins, bld.def(v1), val, Operand::zero(), Operand::c32(8u)));

      //~gfx[^7]! v1: %res1 = v_mul_f32 %a, %b dst_sel:ubyte1 src0_sel:dword src1_sel:dword
      //~gfx[^7]! p_unit_test 1, %res1
      val = bld.vop2(aco_opcode::v_mul_f32, bld.def(v1), inputs[0], inputs[1]);
      writeout(1, bld.pseudo(ins, bld.def(v1), val, Operand::c32(1u), Operand::c32(8u)));

      //~gfx[^7]! v1: %res2 = v_mul_f32 %a, %b dst_sel:ubyte2 src0_sel:dword src1_sel:dword
      //~gfx[^7]! p_unit_test 2, %res2
      val = bld.vop2(aco_opcode::v_mul_f32, bld.def(v1), inputs[0], inputs[1]);
      writeout(2, bld.pseudo(ins, bld.def(v1), val, Operand::c32(2u), Operand::c32(8u)));

      //~gfx[^7]! v1: %res3 = v_mul_f32 %a, %b dst_sel:ubyte3 src0_sel:dword src1_sel:dword
      //~gfx[^7]! p_unit_test 3, %res3
      val = bld.vop2(aco_opcode::v_mul_f32, bld.def(v1), inputs[0], inputs[1]);
      writeout(3, bld.pseudo(ins, bld.def(v1), val, Operand::c32(3u), Operand::c32(8u)));

      //~gfx[^7]! v1: %res4 = v_mul_f32 %a, %b dst_sel:uword0 src0_sel:dword src1_sel:dword
      //~gfx[^7]! p_unit_test 4, %res4
      val = bld.vop2(aco_opcode::v_mul_f32, bld.def(v1), inputs[0], inputs[1]);
      writeout(4, bld.pseudo(ins, bld.def(v1), val, Operand::zero(), Operand::c32(16u)));

      //~gfx[^7]! v1: %res5 = v_mul_f32 %a, %b dst_sel:uword1 src0_sel:dword src1_sel:dword
      //~gfx[^7]! p_unit_test 5, %res5
      val = bld.vop2(aco_opcode::v_mul_f32, bld.def(v1), inputs[0], inputs[1]);
      writeout(5, bld.pseudo(ins, bld.def(v1), val, Operand::c32(1u), Operand::c32(16u)));

      //~gfx[^7]! v1: %res6 = v_mul_f32 %a, %b dst_sel:ubyte0 src0_sel:dword src1_sel:dword
      //~gfx[^7]! p_unit_test 6, %res6
      val = bld.vop2(aco_opcode::v_mul_f32, bld.def(v1), inputs[0], inputs[1]);
      writeout(
         6, bld.pseudo(ext, bld.def(v1), val, Operand::zero(), Operand::c32(8u), Operand::zero()));

      //~gfx[^7]! v1: %res7 = v_mul_f32 %a, %b dst_sel:uword0 src0_sel:dword src1_sel:dword
      //~gfx[^7]! p_unit_test 7, %res7
      val = bld.vop2(aco_opcode::v_mul_f32, bld.def(v1), inputs[0], inputs[1]);
      writeout(
         7, bld.pseudo(ext, bld.def(v1), val, Operand::zero(), Operand::c32(16u), Operand::zero()));

      //~gfx[^7]! v1: %tmp8 = v_mul_f32 %a, %b
      //~gfx[^7]! v1: %res8 = p_extract %tmp8, 2, 8, 0
      //~gfx[^7]! p_unit_test 8, %res8
      val = bld.vop2(aco_opcode::v_mul_f32, bld.def(v1), inputs[0], inputs[1]);
      writeout(
         8, bld.pseudo(ext, bld.def(v1), val, Operand::c32(2u), Operand::c32(8u), Operand::zero()));

      //~gfx[^7]! v1: %tmp9 = v_mul_f32 %a, %b
      //~gfx[^7]! v1: %res9 = p_extract %tmp9, 0, 8, 1
      //~gfx[^7]! p_unit_test 9, %res9
      val = bld.vop2(aco_opcode::v_mul_f32, bld.def(v1), inputs[0], inputs[1]);
      writeout(
         9, bld.pseudo(ext, bld.def(v1), val, Operand::zero(), Operand::c32(8u), Operand::c32(1u)));

      //>> p_unit_test 63
      writeout(63);

      //! v1: %res10 = v_mul_f32 %a, %b
      //! p_unit_test 10, %res10
      val = bld.vop2(aco_opcode::v_mul_f32, bld.def(v1), inputs[0], inputs[1]);
      bld.pseudo(ins, bld.def(v1), val, Operand::c32(1u), Operand::c32(16u));
      writeout(10, val);

      //~gfx[^7]! v1: %tmp11 = v_sub_i16 %a, %b
      //~gfx[^7]! v1: %res11 = p_insert %tmp11, 0, 16
      //~gfx[^7]! p_unit_test 11, %res11
      val = bld.vop3(aco_opcode::v_sub_i16, bld.def(v1), inputs[0], inputs[1]);
      writeout(11, bld.pseudo(ins, bld.def(v1), val, Operand::zero(), Operand::c32(16u)));

      //~gfx[^7]! v1: %tmp12 = v_sub_i16 %a, %b
      //~gfx[^7]! v1: %res12 = p_insert %tmp12, 1, 16
      //~gfx[^7]! p_unit_test 12, %res12
      val = bld.vop3(aco_opcode::v_sub_i16, bld.def(v1), inputs[0], inputs[1]);
      writeout(12, bld.pseudo(ins, bld.def(v1), val, Operand::c32(1u), Operand::c32(16u)));

      //~gfx[^7]! v1: %tmp13 = v_sub_i16 %a, %b
      //~gfx[^7]! v1: %res13 = p_insert %tmp13, 0, 8
      //~gfx[^7]! p_unit_test 13, %res13
      val = bld.vop3(aco_opcode::v_sub_i16, bld.def(v1), inputs[0], inputs[1]);
      writeout(13, bld.pseudo(ins, bld.def(v1), val, Operand::zero(), Operand::c32(8u)));

      finish_opt_test();
   }
END_TEST

BEGIN_TEST(optimize.sdwa.insert_modifiers)
   for (unsigned i = GFX8; i <= GFX9; i++) {
      //>> v1: %a = p_startpgm
      if (!setup_cs("v1", (amd_gfx_level)i))
         continue;

      aco_opcode ins = aco_opcode::p_insert;

      //~gfx8! v1: %tmp0 = v_rcp_f32 %a *2
      //~gfx8! v1: %res0 = p_insert %tmp0, 0, 8
      //~gfx9! v1: %res0 = v_rcp_f32 %a *2 dst_sel:ubyte0 src0_sel:dword
      //! p_unit_test 0, %res0
      Temp val = bld.vop1(aco_opcode::v_rcp_f32, bld.def(v1), inputs[0]);
      val = bld.vop2_e64(aco_opcode::v_mul_f32, bld.def(v1), val, Operand::c32(0x40000000u));
      writeout(0, bld.pseudo(ins, bld.def(v1), val, Operand::zero(), Operand::c32(8u)));

      //! v1: %res1 = v_rcp_f32 %a clamp dst_sel:ubyte0 src0_sel:dword
      //! p_unit_test 1, %res1
      val = bld.vop1(aco_opcode::v_rcp_f32, bld.def(v1), inputs[0]);
      val = bld.vop3(aco_opcode::v_med3_f32, bld.def(v1), val, Operand::zero(),
                     Operand::c32(0x3f800000u));
      writeout(1, bld.pseudo(ins, bld.def(v1), val, Operand::zero(), Operand::c32(8u)));

      //! v1: %tmp2 = v_rcp_f32 %a dst_sel:ubyte0 src0_sel:dword
      //! v1: %res2 = v_mul_f32 %tmp2, 2.0
      //! p_unit_test 2, %res2
      val = bld.vop1(aco_opcode::v_rcp_f32, bld.def(v1), inputs[0]);
      val = bld.pseudo(ins, bld.def(v1), val, Operand::zero(), Operand::c32(8u));
      val = bld.vop2_e64(aco_opcode::v_mul_f32, bld.def(v1), val, Operand::c32(0x40000000u));
      writeout(2, val);

      //! v1: %tmp3 = v_rcp_f32 %a dst_sel:ubyte0 src0_sel:dword
      //! v1: %res3 = v_med3_f32 %tmp3, 0, 1.0
      //! p_unit_test 3, %res3
      val = bld.vop1(aco_opcode::v_rcp_f32, bld.def(v1), inputs[0]);
      val = bld.pseudo(ins, bld.def(v1), val, Operand::zero(), Operand::c32(8u));
      val = bld.vop3(aco_opcode::v_med3_f32, bld.def(v1), val, Operand::zero(),
                     Operand::c32(0x3f800000u));
      writeout(3, val);

      //~gfx8! v1: %tmp4 = v_rcp_f32 %a *2 clamp
      //~gfx8! v1: %res4 = p_insert %tmp4, 0, 8
      //~gfx9! v1: %res4 = v_rcp_f32 %a *2 clamp dst_sel:ubyte0 src0_sel:dword
      //! p_unit_test 4, %res4
      val = bld.vop1(aco_opcode::v_rcp_f32, bld.def(v1), inputs[0]);
      val = bld.vop2_e64(aco_opcode::v_mul_f32, bld.def(v1), val, Operand::c32(0x40000000u));
      val = bld.vop3(aco_opcode::v_med3_f32, bld.def(v1), val, Operand::zero(),
                     Operand::c32(0x3f800000u));
      writeout(4, bld.pseudo(ins, bld.def(v1), val, Operand::zero(), Operand::c32(8u)));

      finish_opt_test();
   }
END_TEST