diff options
Diffstat (limited to 'llvm/test/MC/AMDGPU/gfx12_asm_vop3_from_vop2_dpp8.s')
| -rw-r--r-- | llvm/test/MC/AMDGPU/gfx12_asm_vop3_from_vop2_dpp8.s | 241 |
1 files changed, 241 insertions, 0 deletions
diff --git a/llvm/test/MC/AMDGPU/gfx12_asm_vop3_from_vop2_dpp8.s b/llvm/test/MC/AMDGPU/gfx12_asm_vop3_from_vop2_dpp8.s index dc151d66..b18029d 100644 --- a/llvm/test/MC/AMDGPU/gfx12_asm_vop3_from_vop2_dpp8.s +++ b/llvm/test/MC/AMDGPU/gfx12_asm_vop3_from_vop2_dpp8.s @@ -45,6 +45,12 @@ v_add_co_ci_u32_e64_dpp v255, null, v255, v255, null clamp dpp8:[0,0,0,0,0,0,0,0 v_add_f16_e64_dpp v5, v1, v2 dpp8:[7,6,5,4,3,2,1,0] // GFX12: [0x05,0x00,0x32,0xd5,0xe9,0x04,0x02,0x00,0x01,0x77,0x39,0x05] +v_add_f16_e64_dpp v5, v1, s2 dpp8:[7,6,5,4,3,2,1,0] +// GFX12: [0x05,0x00,0x32,0xd5,0xe9,0x04,0x00,0x00,0x01,0x77,0x39,0x05] + +v_add_f16_e64_dpp v5, v1, 2.0 dpp8:[7,6,5,4,3,2,1,0] +// GFX12: [0x05,0x00,0x32,0xd5,0xe9,0xe8,0x01,0x00,0x01,0x77,0x39,0x05] + v_add_f16_e64_dpp v5, |v1|, -v2 mul:2 dpp8:[7,6,5,4,3,2,1,0] // GFX12: [0x05,0x01,0x32,0xd5,0xe9,0x04,0x02,0x48,0x01,0x77,0x39,0x05] @@ -57,6 +63,12 @@ v_add_f16_e64_dpp v255, -|v255|, -|v255| clamp div:2 dpp8:[0,0,0,0,0,0,0,0] fi:0 v_add_f32_e64_dpp v5, v1, v2 dpp8:[7,6,5,4,3,2,1,0] // GFX12: [0x05,0x00,0x03,0xd5,0xe9,0x04,0x02,0x00,0x01,0x77,0x39,0x05] +v_add_f32_e64_dpp v5, v1, s2 dpp8:[7,6,5,4,3,2,1,0] +// GFX12: [0x05,0x00,0x03,0xd5,0xe9,0x04,0x00,0x00,0x01,0x77,0x39,0x05] + +v_add_f32_e64_dpp v5, v1, 2.0 dpp8:[7,6,5,4,3,2,1,0] +// GFX12: [0x05,0x00,0x03,0xd5,0xe9,0xe8,0x01,0x00,0x01,0x77,0x39,0x05] + v_add_f32_e64_dpp v5, |v1|, -v2 mul:2 dpp8:[7,6,5,4,3,2,1,0] // GFX12: [0x05,0x01,0x03,0xd5,0xe9,0x04,0x02,0x48,0x01,0x77,0x39,0x05] @@ -69,6 +81,12 @@ v_add_f32_e64_dpp v255, -|v255|, -|v255| clamp div:2 dpp8:[0,0,0,0,0,0,0,0] fi:0 v_add_nc_u32_e64_dpp v5, v1, v2 dpp8:[7,6,5,4,3,2,1,0] // GFX12: [0x05,0x00,0x25,0xd5,0xe9,0x04,0x02,0x00,0x01,0x77,0x39,0x05] +v_add_nc_u32_e64_dpp v5, v1, s2 dpp8:[7,6,5,4,3,2,1,0] +// GFX12: [0x05,0x00,0x25,0xd5,0xe9,0x04,0x00,0x00,0x01,0x77,0x39,0x05] + +v_add_nc_u32_e64_dpp v5, v1, 10 dpp8:[7,6,5,4,3,2,1,0] +// GFX12: [0x05,0x00,0x25,0xd5,0xe9,0x14,0x01,0x00,0x01,0x77,0x39,0x05] + v_add_nc_u32_e64_dpp v5, v1, v2 dpp8:[7,6,5,4,3,2,1,0] fi:1 // GFX12: [0x05,0x00,0x25,0xd5,0xea,0x04,0x02,0x00,0x01,0x77,0x39,0x05] @@ -78,6 +96,12 @@ v_add_nc_u32_e64_dpp v255, v255, v255 clamp dpp8:[0,0,0,0,0,0,0,0] fi:0 v_and_b32_e64_dpp v5, v1, v2 dpp8:[7,6,5,4,3,2,1,0] // GFX12: [0x05,0x00,0x1b,0xd5,0xe9,0x04,0x02,0x00,0x01,0x77,0x39,0x05] +v_and_b32_e64_dpp v5, v1, s2 dpp8:[7,6,5,4,3,2,1,0] +// GFX12: [0x05,0x00,0x1b,0xd5,0xe9,0x04,0x00,0x00,0x01,0x77,0x39,0x05] + +v_and_b32_e64_dpp v5, v1, 10 dpp8:[7,6,5,4,3,2,1,0] +// GFX12: [0x05,0x00,0x1b,0xd5,0xe9,0x14,0x01,0x00,0x01,0x77,0x39,0x05] + v_and_b32_e64_dpp v5, v1, v2 dpp8:[7,6,5,4,3,2,1,0] fi:1 // GFX12: [0x05,0x00,0x1b,0xd5,0xea,0x04,0x02,0x00,0x01,0x77,0x39,0x05] @@ -87,6 +111,12 @@ v_and_b32_e64_dpp v255, v255, v255 dpp8:[0,0,0,0,0,0,0,0] fi:0 v_ashrrev_i32_e64_dpp v5, v1, v2 dpp8:[7,6,5,4,3,2,1,0] // GFX12: [0x05,0x00,0x1a,0xd5,0xe9,0x04,0x02,0x00,0x01,0x77,0x39,0x05] +v_ashrrev_i32_e64_dpp v5, v1, s2 dpp8:[7,6,5,4,3,2,1,0] +// GFX12: [0x05,0x00,0x1a,0xd5,0xe9,0x04,0x00,0x00,0x01,0x77,0x39,0x05] + +v_ashrrev_i32_e64_dpp v5, v1, 10 dpp8:[7,6,5,4,3,2,1,0] +// GFX12: [0x05,0x00,0x1a,0xd5,0xe9,0x14,0x01,0x00,0x01,0x77,0x39,0x05] + v_ashrrev_i32_e64_dpp v5, v1, v2 dpp8:[7,6,5,4,3,2,1,0] fi:1 // GFX12: [0x05,0x00,0x1a,0xd5,0xea,0x04,0x02,0x00,0x01,0x77,0x39,0x05] @@ -97,14 +127,30 @@ v_cndmask_b32_e64_dpp v5, v1, v2, s3 dpp8:[7,6,5,4,3,2,1,0] // W32: [0x05,0x00,0x01,0xd5,0xe9,0x04,0x0e,0x00,0x01,0x77,0x39,0x05] // W64-ERR: :[[@LINE-2]]:{{[0-9]+}}: error: invalid operand for instruction +v_cndmask_b32_e64_dpp v5, v1, s2, s3 dpp8:[7,6,5,4,3,2,1,0] +// W32: [0x05,0x00,0x01,0xd5,0xe9,0x04,0x0c,0x00,0x01,0x77,0x39,0x05] +// W64-ERR: :[[@LINE-2]]:{{[0-9]+}}: error: invalid operand for instruction + +v_cndmask_b32_e64_dpp v5, v1, 10, s3 dpp8:[7,6,5,4,3,2,1,0] +// W32: [0x05,0x00,0x01,0xd5,0xe9,0x14,0x0d,0x00,0x01,0x77,0x39,0x05] +// W64-ERR: :[[@LINE-2]]:{{[0-9]+}}: error: invalid operand for instruction + v_cndmask_b32_e64_dpp v5, v1, v2, s105 dpp8:[7,6,5,4,3,2,1,0] // W32: [0x05,0x00,0x01,0xd5,0xe9,0x04,0xa6,0x01,0x01,0x77,0x39,0x05] // W64-ERR: :[[@LINE-2]]:{{[0-9]+}}: error: invalid operand for instruction +v_cndmask_b32_e64_dpp v5, v1, s2, s105 dpp8:[7,6,5,4,3,2,1,0] +// W32: [0x05,0x00,0x01,0xd5,0xe9,0x04,0xa4,0x01,0x01,0x77,0x39,0x05] +// W64-ERR: :[[@LINE-2]]:{{[0-9]+}}: error: invalid operand for instruction + v_cndmask_b32_e64_dpp v5, v1, v2, vcc_hi dpp8:[7,6,5,4,3,2,1,0] // W32: [0x05,0x00,0x01,0xd5,0xe9,0x04,0xae,0x01,0x01,0x77,0x39,0x05] // W64-ERR: :[[@LINE-2]]:{{[0-9]+}}: error: invalid operand for instruction +v_cndmask_b32_e64_dpp v5, v1, s2, vcc_hi dpp8:[7,6,5,4,3,2,1,0] +// W32: [0x05,0x00,0x01,0xd5,0xe9,0x04,0xac,0x01,0x01,0x77,0x39,0x05] +// W64-ERR: :[[@LINE-2]]:{{[0-9]+}}: error: invalid operand for instruction + v_cndmask_b32_e64_dpp v5, |v1|, -v2, vcc_lo dpp8:[7,6,5,4,3,2,1,0] // W32: [0x05,0x01,0x01,0xd5,0xe9,0x04,0xaa,0x41,0x01,0x77,0x39,0x05] // W64-ERR: :[[@LINE-2]]:{{[0-9]+}}: error: invalid operand for instruction @@ -117,10 +163,22 @@ v_cndmask_b32_e64_dpp v5, v1, v2, s[6:7] dpp8:[7,6,5,4,3,2,1,0] // W64: [0x05,0x00,0x01,0xd5,0xe9,0x04,0x1a,0x00,0x01,0x77,0x39,0x05] // W32-ERR: :[[@LINE-2]]:{{[0-9]+}}: error: invalid operand for instruction +v_cndmask_b32_e64_dpp v5, v1, s2, s[6:7] dpp8:[7,6,5,4,3,2,1,0] +// W64: [0x05,0x00,0x01,0xd5,0xe9,0x04,0x18,0x00,0x01,0x77,0x39,0x05] +// W32-ERR: :[[@LINE-2]]:{{[0-9]+}}: error: invalid operand for instruction + v_cndmask_b32_e64_dpp v5, v1, v2, s[104:105] dpp8:[7,6,5,4,3,2,1,0] // W64: [0x05,0x00,0x01,0xd5,0xe9,0x04,0xa2,0x01,0x01,0x77,0x39,0x05] // W32-ERR: :[[@LINE-2]]:{{[0-9]+}}: error: invalid operand for instruction +v_cndmask_b32_e64_dpp v5, v1, s2, s[104:105] dpp8:[7,6,5,4,3,2,1,0] +// W64: [0x05,0x00,0x01,0xd5,0xe9,0x04,0xa0,0x01,0x01,0x77,0x39,0x05] +// W32-ERR: :[[@LINE-2]]:{{[0-9]+}}: error: invalid operand for instruction + +v_cndmask_b32_e64_dpp v5, v1, 10, s[104:105] dpp8:[7,6,5,4,3,2,1,0] +// W64: [0x05,0x00,0x01,0xd5,0xe9,0x14,0xa1,0x01,0x01,0x77,0x39,0x05] +// W32-ERR: :[[@LINE-2]]:{{[0-9]+}}: error: invalid operand for instruction + v_cndmask_b32_e64_dpp v5, |v1|, -v2, vcc dpp8:[7,6,5,4,3,2,1,0] // W64: [0x05,0x01,0x01,0xd5,0xe9,0x04,0xaa,0x41,0x01,0x77,0x39,0x05] // W32-ERR: :[[@LINE-2]]:{{[0-9]+}}: error: invalid operand for instruction @@ -135,6 +193,12 @@ v_cndmask_b32_e64_dpp v255, -|v255|, -|v255|, null dpp8:[0,0,0,0,0,0,0,0] fi:0 v_cvt_pk_rtz_f16_f32_e64_dpp v5, v1, v2 dpp8:[7,6,5,4,3,2,1,0] // GFX12: [0x05,0x00,0x2f,0xd5,0xe9,0x04,0x02,0x00,0x01,0x77,0x39,0x05] +v_cvt_pk_rtz_f16_f32_e64_dpp v5, v1, s2 dpp8:[7,6,5,4,3,2,1,0] +// GFX12: [0x05,0x00,0x2f,0xd5,0xe9,0x04,0x00,0x00,0x01,0x77,0x39,0x05] + +v_cvt_pk_rtz_f16_f32_e64_dpp v5, v1, 2.0 dpp8:[7,6,5,4,3,2,1,0] +// GFX12: [0x05,0x00,0x2f,0xd5,0xe9,0xe8,0x01,0x00,0x01,0x77,0x39,0x05] + v_cvt_pk_rtz_f16_f32_e64_dpp v5, |v1|, -v2 dpp8:[7,6,5,4,3,2,1,0] // GFX12: [0x05,0x01,0x2f,0xd5,0xe9,0x04,0x02,0x40,0x01,0x77,0x39,0x05] @@ -147,6 +211,12 @@ v_cvt_pk_rtz_f16_f32_e64_dpp v255, -|v255|, -|v255| clamp dpp8:[0,0,0,0,0,0,0,0] v_cvt_pkrtz_f16_f32_e64_dpp v5, v1, v2 dpp8:[7,6,5,4,3,2,1,0] // GFX12: [0x05,0x00,0x2f,0xd5,0xe9,0x04,0x02,0x00,0x01,0x77,0x39,0x05] +v_cvt_pkrtz_f16_f32_e64_dpp v5, v1, s2 dpp8:[7,6,5,4,3,2,1,0] +// GFX12: [0x05,0x00,0x2f,0xd5,0xe9,0x04,0x00,0x00,0x01,0x77,0x39,0x05] + +v_cvt_pkrtz_f16_f32_e64_dpp v5, v1, 2.0 dpp8:[7,6,5,4,3,2,1,0] +// GFX12: [0x05,0x00,0x2f,0xd5,0xe9,0xe8,0x01,0x00,0x01,0x77,0x39,0x05] + v_cvt_pkrtz_f16_f32_e64_dpp v5, |v1|, -v2 dpp8:[7,6,5,4,3,2,1,0] // GFX12: [0x05,0x01,0x2f,0xd5,0xe9,0x04,0x02,0x40,0x01,0x77,0x39,0x05] @@ -159,9 +229,18 @@ v_cvt_pkrtz_f16_f32_e64_dpp v255, -|v255|, -|v255| clamp dpp8:[0,0,0,0,0,0,0,0] v_ldexp_f16_e64_dpp v5, v1, v2 dpp8:[7,6,5,4,3,2,1,0] // GFX12: [0x05,0x00,0x3b,0xd5,0xe9,0x04,0x02,0x00,0x01,0x77,0x39,0x05] +v_ldexp_f16_e64_dpp v5, v1, s2 dpp8:[7,6,5,4,3,2,1,0] +// GFX12: [0x05,0x00,0x3b,0xd5,0xe9,0x04,0x00,0x00,0x01,0x77,0x39,0x05] + v_ldexp_f16_e64_dpp v5, v1, v2 mul:2 dpp8:[7,6,5,4,3,2,1,0] // GFX12: [0x05,0x00,0x3b,0xd5,0xe9,0x04,0x02,0x08,0x01,0x77,0x39,0x05] +v_ldexp_f16_e64_dpp v5, v1, s2 mul:2 dpp8:[7,6,5,4,3,2,1,0] +// GFX12: [0x05,0x00,0x3b,0xd5,0xe9,0x04,0x00,0x08,0x01,0x77,0x39,0x05] + +v_ldexp_f16_e64_dpp v5, v1, 2.0 mul:2 dpp8:[7,6,5,4,3,2,1,0] +// GFX12: [0x05,0x00,0x3b,0xd5,0xe9,0xe8,0x01,0x08,0x01,0x77,0x39,0x05] + v_ldexp_f16_e64_dpp v5, v1, v2 mul:4 dpp8:[7,6,5,4,3,2,1,0] fi:1 // GFX12: [0x05,0x00,0x3b,0xd5,0xea,0x04,0x02,0x10,0x01,0x77,0x39,0x05] @@ -171,6 +250,12 @@ v_ldexp_f16_e64_dpp v255, -|v255|, v255 clamp div:2 dpp8:[0,0,0,0,0,0,0,0] fi:0 v_lshlrev_b32_e64_dpp v5, v1, v2 dpp8:[7,6,5,4,3,2,1,0] // GFX12: [0x05,0x00,0x18,0xd5,0xe9,0x04,0x02,0x00,0x01,0x77,0x39,0x05] +v_lshlrev_b32_e64_dpp v5, v1, s2 dpp8:[7,6,5,4,3,2,1,0] +// GFX12: [0x05,0x00,0x18,0xd5,0xe9,0x04,0x00,0x00,0x01,0x77,0x39,0x05] + +v_lshlrev_b32_e64_dpp v5, v1, 10 dpp8:[7,6,5,4,3,2,1,0] +// GFX12: [0x05,0x00,0x18,0xd5,0xe9,0x14,0x01,0x00,0x01,0x77,0x39,0x05] + v_lshlrev_b32_e64_dpp v5, v1, v2 dpp8:[7,6,5,4,3,2,1,0] fi:1 // GFX12: [0x05,0x00,0x18,0xd5,0xea,0x04,0x02,0x00,0x01,0x77,0x39,0x05] @@ -180,6 +265,12 @@ v_lshlrev_b32_e64_dpp v255, v255, v255 dpp8:[0,0,0,0,0,0,0,0] fi:0 v_lshrrev_b32_e64_dpp v5, v1, v2 dpp8:[7,6,5,4,3,2,1,0] // GFX12: [0x05,0x00,0x19,0xd5,0xe9,0x04,0x02,0x00,0x01,0x77,0x39,0x05] +v_lshrrev_b32_e64_dpp v5, v1, s2 dpp8:[7,6,5,4,3,2,1,0] +// GFX12: [0x05,0x00,0x19,0xd5,0xe9,0x04,0x00,0x00,0x01,0x77,0x39,0x05] + +v_lshrrev_b32_e64_dpp v5, v1, 10 dpp8:[7,6,5,4,3,2,1,0] +// GFX12: [0x05,0x00,0x19,0xd5,0xe9,0x14,0x01,0x00,0x01,0x77,0x39,0x05] + v_lshrrev_b32_e64_dpp v5, v1, v2 dpp8:[7,6,5,4,3,2,1,0] fi:1 // GFX12: [0x05,0x00,0x19,0xd5,0xea,0x04,0x02,0x00,0x01,0x77,0x39,0x05] @@ -189,6 +280,12 @@ v_lshrrev_b32_e64_dpp v255, v255, v255 dpp8:[0,0,0,0,0,0,0,0] fi:0 v_max_num_f16_e64_dpp v5, v1, v2 dpp8:[7,6,5,4,3,2,1,0] // GFX12: [0x05,0x00,0x31,0xd5,0xe9,0x04,0x02,0x00,0x01,0x77,0x39,0x05] +v_max_num_f16_e64_dpp v5, v1, s2 dpp8:[7,6,5,4,3,2,1,0] +// GFX12: [0x05,0x00,0x31,0xd5,0xe9,0x04,0x00,0x00,0x01,0x77,0x39,0x05] + +v_max_num_f16_e64_dpp v5, v1, 2.0 dpp8:[7,6,5,4,3,2,1,0] +// GFX12: [0x05,0x00,0x31,0xd5,0xe9,0xe8,0x01,0x00,0x01,0x77,0x39,0x05] + v_max_num_f16_e64_dpp v5, |v1|, -v2 mul:2 dpp8:[7,6,5,4,3,2,1,0] // GFX12: [0x05,0x01,0x31,0xd5,0xe9,0x04,0x02,0x48,0x01,0x77,0x39,0x05] @@ -201,6 +298,12 @@ v_max_num_f16_e64_dpp v255, -|v255|, -|v255| clamp div:2 dpp8:[0,0,0,0,0,0,0,0] v_max_num_f32_e64_dpp v5, v1, v2 dpp8:[7,6,5,4,3,2,1,0] // GFX12: [0x05,0x00,0x16,0xd5,0xe9,0x04,0x02,0x00,0x01,0x77,0x39,0x05] +v_max_num_f32_e64_dpp v5, v1, s2 dpp8:[7,6,5,4,3,2,1,0] +// GFX12: [0x05,0x00,0x16,0xd5,0xe9,0x04,0x00,0x00,0x01,0x77,0x39,0x05] + +v_max_num_f32_e64_dpp v5, v1, 2.0 dpp8:[7,6,5,4,3,2,1,0] +// GFX12: [0x05,0x00,0x16,0xd5,0xe9,0xe8,0x01,0x00,0x01,0x77,0x39,0x05] + v_max_num_f32_e64_dpp v5, |v1|, -v2 mul:2 dpp8:[7,6,5,4,3,2,1,0] // GFX12: [0x05,0x01,0x16,0xd5,0xe9,0x04,0x02,0x48,0x01,0x77,0x39,0x05] @@ -213,6 +316,12 @@ v_max_num_f32_e64_dpp v255, -|v255|, -|v255| clamp div:2 dpp8:[0,0,0,0,0,0,0,0] v_max_i32_e64_dpp v5, v1, v2 dpp8:[7,6,5,4,3,2,1,0] // GFX12: [0x05,0x00,0x12,0xd5,0xe9,0x04,0x02,0x00,0x01,0x77,0x39,0x05] +v_max_i32_e64_dpp v5, v1, s2 dpp8:[7,6,5,4,3,2,1,0] +// GFX12: [0x05,0x00,0x12,0xd5,0xe9,0x04,0x00,0x00,0x01,0x77,0x39,0x05] + +v_max_i32_e64_dpp v5, v1, 10 dpp8:[7,6,5,4,3,2,1,0] +// GFX12: [0x05,0x00,0x12,0xd5,0xe9,0x14,0x01,0x00,0x01,0x77,0x39,0x05] + v_max_i32_e64_dpp v5, v1, v2 dpp8:[7,6,5,4,3,2,1,0] fi:1 // GFX12: [0x05,0x00,0x12,0xd5,0xea,0x04,0x02,0x00,0x01,0x77,0x39,0x05] @@ -222,6 +331,12 @@ v_max_i32_e64_dpp v255, v255, v255 dpp8:[0,0,0,0,0,0,0,0] fi:0 v_max_u32_e64_dpp v5, v1, v2 dpp8:[7,6,5,4,3,2,1,0] // GFX12: [0x05,0x00,0x14,0xd5,0xe9,0x04,0x02,0x00,0x01,0x77,0x39,0x05] +v_max_u32_e64_dpp v5, v1, s2 dpp8:[7,6,5,4,3,2,1,0] +// GFX12: [0x05,0x00,0x14,0xd5,0xe9,0x04,0x00,0x00,0x01,0x77,0x39,0x05] + +v_max_u32_e64_dpp v5, v1, 10 dpp8:[7,6,5,4,3,2,1,0] +// GFX12: [0x05,0x00,0x14,0xd5,0xe9,0x14,0x01,0x00,0x01,0x77,0x39,0x05] + v_max_u32_e64_dpp v5, v1, v2 dpp8:[7,6,5,4,3,2,1,0] fi:1 // GFX12: [0x05,0x00,0x14,0xd5,0xea,0x04,0x02,0x00,0x01,0x77,0x39,0x05] @@ -231,6 +346,12 @@ v_max_u32_e64_dpp v255, v255, v255 dpp8:[0,0,0,0,0,0,0,0] fi:0 v_min_num_f16_e64_dpp v5, v1, v2 dpp8:[7,6,5,4,3,2,1,0] // GFX12: [0x05,0x00,0x30,0xd5,0xe9,0x04,0x02,0x00,0x01,0x77,0x39,0x05] +v_min_num_f16_e64_dpp v5, v1, s2 dpp8:[7,6,5,4,3,2,1,0] +// GFX12: [0x05,0x00,0x30,0xd5,0xe9,0x04,0x00,0x00,0x01,0x77,0x39,0x05] + +v_min_num_f16_e64_dpp v5, v1, 2.0 dpp8:[7,6,5,4,3,2,1,0] +// GFX12: [0x05,0x00,0x30,0xd5,0xe9,0xe8,0x01,0x00,0x01,0x77,0x39,0x05] + v_min_num_f16_e64_dpp v5, |v1|, -v2 mul:2 dpp8:[7,6,5,4,3,2,1,0] // GFX12: [0x05,0x01,0x30,0xd5,0xe9,0x04,0x02,0x48,0x01,0x77,0x39,0x05] @@ -243,6 +364,12 @@ v_min_num_f16_e64_dpp v255, -|v255|, -|v255| clamp div:2 dpp8:[0,0,0,0,0,0,0,0] v_min_num_f32_e64_dpp v5, v1, v2 dpp8:[7,6,5,4,3,2,1,0] // GFX12: [0x05,0x00,0x15,0xd5,0xe9,0x04,0x02,0x00,0x01,0x77,0x39,0x05] +v_min_num_f32_e64_dpp v5, v1, s2 dpp8:[7,6,5,4,3,2,1,0] +// GFX12: [0x05,0x00,0x15,0xd5,0xe9,0x04,0x00,0x00,0x01,0x77,0x39,0x05] + +v_min_num_f32_e64_dpp v5, v1, 2.0 dpp8:[7,6,5,4,3,2,1,0] +// GFX12: [0x05,0x00,0x15,0xd5,0xe9,0xe8,0x01,0x00,0x01,0x77,0x39,0x05] + v_min_num_f32_e64_dpp v5, |v1|, -v2 mul:2 dpp8:[7,6,5,4,3,2,1,0] // GFX12: [0x05,0x01,0x15,0xd5,0xe9,0x04,0x02,0x48,0x01,0x77,0x39,0x05] @@ -255,6 +382,12 @@ v_min_num_f32_e64_dpp v255, -|v255|, -|v255| clamp div:2 dpp8:[0,0,0,0,0,0,0,0] v_min_i32_e64_dpp v5, v1, v2 dpp8:[7,6,5,4,3,2,1,0] // GFX12: [0x05,0x00,0x11,0xd5,0xe9,0x04,0x02,0x00,0x01,0x77,0x39,0x05] +v_min_i32_e64_dpp v5, v1, s2 dpp8:[7,6,5,4,3,2,1,0] +// GFX12: [0x05,0x00,0x11,0xd5,0xe9,0x04,0x00,0x00,0x01,0x77,0x39,0x05] + +v_min_i32_e64_dpp v5, v1, 10 dpp8:[7,6,5,4,3,2,1,0] +// GFX12: [0x05,0x00,0x11,0xd5,0xe9,0x14,0x01,0x00,0x01,0x77,0x39,0x05] + v_min_i32_e64_dpp v5, v1, v2 dpp8:[7,6,5,4,3,2,1,0] fi:1 // GFX12: [0x05,0x00,0x11,0xd5,0xea,0x04,0x02,0x00,0x01,0x77,0x39,0x05] @@ -264,6 +397,12 @@ v_min_i32_e64_dpp v255, v255, v255 dpp8:[0,0,0,0,0,0,0,0] fi:0 v_min_u32_e64_dpp v5, v1, v2 dpp8:[7,6,5,4,3,2,1,0] // GFX12: [0x05,0x00,0x13,0xd5,0xe9,0x04,0x02,0x00,0x01,0x77,0x39,0x05] +v_min_u32_e64_dpp v5, v1, s2 dpp8:[7,6,5,4,3,2,1,0] +// GFX12: [0x05,0x00,0x13,0xd5,0xe9,0x04,0x00,0x00,0x01,0x77,0x39,0x05] + +v_min_u32_e64_dpp v5, v1, 10 dpp8:[7,6,5,4,3,2,1,0] +// GFX12: [0x05,0x00,0x13,0xd5,0xe9,0x14,0x01,0x00,0x01,0x77,0x39,0x05] + v_min_u32_e64_dpp v5, v1, v2 dpp8:[7,6,5,4,3,2,1,0] fi:1 // GFX12: [0x05,0x00,0x13,0xd5,0xea,0x04,0x02,0x00,0x01,0x77,0x39,0x05] @@ -273,6 +412,12 @@ v_min_u32_e64_dpp v255, v255, v255 dpp8:[0,0,0,0,0,0,0,0] fi:0 v_mul_dx9_zero_f32_e64_dpp v5, v1, v2 dpp8:[7,6,5,4,3,2,1,0] // GFX12: [0x05,0x00,0x07,0xd5,0xe9,0x04,0x02,0x00,0x01,0x77,0x39,0x05] +v_mul_dx9_zero_f32_e64_dpp v5, v1, s2 dpp8:[7,6,5,4,3,2,1,0] +// GFX12: [0x05,0x00,0x07,0xd5,0xe9,0x04,0x00,0x00,0x01,0x77,0x39,0x05] + +v_mul_dx9_zero_f32_e64_dpp v5, v1, 2.0 dpp8:[7,6,5,4,3,2,1,0] +// GFX12: [0x05,0x00,0x07,0xd5,0xe9,0xe8,0x01,0x00,0x01,0x77,0x39,0x05] + v_mul_dx9_zero_f32_e64_dpp v5, |v1|, -v2 mul:2 dpp8:[7,6,5,4,3,2,1,0] // GFX12: [0x05,0x01,0x07,0xd5,0xe9,0x04,0x02,0x48,0x01,0x77,0x39,0x05] @@ -285,6 +430,12 @@ v_mul_dx9_zero_f32_e64_dpp v255, -|v255|, -|v255| clamp div:2 dpp8:[0,0,0,0,0,0, v_mul_f16_e64_dpp v5, v1, v2 dpp8:[7,6,5,4,3,2,1,0] // GFX12: [0x05,0x00,0x35,0xd5,0xe9,0x04,0x02,0x00,0x01,0x77,0x39,0x05] +v_mul_f16_e64_dpp v5, v1, s2 dpp8:[7,6,5,4,3,2,1,0] +// GFX12: [0x05,0x00,0x35,0xd5,0xe9,0x04,0x00,0x00,0x01,0x77,0x39,0x05] + +v_mul_f16_e64_dpp v5, v1, 2.0 dpp8:[7,6,5,4,3,2,1,0] +// GFX12: [0x05,0x00,0x35,0xd5,0xe9,0xe8,0x01,0x00,0x01,0x77,0x39,0x05] + v_mul_f16_e64_dpp v5, |v1|, -v2 mul:2 dpp8:[7,6,5,4,3,2,1,0] // GFX12: [0x05,0x01,0x35,0xd5,0xe9,0x04,0x02,0x48,0x01,0x77,0x39,0x05] @@ -297,6 +448,12 @@ v_mul_f16_e64_dpp v255, -|v255|, -|v255| clamp div:2 dpp8:[0,0,0,0,0,0,0,0] fi:0 v_mul_f32_e64_dpp v5, v1, v2 dpp8:[7,6,5,4,3,2,1,0] // GFX12: [0x05,0x00,0x08,0xd5,0xe9,0x04,0x02,0x00,0x01,0x77,0x39,0x05] +v_mul_f32_e64_dpp v5, v1, s2 dpp8:[7,6,5,4,3,2,1,0] +// GFX12: [0x05,0x00,0x08,0xd5,0xe9,0x04,0x00,0x00,0x01,0x77,0x39,0x05] + +v_mul_f32_e64_dpp v5, v1, 2.0 dpp8:[7,6,5,4,3,2,1,0] +// GFX12: [0x05,0x00,0x08,0xd5,0xe9,0xe8,0x01,0x00,0x01,0x77,0x39,0x05] + v_mul_f32_e64_dpp v5, |v1|, -v2 mul:2 dpp8:[7,6,5,4,3,2,1,0] // GFX12: [0x05,0x01,0x08,0xd5,0xe9,0x04,0x02,0x48,0x01,0x77,0x39,0x05] @@ -309,6 +466,12 @@ v_mul_f32_e64_dpp v255, -|v255|, -|v255| clamp div:2 dpp8:[0,0,0,0,0,0,0,0] fi:0 v_mul_hi_i32_i24_e64_dpp v5, v1, v2 dpp8:[7,6,5,4,3,2,1,0] // GFX12: [0x05,0x00,0x0a,0xd5,0xe9,0x04,0x02,0x00,0x01,0x77,0x39,0x05] +v_mul_hi_i32_i24_e64_dpp v5, v1, s2 dpp8:[7,6,5,4,3,2,1,0] +// GFX12: [0x05,0x00,0x0a,0xd5,0xe9,0x04,0x00,0x00,0x01,0x77,0x39,0x05] + +v_mul_hi_i32_i24_e64_dpp v5, v1, 10 dpp8:[7,6,5,4,3,2,1,0] +// GFX12: [0x05,0x00,0x0a,0xd5,0xe9,0x14,0x01,0x00,0x01,0x77,0x39,0x05] + v_mul_hi_i32_i24_e64_dpp v5, v1, v2 dpp8:[7,6,5,4,3,2,1,0] fi:1 // GFX12: [0x05,0x00,0x0a,0xd5,0xea,0x04,0x02,0x00,0x01,0x77,0x39,0x05] @@ -318,6 +481,12 @@ v_mul_hi_i32_i24_e64_dpp v255, v255, v255 dpp8:[0,0,0,0,0,0,0,0] fi:0 v_mul_hi_u32_u24_e64_dpp v5, v1, v2 dpp8:[7,6,5,4,3,2,1,0] // GFX12: [0x05,0x00,0x0c,0xd5,0xe9,0x04,0x02,0x00,0x01,0x77,0x39,0x05] +v_mul_hi_u32_u24_e64_dpp v5, v1, s2 dpp8:[7,6,5,4,3,2,1,0] +// GFX12: [0x05,0x00,0x0c,0xd5,0xe9,0x04,0x00,0x00,0x01,0x77,0x39,0x05] + +v_mul_hi_u32_u24_e64_dpp v5, v1, 10 dpp8:[7,6,5,4,3,2,1,0] +// GFX12: [0x05,0x00,0x0c,0xd5,0xe9,0x14,0x01,0x00,0x01,0x77,0x39,0x05] + v_mul_hi_u32_u24_e64_dpp v5, v1, v2 dpp8:[7,6,5,4,3,2,1,0] fi:1 // GFX12: [0x05,0x00,0x0c,0xd5,0xea,0x04,0x02,0x00,0x01,0x77,0x39,0x05] @@ -327,6 +496,12 @@ v_mul_hi_u32_u24_e64_dpp v255, v255, v255 dpp8:[0,0,0,0,0,0,0,0] fi:0 v_mul_i32_i24_e64_dpp v5, v1, v2 dpp8:[7,6,5,4,3,2,1,0] // GFX12: [0x05,0x00,0x09,0xd5,0xe9,0x04,0x02,0x00,0x01,0x77,0x39,0x05] +v_mul_i32_i24_e64_dpp v5, v1, s2 dpp8:[7,6,5,4,3,2,1,0] +// GFX12: [0x05,0x00,0x09,0xd5,0xe9,0x04,0x00,0x00,0x01,0x77,0x39,0x05] + +v_mul_i32_i24_e64_dpp v5, v1, 10 dpp8:[7,6,5,4,3,2,1,0] +// GFX12: [0x05,0x00,0x09,0xd5,0xe9,0x14,0x01,0x00,0x01,0x77,0x39,0x05] + v_mul_i32_i24_e64_dpp v5, v1, v2 dpp8:[7,6,5,4,3,2,1,0] fi:1 // GFX12: [0x05,0x00,0x09,0xd5,0xea,0x04,0x02,0x00,0x01,0x77,0x39,0x05] @@ -336,6 +511,12 @@ v_mul_i32_i24_e64_dpp v255, v255, v255 clamp dpp8:[0,0,0,0,0,0,0,0] fi:0 v_mul_legacy_f32_e64_dpp v5, v1, v2 dpp8:[7,6,5,4,3,2,1,0] // GFX12: [0x05,0x00,0x07,0xd5,0xe9,0x04,0x02,0x00,0x01,0x77,0x39,0x05] +v_mul_legacy_f32_e64_dpp v5, v1, s2 dpp8:[7,6,5,4,3,2,1,0] +// GFX12: [0x05,0x00,0x07,0xd5,0xe9,0x04,0x00,0x00,0x01,0x77,0x39,0x05] + +v_mul_legacy_f32_e64_dpp v5, v1, 2.0 dpp8:[7,6,5,4,3,2,1,0] +// GFX12: [0x05,0x00,0x07,0xd5,0xe9,0xe8,0x01,0x00,0x01,0x77,0x39,0x05] + v_mul_legacy_f32_e64_dpp v5, |v1|, -v2 mul:2 dpp8:[7,6,5,4,3,2,1,0] // GFX12: [0x05,0x01,0x07,0xd5,0xe9,0x04,0x02,0x48,0x01,0x77,0x39,0x05] @@ -348,6 +529,12 @@ v_mul_legacy_f32_e64_dpp v255, -|v255|, -|v255| clamp div:2 dpp8:[0,0,0,0,0,0,0, v_mul_u32_u24_e64_dpp v5, v1, v2 dpp8:[7,6,5,4,3,2,1,0] // GFX12: [0x05,0x00,0x0b,0xd5,0xe9,0x04,0x02,0x00,0x01,0x77,0x39,0x05] +v_mul_u32_u24_e64_dpp v5, v1, s2 dpp8:[7,6,5,4,3,2,1,0] +// GFX12: [0x05,0x00,0x0b,0xd5,0xe9,0x04,0x00,0x00,0x01,0x77,0x39,0x05] + +v_mul_u32_u24_e64_dpp v5, v1, 10 dpp8:[7,6,5,4,3,2,1,0] +// GFX12: [0x05,0x00,0x0b,0xd5,0xe9,0x14,0x01,0x00,0x01,0x77,0x39,0x05] + v_mul_u32_u24_e64_dpp v5, v1, v2 dpp8:[7,6,5,4,3,2,1,0] fi:1 // GFX12: [0x05,0x00,0x0b,0xd5,0xea,0x04,0x02,0x00,0x01,0x77,0x39,0x05] @@ -357,6 +544,12 @@ v_mul_u32_u24_e64_dpp v255, v255, v255 clamp dpp8:[0,0,0,0,0,0,0,0] fi:0 v_or_b32_e64_dpp v5, v1, v2 dpp8:[7,6,5,4,3,2,1,0] // GFX12: [0x05,0x00,0x1c,0xd5,0xe9,0x04,0x02,0x00,0x01,0x77,0x39,0x05] +v_or_b32_e64_dpp v5, v1, s2 dpp8:[7,6,5,4,3,2,1,0] +// GFX12: [0x05,0x00,0x1c,0xd5,0xe9,0x04,0x00,0x00,0x01,0x77,0x39,0x05] + +v_or_b32_e64_dpp v5, v1, 10 dpp8:[7,6,5,4,3,2,1,0] +// GFX12: [0x05,0x00,0x1c,0xd5,0xe9,0x14,0x01,0x00,0x01,0x77,0x39,0x05] + v_or_b32_e64_dpp v5, v1, v2 dpp8:[7,6,5,4,3,2,1,0] fi:1 // GFX12: [0x05,0x00,0x1c,0xd5,0xea,0x04,0x02,0x00,0x01,0x77,0x39,0x05] @@ -405,6 +598,12 @@ v_sub_co_ci_u32_e64_dpp v255, null, v255, v255, null clamp dpp8:[0,0,0,0,0,0,0,0 v_sub_f16_e64_dpp v5, v1, v2 dpp8:[7,6,5,4,3,2,1,0] // GFX12: [0x05,0x00,0x33,0xd5,0xe9,0x04,0x02,0x00,0x01,0x77,0x39,0x05] +v_sub_f16_e64_dpp v5, v1, s2 dpp8:[7,6,5,4,3,2,1,0] +// GFX12: [0x05,0x00,0x33,0xd5,0xe9,0x04,0x00,0x00,0x01,0x77,0x39,0x05] + +v_sub_f16_e64_dpp v5, v1, 2.0 dpp8:[7,6,5,4,3,2,1,0] +// GFX12: [0x05,0x00,0x33,0xd5,0xe9,0xe8,0x01,0x00,0x01,0x77,0x39,0x05] + v_sub_f16_e64_dpp v5, |v1|, -v2 mul:2 dpp8:[7,6,5,4,3,2,1,0] // GFX12: [0x05,0x01,0x33,0xd5,0xe9,0x04,0x02,0x48,0x01,0x77,0x39,0x05] @@ -417,6 +616,12 @@ v_sub_f16_e64_dpp v255, -|v255|, -|v255| clamp div:2 dpp8:[0,0,0,0,0,0,0,0] fi:0 v_sub_f32_e64_dpp v5, v1, v2 dpp8:[7,6,5,4,3,2,1,0] // GFX12: [0x05,0x00,0x04,0xd5,0xe9,0x04,0x02,0x00,0x01,0x77,0x39,0x05] +v_sub_f32_e64_dpp v5, v1, s2 dpp8:[7,6,5,4,3,2,1,0] +// GFX12: [0x05,0x00,0x04,0xd5,0xe9,0x04,0x00,0x00,0x01,0x77,0x39,0x05] + +v_sub_f32_e64_dpp v5, v1, 2.0 dpp8:[7,6,5,4,3,2,1,0] +// GFX12: [0x05,0x00,0x04,0xd5,0xe9,0xe8,0x01,0x00,0x01,0x77,0x39,0x05] + v_sub_f32_e64_dpp v5, |v1|, -v2 mul:2 dpp8:[7,6,5,4,3,2,1,0] // GFX12: [0x05,0x01,0x04,0xd5,0xe9,0x04,0x02,0x48,0x01,0x77,0x39,0x05] @@ -429,6 +634,12 @@ v_sub_f32_e64_dpp v255, -|v255|, -|v255| clamp div:2 dpp8:[0,0,0,0,0,0,0,0] fi:0 v_sub_nc_u32_e64_dpp v5, v1, v2 dpp8:[7,6,5,4,3,2,1,0] // GFX12: [0x05,0x00,0x26,0xd5,0xe9,0x04,0x02,0x00,0x01,0x77,0x39,0x05] +v_sub_nc_u32_e64_dpp v5, v1, s2 dpp8:[7,6,5,4,3,2,1,0] +// GFX12: [0x05,0x00,0x26,0xd5,0xe9,0x04,0x00,0x00,0x01,0x77,0x39,0x05] + +v_sub_nc_u32_e64_dpp v5, v1, 10 dpp8:[7,6,5,4,3,2,1,0] +// GFX12: [0x05,0x00,0x26,0xd5,0xe9,0x14,0x01,0x00,0x01,0x77,0x39,0x05] + v_sub_nc_u32_e64_dpp v5, v1, v2 dpp8:[7,6,5,4,3,2,1,0] fi:1 // GFX12: [0x05,0x00,0x26,0xd5,0xea,0x04,0x02,0x00,0x01,0x77,0x39,0x05] @@ -477,6 +688,12 @@ v_subrev_co_ci_u32_e64_dpp v255, null, v255, v255, null clamp dpp8:[0,0,0,0,0,0, v_subrev_f16_e64_dpp v5, v1, v2 dpp8:[7,6,5,4,3,2,1,0] // GFX12: [0x05,0x00,0x34,0xd5,0xe9,0x04,0x02,0x00,0x01,0x77,0x39,0x05] +v_subrev_f16_e64_dpp v5, v1, s2 dpp8:[7,6,5,4,3,2,1,0] +// GFX12: [0x05,0x00,0x34,0xd5,0xe9,0x04,0x00,0x00,0x01,0x77,0x39,0x05] + +v_subrev_f16_e64_dpp v5, v1, 2.0 dpp8:[7,6,5,4,3,2,1,0] +// GFX12: [0x05,0x00,0x34,0xd5,0xe9,0xe8,0x01,0x00,0x01,0x77,0x39,0x05] + v_subrev_f16_e64_dpp v5, |v1|, -v2 mul:2 dpp8:[7,6,5,4,3,2,1,0] // GFX12: [0x05,0x01,0x34,0xd5,0xe9,0x04,0x02,0x48,0x01,0x77,0x39,0x05] @@ -489,6 +706,12 @@ v_subrev_f16_e64_dpp v255, -|v255|, -|v255| clamp div:2 dpp8:[0,0,0,0,0,0,0,0] f v_subrev_f32_e64_dpp v5, v1, v2 dpp8:[7,6,5,4,3,2,1,0] // GFX12: [0x05,0x00,0x05,0xd5,0xe9,0x04,0x02,0x00,0x01,0x77,0x39,0x05] +v_subrev_f32_e64_dpp v5, v1, s2 dpp8:[7,6,5,4,3,2,1,0] +// GFX12: [0x05,0x00,0x05,0xd5,0xe9,0x04,0x00,0x00,0x01,0x77,0x39,0x05] + +v_subrev_f32_e64_dpp v5, v1, 2.0 dpp8:[7,6,5,4,3,2,1,0] +// GFX12: [0x05,0x00,0x05,0xd5,0xe9,0xe8,0x01,0x00,0x01,0x77,0x39,0x05] + v_subrev_f32_e64_dpp v5, |v1|, -v2 mul:2 dpp8:[7,6,5,4,3,2,1,0] // GFX12: [0x05,0x01,0x05,0xd5,0xe9,0x04,0x02,0x48,0x01,0x77,0x39,0x05] @@ -501,6 +724,12 @@ v_subrev_f32_e64_dpp v255, -|v255|, -|v255| clamp div:2 dpp8:[0,0,0,0,0,0,0,0] f v_subrev_nc_u32_e64_dpp v5, v1, v2 dpp8:[7,6,5,4,3,2,1,0] // GFX12: [0x05,0x00,0x27,0xd5,0xe9,0x04,0x02,0x00,0x01,0x77,0x39,0x05] +v_subrev_nc_u32_e64_dpp v5, v1, s2 dpp8:[7,6,5,4,3,2,1,0] +// GFX12: [0x05,0x00,0x27,0xd5,0xe9,0x04,0x00,0x00,0x01,0x77,0x39,0x05] + +v_subrev_nc_u32_e64_dpp v5, v1, 10 dpp8:[7,6,5,4,3,2,1,0] +// GFX12: [0x05,0x00,0x27,0xd5,0xe9,0x14,0x01,0x00,0x01,0x77,0x39,0x05] + v_subrev_nc_u32_e64_dpp v5, v1, v2 dpp8:[7,6,5,4,3,2,1,0] fi:1 // GFX12: [0x05,0x00,0x27,0xd5,0xea,0x04,0x02,0x00,0x01,0x77,0x39,0x05] @@ -510,6 +739,12 @@ v_subrev_nc_u32_e64_dpp v255, v255, v255 clamp dpp8:[0,0,0,0,0,0,0,0] fi:0 v_xnor_b32_e64_dpp v5, v1, v2 dpp8:[7,6,5,4,3,2,1,0] // GFX12: [0x05,0x00,0x1e,0xd5,0xe9,0x04,0x02,0x00,0x01,0x77,0x39,0x05] +v_xnor_b32_e64_dpp v5, v1, s2 dpp8:[7,6,5,4,3,2,1,0] +// GFX12: [0x05,0x00,0x1e,0xd5,0xe9,0x04,0x00,0x00,0x01,0x77,0x39,0x05] + +v_xnor_b32_e64_dpp v5, v1, 10 dpp8:[7,6,5,4,3,2,1,0] +// GFX12: [0x05,0x00,0x1e,0xd5,0xe9,0x14,0x01,0x00,0x01,0x77,0x39,0x05] + v_xnor_b32_e64_dpp v5, v1, v2 dpp8:[7,6,5,4,3,2,1,0] fi:1 // GFX12: [0x05,0x00,0x1e,0xd5,0xea,0x04,0x02,0x00,0x01,0x77,0x39,0x05] @@ -519,6 +754,12 @@ v_xnor_b32_e64_dpp v255, v255, v255 dpp8:[0,0,0,0,0,0,0,0] fi:0 v_xor_b32_e64_dpp v5, v1, v2 dpp8:[7,6,5,4,3,2,1,0] // GFX12: [0x05,0x00,0x1d,0xd5,0xe9,0x04,0x02,0x00,0x01,0x77,0x39,0x05] +v_xor_b32_e64_dpp v5, v1, s2 dpp8:[7,6,5,4,3,2,1,0] +// GFX12: [0x05,0x00,0x1d,0xd5,0xe9,0x04,0x00,0x00,0x01,0x77,0x39,0x05] + +v_xor_b32_e64_dpp v5, v1, 10 dpp8:[7,6,5,4,3,2,1,0] +// GFX12: [0x05,0x00,0x1d,0xd5,0xe9,0x14,0x01,0x00,0x01,0x77,0x39,0x05] + v_xor_b32_e64_dpp v5, v1, v2 dpp8:[7,6,5,4,3,2,1,0] fi:1 // GFX12: [0x05,0x00,0x1d,0xd5,0xea,0x04,0x02,0x00,0x01,0x77,0x39,0x05] |