Skip to content

[AMDGPU][MC] Add dpp for V_PK_FMAC_F16 for GFX10 #79598

New issue

Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.

Sign up for GitHub

By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.

Already on GitHub? Sign in to your account

Open
ankurepa wants to merge 5 commits into
base: main
Choose a base branch
from
Open

[AMDGPU][MC] Add dpp for V_PK_FMAC_F16 for GFX10 #79598

Show file tree
Hide file tree
Changes from 1 commit
Commits
Show all changes
5 commits
Select commit Hold shift + click to select a range
d3e1954
[AMDGPU][MC] Add dpp for V_PK_FMAC_F16 for GFX10
Jan 26, 2024
7bf323c
Fix test
Feb 7, 2024
a6998ce
codegen dpp combine test
Feb 9, 2024
9bfd6fa
fix
Feb 9, 2024
9ee09fe
Added gfx10 dpp combine test
Feb 9, 2024
File filter

Filter by extension

Filter by extension
Conversations
Failed to load comments.
Loading
Jump to
Jump to file
Failed to load files.
Loading
Diff view
Diff view
15 changes: 11 additions & 4 deletions llvm/lib/Target/AMDGPU/VOP2Instructions.td
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -1615,6 +1615,9 @@ multiclass VOP2_Real_FULL_with_name_gfx11_gfx12<bits<6> op, string opName,
multiclass VOP2_Real_e32_gfx11_gfx12<bits<6> op> :
VOP2Only_Real<GFX11Gen, op>, VOP2Only_Real<GFX12Gen, op>;

multiclass VOP2_V_PK_FMAC_F16_gfx11_gfx12<bits<6> op> :
VOP2Only_Real_e32<GFX11Gen, op>, VOP2Only_Real_e32<GFX12Gen, op>;

multiclass VOP3Only_Realtriple_gfx11_gfx12<bits<10> op> :
VOP3Only_Realtriple<GFX11Gen, op>, VOP3Only_Realtriple<GFX12Gen, op>;

Expand Down Expand Up @@ -1661,7 +1664,8 @@ defm V_SUBREV_CO_CI_U32 :

defm V_CVT_PK_RTZ_F16_F32 : VOP2_Real_FULL_with_name_gfx11_gfx12<0x02f,
"V_CVT_PKRTZ_F16_F32", "v_cvt_pk_rtz_f16_f32">;
defm V_PK_FMAC_F16 : VOP2_Real_e32_gfx11_gfx12<0x03c>;

defm V_PK_FMAC_F16 : VOP2_V_PK_FMAC_F16_gfx11_gfx12<0x03c>;

defm V_ADD_F16_t16 : VOP2_Real_FULL_t16_gfx11_gfx12<0x032, "v_add_f16">;
defm V_ADD_F16_fake16 : VOP2_Real_FULL_t16_gfx11_gfx12<0x032, "v_add_f16">;
Expand Down Expand Up @@ -1945,6 +1949,11 @@ multiclass VOP2e_Real_gfx10<bits<6> op, string opName, string asmName> :
VOP2be_Real_dpp_gfx10<op, opName, asmName>,
VOP2be_Real_dpp8_gfx10<op, opName, asmName>;

multiclass VOP2_FMAC_Real<bits<6> op> :
VOP2_Real_e32_gfx10<op>,
VOP2_Real_dpp_gfx10<op>,
VOP2_Real_dpp8_gfx10<op>;

multiclass VOP2_Real_gfx10<bits<6> op> :
VOP2_Real_e32_gfx10<op>, VOP2_Real_e64_gfx10<op>,
VOP2_Real_sdwa_gfx10<op>, VOP2_Real_dpp_gfx10<op>, VOP2_Real_dpp8_gfx10<op>;
Expand Down Expand Up @@ -1988,9 +1997,7 @@ defm V_MAX_F16 : VOP2_Real_gfx10<0x039>;
defm V_MIN_F16 : VOP2_Real_gfx10<0x03a>;
defm V_LDEXP_F16 : VOP2_Real_gfx10<0x03b>;

let IsSingle = 1 in {
defm V_PK_FMAC_F16 : VOP2_Real_e32_gfx10<0x03c>;
}
defm V_PK_FMAC_F16 : VOP2_FMAC_Real<0x03c>;

// VOP2 no carry-in, carry-out.
defm V_ADD_NC_U32 :
Expand Down
6 changes: 6 additions & 0 deletions llvm/test/MC/AMDGPU/gfx10_asm_vop2.s
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -13185,3 +13185,9 @@ v_pk_fmac_f16 v5, -4.0, v2

v_pk_fmac_f16 v5, v1, v255
// GFX10: encoding: [0x01,0xff,0x0b,0x78]

v_pk_fmac_f16_dpp v5, v1, v2 quad_perm:[0,1,2,3]
// GFX10: encoding: [0xfa,0x04,0x0a,0x78,0x01,0xe4,0x00,0xff]

v_pk_fmac_f16_dpp v5, v1, v2 quad_perm:[0,1,2,3] row_mask:0x0 bank_mask:0x3
Copy link
Contributor

@Sisyph Sisyph Feb 23, 2024

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

Please add gfx10 error tests that DPP16 NEG and ABS modifiers are not supported by this instruction

// GFX10: encoding: [0xfa,0x04,0x0a,0x78,0x01,0xe4,0x00,0x03]
2 changes: 1 addition & 1 deletion llvm/test/MC/AMDGPU/literalv216.s
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -291,4 +291,4 @@ v_pk_add_u16 v5, v1, 123456.0
// FIXME: v_pk_fmac_f16 cannot be promoted to VOP3 so '_e32' suffix is not valid
v_pk_fmac_f16 v5, 0x12345678, v2
// NOGFX9: :[[@LINE-1]]:{{[0-9]+}}: error: instruction not supported on this GPU
// GFX10: v_pk_fmac_f16 v5, 0x12345678, v2 ; encoding: [0xff,0x04,0x0a,0x78,0x78,0x56,0x34,0x12]
// GFX10: v_pk_fmac_f16_e32 v5, 0x12345678, v2 ; encoding: [0xff,0x04,0x0a,0x78,0x78,0x56,0x34,0x12]
Copy link
Contributor

@Sisyph Sisyph Feb 23, 2024

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

If I have the convention right, as stated in this patch 0f5ebbc, instructions without a VOP3 form should not have _e32. It looks like removing IsSingle caused _e32 to be added to the mnemonic. Please change it back.

38 changes: 22 additions & 16 deletions llvm/test/MC/Disassembler/AMDGPU/gfx10_vop2.txt
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -1779,54 +1779,60 @@
# GFX10: v_or_b32_e32 v5, vcc_lo, v2 ; encoding: [0x6a,0x04,0x0a,0x38]
0x6a,0x04,0x0a,0x38

# GFX10: v_pk_fmac_f16 v255, v1, v2 ; encoding: [0x01,0x05,0xfe,0x79]
# GFX10: v_pk_fmac_f16_e32 v255, v1, v2 ; encoding: [0x01,0x05,0xfe,0x79]
0x01,0x05,0xfe,0x79

# GFX10: v_pk_fmac_f16 v5, -1, v2 ; encoding: [0xc1,0x04,0x0a,0x78]
# GFX10: v_pk_fmac_f16_e32 v5, -1, v2 ; encoding: [0xc1,0x04,0x0a,0x78]
0xc1,0x04,0x0a,0x78

# GFX10: v_pk_fmac_f16 v5, -4.0, v2 ; encoding: [0xf7,0x04,0x0a,0x78]
# GFX10: v_pk_fmac_f16_e32 v5, -4.0, v2 ; encoding: [0xf7,0x04,0x0a,0x78]
0xf7,0x04,0x0a,0x78

# GFX10: v_pk_fmac_f16 v5, 0, v2 ; encoding: [0x80,0x04,0x0a,0x78]
# GFX10: v_pk_fmac_f16_e32 v5, 0, v2 ; encoding: [0x80,0x04,0x0a,0x78]
0x80,0x04,0x0a,0x78

# GFX10: v_pk_fmac_f16 v5, 0.5, v2 ; encoding: [0xf0,0x04,0x0a,0x78]
# GFX10: v_pk_fmac_f16_e32 v5, 0.5, v2 ; encoding: [0xf0,0x04,0x0a,0x78]
0xf0,0x04,0x0a,0x78

# GFX10: v_pk_fmac_f16 v5, exec_hi, v2 ; encoding: [0x7f,0x04,0x0a,0x78]
# GFX10: v_pk_fmac_f16_e32 v5, exec_hi, v2 ; encoding: [0x7f,0x04,0x0a,0x78]
0x7f,0x04,0x0a,0x78

# GFX10: v_pk_fmac_f16 v5, exec_lo, v2 ; encoding: [0x7e,0x04,0x0a,0x78]
# GFX10: v_pk_fmac_f16_e32 v5, exec_lo, v2 ; encoding: [0x7e,0x04,0x0a,0x78]
0x7e,0x04,0x0a,0x78

# GFX10: v_pk_fmac_f16 v5, m0, v2 ; encoding: [0x7c,0x04,0x0a,0x78]
# GFX10: v_pk_fmac_f16_e32 v5, m0, v2 ; encoding: [0x7c,0x04,0x0a,0x78]
0x7c,0x04,0x0a,0x78

# GFX10: v_pk_fmac_f16 v5, s1, v2 ; encoding: [0x01,0x04,0x0a,0x78]
# GFX10: v_pk_fmac_f16_e32 v5, s1, v2 ; encoding: [0x01,0x04,0x0a,0x78]
0x01,0x04,0x0a,0x78

# GFX10: v_pk_fmac_f16 v5, s103, v2 ; encoding: [0x67,0x04,0x0a,0x78]
# GFX10: v_pk_fmac_f16_e32 v5, s103, v2 ; encoding: [0x67,0x04,0x0a,0x78]
0x67,0x04,0x0a,0x78

# GFX10: v_pk_fmac_f16 v5, ttmp11, v2 ; encoding: [0x77,0x04,0x0a,0x78]
# GFX10: v_pk_fmac_f16_e32 v5, ttmp11, v2 ; encoding: [0x77,0x04,0x0a,0x78]
0x77,0x04,0x0a,0x78

# GFX10: v_pk_fmac_f16 v5, v1, v2 ; encoding: [0x01,0x05,0x0a,0x78]
# GFX10: v_pk_fmac_f16_e32 v5, v1, v2 ; encoding: [0x01,0x05,0x0a,0x78]
0x01,0x05,0x0a,0x78

# GFX10: v_pk_fmac_f16 v5, v1, v255 ; encoding: [0x01,0xff,0x0b,0x78]
# GFX10: v_pk_fmac_f16_e32 v5, v1, v255 ; encoding: [0x01,0xff,0x0b,0x78]
0x01,0xff,0x0b,0x78

# GFX10: v_pk_fmac_f16 v5, v255, v2 ; encoding: [0xff,0x05,0x0a,0x78]
# GFX10: v_pk_fmac_f16_e32 v5, v255, v2 ; encoding: [0xff,0x05,0x0a,0x78]
0xff,0x05,0x0a,0x78

# GFX10: v_pk_fmac_f16 v5, vcc_hi, v2 ; encoding: [0x6b,0x04,0x0a,0x78]
# GFX10: v_pk_fmac_f16_e32 v5, vcc_hi, v2 ; encoding: [0x6b,0x04,0x0a,0x78]
0x6b,0x04,0x0a,0x78

# GFX10: v_pk_fmac_f16 v5, vcc_lo, v2 ; encoding: [0x6a,0x04,0x0a,0x78]
# GFX10: v_pk_fmac_f16_e32 v5, vcc_lo, v2 ; encoding: [0x6a,0x04,0x0a,0x78]
0x6a,0x04,0x0a,0x78

#GFX10: v_pk_fmac_f16_dpp v5, v1, v2 quad_perm:[0,1,2,3] row_mask:0xf bank_mask:0xf ; encoding: [0xfa,0x04,0x0a,0x78,0x01,0xe4,0x00,0xff]
0xfa,0x04,0x0a,0x78,0x01,0xe4,0x00,0xff

#GFX10: v_pk_fmac_f16_dpp v5, v1, v2 quad_perm:[0,1,2,3] row_mask:0x0 bank_mask:0x3 ; encoding: [0xfa,0x04,0x0a,0x78,0x01,0xe4,0x00,0x03]
0xfa,0x04,0x0a,0x78,0x01,0xe4,0x00,0x03

# W32: v_sub_co_ci_u32_e32 v255, vcc_lo, v1, v2, vcc_lo ; encoding: [0x01,0x05,0xfe,0x53]
# W64: v_sub_co_ci_u32_e32 v255, vcc, v1, v2, vcc ; encoding: [0x01,0x05,0xfe,0x53]
0x01,0x05,0xfe,0x53
Expand Down
2 changes: 1 addition & 1 deletion llvm/test/MC/Disassembler/AMDGPU/gfx10_vop3p_literalv216.txt
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -144,5 +144,5 @@
# Packed VOP2
#===----------------------------------------------------------------------===//

# GFX10: v_pk_fmac_f16 v5, 0x12345678, v2 ; encoding: [0xff,0x04,0x0a,0x78,0x78,0x56,0x34,0x12]
# GFX10: v_pk_fmac_f16_e32 v5, 0x12345678, v2 ; encoding: [0xff,0x04,0x0a,0x78,0x78,0x56,0x34,0x12]
0xff,0x04,0x0a,0x78,0x78,0x56,0x34,0x12