[llvm] [AMDGPU] Allow dpp in v_pk_fmac_f16 for GFX9 and GFX10 (PR #144782)
Jun Wang via llvm-commits
llvm-commits at lists.llvm.org
Thu Jun 19 09:42:17 PDT 2025
https://github.com/jwanggit86 updated https://github.com/llvm/llvm-project/pull/144782
>From f81287454323862f75863f61f042348aa16486a8 Mon Sep 17 00:00:00 2001
From: Jun Wang <jwang86 at yahoo.com>
Date: Wed, 18 Jun 2025 11:59:39 -0700
Subject: [PATCH 1/3] [AMDGPU] Allow dpp in v_pk_fmac_f16 for GFX9 and GFX10
Allows dpp in v_pk_fmac-f16 for GFX9, and both dpp and dpp8
for GFX10.
---
llvm/lib/Target/AMDGPU/VOP2Instructions.td | 8 ++++++++
llvm/test/MC/AMDGPU/gfx10_asm_vop2.s | 12 ++++++++++++
llvm/test/MC/AMDGPU/gfx9_asm_vop2_features.s | 13 +++++++++++++
.../MC/Disassembler/AMDGPU/gfx10_vop2_dpp16.txt | 7 +++++++
.../test/MC/Disassembler/AMDGPU/gfx10_vop2_dpp8.txt | 3 +++
.../MC/Disassembler/AMDGPU/gfx9_vop2_features.txt | 10 ++++++++++
6 files changed, 53 insertions(+)
create mode 100644 llvm/test/MC/AMDGPU/gfx9_asm_vop2_features.s
create mode 100644 llvm/test/MC/Disassembler/AMDGPU/gfx9_vop2_features.txt
diff --git a/llvm/lib/Target/AMDGPU/VOP2Instructions.td b/llvm/lib/Target/AMDGPU/VOP2Instructions.td
index 0c7e20fc1ebf3..c459c4df11e9e 100644
--- a/llvm/lib/Target/AMDGPU/VOP2Instructions.td
+++ b/llvm/lib/Target/AMDGPU/VOP2Instructions.td
@@ -2172,6 +2172,7 @@ 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_Real_dpp_gfx10<0x03c>, VOP2_Real_dpp8_gfx10<0x03c>;
// VOP2 no carry-in, carry-out.
defm V_ADD_NC_U32 :
@@ -2504,6 +2505,11 @@ multiclass VOP2_Real_e32e64_gfx9 <bits<6> op> {
VOP2_DPPe<op, !cast<VOP2_DPP_Pseudo>(NAME#"_dpp")>;
}
+ multiclass VOP2_Real_dpp_gfx9<bits<6> op> {
+ if !cast<VOP2_Pseudo>(NAME#"_e32").Pfl.HasExtDPP then
+ def _dpp_gfx9 : VOP2_DPP16<op, !cast<VOP2_DPP_Pseudo>(NAME#"_dpp"), SIEncodingFamily.GFX9>;
+ }
+
} // End DecoderNamespace = "GFX9"
multiclass VOP2_Real_e32e64_vi <bits<6> op> :
@@ -2560,6 +2566,8 @@ defm V_SUBBREV_CO_U32 : VOP2be_Real_e32e64_gfx9 <0x1e, "V_SUBBREV_U32", "v_s
defm V_ADD_U32 : VOP2_Real_e32e64_gfx9 <0x34>;
defm V_SUB_U32 : VOP2_Real_e32e64_gfx9 <0x35>;
defm V_SUBREV_U32 : VOP2_Real_e32e64_gfx9 <0x36>;
+
+defm V_PK_FMAC_F16 : VOP2_Real_dpp_gfx9<0x03c>;
} // End AssemblerPredicate = isGFX9Only
defm V_BFM_B32 : VOP2_Real_e64only_vi <0x293>;
diff --git a/llvm/test/MC/AMDGPU/gfx10_asm_vop2.s b/llvm/test/MC/AMDGPU/gfx10_asm_vop2.s
index 3dcf288bbbaa5..bbd36a9140b96 100644
--- a/llvm/test/MC/AMDGPU/gfx10_asm_vop2.s
+++ b/llvm/test/MC/AMDGPU/gfx10_asm_vop2.s
@@ -13185,3 +13185,15 @@ 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 v5, v1, v2
+// GFX10: encoding: [0x01,0x05,0x0a,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:0x0
+// GFX10: encoding: [0xfa,0x04,0x0a,0x78,0x01,0xe4,0x00,0x00]
+
+v_pk_fmac_f16_dpp v5, v1, v2 dpp8:[7,6,5,4,3,2,1,0]
+// GFX10: encoding: [0xe9,0x04,0x0a,0x78,0x01,0x77,0x39,0x05]
diff --git a/llvm/test/MC/AMDGPU/gfx9_asm_vop2_features.s b/llvm/test/MC/AMDGPU/gfx9_asm_vop2_features.s
new file mode 100644
index 0000000000000..f7dab2d0359dc
--- /dev/null
+++ b/llvm/test/MC/AMDGPU/gfx9_asm_vop2_features.s
@@ -0,0 +1,13 @@
+// RUN: llvm-mc -triple=amdgcn -mcpu=gfx908 -show-encoding %s | FileCheck --check-prefix=CHECK-MI %s
+// RUN: llvm-mc -triple=amdgcn -mcpu=gfx90a -show-encoding %s | FileCheck --check-prefix=CHECK-MI %s
+// RUN: llvm-mc -triple=amdgcn -mcpu=gfx942 -show-encoding %s | FileCheck --check-prefix=CHECK-MI %s
+// RUN: llvm-mc -triple=amdgcn -mcpu=gfx950 -show-encoding %s | FileCheck --check-prefix=CHECK-MI %s
+
+v_pk_fmac_f16 v5, v1, v2
+// CHECK-MI: [0x01,0x05,0x0a,0x78]
+
+v_pk_fmac_f16 v5, v1, v2 quad_perm:[0,1,2,3]
+// CHECK-MI: [0xfa,0x04,0x0a,0x78,0x01,0xe4,0x00,0xff]
+
+v_pk_fmac_f16 v5, v1, v2 quad_perm:[0,1,2,3] row_mask:0x0 bank_mask:0x0
+// CHECK-MI: [0xfa,0x04,0x0a,0x78,0x01,0xe4,0x00,0x00]
diff --git a/llvm/test/MC/Disassembler/AMDGPU/gfx10_vop2_dpp16.txt b/llvm/test/MC/Disassembler/AMDGPU/gfx10_vop2_dpp16.txt
index 1774efa4a65c7..4a7471e6c6f98 100644
--- a/llvm/test/MC/Disassembler/AMDGPU/gfx10_vop2_dpp16.txt
+++ b/llvm/test/MC/Disassembler/AMDGPU/gfx10_vop2_dpp16.txt
@@ -2476,3 +2476,10 @@
# W32: v_cndmask_b32_dpp v5, -|v1|, -|v2|, vcc_lo quad_perm:[0,1,2,3] row_mask:0xf bank_mask:0xf ; encoding: [0xfa,0x04,0x0a,0x02,0x01,0xe4,0xf0,0xff]
# W64: v_cndmask_b32_dpp v5, -|v1|, -|v2|, vcc quad_perm:[0,1,2,3] row_mask:0xf bank_mask:0xf ; encoding: [0xfa,0x04,0x0a,0x02,0x01,0xe4,0xf0,0xff]
0xfa,0x04,0x0a,0x02,0x01,0xe4,0xf0,0xff
+
+# 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:0x0 ; encoding: [0xfa,0x04,0x0a,0x78,0x01,0xe4,0x00,0x00]
+0xfa,0x04,0x0a,0x78,0x01,0xe4,0x00,0x00
+
diff --git a/llvm/test/MC/Disassembler/AMDGPU/gfx10_vop2_dpp8.txt b/llvm/test/MC/Disassembler/AMDGPU/gfx10_vop2_dpp8.txt
index 40b8f24e4d72f..233f93a5b8e7d 100644
--- a/llvm/test/MC/Disassembler/AMDGPU/gfx10_vop2_dpp8.txt
+++ b/llvm/test/MC/Disassembler/AMDGPU/gfx10_vop2_dpp8.txt
@@ -222,3 +222,6 @@
# W32: v_cndmask_b32_dpp v0, v1, v2, vcc_lo dpp8:[0,1,2,3,4,5,6,7] fi:1 ; encoding: [0xea,0x04,0x00,0x02,0x01,0x88,0xc6,0xfa]
# W64: v_cndmask_b32_dpp v0, v1, v2, vcc dpp8:[0,1,2,3,4,5,6,7] fi:1 ; encoding: [0xea,0x04,0x00,0x02,0x01,0x88,0xc6,0xfa]
0xea,0x04,0x00,0x02,0x01,0x88,0xc6,0xfa
+
+# GFX10: v_pk_fmac_f16_dpp v5, v1, v2 dpp8:[7,6,5,4,3,2,1,0] ; encoding: [0xe9,0x04,0x0a,0x78,0x01,0x77,0x39,0x05]
+0xe9,0x04,0x0a,0x78,0x01,0x77,0x39,0x05
diff --git a/llvm/test/MC/Disassembler/AMDGPU/gfx9_vop2_features.txt b/llvm/test/MC/Disassembler/AMDGPU/gfx9_vop2_features.txt
new file mode 100644
index 0000000000000..ac1ef4baa9aa4
--- /dev/null
+++ b/llvm/test/MC/Disassembler/AMDGPU/gfx9_vop2_features.txt
@@ -0,0 +1,10 @@
+# RUN: llvm-mc -triple=amdgcn -mcpu=gfx908 -disassemble -show-encoding < %s | FileCheck -check-prefix=CHECK-MI %s
+
+# CHECK-MI: v_pk_fmac_f16_e32 v5, v1, v2
+0x01,0x05,0x0a,0x78
+
+# CHECK-MI: v_pk_fmac_f16_dpp v5, v1, v2 quad_perm:[0,1,2,3] row_mask:0xf bank_mask:0xf
+0xfa,0x04,0x0a,0x78,0x01,0xe4,0x00,0xff
+
+# CHECK-MI: v_pk_fmac_f16_dpp v5, v1, v2 quad_perm:[0,1,2,3] row_mask:0x0 bank_mask:0x0
+0xfa,0x04,0x0a,0x78,0x01,0xe4,0x00,0x00
>From 4d6975543f86f71939914e39501f10674c316661 Mon Sep 17 00:00:00 2001
From: Jun Wang <jwang86 at yahoo.com>
Date: Wed, 18 Jun 2025 17:48:13 -0700
Subject: [PATCH 2/3] for gfx9, use existing VOP2_Real_e32e64_gfx9
---
llvm/lib/Target/AMDGPU/VOP2Instructions.td | 7 +------
1 file changed, 1 insertion(+), 6 deletions(-)
diff --git a/llvm/lib/Target/AMDGPU/VOP2Instructions.td b/llvm/lib/Target/AMDGPU/VOP2Instructions.td
index c459c4df11e9e..8c6473a43ec4e 100644
--- a/llvm/lib/Target/AMDGPU/VOP2Instructions.td
+++ b/llvm/lib/Target/AMDGPU/VOP2Instructions.td
@@ -2505,11 +2505,6 @@ multiclass VOP2_Real_e32e64_gfx9 <bits<6> op> {
VOP2_DPPe<op, !cast<VOP2_DPP_Pseudo>(NAME#"_dpp")>;
}
- multiclass VOP2_Real_dpp_gfx9<bits<6> op> {
- if !cast<VOP2_Pseudo>(NAME#"_e32").Pfl.HasExtDPP then
- def _dpp_gfx9 : VOP2_DPP16<op, !cast<VOP2_DPP_Pseudo>(NAME#"_dpp"), SIEncodingFamily.GFX9>;
- }
-
} // End DecoderNamespace = "GFX9"
multiclass VOP2_Real_e32e64_vi <bits<6> op> :
@@ -2567,7 +2562,7 @@ defm V_ADD_U32 : VOP2_Real_e32e64_gfx9 <0x34>;
defm V_SUB_U32 : VOP2_Real_e32e64_gfx9 <0x35>;
defm V_SUBREV_U32 : VOP2_Real_e32e64_gfx9 <0x36>;
-defm V_PK_FMAC_F16 : VOP2_Real_dpp_gfx9<0x03c>;
+defm V_PK_FMAC_F16 : VOP2_Real_e32e64_gfx9<0x03c>;
} // End AssemblerPredicate = isGFX9Only
defm V_BFM_B32 : VOP2_Real_e64only_vi <0x293>;
>From f2d51a235dc7f85e99c2f9a54376d5530a9daddc Mon Sep 17 00:00:00 2001
From: Jun Wang <jwang86 at yahoo.com>
Date: Thu, 19 Jun 2025 09:41:31 -0700
Subject: [PATCH 3/3] for gfx10, use existing VOP2_Real_gfx10
---
llvm/lib/Target/AMDGPU/VOP2Instructions.td | 5 +--
llvm/test/MC/AMDGPU/literalv216.s | 2 +-
.../MC/Disassembler/AMDGPU/gfx10_vop2.txt | 32 +++++++++----------
.../AMDGPU/gfx10_vop3p_literalv216.txt | 2 +-
4 files changed, 19 insertions(+), 22 deletions(-)
diff --git a/llvm/lib/Target/AMDGPU/VOP2Instructions.td b/llvm/lib/Target/AMDGPU/VOP2Instructions.td
index 8c6473a43ec4e..4ab8144b45b74 100644
--- a/llvm/lib/Target/AMDGPU/VOP2Instructions.td
+++ b/llvm/lib/Target/AMDGPU/VOP2Instructions.td
@@ -2169,10 +2169,7 @@ 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_Real_dpp_gfx10<0x03c>, VOP2_Real_dpp8_gfx10<0x03c>;
+defm V_PK_FMAC_F16 : VOP2_Real_gfx10<0x03c>;
// VOP2 no carry-in, carry-out.
defm V_ADD_NC_U32 :
diff --git a/llvm/test/MC/AMDGPU/literalv216.s b/llvm/test/MC/AMDGPU/literalv216.s
index c695bc3600c38..f5afaa6bd6181 100644
--- a/llvm/test/MC/AMDGPU/literalv216.s
+++ b/llvm/test/MC/AMDGPU/literalv216.s
@@ -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]
diff --git a/llvm/test/MC/Disassembler/AMDGPU/gfx10_vop2.txt b/llvm/test/MC/Disassembler/AMDGPU/gfx10_vop2.txt
index fb1099d709940..b6cb64fad9d84 100644
--- a/llvm/test/MC/Disassembler/AMDGPU/gfx10_vop2.txt
+++ b/llvm/test/MC/Disassembler/AMDGPU/gfx10_vop2.txt
@@ -1779,52 +1779,52 @@
# 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
# W32: v_sub_co_ci_u32_e32 v255, vcc_lo, v1, v2, vcc_lo ; encoding: [0x01,0x05,0xfe,0x53]
diff --git a/llvm/test/MC/Disassembler/AMDGPU/gfx10_vop3p_literalv216.txt b/llvm/test/MC/Disassembler/AMDGPU/gfx10_vop3p_literalv216.txt
index a022c79fe97e6..97c81ed1a629a 100644
--- a/llvm/test/MC/Disassembler/AMDGPU/gfx10_vop3p_literalv216.txt
+++ b/llvm/test/MC/Disassembler/AMDGPU/gfx10_vop3p_literalv216.txt
@@ -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
More information about the llvm-commits
mailing list