Skip to content

[AMDGPU] Generate s_absdiff_i32 #164835

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

Merged
LU-JOHN merged 3 commits into from
Oct 27, 2025
Merged

[AMDGPU] Generate s_absdiff_i32 #164835

Show file tree
Hide file tree
Changes from all commits
Commits
Show all changes
3 commits
Select commit Hold shift + click to select a range
a371ba3
Generate s_absdiff_i32
LU-JOHN Oct 21, 2025
81b0703
Remove unnecessary braces
LU-JOHN Oct 23, 2025
6720b3d
Add absdiff test variations
LU-JOHN Oct 24, 2025
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
7 changes: 4 additions & 3 deletions llvm/lib/Target/AMDGPU/SOPInstructions.td
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -838,9 +838,10 @@ def S_CBRANCH_G_FORK : SOP2_Pseudo <
let SubtargetPredicate = isGFX6GFX7GFX8GFX9;
}

let Defs = [SCC] in {
def S_ABSDIFF_I32 : SOP2_32 <"s_absdiff_i32">;
} // End Defs = [SCC]
let isCommutable = 1, Defs = [SCC] in
def S_ABSDIFF_I32 : SOP2_32 <"s_absdiff_i32",
[(set i32:$sdst, (UniformUnaryFrag<abs> (sub_oneuse i32:$src0, i32:$src1)))]
>;

let SubtargetPredicate = isGFX8GFX9 in {
def S_RFE_RESTORE_B64 : SOP2_Pseudo <
Expand Down
104 changes: 104 additions & 0 deletions llvm/test/CodeGen/AMDGPU/absdiff.ll
View file
Open in desktop
Original file line number Diff line number Diff line change
@@ -0,0 +1,104 @@
; NOTE: Assertions have been autogenerated by utils/update_llc_test_checks.py
; RUN: llc -mtriple=amdgcn-amd-amdpal -mcpu=gfx900 < %s | FileCheck %s

Copy link
Contributor

@arsenm arsenm Oct 24, 2025

Choose a reason for hiding this comment

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

Also test 16-bit promoted case

Copy link
Contributor Author

@LU-JOHN LU-JOHN Oct 24, 2025 •
edited
Loading

Choose a reason for hiding this comment

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

Added 16-bit test case.

define amdgpu_ps i16 @absdiff_i16_false(i16 inreg %arg0, i16 inreg %arg1) {
; CHECK-LABEL: absdiff_i16_false:
; CHECK: ; %bb.0:
; CHECK-NEXT: s_sub_i32 s0, s0, s1
; CHECK-NEXT: s_sext_i32_i16 s1, s0
; CHECK-NEXT: s_sub_i32 s0, 0, s0
; CHECK-NEXT: s_sext_i32_i16 s0, s0
; CHECK-NEXT: s_max_i32 s0, s1, s0
; CHECK-NEXT: ; return to shader part epilog
%diff = sub i16 %arg0, %arg1
%res = call i16 @llvm.abs.i16(i16 %diff, i1 false) ; INT_MIN input returns INT_MIN
ret i16 %res
}

define amdgpu_ps i16 @absdiff_i16_true(i16 inreg %arg0, i16 inreg %arg1) {
; CHECK-LABEL: absdiff_i16_true:
; CHECK: ; %bb.0:
; CHECK-NEXT: s_sub_i32 s0, s0, s1
; CHECK-NEXT: s_sext_i32_i16 s1, s0
; CHECK-NEXT: s_sub_i32 s0, 0, s0
; CHECK-NEXT: s_sext_i32_i16 s0, s0
; CHECK-NEXT: s_max_i32 s0, s1, s0
; CHECK-NEXT: ; return to shader part epilog
%diff = sub i16 %arg0, %arg1
%res = call i16 @llvm.abs.i16(i16 %diff, i1 true) ; INT_MIN input returns poison
ret i16 %res
}

define amdgpu_ps i32 @absdiff_i32_false(i32 inreg %arg0, i32 inreg %arg1) {
; CHECK-LABEL: absdiff_i32_false:
; CHECK: ; %bb.0:
; CHECK-NEXT: s_absdiff_i32 s0, s0, s1
; CHECK-NEXT: ; return to shader part epilog
%diff = sub i32 %arg0, %arg1
%res = call i32 @llvm.abs.i32(i32 %diff, i1 false) ; INT_MIN input returns INT_MIN
ret i32 %res
}
Copy link
Contributor

@arsenm arsenm Oct 24, 2025

Choose a reason for hiding this comment

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

Negative test for the multi use case?

Copy link
Contributor Author

@LU-JOHN LU-JOHN Oct 24, 2025

Choose a reason for hiding this comment

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

Added negative test for multi-use case.


define amdgpu_ps i32 @absdiff_i32_true(i32 inreg %arg0, i32 inreg %arg1) {
; CHECK-LABEL: absdiff_i32_true:
; CHECK: ; %bb.0:
; CHECK-NEXT: s_absdiff_i32 s0, s0, s1
; CHECK-NEXT: ; return to shader part epilog
%diff = sub i32 %arg0, %arg1
%res = call i32 @llvm.abs.i32(i32 %diff, i1 true) ; INT_MIN input returns poison
ret i32 %res
}

; Multiple uses of %diff. No benefit for using s_absdiff_i32.
define amdgpu_ps i32 @absdiff_i32_false_multi_use(i32 inreg %arg0, i32 inreg %arg1) {
; CHECK-LABEL: absdiff_i32_false_multi_use:
; CHECK: ; %bb.0:
; CHECK-NEXT: s_sub_i32 s1, s0, s1
; CHECK-NEXT: s_abs_i32 s0, s1
; CHECK-NEXT: ;;#ASMSTART
; CHECK-NEXT: ; use s1
; CHECK-NEXT: ;;#ASMEND
; CHECK-NEXT: ; return to shader part epilog
%diff = sub i32 %arg0, %arg1
%res = call i32 @llvm.abs.i32(i32 %diff, i1 false) ; INT_MIN input returns INT_MIN
call void asm "; use $0", "s"(i32 %diff)
ret i32 %res
}

define <2 x i32> @absdiff_2xi32_false(<2 x i32> %arg0, <2 x i32> %arg1) {
; CHECK-LABEL: absdiff_2xi32_false:
; CHECK: ; %bb.0:
; CHECK-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; CHECK-NEXT: v_sub_u32_e32 v0, v0, v2
; CHECK-NEXT: v_sub_u32_e32 v1, v1, v3
; CHECK-NEXT: v_sub_u32_e32 v2, 0, v0
; CHECK-NEXT: v_max_i32_e32 v0, v2, v0
; CHECK-NEXT: v_sub_u32_e32 v2, 0, v1
; CHECK-NEXT: v_max_i32_e32 v1, v2, v1
; CHECK-NEXT: s_setpc_b64 s[30:31]
%diff = sub <2 x i32> %arg0, %arg1
%res = call <2 x i32> @llvm.abs.v2i32(<2 x i32> %diff, i1 false) ; INT_MIN input returns INT_MIN
ret <2 x i32> %res
}

define <4 x i32> @absdiff_4xi32_false(<4 x i32> %arg0, <4 x i32> %arg1) {
; CHECK-LABEL: absdiff_4xi32_false:
; CHECK: ; %bb.0:
; CHECK-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; CHECK-NEXT: v_sub_u32_e32 v0, v0, v4
; CHECK-NEXT: v_sub_u32_e32 v1, v1, v5
; CHECK-NEXT: v_sub_u32_e32 v4, 0, v0
; CHECK-NEXT: v_sub_u32_e32 v2, v2, v6
; CHECK-NEXT: v_max_i32_e32 v0, v4, v0
; CHECK-NEXT: v_sub_u32_e32 v4, 0, v1
; CHECK-NEXT: v_sub_u32_e32 v3, v3, v7
; CHECK-NEXT: v_max_i32_e32 v1, v4, v1
; CHECK-NEXT: v_sub_u32_e32 v4, 0, v2
; CHECK-NEXT: v_max_i32_e32 v2, v4, v2
; CHECK-NEXT: v_sub_u32_e32 v4, 0, v3
; CHECK-NEXT: v_max_i32_e32 v3, v4, v3
; CHECK-NEXT: s_setpc_b64 s[30:31]
%diff = sub <4 x i32> %arg0, %arg1
%res = call <4 x i32> @llvm.abs.v4i32(<4 x i32> %diff, i1 false) ; INT_MIN input returns INT_MIN
ret <4 x i32> %res
}
25 changes: 20 additions & 5 deletions llvm/test/CodeGen/AMDGPU/s_cmp_0.ll
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -110,6 +110,21 @@ define amdgpu_ps i32 @abs32(i32 inreg %val0) {
ret i32 %zext
}

define amdgpu_ps i32 @absdiff32(i32 inreg %val0, i32 inreg %val1) {
; CHECK-LABEL: absdiff32:
; CHECK: ; %bb.0:
; CHECK-NEXT: s_absdiff_i32 s0, s0, s1
; CHECK-NEXT: s_cselect_b64 s[0:1], -1, 0
; CHECK-NEXT: v_cndmask_b32_e64 v0, 0, 1, s[0:1]
; CHECK-NEXT: v_readfirstlane_b32 s0, v0
Comment on lines +117 to +119
Copy link
Contributor

@jayfoad jayfoad Oct 23, 2025

Choose a reason for hiding this comment

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

Codegen is terrible here, but that's not your fault. Should be:

Suggested change
; CHECK-NEXT: s_cselect_b64 s[0:1], -1, 0
; CHECK-NEXT: v_cndmask_b32_e64 v0, 0, 1, s[0:1]
; CHECK-NEXT: v_readfirstlane_b32 s0, v0
; CHECK-NEXT: s_cselect_b32 s0, 1, 0

; CHECK-NEXT: ; return to shader part epilog
%diff = sub i32 %val0, %val1
%result = call i32 @llvm.abs.i32(i32 %diff, i1 false)
%cmp = icmp ne i32 %result, 0
%zext = zext i1 %cmp to i32
ret i32 %zext
}

define amdgpu_ps i32 @and32(i32 inreg %val0, i32 inreg %val1) {
; CHECK-LABEL: and32:
; CHECK: ; %bb.0:
Expand Down Expand Up @@ -608,14 +623,14 @@ define amdgpu_ps i32 @si_pc_add_rel_offset_must_not_optimize() {
; CHECK-NEXT: s_add_u32 s0, s0, __unnamed_1@rel32@lo+4
; CHECK-NEXT: s_addc_u32 s1, s1, __unnamed_1@rel32@hi+12
; CHECK-NEXT: s_cmp_lg_u64 s[0:1], 0
; CHECK-NEXT: s_cbranch_scc0 .LBB35_2
; CHECK-NEXT: s_cbranch_scc0 .LBB36_2
; CHECK-NEXT: ; %bb.1: ; %endif
; CHECK-NEXT: s_mov_b32 s0, 1
; CHECK-NEXT: s_branch .LBB35_3
; CHECK-NEXT: .LBB35_2: ; %if
; CHECK-NEXT: s_branch .LBB36_3
; CHECK-NEXT: .LBB36_2: ; %if
; CHECK-NEXT: s_mov_b32 s0, 0
; CHECK-NEXT: s_branch .LBB35_3
; CHECK-NEXT: .LBB35_3:
; CHECK-NEXT: s_branch .LBB36_3
; CHECK-NEXT: .LBB36_3:
%cmp = icmp ne ptr addrspace(4) @1, null
br i1 %cmp, label %endif, label %if

Expand Down