Skip to content

Commit eaf737a

Browse files
committed
[AMDGPU] Remove the GFX11 runs in CodeGen/AMDGPU/fma.f16.ll.
It still fails with expensive checks enabled. This partially reverts: a1e38e0 [AMDGPU][GFX11] Add more test coverage for FMA instructions.
1 parent e791535 commit eaf737a

File tree

1 file changed

+0
-91
lines changed

1 file changed

+0
-91
lines changed

llvm/test/CodeGen/AMDGPU/fma.f16.ll

Lines changed: 0 additions & 91 deletions
Original file line numberDiff line numberDiff line change
@@ -3,8 +3,6 @@
33
; RUN: llc -global-isel=1 -march=amdgcn -mcpu=gfx900 < %s | FileCheck %s -check-prefixes=GFX9,GFX9-GISEL
44
; RUN: llc -global-isel=0 -march=amdgcn -mcpu=gfx1010 < %s | FileCheck %s -check-prefixes=GFX10,GFX10-SDAG
55
; RUN: llc -global-isel=1 -march=amdgcn -mcpu=gfx1010 < %s | FileCheck %s -check-prefixes=GFX10,GFX10-GISEL
6-
; RUN: llc -global-isel=0 -march=amdgcn -mcpu=gfx1100 < %s | FileCheck %s -check-prefixes=GFX11,GFX11-SDAG
7-
; RUN: llc -global-isel=1 -march=amdgcn -mcpu=gfx1100 < %s | FileCheck %s -check-prefixes=GFX11,GFX11-GISEL
86

97
declare half @llvm.fma.f16(half, half, half)
108
declare half @llvm.maxnum.f16(half, half)
@@ -21,12 +19,6 @@ define half @test_fma(half %x, half %y, half %z) {
2119
; GFX10-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
2220
; GFX10-NEXT: v_fma_f16 v0, v0, v1, v2
2321
; GFX10-NEXT: s_setpc_b64 s[30:31]
24-
;
25-
; GFX11-LABEL: test_fma:
26-
; GFX11: ; %bb.0:
27-
; GFX11-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
28-
; GFX11-NEXT: v_fma_f16 v0, v0, v1, v2
29-
; GFX11-NEXT: s_setpc_b64 s[30:31]
3022
%r = call half @llvm.fma.f16(half %x, half %y, half %z)
3123
ret half %r
3224
}
@@ -44,12 +36,6 @@ define half @test_fmac(half %x, half %y, half %z) {
4436
; GFX10-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
4537
; GFX10-NEXT: v_fmac_f16_e32 v0, v1, v2
4638
; GFX10-NEXT: s_setpc_b64 s[30:31]
47-
;
48-
; GFX11-LABEL: test_fmac:
49-
; GFX11: ; %bb.0:
50-
; GFX11-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
51-
; GFX11-NEXT: v_fmac_f16_e32 v0, v1, v2
52-
; GFX11-NEXT: s_setpc_b64 s[30:31]
5339
%r = call half @llvm.fma.f16(half %y, half %z, half %x)
5440
ret half %r
5541
}
@@ -75,12 +61,6 @@ define half @test_fmaak(half %x, half %y, half %z) {
7561
; GFX10-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
7662
; GFX10-NEXT: v_fmaak_f16 v0, v0, v1, 0x4200
7763
; GFX10-NEXT: s_setpc_b64 s[30:31]
78-
;
79-
; GFX11-LABEL: test_fmaak:
80-
; GFX11: ; %bb.0:
81-
; GFX11-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
82-
; GFX11-NEXT: v_fmaak_f16 v0, v0, v1, 0x4200
83-
; GFX11-NEXT: s_setpc_b64 s[30:31]
8464
%r = call half @llvm.fma.f16(half %x, half %y, half 0xH4200)
8565
ret half %r
8666
}
@@ -106,12 +86,6 @@ define half @test_fmamk(half %x, half %y, half %z) {
10686
; GFX10-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
10787
; GFX10-NEXT: v_fmamk_f16 v0, v0, 0x4200, v2
10888
; GFX10-NEXT: s_setpc_b64 s[30:31]
109-
;
110-
; GFX11-LABEL: test_fmamk:
111-
; GFX11: ; %bb.0:
112-
; GFX11-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
113-
; GFX11-NEXT: v_fmamk_f16 v0, v0, 0x4200, v2
114-
; GFX11-NEXT: s_setpc_b64 s[30:31]
11589
%r = call half @llvm.fma.f16(half %x, half 0xH4200, half %z)
11690
ret half %r
11791
}
@@ -165,33 +139,6 @@ define i32 @test_D139469_f16(half %arg) {
165139
; GFX10-GISEL-NEXT: s_or_b32 s4, vcc_lo, s4
166140
; GFX10-GISEL-NEXT: v_cndmask_b32_e64 v0, 0, 1, s4
167141
; GFX10-GISEL-NEXT: s_setpc_b64 s[30:31]
168-
;
169-
; GFX11-SDAG-LABEL: test_D139469_f16:
170-
; GFX11-SDAG: ; %bb.0: ; %bb
171-
; GFX11-SDAG-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
172-
; GFX11-SDAG-NEXT: v_mov_b32_e32 v1, 0x211e
173-
; GFX11-SDAG-NEXT: v_mul_f16_e32 v2, 0x291e, v0
174-
; GFX11-SDAG-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_1)
175-
; GFX11-SDAG-NEXT: v_fmac_f16_e32 v1, 0x291e, v0
176-
; GFX11-SDAG-NEXT: v_min_f16_e32 v0, v2, v1
177-
; GFX11-SDAG-NEXT: s_delay_alu instid0(VALU_DEP_1)
178-
; GFX11-SDAG-NEXT: v_cmp_gt_f16_e32 vcc_lo, 0, v0
179-
; GFX11-SDAG-NEXT: v_cndmask_b32_e64 v0, 0, 1, vcc_lo
180-
; GFX11-SDAG-NEXT: s_setpc_b64 s[30:31]
181-
;
182-
; GFX11-GISEL-LABEL: test_D139469_f16:
183-
; GFX11-GISEL: ; %bb.0: ; %bb
184-
; GFX11-GISEL-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
185-
; GFX11-GISEL-NEXT: s_movk_i32 s0, 0x291e
186-
; GFX11-GISEL-NEXT: v_mul_f16_e32 v1, 0x291e, v0
187-
; GFX11-GISEL-NEXT: v_fmaak_f16 v0, s0, v0, 0x211e
188-
; GFX11-GISEL-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_2)
189-
; GFX11-GISEL-NEXT: v_cmp_gt_f16_e32 vcc_lo, 0, v1
190-
; GFX11-GISEL-NEXT: v_cmp_gt_f16_e64 s0, 0, v0
191-
; GFX11-GISEL-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(SALU_CYCLE_1)
192-
; GFX11-GISEL-NEXT: s_or_b32 s0, vcc_lo, s0
193-
; GFX11-GISEL-NEXT: v_cndmask_b32_e64 v0, 0, 1, s0
194-
; GFX11-GISEL-NEXT: s_setpc_b64 s[30:31]
195142
bb:
196143
%i = fmul contract half %arg, 0xH291E
197144
%i1 = fcmp olt half %i, 0xH0000
@@ -266,44 +213,6 @@ define <2 x i32> @test_D139469_v2f16(<2 x half> %arg) {
266213
; GFX10-GISEL-NEXT: s_or_b32 s4, s6, s5
267214
; GFX10-GISEL-NEXT: v_cndmask_b32_e64 v1, 0, 1, s4
268215
; GFX10-GISEL-NEXT: s_setpc_b64 s[30:31]
269-
;
270-
; GFX11-SDAG-LABEL: test_D139469_v2f16:
271-
; GFX11-SDAG: ; %bb.0: ; %bb
272-
; GFX11-SDAG-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
273-
; GFX11-SDAG-NEXT: s_movk_i32 s0, 0x211e
274-
; GFX11-SDAG-NEXT: v_pk_mul_f16 v1, 0x291e, v0 op_sel_hi:[0,1]
275-
; GFX11-SDAG-NEXT: v_pk_fma_f16 v0, 0x291e, v0, s0 op_sel_hi:[0,1,0]
276-
; GFX11-SDAG-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
277-
; GFX11-SDAG-NEXT: v_pk_min_f16 v0, v1, v0
278-
; GFX11-SDAG-NEXT: v_lshrrev_b32_e32 v1, 16, v0
279-
; GFX11-SDAG-NEXT: v_cmp_gt_f16_e32 vcc_lo, 0, v0
280-
; GFX11-SDAG-NEXT: v_cndmask_b32_e64 v0, 0, 1, vcc_lo
281-
; GFX11-SDAG-NEXT: s_delay_alu instid0(VALU_DEP_3)
282-
; GFX11-SDAG-NEXT: v_cmp_gt_f16_e32 vcc_lo, 0, v1
283-
; GFX11-SDAG-NEXT: v_cndmask_b32_e64 v1, 0, 1, vcc_lo
284-
; GFX11-SDAG-NEXT: s_setpc_b64 s[30:31]
285-
;
286-
; GFX11-GISEL-LABEL: test_D139469_v2f16:
287-
; GFX11-GISEL: ; %bb.0: ; %bb
288-
; GFX11-GISEL-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
289-
; GFX11-GISEL-NEXT: s_mov_b32 s0, 0x291e291e
290-
; GFX11-GISEL-NEXT: v_pk_mul_f16 v1, v0, 0x291e op_sel_hi:[1,0]
291-
; GFX11-GISEL-NEXT: v_pk_fma_f16 v0, v0, s0, 0x211e op_sel_hi:[1,1,0]
292-
; GFX11-GISEL-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_2)
293-
; GFX11-GISEL-NEXT: v_lshrrev_b32_e32 v2, 16, v1
294-
; GFX11-GISEL-NEXT: v_lshrrev_b32_e32 v3, 16, v0
295-
; GFX11-GISEL-NEXT: v_cmp_gt_f16_e32 vcc_lo, 0, v1
296-
; GFX11-GISEL-NEXT: v_cmp_gt_f16_e64 s0, 0, v0
297-
; GFX11-GISEL-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_4)
298-
; GFX11-GISEL-NEXT: v_cmp_gt_f16_e64 s1, 0, v2
299-
; GFX11-GISEL-NEXT: v_cmp_gt_f16_e64 s2, 0, v3
300-
; GFX11-GISEL-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(SALU_CYCLE_1)
301-
; GFX11-GISEL-NEXT: s_or_b32 s0, vcc_lo, s0
302-
; GFX11-GISEL-NEXT: v_cndmask_b32_e64 v0, 0, 1, s0
303-
; GFX11-GISEL-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(SALU_CYCLE_1)
304-
; GFX11-GISEL-NEXT: s_or_b32 s0, s1, s2
305-
; GFX11-GISEL-NEXT: v_cndmask_b32_e64 v1, 0, 1, s0
306-
; GFX11-GISEL-NEXT: s_setpc_b64 s[30:31]
307216
bb:
308217
%i = fmul contract <2 x half> %arg, <half 0xH291E, half 0xH291E>
309218
%i1 = fcmp olt <2 x half> %i, <half 0xH0000, half 0xH0000>

0 commit comments

Comments
 (0)