replica
/
hanchenye-llvm-project
2
0
Fork 0
Code Issues Pull Requests Projects Releases Wiki Activity
hanchenye-llvm-project/llvm/test/CodeGen/AMDGPU/llvm.maxnum.f16.ll

755 lines
28 KiB
LLVM
Raw Normal View History

DAG: Change behavior of fminnum/fmaxnum nodes Introduce new versions that follow the IEEE semantics to help with legalization that may need quieted inputs. There are some regressions from inserting unnecessary canonicalizes when these are matched from fast math fcmp + select which should be fixed in a future commit. llvm-svn: 344914
2018-10-23 00:27:27 +08:00
; NOTE: Assertions have been autogenerated by utils/update_llc_test_checks.py
; RUN: llc -mtriple=amdgcn-- -mcpu=tahiti -verify-machineinstrs < %s | FileCheck -enable-var-scope -check-prefixes=SI %s
; RUN: llc -mtriple=amdgcn-- -mcpu=fiji -mattr=-flat-for-global -verify-machineinstrs < %s | FileCheck -enable-var-scope -check-prefixes=VI,SIVI %s
; RUN: llc -mtriple=amdgcn-- -mcpu=gfx900 -mattr=-flat-for-global -verify-machineinstrs < %s | FileCheck -enable-var-scope -check-prefixes=GFX9 %s
[AMDGPU] Add f16 support (VI+) Differential Revision: https://reviews.llvm.org/D25975 llvm-svn: 286753
2016-11-13 15:01:11 +08:00
declare half @llvm.maxnum.f16(half %a, half %b)
declare <2 x half> @llvm.maxnum.v2f16(<2 x half> %a, <2 x half> %b)
AMDGPU: Fix missing test coverage for some 16-bit and packed ops llvm-svn: 333024
2018-05-23 04:42:00 +08:00
declare <3 x half> @llvm.maxnum.v3f16(<3 x half> %a, <3 x half> %b)
declare <4 x half> @llvm.maxnum.v4f16(<4 x half> %a, <4 x half> %b)
[AMDGPU] Add f16 support (VI+) Differential Revision: https://reviews.llvm.org/D25975 llvm-svn: 286753
2016-11-13 15:01:11 +08:00
AMDGPU: Mark all unspecified CC functions in tests as amdgpu_kernel Currently the default C calling convention functions are treated the same as compute kernels. Make this explicit so the default calling convention can be changed to a non-kernel. Converted with perl -pi -e 's/define void/define amdgpu_kernel void/' on the relevant test directories (and undoing in one place that actually wanted a non-kernel). llvm-svn: 298444
2017-03-22 05:39:51 +08:00
define amdgpu_kernel void @maxnum_f16(
DAG: Change behavior of fminnum/fmaxnum nodes Introduce new versions that follow the IEEE semantics to help with legalization that may need quieted inputs. There are some regressions from inserting unnecessary canonicalizes when these are matched from fast math fcmp + select which should be fixed in a future commit. llvm-svn: 344914
2018-10-23 00:27:27 +08:00
; SI-LABEL: maxnum_f16:
; SI: ; %bb.0: ; %entry
; SI-NEXT: s_load_dwordx4 s[4:7], s[0:1], 0x9
; SI-NEXT: s_load_dwordx2 s[0:1], s[0:1], 0xd
; SI-NEXT: s_mov_b32 s11, 0xf000
; SI-NEXT: s_mov_b32 s10, -1
; SI-NEXT: s_mov_b32 s2, s10
; SI-NEXT: s_mov_b32 s3, s11
; SI-NEXT: s_waitcnt lgkmcnt(0)
; SI-NEXT: s_mov_b32 s12, s6
; SI-NEXT: s_mov_b32 s13, s7
; SI-NEXT: s_mov_b32 s14, s10
; SI-NEXT: s_mov_b32 s15, s11
; SI-NEXT: buffer_load_ushort v0, off, s[12:15], 0
; SI-NEXT: buffer_load_ushort v1, off, s[0:3], 0
; SI-NEXT: s_mov_b32 s8, s4
; SI-NEXT: s_mov_b32 s9, s5
; SI-NEXT: s_waitcnt vmcnt(1)
; SI-NEXT: v_cvt_f32_f16_e32 v0, v0
; SI-NEXT: s_waitcnt vmcnt(0)
; SI-NEXT: v_cvt_f32_f16_e32 v1, v1
; SI-NEXT: v_mul_f32_e32 v0, 1.0, v0
; SI-NEXT: v_mul_f32_e32 v1, 1.0, v1
; SI-NEXT: v_max_f32_e32 v0, v0, v1
; SI-NEXT: v_cvt_f16_f32_e32 v0, v0
; SI-NEXT: buffer_store_short v0, off, s[8:11], 0
; SI-NEXT: s_endpgm
;
; VI-LABEL: maxnum_f16:
; VI: ; %bb.0: ; %entry
; VI-NEXT: s_load_dwordx4 s[4:7], s[0:1], 0x24
; VI-NEXT: s_load_dwordx2 s[8:9], s[0:1], 0x34
; VI-NEXT: s_mov_b32 s3, 0xf000
; VI-NEXT: s_mov_b32 s2, -1
; VI-NEXT: s_mov_b32 s10, s2
; VI-NEXT: s_waitcnt lgkmcnt(0)
; VI-NEXT: s_mov_b32 s0, s4
; VI-NEXT: s_mov_b32 s1, s5
; VI-NEXT: s_mov_b32 s4, s6
; VI-NEXT: s_mov_b32 s5, s7
; VI-NEXT: s_mov_b32 s11, s3
; VI-NEXT: s_mov_b32 s6, s2
; VI-NEXT: s_mov_b32 s7, s3
; VI-NEXT: buffer_load_ushort v0, off, s[4:7], 0
; VI-NEXT: buffer_load_ushort v1, off, s[8:11], 0
; VI-NEXT: s_waitcnt vmcnt(1)
; VI-NEXT: v_max_f16_e32 v0, v0, v0
; VI-NEXT: s_waitcnt vmcnt(0)
; VI-NEXT: v_max_f16_e32 v1, v1, v1
; VI-NEXT: v_max_f16_e32 v0, v0, v1
; VI-NEXT: buffer_store_short v0, off, s[0:3], 0
; VI-NEXT: s_endpgm
;
; GFX9-LABEL: maxnum_f16:
; GFX9: ; %bb.0: ; %entry
; GFX9-NEXT: s_load_dwordx4 s[4:7], s[0:1], 0x24
; GFX9-NEXT: s_load_dwordx2 s[8:9], s[0:1], 0x34
; GFX9-NEXT: s_mov_b32 s3, 0xf000
; GFX9-NEXT: s_mov_b32 s2, -1
; GFX9-NEXT: s_mov_b32 s10, s2
; GFX9-NEXT: s_waitcnt lgkmcnt(0)
; GFX9-NEXT: s_mov_b32 s0, s4
; GFX9-NEXT: s_mov_b32 s1, s5
; GFX9-NEXT: s_mov_b32 s4, s6
; GFX9-NEXT: s_mov_b32 s5, s7
; GFX9-NEXT: s_mov_b32 s11, s3
; GFX9-NEXT: s_mov_b32 s6, s2
; GFX9-NEXT: s_mov_b32 s7, s3
; GFX9-NEXT: buffer_load_ushort v0, off, s[4:7], 0
; GFX9-NEXT: buffer_load_ushort v1, off, s[8:11], 0
; GFX9-NEXT: s_waitcnt vmcnt(1)
; GFX9-NEXT: v_max_f16_e32 v0, v0, v0
; GFX9-NEXT: s_waitcnt vmcnt(0)
; GFX9-NEXT: v_max_f16_e32 v1, v1, v1
; GFX9-NEXT: v_max_f16_e32 v0, v0, v1
; GFX9-NEXT: buffer_store_short v0, off, s[0:3], 0
; GFX9-NEXT: s_endpgm
[AMDGPU] Add f16 support (VI+) Differential Revision: https://reviews.llvm.org/D25975 llvm-svn: 286753
2016-11-13 15:01:11 +08:00
half addrspace(1)* %r,
half addrspace(1)* %a,
half addrspace(1)* %b) {
entry:
AMDGPU: Add pass to lower kernel arguments to loads This replaces most argument uses with loads, but for now not all. The code in SelectionDAG for calling convention lowering is actively harmful for amdgpu_kernel. It attempts to split the argument types into register legal types, which results in low quality code for arbitary types. Since all kernel arguments are passed in memory, we just want the raw types. I've tried a couple of methods of mitigating this in SelectionDAG, but it's easier to just bypass this problem alltogether. It's possible to hack around the problem in the initial lowering, but the real problem is the DAG then expects to be able to use CopyToReg/CopyFromReg for uses of the arguments outside the block. Exposing the argument loads in the IR also has the advantage that the LoadStoreVectorizer can merge them. I'm not sure the best approach to dealing with the IR argument list is. The patch as-is just leaves the IR arguments in place, so all the existing code will still compute the same kernarg size and pointlessly lowers the arguments. Arguably the frontend should emit kernels with an empty argument list in the first place. Alternatively a dummy array could be inserted as a single argument just to reserve space. This does have some disadvantages. Local pointer kernel arguments can no longer have AssertZext placed on them as the equivalent !range metadata is not valid on pointer typed loads. This is mostly bad for SI which needs to know about the known bits in order to use the DS instruction offset, so in this case this is not done. More importantly, this skips noalias arguments since this pass does not yet convert this to the equivalent !alias.scope and !noalias metadata. Producing this metadata correctly seems to be tricky, although this logically is the same as inlining into a function which doesn't exist. Additionally, exposing these loads to the vectorizer may result in degraded aliasing information if a pointer load is merged with another argument load. I'm also not entirely sure this is preserving the current clover ABI, although I would greatly prefer if it would stop widening arguments and match the HSA ABI. As-is I think it is extending < 4-byte arguments to 4-bytes but doesn't align them to 4-bytes. llvm-svn: 335650
2018-06-27 03:10:00 +08:00
%a.val = load volatile half, half addrspace(1)* %a
%b.val = load volatile half, half addrspace(1)* %b
[AMDGPU] Add f16 support (VI+) Differential Revision: https://reviews.llvm.org/D25975 llvm-svn: 286753
2016-11-13 15:01:11 +08:00
%r.val = call half @llvm.maxnum.f16(half %a.val, half %b.val)
store half %r.val, half addrspace(1)* %r
ret void
}
AMDGPU: Mark all unspecified CC functions in tests as amdgpu_kernel Currently the default C calling convention functions are treated the same as compute kernels. Make this explicit so the default calling convention can be changed to a non-kernel. Converted with perl -pi -e 's/define void/define amdgpu_kernel void/' on the relevant test directories (and undoing in one place that actually wanted a non-kernel). llvm-svn: 298444
2017-03-22 05:39:51 +08:00
define amdgpu_kernel void @maxnum_f16_imm_a(
DAG: Change behavior of fminnum/fmaxnum nodes Introduce new versions that follow the IEEE semantics to help with legalization that may need quieted inputs. There are some regressions from inserting unnecessary canonicalizes when these are matched from fast math fcmp + select which should be fixed in a future commit. llvm-svn: 344914
2018-10-23 00:27:27 +08:00
; SI-LABEL: maxnum_f16_imm_a:
; SI: ; %bb.0: ; %entry
; SI-NEXT: s_load_dwordx4 s[4:7], s[0:1], 0x9
; SI-NEXT: s_mov_b32 s3, 0xf000
; SI-NEXT: s_mov_b32 s2, -1
; SI-NEXT: s_mov_b32 s10, s2
; SI-NEXT: s_mov_b32 s11, s3
; SI-NEXT: s_waitcnt lgkmcnt(0)
; SI-NEXT: s_mov_b32 s8, s6
; SI-NEXT: s_mov_b32 s9, s7
; SI-NEXT: buffer_load_ushort v0, off, s[8:11], 0
; SI-NEXT: s_mov_b32 s0, s4
; SI-NEXT: s_mov_b32 s1, s5
; SI-NEXT: s_waitcnt vmcnt(0)
; SI-NEXT: v_cvt_f32_f16_e32 v0, v0
; SI-NEXT: v_mul_f32_e32 v0, 1.0, v0
; SI-NEXT: v_max_f32_e32 v0, 0x40400000, v0
; SI-NEXT: v_cvt_f16_f32_e32 v0, v0
; SI-NEXT: buffer_store_short v0, off, s[0:3], 0
; SI-NEXT: s_endpgm
;
; VI-LABEL: maxnum_f16_imm_a:
; VI: ; %bb.0: ; %entry
; VI-NEXT: s_load_dwordx4 s[4:7], s[0:1], 0x24
; VI-NEXT: s_mov_b32 s3, 0xf000
; VI-NEXT: s_mov_b32 s2, -1
; VI-NEXT: s_waitcnt lgkmcnt(0)
; VI-NEXT: s_mov_b32 s0, s4
; VI-NEXT: s_mov_b32 s1, s5
; VI-NEXT: s_mov_b32 s4, s6
; VI-NEXT: s_mov_b32 s5, s7
; VI-NEXT: s_mov_b32 s6, s2
; VI-NEXT: s_mov_b32 s7, s3
; VI-NEXT: buffer_load_ushort v0, off, s[4:7], 0
; VI-NEXT: s_waitcnt vmcnt(0)
; VI-NEXT: v_max_f16_e32 v0, v0, v0
; VI-NEXT: v_max_f16_e32 v0, 0x4200, v0
; VI-NEXT: buffer_store_short v0, off, s[0:3], 0
; VI-NEXT: s_endpgm
;
; GFX9-LABEL: maxnum_f16_imm_a:
; GFX9: ; %bb.0: ; %entry
; GFX9-NEXT: s_load_dwordx4 s[4:7], s[0:1], 0x24
; GFX9-NEXT: s_mov_b32 s3, 0xf000
; GFX9-NEXT: s_mov_b32 s2, -1
; GFX9-NEXT: s_waitcnt lgkmcnt(0)
; GFX9-NEXT: s_mov_b32 s0, s4
; GFX9-NEXT: s_mov_b32 s1, s5
; GFX9-NEXT: s_mov_b32 s4, s6
; GFX9-NEXT: s_mov_b32 s5, s7
; GFX9-NEXT: s_mov_b32 s6, s2
; GFX9-NEXT: s_mov_b32 s7, s3
; GFX9-NEXT: buffer_load_ushort v0, off, s[4:7], 0
; GFX9-NEXT: s_waitcnt vmcnt(0)
; GFX9-NEXT: v_max_f16_e32 v0, v0, v0
; GFX9-NEXT: v_max_f16_e32 v0, 0x4200, v0
; GFX9-NEXT: buffer_store_short v0, off, s[0:3], 0
; GFX9-NEXT: s_endpgm
[AMDGPU] Add f16 support (VI+) Differential Revision: https://reviews.llvm.org/D25975 llvm-svn: 286753
2016-11-13 15:01:11 +08:00
half addrspace(1)* %r,
half addrspace(1)* %b) {
entry:
%b.val = load half, half addrspace(1)* %b
%r.val = call half @llvm.maxnum.f16(half 3.0, half %b.val)
store half %r.val, half addrspace(1)* %r
ret void
}
AMDGPU: Mark all unspecified CC functions in tests as amdgpu_kernel Currently the default C calling convention functions are treated the same as compute kernels. Make this explicit so the default calling convention can be changed to a non-kernel. Converted with perl -pi -e 's/define void/define amdgpu_kernel void/' on the relevant test directories (and undoing in one place that actually wanted a non-kernel). llvm-svn: 298444
2017-03-22 05:39:51 +08:00
define amdgpu_kernel void @maxnum_f16_imm_b(
DAG: Change behavior of fminnum/fmaxnum nodes Introduce new versions that follow the IEEE semantics to help with legalization that may need quieted inputs. There are some regressions from inserting unnecessary canonicalizes when these are matched from fast math fcmp + select which should be fixed in a future commit. llvm-svn: 344914
2018-10-23 00:27:27 +08:00
; SI-LABEL: maxnum_f16_imm_b:
; SI: ; %bb.0: ; %entry
; SI-NEXT: s_load_dwordx4 s[4:7], s[0:1], 0x9
; SI-NEXT: s_mov_b32 s3, 0xf000
; SI-NEXT: s_mov_b32 s2, -1
; SI-NEXT: s_mov_b32 s10, s2
; SI-NEXT: s_mov_b32 s11, s3
; SI-NEXT: s_waitcnt lgkmcnt(0)
; SI-NEXT: s_mov_b32 s8, s6
; SI-NEXT: s_mov_b32 s9, s7
; SI-NEXT: buffer_load_ushort v0, off, s[8:11], 0
; SI-NEXT: s_mov_b32 s0, s4
; SI-NEXT: s_mov_b32 s1, s5
; SI-NEXT: s_waitcnt vmcnt(0)
; SI-NEXT: v_cvt_f32_f16_e32 v0, v0
; SI-NEXT: v_mul_f32_e32 v0, 1.0, v0
; SI-NEXT: v_max_f32_e32 v0, 4.0, v0
; SI-NEXT: v_cvt_f16_f32_e32 v0, v0
; SI-NEXT: buffer_store_short v0, off, s[0:3], 0
; SI-NEXT: s_endpgm
;
; VI-LABEL: maxnum_f16_imm_b:
; VI: ; %bb.0: ; %entry
; VI-NEXT: s_load_dwordx4 s[4:7], s[0:1], 0x24
; VI-NEXT: s_mov_b32 s3, 0xf000
; VI-NEXT: s_mov_b32 s2, -1
; VI-NEXT: s_waitcnt lgkmcnt(0)
; VI-NEXT: s_mov_b32 s0, s4
; VI-NEXT: s_mov_b32 s1, s5
; VI-NEXT: s_mov_b32 s4, s6
; VI-NEXT: s_mov_b32 s5, s7
; VI-NEXT: s_mov_b32 s6, s2
; VI-NEXT: s_mov_b32 s7, s3
; VI-NEXT: buffer_load_ushort v0, off, s[4:7], 0
; VI-NEXT: s_waitcnt vmcnt(0)
; VI-NEXT: v_max_f16_e32 v0, v0, v0
; VI-NEXT: v_max_f16_e32 v0, 4.0, v0
; VI-NEXT: buffer_store_short v0, off, s[0:3], 0
; VI-NEXT: s_endpgm
;
; GFX9-LABEL: maxnum_f16_imm_b:
; GFX9: ; %bb.0: ; %entry
; GFX9-NEXT: s_load_dwordx4 s[4:7], s[0:1], 0x24
; GFX9-NEXT: s_mov_b32 s3, 0xf000
; GFX9-NEXT: s_mov_b32 s2, -1
; GFX9-NEXT: s_waitcnt lgkmcnt(0)
; GFX9-NEXT: s_mov_b32 s0, s4
; GFX9-NEXT: s_mov_b32 s1, s5
; GFX9-NEXT: s_mov_b32 s4, s6
; GFX9-NEXT: s_mov_b32 s5, s7
; GFX9-NEXT: s_mov_b32 s6, s2
; GFX9-NEXT: s_mov_b32 s7, s3
; GFX9-NEXT: buffer_load_ushort v0, off, s[4:7], 0
; GFX9-NEXT: s_waitcnt vmcnt(0)
; GFX9-NEXT: v_max_f16_e32 v0, v0, v0
; GFX9-NEXT: v_max_f16_e32 v0, 4.0, v0
; GFX9-NEXT: buffer_store_short v0, off, s[0:3], 0
; GFX9-NEXT: s_endpgm
[AMDGPU] Add f16 support (VI+) Differential Revision: https://reviews.llvm.org/D25975 llvm-svn: 286753
2016-11-13 15:01:11 +08:00
half addrspace(1)* %r,
half addrspace(1)* %a) {
entry:
%a.val = load half, half addrspace(1)* %a
%r.val = call half @llvm.maxnum.f16(half %a.val, half 4.0)
store half %r.val, half addrspace(1)* %r
ret void
}
AMDGPU: Mark all unspecified CC functions in tests as amdgpu_kernel Currently the default C calling convention functions are treated the same as compute kernels. Make this explicit so the default calling convention can be changed to a non-kernel. Converted with perl -pi -e 's/define void/define amdgpu_kernel void/' on the relevant test directories (and undoing in one place that actually wanted a non-kernel). llvm-svn: 298444
2017-03-22 05:39:51 +08:00
define amdgpu_kernel void @maxnum_v2f16(
DAG: Change behavior of fminnum/fmaxnum nodes Introduce new versions that follow the IEEE semantics to help with legalization that may need quieted inputs. There are some regressions from inserting unnecessary canonicalizes when these are matched from fast math fcmp + select which should be fixed in a future commit. llvm-svn: 344914
2018-10-23 00:27:27 +08:00
; SI-LABEL: maxnum_v2f16:
; SI: ; %bb.0: ; %entry
; SI-NEXT: s_load_dwordx4 s[4:7], s[0:1], 0x9
; SI-NEXT: s_load_dwordx2 s[0:1], s[0:1], 0xd
; SI-NEXT: s_mov_b32 s3, 0xf000
; SI-NEXT: s_mov_b32 s2, -1
; SI-NEXT: s_waitcnt lgkmcnt(0)
; SI-NEXT: s_load_dword s6, s[6:7], 0x0
; SI-NEXT: s_load_dword s0, s[0:1], 0x0
; SI-NEXT: s_waitcnt lgkmcnt(0)
; SI-NEXT: s_lshr_b32 s1, s6, 16
; SI-NEXT: v_cvt_f32_f16_e32 v1, s0
; SI-NEXT: s_lshr_b32 s0, s0, 16
; SI-NEXT: v_cvt_f32_f16_e32 v2, s0
; SI-NEXT: v_cvt_f32_f16_e32 v3, s1
; SI-NEXT: v_cvt_f32_f16_e32 v0, s6
; SI-NEXT: v_mul_f32_e32 v1, 1.0, v1
; SI-NEXT: v_mul_f32_e32 v2, 1.0, v2
; SI-NEXT: v_mul_f32_e32 v3, 1.0, v3
; SI-NEXT: v_mul_f32_e32 v0, 1.0, v0
; SI-NEXT: v_max_f32_e32 v2, v3, v2
; SI-NEXT: v_max_f32_e32 v0, v0, v1
; SI-NEXT: v_cvt_f16_f32_e32 v2, v2
; SI-NEXT: v_cvt_f16_f32_e32 v0, v0
; SI-NEXT: s_mov_b32 s0, s4
; SI-NEXT: s_mov_b32 s1, s5
; SI-NEXT: v_lshlrev_b32_e32 v1, 16, v2
; SI-NEXT: v_or_b32_e32 v0, v0, v1
; SI-NEXT: buffer_store_dword v0, off, s[0:3], 0
; SI-NEXT: s_endpgm
;
; VI-LABEL: maxnum_v2f16:
; VI: ; %bb.0: ; %entry
; VI-NEXT: s_load_dwordx4 s[4:7], s[0:1], 0x24
; VI-NEXT: s_load_dwordx2 s[8:9], s[0:1], 0x34
; VI-NEXT: s_mov_b32 s3, 0xf000
; VI-NEXT: s_mov_b32 s2, -1
; VI-NEXT: s_waitcnt lgkmcnt(0)
; VI-NEXT: s_mov_b32 s0, s4
; VI-NEXT: s_mov_b32 s1, s5
; VI-NEXT: s_load_dword s4, s[6:7], 0x0
; VI-NEXT: s_load_dword s5, s[8:9], 0x0
; VI-NEXT: s_waitcnt lgkmcnt(0)
; VI-NEXT: v_max_f16_e64 v1, s4, s4
; VI-NEXT: v_max_f16_e64 v0, s5, s5
; VI-NEXT: s_lshr_b32 s4, s4, 16
; VI-NEXT: s_lshr_b32 s5, s5, 16
; VI-NEXT: v_max_f16_e32 v0, v1, v0
; VI-NEXT: v_max_f16_e64 v1, s5, s5
; VI-NEXT: v_max_f16_e64 v2, s4, s4
; VI-NEXT: v_max_f16_sdwa v1, v2, v1 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
; VI-NEXT: v_or_b32_sdwa v0, v0, v1 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
; VI-NEXT: buffer_store_dword v0, off, s[0:3], 0
; VI-NEXT: s_endpgm
;
; GFX9-LABEL: maxnum_v2f16:
; GFX9: ; %bb.0: ; %entry
; GFX9-NEXT: s_load_dwordx4 s[4:7], s[0:1], 0x24
; GFX9-NEXT: s_load_dwordx2 s[8:9], s[0:1], 0x34
; GFX9-NEXT: s_mov_b32 s3, 0xf000
; GFX9-NEXT: s_mov_b32 s2, -1
; GFX9-NEXT: s_waitcnt lgkmcnt(0)
; GFX9-NEXT: s_mov_b32 s0, s4
; GFX9-NEXT: s_mov_b32 s1, s5
; GFX9-NEXT: s_load_dword s4, s[6:7], 0x0
; GFX9-NEXT: s_load_dword s5, s[8:9], 0x0
; GFX9-NEXT: s_waitcnt lgkmcnt(0)
; GFX9-NEXT: v_pk_max_f16 v1, s4, s4
; GFX9-NEXT: v_pk_max_f16 v0, s5, s5
; GFX9-NEXT: v_pk_max_f16 v0, v1, v0
; GFX9-NEXT: buffer_store_dword v0, off, s[0:3], 0
; GFX9-NEXT: s_endpgm
[AMDGPU] Add f16 support (VI+) Differential Revision: https://reviews.llvm.org/D25975 llvm-svn: 286753
2016-11-13 15:01:11 +08:00
<2 x half> addrspace(1)* %r,
<2 x half> addrspace(1)* %a,
<2 x half> addrspace(1)* %b) {
entry:
%a.val = load <2 x half>, <2 x half> addrspace(1)* %a
%b.val = load <2 x half>, <2 x half> addrspace(1)* %b
%r.val = call <2 x half> @llvm.maxnum.v2f16(<2 x half> %a.val, <2 x half> %b.val)
store <2 x half> %r.val, <2 x half> addrspace(1)* %r
ret void
}
AMDGPU: Mark all unspecified CC functions in tests as amdgpu_kernel Currently the default C calling convention functions are treated the same as compute kernels. Make this explicit so the default calling convention can be changed to a non-kernel. Converted with perl -pi -e 's/define void/define amdgpu_kernel void/' on the relevant test directories (and undoing in one place that actually wanted a non-kernel). llvm-svn: 298444
2017-03-22 05:39:51 +08:00
define amdgpu_kernel void @maxnum_v2f16_imm_a(
DAG: Change behavior of fminnum/fmaxnum nodes Introduce new versions that follow the IEEE semantics to help with legalization that may need quieted inputs. There are some regressions from inserting unnecessary canonicalizes when these are matched from fast math fcmp + select which should be fixed in a future commit. llvm-svn: 344914
2018-10-23 00:27:27 +08:00
; SI-LABEL: maxnum_v2f16_imm_a:
; SI: ; %bb.0: ; %entry
; SI-NEXT: s_load_dwordx4 s[0:3], s[0:1], 0x9
; SI-NEXT: s_waitcnt lgkmcnt(0)
; SI-NEXT: s_load_dword s2, s[2:3], 0x0
; SI-NEXT: s_mov_b32 s3, 0xf000
; SI-NEXT: s_waitcnt lgkmcnt(0)
; SI-NEXT: v_cvt_f32_f16_e32 v0, s2
; SI-NEXT: s_lshr_b32 s2, s2, 16
; SI-NEXT: v_cvt_f32_f16_e32 v1, s2
; SI-NEXT: s_mov_b32 s2, -1
; SI-NEXT: v_mul_f32_e32 v0, 1.0, v0
; SI-NEXT: v_max_f32_e32 v0, 0x40400000, v0
; SI-NEXT: v_mul_f32_e32 v1, 1.0, v1
; SI-NEXT: v_max_f32_e32 v1, 4.0, v1
; SI-NEXT: v_cvt_f16_f32_e32 v1, v1
; SI-NEXT: v_cvt_f16_f32_e32 v0, v0
; SI-NEXT: v_lshlrev_b32_e32 v1, 16, v1
; SI-NEXT: v_or_b32_e32 v0, v0, v1
; SI-NEXT: buffer_store_dword v0, off, s[0:3], 0
; SI-NEXT: s_endpgm
;
; VI-LABEL: maxnum_v2f16_imm_a:
; VI: ; %bb.0: ; %entry
; VI-NEXT: s_load_dwordx4 s[4:7], s[0:1], 0x24
; VI-NEXT: v_mov_b32_e32 v2, 0x4400
; VI-NEXT: s_mov_b32 s3, 0xf000
; VI-NEXT: s_mov_b32 s2, -1
; VI-NEXT: s_waitcnt lgkmcnt(0)
; VI-NEXT: s_mov_b32 s0, s4
; VI-NEXT: s_load_dword s4, s[6:7], 0x0
; VI-NEXT: s_mov_b32 s1, s5
; VI-NEXT: s_waitcnt lgkmcnt(0)
; VI-NEXT: v_max_f16_e64 v0, s4, s4
; VI-NEXT: s_lshr_b32 s4, s4, 16
; VI-NEXT: v_max_f16_e64 v1, s4, s4
; VI-NEXT: v_max_f16_e32 v0, 0x4200, v0
; VI-NEXT: v_max_f16_sdwa v1, v1, v2 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
; VI-NEXT: v_or_b32_sdwa v0, v0, v1 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
; VI-NEXT: buffer_store_dword v0, off, s[0:3], 0
; VI-NEXT: s_endpgm
;
; GFX9-LABEL: maxnum_v2f16_imm_a:
; GFX9: ; %bb.0: ; %entry
; GFX9-NEXT: s_load_dwordx4 s[4:7], s[0:1], 0x24
; GFX9-NEXT: s_mov_b32 s3, 0xf000
; GFX9-NEXT: s_mov_b32 s2, -1
; GFX9-NEXT: s_waitcnt lgkmcnt(0)
; GFX9-NEXT: s_mov_b32 s0, s4
; GFX9-NEXT: s_load_dword s4, s[6:7], 0x0
; GFX9-NEXT: s_mov_b32 s1, s5
; GFX9-NEXT: s_waitcnt lgkmcnt(0)
; GFX9-NEXT: v_pk_max_f16 v0, s4, s4
; GFX9-NEXT: s_mov_b32 s4, 0x44004200
; GFX9-NEXT: v_pk_max_f16 v0, v0, s4
; GFX9-NEXT: buffer_store_dword v0, off, s[0:3], 0
; GFX9-NEXT: s_endpgm
[AMDGPU] Add f16 support (VI+) Differential Revision: https://reviews.llvm.org/D25975 llvm-svn: 286753
2016-11-13 15:01:11 +08:00
<2 x half> addrspace(1)* %r,
<2 x half> addrspace(1)* %b) {
entry:
%b.val = load <2 x half>, <2 x half> addrspace(1)* %b
%r.val = call <2 x half> @llvm.maxnum.v2f16(<2 x half> <half 3.0, half 4.0>, <2 x half> %b.val)
store <2 x half> %r.val, <2 x half> addrspace(1)* %r
ret void
}
AMDGPU: Mark all unspecified CC functions in tests as amdgpu_kernel Currently the default C calling convention functions are treated the same as compute kernels. Make this explicit so the default calling convention can be changed to a non-kernel. Converted with perl -pi -e 's/define void/define amdgpu_kernel void/' on the relevant test directories (and undoing in one place that actually wanted a non-kernel). llvm-svn: 298444
2017-03-22 05:39:51 +08:00
define amdgpu_kernel void @maxnum_v2f16_imm_b(
DAG: Change behavior of fminnum/fmaxnum nodes Introduce new versions that follow the IEEE semantics to help with legalization that may need quieted inputs. There are some regressions from inserting unnecessary canonicalizes when these are matched from fast math fcmp + select which should be fixed in a future commit. llvm-svn: 344914
2018-10-23 00:27:27 +08:00
; SI-LABEL: maxnum_v2f16_imm_b:
; SI: ; %bb.0: ; %entry
; SI-NEXT: s_load_dwordx4 s[0:3], s[0:1], 0x9
; SI-NEXT: s_waitcnt lgkmcnt(0)
; SI-NEXT: s_load_dword s2, s[2:3], 0x0
; SI-NEXT: s_mov_b32 s3, 0xf000
; SI-NEXT: s_waitcnt lgkmcnt(0)
; SI-NEXT: v_cvt_f32_f16_e32 v0, s2
; SI-NEXT: s_lshr_b32 s2, s2, 16
; SI-NEXT: v_cvt_f32_f16_e32 v1, s2
; SI-NEXT: s_mov_b32 s2, -1
; SI-NEXT: v_mul_f32_e32 v0, 1.0, v0
; SI-NEXT: v_max_f32_e32 v0, 4.0, v0
; SI-NEXT: v_mul_f32_e32 v1, 1.0, v1
; SI-NEXT: v_max_f32_e32 v1, 0x40400000, v1
; SI-NEXT: v_cvt_f16_f32_e32 v1, v1
; SI-NEXT: v_cvt_f16_f32_e32 v0, v0
; SI-NEXT: v_lshlrev_b32_e32 v1, 16, v1
; SI-NEXT: v_or_b32_e32 v0, v0, v1
; SI-NEXT: buffer_store_dword v0, off, s[0:3], 0
; SI-NEXT: s_endpgm
;
; VI-LABEL: maxnum_v2f16_imm_b:
; VI: ; %bb.0: ; %entry
; VI-NEXT: s_load_dwordx4 s[4:7], s[0:1], 0x24
; VI-NEXT: v_mov_b32_e32 v2, 0x4200
; VI-NEXT: s_mov_b32 s3, 0xf000
; VI-NEXT: s_mov_b32 s2, -1
; VI-NEXT: s_waitcnt lgkmcnt(0)
; VI-NEXT: s_mov_b32 s0, s4
; VI-NEXT: s_load_dword s4, s[6:7], 0x0
; VI-NEXT: s_mov_b32 s1, s5
; VI-NEXT: s_waitcnt lgkmcnt(0)
; VI-NEXT: v_max_f16_e64 v0, s4, s4
; VI-NEXT: s_lshr_b32 s4, s4, 16
; VI-NEXT: v_max_f16_e64 v1, s4, s4
; VI-NEXT: v_max_f16_e32 v0, 4.0, v0
; VI-NEXT: v_max_f16_sdwa v1, v1, v2 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
; VI-NEXT: v_or_b32_sdwa v0, v0, v1 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
; VI-NEXT: buffer_store_dword v0, off, s[0:3], 0
; VI-NEXT: s_endpgm
;
; GFX9-LABEL: maxnum_v2f16_imm_b:
; GFX9: ; %bb.0: ; %entry
; GFX9-NEXT: s_load_dwordx4 s[4:7], s[0:1], 0x24
; GFX9-NEXT: s_mov_b32 s3, 0xf000
; GFX9-NEXT: s_mov_b32 s2, -1
; GFX9-NEXT: s_waitcnt lgkmcnt(0)
; GFX9-NEXT: s_mov_b32 s0, s4
; GFX9-NEXT: s_load_dword s4, s[6:7], 0x0
; GFX9-NEXT: s_mov_b32 s1, s5
; GFX9-NEXT: s_waitcnt lgkmcnt(0)
; GFX9-NEXT: v_pk_max_f16 v0, s4, s4
; GFX9-NEXT: s_mov_b32 s4, 0x42004400
; GFX9-NEXT: v_pk_max_f16 v0, v0, s4
; GFX9-NEXT: buffer_store_dword v0, off, s[0:3], 0
; GFX9-NEXT: s_endpgm
[AMDGPU] Add f16 support (VI+) Differential Revision: https://reviews.llvm.org/D25975 llvm-svn: 286753
2016-11-13 15:01:11 +08:00
<2 x half> addrspace(1)* %r,
<2 x half> addrspace(1)* %a) {
entry:
%a.val = load <2 x half>, <2 x half> addrspace(1)* %a
%r.val = call <2 x half> @llvm.maxnum.v2f16(<2 x half> %a.val, <2 x half> <half 4.0, half 3.0>)
store <2 x half> %r.val, <2 x half> addrspace(1)* %r
ret void
}
AMDGPU: Fix missing test coverage for some 16-bit and packed ops llvm-svn: 333024
2018-05-23 04:42:00 +08:00
; FIXME: Scalarize with undef half
define amdgpu_kernel void @maxnum_v3f16(
DAG: Change behavior of fminnum/fmaxnum nodes Introduce new versions that follow the IEEE semantics to help with legalization that may need quieted inputs. There are some regressions from inserting unnecessary canonicalizes when these are matched from fast math fcmp + select which should be fixed in a future commit. llvm-svn: 344914
2018-10-23 00:27:27 +08:00
; SI-LABEL: maxnum_v3f16:
; SI: ; %bb.0: ; %entry
; SI-NEXT: s_load_dwordx4 s[4:7], s[0:1], 0x9
; SI-NEXT: s_load_dwordx2 s[8:9], s[0:1], 0xd
; SI-NEXT: s_mov_b32 s3, 0xf000
; SI-NEXT: s_mov_b32 s2, -1
; SI-NEXT: s_waitcnt lgkmcnt(0)
; SI-NEXT: s_load_dwordx2 s[6:7], s[6:7], 0x0
; SI-NEXT: s_load_dwordx2 s[8:9], s[8:9], 0x0
; SI-NEXT: s_mov_b32 s0, s4
; SI-NEXT: s_waitcnt lgkmcnt(0)
; SI-NEXT: s_lshr_b32 s1, s6, 16
; SI-NEXT: s_lshr_b32 s4, s8, 16
; SI-NEXT: v_cvt_f32_f16_e32 v3, s1
; SI-NEXT: v_cvt_f32_f16_e32 v2, s4
; SI-NEXT: v_cvt_f32_f16_e32 v1, s6
; SI-NEXT: v_cvt_f32_f16_e32 v5, s8
; SI-NEXT: v_cvt_f32_f16_e32 v0, s7
; SI-NEXT: v_cvt_f32_f16_e32 v4, s9
; SI-NEXT: v_mul_f32_e32 v2, 1.0, v2
; SI-NEXT: v_mul_f32_e32 v3, 1.0, v3
; SI-NEXT: v_max_f32_e32 v2, v3, v2
; SI-NEXT: v_mul_f32_e32 v3, 1.0, v5
; SI-NEXT: v_mul_f32_e32 v1, 1.0, v1
; SI-NEXT: v_max_f32_e32 v1, v1, v3
; SI-NEXT: v_cvt_f16_f32_e32 v2, v2
; SI-NEXT: v_mul_f32_e32 v3, 1.0, v4
; SI-NEXT: v_mul_f32_e32 v0, 1.0, v0
; SI-NEXT: v_max_f32_e32 v0, v0, v3
; SI-NEXT: v_cvt_f16_f32_e32 v1, v1
; SI-NEXT: v_cvt_f16_f32_e32 v0, v0
; SI-NEXT: v_lshlrev_b32_e32 v2, 16, v2
; SI-NEXT: s_mov_b32 s1, s5
; SI-NEXT: v_or_b32_e32 v1, v1, v2
; SI-NEXT: buffer_store_short v0, off, s[0:3], 0 offset:4
; SI-NEXT: buffer_store_dword v1, off, s[0:3], 0
; SI-NEXT: s_endpgm
;
; VI-LABEL: maxnum_v3f16:
; VI: ; %bb.0: ; %entry
; VI-NEXT: s_load_dwordx4 s[4:7], s[0:1], 0x24
; VI-NEXT: s_load_dwordx2 s[8:9], s[0:1], 0x34
; VI-NEXT: s_mov_b32 s3, 0xf000
; VI-NEXT: s_mov_b32 s2, -1
; VI-NEXT: s_waitcnt lgkmcnt(0)
; VI-NEXT: s_mov_b32 s0, s4
; VI-NEXT: s_mov_b32 s1, s5
; VI-NEXT: s_load_dwordx2 s[4:5], s[6:7], 0x0
; VI-NEXT: s_load_dwordx2 s[6:7], s[8:9], 0x0
; VI-NEXT: s_waitcnt lgkmcnt(0)
; VI-NEXT: v_max_f16_e64 v1, s4, s4
; VI-NEXT: v_max_f16_e64 v0, s6, s6
; VI-NEXT: s_lshr_b32 s4, s4, 16
; VI-NEXT: s_lshr_b32 s6, s6, 16
; VI-NEXT: v_max_f16_e32 v0, v1, v0
; VI-NEXT: v_max_f16_e64 v1, s6, s6
; VI-NEXT: v_max_f16_e64 v2, s4, s4
; VI-NEXT: v_max_f16_sdwa v1, v2, v1 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
; VI-NEXT: v_or_b32_sdwa v0, v0, v1 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
; VI-NEXT: v_max_f16_e64 v1, s7, s7
; VI-NEXT: v_max_f16_e64 v2, s5, s5
; VI-NEXT: v_max_f16_e32 v1, v2, v1
; VI-NEXT: buffer_store_short v1, off, s[0:3], 0 offset:4
; VI-NEXT: buffer_store_dword v0, off, s[0:3], 0
; VI-NEXT: s_endpgm
;
; GFX9-LABEL: maxnum_v3f16:
; GFX9: ; %bb.0: ; %entry
; GFX9-NEXT: s_load_dwordx4 s[4:7], s[0:1], 0x24
; GFX9-NEXT: s_load_dwordx2 s[8:9], s[0:1], 0x34
; GFX9-NEXT: s_mov_b32 s3, 0xf000
; GFX9-NEXT: s_mov_b32 s2, -1
; GFX9-NEXT: s_waitcnt lgkmcnt(0)
; GFX9-NEXT: s_mov_b32 s0, s4
; GFX9-NEXT: s_mov_b32 s1, s5
; GFX9-NEXT: s_load_dwordx2 s[4:5], s[6:7], 0x0
; GFX9-NEXT: s_load_dwordx2 s[6:7], s[8:9], 0x0
; GFX9-NEXT: s_waitcnt lgkmcnt(0)
; GFX9-NEXT: v_pk_max_f16 v1, s4, s4
; GFX9-NEXT: v_pk_max_f16 v0, s6, s6
; GFX9-NEXT: v_pk_max_f16 v0, v1, v0
; GFX9-NEXT: v_pk_max_f16 v2, s7, s7
; GFX9-NEXT: v_pk_max_f16 v1, s5, s5
; GFX9-NEXT: v_pk_max_f16 v1, v1, v2
; GFX9-NEXT: buffer_store_short v1, off, s[0:3], 0 offset:4
; GFX9-NEXT: buffer_store_dword v0, off, s[0:3], 0
; GFX9-NEXT: s_endpgm
AMDGPU: Fix missing test coverage for some 16-bit and packed ops llvm-svn: 333024
2018-05-23 04:42:00 +08:00
<3 x half> addrspace(1)* %r,
<3 x half> addrspace(1)* %a,
<3 x half> addrspace(1)* %b) {
entry:
%a.val = load <3 x half>, <3 x half> addrspace(1)* %a
%b.val = load <3 x half>, <3 x half> addrspace(1)* %b
%r.val = call <3 x half> @llvm.maxnum.v3f16(<3 x half> %a.val, <3 x half> %b.val)
store <3 x half> %r.val, <3 x half> addrspace(1)* %r
ret void
}
define amdgpu_kernel void @maxnum_v4f16(
DAG: Change behavior of fminnum/fmaxnum nodes Introduce new versions that follow the IEEE semantics to help with legalization that may need quieted inputs. There are some regressions from inserting unnecessary canonicalizes when these are matched from fast math fcmp + select which should be fixed in a future commit. llvm-svn: 344914
2018-10-23 00:27:27 +08:00
; SI-LABEL: maxnum_v4f16:
; SI: ; %bb.0: ; %entry
; SI-NEXT: s_load_dwordx4 s[4:7], s[0:1], 0x9
; SI-NEXT: s_load_dwordx2 s[8:9], s[0:1], 0xd
; SI-NEXT: s_mov_b32 s3, 0xf000
; SI-NEXT: s_mov_b32 s2, -1
; SI-NEXT: s_waitcnt lgkmcnt(0)
; SI-NEXT: s_mov_b32 s0, s4
; SI-NEXT: s_mov_b32 s1, s5
; SI-NEXT: s_load_dwordx2 s[4:5], s[6:7], 0x0
; SI-NEXT: s_load_dwordx2 s[6:7], s[8:9], 0x0
; SI-NEXT: s_waitcnt lgkmcnt(0)
; SI-NEXT: v_cvt_f32_f16_e32 v0, s4
; SI-NEXT: s_lshr_b32 s4, s4, 16
; SI-NEXT: v_cvt_f32_f16_e32 v2, s4
; SI-NEXT: s_lshr_b32 s4, s5, 16
; SI-NEXT: v_cvt_f32_f16_e32 v3, s4
; SI-NEXT: s_lshr_b32 s4, s7, 16
; SI-NEXT: v_cvt_f32_f16_e32 v5, s4
; SI-NEXT: v_cvt_f32_f16_e32 v1, s5
; SI-NEXT: s_lshr_b32 s4, s6, 16
; SI-NEXT: v_cvt_f32_f16_e32 v7, s7
; SI-NEXT: v_cvt_f32_f16_e32 v6, s4
; SI-NEXT: v_cvt_f32_f16_e32 v4, s6
; SI-NEXT: v_mul_f32_e32 v5, 1.0, v5
; SI-NEXT: v_mul_f32_e32 v3, 1.0, v3
; SI-NEXT: v_max_f32_e32 v3, v3, v5
; SI-NEXT: v_mul_f32_e32 v5, 1.0, v7
; SI-NEXT: v_mul_f32_e32 v1, 1.0, v1
; SI-NEXT: v_max_f32_e32 v1, v1, v5
; SI-NEXT: v_mul_f32_e32 v5, 1.0, v6
; SI-NEXT: v_mul_f32_e32 v2, 1.0, v2
; SI-NEXT: v_max_f32_e32 v2, v2, v5
; SI-NEXT: v_mul_f32_e32 v4, 1.0, v4
; SI-NEXT: v_mul_f32_e32 v0, 1.0, v0
; SI-NEXT: v_cvt_f16_f32_e32 v3, v3
; SI-NEXT: v_cvt_f16_f32_e32 v2, v2
; SI-NEXT: v_max_f32_e32 v0, v0, v4
; SI-NEXT: v_cvt_f16_f32_e32 v1, v1
; SI-NEXT: v_cvt_f16_f32_e32 v0, v0
; SI-NEXT: v_lshlrev_b32_e32 v3, 16, v3
; SI-NEXT: v_lshlrev_b32_e32 v2, 16, v2
; SI-NEXT: v_or_b32_e32 v1, v1, v3
; SI-NEXT: v_or_b32_e32 v0, v0, v2
; SI-NEXT: buffer_store_dwordx2 v[0:1], off, s[0:3], 0
; SI-NEXT: s_endpgm
;
; VI-LABEL: maxnum_v4f16:
; VI: ; %bb.0: ; %entry
; VI-NEXT: s_load_dwordx4 s[4:7], s[0:1], 0x24
; VI-NEXT: s_load_dwordx2 s[8:9], s[0:1], 0x34
; VI-NEXT: s_mov_b32 s3, 0xf000
; VI-NEXT: s_mov_b32 s2, -1
; VI-NEXT: s_waitcnt lgkmcnt(0)
; VI-NEXT: s_mov_b32 s0, s4
; VI-NEXT: s_mov_b32 s1, s5
; VI-NEXT: s_load_dwordx2 s[4:5], s[6:7], 0x0
; VI-NEXT: s_load_dwordx2 s[6:7], s[8:9], 0x0
; VI-NEXT: s_waitcnt lgkmcnt(0)
; VI-NEXT: v_max_f16_e64 v1, s5, s5
; VI-NEXT: v_max_f16_e64 v0, s7, s7
; VI-NEXT: s_lshr_b32 s5, s5, 16
; VI-NEXT: s_lshr_b32 s7, s7, 16
; VI-NEXT: v_max_f16_e32 v0, v1, v0
; VI-NEXT: v_max_f16_e64 v2, s5, s5
; VI-NEXT: v_max_f16_e64 v1, s7, s7
; VI-NEXT: v_max_f16_sdwa v1, v2, v1 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
; VI-NEXT: v_or_b32_sdwa v1, v0, v1 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
; VI-NEXT: v_max_f16_e64 v2, s4, s4
; VI-NEXT: v_max_f16_e64 v0, s6, s6
; VI-NEXT: s_lshr_b32 s4, s4, 16
; VI-NEXT: s_lshr_b32 s5, s6, 16
; VI-NEXT: v_max_f16_e32 v0, v2, v0
; VI-NEXT: v_max_f16_e64 v2, s5, s5
; VI-NEXT: v_max_f16_e64 v3, s4, s4
; VI-NEXT: v_max_f16_sdwa v2, v3, v2 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
; VI-NEXT: v_or_b32_sdwa v0, v0, v2 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
; VI-NEXT: buffer_store_dwordx2 v[0:1], off, s[0:3], 0
; VI-NEXT: s_endpgm
;
; GFX9-LABEL: maxnum_v4f16:
; GFX9: ; %bb.0: ; %entry
; GFX9-NEXT: s_load_dwordx4 s[4:7], s[0:1], 0x24
; GFX9-NEXT: s_load_dwordx2 s[8:9], s[0:1], 0x34
; GFX9-NEXT: s_mov_b32 s3, 0xf000
; GFX9-NEXT: s_mov_b32 s2, -1
; GFX9-NEXT: s_waitcnt lgkmcnt(0)
; GFX9-NEXT: s_mov_b32 s0, s4
; GFX9-NEXT: s_mov_b32 s1, s5
; GFX9-NEXT: s_load_dwordx2 s[4:5], s[6:7], 0x0
; GFX9-NEXT: s_load_dwordx2 s[6:7], s[8:9], 0x0
; GFX9-NEXT: s_waitcnt lgkmcnt(0)
; GFX9-NEXT: v_pk_max_f16 v1, s5, s5
; GFX9-NEXT: v_pk_max_f16 v0, s7, s7
; GFX9-NEXT: v_pk_max_f16 v1, v1, v0
; GFX9-NEXT: v_pk_max_f16 v2, s6, s6
; GFX9-NEXT: v_pk_max_f16 v0, s4, s4
; GFX9-NEXT: v_pk_max_f16 v0, v0, v2
; GFX9-NEXT: buffer_store_dwordx2 v[0:1], off, s[0:3], 0
; GFX9-NEXT: s_endpgm
AMDGPU: Fix missing test coverage for some 16-bit and packed ops llvm-svn: 333024
2018-05-23 04:42:00 +08:00
<4 x half> addrspace(1)* %r,
<4 x half> addrspace(1)* %a,
<4 x half> addrspace(1)* %b) {
entry:
%a.val = load <4 x half>, <4 x half> addrspace(1)* %a
%b.val = load <4 x half>, <4 x half> addrspace(1)* %b
%r.val = call <4 x half> @llvm.maxnum.v4f16(<4 x half> %a.val, <4 x half> %b.val)
store <4 x half> %r.val, <4 x half> addrspace(1)* %r
ret void
}
define amdgpu_kernel void @fmax_v4f16_imm_a(
DAG: Change behavior of fminnum/fmaxnum nodes Introduce new versions that follow the IEEE semantics to help with legalization that may need quieted inputs. There are some regressions from inserting unnecessary canonicalizes when these are matched from fast math fcmp + select which should be fixed in a future commit. llvm-svn: 344914
2018-10-23 00:27:27 +08:00
; SI-LABEL: fmax_v4f16_imm_a:
; SI: ; %bb.0: ; %entry
; SI-NEXT: s_load_dwordx4 s[4:7], s[0:1], 0x9
; SI-NEXT: s_mov_b32 s3, 0xf000
; SI-NEXT: s_mov_b32 s2, -1
; SI-NEXT: s_waitcnt lgkmcnt(0)
; SI-NEXT: s_mov_b32 s0, s4
; SI-NEXT: s_mov_b32 s1, s5
; SI-NEXT: s_load_dwordx2 s[4:5], s[6:7], 0x0
; SI-NEXT: s_waitcnt lgkmcnt(0)
; SI-NEXT: v_cvt_f32_f16_e32 v1, s5
; SI-NEXT: s_lshr_b32 s5, s5, 16
; SI-NEXT: v_cvt_f32_f16_e32 v0, s4
; SI-NEXT: v_cvt_f32_f16_e32 v2, s5
; SI-NEXT: s_lshr_b32 s4, s4, 16
; SI-NEXT: v_cvt_f32_f16_e32 v3, s4
; SI-NEXT: v_mul_f32_e32 v1, 1.0, v1
; SI-NEXT: v_mul_f32_e32 v2, 1.0, v2
; SI-NEXT: v_max_f32_e32 v2, 4.0, v2
; SI-NEXT: v_mul_f32_e32 v3, 1.0, v3
; SI-NEXT: v_max_f32_e32 v1, 0x40400000, v1
; SI-NEXT: v_cvt_f16_f32_e32 v2, v2
; SI-NEXT: v_max_f32_e32 v3, 2.0, v3
; SI-NEXT: v_mul_f32_e32 v0, 1.0, v0
; SI-NEXT: v_max_f32_e32 v0, 0x41000000, v0
; SI-NEXT: v_cvt_f16_f32_e32 v1, v1
; SI-NEXT: v_cvt_f16_f32_e32 v3, v3
; SI-NEXT: v_cvt_f16_f32_e32 v0, v0
; SI-NEXT: v_lshlrev_b32_e32 v2, 16, v2
; SI-NEXT: v_or_b32_e32 v1, v1, v2
; SI-NEXT: v_lshlrev_b32_e32 v2, 16, v3
; SI-NEXT: v_or_b32_e32 v0, v0, v2
; SI-NEXT: buffer_store_dwordx2 v[0:1], off, s[0:3], 0
; SI-NEXT: s_endpgm
;
; VI-LABEL: fmax_v4f16_imm_a:
; VI: ; %bb.0: ; %entry
; VI-NEXT: s_load_dwordx4 s[4:7], s[0:1], 0x24
; VI-NEXT: v_mov_b32_e32 v0, 0x4400
; VI-NEXT: s_mov_b32 s3, 0xf000
; VI-NEXT: s_mov_b32 s2, -1
; VI-NEXT: s_waitcnt lgkmcnt(0)
; VI-NEXT: s_mov_b32 s0, s4
; VI-NEXT: s_mov_b32 s1, s5
; VI-NEXT: s_load_dwordx2 s[4:5], s[6:7], 0x0
; VI-NEXT: s_waitcnt lgkmcnt(0)
; VI-NEXT: v_max_f16_e64 v1, s5, s5
; VI-NEXT: s_lshr_b32 s5, s5, 16
; VI-NEXT: v_max_f16_e64 v3, s5, s5
; VI-NEXT: v_max_f16_e64 v2, s4, s4
; VI-NEXT: v_max_f16_sdwa v0, v3, v0 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
; VI-NEXT: v_max_f16_e32 v1, 0x4200, v1
; VI-NEXT: s_lshr_b32 s4, s4, 16
; VI-NEXT: v_or_b32_sdwa v1, v1, v0 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
; VI-NEXT: v_max_f16_e32 v0, 0x4800, v2
; VI-NEXT: v_max_f16_e64 v2, s4, s4
; VI-NEXT: v_mov_b32_e32 v3, 0x4000
; VI-NEXT: v_max_f16_sdwa v2, v2, v3 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
; VI-NEXT: v_or_b32_sdwa v0, v0, v2 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
; VI-NEXT: buffer_store_dwordx2 v[0:1], off, s[0:3], 0
; VI-NEXT: s_endpgm
;
; GFX9-LABEL: fmax_v4f16_imm_a:
; GFX9: ; %bb.0: ; %entry
; GFX9-NEXT: s_load_dwordx4 s[4:7], s[0:1], 0x24
; GFX9-NEXT: s_mov_b32 s8, 0x44004200
; GFX9-NEXT: s_mov_b32 s9, 0x40004800
; GFX9-NEXT: s_mov_b32 s3, 0xf000
; GFX9-NEXT: s_mov_b32 s2, -1
; GFX9-NEXT: s_waitcnt lgkmcnt(0)
; GFX9-NEXT: s_mov_b32 s0, s4
; GFX9-NEXT: s_mov_b32 s1, s5
; GFX9-NEXT: s_load_dwordx2 s[4:5], s[6:7], 0x0
; GFX9-NEXT: s_waitcnt lgkmcnt(0)
; GFX9-NEXT: v_pk_max_f16 v0, s5, s5
; GFX9-NEXT: v_pk_max_f16 v2, s4, s4
; GFX9-NEXT: v_pk_max_f16 v1, v0, s8
; GFX9-NEXT: v_pk_max_f16 v0, v2, s9
; GFX9-NEXT: buffer_store_dwordx2 v[0:1], off, s[0:3], 0
; GFX9-NEXT: s_endpgm
AMDGPU: Fix missing test coverage for some 16-bit and packed ops llvm-svn: 333024
2018-05-23 04:42:00 +08:00
<4 x half> addrspace(1)* %r,
<4 x half> addrspace(1)* %b) {
entry:
%b.val = load <4 x half>, <4 x half> addrspace(1)* %b
%r.val = call <4 x half> @llvm.maxnum.v4f16(<4 x half> <half 8.0, half 2.0, half 3.0, half 4.0>, <4 x half> %b.val)
store <4 x half> %r.val, <4 x half> addrspace(1)* %r
ret void
}