Skip to content

Commit dcc76b4

Browse files
author
Greg Roth
committed
Add IR test for dxilgen pass
1 parent 907cdba commit dcc76b4

2 files changed

Lines changed: 620 additions & 0 deletions

File tree

‎tools/clang/test/CodeGenDXIL/passes/longvec-intrinsics.hlsl‎

Lines changed: 186 additions & 0 deletions
Original file line numberDiff line numberDiff line change
@@ -0,0 +1,186 @@
1+
// RUN: %dxc -T cs_6_9 -enable-16bit-types -DNUM=13 %s | FileCheck %s
2+
3+
// Source for dxilgen test CodeGenDXIL/passes/longvec-intrinsics.ll.
4+
// Some targetted filecheck testing as an incidental.
5+
6+
RWStructuredBuffer<vector<float16_t, NUM> > hBuf;
7+
RWStructuredBuffer<vector<float, NUM> > fBuf;
8+
RWStructuredBuffer<vector<double, NUM> > dBuf;
9+
10+
RWStructuredBuffer<vector<bool, NUM> > bBuf;
11+
RWStructuredBuffer<vector<uint, NUM> > uBuf;
12+
RWStructuredBuffer<vector<int64_t, NUM> > lBuf;
13+
14+
[numthreads(8,1,1)]
15+
void main() {
16+
17+
// CHECK: [[ld:%.*]] = call %dx.types.ResRet.v13f32 @dx.op.rawBufferVectorLoad.v13f32(i32 303, %dx.types.Handle {{%.*}}, i32 11, i32 0, i32 4)
18+
// CHECK: [[fvec1:%.*]] = extractvalue %dx.types.ResRet.v13f32 [[ld]], 0
19+
// CHECK: [[ld:%.*]] = call %dx.types.ResRet.v13f32 @dx.op.rawBufferVectorLoad.v13f32(i32 303, %dx.types.Handle {{%.*}}, i32 12, i32 0, i32 4)
20+
// CHECK: [[fvec2:%.*]] = extractvalue %dx.types.ResRet.v13f32 [[ld]], 0
21+
// CHECK: [[ld:%.*]] = call %dx.types.ResRet.v13f32 @dx.op.rawBufferVectorLoad.v13f32(i32 303, %dx.types.Handle {{%.*}}, i32 13, i32 0, i32 4)
22+
// CHECK: [[fvec3:%.*]] = extractvalue %dx.types.ResRet.v13f32 [[ld]], 0
23+
vector<float, NUM> fVec1 = fBuf[11];
24+
vector<float, NUM> fVec2 = fBuf[12];
25+
vector<float, NUM> fVec3 = fBuf[13];
26+
27+
// CHECK: [[tmp:%.*]] = call <13 x float> @dx.op.binary.v13f32(i32 35, <13 x float> [[fvec1]], <13 x float> [[fvec2]]) ; FMax(a,b)
28+
// CHECK: call <13 x float> @dx.op.binary.v13f32(i32 36, <13 x float> [[tmp]], <13 x float> [[fvec3]]) ; FMin(a,b)
29+
vector<float, NUM> fRes = clamp(fVec1, fVec2, fVec3);
30+
31+
// CHECK: [[ld:%.*]] = call %dx.types.ResRet.v13f16 @dx.op.rawBufferVectorLoad.v13f16(i32 303, %dx.types.Handle {{%.*}}, i32 14, i32 0, i32 2)
32+
// CHECK: [[hvec1:%.*]] = extractvalue %dx.types.ResRet.v13f16 [[ld]], 0
33+
// CHECK: [[ld:%.*]] = call %dx.types.ResRet.v13f16 @dx.op.rawBufferVectorLoad.v13f16(i32 303, %dx.types.Handle {{%.*}}, i32 15, i32 0, i32 2)
34+
// CHECK: [[hvec2:%.*]] = extractvalue %dx.types.ResRet.v13f16 [[ld]], 0
35+
// CHECK: [[ld:%.*]] = call %dx.types.ResRet.v13f16 @dx.op.rawBufferVectorLoad.v13f16(i32 303, %dx.types.Handle {{%.*}}, i32 16, i32 0, i32 2)
36+
// CHECK: [[hvec3:%.*]] = extractvalue %dx.types.ResRet.v13f16 [[ld]], 0
37+
vector<float16_t, NUM> hVec1 = hBuf[14];
38+
vector<float16_t, NUM> hVec2 = hBuf[15];
39+
vector<float16_t, NUM> hVec3 = hBuf[16];
40+
41+
// CHECK: [[tmp:%.*]] = fcmp fast olt <13 x half> [[hvec2]], [[hvec1]]
42+
// CHECK: select <13 x i1> [[tmp]], <13 x half> zeroinitializer, <13 x half> <half 0xH3C00
43+
vector<float16_t, NUM> hRes = step(hVec1, hVec2);
44+
45+
// CHECK: [[tmp:%.*]] = fmul fast <13 x float> [[fvec1]], <float 0x
46+
// CHECK: call <13 x float> @dx.op.unary.v13f32(i32 21, <13 x float> [[tmp]]) ; Exp(value)
47+
fRes += exp(fVec1);
48+
49+
// CHECK: [[tmp:%.*]] = call <13 x half> @dx.op.unary.v13f16(i32 23, <13 x half> [[hvec1]]) ; Log(value)
50+
// CHECK: fmul fast <13 x half> [[tmp]], <half 0xH398C
51+
hRes += log(hVec1);
52+
53+
// CHECK: [[sub:%.*]] = fsub fast <13 x float> [[fvec2]], [[fvec1]]
54+
// CHECK: [[xsub:%.*]] = fsub fast <13 x float> [[fvec3]], [[fvec1]]
55+
// CHECK: [[div:%.*]] = fdiv fast <13 x float> [[xsub]], [[sub]]
56+
// CHECK: [[sat:%.*]] = call <13 x float> @dx.op.unary.v13f32(i32 7, <13 x float> [[div]]) ; Saturate(value)
57+
// CHECK: [[mul:%.*]] = fmul fast <13 x float> [[sat]], <float 2.000000e+00,
58+
// CHECK: [[sub:%.*]] = fsub fast <13 x float> <float 3.000000e+00, {{.*}}>, [[mul]]
59+
// CHECK: [[mul:%.*]] = fmul fast <13 x float> [[sat]], [[sat]]
60+
// CHECK: fmul fast <13 x float> [[mul]], [[sub]]
61+
fRes += smoothstep(fVec1, fVec2, fVec3);
62+
63+
// Intrinsics that expand into llvm ops.
64+
65+
// CHECK: fmul fast <13 x float> [[fvec3]], <float 0x3F91DF46A0000000
66+
fRes += radians(fVec3);
67+
68+
// CHECK: [[cmp:%.*]] = fcmp fast une <13 x float> [[fvec1]], zeroinitializer
69+
// CHECK: [[f2i:%.*]] = bitcast <13 x float> [[fvec1]] to <13 x i32>
70+
// CHECK: [[and:%.*]] = and <13 x i32> [[f2i]], <i32 2139095040
71+
// CHECK: [[add:%.*]] = add nsw <13 x i32> [[and]], <i32 -1056964608
72+
// CHECK: [[shr:%.*]] = ashr <13 x i32> [[add]], <i32 23
73+
// CHECK: [[i2f:%.*]] = sitofp <13 x i32> [[shr]] to <13 x float>
74+
// CHECK: [[sel:%.*]] = select <13 x i1> [[cmp]], <13 x float> [[i2f]], <13 x float> zeroinitializer
75+
// CHECK: [[and:%.*]] = and <13 x i32> [[f2i]], <i32 8388607
76+
// CHECK: or <13 x i32> [[and]], <i32 1056964608
77+
vector<float, NUM> exp = fVec3;
78+
fRes += frexp(fVec1, exp);
79+
fRes += exp;
80+
81+
// CHECK: [[tmp:%.*]] = fsub fast <13 x half> [[hvec3]], [[hvec2]]
82+
// CHECK: fmul fast <13 x half> [[tmp]], [[hvec1]]
83+
hRes += lerp(hVec2, hVec3, hVec1);
84+
85+
86+
// CHECK: [[ld:%.*]] = call %dx.types.ResRet.v13i32 @dx.op.rawBufferVectorLoad.v13i32(i32 303, %dx.types.Handle {{%.*}}, i32 17, i32 0, i32 4)
87+
// CHECK: [[uvec1:%.*]] = extractvalue %dx.types.ResRet.v13i32 [[ld]], 0
88+
// CHECK: [[ld:%.*]] = call %dx.types.ResRet.v13i32 @dx.op.rawBufferVectorLoad.v13i32(i32 303, %dx.types.Handle {{%.*}}, i32 18, i32 0, i32 4)
89+
// CHECK: [[uvec2:%.*]] = extractvalue %dx.types.ResRet.v13i32 [[ld]], 0
90+
vector<uint, NUM> uVec1 = uBuf[17];
91+
vector<uint, NUM> uVec2 = uBuf[18];
92+
93+
vector<uint, NUM> signs = 1;
94+
// CHECK: [[cmp:%.*]] = icmp ne <13 x i32> [[uvec2]], zeroinitializer
95+
// CHECK: zext <13 x i1> [[cmp]] to <13 x i32>
96+
signs *= sign(uVec2);
97+
98+
// CHECK: [[ld:%.*]] = call %dx.types.ResRet.v13i64 @dx.op.rawBufferVectorLoad.v13i64(i32 303, %dx.types.Handle {{%.*}}, i32 19, i32 0, i32 8)
99+
// CHECK: [[lvec1:%.*]] = extractvalue %dx.types.ResRet.v13i64 [[ld]], 0
100+
// CHECK: [[ld:%.*]] = call %dx.types.ResRet.v13i64 @dx.op.rawBufferVectorLoad.v13i64(i32 303, %dx.types.Handle {{%.*}}, i32 20, i32 0, i32 8)
101+
// CHECK: [[lvec2:%.*]] = extractvalue %dx.types.ResRet.v13i64 [[ld]], 0
102+
vector<int64_t, NUM> lVec1 = lBuf[19];
103+
vector<int64_t, NUM> lVec2 = lBuf[20];
104+
105+
// CHECK: [[gt:%.*]] = icmp sgt <13 x i64> [[lvec2]], zeroinitializer
106+
// CHECK: [[lt:%.*]] = icmp slt <13 x i64> [[lvec2]], zeroinitializer
107+
// CHECK: [[igt:%.*]] = zext <13 x i1> [[gt]] to <13 x i32>
108+
// CHECK: [[ilt:%.*]] = zext <13 x i1> [[lt]] to <13 x i32>
109+
// CHECK: sub nsw <13 x i32> [[igt]], [[ilt]]
110+
signs *= sign(lVec2);
111+
112+
vector<uint, NUM> uRes = signs;
113+
114+
// CHECK: [[ld:%.*]] = call %dx.types.ResRet.v13i32 @dx.op.rawBufferVectorLoad.v13i32(i32 303, %dx.types.Handle {{%.*}}, i32 21, i32 0, i32 4)
115+
// CHECK: [[vec:%.*]] = extractvalue %dx.types.ResRet.v13i32 [[ld]], 0
116+
// CHECK: [[bvec:%.*]] = icmp ne <13 x i32> [[vec]], zeroinitializer
117+
// CHECK: [[vec1:%.*]] = zext <13 x i1> [[bvec]] to <13 x i32>
118+
// CHECK: [[ld:%.*]] = call %dx.types.ResRet.v13i32 @dx.op.rawBufferVectorLoad.v13i32(i32 303, %dx.types.Handle {{%.*}}, i32 22, i32 0, i32 4)
119+
// CHECK: [[vec:%.*]] = extractvalue %dx.types.ResRet.v13i32 [[ld]], 0
120+
// CHECK: [[bvec:%.*]] = icmp ne <13 x i32> [[vec]], zeroinitializer
121+
// CHECK: [[vec2:%.*]] = zext <13 x i1> [[bvec]] to <13 x i32>
122+
// CHECK: [[ld:%.*]] = call %dx.types.ResRet.v13i32 @dx.op.rawBufferVectorLoad.v13i32(i32 303, %dx.types.Handle {{%.*}}, i32 23, i32 0, i32 4)
123+
// CHECK: [[vec:%.*]] = extractvalue %dx.types.ResRet.v13i32 [[ld]], 0
124+
// CHECK: [[bvec:%.*]] = icmp ne <13 x i32> [[vec]], zeroinitializer
125+
// CHECK: [[vec3:%.*]] = zext <13 x i1> [[bvec]] to <13 x i32>
126+
vector<bool, NUM> bVec1 = bBuf[21];
127+
vector<bool, NUM> bVec2 = bBuf[22];
128+
vector<bool, NUM> bVec3 = bBuf[23];
129+
130+
// CHECK: [[bvec2:%.*]] = icmp ne <13 x i32> [[vec2]], zeroinitializer
131+
// CHECK: [[bvec1:%.*]] = icmp ne <13 x i32> [[vec1]], zeroinitializer
132+
// CHECK: or <13 x i1> [[bvec2]], [[bvec1]]
133+
uRes += or(bVec1, bVec2);
134+
135+
// CHECK: [[bvec3:%.*]] = icmp ne <13 x i32> [[vec3]], zeroinitializer
136+
// CHECK: and <13 x i1> [[bvec3]], [[bvec2]]
137+
uRes += and(bVec2, bVec3);
138+
139+
// CHECK: select <13 x i1> [[bvec3]], <13 x i64> [[lvec1]], <13 x i64> [[lvec2]]
140+
vector<int64_t, NUM> lRes = select(bVec3, lVec1, lVec2);
141+
142+
// CHECK: [[el1:%.*]] = extractelement <13 x float> [[fvec1]]
143+
// CHECK: [[el2:%.*]] = extractelement <13 x float> [[fvec2]]
144+
// CHECK: [[mul:%.*]] = fmul fast float [[el2]], [[el1]]
145+
// CHECK: [[mad1:%.*]] = call float @dx.op.tertiary.f32(i32 46, float %{{.*}}, float %{{.*}}, float [[mul]]) ; FMad(a,b,c)
146+
// CHECK: [[mad2:%.*]] = call float @dx.op.tertiary.f32(i32 46, float %{{.*}}, float %{{.*}}, float [[mad1]]) ; FMad(a,b,c)
147+
// CHECK: [[mad3:%.*]] = call float @dx.op.tertiary.f32(i32 46, float %{{.*}}, float %{{.*}}, float [[mad2]]) ; FMad(a,b,c)
148+
// CHECK: [[mad4:%.*]] = call float @dx.op.tertiary.f32(i32 46, float %{{.*}}, float %{{.*}}, float [[mad3]]) ; FMad(a,b,c)
149+
// CHECK: [[mad5:%.*]] = call float @dx.op.tertiary.f32(i32 46, float %{{.*}}, float %{{.*}}, float [[mad4]]) ; FMad(a,b,c)
150+
// CHECK: [[mad6:%.*]] = call float @dx.op.tertiary.f32(i32 46, float %{{.*}}, float %{{.*}}, float [[mad5]]) ; FMad(a,b,c)
151+
// CHECK: [[mad7:%.*]] = call float @dx.op.tertiary.f32(i32 46, float %{{.*}}, float %{{.*}}, float [[mad6]]) ; FMad(a,b,c)
152+
// CHECK: [[mad8:%.*]] = call float @dx.op.tertiary.f32(i32 46, float %{{.*}}, float %{{.*}}, float [[mad7]]) ; FMad(a,b,c)
153+
// CHECK: [[mad9:%.*]] = call float @dx.op.tertiary.f32(i32 46, float %{{.*}}, float %{{.*}}, float [[mad8]]) ; FMad(a,b,c)
154+
// CHECK: [[mad10:%.*]] = call float @dx.op.tertiary.f32(i32 46, float %{{.*}}, float %{{.*}}, float [[mad9]]) ; FMad(a,b,c)
155+
// CHECK: [[mad11:%.*]] = call float @dx.op.tertiary.f32(i32 46, float %{{.*}}, float %{{.*}}, float [[mad10]]) ; FMad(a,b,c)
156+
// CHECK: [[mad12:%.*]] = call float @dx.op.tertiary.f32(i32 46, float %{{.*}}, float %{{.*}}, float [[mad11]]) ; FMad(a,b,c)
157+
fRes += dot(fVec1, fVec2);
158+
159+
// CHECK: call <13 x float> @dx.op.unary.v13f32(i32 17, <13 x float> [[fvec1]]) ; Atan(value)
160+
fRes += atan(fVec1);
161+
162+
// CHECK: call <13 x i32> @dx.op.binary.v13i32(i32 40, <13 x i32> [[uvec1]], <13 x i32> [[uvec2]]) ; UMin(a,b)
163+
uRes += min(uVec1, uVec2);
164+
165+
// CHECK: call <13 x float> @dx.op.tertiary.v13f32(i32 46, <13 x float> [[fvec1]], <13 x float> [[fvec2]], <13 x float> [[fvec3]]) ; FMad(a,b,c)
166+
fRes += mad(fVec1, fVec2, fVec3);
167+
168+
// CHECK: [[ld:%.*]] = call %dx.types.ResRet.v13f64 @dx.op.rawBufferVectorLoad.v13f64(i32 303, %dx.types.Handle {{%.*}}, i32 24, i32 0, i32 8)
169+
// CHECK: [[dvec1:%.*]] = extractvalue %dx.types.ResRet.v13f64 [[ld]], 0
170+
// CHECK: [[ld:%.*]] = call %dx.types.ResRet.v13f64 @dx.op.rawBufferVectorLoad.v13f64(i32 303, %dx.types.Handle {{%.*}}, i32 25, i32 0, i32 8)
171+
// CHECK: [[dvec2:%.*]] = extractvalue %dx.types.ResRet.v13f64 [[ld]], 0
172+
// CHECK: [[ld:%.*]] = call %dx.types.ResRet.v13f64 @dx.op.rawBufferVectorLoad.v13f64(i32 303, %dx.types.Handle {{%.*}}, i32 26, i32 0, i32 8)
173+
// CHECK: [[dvec3:%.*]] = extractvalue %dx.types.ResRet.v13f64 [[ld]], 0
174+
vector<double, NUM> dVec1 = dBuf[24];
175+
vector<double, NUM> dVec2 = dBuf[25];
176+
vector<double, NUM> dVec3 = dBuf[26];
177+
178+
// CHECK: call <13 x double> @dx.op.tertiary.v13f64(i32 47, <13 x double> [[dvec1]], <13 x double> [[dvec2]], <13 x double> [[dvec3]])
179+
vector<double, NUM> dRes = fma(dVec1, dVec2, dVec3);
180+
181+
hBuf[0] = hRes;
182+
fBuf[0] = fRes;
183+
dBuf[0] = dRes;
184+
uBuf[0] = uRes;
185+
lBuf[0] = lRes;
186+
}

0 commit comments

Comments
 (0)