Skip to content

Commit eb04956

Browse files
mariusz-sikora-at-amdmariusz-sikora-at-amd
committed
[AMDGPU][GFX12] Add tests for unsupported builtins
__builtin_amdgcn_mfma* and __builtin_amdgcn_smfmac*
1 parent 54952e9 commit eb04956

File tree

1 file changed

+105
-1
lines changed

1 file changed

+105
-1
lines changed

‎clang/test/CodeGenOpenCL/builtins-amdgcn-gfx12-err.cl

Lines changed: 105 additions & 1 deletion
Original file line numberDiff line numberDiff line change
@@ -4,10 +4,114 @@
44

55
typedef unsigned int uint;
66

7-
kernel void test_builtins_amdgcn_gws_insts(uint a, uint b) {
7+
#pragma OPENCL EXTENSION cl_khr_fp64:enable
8+
9+
typedef float v2f __attribute__((ext_vector_type(2)));
10+
typedef float v4f __attribute__((ext_vector_type(4)));
11+
typedef float v16f __attribute__((ext_vector_type(16)));
12+
typedef float v32f __attribute__((ext_vector_type(32)));
13+
typedef half v4h __attribute__((ext_vector_type(4)));
14+
typedef half v8h __attribute__((ext_vector_type(8)));
15+
typedef half v16h __attribute__((ext_vector_type(16)));
16+
typedef half v32h __attribute__((ext_vector_type(32)));
17+
typedef int v2i __attribute__((ext_vector_type(2)));
18+
typedef int v4i __attribute__((ext_vector_type(4)));
19+
typedef int v16i __attribute__((ext_vector_type(16)));
20+
typedef int v32i __attribute__((ext_vector_type(32)));
21+
typedef short v2s __attribute__((ext_vector_type(2)));
22+
typedef short v4s __attribute__((ext_vector_type(4)));
23+
typedef short v8s __attribute__((ext_vector_type(8)));
24+
typedef short v16s __attribute__((ext_vector_type(16)));
25+
typedef short v32s __attribute__((ext_vector_type(32)));
26+
typedef double v4d __attribute__((ext_vector_type(4)));
27+
28+
void builtin_test_unsupported(global v32f* out_v32f,
29+
global v16f* out_v16f,
30+
global v4f* out_v4f,
31+
global v32i* out_v32i,
32+
global v16i* out_v16i,
33+
global v4i* out_v4i,
34+
global v4d* out_v4d,
35+
global double* out_double,
36+
double a_double , double b_double , double c_double,
37+
float a_float , float b_float , float c_float,
38+
int a_int , int b_int , int c_int,
39+
long a_long , long b_long , long c_long,
40+
v4d a_v4d , v4d b_v4d , v4d c_v4d,
41+
v8s a_v8s , v8s b_v8s , v8s c_v8s,
42+
v4s a_v4s , v4s b_v4s , v4s c_v4s,
43+
v2s a_v2s , v2s b_v2s , v2s c_v2s,
44+
v2i a_v2i , v2i b_v2i , v2i c_v2i,
45+
v16i a_v16i , v16i b_v16i , v16i c_v16i,
46+
v32i a_v32i , v32i b_v32i , v32i c_v32i,
47+
v4i a_v4i , v4i b_v4i , v4i c_v4i,
48+
v2f a_v2f , v2f b_v2f , v2f c_v2f,
49+
v4f a_v4f , v4f b_v4f , v4f c_v4f,
50+
v16f a_v16f , v16f b_v16f , v16f c_v16f,
51+
v32f a_v32f , v32f b_v32f , v32f c_v32f,
52+
v4h a_v4h , v4h b_v4h , v4h c_v4h,
53+
v8h a_v8h , v8h b_v8h , v8h c_v8h,
54+
int idx,
55+
56+
uint a, uint b) {
57+
858
__builtin_amdgcn_ds_gws_init(a, b); // expected-error {{'__builtin_amdgcn_ds_gws_init' needs target feature gws}}
959
__builtin_amdgcn_ds_gws_barrier(a, b); // expected-error {{'__builtin_amdgcn_ds_gws_barrier' needs target feature gws}}
1060
__builtin_amdgcn_ds_gws_sema_v(a); // expected-error {{'__builtin_amdgcn_ds_gws_sema_v' needs target feature gws}}
1161
__builtin_amdgcn_ds_gws_sema_br(a, b); // expected-error {{'__builtin_amdgcn_ds_gws_sema_br' needs target feature gws}}
1262
__builtin_amdgcn_ds_gws_sema_p(a); // expected-error {{'__builtin_amdgcn_ds_gws_sema_p' needs target feature gws}}
63+
64+
*out_v32f = __builtin_amdgcn_mfma_f32_32x32x1f32(a_float, b_float, c_v32f, 0, 0, 0); // expected-error {{'__builtin_amdgcn_mfma_f32_32x32x1f32' needs target feature mai-insts}}
65+
*out_v16f = __builtin_amdgcn_mfma_f32_16x16x1f32(a_float, b_float, c_v16f, 0, 0, 0); // expected-error {{'__builtin_amdgcn_mfma_f32_16x16x1f32' needs target feature mai-insts}}
66+
*out_v4f = __builtin_amdgcn_mfma_f32_4x4x1f32(a_float, b_float, c_v4f, 0, 0, 0); // expected-error {{'__builtin_amdgcn_mfma_f32_4x4x1f32' needs target feature mai-insts}}
67+
*out_v16f = __builtin_amdgcn_mfma_f32_32x32x2f32(a_float, b_float, c_v16f, 0, 0, 0); // expected-error {{'__builtin_amdgcn_mfma_f32_32x32x2f32' needs target feature mai-insts}}
68+
*out_v4f = __builtin_amdgcn_mfma_f32_16x16x4f32(a_float, b_float, c_v4f, 0, 0, 0); // expected-error {{'__builtin_amdgcn_mfma_f32_16x16x4f32' needs target feature mai-insts}}
69+
*out_v32f = __builtin_amdgcn_mfma_f32_32x32x4f16(a_v4h, b_v4h, c_v32f, 0, 0, 0); // expected-error {{'__builtin_amdgcn_mfma_f32_32x32x4f16' needs target feature mai-insts}}
70+
*out_v16f = __builtin_amdgcn_mfma_f32_16x16x4f16(a_v4h, b_v4h, c_v16f, 0, 0, 0); // expected-error {{'__builtin_amdgcn_mfma_f32_16x16x4f16' needs target feature mai-insts}}
71+
*out_v4f = __builtin_amdgcn_mfma_f32_4x4x4f16(a_v4h, b_v4h, c_v4f, 0, 0, 0); // expected-error {{'__builtin_amdgcn_mfma_f32_4x4x4f16' needs target feature mai-insts}}
72+
*out_v16f = __builtin_amdgcn_mfma_f32_32x32x8f16(a_v4h, b_v4h, c_v16f, 0, 0, 0); // expected-error {{'__builtin_amdgcn_mfma_f32_32x32x8f16' needs target feature mai-insts}}
73+
*out_v4f = __builtin_amdgcn_mfma_f32_16x16x16f16(a_v4h, b_v4h, c_v4f, 0, 0, 0); // expected-error {{'__builtin_amdgcn_mfma_f32_16x16x16f16' needs target feature mai-insts}}
74+
*out_v32i = __builtin_amdgcn_mfma_i32_32x32x4i8(a_int, b_int, c_v32i, 0, 0, 0); // expected-error {{'__builtin_amdgcn_mfma_i32_32x32x4i8' needs target feature mai-insts}}
75+
*out_v16i = __builtin_amdgcn_mfma_i32_16x16x4i8(a_int, b_int, c_v16i, 0, 0, 0); // expected-error {{'__builtin_amdgcn_mfma_i32_16x16x4i8' needs target feature mai-insts}}
76+
*out_v4i = __builtin_amdgcn_mfma_i32_4x4x4i8(a_int, b_int, c_v4i, 0, 0, 0); // expected-error {{'__builtin_amdgcn_mfma_i32_4x4x4i8' needs target feature mai-insts}}
77+
*out_v16i = __builtin_amdgcn_mfma_i32_32x32x8i8(a_int, b_int, c_v16i, 0, 0, 0); // expected-error {{'__builtin_amdgcn_mfma_i32_32x32x8i8' needs target feature mai-insts}}
78+
*out_v4i = __builtin_amdgcn_mfma_i32_16x16x16i8(a_int, b_int, c_v4i, 0, 0, 0); // expected-error {{'__builtin_amdgcn_mfma_i32_16x16x16i8' needs target feature mai-insts}}
79+
*out_v32f = __builtin_amdgcn_mfma_f32_32x32x2bf16(a_v2s, b_v2s, c_v32f, 0, 0, 0); // expected-error {{'__builtin_amdgcn_mfma_f32_32x32x2bf16' needs target feature mai-insts}}
80+
*out_v16f = __builtin_amdgcn_mfma_f32_16x16x2bf16(a_v2s, b_v2s, c_v16f, 0, 0, 0); // expected-error {{'__builtin_amdgcn_mfma_f32_16x16x2bf16' needs target feature mai-insts}}
81+
*out_v4f = __builtin_amdgcn_mfma_f32_4x4x2bf16(a_v2s, b_v2s, c_v4f, 0, 0, 0); // expected-error {{'__builtin_amdgcn_mfma_f32_4x4x2bf16' needs target feature mai-insts}}
82+
*out_v16f = __builtin_amdgcn_mfma_f32_32x32x4bf16(a_v2s, b_v2s, c_v16f, 0, 0, 0); // expected-error {{'__builtin_amdgcn_mfma_f32_32x32x4bf16' needs target feature mai-insts}}
83+
*out_v4f = __builtin_amdgcn_mfma_f32_16x16x8bf16(a_v2s, b_v2s, c_v4f, 0, 0, 0); // expected-error {{'__builtin_amdgcn_mfma_f32_16x16x8bf16' needs target feature mai-insts}}
84+
*out_v32f = __builtin_amdgcn_mfma_f32_32x32x4bf16_1k(a_v4s, b_v4s, c_v32f, 0, 0, 0); // expected-error {{'__builtin_amdgcn_mfma_f32_32x32x4bf16_1k' needs target feature mai-insts}}
85+
*out_v16f = __builtin_amdgcn_mfma_f32_16x16x4bf16_1k(a_v4s, b_v4s, c_v16f, 0, 0, 0); // expected-error {{'__builtin_amdgcn_mfma_f32_16x16x4bf16_1k' needs target feature mai-insts}}
86+
*out_v4f = __builtin_amdgcn_mfma_f32_4x4x4bf16_1k(a_v4s, b_v4s, c_v4f, 0, 0, 0); // expected-error {{'__builtin_amdgcn_mfma_f32_4x4x4bf16_1k' needs target feature mai-insts}}
87+
*out_v16f = __builtin_amdgcn_mfma_f32_32x32x8bf16_1k(a_v4s, b_v4s, c_v16f, 0, 0, 0); // expected-error {{'__builtin_amdgcn_mfma_f32_32x32x8bf16_1k' needs target feature mai-insts}}
88+
*out_v4f = __builtin_amdgcn_mfma_f32_16x16x16bf16_1k(a_v4s, b_v4s, c_v4f, 0, 0, 0); // expected-error {{'__builtin_amdgcn_mfma_f32_16x16x16bf16_1k' needs target feature mai-insts}}
89+
*out_v4d = __builtin_amdgcn_mfma_f64_16x16x4f64(a_double, b_double, c_v4d, 0, 0, 0); // expected-error {{'__builtin_amdgcn_mfma_f64_16x16x4f64' needs target feature mai-insts}}
90+
*out_double = __builtin_amdgcn_mfma_f64_4x4x4f64(a_double, b_double, c_double, 0, 0, 0); // expected-error {{'__builtin_amdgcn_mfma_f64_4x4x4f64' needs target feature mai-insts}}
91+
*out_v4i = __builtin_amdgcn_mfma_i32_16x16x32_i8(a_long, b_long, c_v4i, 0, 0, 0); // expected-error {{'__builtin_amdgcn_mfma_i32_16x16x32_i8' needs target feature mai-insts}}
92+
*out_v16i = __builtin_amdgcn_mfma_i32_32x32x16_i8(a_long, b_long, c_v16i, 0, 0, 0); // expected-error {{'__builtin_amdgcn_mfma_i32_32x32x16_i8' needs target feature mai-insts}}
93+
*out_v4f = __builtin_amdgcn_mfma_f32_16x16x8_xf32(a_v2f, b_v2f, c_v4f, 0, 0, 0); // expected-error {{'__builtin_amdgcn_mfma_f32_16x16x8_xf32' needs target feature mai-insts}}
94+
*out_v16f = __builtin_amdgcn_mfma_f32_32x32x4_xf32(a_v2f, b_v2f, c_v16f, 0, 0, 0); // expected-error {{'__builtin_amdgcn_mfma_f32_32x32x4_xf32' needs target feature mai-insts}}
95+
*out_v4f = __builtin_amdgcn_mfma_f32_16x16x32_bf8_bf8(a_long, b_long, c_v4f, 0, 0, 0); // expected-error {{'__builtin_amdgcn_mfma_f32_16x16x32_bf8_bf8' needs target feature fp8-insts}}
96+
*out_v4f = __builtin_amdgcn_mfma_f32_16x16x32_bf8_fp8(a_long, b_long, c_v4f, 0, 0, 0); // expected-error {{'__builtin_amdgcn_mfma_f32_16x16x32_bf8_fp8' needs target feature fp8-insts}}
97+
*out_v4f = __builtin_amdgcn_mfma_f32_16x16x32_fp8_bf8(a_long, b_long, c_v4f, 0, 0, 0); // expected-error {{'__builtin_amdgcn_mfma_f32_16x16x32_fp8_bf8' needs target feature fp8-insts}}
98+
*out_v4f = __builtin_amdgcn_mfma_f32_16x16x32_fp8_fp8(a_long, b_long, c_v4f, 0, 0, 0); // expected-error {{'__builtin_amdgcn_mfma_f32_16x16x32_fp8_fp8' needs target feature fp8-insts}}
99+
*out_v16f = __builtin_amdgcn_mfma_f32_32x32x16_bf8_bf8(a_long, b_long, c_v16f, 0, 0, 0); // expected-error {{'__builtin_amdgcn_mfma_f32_32x32x16_bf8_bf8' needs target feature fp8-insts}}
100+
*out_v16f = __builtin_amdgcn_mfma_f32_32x32x16_bf8_fp8(a_long, b_long, c_v16f, 0, 0, 0); // expected-error {{'__builtin_amdgcn_mfma_f32_32x32x16_bf8_fp8' needs target feature fp8-insts}}
101+
*out_v16f = __builtin_amdgcn_mfma_f32_32x32x16_fp8_bf8(a_long, b_long, c_v16f, 0, 0, 0); // expected-error {{'__builtin_amdgcn_mfma_f32_32x32x16_fp8_bf8' needs target feature fp8-insts}}
102+
*out_v16f = __builtin_amdgcn_mfma_f32_32x32x16_fp8_fp8(a_long, b_long, c_v16f, 0, 0, 0); // expected-error {{'__builtin_amdgcn_mfma_f32_32x32x16_fp8_fp8' needs target feature fp8-insts}}
103+
*out_v4f = __builtin_amdgcn_smfmac_f32_16x16x32_f16(a_v4h, b_v8h, c_v4f, idx, 0, 0); // expected-error {{'__builtin_amdgcn_smfmac_f32_16x16x32_f16' needs target feature mai-insts}}
104+
*out_v16f = __builtin_amdgcn_smfmac_f32_32x32x16_f16(a_v4h, b_v8h, c_v16f, idx, 0, 0); // expected-error {{'__builtin_amdgcn_smfmac_f32_32x32x16_f16' needs target feature mai-insts}}
105+
*out_v4f = __builtin_amdgcn_smfmac_f32_16x16x32_bf16(a_v4s, b_v8s, c_v4f, idx, 0, 0); // expected-error {{'__builtin_amdgcn_smfmac_f32_16x16x32_bf16' needs target feature mai-insts}}
106+
*out_v16f = __builtin_amdgcn_smfmac_f32_32x32x16_bf16(a_v4s, b_v8s, c_v16f, idx, 0, 0); // expected-error {{'__builtin_amdgcn_smfmac_f32_32x32x16_bf16' needs target feature mai-insts}}
107+
*out_v4i = __builtin_amdgcn_smfmac_i32_16x16x64_i8(a_v2i, b_v4i, c_v4i, idx, 0, 0); // expected-error {{'__builtin_amdgcn_smfmac_i32_16x16x64_i8' needs target feature mai-insts}}
108+
*out_v16i = __builtin_amdgcn_smfmac_i32_32x32x32_i8(a_v2i, b_v4i, c_v16i, idx, 0, 0); // expected-error {{'__builtin_amdgcn_smfmac_i32_32x32x32_i8' needs target feature mai-insts}}
109+
*out_v4f = __builtin_amdgcn_smfmac_f32_16x16x64_bf8_bf8(a_v2i, b_v4i, c_v4f, idx, 0, 0); // expected-error {{'__builtin_amdgcn_smfmac_f32_16x16x64_bf8_bf8' needs target feature fp8-insts}}
110+
*out_v4f = __builtin_amdgcn_smfmac_f32_16x16x64_bf8_fp8(a_v2i, b_v4i, c_v4f, idx, 0, 0); // expected-error {{'__builtin_amdgcn_smfmac_f32_16x16x64_bf8_fp8' needs target feature fp8-insts}}
111+
*out_v4f = __builtin_amdgcn_smfmac_f32_16x16x64_fp8_bf8(a_v2i, b_v4i, c_v4f, idx, 0, 0); // expected-error {{'__builtin_amdgcn_smfmac_f32_16x16x64_fp8_bf8' needs target feature fp8-insts}}
112+
*out_v4f = __builtin_amdgcn_smfmac_f32_16x16x64_fp8_fp8(a_v2i, b_v4i, c_v4f, idx, 0, 0); // expected-error {{'__builtin_amdgcn_smfmac_f32_16x16x64_fp8_fp8' needs target feature fp8-insts}}
113+
*out_v16f = __builtin_amdgcn_smfmac_f32_32x32x32_bf8_bf8(a_v2i, b_v4i, c_v16f, idx, 0, 0); // expected-error {{'__builtin_amdgcn_smfmac_f32_32x32x32_bf8_bf8' needs target feature fp8-insts}}
114+
*out_v16f = __builtin_amdgcn_smfmac_f32_32x32x32_bf8_fp8(a_v2i, b_v4i, c_v16f, idx, 0, 0); // expected-error {{'__builtin_amdgcn_smfmac_f32_32x32x32_bf8_fp8' needs target feature fp8-insts}}
115+
*out_v16f = __builtin_amdgcn_smfmac_f32_32x32x32_fp8_bf8(a_v2i, b_v4i, c_v16f, idx, 0, 0); // expected-error {{'__builtin_amdgcn_smfmac_f32_32x32x32_fp8_bf8' needs target feature fp8-insts}}
116+
*out_v16f = __builtin_amdgcn_smfmac_f32_32x32x32_fp8_fp8(a_v2i, b_v4i, c_v16f, idx, 0, 0); // expected-error {{'__builtin_amdgcn_smfmac_f32_32x32x32_fp8_fp8' needs target feature fp8-insts}}
13117
}

0 commit comments

Comments
 (0)