annotate clang/test/CodeGen/builtins-nvptx-sm_70.cu @ 207:2e18cbf3894f

LLVM12
author Shinji KONO <kono@ie.u-ryukyu.ac.jp>
date Tue, 08 Jun 2021 06:07:14 +0900
parents 1d019706d866
children
Ignore whitespace changes - Everywhere: Within whitespace: At end of lines:
rev   line source
150
anatofuz
parents:
diff changeset
1 // RUN: %clang_cc1 -triple nvptx64-unknown-unknown -target-cpu sm_70 \
anatofuz
parents:
diff changeset
2 // RUN: -fcuda-is-device -target-feature +ptx60 \
anatofuz
parents:
diff changeset
3 // RUN: -S -emit-llvm -o - -x cuda %s \
anatofuz
parents:
diff changeset
4 // RUN: | FileCheck -check-prefix=CHECK_M16 %s
anatofuz
parents:
diff changeset
5 // RUN: %clang_cc1 -triple nvptx64-unknown-unknown -target-cpu sm_70 \
anatofuz
parents:
diff changeset
6 // RUN: -fcuda-is-device -target-feature +ptx61 -DPTX61 \
anatofuz
parents:
diff changeset
7 // RUN: -S -emit-llvm -o - -x cuda %s \
anatofuz
parents:
diff changeset
8 // RUN: | FileCheck -check-prefixes=CHECK_M16,CHECK_M32_M8 %s
207
Shinji KONO <kono@ie.u-ryukyu.ac.jp>
parents: 150
diff changeset
9 // Make sure builtins still work with the latest combination of GPU & PTX.
Shinji KONO <kono@ie.u-ryukyu.ac.jp>
parents: 150
diff changeset
10 // RUN: %clang_cc1 -triple nvptx64-unknown-unknown -target-cpu sm_86 \
Shinji KONO <kono@ie.u-ryukyu.ac.jp>
parents: 150
diff changeset
11 // RUN: -fcuda-is-device -target-feature +ptx72 -DPTX61 \
Shinji KONO <kono@ie.u-ryukyu.ac.jp>
parents: 150
diff changeset
12 // RUN: -S -emit-llvm -o - -x cuda %s \
Shinji KONO <kono@ie.u-ryukyu.ac.jp>
parents: 150
diff changeset
13 // RUN: | FileCheck -check-prefixes=CHECK_M16,CHECK_M32_M8 %s
150
anatofuz
parents:
diff changeset
14 // RUN: %clang_cc1 -triple nvptx-unknown-unknown -target-cpu sm_60 \
anatofuz
parents:
diff changeset
15 // RUN: -DPTX61 -fcuda-is-device -S -o /dev/null -x cuda -verify=pre-sm_70 %s
anatofuz
parents:
diff changeset
16 // RUN: %clang_cc1 -triple nvptx-unknown-unknown \
anatofuz
parents:
diff changeset
17 // RUN: -target-cpu sm_70 -target-feature +ptx60 \
anatofuz
parents:
diff changeset
18 // RUN: -DPTX61 -fcuda-is-device -S -o /dev/null -x cuda -verify=pre-ptx61 %s
anatofuz
parents:
diff changeset
19
anatofuz
parents:
diff changeset
20 #if !defined(CUDA_VERSION)
anatofuz
parents:
diff changeset
21 #define __device__ __attribute__((device))
anatofuz
parents:
diff changeset
22 #define __global__ __attribute__((global))
anatofuz
parents:
diff changeset
23 #define __shared__ __attribute__((shared))
anatofuz
parents:
diff changeset
24 #define __constant__ __attribute__((constant))
anatofuz
parents:
diff changeset
25
anatofuz
parents:
diff changeset
26 typedef unsigned long long uint64_t;
anatofuz
parents:
diff changeset
27 #endif
anatofuz
parents:
diff changeset
28 // We have to keep all builtins that depend on particular target feature in the
anatofuz
parents:
diff changeset
29 // same function, because the codegen will stop after the very first function
anatofuz
parents:
diff changeset
30 // that encounters an error, so -verify will not be able to find errors in
anatofuz
parents:
diff changeset
31 // subsequent functions.
anatofuz
parents:
diff changeset
32
anatofuz
parents:
diff changeset
33 // CHECK-LABEL: nvvm_wmma_m16n16k16
anatofuz
parents:
diff changeset
34 __device__ void nvvm_wmma_m16n16k16(int *src, int *dst,
anatofuz
parents:
diff changeset
35 float *fsrc, float *fdst,
anatofuz
parents:
diff changeset
36 int ldm) {
anatofuz
parents:
diff changeset
37 // CHECK_M16: call {{.*}} @llvm.nvvm.wmma.m16n16k16.load.a.row.stride.f16
207
Shinji KONO <kono@ie.u-ryukyu.ac.jp>
parents: 150
diff changeset
38 // pre-sm_70-error-re@+1 {{'__hmma_m16n16k16_ld_a' needs target feature (sm_70{{.*}},(ptx60{{.*}}}}
150
anatofuz
parents:
diff changeset
39 __hmma_m16n16k16_ld_a(dst, src, ldm, 0);
anatofuz
parents:
diff changeset
40 // CHECK_M16: call {{.*}} @llvm.nvvm.wmma.m16n16k16.load.a.col.stride.f16
207
Shinji KONO <kono@ie.u-ryukyu.ac.jp>
parents: 150
diff changeset
41 // pre-sm_70-error-re@+1 {{'__hmma_m16n16k16_ld_a' needs target feature (sm_70{{.*}},(ptx60{{.*}}}}
150
anatofuz
parents:
diff changeset
42 __hmma_m16n16k16_ld_a(dst, src+1, ldm, 1);
anatofuz
parents:
diff changeset
43
anatofuz
parents:
diff changeset
44 // CHECK_M16: call {{.*}} @llvm.nvvm.wmma.m16n16k16.load.b.row.stride.f16
207
Shinji KONO <kono@ie.u-ryukyu.ac.jp>
parents: 150
diff changeset
45 // pre-sm_70-error-re@+1 {{'__hmma_m16n16k16_ld_b' needs target feature (sm_70{{.*}},(ptx60{{.*}}}}
150
anatofuz
parents:
diff changeset
46 __hmma_m16n16k16_ld_b(dst, src, ldm, 0);
anatofuz
parents:
diff changeset
47 // CHECK_M16: call {{.*}} @llvm.nvvm.wmma.m16n16k16.load.b.col.stride.f16
207
Shinji KONO <kono@ie.u-ryukyu.ac.jp>
parents: 150
diff changeset
48 // pre-sm_70-error-re@+1 {{'__hmma_m16n16k16_ld_b' needs target feature (sm_70{{.*}},(ptx60{{.*}}}}
150
anatofuz
parents:
diff changeset
49 __hmma_m16n16k16_ld_b(dst, src+2, ldm, 1);
anatofuz
parents:
diff changeset
50
anatofuz
parents:
diff changeset
51 // CHECK_M16: call {{.*}} @llvm.nvvm.wmma.m16n16k16.load.c.row.stride.f16
207
Shinji KONO <kono@ie.u-ryukyu.ac.jp>
parents: 150
diff changeset
52 // pre-sm_70-error-re@+1 {{'__hmma_m16n16k16_ld_c_f16' needs target feature (sm_70{{.*}},(ptx60{{.*}}}}
150
anatofuz
parents:
diff changeset
53 __hmma_m16n16k16_ld_c_f16(dst, src, ldm, 0);
anatofuz
parents:
diff changeset
54 // CHECK_M16: call {{.*}} @llvm.nvvm.wmma.m16n16k16.load.c.col.stride.f16
207
Shinji KONO <kono@ie.u-ryukyu.ac.jp>
parents: 150
diff changeset
55 // pre-sm_70-error-re@+1 {{'__hmma_m16n16k16_ld_c_f16' needs target feature (sm_70{{.*}},(ptx60{{.*}}}}
150
anatofuz
parents:
diff changeset
56 __hmma_m16n16k16_ld_c_f16(dst, src, ldm, 1);
anatofuz
parents:
diff changeset
57
anatofuz
parents:
diff changeset
58 // CHECK_M16: call {{.*}} @llvm.nvvm.wmma.m16n16k16.load.c.row.stride.f32
207
Shinji KONO <kono@ie.u-ryukyu.ac.jp>
parents: 150
diff changeset
59 // pre-sm_70-error-re@+1 {{'__hmma_m16n16k16_ld_c_f32' needs target feature (sm_70{{.*}},(ptx60{{.*}}}}
150
anatofuz
parents:
diff changeset
60 __hmma_m16n16k16_ld_c_f32(fdst, fsrc, ldm, 0);
anatofuz
parents:
diff changeset
61 // CHECK_M16: call {{.*}} @llvm.nvvm.wmma.m16n16k16.load.c.col.stride.f32
207
Shinji KONO <kono@ie.u-ryukyu.ac.jp>
parents: 150
diff changeset
62 // pre-sm_70-error-re@+1 {{'__hmma_m16n16k16_ld_c_f32' needs target feature (sm_70{{.*}},(ptx60{{.*}}}}
150
anatofuz
parents:
diff changeset
63 __hmma_m16n16k16_ld_c_f32(fdst, fsrc, ldm, 1);
anatofuz
parents:
diff changeset
64
anatofuz
parents:
diff changeset
65 // CHECK_M16: call {{.*}} @llvm.nvvm.wmma.m16n16k16.store.d.row.stride.f16
207
Shinji KONO <kono@ie.u-ryukyu.ac.jp>
parents: 150
diff changeset
66 // pre-sm_70-error-re@+1 {{'__hmma_m16n16k16_st_c_f16' needs target feature (sm_70{{.*}},(ptx60{{.*}}}}
150
anatofuz
parents:
diff changeset
67 __hmma_m16n16k16_st_c_f16(dst, src, ldm, 0);
anatofuz
parents:
diff changeset
68 // CHECK_M16: call {{.*}} @llvm.nvvm.wmma.m16n16k16.store.d.col.stride.f16
207
Shinji KONO <kono@ie.u-ryukyu.ac.jp>
parents: 150
diff changeset
69 // pre-sm_70-error-re@+1 {{'__hmma_m16n16k16_st_c_f16' needs target feature (sm_70{{.*}},(ptx60{{.*}}}}
150
anatofuz
parents:
diff changeset
70 __hmma_m16n16k16_st_c_f16(dst, src, ldm, 1);
anatofuz
parents:
diff changeset
71
anatofuz
parents:
diff changeset
72 // CHECK_M16: call {{.*}} @llvm.nvvm.wmma.m16n16k16.store.d.row.stride.f32
207
Shinji KONO <kono@ie.u-ryukyu.ac.jp>
parents: 150
diff changeset
73 // pre-sm_70-error-re@+1 {{'__hmma_m16n16k16_st_c_f32' needs target feature (sm_70{{.*}},(ptx60{{.*}}}}
150
anatofuz
parents:
diff changeset
74 __hmma_m16n16k16_st_c_f32(fdst, fsrc, ldm, 0);
anatofuz
parents:
diff changeset
75 // CHECK_M16: call {{.*}} @llvm.nvvm.wmma.m16n16k16.store.d.col.stride.f32
207
Shinji KONO <kono@ie.u-ryukyu.ac.jp>
parents: 150
diff changeset
76 // pre-sm_70-error-re@+1 {{'__hmma_m16n16k16_st_c_f32' needs target feature (sm_70{{.*}},(ptx60{{.*}}}}
150
anatofuz
parents:
diff changeset
77 __hmma_m16n16k16_st_c_f32(fdst, fsrc, ldm, 1);
anatofuz
parents:
diff changeset
78
anatofuz
parents:
diff changeset
79 // CHECK_M16: call {{.*}} @llvm.nvvm.wmma.m16n16k16.mma.row.row.f16.f16
207
Shinji KONO <kono@ie.u-ryukyu.ac.jp>
parents: 150
diff changeset
80 // pre-sm_70-error-re@+1 {{'__hmma_m16n16k16_mma_f16f16' needs target feature (sm_70{{.*}},(ptx60{{.*}}}}
150
anatofuz
parents:
diff changeset
81 __hmma_m16n16k16_mma_f16f16(dst, src, src, src, 0, 0);
anatofuz
parents:
diff changeset
82 // CHECK_M16: call {{.*}} @llvm.nvvm.wmma.m16n16k16.mma.row.row.f16.f16.satfinite
207
Shinji KONO <kono@ie.u-ryukyu.ac.jp>
parents: 150
diff changeset
83 // pre-sm_70-error-re@+1 {{'__hmma_m16n16k16_mma_f16f16' needs target feature (sm_70{{.*}},(ptx60{{.*}}}}
150
anatofuz
parents:
diff changeset
84 __hmma_m16n16k16_mma_f16f16(dst, src, src, src, 0, 1);
anatofuz
parents:
diff changeset
85 // CHECK_M16: call {{.*}} @llvm.nvvm.wmma.m16n16k16.mma.row.col.f16.f16
207
Shinji KONO <kono@ie.u-ryukyu.ac.jp>
parents: 150
diff changeset
86 // pre-sm_70-error-re@+1 {{'__hmma_m16n16k16_mma_f16f16' needs target feature (sm_70{{.*}},(ptx60{{.*}}}}
150
anatofuz
parents:
diff changeset
87 __hmma_m16n16k16_mma_f16f16(dst, src, src, src, 1, 0);
anatofuz
parents:
diff changeset
88 // CHECK_M16: call {{.*}} @llvm.nvvm.wmma.m16n16k16.mma.row.col.f16.f16.satfinite
207
Shinji KONO <kono@ie.u-ryukyu.ac.jp>
parents: 150
diff changeset
89 // pre-sm_70-error-re@+1 {{'__hmma_m16n16k16_mma_f16f16' needs target feature (sm_70{{.*}},(ptx60{{.*}}}}
150
anatofuz
parents:
diff changeset
90 __hmma_m16n16k16_mma_f16f16(dst, src, src, src, 1, 1);
anatofuz
parents:
diff changeset
91 // CHECK_M16: call {{.*}} @llvm.nvvm.wmma.m16n16k16.mma.col.row.f16.f16
207
Shinji KONO <kono@ie.u-ryukyu.ac.jp>
parents: 150
diff changeset
92 // pre-sm_70-error-re@+1 {{'__hmma_m16n16k16_mma_f16f16' needs target feature (sm_70{{.*}},(ptx60{{.*}}}}
150
anatofuz
parents:
diff changeset
93 __hmma_m16n16k16_mma_f16f16(dst, src, src, src, 2, 0);
anatofuz
parents:
diff changeset
94 // CHECK_M16: call {{.*}} @llvm.nvvm.wmma.m16n16k16.mma.col.row.f16.f16.satfinite
207
Shinji KONO <kono@ie.u-ryukyu.ac.jp>
parents: 150
diff changeset
95 // pre-sm_70-error-re@+1 {{'__hmma_m16n16k16_mma_f16f16' needs target feature (sm_70{{.*}},(ptx60{{.*}}}}
150
anatofuz
parents:
diff changeset
96 __hmma_m16n16k16_mma_f16f16(dst, src, src, src, 2, 1);
anatofuz
parents:
diff changeset
97 // CHECK_M16: call {{.*}} @llvm.nvvm.wmma.m16n16k16.mma.col.col.f16.f16
207
Shinji KONO <kono@ie.u-ryukyu.ac.jp>
parents: 150
diff changeset
98 // pre-sm_70-error-re@+1 {{'__hmma_m16n16k16_mma_f16f16' needs target feature (sm_70{{.*}},(ptx60{{.*}}}}
150
anatofuz
parents:
diff changeset
99 __hmma_m16n16k16_mma_f16f16(dst, src, src, src, 3, 0);
anatofuz
parents:
diff changeset
100 // CHECK_M16: call {{.*}} @llvm.nvvm.wmma.m16n16k16.mma.col.col.f16.f16.satfinite
207
Shinji KONO <kono@ie.u-ryukyu.ac.jp>
parents: 150
diff changeset
101 // pre-sm_70-error-re@+1 {{'__hmma_m16n16k16_mma_f16f16' needs target feature (sm_70{{.*}},(ptx60{{.*}}}}
150
anatofuz
parents:
diff changeset
102 __hmma_m16n16k16_mma_f16f16(dst, src, src, src, 3, 1);
anatofuz
parents:
diff changeset
103
anatofuz
parents:
diff changeset
104 // CHECK_M16: call {{.*}} @llvm.nvvm.wmma.m16n16k16.mma.row.row.f16.f32
207
Shinji KONO <kono@ie.u-ryukyu.ac.jp>
parents: 150
diff changeset
105 // pre-sm_70-error-re@+1 {{'__hmma_m16n16k16_mma_f16f32' needs target feature (sm_70{{.*}},(ptx60{{.*}}}}
150
anatofuz
parents:
diff changeset
106 __hmma_m16n16k16_mma_f16f32(dst, src, src, fsrc, 0, 0);
anatofuz
parents:
diff changeset
107 // CHECK_M16: call {{.*}} @llvm.nvvm.wmma.m16n16k16.mma.row.row.f16.f32.satfinite
207
Shinji KONO <kono@ie.u-ryukyu.ac.jp>
parents: 150
diff changeset
108 // pre-sm_70-error-re@+1 {{'__hmma_m16n16k16_mma_f16f32' needs target feature (sm_70{{.*}},(ptx60{{.*}}}}
150
anatofuz
parents:
diff changeset
109 __hmma_m16n16k16_mma_f16f32(dst, src, src, fsrc, 0, 1);
anatofuz
parents:
diff changeset
110 // CHECK_M16: call {{.*}} @llvm.nvvm.wmma.m16n16k16.mma.row.col.f16.f32
207
Shinji KONO <kono@ie.u-ryukyu.ac.jp>
parents: 150
diff changeset
111 // pre-sm_70-error-re@+1 {{'__hmma_m16n16k16_mma_f16f32' needs target feature (sm_70{{.*}},(ptx60{{.*}}}}
150
anatofuz
parents:
diff changeset
112 __hmma_m16n16k16_mma_f16f32(dst, src, src, fsrc, 1, 0);
anatofuz
parents:
diff changeset
113 // CHECK_M16: call {{.*}} @llvm.nvvm.wmma.m16n16k16.mma.row.col.f16.f32.satfinite
207
Shinji KONO <kono@ie.u-ryukyu.ac.jp>
parents: 150
diff changeset
114 // pre-sm_70-error-re@+1 {{'__hmma_m16n16k16_mma_f16f32' needs target feature (sm_70{{.*}},(ptx60{{.*}}}}
150
anatofuz
parents:
diff changeset
115 __hmma_m16n16k16_mma_f16f32(dst, src, src, fsrc, 1, 1);
anatofuz
parents:
diff changeset
116 // CHECK_M16: call {{.*}} @llvm.nvvm.wmma.m16n16k16.mma.col.row.f16.f32
207
Shinji KONO <kono@ie.u-ryukyu.ac.jp>
parents: 150
diff changeset
117 // pre-sm_70-error-re@+1 {{'__hmma_m16n16k16_mma_f16f32' needs target feature (sm_70{{.*}},(ptx60{{.*}}}}
150
anatofuz
parents:
diff changeset
118 __hmma_m16n16k16_mma_f16f32(dst, src, src, fsrc, 2, 0);
anatofuz
parents:
diff changeset
119 // CHECK_M16: call {{.*}} @llvm.nvvm.wmma.m16n16k16.mma.col.row.f16.f32.satfinite
207
Shinji KONO <kono@ie.u-ryukyu.ac.jp>
parents: 150
diff changeset
120 // pre-sm_70-error-re@+1 {{'__hmma_m16n16k16_mma_f16f32' needs target feature (sm_70{{.*}},(ptx60{{.*}}}}
150
anatofuz
parents:
diff changeset
121 __hmma_m16n16k16_mma_f16f32(dst, src, src, fsrc, 2, 1);
anatofuz
parents:
diff changeset
122 // CHECK_M16: call {{.*}} @llvm.nvvm.wmma.m16n16k16.mma.col.col.f16.f32
207
Shinji KONO <kono@ie.u-ryukyu.ac.jp>
parents: 150
diff changeset
123 // pre-sm_70-error-re@+1 {{'__hmma_m16n16k16_mma_f16f32' needs target feature (sm_70{{.*}},(ptx60{{.*}}}}
150
anatofuz
parents:
diff changeset
124 __hmma_m16n16k16_mma_f16f32(dst, src, src, fsrc, 3, 0);
anatofuz
parents:
diff changeset
125 // CHECK_M16: call {{.*}} @llvm.nvvm.wmma.m16n16k16.mma.col.col.f16.f32.satfinite
207
Shinji KONO <kono@ie.u-ryukyu.ac.jp>
parents: 150
diff changeset
126 // pre-sm_70-error-re@+1 {{'__hmma_m16n16k16_mma_f16f32' needs target feature (sm_70{{.*}},(ptx60{{.*}}}}
150
anatofuz
parents:
diff changeset
127 __hmma_m16n16k16_mma_f16f32(dst, src, src, fsrc, 3, 1);
anatofuz
parents:
diff changeset
128
anatofuz
parents:
diff changeset
129 // CHECK_M16: call {{.*}} @llvm.nvvm.wmma.m16n16k16.mma.row.row.f32.f16
207
Shinji KONO <kono@ie.u-ryukyu.ac.jp>
parents: 150
diff changeset
130 // pre-sm_70-error-re@+1 {{'__hmma_m16n16k16_mma_f32f16' needs target feature (sm_70{{.*}},(ptx60{{.*}}}}
150
anatofuz
parents:
diff changeset
131 __hmma_m16n16k16_mma_f32f16(fdst, src, src, src, 0, 0);
anatofuz
parents:
diff changeset
132 // CHECK_M16: call {{.*}} @llvm.nvvm.wmma.m16n16k16.mma.row.row.f32.f16.satfinite
207
Shinji KONO <kono@ie.u-ryukyu.ac.jp>
parents: 150
diff changeset
133 // pre-sm_70-error-re@+1 {{'__hmma_m16n16k16_mma_f32f16' needs target feature (sm_70{{.*}},(ptx60{{.*}}}}
150
anatofuz
parents:
diff changeset
134 __hmma_m16n16k16_mma_f32f16(fdst, src, src, src, 0, 1);
anatofuz
parents:
diff changeset
135 // CHECK_M16: call {{.*}} @llvm.nvvm.wmma.m16n16k16.mma.row.col.f32.f16
207
Shinji KONO <kono@ie.u-ryukyu.ac.jp>
parents: 150
diff changeset
136 // pre-sm_70-error-re@+1 {{'__hmma_m16n16k16_mma_f32f16' needs target feature (sm_70{{.*}},(ptx60{{.*}}}}
150
anatofuz
parents:
diff changeset
137 __hmma_m16n16k16_mma_f32f16(fdst, src, src, src, 1, 0);
anatofuz
parents:
diff changeset
138 // CHECK_M16: call {{.*}} @llvm.nvvm.wmma.m16n16k16.mma.row.col.f32.f16.satfinite
207
Shinji KONO <kono@ie.u-ryukyu.ac.jp>
parents: 150
diff changeset
139 // pre-sm_70-error-re@+1 {{'__hmma_m16n16k16_mma_f32f16' needs target feature (sm_70{{.*}},(ptx60{{.*}}}}
150
anatofuz
parents:
diff changeset
140 __hmma_m16n16k16_mma_f32f16(fdst, src, src, src, 1, 1);
anatofuz
parents:
diff changeset
141 // CHECK_M16: call {{.*}} @llvm.nvvm.wmma.m16n16k16.mma.col.row.f32.f16
207
Shinji KONO <kono@ie.u-ryukyu.ac.jp>
parents: 150
diff changeset
142 // pre-sm_70-error-re@+1 {{'__hmma_m16n16k16_mma_f32f16' needs target feature (sm_70{{.*}},(ptx60{{.*}}}}
150
anatofuz
parents:
diff changeset
143 __hmma_m16n16k16_mma_f32f16(fdst, src, src, src, 2, 0);
anatofuz
parents:
diff changeset
144 // CHECK_M16: call {{.*}} @llvm.nvvm.wmma.m16n16k16.mma.col.row.f32.f16.satfinite
207
Shinji KONO <kono@ie.u-ryukyu.ac.jp>
parents: 150
diff changeset
145 // pre-sm_70-error-re@+1 {{'__hmma_m16n16k16_mma_f32f16' needs target feature (sm_70{{.*}},(ptx60{{.*}}}}
150
anatofuz
parents:
diff changeset
146 __hmma_m16n16k16_mma_f32f16(fdst, src, src, src, 2, 1);
anatofuz
parents:
diff changeset
147 // CHECK_M16: call {{.*}} @llvm.nvvm.wmma.m16n16k16.mma.col.col.f32.f16
207
Shinji KONO <kono@ie.u-ryukyu.ac.jp>
parents: 150
diff changeset
148 // pre-sm_70-error-re@+1 {{'__hmma_m16n16k16_mma_f32f16' needs target feature (sm_70{{.*}},(ptx60{{.*}}}}
150
anatofuz
parents:
diff changeset
149 __hmma_m16n16k16_mma_f32f16(fdst, src, src, src, 3, 0);
anatofuz
parents:
diff changeset
150 // CHECK_M16: call {{.*}} @llvm.nvvm.wmma.m16n16k16.mma.col.col.f32.f16.satfinite
207
Shinji KONO <kono@ie.u-ryukyu.ac.jp>
parents: 150
diff changeset
151 // pre-sm_70-error-re@+1 {{'__hmma_m16n16k16_mma_f32f16' needs target feature (sm_70{{.*}},(ptx60{{.*}}}}
150
anatofuz
parents:
diff changeset
152 __hmma_m16n16k16_mma_f32f16(fdst, src, src, src, 3, 1);
anatofuz
parents:
diff changeset
153
anatofuz
parents:
diff changeset
154 // CHECK_M16: call {{.*}} @llvm.nvvm.wmma.m16n16k16.mma.row.row.f32.f32
207
Shinji KONO <kono@ie.u-ryukyu.ac.jp>
parents: 150
diff changeset
155 // pre-sm_70-error-re@+1 {{'__hmma_m16n16k16_mma_f32f32' needs target feature (sm_70{{.*}},(ptx60{{.*}}}}
150
anatofuz
parents:
diff changeset
156 __hmma_m16n16k16_mma_f32f32(fdst, src, src, fsrc, 0, 0);
anatofuz
parents:
diff changeset
157 // CHECK_M16: call {{.*}} @llvm.nvvm.wmma.m16n16k16.mma.row.row.f32.f32.satfinite
207
Shinji KONO <kono@ie.u-ryukyu.ac.jp>
parents: 150
diff changeset
158 // pre-sm_70-error-re@+1 {{'__hmma_m16n16k16_mma_f32f32' needs target feature (sm_70{{.*}},(ptx60{{.*}}}}
150
anatofuz
parents:
diff changeset
159 __hmma_m16n16k16_mma_f32f32(fdst, src, src, fsrc, 0, 1);
anatofuz
parents:
diff changeset
160 // CHECK_M16: call {{.*}} @llvm.nvvm.wmma.m16n16k16.mma.row.col.f32.f32
207
Shinji KONO <kono@ie.u-ryukyu.ac.jp>
parents: 150
diff changeset
161 // pre-sm_70-error-re@+1 {{'__hmma_m16n16k16_mma_f32f32' needs target feature (sm_70{{.*}},(ptx60{{.*}}}}
150
anatofuz
parents:
diff changeset
162 __hmma_m16n16k16_mma_f32f32(fdst, src, src, fsrc, 1, 0);
anatofuz
parents:
diff changeset
163 // CHECK_M16: call {{.*}} @llvm.nvvm.wmma.m16n16k16.mma.row.col.f32.f32.satfinite
207
Shinji KONO <kono@ie.u-ryukyu.ac.jp>
parents: 150
diff changeset
164 // pre-sm_70-error-re@+1 {{'__hmma_m16n16k16_mma_f32f32' needs target feature (sm_70{{.*}},(ptx60{{.*}}}}
150
anatofuz
parents:
diff changeset
165 __hmma_m16n16k16_mma_f32f32(fdst, src, src, fsrc, 1, 1);
anatofuz
parents:
diff changeset
166 // CHECK_M16: call {{.*}} @llvm.nvvm.wmma.m16n16k16.mma.col.row.f32.f32
207
Shinji KONO <kono@ie.u-ryukyu.ac.jp>
parents: 150
diff changeset
167 // pre-sm_70-error-re@+1 {{'__hmma_m16n16k16_mma_f32f32' needs target feature (sm_70{{.*}},(ptx60{{.*}}}}
150
anatofuz
parents:
diff changeset
168 __hmma_m16n16k16_mma_f32f32(fdst, src, src, fsrc, 2, 0);
anatofuz
parents:
diff changeset
169 // CHECK_M16: call {{.*}} @llvm.nvvm.wmma.m16n16k16.mma.col.row.f32.f32.satfinite
207
Shinji KONO <kono@ie.u-ryukyu.ac.jp>
parents: 150
diff changeset
170 // pre-sm_70-error-re@+1 {{'__hmma_m16n16k16_mma_f32f32' needs target feature (sm_70{{.*}},(ptx60{{.*}}}}
150
anatofuz
parents:
diff changeset
171 __hmma_m16n16k16_mma_f32f32(fdst, src, src, fsrc, 2, 1);
anatofuz
parents:
diff changeset
172 // CHECK_M16: call {{.*}} @llvm.nvvm.wmma.m16n16k16.mma.col.col.f32.f32
207
Shinji KONO <kono@ie.u-ryukyu.ac.jp>
parents: 150
diff changeset
173 // pre-sm_70-error-re@+1 {{'__hmma_m16n16k16_mma_f32f32' needs target feature (sm_70{{.*}},(ptx60{{.*}}}}
150
anatofuz
parents:
diff changeset
174 __hmma_m16n16k16_mma_f32f32(fdst, src, src, fsrc, 3, 0);
anatofuz
parents:
diff changeset
175 // CHECK_M16: call {{.*}} @llvm.nvvm.wmma.m16n16k16.mma.col.col.f32.f32.satfinite
207
Shinji KONO <kono@ie.u-ryukyu.ac.jp>
parents: 150
diff changeset
176 // pre-sm_70-error-re@+1 {{'__hmma_m16n16k16_mma_f32f32' needs target feature (sm_70{{.*}},(ptx60{{.*}}}}
150
anatofuz
parents:
diff changeset
177 __hmma_m16n16k16_mma_f32f32(fdst, src, src, fsrc, 3, 1);
anatofuz
parents:
diff changeset
178 }
anatofuz
parents:
diff changeset
179
anatofuz
parents:
diff changeset
180 #ifdef PTX61
anatofuz
parents:
diff changeset
181 // CHECK-LABEL: nvvm_wmma_m32n8k16
anatofuz
parents:
diff changeset
182 __device__ void nvvm_wmma_m32n8k16(int *src, int *dst,
anatofuz
parents:
diff changeset
183 float *fsrc, float *fdst,
anatofuz
parents:
diff changeset
184 int ldm) {
anatofuz
parents:
diff changeset
185 // CHECK_M32_M8: call {{.*}} @llvm.nvvm.wmma.m32n8k16.load.a.row.stride.f16
207
Shinji KONO <kono@ie.u-ryukyu.ac.jp>
parents: 150
diff changeset
186 // pre-ptx61-error-re@+1 {{'__hmma_m32n8k16_ld_a' needs target feature (sm_70{{.*}},(ptx61{{.*}}}}
150
anatofuz
parents:
diff changeset
187 __hmma_m32n8k16_ld_a(dst, src, ldm, 0);
anatofuz
parents:
diff changeset
188 // CHECK_M32_M8: call {{.*}} @llvm.nvvm.wmma.m32n8k16.load.a.col.stride.f16
207
Shinji KONO <kono@ie.u-ryukyu.ac.jp>
parents: 150
diff changeset
189 // pre-ptx61-error-re@+1 {{'__hmma_m32n8k16_ld_a' needs target feature (sm_70{{.*}},(ptx61{{.*}}}}
150
anatofuz
parents:
diff changeset
190 __hmma_m32n8k16_ld_a(dst, src+1, ldm, 1);
anatofuz
parents:
diff changeset
191
anatofuz
parents:
diff changeset
192 // CHECK_M32_M8: call {{.*}} @llvm.nvvm.wmma.m32n8k16.load.b.row.stride.f16
207
Shinji KONO <kono@ie.u-ryukyu.ac.jp>
parents: 150
diff changeset
193 // pre-ptx61-error-re@+1 {{'__hmma_m32n8k16_ld_b' needs target feature (sm_70{{.*}},(ptx61{{.*}}}}
150
anatofuz
parents:
diff changeset
194 __hmma_m32n8k16_ld_b(dst, src, ldm, 0);
anatofuz
parents:
diff changeset
195 // CHECK_M32_M8: call {{.*}} @llvm.nvvm.wmma.m32n8k16.load.b.col.stride.f16
207
Shinji KONO <kono@ie.u-ryukyu.ac.jp>
parents: 150
diff changeset
196 // pre-ptx61-error-re@+1 {{'__hmma_m32n8k16_ld_b' needs target feature (sm_70{{.*}},(ptx61{{.*}}}}
150
anatofuz
parents:
diff changeset
197 __hmma_m32n8k16_ld_b(dst, src+2, ldm, 1);
anatofuz
parents:
diff changeset
198
anatofuz
parents:
diff changeset
199 // CHECK_M32_M8: call {{.*}} @llvm.nvvm.wmma.m32n8k16.load.c.row.stride.f16
207
Shinji KONO <kono@ie.u-ryukyu.ac.jp>
parents: 150
diff changeset
200 // pre-ptx61-error-re@+1 {{'__hmma_m32n8k16_ld_c_f16' needs target feature (sm_70{{.*}},(ptx61{{.*}}}}
150
anatofuz
parents:
diff changeset
201 __hmma_m32n8k16_ld_c_f16(dst, src, ldm, 0);
anatofuz
parents:
diff changeset
202 // CHECK_M32_M8: call {{.*}} @llvm.nvvm.wmma.m32n8k16.load.c.col.stride.f16
207
Shinji KONO <kono@ie.u-ryukyu.ac.jp>
parents: 150
diff changeset
203 // pre-ptx61-error-re@+1 {{'__hmma_m32n8k16_ld_c_f16' needs target feature (sm_70{{.*}},(ptx61{{.*}}}}
150
anatofuz
parents:
diff changeset
204 __hmma_m32n8k16_ld_c_f16(dst, src, ldm, 1);
anatofuz
parents:
diff changeset
205
anatofuz
parents:
diff changeset
206 // CHECK_M32_M8: call {{.*}} @llvm.nvvm.wmma.m32n8k16.load.c.row.stride.f32
207
Shinji KONO <kono@ie.u-ryukyu.ac.jp>
parents: 150
diff changeset
207 // pre-ptx61-error-re@+1 {{'__hmma_m32n8k16_ld_c_f32' needs target feature (sm_70{{.*}},(ptx61{{.*}}}}
150
anatofuz
parents:
diff changeset
208 __hmma_m32n8k16_ld_c_f32(fdst, fsrc, ldm, 0);
anatofuz
parents:
diff changeset
209 // CHECK_M32_M8: call {{.*}} @llvm.nvvm.wmma.m32n8k16.load.c.col.stride.f32
207
Shinji KONO <kono@ie.u-ryukyu.ac.jp>
parents: 150
diff changeset
210 // pre-ptx61-error-re@+1 {{'__hmma_m32n8k16_ld_c_f32' needs target feature (sm_70{{.*}},(ptx61{{.*}}}}
150
anatofuz
parents:
diff changeset
211 __hmma_m32n8k16_ld_c_f32(fdst, fsrc, ldm, 1);
anatofuz
parents:
diff changeset
212
anatofuz
parents:
diff changeset
213 // CHECK_M32_M8: call {{.*}} @llvm.nvvm.wmma.m32n8k16.store.d.row.stride.f16
207
Shinji KONO <kono@ie.u-ryukyu.ac.jp>
parents: 150
diff changeset
214 // pre-ptx61-error-re@+1 {{'__hmma_m32n8k16_st_c_f16' needs target feature (sm_70{{.*}},(ptx61{{.*}}}}
150
anatofuz
parents:
diff changeset
215 __hmma_m32n8k16_st_c_f16(dst, src, ldm, 0);
anatofuz
parents:
diff changeset
216 // CHECK_M32_M8: call {{.*}} @llvm.nvvm.wmma.m32n8k16.store.d.col.stride.f16
207
Shinji KONO <kono@ie.u-ryukyu.ac.jp>
parents: 150
diff changeset
217 // pre-ptx61-error-re@+1 {{'__hmma_m32n8k16_st_c_f16' needs target feature (sm_70{{.*}},(ptx61{{.*}}}}
150
anatofuz
parents:
diff changeset
218 __hmma_m32n8k16_st_c_f16(dst, src, ldm, 1);
anatofuz
parents:
diff changeset
219
anatofuz
parents:
diff changeset
220 // CHECK_M32_M8: call {{.*}} @llvm.nvvm.wmma.m32n8k16.store.d.row.stride.f32
207
Shinji KONO <kono@ie.u-ryukyu.ac.jp>
parents: 150
diff changeset
221 // pre-ptx61-error-re@+1 {{'__hmma_m32n8k16_st_c_f32' needs target feature (sm_70{{.*}},(ptx61{{.*}}}}
150
anatofuz
parents:
diff changeset
222 __hmma_m32n8k16_st_c_f32(fdst, fsrc, ldm, 0);
anatofuz
parents:
diff changeset
223 // CHECK_M32_M8: call {{.*}} @llvm.nvvm.wmma.m32n8k16.store.d.col.stride.f32
207
Shinji KONO <kono@ie.u-ryukyu.ac.jp>
parents: 150
diff changeset
224 // pre-ptx61-error-re@+1 {{'__hmma_m32n8k16_st_c_f32' needs target feature (sm_70{{.*}},(ptx61{{.*}}}}
150
anatofuz
parents:
diff changeset
225 __hmma_m32n8k16_st_c_f32(fdst, fsrc, ldm, 1);
anatofuz
parents:
diff changeset
226
anatofuz
parents:
diff changeset
227 // CHECK_M32_M8: call {{.*}} @llvm.nvvm.wmma.m32n8k16.mma.row.row.f16.f16
207
Shinji KONO <kono@ie.u-ryukyu.ac.jp>
parents: 150
diff changeset
228 // pre-ptx61-error-re@+1 {{'__hmma_m32n8k16_mma_f16f16' needs target feature (sm_70{{.*}},(ptx61{{.*}}}}
150
anatofuz
parents:
diff changeset
229 __hmma_m32n8k16_mma_f16f16(dst, src, src, src, 0, 0);
anatofuz
parents:
diff changeset
230 // CHECK_M32_M8: call {{.*}} @llvm.nvvm.wmma.m32n8k16.mma.row.row.f16.f16.satfinite
207
Shinji KONO <kono@ie.u-ryukyu.ac.jp>
parents: 150
diff changeset
231 // pre-ptx61-error-re@+1 {{'__hmma_m32n8k16_mma_f16f16' needs target feature (sm_70{{.*}},(ptx61{{.*}}}}
150
anatofuz
parents:
diff changeset
232 __hmma_m32n8k16_mma_f16f16(dst, src, src, src, 0, 1);
anatofuz
parents:
diff changeset
233 // CHECK_M32_M8: call {{.*}} @llvm.nvvm.wmma.m32n8k16.mma.row.col.f16.f16
207
Shinji KONO <kono@ie.u-ryukyu.ac.jp>
parents: 150
diff changeset
234 // pre-ptx61-error-re@+1 {{'__hmma_m32n8k16_mma_f16f16' needs target feature (sm_70{{.*}},(ptx61{{.*}}}}
150
anatofuz
parents:
diff changeset
235 __hmma_m32n8k16_mma_f16f16(dst, src, src, src, 1, 0);
anatofuz
parents:
diff changeset
236 // CHECK_M32_M8: call {{.*}} @llvm.nvvm.wmma.m32n8k16.mma.row.col.f16.f16.satfinite
207
Shinji KONO <kono@ie.u-ryukyu.ac.jp>
parents: 150
diff changeset
237 // pre-ptx61-error-re@+1 {{'__hmma_m32n8k16_mma_f16f16' needs target feature (sm_70{{.*}},(ptx61{{.*}}}}
150
anatofuz
parents:
diff changeset
238 __hmma_m32n8k16_mma_f16f16(dst, src, src, src, 1, 1);
anatofuz
parents:
diff changeset
239 // CHECK_M32_M8: call {{.*}} @llvm.nvvm.wmma.m32n8k16.mma.col.row.f16.f16
207
Shinji KONO <kono@ie.u-ryukyu.ac.jp>
parents: 150
diff changeset
240 // pre-ptx61-error-re@+1 {{'__hmma_m32n8k16_mma_f16f16' needs target feature (sm_70{{.*}},(ptx61{{.*}}}}
150
anatofuz
parents:
diff changeset
241 __hmma_m32n8k16_mma_f16f16(dst, src, src, src, 2, 0);
anatofuz
parents:
diff changeset
242 // CHECK_M32_M8: call {{.*}} @llvm.nvvm.wmma.m32n8k16.mma.col.row.f16.f16.satfinite
207
Shinji KONO <kono@ie.u-ryukyu.ac.jp>
parents: 150
diff changeset
243 // pre-ptx61-error-re@+1 {{'__hmma_m32n8k16_mma_f16f16' needs target feature (sm_70{{.*}},(ptx61{{.*}}}}
150
anatofuz
parents:
diff changeset
244 __hmma_m32n8k16_mma_f16f16(dst, src, src, src, 2, 1);
anatofuz
parents:
diff changeset
245 // CHECK_M32_M8: call {{.*}} @llvm.nvvm.wmma.m32n8k16.mma.col.col.f16.f16
207
Shinji KONO <kono@ie.u-ryukyu.ac.jp>
parents: 150
diff changeset
246 // pre-ptx61-error-re@+1 {{'__hmma_m32n8k16_mma_f16f16' needs target feature (sm_70{{.*}},(ptx61{{.*}}}}
150
anatofuz
parents:
diff changeset
247 __hmma_m32n8k16_mma_f16f16(dst, src, src, src, 3, 0);
anatofuz
parents:
diff changeset
248 // CHECK_M32_M8: call {{.*}} @llvm.nvvm.wmma.m32n8k16.mma.col.col.f16.f16.satfinite
207
Shinji KONO <kono@ie.u-ryukyu.ac.jp>
parents: 150
diff changeset
249 // pre-ptx61-error-re@+1 {{'__hmma_m32n8k16_mma_f16f16' needs target feature (sm_70{{.*}},(ptx61{{.*}}}}
150
anatofuz
parents:
diff changeset
250 __hmma_m32n8k16_mma_f16f16(dst, src, src, src, 3, 1);
anatofuz
parents:
diff changeset
251
anatofuz
parents:
diff changeset
252 // CHECK_M32_M8: call {{.*}} @llvm.nvvm.wmma.m32n8k16.mma.row.row.f16.f32
207
Shinji KONO <kono@ie.u-ryukyu.ac.jp>
parents: 150
diff changeset
253 // pre-ptx61-error-re@+1 {{'__hmma_m32n8k16_mma_f16f32' needs target feature (sm_70{{.*}},(ptx61{{.*}}}}
150
anatofuz
parents:
diff changeset
254 __hmma_m32n8k16_mma_f16f32(dst, src, src, fsrc, 0, 0);
anatofuz
parents:
diff changeset
255 // CHECK_M32_M8: call {{.*}} @llvm.nvvm.wmma.m32n8k16.mma.row.row.f16.f32.satfinite
207
Shinji KONO <kono@ie.u-ryukyu.ac.jp>
parents: 150
diff changeset
256 // pre-ptx61-error-re@+1 {{'__hmma_m32n8k16_mma_f16f32' needs target feature (sm_70{{.*}},(ptx61{{.*}}}}
150
anatofuz
parents:
diff changeset
257 __hmma_m32n8k16_mma_f16f32(dst, src, src, fsrc, 0, 1);
anatofuz
parents:
diff changeset
258 // CHECK_M32_M8: call {{.*}} @llvm.nvvm.wmma.m32n8k16.mma.row.col.f16.f32
207
Shinji KONO <kono@ie.u-ryukyu.ac.jp>
parents: 150
diff changeset
259 // pre-ptx61-error-re@+1 {{'__hmma_m32n8k16_mma_f16f32' needs target feature (sm_70{{.*}},(ptx61{{.*}}}}
150
anatofuz
parents:
diff changeset
260 __hmma_m32n8k16_mma_f16f32(dst, src, src, fsrc, 1, 0);
anatofuz
parents:
diff changeset
261 // CHECK_M32_M8: call {{.*}} @llvm.nvvm.wmma.m32n8k16.mma.row.col.f16.f32.satfinite
207
Shinji KONO <kono@ie.u-ryukyu.ac.jp>
parents: 150
diff changeset
262 // pre-ptx61-error-re@+1 {{'__hmma_m32n8k16_mma_f16f32' needs target feature (sm_70{{.*}},(ptx61{{.*}}}}
150
anatofuz
parents:
diff changeset
263 __hmma_m32n8k16_mma_f16f32(dst, src, src, fsrc, 1, 1);
anatofuz
parents:
diff changeset
264 // CHECK_M32_M8: call {{.*}} @llvm.nvvm.wmma.m32n8k16.mma.col.row.f16.f32
207
Shinji KONO <kono@ie.u-ryukyu.ac.jp>
parents: 150
diff changeset
265 // pre-ptx61-error-re@+1 {{'__hmma_m32n8k16_mma_f16f32' needs target feature (sm_70{{.*}},(ptx61{{.*}}}}
150
anatofuz
parents:
diff changeset
266 __hmma_m32n8k16_mma_f16f32(dst, src, src, fsrc, 2, 0);
anatofuz
parents:
diff changeset
267 // CHECK_M32_M8: call {{.*}} @llvm.nvvm.wmma.m32n8k16.mma.col.row.f16.f32.satfinite
207
Shinji KONO <kono@ie.u-ryukyu.ac.jp>
parents: 150
diff changeset
268 // pre-ptx61-error-re@+1 {{'__hmma_m32n8k16_mma_f16f32' needs target feature (sm_70{{.*}},(ptx61{{.*}}}}
150
anatofuz
parents:
diff changeset
269 __hmma_m32n8k16_mma_f16f32(dst, src, src, fsrc, 2, 1);
anatofuz
parents:
diff changeset
270 // CHECK_M32_M8: call {{.*}} @llvm.nvvm.wmma.m32n8k16.mma.col.col.f16.f32
207
Shinji KONO <kono@ie.u-ryukyu.ac.jp>
parents: 150
diff changeset
271 // pre-ptx61-error-re@+1 {{'__hmma_m32n8k16_mma_f16f32' needs target feature (sm_70{{.*}},(ptx61{{.*}}}}
150
anatofuz
parents:
diff changeset
272 __hmma_m32n8k16_mma_f16f32(dst, src, src, fsrc, 3, 0);
anatofuz
parents:
diff changeset
273 // CHECK_M32_M8: call {{.*}} @llvm.nvvm.wmma.m32n8k16.mma.col.col.f16.f32.satfinite
207
Shinji KONO <kono@ie.u-ryukyu.ac.jp>
parents: 150
diff changeset
274 // pre-ptx61-error-re@+1 {{'__hmma_m32n8k16_mma_f16f32' needs target feature (sm_70{{.*}},(ptx61{{.*}}}}
150
anatofuz
parents:
diff changeset
275 __hmma_m32n8k16_mma_f16f32(dst, src, src, fsrc, 3, 1);
anatofuz
parents:
diff changeset
276
anatofuz
parents:
diff changeset
277 // CHECK_M32_M8: call {{.*}} @llvm.nvvm.wmma.m32n8k16.mma.row.row.f32.f16
207
Shinji KONO <kono@ie.u-ryukyu.ac.jp>
parents: 150
diff changeset
278 // pre-ptx61-error-re@+1 {{'__hmma_m32n8k16_mma_f32f16' needs target feature (sm_70{{.*}},(ptx61{{.*}}}}
150
anatofuz
parents:
diff changeset
279 __hmma_m32n8k16_mma_f32f16(fdst, src, src, src, 0, 0);
anatofuz
parents:
diff changeset
280 // CHECK_M32_M8: call {{.*}} @llvm.nvvm.wmma.m32n8k16.mma.row.row.f32.f16.satfinite
207
Shinji KONO <kono@ie.u-ryukyu.ac.jp>
parents: 150
diff changeset
281 // pre-ptx61-error-re@+1 {{'__hmma_m32n8k16_mma_f32f16' needs target feature (sm_70{{.*}},(ptx61{{.*}}}}
150
anatofuz
parents:
diff changeset
282 __hmma_m32n8k16_mma_f32f16(fdst, src, src, src, 0, 1);
anatofuz
parents:
diff changeset
283 // CHECK_M32_M8: call {{.*}} @llvm.nvvm.wmma.m32n8k16.mma.row.col.f32.f16
207
Shinji KONO <kono@ie.u-ryukyu.ac.jp>
parents: 150
diff changeset
284 // pre-ptx61-error-re@+1 {{'__hmma_m32n8k16_mma_f32f16' needs target feature (sm_70{{.*}},(ptx61{{.*}}}}
150
anatofuz
parents:
diff changeset
285 __hmma_m32n8k16_mma_f32f16(fdst, src, src, src, 1, 0);
anatofuz
parents:
diff changeset
286 // CHECK_M32_M8: call {{.*}} @llvm.nvvm.wmma.m32n8k16.mma.row.col.f32.f16.satfinite
207
Shinji KONO <kono@ie.u-ryukyu.ac.jp>
parents: 150
diff changeset
287 // pre-ptx61-error-re@+1 {{'__hmma_m32n8k16_mma_f32f16' needs target feature (sm_70{{.*}},(ptx61{{.*}}}}
150
anatofuz
parents:
diff changeset
288 __hmma_m32n8k16_mma_f32f16(fdst, src, src, src, 1, 1);
anatofuz
parents:
diff changeset
289 // CHECK_M32_M8: call {{.*}} @llvm.nvvm.wmma.m32n8k16.mma.col.row.f32.f16
207
Shinji KONO <kono@ie.u-ryukyu.ac.jp>
parents: 150
diff changeset
290 // pre-ptx61-error-re@+1 {{'__hmma_m32n8k16_mma_f32f16' needs target feature (sm_70{{.*}},(ptx61{{.*}}}}
150
anatofuz
parents:
diff changeset
291 __hmma_m32n8k16_mma_f32f16(fdst, src, src, src, 2, 0);
anatofuz
parents:
diff changeset
292 // CHECK_M32_M8: call {{.*}} @llvm.nvvm.wmma.m32n8k16.mma.col.row.f32.f16.satfinite
207
Shinji KONO <kono@ie.u-ryukyu.ac.jp>
parents: 150
diff changeset
293 // pre-ptx61-error-re@+1 {{'__hmma_m32n8k16_mma_f32f16' needs target feature (sm_70{{.*}},(ptx61{{.*}}}}
150
anatofuz
parents:
diff changeset
294 __hmma_m32n8k16_mma_f32f16(fdst, src, src, src, 2, 1);
anatofuz
parents:
diff changeset
295 // CHECK_M32_M8: call {{.*}} @llvm.nvvm.wmma.m32n8k16.mma.col.col.f32.f16
207
Shinji KONO <kono@ie.u-ryukyu.ac.jp>
parents: 150
diff changeset
296 // pre-ptx61-error-re@+1 {{'__hmma_m32n8k16_mma_f32f16' needs target feature (sm_70{{.*}},(ptx61{{.*}}}}
150
anatofuz
parents:
diff changeset
297 __hmma_m32n8k16_mma_f32f16(fdst, src, src, src, 3, 0);
anatofuz
parents:
diff changeset
298 // CHECK_M32_M8: call {{.*}} @llvm.nvvm.wmma.m32n8k16.mma.col.col.f32.f16.satfinite
207
Shinji KONO <kono@ie.u-ryukyu.ac.jp>
parents: 150
diff changeset
299 // pre-ptx61-error-re@+1 {{'__hmma_m32n8k16_mma_f32f16' needs target feature (sm_70{{.*}},(ptx61{{.*}}}}
150
anatofuz
parents:
diff changeset
300 __hmma_m32n8k16_mma_f32f16(fdst, src, src, src, 3, 1);
anatofuz
parents:
diff changeset
301
anatofuz
parents:
diff changeset
302 // CHECK_M32_M8: call {{.*}} @llvm.nvvm.wmma.m32n8k16.mma.row.row.f32.f32
207
Shinji KONO <kono@ie.u-ryukyu.ac.jp>
parents: 150
diff changeset
303 // pre-ptx61-error-re@+1 {{'__hmma_m32n8k16_mma_f32f32' needs target feature (sm_70{{.*}},(ptx61{{.*}}}}
150
anatofuz
parents:
diff changeset
304 __hmma_m32n8k16_mma_f32f32(fdst, src, src, fsrc, 0, 0);
anatofuz
parents:
diff changeset
305 // CHECK_M32_M8: call {{.*}} @llvm.nvvm.wmma.m32n8k16.mma.row.row.f32.f32.satfinite
207
Shinji KONO <kono@ie.u-ryukyu.ac.jp>
parents: 150
diff changeset
306 // pre-ptx61-error-re@+1 {{'__hmma_m32n8k16_mma_f32f32' needs target feature (sm_70{{.*}},(ptx61{{.*}}}}
150
anatofuz
parents:
diff changeset
307 __hmma_m32n8k16_mma_f32f32(fdst, src, src, fsrc, 0, 1);
anatofuz
parents:
diff changeset
308 // CHECK_M32_M8: call {{.*}} @llvm.nvvm.wmma.m32n8k16.mma.row.col.f32.f32
207
Shinji KONO <kono@ie.u-ryukyu.ac.jp>
parents: 150
diff changeset
309 // pre-ptx61-error-re@+1 {{'__hmma_m32n8k16_mma_f32f32' needs target feature (sm_70{{.*}},(ptx61{{.*}}}}
150
anatofuz
parents:
diff changeset
310 __hmma_m32n8k16_mma_f32f32(fdst, src, src, fsrc, 1, 0);
anatofuz
parents:
diff changeset
311 // CHECK_M32_M8: call {{.*}} @llvm.nvvm.wmma.m32n8k16.mma.row.col.f32.f32.satfinite
207
Shinji KONO <kono@ie.u-ryukyu.ac.jp>
parents: 150
diff changeset
312 // pre-ptx61-error-re@+1 {{'__hmma_m32n8k16_mma_f32f32' needs target feature (sm_70{{.*}},(ptx61{{.*}}}}
150
anatofuz
parents:
diff changeset
313 __hmma_m32n8k16_mma_f32f32(fdst, src, src, fsrc, 1, 1);
anatofuz
parents:
diff changeset
314 // CHECK_M32_M8: call {{.*}} @llvm.nvvm.wmma.m32n8k16.mma.col.row.f32.f32
207
Shinji KONO <kono@ie.u-ryukyu.ac.jp>
parents: 150
diff changeset
315 // pre-ptx61-error-re@+1 {{'__hmma_m32n8k16_mma_f32f32' needs target feature (sm_70{{.*}},(ptx61{{.*}}}}
150
anatofuz
parents:
diff changeset
316 __hmma_m32n8k16_mma_f32f32(fdst, src, src, fsrc, 2, 0);
anatofuz
parents:
diff changeset
317 // CHECK_M32_M8: call {{.*}} @llvm.nvvm.wmma.m32n8k16.mma.col.row.f32.f32.satfinite
207
Shinji KONO <kono@ie.u-ryukyu.ac.jp>
parents: 150
diff changeset
318 // pre-ptx61-error-re@+1 {{'__hmma_m32n8k16_mma_f32f32' needs target feature (sm_70{{.*}},(ptx61{{.*}}}}
150
anatofuz
parents:
diff changeset
319 __hmma_m32n8k16_mma_f32f32(fdst, src, src, fsrc, 2, 1);
anatofuz
parents:
diff changeset
320 // CHECK_M32_M8: call {{.*}} @llvm.nvvm.wmma.m32n8k16.mma.col.col.f32.f32
207
Shinji KONO <kono@ie.u-ryukyu.ac.jp>
parents: 150
diff changeset
321 // pre-ptx61-error-re@+1 {{'__hmma_m32n8k16_mma_f32f32' needs target feature (sm_70{{.*}},(ptx61{{.*}}}}
150
anatofuz
parents:
diff changeset
322 __hmma_m32n8k16_mma_f32f32(fdst, src, src, fsrc, 3, 0);
anatofuz
parents:
diff changeset
323 // CHECK_M32_M8: call {{.*}} @llvm.nvvm.wmma.m32n8k16.mma.col.col.f32.f32.satfinite
207
Shinji KONO <kono@ie.u-ryukyu.ac.jp>
parents: 150
diff changeset
324 // pre-ptx61-error-re@+1 {{'__hmma_m32n8k16_mma_f32f32' needs target feature (sm_70{{.*}},(ptx61{{.*}}}}
150
anatofuz
parents:
diff changeset
325 __hmma_m32n8k16_mma_f32f32(fdst, src, src, fsrc, 3, 1);
anatofuz
parents:
diff changeset
326
anatofuz
parents:
diff changeset
327
anatofuz
parents:
diff changeset
328 // m8n32k16 variants.
anatofuz
parents:
diff changeset
329
anatofuz
parents:
diff changeset
330 // CHECK_M32_M8: call {{.*}} @llvm.nvvm.wmma.m8n32k16.load.a.row.stride.f16
207
Shinji KONO <kono@ie.u-ryukyu.ac.jp>
parents: 150
diff changeset
331 // pre-ptx61-error-re@+1 {{'__hmma_m8n32k16_ld_a' needs target feature (sm_70{{.*}},(ptx61{{.*}}}}
150
anatofuz
parents:
diff changeset
332 __hmma_m8n32k16_ld_a(dst, src, ldm, 0);
anatofuz
parents:
diff changeset
333 // CHECK_M32_M8: call {{.*}} @llvm.nvvm.wmma.m8n32k16.load.a.col.stride.f16
207
Shinji KONO <kono@ie.u-ryukyu.ac.jp>
parents: 150
diff changeset
334 // pre-ptx61-error-re@+1 {{'__hmma_m8n32k16_ld_a' needs target feature (sm_70{{.*}},(ptx61{{.*}}}}
150
anatofuz
parents:
diff changeset
335 __hmma_m8n32k16_ld_a(dst, src+1, ldm, 1);
anatofuz
parents:
diff changeset
336
anatofuz
parents:
diff changeset
337 // CHECK_M32_M8: call {{.*}} @llvm.nvvm.wmma.m8n32k16.load.b.row.stride.f16
207
Shinji KONO <kono@ie.u-ryukyu.ac.jp>
parents: 150
diff changeset
338 // pre-ptx61-error-re@+1 {{'__hmma_m8n32k16_ld_b' needs target feature (sm_70{{.*}},(ptx61{{.*}}}}
150
anatofuz
parents:
diff changeset
339 __hmma_m8n32k16_ld_b(dst, src, ldm, 0);
anatofuz
parents:
diff changeset
340 // CHECK_M32_M8: call {{.*}} @llvm.nvvm.wmma.m8n32k16.load.b.col.stride.f16
207
Shinji KONO <kono@ie.u-ryukyu.ac.jp>
parents: 150
diff changeset
341 // pre-ptx61-error-re@+1 {{'__hmma_m8n32k16_ld_b' needs target feature (sm_70{{.*}},(ptx61{{.*}}}}
150
anatofuz
parents:
diff changeset
342 __hmma_m8n32k16_ld_b(dst, src+2, ldm, 1);
anatofuz
parents:
diff changeset
343
anatofuz
parents:
diff changeset
344 // CHECK_M32_M8: call {{.*}} @llvm.nvvm.wmma.m8n32k16.load.c.row.stride.f16
207
Shinji KONO <kono@ie.u-ryukyu.ac.jp>
parents: 150
diff changeset
345 // pre-ptx61-error-re@+1 {{'__hmma_m8n32k16_ld_c_f16' needs target feature (sm_70{{.*}},(ptx61{{.*}}}}
150
anatofuz
parents:
diff changeset
346 __hmma_m8n32k16_ld_c_f16(dst, src, ldm, 0);
anatofuz
parents:
diff changeset
347 // CHECK_M32_M8: call {{.*}} @llvm.nvvm.wmma.m8n32k16.load.c.col.stride.f16
207
Shinji KONO <kono@ie.u-ryukyu.ac.jp>
parents: 150
diff changeset
348 // pre-ptx61-error-re@+1 {{'__hmma_m8n32k16_ld_c_f16' needs target feature (sm_70{{.*}},(ptx61{{.*}}}}
150
anatofuz
parents:
diff changeset
349 __hmma_m8n32k16_ld_c_f16(dst, src, ldm, 1);
anatofuz
parents:
diff changeset
350
anatofuz
parents:
diff changeset
351 // CHECK_M32_M8: call {{.*}} @llvm.nvvm.wmma.m8n32k16.load.c.row.stride.f32
207
Shinji KONO <kono@ie.u-ryukyu.ac.jp>
parents: 150
diff changeset
352 // pre-ptx61-error-re@+1 {{'__hmma_m8n32k16_ld_c_f32' needs target feature (sm_70{{.*}},(ptx61{{.*}}}}
150
anatofuz
parents:
diff changeset
353 __hmma_m8n32k16_ld_c_f32(fdst, fsrc, ldm, 0);
anatofuz
parents:
diff changeset
354 // CHECK_M32_M8: call {{.*}} @llvm.nvvm.wmma.m8n32k16.load.c.col.stride.f32
207
Shinji KONO <kono@ie.u-ryukyu.ac.jp>
parents: 150
diff changeset
355 // pre-ptx61-error-re@+1 {{'__hmma_m8n32k16_ld_c_f32' needs target feature (sm_70{{.*}},(ptx61{{.*}}}}
150
anatofuz
parents:
diff changeset
356 __hmma_m8n32k16_ld_c_f32(fdst, fsrc, ldm, 1);
anatofuz
parents:
diff changeset
357
anatofuz
parents:
diff changeset
358 // CHECK_M32_M8: call {{.*}} @llvm.nvvm.wmma.m8n32k16.store.d.row.stride.f16
207
Shinji KONO <kono@ie.u-ryukyu.ac.jp>
parents: 150
diff changeset
359 // pre-ptx61-error-re@+1 {{'__hmma_m8n32k16_st_c_f16' needs target feature (sm_70{{.*}},(ptx61{{.*}}}}
150
anatofuz
parents:
diff changeset
360 __hmma_m8n32k16_st_c_f16(dst, src, ldm, 0);
anatofuz
parents:
diff changeset
361 // CHECK_M32_M8: call {{.*}} @llvm.nvvm.wmma.m8n32k16.store.d.col.stride.f16
207
Shinji KONO <kono@ie.u-ryukyu.ac.jp>
parents: 150
diff changeset
362 // pre-ptx61-error-re@+1 {{'__hmma_m8n32k16_st_c_f16' needs target feature (sm_70{{.*}},(ptx61{{.*}}}}
150
anatofuz
parents:
diff changeset
363 __hmma_m8n32k16_st_c_f16(dst, src, ldm, 1);
anatofuz
parents:
diff changeset
364
anatofuz
parents:
diff changeset
365 // CHECK_M32_M8: call {{.*}} @llvm.nvvm.wmma.m8n32k16.store.d.row.stride.f32
207
Shinji KONO <kono@ie.u-ryukyu.ac.jp>
parents: 150
diff changeset
366 // pre-ptx61-error-re@+1 {{'__hmma_m8n32k16_st_c_f32' needs target feature (sm_70{{.*}},(ptx61{{.*}}}}
150
anatofuz
parents:
diff changeset
367 __hmma_m8n32k16_st_c_f32(fdst, fsrc, ldm, 0);
anatofuz
parents:
diff changeset
368 // CHECK_M32_M8: call {{.*}} @llvm.nvvm.wmma.m8n32k16.store.d.col.stride.f32
207
Shinji KONO <kono@ie.u-ryukyu.ac.jp>
parents: 150
diff changeset
369 // pre-ptx61-error-re@+1 {{'__hmma_m8n32k16_st_c_f32' needs target feature (sm_70{{.*}},(ptx61{{.*}}}}
150
anatofuz
parents:
diff changeset
370 __hmma_m8n32k16_st_c_f32(fdst, fsrc, ldm, 1);
anatofuz
parents:
diff changeset
371
anatofuz
parents:
diff changeset
372 // CHECK_M32_M8: call {{.*}} @llvm.nvvm.wmma.m8n32k16.mma.row.row.f16.f16
207
Shinji KONO <kono@ie.u-ryukyu.ac.jp>
parents: 150
diff changeset
373 // pre-ptx61-error-re@+1 {{'__hmma_m8n32k16_mma_f16f16' needs target feature (sm_70{{.*}},(ptx61{{.*}}}}
150
anatofuz
parents:
diff changeset
374 __hmma_m8n32k16_mma_f16f16(dst, src, src, src, 0, 0);
anatofuz
parents:
diff changeset
375 // CHECK_M32_M8: call {{.*}} @llvm.nvvm.wmma.m8n32k16.mma.row.row.f16.f16.satfinite
207
Shinji KONO <kono@ie.u-ryukyu.ac.jp>
parents: 150
diff changeset
376 // pre-ptx61-error-re@+1 {{'__hmma_m8n32k16_mma_f16f16' needs target feature (sm_70{{.*}},(ptx61{{.*}}}}
150
anatofuz
parents:
diff changeset
377 __hmma_m8n32k16_mma_f16f16(dst, src, src, src, 0, 1);
anatofuz
parents:
diff changeset
378 // CHECK_M32_M8: call {{.*}} @llvm.nvvm.wmma.m8n32k16.mma.row.col.f16.f16
207
Shinji KONO <kono@ie.u-ryukyu.ac.jp>
parents: 150
diff changeset
379 // pre-ptx61-error-re@+1 {{'__hmma_m8n32k16_mma_f16f16' needs target feature (sm_70{{.*}},(ptx61{{.*}}}}
150
anatofuz
parents:
diff changeset
380 __hmma_m8n32k16_mma_f16f16(dst, src, src, src, 1, 0);
anatofuz
parents:
diff changeset
381 // CHECK_M32_M8: call {{.*}} @llvm.nvvm.wmma.m8n32k16.mma.row.col.f16.f16.satfinite
207
Shinji KONO <kono@ie.u-ryukyu.ac.jp>
parents: 150
diff changeset
382 // pre-ptx61-error-re@+1 {{'__hmma_m8n32k16_mma_f16f16' needs target feature (sm_70{{.*}},(ptx61{{.*}}}}
150
anatofuz
parents:
diff changeset
383 __hmma_m8n32k16_mma_f16f16(dst, src, src, src, 1, 1);
anatofuz
parents:
diff changeset
384 // CHECK_M32_M8: call {{.*}} @llvm.nvvm.wmma.m8n32k16.mma.col.row.f16.f16
207
Shinji KONO <kono@ie.u-ryukyu.ac.jp>
parents: 150
diff changeset
385 // pre-ptx61-error-re@+1 {{'__hmma_m8n32k16_mma_f16f16' needs target feature (sm_70{{.*}},(ptx61{{.*}}}}
150
anatofuz
parents:
diff changeset
386 __hmma_m8n32k16_mma_f16f16(dst, src, src, src, 2, 0);
anatofuz
parents:
diff changeset
387 // CHECK_M32_M8: call {{.*}} @llvm.nvvm.wmma.m8n32k16.mma.col.row.f16.f16.satfinite
207
Shinji KONO <kono@ie.u-ryukyu.ac.jp>
parents: 150
diff changeset
388 // pre-ptx61-error-re@+1 {{'__hmma_m8n32k16_mma_f16f16' needs target feature (sm_70{{.*}},(ptx61{{.*}}}}
150
anatofuz
parents:
diff changeset
389 __hmma_m8n32k16_mma_f16f16(dst, src, src, src, 2, 1);
anatofuz
parents:
diff changeset
390 // CHECK_M32_M8: call {{.*}} @llvm.nvvm.wmma.m8n32k16.mma.col.col.f16.f16
207
Shinji KONO <kono@ie.u-ryukyu.ac.jp>
parents: 150
diff changeset
391 // pre-ptx61-error-re@+1 {{'__hmma_m8n32k16_mma_f16f16' needs target feature (sm_70{{.*}},(ptx61{{.*}}}}
150
anatofuz
parents:
diff changeset
392 __hmma_m8n32k16_mma_f16f16(dst, src, src, src, 3, 0);
anatofuz
parents:
diff changeset
393 // CHECK_M32_M8: call {{.*}} @llvm.nvvm.wmma.m8n32k16.mma.col.col.f16.f16.satfinite
207
Shinji KONO <kono@ie.u-ryukyu.ac.jp>
parents: 150
diff changeset
394 // pre-ptx61-error-re@+1 {{'__hmma_m8n32k16_mma_f16f16' needs target feature (sm_70{{.*}},(ptx61{{.*}}}}
150
anatofuz
parents:
diff changeset
395 __hmma_m8n32k16_mma_f16f16(dst, src, src, src, 3, 1);
anatofuz
parents:
diff changeset
396
anatofuz
parents:
diff changeset
397 // CHECK_M32_M8: call {{.*}} @llvm.nvvm.wmma.m8n32k16.mma.row.row.f16.f32
207
Shinji KONO <kono@ie.u-ryukyu.ac.jp>
parents: 150
diff changeset
398 // pre-ptx61-error-re@+1 {{'__hmma_m8n32k16_mma_f16f32' needs target feature (sm_70{{.*}},(ptx61{{.*}}}}
150
anatofuz
parents:
diff changeset
399 __hmma_m8n32k16_mma_f16f32(dst, src, src, fsrc, 0, 0);
anatofuz
parents:
diff changeset
400 // CHECK_M32_M8: call {{.*}} @llvm.nvvm.wmma.m8n32k16.mma.row.row.f16.f32.satfinite
207
Shinji KONO <kono@ie.u-ryukyu.ac.jp>
parents: 150
diff changeset
401 // pre-ptx61-error-re@+1 {{'__hmma_m8n32k16_mma_f16f32' needs target feature (sm_70{{.*}},(ptx61{{.*}}}}
150
anatofuz
parents:
diff changeset
402 __hmma_m8n32k16_mma_f16f32(dst, src, src, fsrc, 0, 1);
anatofuz
parents:
diff changeset
403 // CHECK_M32_M8: call {{.*}} @llvm.nvvm.wmma.m8n32k16.mma.row.col.f16.f32
207
Shinji KONO <kono@ie.u-ryukyu.ac.jp>
parents: 150
diff changeset
404 // pre-ptx61-error-re@+1 {{'__hmma_m8n32k16_mma_f16f32' needs target feature (sm_70{{.*}},(ptx61{{.*}}}}
150
anatofuz
parents:
diff changeset
405 __hmma_m8n32k16_mma_f16f32(dst, src, src, fsrc, 1, 0);
anatofuz
parents:
diff changeset
406 // CHECK_M32_M8: call {{.*}} @llvm.nvvm.wmma.m8n32k16.mma.row.col.f16.f32.satfinite
207
Shinji KONO <kono@ie.u-ryukyu.ac.jp>
parents: 150
diff changeset
407 // pre-ptx61-error-re@+1 {{'__hmma_m8n32k16_mma_f16f32' needs target feature (sm_70{{.*}},(ptx61{{.*}}}}
150
anatofuz
parents:
diff changeset
408 __hmma_m8n32k16_mma_f16f32(dst, src, src, fsrc, 1, 1);
anatofuz
parents:
diff changeset
409 // CHECK_M32_M8: call {{.*}} @llvm.nvvm.wmma.m8n32k16.mma.col.row.f16.f32
207
Shinji KONO <kono@ie.u-ryukyu.ac.jp>
parents: 150
diff changeset
410 // pre-ptx61-error-re@+1 {{'__hmma_m8n32k16_mma_f16f32' needs target feature (sm_70{{.*}},(ptx61{{.*}}}}
150
anatofuz
parents:
diff changeset
411 __hmma_m8n32k16_mma_f16f32(dst, src, src, fsrc, 2, 0);
anatofuz
parents:
diff changeset
412 // CHECK_M32_M8: call {{.*}} @llvm.nvvm.wmma.m8n32k16.mma.col.row.f16.f32.satfinite
207
Shinji KONO <kono@ie.u-ryukyu.ac.jp>
parents: 150
diff changeset
413 // pre-ptx61-error-re@+1 {{'__hmma_m8n32k16_mma_f16f32' needs target feature (sm_70{{.*}},(ptx61{{.*}}}}
150
anatofuz
parents:
diff changeset
414 __hmma_m8n32k16_mma_f16f32(dst, src, src, fsrc, 2, 1);
anatofuz
parents:
diff changeset
415 // CHECK_M32_M8: call {{.*}} @llvm.nvvm.wmma.m8n32k16.mma.col.col.f16.f32
207
Shinji KONO <kono@ie.u-ryukyu.ac.jp>
parents: 150
diff changeset
416 // pre-ptx61-error-re@+1 {{'__hmma_m8n32k16_mma_f16f32' needs target feature (sm_70{{.*}},(ptx61{{.*}}}}
150
anatofuz
parents:
diff changeset
417 __hmma_m8n32k16_mma_f16f32(dst, src, src, fsrc, 3, 0);
anatofuz
parents:
diff changeset
418 // CHECK_M32_M8: call {{.*}} @llvm.nvvm.wmma.m8n32k16.mma.col.col.f16.f32.satfinite
207
Shinji KONO <kono@ie.u-ryukyu.ac.jp>
parents: 150
diff changeset
419 // pre-ptx61-error-re@+1 {{'__hmma_m8n32k16_mma_f16f32' needs target feature (sm_70{{.*}},(ptx61{{.*}}}}
150
anatofuz
parents:
diff changeset
420 __hmma_m8n32k16_mma_f16f32(dst, src, src, fsrc, 3, 1);
anatofuz
parents:
diff changeset
421
anatofuz
parents:
diff changeset
422 // CHECK_M32_M8: call {{.*}} @llvm.nvvm.wmma.m8n32k16.mma.row.row.f32.f16
207
Shinji KONO <kono@ie.u-ryukyu.ac.jp>
parents: 150
diff changeset
423 // pre-ptx61-error-re@+1 {{'__hmma_m8n32k16_mma_f32f16' needs target feature (sm_70{{.*}},(ptx61{{.*}}}}
150
anatofuz
parents:
diff changeset
424 __hmma_m8n32k16_mma_f32f16(fdst, src, src, src, 0, 0);
anatofuz
parents:
diff changeset
425 // CHECK_M32_M8: call {{.*}} @llvm.nvvm.wmma.m8n32k16.mma.row.row.f32.f16.satfinite
207
Shinji KONO <kono@ie.u-ryukyu.ac.jp>
parents: 150
diff changeset
426 // pre-ptx61-error-re@+1 {{'__hmma_m8n32k16_mma_f32f16' needs target feature (sm_70{{.*}},(ptx61{{.*}}}}
150
anatofuz
parents:
diff changeset
427 __hmma_m8n32k16_mma_f32f16(fdst, src, src, src, 0, 1);
anatofuz
parents:
diff changeset
428 // CHECK_M32_M8: call {{.*}} @llvm.nvvm.wmma.m8n32k16.mma.row.col.f32.f16
207
Shinji KONO <kono@ie.u-ryukyu.ac.jp>
parents: 150
diff changeset
429 // pre-ptx61-error-re@+1 {{'__hmma_m8n32k16_mma_f32f16' needs target feature (sm_70{{.*}},(ptx61{{.*}}}}
150
anatofuz
parents:
diff changeset
430 __hmma_m8n32k16_mma_f32f16(fdst, src, src, src, 1, 0);
anatofuz
parents:
diff changeset
431 // CHECK_M32_M8: call {{.*}} @llvm.nvvm.wmma.m8n32k16.mma.row.col.f32.f16.satfinite
207
Shinji KONO <kono@ie.u-ryukyu.ac.jp>
parents: 150
diff changeset
432 // pre-ptx61-error-re@+1 {{'__hmma_m8n32k16_mma_f32f16' needs target feature (sm_70{{.*}},(ptx61{{.*}}}}
150
anatofuz
parents:
diff changeset
433 __hmma_m8n32k16_mma_f32f16(fdst, src, src, src, 1, 1);
anatofuz
parents:
diff changeset
434 // CHECK_M32_M8: call {{.*}} @llvm.nvvm.wmma.m8n32k16.mma.col.row.f32.f16
207
Shinji KONO <kono@ie.u-ryukyu.ac.jp>
parents: 150
diff changeset
435 // pre-ptx61-error-re@+1 {{'__hmma_m8n32k16_mma_f32f16' needs target feature (sm_70{{.*}},(ptx61{{.*}}}}
150
anatofuz
parents:
diff changeset
436 __hmma_m8n32k16_mma_f32f16(fdst, src, src, src, 2, 0);
anatofuz
parents:
diff changeset
437 // CHECK_M32_M8: call {{.*}} @llvm.nvvm.wmma.m8n32k16.mma.col.row.f32.f16.satfinite
207
Shinji KONO <kono@ie.u-ryukyu.ac.jp>
parents: 150
diff changeset
438 // pre-ptx61-error-re@+1 {{'__hmma_m8n32k16_mma_f32f16' needs target feature (sm_70{{.*}},(ptx61{{.*}}}}
150
anatofuz
parents:
diff changeset
439 __hmma_m8n32k16_mma_f32f16(fdst, src, src, src, 2, 1);
anatofuz
parents:
diff changeset
440 // CHECK_M32_M8: call {{.*}} @llvm.nvvm.wmma.m8n32k16.mma.col.col.f32.f16
207
Shinji KONO <kono@ie.u-ryukyu.ac.jp>
parents: 150
diff changeset
441 // pre-ptx61-error-re@+1 {{'__hmma_m8n32k16_mma_f32f16' needs target feature (sm_70{{.*}},(ptx61{{.*}}}}
150
anatofuz
parents:
diff changeset
442 __hmma_m8n32k16_mma_f32f16(fdst, src, src, src, 3, 0);
anatofuz
parents:
diff changeset
443 // CHECK_M32_M8: call {{.*}} @llvm.nvvm.wmma.m8n32k16.mma.col.col.f32.f16.satfinite
207
Shinji KONO <kono@ie.u-ryukyu.ac.jp>
parents: 150
diff changeset
444 // pre-ptx61-error-re@+1 {{'__hmma_m8n32k16_mma_f32f16' needs target feature (sm_70{{.*}},(ptx61{{.*}}}}
150
anatofuz
parents:
diff changeset
445 __hmma_m8n32k16_mma_f32f16(fdst, src, src, src, 3, 1);
anatofuz
parents:
diff changeset
446
anatofuz
parents:
diff changeset
447 // CHECK_M32_M8: call {{.*}} @llvm.nvvm.wmma.m8n32k16.mma.row.row.f32.f32
207
Shinji KONO <kono@ie.u-ryukyu.ac.jp>
parents: 150
diff changeset
448 // pre-ptx61-error-re@+1 {{'__hmma_m8n32k16_mma_f32f32' needs target feature (sm_70{{.*}},(ptx61{{.*}}}}
150
anatofuz
parents:
diff changeset
449 __hmma_m8n32k16_mma_f32f32(fdst, src, src, fsrc, 0, 0);
anatofuz
parents:
diff changeset
450 // CHECK_M32_M8: call {{.*}} @llvm.nvvm.wmma.m8n32k16.mma.row.row.f32.f32.satfinite
207
Shinji KONO <kono@ie.u-ryukyu.ac.jp>
parents: 150
diff changeset
451 // pre-ptx61-error-re@+1 {{'__hmma_m8n32k16_mma_f32f32' needs target feature (sm_70{{.*}},(ptx61{{.*}}}}
150
anatofuz
parents:
diff changeset
452 __hmma_m8n32k16_mma_f32f32(fdst, src, src, fsrc, 0, 1);
anatofuz
parents:
diff changeset
453 // CHECK_M32_M8: call {{.*}} @llvm.nvvm.wmma.m8n32k16.mma.row.col.f32.f32
207
Shinji KONO <kono@ie.u-ryukyu.ac.jp>
parents: 150
diff changeset
454 // pre-ptx61-error-re@+1 {{'__hmma_m8n32k16_mma_f32f32' needs target feature (sm_70{{.*}},(ptx61{{.*}}}}
150
anatofuz
parents:
diff changeset
455 __hmma_m8n32k16_mma_f32f32(fdst, src, src, fsrc, 1, 0);
anatofuz
parents:
diff changeset
456 // CHECK_M32_M8: call {{.*}} @llvm.nvvm.wmma.m8n32k16.mma.row.col.f32.f32.satfinite
207
Shinji KONO <kono@ie.u-ryukyu.ac.jp>
parents: 150
diff changeset
457 // pre-ptx61-error-re@+1 {{'__hmma_m8n32k16_mma_f32f32' needs target feature (sm_70{{.*}},(ptx61{{.*}}}}
150
anatofuz
parents:
diff changeset
458 __hmma_m8n32k16_mma_f32f32(fdst, src, src, fsrc, 1, 1);
anatofuz
parents:
diff changeset
459 // CHECK_M32_M8: call {{.*}} @llvm.nvvm.wmma.m8n32k16.mma.col.row.f32.f32
207
Shinji KONO <kono@ie.u-ryukyu.ac.jp>
parents: 150
diff changeset
460 // pre-ptx61-error-re@+1 {{'__hmma_m8n32k16_mma_f32f32' needs target feature (sm_70{{.*}},(ptx61{{.*}}}}
150
anatofuz
parents:
diff changeset
461 __hmma_m8n32k16_mma_f32f32(fdst, src, src, fsrc, 2, 0);
anatofuz
parents:
diff changeset
462 // CHECK_M32_M8: call {{.*}} @llvm.nvvm.wmma.m8n32k16.mma.col.row.f32.f32.satfinite
207
Shinji KONO <kono@ie.u-ryukyu.ac.jp>
parents: 150
diff changeset
463 // pre-ptx61-error-re@+1 {{'__hmma_m8n32k16_mma_f32f32' needs target feature (sm_70{{.*}},(ptx61{{.*}}}}
150
anatofuz
parents:
diff changeset
464 __hmma_m8n32k16_mma_f32f32(fdst, src, src, fsrc, 2, 1);
anatofuz
parents:
diff changeset
465 // CHECK_M32_M8: call {{.*}} @llvm.nvvm.wmma.m8n32k16.mma.col.col.f32.f32
207
Shinji KONO <kono@ie.u-ryukyu.ac.jp>
parents: 150
diff changeset
466 // pre-ptx61-error-re@+1 {{'__hmma_m8n32k16_mma_f32f32' needs target feature (sm_70{{.*}},(ptx61{{.*}}}}
150
anatofuz
parents:
diff changeset
467 __hmma_m8n32k16_mma_f32f32(fdst, src, src, fsrc, 3, 0);
anatofuz
parents:
diff changeset
468 // CHECK_M32_M8: call {{.*}} @llvm.nvvm.wmma.m8n32k16.mma.col.col.f32.f32.satfinite
207
Shinji KONO <kono@ie.u-ryukyu.ac.jp>
parents: 150
diff changeset
469 // pre-ptx61-error-re@+1 {{'__hmma_m8n32k16_mma_f32f32' needs target feature (sm_70{{.*}},(ptx61{{.*}}}}
150
anatofuz
parents:
diff changeset
470 __hmma_m8n32k16_mma_f32f32(fdst, src, src, fsrc, 3, 1);
anatofuz
parents:
diff changeset
471 }
anatofuz
parents:
diff changeset
472 #endif