cross-project-tests/amdgpu/builtins-amdgcn-wmma-w64.cl

   1 // NOTE: Assertions have been autogenerated by utils/update_cc_test_checks.py
   2 // RUN: %clang_cc1 -triple amdgcn-unknown-unknown -target-cpu gfx1100 -target-feature +wavefrontsize64 -DWMMA_GFX1100_TESTS -S -o - %s | FileCheck %s --check-prefix=CHECK-GFX1100
   3
   4 typedef float  v4f   __attribute__((ext_vector_type(4)));
   5 typedef float  v8f   __attribute__((ext_vector_type(8)));
   6 typedef half   v8h   __attribute__((ext_vector_type(8)));
   7 typedef half   v16h  __attribute__((ext_vector_type(16)));
   8 typedef int    v2i   __attribute__((ext_vector_type(2)));
   9 typedef int    v4i   __attribute__((ext_vector_type(4)));
  10 typedef int    v8i   __attribute__((ext_vector_type(8)));
  11 typedef short  v8s   __attribute__((ext_vector_type(8)));
  12 typedef short  v16s  __attribute__((ext_vector_type(16)));
  13
  14 #ifdef WMMA_GFX1100_TESTS
  15
  16 // Wave64
  17
  18
  19 // CHECK-GFX1100-LABEL: test_amdgcn_wmma_f32_16x16x16_f16_w64:
  20 // CHECK-GFX1100: v_wmma_f32_16x16x16_f16 v[{{.*}}], v[{{.*}} v[{{.*}}], v[{{.*}}]
  21 //
  22 void test_amdgcn_wmma_f32_16x16x16_f16_w64(global v4f* out, v16h a, v16h b, v4f c)
  23 {
  24   *out = __builtin_amdgcn_wmma_f32_16x16x16_f16_w64(a, b, c);
  25 }
  26
  27
  28 // CHECK-GFX1100-LABEL: test_amdgcn_wmma_f32_16x16x16_bf16_w64:
  29 // CHECK-GFX1100: v_wmma_f32_16x16x16_bf16 v[{{.*}}], v[{{.*}} v[{{.*}}], v[{{.*}}]
  30 //
  31 void test_amdgcn_wmma_f32_16x16x16_bf16_w64(global v4f* out, v16s a, v16s b, v4f c)
  32 {
  33   *out = __builtin_amdgcn_wmma_f32_16x16x16_bf16_w64(a, b, c);
  34 }
  35
  36
  37 // CHECK-GFX1100-LABEL: test_amdgcn_wmma_f16_16x16x16_f16_w64:
  38 // CHECK-GFX1100: v_wmma_f16_16x16x16_f16 v[{{.*}}], v[{{.*}} v[{{.*}}], v[{{.*}}] op_sel:[0,0,1]
  39 //
  40 void test_amdgcn_wmma_f16_16x16x16_f16_w64(global v8h* out, v16h a, v16h b, v8h c)
  41 {
  42   *out = __builtin_amdgcn_wmma_f16_16x16x16_f16_w64(a, b, c, true);
  43 }
  44
  45
  46 // CHECK-GFX1100-LABEL: test_amdgcn_wmma_bf16_16x16x16_bf16_w64:
  47 // CHECK-GFX1100: v_wmma_bf16_16x16x16_bf16 v[{{.*}}], v[{{.*}} v[{{.*}}], v[{{.*}}] op_sel:[0,0,1]
  48 //
  49 void test_amdgcn_wmma_bf16_16x16x16_bf16_w64(global v8s* out, v16s a, v16s b, v8s c)
  50 {
  51   *out = __builtin_amdgcn_wmma_bf16_16x16x16_bf16_w64(a, b, c, true);
  52 }
  53
  54
  55 // CHECK-GFX1100-LABEL: test_amdgcn_wmma_i32_16x16x16_iu8_w64:
  56 // CHECK-GFX1100: v_wmma_i32_16x16x16_iu8 v[{{.*}}], v[{{.*}} v[{{.*}} v[{{.*}}] neg_lo:[1,1,0]
  57 //
  58 void test_amdgcn_wmma_i32_16x16x16_iu8_w64(global v4i* out, v4i a, v4i b, v4i c)
  59 {
  60   *out = __builtin_amdgcn_wmma_i32_16x16x16_iu8_w64(true, a, true, b, c, false);
  61 }
  62
  63
  64 // CHECK-GFX1100-LABEL: test_amdgcn_wmma_i32_16x16x16_iu4_w64:
  65 // CHECK-GFX1100: v_wmma_i32_16x16x16_iu4 v[{{.*}} v[{{.*}} v[{{.*}} v[{{.*}}neg_lo:[1,1,0]
  66 //
  67 void test_amdgcn_wmma_i32_16x16x16_iu4_w64(global v4i* out, v2i a, v2i b, v4i c)
  68 {
  69   *out = __builtin_amdgcn_wmma_i32_16x16x16_iu4_w64(true, a, true, b, c, false);
  70 }
  71
  72 #endif