clang/test/CodeGenOpenCL/builtins-amdgcn-wmma-w64.cl

   1 // REQUIRES: amdgpu-registered-target
   2 // RUN: %clang_cc1 -triple amdgcn-unknown-unknown -target-cpu gfx1100 -DWMMA_GFX1100_TESTS -S -emit-llvm -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 // amdgcn_wmma_f32_16x16x16_f16
  20 //
  21
  22 // CHECK-GFX1100-LABEL: @test_amdgcn_wmma_f32_16x16x16_f16_w64
  23 // CHECK-GFX1100: call <4 x float> @llvm.amdgcn.wmma.f32.16x16x16.f16.v4f32(<16 x half> %a, <16 x half> %b, <4 x float> %c)
  24 void test_amdgcn_wmma_f32_16x16x16_f16_w64(global v4f* out, v16h a, v16h b, v4f c)
  25 {
  26   *out = __builtin_amdgcn_wmma_f32_16x16x16_f16_w64(a, b, c);
  27 }
  28
  29 //
  30 // amdgcn_wmma_f32_16x16x16_bf16
  31 //
  32
  33 // CHECK-GFX1100-LABEL: @test_amdgcn_wmma_f32_16x16x16_bf16_w64
  34 // CHECK-GFX1100: call <4 x float> @llvm.amdgcn.wmma.f32.16x16x16.bf16.v4f32(<16 x i16> %a, <16 x i16> %b, <4 x float> %c)
  35 void test_amdgcn_wmma_f32_16x16x16_bf16_w64(global v4f* out, v16s a, v16s b, v4f c)
  36 {
  37   *out = __builtin_amdgcn_wmma_f32_16x16x16_bf16_w64(a, b, c);
  38 }
  39
  40 //
  41 // amdgcn_wmma_f16_16x16x16_f16
  42 //
  43
  44 // CHECK-GFX1100-LABEL: @test_amdgcn_wmma_f16_16x16x16_f16_w64
  45 // CHECK-GFX1100: call <8 x half> @llvm.amdgcn.wmma.f16.16x16x16.f16.v8f16(<16 x half> %a, <16 x half> %b, <8 x half> %c, i1 true)
  46 void test_amdgcn_wmma_f16_16x16x16_f16_w64(global v8h* out, v16h a, v16h b, v8h c)
  47 {
  48   *out = __builtin_amdgcn_wmma_f16_16x16x16_f16_w64(a, b, c, true);
  49 }
  50
  51 //
  52 // amdgcn_wmma_bf16_16x16x16_bf16
  53 //
  54
  55 // CHECK-GFX1100-LABEL: @test_amdgcn_wmma_bf16_16x16x16_bf16_w64
  56 // CHECK-GFX1100: call <8 x i16> @llvm.amdgcn.wmma.bf16.16x16x16.bf16.v8i16(<16 x i16> %a, <16 x i16> %b, <8 x i16> %c, i1 true)
  57 void test_amdgcn_wmma_bf16_16x16x16_bf16_w64(global v8s* out, v16s a, v16s b, v8s c)
  58 {
  59   *out = __builtin_amdgcn_wmma_bf16_16x16x16_bf16_w64(a, b, c, true);
  60 }
  61
  62 //
  63 // amdgcn_wmma_i32_16x16x16_iu8
  64 //
  65
  66 // CHECK-GFX1100-LABEL: @test_amdgcn_wmma_i32_16x16x16_iu8_w64
  67 // CHECK-GFX1100: call <4 x i32> @llvm.amdgcn.wmma.i32.16x16x16.iu8.v4i32(i1 true, <4 x i32> %a, i1 true, <4 x i32> %b, <4 x i32> %c, i1 false)
  68 void test_amdgcn_wmma_i32_16x16x16_iu8_w64(global v4i* out, v4i a, v4i b, v4i c)
  69 {
  70   *out = __builtin_amdgcn_wmma_i32_16x16x16_iu8_w64(true, a, true, b, c, false);
  71 }
  72
  73 //
  74 // amdgcn_wmma_i32_16x16x16_iu4
  75 //
  76
  77 // CHECK-GFX1100-LABEL: @test_amdgcn_wmma_i32_16x16x16_iu4_w64
  78 // CHECK-GFX1100: call <4 x i32> @llvm.amdgcn.wmma.i32.16x16x16.iu4.v4i32(i1 true, <2 x i32> %a, i1 true, <2 x i32> %b, <4 x i32> %c, i1 false)
  79 void test_amdgcn_wmma_i32_16x16x16_iu4_w64(global v4i* out, v2i a, v2i b, v4i c)
  80 {
  81   *out = __builtin_amdgcn_wmma_i32_16x16x16_iu4_w64(true, a, true, b, c, false);
  82 }
  83
  84 #endif