clang/test/CodeGenOpenCL/builtins-amdgcn-wmma-w32.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   v16h  __attribute__((ext_vector_type(16)));
   7 typedef int    v2i   __attribute__((ext_vector_type(2)));
   8 typedef int    v4i   __attribute__((ext_vector_type(4)));
   9 typedef int    v8i   __attribute__((ext_vector_type(8)));
  10 typedef short  v16s  __attribute__((ext_vector_type(16)));
  11
  12 #ifdef WMMA_GFX1100_TESTS
  13
  14 // Wave32
  15
  16 //
  17 // amdgcn_wmma_f32_16x16x16_f16
  18 //
  19
  20 // CHECK-GFX1100-LABEL: @test_amdgcn_wmma_f32_16x16x16_f16_w32
  21 // CHECK-GFX1100: call <8 x float> @llvm.amdgcn.wmma.f32.16x16x16.f16.v8f32(<16 x half> %a, <16 x half> %b, <8 x float> %c)
  22 void test_amdgcn_wmma_f32_16x16x16_f16_w32(global v8f* out, v16h a, v16h b, v8f c)
  23 {
  24   *out = __builtin_amdgcn_wmma_f32_16x16x16_f16_w32(a, b, c);
  25 }
  26
  27 //
  28 // amdgcn_wmma_f32_16x16x16_bf16
  29 //
  30
  31 // CHECK-GFX1100-LABEL: @test_amdgcn_wmma_f32_16x16x16_bf16_w32
  32 // CHECK-GFX1100: call <8 x float> @llvm.amdgcn.wmma.f32.16x16x16.bf16.v8f32(<16 x i16> %a, <16 x i16> %b, <8 x float> %c)
  33 void test_amdgcn_wmma_f32_16x16x16_bf16_w32(global v8f* out, v16s a, v16s b, v8f c)
  34 {
  35   *out = __builtin_amdgcn_wmma_f32_16x16x16_bf16_w32(a, b, c);
  36 }
  37
  38 //
  39 // amdgcn_wmma_f16_16x16x16_f16
  40 //
  41
  42 // CHECK-GFX1100-LABEL: @test_amdgcn_wmma_f16_16x16x16_f16_w32
  43 // CHECK-GFX1100: call <16 x half> @llvm.amdgcn.wmma.f16.16x16x16.f16.v16f16(<16 x half> %a, <16 x half> %b, <16 x half> %c, i1 true)
  44 void test_amdgcn_wmma_f16_16x16x16_f16_w32(global v16h* out, v16h a, v16h b, v16h c)
  45 {
  46   *out = __builtin_amdgcn_wmma_f16_16x16x16_f16_w32(a, b, c, true);
  47 }
  48
  49 //
  50 // amdgcn_wmma_bf16_16x16x16_bf16
  51 //
  52
  53 // CHECK-GFX1100-LABEL: @test_amdgcn_wmma_bf16_16x16x16_bf16_w32
  54 // CHECK-GFX1100: call <16 x i16> @llvm.amdgcn.wmma.bf16.16x16x16.bf16.v16i16(<16 x i16> %a, <16 x i16> %b, <16 x i16> %c, i1 true)
  55 void test_amdgcn_wmma_bf16_16x16x16_bf16_w32(global v16s* out, v16s a, v16s b, v16s c)
  56 {
  57   *out = __builtin_amdgcn_wmma_bf16_16x16x16_bf16_w32(a, b, c, true);
  58 }
  59
  60 //
  61 // amdgcn_wmma_i32_16x16x16_iu8
  62 //
  63
  64 // CHECK-GFX1100-LABEL: @test_amdgcn_wmma_i32_16x16x16_iu8_w32
  65 // CHECK-GFX1100: call <8 x i32> @llvm.amdgcn.wmma.i32.16x16x16.iu8.v8i32(i1 true, <4 x i32> %a, i1 true, <4 x i32> %b, <8 x i32> %c, i1 false)
  66 void test_amdgcn_wmma_i32_16x16x16_iu8_w32(global v8i* out, v4i a, v4i b, v8i c)
  67 {
  68   *out = __builtin_amdgcn_wmma_i32_16x16x16_iu8_w32(true, a, true, b, c, false);
  69 }
  70
  71 //
  72 // amdgcn_wmma_i32_16x16x16_iu4
  73 //
  74
  75 // CHECK-GFX1100-LABEL: @test_amdgcn_wmma_i32_16x16x16_iu4_w32
  76 // CHECK-GFX1100: call <8 x i32> @llvm.amdgcn.wmma.i32.16x16x16.iu4.v8i32(i1 true, <2 x i32> %a, i1 true, <2 x i32> %b, <8 x i32> %c, i1 false)
  77 void test_amdgcn_wmma_i32_16x16x16_iu4_w32(global v8i* out, v2i a, v2i b, v8i c)
  78 {
  79   *out = __builtin_amdgcn_wmma_i32_16x16x16_iu4_w32(true, a, true, b, c, false);
  80 }
  81
  82 #endif