From 6f0b71692f15c81f63c8d478738372704b230fea Mon Sep 17 00:00:00 2001
From: "Zhong, Zhicong" <zhicong.zhong@intel.com>
Date: Tue, 15 Oct 2024 02:58:28 +0000
Subject: [PATCH] add bf16 test

---
 .../Dialect/Gpu/EltwiseAdd_BF16.mlir          | 39 +++++++++++++++++++
 1 file changed, 39 insertions(+)

diff --git a/test/Integration/Dialect/Gpu/EltwiseAdd_BF16.mlir b/test/Integration/Dialect/Gpu/EltwiseAdd_BF16.mlir
index e76b9b7c7..0b253df30 100644
--- a/test/Integration/Dialect/Gpu/EltwiseAdd_BF16.mlir
+++ b/test/Integration/Dialect/Gpu/EltwiseAdd_BF16.mlir
@@ -51,3 +51,42 @@ module @eltwise_add attributes {gpu.container_module} {
   }
    func.func private @printMemrefBF16(memref<*xbf16>)  attributes {llvm.emit_c_interface}
 }
+
+module @eltwise_add_usm attributes {gpu.container_module} {
+  memref.global "private" constant @__constant_10x20xbf16 : memref<10x20xbf16> = dense<5.000000e-01>
+  func.func @test(%arg0: memref<10x20xbf16>, %arg1: memref<10x20xbf16>) -> memref<10x20xbf16> {
+    %c20 = arith.constant 20 : index
+    %c10 = arith.constant 10 : index
+    %c1 = arith.constant 1 : index
+    %memref = gpu.alloc  host_shared () : memref<10x20xbf16>
+    gpu.launch_func  @test_kernel::@test_kernel blocks in (%c10, %c20, %c1) threads in (%c1, %c1, %c1)  args(%arg0 : memref<10x20xbf16>, %arg1 : memref<10x20xbf16>, %memref : memref<10x20xbf16>)
+    %alloc = memref.alloc() : memref<10x20xbf16>
+    memref.copy %memref, %alloc : memref<10x20xbf16> to memref<10x20xbf16>
+    gpu.dealloc  %memref : memref<10x20xbf16>
+    return %alloc : memref<10x20xbf16>
+  }
+  gpu.module @test_kernel attributes {spirv.target_env = #spirv.target_env<#spirv.vce<v1.0, [Addresses, Float16Buffer, Int64, Int16, Int8, Kernel, Linkage, Vector16, GenericPointer, Groups, Float16, Float64, AtomicFloat32AddEXT, ExpectAssumeKHR, SubgroupDispatch, VectorComputeINTEL, VectorAnyINTEL, Bfloat16ConversionINTEL], [SPV_EXT_shader_atomic_float_add, SPV_KHR_expect_assume, SPV_INTEL_vector_compute, SPV_INTEL_bfloat16_conversion]>, api=OpenCL, #spirv.resource_limits<>>} {
+    gpu.func @test_kernel(%arg0: memref<10x20xbf16>, %arg1: memref<10x20xbf16>, %arg2: memref<10x20xbf16>) kernel attributes {VectorComputeFunctionINTEL, gpu.known_block_size = array<i32: 1, 1, 1>, gpu.known_grid_size = array<i32: 10, 20, 1>, spirv.entry_point_abi = #spirv.entry_point_abi<>} {
+      %block_id_x = gpu.block_id  x
+      %block_id_y = gpu.block_id  y
+      %cst = arith.constant 0.5 : bf16
+      %0 = memref.load %arg0[%block_id_x, %block_id_y] : memref<10x20xbf16>
+      %1 = memref.load %arg1[%block_id_x, %block_id_y] : memref<10x20xbf16>
+      %2 = arith.addf %0, %1 : bf16
+      %3 = arith.addf %2, %cst : bf16
+      memref.store %3, %arg2[%block_id_x, %block_id_y] : memref<10x20xbf16>
+      gpu.return
+    }
+  }
+  func.func @main() {
+    %0 = memref.get_global @__constant_10x20xbf16 : memref<10x20xbf16>
+    %1 = memref.get_global @__constant_10x20xbf16 : memref<10x20xbf16>
+    %2 = call @test(%0, %1) : (memref<10x20xbf16>, memref<10x20xbf16>) -> memref<10x20xbf16>
+    %cast = memref.cast %2 : memref<10x20xbf16> to memref<*xbf16>
+    // CHECK: Unranked Memref base@ = {{(0x)?[-9a-f]*}}
+    // CHECK-COUNT-200: 1.5
+    call @printMemrefbf16(%cast) : (memref<*xbf16>) -> ()
+    return
+  }
+  func.func private @printMemrefBF16(memref<*xbf16>)  attributes {llvm.emit_c_interface}
+}