From 9c53a0db19c8a113329d924cf6c02ce4759c4113 Mon Sep 17 00:00:00 2001
From: Yi Sun <yi.sun@intel.com>
Date: Mon, 9 Sep 2013 16:55:47 +0800
Subject: [PATCH] utest: Add test case for built-in function pow.

Signed-off-by: Yi Sun <yi.sun@intel.com>
Reviewed-by: Zhigang Gong <zhigang.gong@linux.intel.com>
---
 kernels/builtin_pow.cl |  7 ++++
 utests/CMakeLists.txt  |  1 +
 utests/builtin_pow.cpp | 92 ++++++++++++++++++++++++++++++++++++++++++++++++++
 3 files changed, 100 insertions(+)
 create mode 100644 kernels/builtin_pow.cl
 create mode 100644 utests/builtin_pow.cpp

diff --git a/kernels/builtin_pow.cl b/kernels/builtin_pow.cl
new file mode 100644
index 0000000..17d753e
--- /dev/null
+++ b/kernels/builtin_pow.cl
@@ -0,0 +1,7 @@
+kernel void builtin_pow(global float *dst, global float *src1, global float *src2, global int *max_func) {
+
+  int i = get_global_id(0);
+  dst[i * (*max_func) + 0] = pow(src1[i], src2[i]);
+  dst[i * (*max_func) + 1] = src1[i];
+
+}
diff --git a/utests/CMakeLists.txt b/utests/CMakeLists.txt
index 37240fe..b619d52 100644
--- a/utests/CMakeLists.txt
+++ b/utests/CMakeLists.txt
@@ -129,6 +129,7 @@ set (utests_sources
   builtin_num_groups.cpp
   builtin_local_id.cpp
   builtin_acos_asin.cpp
+  builtin_pow.cpp
   builtin_convert_sat.cpp
   runtime_createcontext.cpp
   runtime_null_kernel_arg.cpp
diff --git a/utests/builtin_pow.cpp b/utests/builtin_pow.cpp
new file mode 100644
index 0000000..8ed17ed
--- /dev/null
+++ b/utests/builtin_pow.cpp
@@ -0,0 +1,92 @@
+#include "utest_helper.hpp"
+#include <cmath>
+#include <algorithm>
+
+#define udebug 0
+#define printf_c(...) \
+{\
+  printf("\033[1m\033[40;31m");\
+  printf( __VA_ARGS__ );\
+  printf("\033[0m");\
+}
+const float ori_data[] = {-20.5, -1, -0.9, -0.01, 0, 0.01, 0.9, 1.0, 20.5};
+const int count_input_ori = sizeof(ori_data) / sizeof(ori_data[0]);
+const int count_input = count_input_ori * count_input_ori;
+
+float input_data1[count_input];
+float input_data2[count_input];
+const int max_function = 1;
+
+static void cpu_compiler_math(const float *src1, const float *src2, float *dst)
+{
+  dst[0] = powf(src1[0], src2[0]);
+//  dst[1] = src1[0];
+}
+
+static void builtin_pow(void)
+{
+  // Setup kernel and buffers
+  int k, i, index_cur;
+  float gpu_data[max_function * count_input] = {0}, cpu_data[max_function * count_input] = {0};
+
+  for(i=0; i<count_input_ori;i++)
+    for(k=0; k<count_input_ori;k++)
+    {
+      input_data1[i*count_input_ori+k] = ori_data[i];
+      input_data2[i*count_input_ori+k] = ori_data[k];
+    }
+
+  OCL_CREATE_KERNEL("builtin_pow");
+
+  OCL_CREATE_BUFFER(buf[0], CL_MEM_READ_WRITE, count_input * max_function * sizeof(float), NULL);
+  OCL_CREATE_BUFFER(buf[1], CL_MEM_READ_WRITE, count_input * max_function * sizeof(float), NULL);
+  OCL_CREATE_BUFFER(buf[2], CL_MEM_READ_WRITE, count_input * max_function * sizeof(float), NULL);
+  OCL_CREATE_BUFFER(buf[3], CL_MEM_READ_WRITE, sizeof(int), NULL);
+  OCL_SET_ARG(0, sizeof(cl_mem), &buf[0]);
+  OCL_SET_ARG(1, sizeof(cl_mem), &buf[1]);
+  OCL_SET_ARG(2, sizeof(cl_mem), &buf[2]);
+  OCL_SET_ARG(3, sizeof(cl_mem), &buf[3]);
+
+  globals[0] = count_input;
+  locals[0] = 1;
+
+  clEnqueueWriteBuffer( queue, buf[1], CL_TRUE, 0, count_input * sizeof(float), input_data1, 0, NULL, NULL);
+  clEnqueueWriteBuffer( queue, buf[2], CL_TRUE, 0, count_input * sizeof(float), input_data2, 0, NULL, NULL);
+  clEnqueueWriteBuffer( queue, buf[3], CL_TRUE, 0, sizeof(int), &max_function, 0, NULL, NULL);
+
+   // Run the kernel
+  OCL_NDRANGE( 1 );
+
+  clEnqueueReadBuffer( queue, buf[0], CL_TRUE, 0, sizeof(float) * max_function * count_input, gpu_data, 0, NULL, NULL);
+
+  for (k = 0; (uint)k < count_input; k++)
+  {
+    cpu_compiler_math( input_data1 + k, input_data2 + k, cpu_data + k * max_function);
+
+    for (i = 0; i < max_function; i++)
+    {
+      index_cur = k * max_function + i;
+#if udebug
+      if ( (isinf(cpu_data[index_cur]) && !isinf(gpu_data[index_cur])) ||
+           (isnan(cpu_data[index_cur]) && !isnan(gpu_data[index_cur])) ||
+           (fabs(gpu_data[index_cur] - cpu_data[index_cur]) > 1e-5f)   )
+      {
+        printf_c("%d/%d: x:%f, y:%f -> gpu:%f  cpu:%f\n", k, i, input_data1[k], input_data2[k], gpu_data[index_cur], cpu_data[index_cur]);
+      }
+      else
+        printf("%d/%d: x:%f, y:%f -> gpu:%f  cpu:%f\n", k, i, input_data1[k], input_data2[k], gpu_data[index_cur], cpu_data[index_cur]);
+#else
+     if (isinf(cpu_data[index_cur]))
+       OCL_ASSERT(isinf(gpu_data[index_cur]));
+     else if (isnan(cpu_data[index_cur]))
+       OCL_ASSERT(isnan(gpu_data[index_cur]));
+     else
+     {
+       OCL_ASSERT(fabs(gpu_data[index_cur] - cpu_data[index_cur]) < 1e-3f);
+     }
+#endif
+    }
+  }
+}
+
+MAKE_UTEST_FROM_FUNCTION_WITH_ISSUE(builtin_pow)
-- 
2.7.4