diff options
author | rander <rander.wang@intel.com> | 2017-05-19 16:14:40 +0800 |
---|---|---|
committer | Yang Rong <rong.r.yang@intel.com> | 2017-06-09 19:45:55 +0800 |
commit | 8ed276000efe129481ba0f83643bf5a8a2389309 (patch) | |
tree | e5842940f1923da76127619609b517ce871ded80 | |
parent | 0fd1ef5a1643973c1369b203ce596729e00da8ec (diff) |
utests: add utest for sqrt-div optimization
Signed-off-by: rander.wang <rander.wang@intel.com>
Reviewed-by: Yang Rong <rong.r.yang@intel.com>
-rw-r--r-- | kernels/compiler_sqrt_div.cl | 8 | ||||
-rw-r--r-- | utests/CMakeLists.txt | 3 | ||||
-rw-r--r-- | utests/compiler_sqrt_div.cpp | 61 |
3 files changed, 71 insertions, 1 deletions
diff --git a/kernels/compiler_sqrt_div.cl b/kernels/compiler_sqrt_div.cl new file mode 100644 index 00000000..7d5a2f08 --- /dev/null +++ b/kernels/compiler_sqrt_div.cl @@ -0,0 +1,8 @@ +kernel void compiler_sqrt_div(global float *src, global float *dst) { + int i = get_global_id(0); + float tmp = sqrt(src[i]); + dst[i*4] = 1.0f/tmp; + dst[i*4+1] = (float)i/tmp; + dst[i*4+2] = 2.0f/tmp; + dst[i*4+3] = 1.0f/tmp + tmp; +}; diff --git a/utests/CMakeLists.txt b/utests/CMakeLists.txt index b84cdd66..f8abd457 100644 --- a/utests/CMakeLists.txt +++ b/utests/CMakeLists.txt @@ -303,7 +303,8 @@ set (utests_sources compiler_generic_pointer.cpp runtime_pipe_query.cpp compiler_pipe_builtin.cpp - compiler_device_enqueue.cpp) + compiler_device_enqueue.cpp + compiler_sqrt_div.cpp) if (LLVM_VERSION_NODOT VERSION_GREATER 34) SET(utests_sources diff --git a/utests/compiler_sqrt_div.cpp b/utests/compiler_sqrt_div.cpp new file mode 100644 index 00000000..c22c5a39 --- /dev/null +++ b/utests/compiler_sqrt_div.cpp @@ -0,0 +1,61 @@ +#include "utest_helper.hpp" +#include <cmath> + +void compiler_sqrt_div(void) { + const int n = 1024; + float src[n]; + + // Setup kernel and buffers + OCL_CREATE_KERNEL("compiler_sqrt_div"); + OCL_CREATE_BUFFER(buf[0], 0, n * sizeof(float), NULL); + OCL_CREATE_BUFFER(buf[1], 0, n * 4 * sizeof(float), NULL); + OCL_SET_ARG(0, sizeof(cl_mem), &buf[0]); + OCL_SET_ARG(1, sizeof(cl_mem), &buf[1]); + globals[0] = n; + locals[0] = 16; + + for (int j = 0; j < 1024; j++) { + OCL_MAP_BUFFER(0); + for (int i = 0; i < n; ++i) { + src[i] = ((float *)buf_data[0])[i] = (j * n + i + 1) * 0.001f; + } + OCL_UNMAP_BUFFER(0); + + OCL_NDRANGE(1); + + OCL_MAP_BUFFER(1); + float *dst = (float *)buf_data[1]; + for (int i = 0; i < n; ++i) { + float cpu = 1.0f / sqrt(src[i]); + float gpu = dst[4 * i]; + if (fabsf(cpu - gpu) >= 1e-3) { + printf("%f %f %f", src[i], cpu, gpu); + OCL_ASSERT(0); + } + + cpu = i / sqrt(src[i]); + gpu = dst[4 * i + 1]; + if (fabsf(cpu - gpu) >= 1e-3) { + printf("%f %f %f", src[i], cpu, gpu); + OCL_ASSERT(0); + } + + cpu = 2.0f / sqrt(src[i]); + gpu = dst[4 * i + 2]; + if (fabsf(cpu - gpu) >= 1e-3) { + printf("%f %f %f", src[i], cpu, gpu); + OCL_ASSERT(0); + } + + cpu = 1.0f / sqrt(src[i]) + sqrt(src[i]); + gpu = dst[4 * i + 3]; + if (fabsf(cpu - gpu) >= 1e-3) { + printf("%f %f %f", src[i], cpu, gpu); + OCL_ASSERT(0); + } + } + OCL_UNMAP_BUFFER(1); + } +} + +MAKE_UTEST_FROM_FUNCTION(compiler_sqrt_div); |