diff options
| -rw-r--r-- | kernels/compiler_fdiv2rcp.cl | 8 | ||||
| -rw-r--r-- | utests/CMakeLists.txt | 3 | ||||
| -rw-r--r-- | utests/compiler_fdiv2rcp.cpp | 61 |
3 files changed, 71 insertions, 1 deletions
diff --git a/kernels/compiler_fdiv2rcp.cl b/kernels/compiler_fdiv2rcp.cl new file mode 100644 index 00000000..99e0005e --- /dev/null +++ b/kernels/compiler_fdiv2rcp.cl @@ -0,0 +1,8 @@ +kernel void compiler_fdiv2rcp(global float *src, global float *dst) { + int i = get_global_id(0); + float tmp = 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] = 3.0f/tmp; +}; diff --git a/utests/CMakeLists.txt b/utests/CMakeLists.txt index ebbf0f56..655a314b 100644 --- a/utests/CMakeLists.txt +++ b/utests/CMakeLists.txt @@ -306,7 +306,8 @@ set (utests_sources compiler_pipe_builtin.cpp compiler_device_enqueue.cpp compiler_sqrt_div.cpp - compiler_remove_negative_add.cpp) + compiler_remove_negative_add.cpp + compiler_fdiv2rcp.cpp) if (LLVM_VERSION_NODOT VERSION_GREATER 34) SET(utests_sources diff --git a/utests/compiler_fdiv2rcp.cpp b/utests/compiler_fdiv2rcp.cpp new file mode 100644 index 00000000..7a983920 --- /dev/null +++ b/utests/compiler_fdiv2rcp.cpp @@ -0,0 +1,61 @@ +#include "utest_helper.hpp" +#include <cmath> + +void compiler_fdiv2rcp(void) { + const int n = 1024; + float src[n]; + + // Setup kernel and buffers + OCL_CREATE_KERNEL("compiler_fdiv2rcp"); + 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 / (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 / (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 / (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 = 3.0f / (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_fdiv2rcp); |