fixup: fixup: benchdnn: eltwise: improve input data coverage

* 1) With fp16 argument, exp() is dispatched to fp16 version as well,
* overflows and produces inf instead of an fp32 value, resulting in
* 0 in the final answer, while it should be a non-zero value exceeding trh.
* 2) MSVC math functions are invariant to compiler options leaving the
* answer same no matter what. Current threshold is low for it. Slightly
* increasing it...

src/cpu/CMakeLists.txt

@@ -45,7 +45,7 @@ if(CMAKE_CXX_COMPILER_ID STREQUAL "Intel")
     file(GLOB FILES_REQUIRED_PREC_DIV
         ${CMAKE_CURRENT_SOURCE_DIR}/*resampling*.cpp
         ${CMAKE_CURRENT_SOURCE_DIR}/*normalization*.cpp
-        ${CMAKE_CURRENT_SOURCE_DIR}/ref_eltwise.cpp
+        ${CMAKE_CURRENT_SOURCE_DIR}/eltwise/ref_eltwise.cpp
         ${CMAKE_CURRENT_SOURCE_DIR}/ref_rnn.cpp)
     if(WIN32)
         set_source_files_properties(${FILES_REQUIRED_PREC_SQRT}

src/gpu/ocl/ocl_post_ops.h

@@ -78,7 +78,7 @@ POST_OP_DATA_T soft_relu_bwd(POST_OP_DATA_T dd, POST_OP_DATA_T s) {
 }
 
 POST_OP_DATA_T logistic_fwd(POST_OP_DATA_T s) {
-    return 1 / (1 + exp(-s));
+    return 1.0f / (1.0f + exp(-s));
 }
 POST_OP_DATA_T logistic_bwd(POST_OP_DATA_T dd, POST_OP_DATA_T s) {
     POST_OP_DATA_T v = logistic_fwd(s);
@@ -162,7 +162,8 @@ POST_OP_DATA_T gelu_tanh_bwd(POST_OP_DATA_T dd, POST_OP_DATA_T s) {
 }
 
 POST_OP_DATA_T swish_fwd(POST_OP_DATA_T s, POST_OP_DATA_T alpha) {
-    return s / (1.0f + exp(-alpha * s));
+    float w = -alpha * s;
+    return s / (1.0f + exp(w));
 }
 POST_OP_DATA_T swish_bwd(
         POST_OP_DATA_T dd, POST_OP_DATA_T s, POST_OP_DATA_T alpha) {

tests/benchdnn/eltwise/eltwise.cpp

@@ -174,7 +174,7 @@ static int compare(const prb_t *p, const dnn_mem_t &mem_arg_fp,
                 || p->alg == alg_t::SRELU || p->alg == alg_t::LOG
                 || (is_fwd && p->alg == alg_t::ELU_DST)
                 || (is_fwd && p->alg == alg_t::TANH_DST))
-            trh = 3e-5;
+            trh = 4e-5;
         else
             trh = 4e-6;
     }
@@ -187,7 +187,7 @@ static int compare(const prb_t *p, const dnn_mem_t &mem_arg_fp,
         const float dt = mem_dt.get_elem(i);
         const float src = mem_arg_fp.get_elem(i);
         const float fp0 = mem_fp.get_elem(i);
-        const float fp = maybe_saturate(p->dt, fp0);
+        const float fp = round_to_nearest_representable(p->dt, fp0);
 
         const float diff = fabsf(fp - dt);
         const float rel_diff = diff / (fabsf(fp) > FLT_MIN ? fabsf(fp) : 1);