fixed FEDP emulation MACROs

Author: tinebp

ci/regression.sh.in

@@ -494,7 +494,7 @@ tensor()
     make -C hw/unittest/tcu_fedp clean && CONFIGS="-DTCU_TYPE_DPI" NUM_REGS=1 LATENCY=4 make -C hw/unittest/tcu_fedp && hw/unittest/tcu_fedp/tcu_fedp --fmt=1  --no-fused
     make -C hw/unittest/tcu_fedp clean && CONFIGS="-DTCU_TYPE_BHF" NUM_REGS=1 LATENCY=10 make -C hw/unittest/tcu_fedp && hw/unittest/tcu_fedp/tcu_fedp --fmt=1 --no-fused
     make -C hw/unittest/tcu_fedp clean && CONFIGS="-DTCU_TYPE_DSP" NUM_REGS=1 LATENCY=31 make -C hw/unittest/tcu_fedp && hw/unittest/tcu_fedp/tcu_fedp --fmt=1 --no-fused --ulp=3
-    make -C hw/unittest/tcu_fedp clean && CONFIGS="-DTCU_TYPE_DRL -DUSE_FEDP" NUM_REGS=2 LATENCY=4 make -C hw/unittest/tcu_fedp && hw/unittest/tcu_fedp/tcu_fedp --fmt=1 --ulp=2
+    make -C hw/unittest/tcu_fedp clean && CONFIGS="-DTCU_TYPE_DRL -DUSE_FEDP" NUM_REGS=2 LATENCY=4 make -C hw/unittest/tcu_fedp && hw/unittest/tcu_fedp/tcu_fedp --fmt=1 --no-zeros --no-subnormals --no-infinities --no-nans
 
     # test bf16
     make -C hw/unittest/tcu_fedp clean && CONFIGS="-DTCU_TYPE_DPI" NUM_REGS=1 LATENCY=4  make -C hw/unittest/tcu_fedp && hw/unittest/tcu_fedp/tcu_fedp --fmt=2 --no-fused

hw/rtl/tcu/drl/VX_tcu_fedp_drl.sv

@@ -16,7 +16,7 @@
 module VX_tcu_fedp_drl #(
     parameter LATENCY = 0,
     parameter N = 2,
-    parameter W = 53
+    parameter W = 25
 ) (
     input  wire clk,
     input  wire reset,

hw/unittest/tcu_fedp/fedp.h

@@ -22,7 +22,7 @@
 #include <unordered_map>
 #include <vector>
 
-#if FEDP_TRACE
+#ifdef FEDP_TRACE
 #include <cstdio>
 #define LOG(...) std::fprintf(stderr, __VA_ARGS__);
 #else

hw/unittest/tcu_fedp/main.cpp

@@ -43,7 +43,7 @@
 #include <bitmanip.h>
 #include "softfloat_ext.h"
 
-#ifdef FEDP_EMUL
+#ifdef USE_FEDP
 #include "fedp.h"
 #endif
 
@@ -280,7 +280,7 @@ static void pack_elements(const std::vector<uint32_t> &elements, int element_bit
   }
 }
 
-#ifndef FEDP_EMUL
+#ifndef USE_FEDP
 // Calculate expected fp dot product
 static float dot_product(const uint32_t* A, const uint32_t* B, uint32_t C, int n, int eb, int sb, bool fused) {
   auto to_float = [&](uint32_t x, int ebits, int sbits) -> long double {
@@ -679,7 +679,7 @@ class Testbench {
     const uint32_t NF = features_to_test.size();
     const uint32_t tests_per_feature = (NT + NF - 1) / NF;
 
-  #ifdef FEDP_EMUL
+  #ifdef USE_FEDP
     FEDP fedp(config_.exp_bits, config_.sig_bits, NUM_REGS * 2, (int)config_.frm, config_.W, config_.renorm);
   #endif
 
@@ -746,7 +746,7 @@ class Testbench {
       std::memcpy(&dut_result, &dut_result_bits, sizeof(float));
 
       // Calculate expected result
-    #ifdef FEDP_EMUL
+    #ifdef USE_FEDP
       float expected = fedp(a_packed.data(), b_packed.data(), c_value_float, NUM_REGS);
     #else
       float expected = dot_product(a_value_hex.data(), b_value_hex.data(),