fix(runtime): truncate and saturate FPU CVT.W.S like the EE FPU

ps2xRuntime/include/ps2_runtime_macros.h

@@ -311,6 +311,17 @@ inline float ps2_frsqrt(float a, float b)
     return ps2_fall_normal(a, b, r) ? r : ps2_frsqrt_slow(a, b);
 }
 
+// CVT.W.S: the EE always truncates toward zero, whatever FCR31 says. A magnitude of 2^31 or more
+// saturates by sign to 0x7FFFFFFF / 0x80000000. The EE has no Inf/NaN, so exponent-255 patterns are
+// just large numbers and saturate the same way (as PCSX2's interpreter CVT_W does).
+inline int32_t ps2_fcvt_w(float a)
+{
+    const uint32_t bits = std::bit_cast<uint32_t>(a);
+    if ((bits & PS2_FLOAT_EXP_MASK) <= 0x4E800000u) // exponent <= 157: |a| < 2^31
+        return static_cast<int32_t>(a);
+    return (bits & PS2_FLOAT_SIGN_MASK) ? INT32_MIN : INT32_MAX;
+}
+
 enum class Ps2VecOp
 {
     Add,
@@ -853,7 +864,7 @@ inline __m128i ps2_u64_to_epi64_pair(uint64_t value)
 #define FPU_FLOOR_W_S(a) ((int32_t)floorf((float)(a)))
 #define FPU_CVT_S_W(a) ((float)(int32_t)(a))
 #define FPU_CVT_S_L(a) ((float)(int64_t)(a))
-#define FPU_CVT_W_S(a) ((int32_t)nearbyintf((float)(a)))
+#define FPU_CVT_W_S(a) ps2_fcvt_w((float)(a))
 #define FPU_CVT_L_S(a) ((int64_t)(float)(a))
 #define FPU_C_F_S(a, b) (0)
 #define FPU_C_UN_S(a, b) (isnan((float)(a)) || isnan((float)(b)))

ps2xTest/src/ps2_float_semantics_tests.cpp

@@ -224,5 +224,24 @@ void register_ps2_float_semantics_tests()
             t.IsTrue(rsqrt.find("ctx->f[2] = FPU_RSQRT_S(ctx->f[3], ctx->f[4])") != std::string::npos,
                      "RSQRT.S should compute fs / sqrt(ft) but got: " + rsqrt);
         });
+
+        tc.Run("FPU CVT.W.S truncates toward zero and saturates by sign", [](TestCase &t)
+        {
+            const auto cvt = [](float f) { return static_cast<uint32_t>(FPU_CVT_W_S(f)); };
+            // 2.5 and 3.7 differ between round-to-nearest (2/4) and truncation (2/3).
+            t.Equals(cvt(3.7f), 3u, "3.7 should truncate to 3");
+            t.Equals(cvt(-3.7f), static_cast<uint32_t>(-3), "-3.7 should truncate toward zero to -3");
+            t.Equals(cvt(0.999f), 0u, "0.999 should truncate to 0");
+            // A volume scale like 0.9 * 0x3FFF = 14744.7 truncates to 14744, not 14745.
+            t.Equals(cvt(FPU_MUL_S(0.9f, 16383.0f)), 14744u, "0.9 * 0x3FFF should truncate to 14744");
+            t.Equals(cvt(floatOf(0x4EFFFFFFu)), 0x7FFFFF80u, "the largest float below 2^31 should convert exactly");
+            t.Equals(cvt(2147483648.0f), 0x7FFFFFFFu, "2^31 should saturate to 0x7FFFFFFF");
+            t.Equals(cvt(-2147483648.0f), 0x80000000u, "-2^31 should give 0x80000000");
+            t.Equals(cvt(1.0e20f), 0x7FFFFFFFu, "1e20 should saturate to 0x7FFFFFFF");
+            t.Equals(cvt(-1.0e20f), 0x80000000u, "-1e20 should saturate to 0x80000000");
+            t.Equals(cvt(floatOf(0x7FFFFFFFu)), 0x7FFFFFFFu, "+max (a NaN pattern on the host) should saturate to 0x7FFFFFFF");
+            t.Equals(cvt(floatOf(0xFFC00000u)), 0x80000000u, "a negative exponent-255 pattern should saturate to 0x80000000");
+            t.Equals(cvt(floatOf(0x00000001u)), 0u, "a denormal should convert to 0");
+        });
     });
 }