Added unit tests for several SPL macros/functions, and detailed all factors

contributing to bit-not-exact between ARM assembly and generic C versions
in iSAC and SPL, by code comments.
Review URL: https://webrtc-codereview.appspot.com/741004

git-svn-id: http://webrtc.googlecode.com/svn/trunk@2673 4adac7df-926f-26a2-2b94-8c16560cd09d

src/build/common.gypi

@@ -170,9 +170,6 @@
         'msvs_disabled_warnings!': [4189,],
       }],
       ['OS=="android"', {
-        # On Android, we always prefer fixed_point to reduce CPU usage.
-        'prefer_fixed_point%': 1,
-
         'defines': [
           'WEBRTC_LINUX',
           'WEBRTC_ANDROID',

src/common_audio/signal_processing/include/spl_inl_armv7.h

@@ -20,6 +20,10 @@
  * (e.g. __builtin_clz).
  */
 
+/* This function produces result that is not bit exact with that by the generic
+ * C version in some cases, although the former is at least as accurate as the
+ * later.
+ */
 static __inline WebRtc_Word32 WEBRTC_SPL_MUL_16_32_RSFT16(WebRtc_Word16 a,
                                                           WebRtc_Word32 b) {
   WebRtc_Word32 tmp = 0;
@@ -27,6 +31,10 @@ static __inline WebRtc_Word32 WEBRTC_SPL_MUL_16_32_RSFT16(WebRtc_Word16 a,
   return tmp;
 }
 
+/* This function produces result that is not bit exact with that by the generic
+ * C version in some cases, although the former is at least as accurate as the
+ * later.
+ */
 static __inline WebRtc_Word32 WEBRTC_SPL_MUL_32_32_RSFT32(WebRtc_Word16 a,
                                                           WebRtc_Word16 b,
                                                           WebRtc_Word32 c) {

src/common_audio/signal_processing/signal_processing_unittest.cc

@@ -70,9 +70,7 @@ TEST_F(SplTest, MacroTest) {
     EXPECT_EQ(5, WEBRTC_SPL_MUL_32_32_RSFT32(a32a, a32b, A));
     EXPECT_EQ(5, WEBRTC_SPL_MUL_32_32_RSFT32BI(a32, A));
 
-    EXPECT_EQ(-49149, WEBRTC_SPL_MUL_16_16(a, b));
     EXPECT_EQ(-12288, WEBRTC_SPL_MUL_16_16_RSFT(a, b, 2));
-
     EXPECT_EQ(-12287, WEBRTC_SPL_MUL_16_16_RSFT_WITH_ROUND(a, b, 2));
     EXPECT_EQ(-1, WEBRTC_SPL_MUL_16_16_RSFT_WITH_FIXROUND(a, b));
 
@@ -106,6 +104,31 @@ TEST_F(SplTest, MacroTest) {
     EXPECT_EQ(32766u, WEBRTC_SPL_LSHIFT_U32(a, 1));
 
     EXPECT_EQ(1470, WEBRTC_SPL_RAND(A));
+
+    EXPECT_EQ(-49149, WEBRTC_SPL_MUL_16_16(a, b));
+    EXPECT_EQ(1073676289, WEBRTC_SPL_MUL_16_16(WEBRTC_SPL_WORD16_MAX,
+                                               WEBRTC_SPL_WORD16_MAX));
+    EXPECT_EQ(1073709055, WEBRTC_SPL_MUL_16_32_RSFT16(WEBRTC_SPL_WORD16_MAX,
+                                                      WEBRTC_SPL_WORD32_MAX));
+    EXPECT_EQ(1073741824, WEBRTC_SPL_MUL_16_32_RSFT16(WEBRTC_SPL_WORD16_MIN,
+                                                      WEBRTC_SPL_WORD32_MIN));
+#ifdef WEBRTC_ARCH_ARM_V7A
+    EXPECT_EQ(-1073741824,
+              WEBRTC_SPL_MUL_16_32_RSFT16(WEBRTC_SPL_WORD16_MIN,
+                                          WEBRTC_SPL_WORD32_MAX));
+    EXPECT_EQ(0x3fffffff, WEBRTC_SPL_MUL_32_32_RSFT32(WEBRTC_SPL_WORD16_MAX,
+              0xffff, WEBRTC_SPL_WORD32_MAX));
+    EXPECT_EQ(0x3fffffff, WEBRTC_SPL_MUL_32_32_RSFT32BI(WEBRTC_SPL_WORD32_MAX,
+              WEBRTC_SPL_WORD32_MAX));
+#else
+    EXPECT_EQ(-1073741823,
+              WEBRTC_SPL_MUL_16_32_RSFT16(WEBRTC_SPL_WORD16_MIN,
+                                          WEBRTC_SPL_WORD32_MAX));
+    EXPECT_EQ(0x3fff7ffe, WEBRTC_SPL_MUL_32_32_RSFT32(WEBRTC_SPL_WORD16_MAX,
+              0xffff, WEBRTC_SPL_WORD32_MAX));
+    EXPECT_EQ(0x3ffffffd, WEBRTC_SPL_MUL_32_32_RSFT32BI(WEBRTC_SPL_WORD32_MAX,
+                                                        WEBRTC_SPL_WORD32_MAX));
+#endif
 }
 
 TEST_F(SplTest, InlineTest) {

src/modules/audio_coding/codecs/isac/fix/source/lattice.c

@@ -45,11 +45,12 @@ void WebRtcIsacfix_FilterArLoop(int16_t* ar_g_Q0,
                                 int16_t* sth_Q15,
                                 int16_t order_coef);
 
-/* Inner loop used for function WebRtcIsacfix_NormLatticeFilterMa().
+/* Inner loop used for function WebRtcIsacfix_NormLatticeFilterMa(). It does:
-   It does:
    for 0 <= n < HALF_SUBFRAMELEN - 1:
      *ptr2 = input2 * (*ptr2) + input0 * (*ptr0));
      *ptr1 = input1 * (*ptr0) + input0 * (*ptr2);
+   Note, function WebRtcIsacfix_FilterMaLoopNeon and WebRtcIsacfix_FilterMaLoopC
+   are not bit-exact. The accuracy by the ARM Neon function is same or better.
 */
 void WebRtcIsacfix_FilterMaLoopC(int16_t input0,  // Filter coefficient
                                  int16_t input1,  // Filter coefficient