On Tue, 10 Dec 2019, Martin Storsjö wrote:
The stereo_interpolate functions add h_step to the values h BUF_SIZE times. Within the stereo_interpolate C functions, the values h (h0-h3, h00-h13) are declared as local float variables, but the compiler is free to keep them in a register with extra precision.If the accumulation is rounded to 32 bit float precision after each step, the less significant bits of h_step end up ignored and the sum can deviate, affecting the end result more than the currently set EPS. By clearing the log2(BUF_SIZE) lower bits of h_step, we make sure that the accumulation shouldn't differ significantly, regardless of any extra precision in the accmulating register/variable. This fixes the aacpsdsp checkasm test when built with clang for mingw/x86_32. --- tests/checkasm/aacpsdsp.c | 18 ++++++++++++++++++ 1 file changed, 18 insertions(+) diff --git a/tests/checkasm/aacpsdsp.c b/tests/checkasm/aacpsdsp.c index ea68b39fa9..2ceef4341f 100644 --- a/tests/checkasm/aacpsdsp.c +++ b/tests/checkasm/aacpsdsp.c @@ -17,6 +17,7 @@ */ #include "libavcodec/aacpsdsp.h" +#include "libavutil/intfloat.h" #include "checkasm.h" @@ -34,6 +35,16 @@ #define EPS 0.005 +static void clear_less_significant_bits(INTFLOAT *buf, int len, int bits) +{ + int i; + for (i = 0; i < len; i++) { + union av_intfloat32 u = { .f = buf[i] }; + u.i &= (0xffffffff << bits); + buf[i] = u.f; + } +} + static void test_add_squares(void) { LOCAL_ALIGNED_16(INTFLOAT, dst0, [BUF_SIZE]); @@ -198,6 +209,13 @@ static void test_stereo_interpolate(PSDSPContext *psdsp) randomize((INTFLOAT *)h, 2 * 4); randomize((INTFLOAT *)h_step, 2 * 4); + // Clear the least significant 14 bits of h_step, to avoid + // divergence when accumulating h_step BUF_SIZE times into + // a float variable which may or may not have extra intermediate + // precision. Therefore clear roughly log2(BUF_SIZE) less + // significant bits, to get the same result regardless of any + // extra precision in the accumulator. + clear_less_significant_bits((INTFLOAT *)h_step, 2 * 4, 14); call_ref(l0, r0, h, h_step, BUF_SIZE); call_new(l1, r1, h, h_step, BUF_SIZE); -- 2.17.1
Ping // Martin _______________________________________________ ffmpeg-devel mailing list ffmpeg-devel@ffmpeg.org https://ffmpeg.org/mailman/listinfo/ffmpeg-devel To unsubscribe, visit link above, or email ffmpeg-devel-requ...@ffmpeg.org with subject "unsubscribe".
