> -----Original Message----- > From: Beignet [mailto:[email protected]] On Behalf Of > Rebecca N. Palmer > Sent: Thursday, April 23, 2015 10:33 PM > To: [email protected] > Subject: Re: [Beignet] [PATCH] utests: fix test case builtin_tgamma. > > > Should use tgamma instead of tgammaf when generating reference. > That I agree with: we don't want the error of libc tgammaf (~4ulp) appearing > to be our error. > > > And the ulp bound is <= 16ulp > That I wouldn't: it's defined in terms of the infinitely precise value, not > the > correctly-rounded-float value, so (e.g.)16.3ulp would round to 16ulp and > pass the test when it shouldn't. Yeah, this makes sense for me.
> Also, if it's that close to failing for positive inputs, it probably fails for > negative inputs (which this test currently doesn't test). > > > I find that even with your patch the builtin_tgamma still fails. > For me the maximum error is 8ulp on the old test (i.e. positive input and > including tgammaf's error), 4ulp (positive input) 6ulp (negative > input) on my fixed test, so this is a real hardware-dependent error not just > test issues. What result do you get with this patch? On the old test (using tgammaf): What I got was "max error=186.000000 ulp at x=34.633003" Src: 34.633003 cpu: 0x1.e56c58p+125 gpu: 0x1.e56dccp+125 While tgamma returns 0x1.e56dcap+125 for this input. looks like tgammaf on my platform has big ulp error. With your patch: max error=6.000000 ulp at x=-29.857000 > > -------- > > Compare with tgamma instead of tgammaf for better accuracy. > Include negative inputs, and handle the resulting denormals. > Print maximum error found. > > Signed-off-by: Rebecca Palmer <[email protected]> > > diff --git a/utests/builtin_tgamma.cpp b/utests/builtin_tgamma.cpp index > 47cc5f4..b7db69b 100644 > --- a/utests/builtin_tgamma.cpp > +++ b/utests/builtin_tgamma.cpp > @@ -20,10 +20,15 @@ void builtin_tgamma(void) > if (env_strict == NULL || strcmp(env_strict, "0") == 0) > ULPSIZE_FACTOR = 10000.; > > - for (int j = 0; j < 1024; j ++) { > + cl_device_fp_config fp_config; > + clGetDeviceInfo(device, CL_DEVICE_SINGLE_FP_CONFIG, > + sizeof(cl_device_fp_config), &fp_config, 0); bool denormals_supported > + = fp_config & CL_FP_DENORM; float max_ulp = 0, max_ulp_at = 0; > + > + for (int j = 0; j < 128; j ++) { > OCL_MAP_BUFFER(0); > for (int i = 0; i < n; ++i) { > - src[i] = ((float*)buf_data[0])[i] = (j*n+i+1) * 0.001f; > + src[i] = ((float*)buf_data[0])[i] = j - 64 + i*0.001f; > } > OCL_UNMAP_BUFFER(0); > > @@ -32,7 +37,14 @@ void builtin_tgamma(void) > OCL_MAP_BUFFER(1); > float *dst = (float*)buf_data[1]; > for (int i = 0; i < n; ++i) { > - float cpu = tgammaf(src[i]); > + float cpu = tgamma(src[i]); > + if (!denormals_supported && std::fpclassify(cpu)==FP_SUBNORMAL > && dst[i]==0) { > + cpu = 0; > + } > + if (fabsf(cpu - dst[i]) > cl_FLT_ULP(cpu) * max_ulp) { > + max_ulp = fabsf(cpu - dst[i]) / cl_FLT_ULP(cpu); > + max_ulp_at = src[i]; > + } > if (isinf(cpu)) { > OCL_ASSERT(isinf(dst[i])); > } else if (fabsf(cpu - dst[i]) >= cl_FLT_ULP(cpu) * ULPSIZE_FACTOR) > { @@ -42,6 +54,7 @@ void builtin_tgamma(void) > } > OCL_UNMAP_BUFFER(1); > } > + printf("max error=%f ulp at x=%f ", max_ulp, max_ulp_at); > } > > MAKE_UTEST_FROM_FUNCTION(builtin_tgamma); > > _______________________________________________ > Beignet mailing list > [email protected] > http://lists.freedesktop.org/mailman/listinfo/beignet _______________________________________________ Beignet mailing list [email protected] http://lists.freedesktop.org/mailman/listinfo/beignet
