Re: wined3d: Use higher precision to convert floats into text.
On Monday 26 May 2008 12:37:04 pm Vitaliy Margolen wrote:
- shader_addline(buffer, "PARAM srgb_mul_low = {%f, %f, %f, 1.0};\n", + shader_addline(buffer, "PARAM srgb_mul_low = {%.12f, %.12f, %.12f, 1.0};\n",
I don't know if ARB shaders would except the notation (GLSL does), but would %.8e or somesuch work better than %.12f at maintaining accuracy?
Chris Robinson wrote:
On Monday 26 May 2008 12:37:04 pm Vitaliy Margolen wrote:
- shader_addline(buffer, "PARAM srgb_mul_low = {%f, %f, %f, 1.0};\n", + shader_addline(buffer, "PARAM srgb_mul_low = {%.12f, %.12f, %.12f, 1.0};\n",
I don't know if ARB shaders would except the notation (GLSL does), but would %.8e or somesuch work better than %.12f at maintaining accuracy?
I thought they were clamped? In either case .12 should be enough precision to avoid problems. Or the whole local constants needs to be rolled back. Translating binary -> text -> binary is IMHO wrong. Vitaliy.
Am Montag, den 26.05.2008, 19:00 -0600 schrieb Vitaliy Margolen:
Chris Robinson wrote:
On Monday 26 May 2008 12:37:04 pm Vitaliy Margolen wrote:
- shader_addline(buffer, "PARAM srgb_mul_low = {%f, %f, %f, 1.0};\n", + shader_addline(buffer, "PARAM srgb_mul_low = {%.12f, %.12f, %.12f, 1.0};\n",
I don't know if ARB shaders would except the notation (GLSL does), but would %.8e or somesuch work better than %.12f at maintaining accuracy? I thought they were clamped? In either case .12 should be enough precision to avoid problems.
There are two problems with %f that %e readily avoids. Only one of them is irrelevant for clamped values (I take your word for that. I am not experienced in 3D shader languages at all, but I recall that when I skimmed over some documentation yesterday, it seemed to mention a range of -63..+63.) The first problem of %f is unpredictable large strings: char buffer[50]; sprintf(buffer,"%f",1e100); is a nice way of producing a buffer overflow. Now imagine the number coming from a remote system. "%40f" does not help. It just sets a minimum length, not a maximum length, so it will still overflow. Unlike strings, "%.40f" won't help either, as it just causes 40 digits after the decimal dot/comma to be printed, but does not limit what is printed before. This problem is irrelevant for clamped values. The second problem of %f is unpredicted loss of precision: printf("%.6f %.5f\n",1.467e-6,1.467e-6); will print "0.000001 0.00000", so the first value is of by 46%, the second one is lost completely. As I do not know whether such small values are used in Direct3D (but having problems fixed by "%.12f" strongly indicates it!), I am not sure whether this point applies here. Now consider the two examples with "%.8e": In the first case, the buffer will contain "1.00000000e+100" which fits nicely into the 50 allocated characters. In the second case, lets assume a format string of "%.6e %.2e", the output will be "1.467000e-6 1.46e-6" In the second case, precision is still lost, but as one can be sure that the error is smaller than half a percent (extreme case is 1.0049). In this case (passing numbers as text between two programs, not intended for the user), I would definitely plead for using "%.8e" which should even enable round-trips for every 32 bit float pattern. On presenting numbers two the user, %g might be worth considering. %f is IMHO inappropriate if you are not very confident that the number you want to print is in a sensible order of magnitude. (Using "%.3f" is OK, even if the value might be very near to zero, if you just care whether it is less then 0.0005, and smaller values are definitely rounding errors. No one likes to see 2.4 - 10*0.24=1.6e-16 or something like that. On the other hand 2.4e-16 - 9*0.24e-16 should usually not print as zero) Regards, Michael Karcher
As Henri said in the bug report, we can just declare the LC* uniform instead of constant and load them once at program link time. This way we loose the ability to tell the compiler about the constants, but I doubt this changes anything, but we can simply pass a float * to opengl witohut having to convert to a text. For ARB shaders we can use the program-local environment and load it at compilation time.
Stefan Dösinger wrote:
As Henri said in the bug report, we can just declare the LC* uniform instead of constant and load them once at program link time. This way we loose the ability to tell the compiler about the constants, but I doubt this changes anything, but we can simply pass a float * to opengl witohut having to convert to a text. For ARB shaders we can use the program-local environment and load it at compilation time.
Whatever you want to do - I'd like to see this bug fixed before the next version is out. All the discussions about "possible options" doesn't get it anywhere. Vitaliy.
Am Dienstag, 27. Mai 2008 14:57:49 schrieb Vitaliy Margolen:
Whatever you want to do - I'd like to see this bug fixed before the next version is out. All the discussions about "possible options" doesn't get it anywhere. I'll see what I can do. Unfortunately I am quite busy, but I guess in the time spent discussing this bug I could have written a bugfix as well
Stefan Dösinger wrote:
Am Dienstag, 27. Mai 2008 14:57:49 schrieb Vitaliy Margolen:
Whatever you want to do - I'd like to see this bug fixed before the next version is out. All the discussions about "possible options" doesn't get it anywhere. I'll see what I can do. Unfortunately I am quite busy, but I guess in the time spent discussing this bug I could have written a bugfix as well
Thank you for looking into this. Vitaliy.
participants (4)
-
Chris Robinson -
Michael Karcher -
Stefan Dösinger -
Vitaliy Margolen