Support IEEE 754-2008 16-bits floats in bitstrings #2890

josevalim · 2020-11-28T08:29:53Z

This format is often used with computer graphics and
AI, also in cases where data is large enough that the
decreased bandwidth and storage compensates the loss
of precision.

The _Float16 data type is defined in the C11 extension
ISO/IEC TS 18661-3:2015. Recent compilers and platforms
provide support for it and recent CPUs provide specific
instruction sets for them too. In some, _Float16 is storage
only, which is enough for our use cases. If _Float16 is not
available, functions for conversion from float to half-floats
are used, ported from https://github.com/Maratyszcza/FP16.

I am sending this pull request to collect feedback. I will
work on a complete test suite if there is interest in moving
this forward.

jhogberg

Thanks for the PR! I think it's a pretty reasonable feature, we'll take it up on our next technical board meeting.

erts/emulator/beam/erl_bits.c

jhogberg · 2021-01-15T11:20:45Z

Sorry about the late reply, I missed the technical board meeting last month due to illness. We're interested in having this :)

josevalim · 2021-01-15T11:37:15Z

Yes!!! I will work on the test suite today! Thanks for the updates! :)

josevalim · 2021-01-15T17:50:44Z

Tests added and feedback addressed!

erts/emulator/beam/erl_bits.c

lib/stdlib/src/erl_lint.erl

jhogberg · 2021-02-02T11:25:44Z

The debugger test case bs_match_misc_SUITE:t_float/1 has started to fail because 16-bit floats no longer crash, can you change it to try matching 15 bits instead?

josevalim · 2021-02-02T13:57:22Z

@jhogberg done. I changed it to 8 which is what I did in the other similar suites.

erts/emulator/beam/erl_bits_f16.h

jhogberg · 2021-02-03T10:03:02Z

erts/emulator/beam/erl_bits_f16.h

+ * OTHERWISE, ARISING FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER
+ * DEALINGS IN THE SOFTWARE.
+ */
+static ERTS_INLINE float fp32_from_bits(uint32_t w) {


Sorry about the back and forth, but we need to change all instances of uint??_t in this file too. :(

No problem! I should probably get a Windows machine next time I am working on this part of ERTS. :)

I have done the requested changes. I have also added a short notice at the top of this file (below the copyright disclaimer) to list all of the modifications we have done. This will make it easier to track changes or update the code in the future.

jhogberg · 2021-02-03T11:01:38Z

erts/emulator/beam/erl_bits_f16.h

+    exp_bits = (bits >> 13) & UINT32_C(0x00007C00);
+    mantissa_bits = bits & UINT32_C(0x00000FFF);
+    nonsign = exp_bits + mantissa_bits;
+    return (sign >> 16) | (shl1_w > UINT32_C(0xFF000000) ? UINT16_C(0x7E00) : nonsign);


[2021-02-03T10:53:10.718Z] erl_bits.o : error LNK2019: unresolved external symbol UINT32_C referenced in function erts_bs_get_float_2 [2021-02-03T10:53:10.718Z] erl_bits.o : error LNK2019: unresolved external symbol UINT16_C referenced in function fp16_ieee_from_fp32_value

Thanks @jhogberg! Which platform is it coming from? It seems those macros are part of the C99 standard, so I am wondering why they are not available. Although uint16_t is part of the C99 standard too.

Follow up: do we support any 16bit platforms? My understanding is that this will matter only for those platforms, otherwise adding a U suffix would be enough.

A relatively old version of MSVC that doesn't fully support C99 yet. We don't support any 16-bit platforms so it should be safe to use a U prefix.

Updated! I also re-ran the added tests and they continue to pass.

Thank you, it builds fine now :)

This format is often used with computer graphics and AI, also in cases where data is large enough that the decreased bandwidth and storage compensates the loss of precision. The _Float16 data type is defined in the C11 extension ISO/IEC TS 18661-3:2015. Recent compilers and platforms provide support for it and recent CPUs provide specific instruction sets for them too. In some, _Float16 is storage only, which is enough for our use cases. If _Float16 is not available, functions for conversion from float to half-floats are used, ported from https://github.com/Maratyszcza/FP16.

jhogberg · 2021-02-11T13:47:46Z

erts/emulator/beam/erl_bits.c

@@ -1132,17 +1178,21 @@ erts_new_bs_put_float(Process *c_p, Eterm arg, Uint num_bits, int flags)
 #ifdef WORDS_BIGENDIAN


bs_construct_SUITE:fp16/1 fails on our big endian machine, can you look into that?

Can you please post the failure if handy? I will try to figure out how to run this patch on a big endian machine. :)

The first line of the test fails with {badmatch,false}, it doesn't say much more :/

I think the lines below are the culprit as they will always be zero when given an fp16, if I'm reading the code right you want to write the first 16 bits instead:

t[0] = a >> 24; t[1] = a >> 16;

Ah, good point, thanks! I believe I have addressed your changes. I kept it as a separate commit for now as it also has some debugging information for the tests. If it works, then I can squash and we can discuss if we should keep the debugging info or no.

Thanks, the fix looks a bit strange though: why set t[2] and t[3] for a 16-bit value? Shouldn't it be t[0] and t[1]?

Suggested change

#ifdef WORDS_BIGENDIAN

#ifdef WORDS_BIGENDIAN

if (num_bits == 16) {

t[0] = a >> 8;

t[1] = a;

} else if (num_bits >= 32) {

t[0] = a >> 24;

t[1] = a >> 16;

t[2] = a >> 8;

t[3] = a;

if (num_bits == 64) {

t[4] = b >> 24;

t[5] = b >> 16;

t[6] = b >> 8;

t[7] = b;

}

}

@jhogberg of course your fix makes much more sense (or rather my fix made no sense whatsoever :D). I have applied it to both bigendian branches.

Thanks, the tests pass and I'll merge this as soon as I can. There are some bureaucratic hurdles with importing the dependency that I hope will clear soon.

jhogberg · 2021-02-19T10:23:22Z

Merged, thanks for the PR!

josevalim force-pushed the jv-float-16-support branch from 931b023 to ee2c4c0 Compare November 28, 2020 08:32

rickard-green added the team:VM Assigned to OTP team VM label Nov 30, 2020

rickard-green assigned jhogberg Nov 30, 2020

jhogberg added the enhancement label Dec 1, 2020

jhogberg reviewed Dec 1, 2020

View reviewed changes

erts/emulator/beam/erl_bits.c Outdated Show resolved Hide resolved

josevalim mentioned this pull request Dec 28, 2020

Support f16 elixir-nx/nx#117

Closed

josevalim force-pushed the jv-float-16-support branch 2 times, most recently from 2be28a3 to 972bb4e Compare January 13, 2021 11:41

jhogberg added testing currently being tested, tag is used by OTP internal CI and removed testing currently being tested, tag is used by OTP internal CI labels Jan 15, 2021

josevalim force-pushed the jv-float-16-support branch from 972bb4e to 61f7ea5 Compare January 15, 2021 17:50

jhogberg added testing currently being tested, tag is used by OTP internal CI and removed testing currently being tested, tag is used by OTP internal CI labels Feb 1, 2021

jhogberg reviewed Feb 2, 2021

View reviewed changes

erts/emulator/beam/erl_bits.c Outdated Show resolved Hide resolved

josevalim force-pushed the jv-float-16-support branch 2 times, most recently from 95ec166 to 64fe9e7 Compare February 2, 2021 09:33

garazdawi reviewed Feb 2, 2021

View reviewed changes

lib/stdlib/src/erl_lint.erl Outdated Show resolved Hide resolved

josevalim force-pushed the jv-float-16-support branch from 64fe9e7 to c455b32 Compare February 2, 2021 09:41

jhogberg added the testing currently being tested, tag is used by OTP internal CI label Feb 2, 2021

josevalim force-pushed the jv-float-16-support branch from c455b32 to ebf7736 Compare February 2, 2021 13:57

jhogberg reviewed Feb 3, 2021

View reviewed changes

erts/emulator/beam/erl_bits_f16.h Outdated Show resolved Hide resolved

josevalim force-pushed the jv-float-16-support branch from ebf7736 to b79526d Compare February 3, 2021 08:27

jhogberg reviewed Feb 3, 2021

View reviewed changes

josevalim force-pushed the jv-float-16-support branch 2 times, most recently from 386d8c9 to 62d1e08 Compare February 3, 2021 10:13

jhogberg reviewed Feb 3, 2021

View reviewed changes

josevalim force-pushed the jv-float-16-support branch from 62d1e08 to db21c04 Compare February 3, 2021 11:52

jhogberg added the stalled waiting for input by the Erlang/OTP team label Feb 11, 2021

jhogberg reviewed Feb 11, 2021

View reviewed changes

Attempt to address Big Endian system

8b9d974

josevalim force-pushed the jv-float-16-support branch from 4dcf5fb to 8b9d974 Compare February 11, 2021 14:39

jhogberg merged commit 7fc0522 into erlang:master Feb 19, 2021

wojtekmach mentioned this pull request Feb 19, 2021

Support 16bit floats in bitstrings elixir-lang/elixir#10740

Merged

mikpe mentioned this pull request Mar 14, 2024

Inconsistent behavior of Bit Syntax Expression on different platforms #8263

Closed

Provide feedback

Saved searches

Use saved searches to filter your results more quickly

Support IEEE 754-2008 16-bits floats in bitstrings #2890

Support IEEE 754-2008 16-bits floats in bitstrings #2890

josevalim commented Nov 28, 2020 •

edited

Loading

jhogberg left a comment

jhogberg commented Jan 15, 2021

josevalim commented Jan 15, 2021

josevalim commented Jan 15, 2021

jhogberg commented Feb 2, 2021

josevalim commented Feb 2, 2021

jhogberg Feb 3, 2021

josevalim Feb 3, 2021

jhogberg Feb 3, 2021

josevalim Feb 3, 2021

josevalim Feb 3, 2021

jhogberg Feb 3, 2021

josevalim Feb 3, 2021

jhogberg Feb 3, 2021

jhogberg Feb 11, 2021

josevalim Feb 11, 2021

jhogberg Feb 11, 2021

josevalim Feb 11, 2021

jhogberg Feb 11, 2021 •

edited

Loading

josevalim Feb 11, 2021 •

edited

Loading

jhogberg Feb 18, 2021

jhogberg commented Feb 19, 2021

		@@ -1132,17 +1178,21 @@ erts_new_bs_put_float(Process *c_p, Eterm arg, Uint num_bits, int flags)
		#ifdef WORDS_BIGENDIAN

-#ifdef WORDS_BIGENDIAN
+#ifdef WORDS_BIGENDIAN
+	    if (num_bits == 16) {
+	        t[0] = a >> 8;
+	        t[1] = a;
+	    } else if (num_bits >= 32) {
+		t[0] = a >> 24;
+		t[1] = a >> 16;
+	        t[2] = a >> 8;
+	        t[3] = a;
+	        if (num_bits == 64) {
+		    t[4] = b >> 24;
+		    t[5] = b >> 16;
+		    t[6] = b >> 8;
+		    t[7] = b;
+	        }
+	    }

Support IEEE 754-2008 16-bits floats in bitstrings #2890

Support IEEE 754-2008 16-bits floats in bitstrings #2890

Conversation

josevalim commented Nov 28, 2020 • edited Loading

jhogberg left a comment

Choose a reason for hiding this comment

jhogberg commented Jan 15, 2021

josevalim commented Jan 15, 2021

josevalim commented Jan 15, 2021

jhogberg commented Feb 2, 2021

josevalim commented Feb 2, 2021

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

jhogberg Feb 11, 2021 • edited Loading

Choose a reason for hiding this comment

josevalim Feb 11, 2021 • edited Loading

Choose a reason for hiding this comment

Choose a reason for hiding this comment

jhogberg commented Feb 19, 2021

josevalim commented Nov 28, 2020 •

edited

Loading

jhogberg Feb 11, 2021 •

edited

Loading

josevalim Feb 11, 2021 •

edited

Loading