libclc: Add a __builtin to let SPIRV targets select between SW and HW FMA

[daniels]

Summary

The SPIRV targets are supposed to be generic, but some HW still don't
support the Fma instruction. Add the __builtin_has_hw_fma32() builtin
function to let the implementation express whether it has hardware FMA
support or not.

Depends on: D85911

[jenatali]

libclc/generic/lib/math/math.h

44

From discussion on the corresponding Mesa merge request that would handle this builtin function call (https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/6035#note_606348), I think this would be better as defined, non-inlined function. That way, we get the best of both worlds:

• If Mesa understands the builtin, it can choose to override it with its own implementation, rather than using the one provided by libclc, allowing individual drivers to control the behavior.
• If mesa doesn't understand the builtin, then it's just a regular function call, and after non-LLVM optimizations, it will eventually end up inlined, and the dead code for the non-taken path can be removed.

We might need to have a dedicated .cl file for the definition though, rather than defining it inline, just to avoid violating the one-definition-rule here. And we should probably do the same thing for subnormal controls, rather than just setting them to false.

[daniels]

Added noinline to use software by default, letting implementation override

[jenatali]

Yep, I think this is exactly what we need.

[jvesely]

libclc/generic/lib/math/math.h

45

This shouldn't be prefixed __bultin that space is reserved for the compiler. Naming it something like __clc_runtime_has_fma32 would be more fitting.

[daniels]

libclc/generic/lib/math/math.h

45

Fine by me. @jenatali?

[jenatali]

libclc/generic/lib/math/math.h

45

No complaints.

[daniels]

Change builtin to clc_runtime

[daniels]

Done, @jvesely could you please merge this one for us too?

libclc/generic/lib/math/math.h

 #define NZER 0x020
 #define PZER 0x040
 #define PSUB 0x080
 #define PNOR 0x100
 #define PINF 0x200
 
 #if (defined __AMDGCN__ || defined __R600__) && !defined __HAS_FMAF__
 #define HAVE_HW_FMA32() (0)
+#elif defined CLC_SPIRV || defined CLC_SPIRV64
+bool __attribute__((noinline)) __builtin_has_hw_fma32(void);
+#define HAVE_HW_FMA32() __builtin_has_hw_fma32()
 #else
 #define HAVE_HW_FMA32() (1)
 #endif
 
 #define HAVE_BITALIGN() (0)
 #define HAVE_FAST_FMA32() (0)
 
 #define MATH_DIVIDE(X, Y) ((X) / (Y))

libclc/spirv/lib/math/fma.cl

 #include <clc/clc.h>
 #include <math/clc_fma.h>
 
 #define __CLC_BODY <fma.inc>
 #define __FLOAT_ONLY
 #include <clc/math/gentype.inc>
+
+bool __builtin_has_hw_fma32()
+{
+    return false;
+}

libclc/spirv64/lib/math/fma.cl

 #include <clc/clc.h>
 #include <math/clc_fma.h>
 
 #define __CLC_BODY <fma.inc>
 #define __FLOAT_ONLY
 #include <clc/math/gentype.inc>
+
+bool __builtin_has_hw_fma32()
+{
+    return false;
+}