[Sbcl-devel] SB-SIMD

[Heisig, M. <mar...@fa...>]

Hello everyone,

I started working on a high-level SIMD interface for SBCL at SBCL20 (https://github.com/marcoheisig/sb-simd).  So far, I got Lisp wrappers for AVX2 loads, stores, casts and arithmetic operations working.  Before I continue, I'd like to discuss some things:

1. Since 50% of this library is tied to SB-VM or SB-C, I'd like to move the VOP definitions and deftransforms to sbcl/src/compiler, and the remaining code into a contrib.  Is that appropriate?

2. I'd like to introduce more subtypes of SIMD-PACK and SIMD-PACK-256.  Currently there is only support for double-floats, single-floats and unsigned-byte-64s.  I'd like to add the remaining integer types (unsigned-byte-{8,16,32} and signed-byte-{8,16,32,64}.  Would you be willing to merge such a pull request into SBCL?  Or is there a particular reason SBCL only distinguishes single, double and int?

3. The current naming convention for primitive SIMD types is confusing (e.g., SIMD-PACK-DOUBLE and SIMD-PACK-256-UB64).  And also the mapping from type specifiers to primitive types is not obvious (e.g., (SB-EXT:SIMD-PACK INTEGER) has the primitive type SIMD-PACK-INT, which actually contains 64bit unsigned integers).  I'd like to introduce a more uniform naming convention.  I have two proposals:

LONG-NAMES

Every primitive type is named 'SIMD-PACK-' followed by the number of bits, followed by the SIMD pack element type.  Examples: SIMD-PACK-128-DOUBLE-FLOAT,  SIMD-PACK-256-UNSIGNED-BYTE-32, SIMD-PACK-256-SIGNED-BYTE-8

CONCISE-NAMES

SIMD pack element types are abbreviated as U{8,16,32,64}, S{8,16,32,64} and F{32,64}, and the name of each primitive SIMD pack type is then simply the element type, followed by a dot, followed by the number of elements.  Examples: F64.4, U32.8, S8.32

I have no strong opinions on which naming scheme we should adopt, but I think there should be a consistent naming scheme.

4. I encountered a bug when loading AVX2 floating point zero immediates (vxorps takes three arguments):

Bug:

(compile nil '(lambda ()
                        (sb-ext:%simd-pack-256-doubles
                         (sb-ext:%make-simd-pack-256-double 0d0 0d0 0d0 0d0))))

Fix:

diff --git a/src/compiler/x86-64/simd-pack-256.lisp b/src/compiler/x86-64/simd-pack-256.lisp
index 2c688cd11..fd0952cfe 100644
--- a/src/compiler/x86-64/simd-pack-256.lisp
+++ b/src/compiler/x86-64/simd-pack-256.lisp
@@ -48,7 +48,7 @@
          (p2 (%simd-pack-256-2 x))
          (p3 (%simd-pack-256-3 x)))
     (cond ((= p0 p1 p2 p3 0)
-           (inst vxorps y y))
+           (inst vxorps y y y))
           ((= p0 p1 p2 p3 (ldb (byte 64 0) -1))
            (inst vpcmpeqd y y y))
           (t

Best regards,
Marco Heisig