Stream: git-wasmtime

Topic: wasmtime / issue #4462 Cranelift: Missing SIMD `fma` lowe...

Wasmtime GitHub notifications bot (Jul 17 2022 at 13:29):

afonso360 opened issue #4462:

:wave: Hey,

It looks like we don't have any lowering for the fma instruction when used with SIMD types.

.clif Test Case

function %fma_f32x4(f32x4, f32x4, f32x4) -> f32x4 {
block0(v0: f32x4, v1: f32x4, v2: f32x4):
    v3 = fma v0, v1, v2
    return v3
}

Steps to Reproduce

cargo run -- compile --target x86_64 ./the-above.clif

Expected Results

A successful compilation.

Actual Results

We don't have this implemented

thread 'main' panicked at 'internal error: entered unreachable code: implemented in ISLE: inst = `v3 = fma.f32x4 v0, v1, v2`, type = `Some(types::F32X4)`', cranelift\codegen\src\isa\x64\lower.rs:808:9
n

Versions and Environment

Cranelift version or commit: main
Operating system: Windows
Architecture: x86_64

Extra Info

In #4460 I tried to implement this using the vfmadd231ps instruction, but I had issues implementing emit for it and I'd like some help.

It looks like that instruction is only available in VEX encoding (or EVEX for avx512 but I don't have a machine with that). Is our EvexInstruction encoder suitable for emitting VEX instructions, or are they completely different? Do we have a way to emit VEX instructions?

I'm not too familiar with x86 but I'd like to pick this up with some help.

cc: @abrown

Wasmtime GitHub notifications bot (Jul 17 2022 at 13:29):

afonso360 labeled issue #4462:

:wave: Hey,

It looks like we don't have any lowering for the fma instruction when used with SIMD types.

.clif Test Case

function %fma_f32x4(f32x4, f32x4, f32x4) -> f32x4 {
block0(v0: f32x4, v1: f32x4, v2: f32x4):
    v3 = fma v0, v1, v2
    return v3
}

Steps to Reproduce

cargo run -- compile --target x86_64 ./the-above.clif

Expected Results

A successful compilation.

Actual Results

We don't have this implemented

thread 'main' panicked at 'internal error: entered unreachable code: implemented in ISLE: inst = `v3 = fma.f32x4 v0, v1, v2`, type = `Some(types::F32X4)`', cranelift\codegen\src\isa\x64\lower.rs:808:9
n

Versions and Environment

Cranelift version or commit: main
Operating system: Windows
Architecture: x86_64

Extra Info

In #4460 I tried to implement this using the vfmadd231ps instruction, but I had issues implementing emit for it and I'd like some help.

It looks like that instruction is only available in VEX encoding (or EVEX for avx512 but I don't have a machine with that). Is our EvexInstruction encoder suitable for emitting VEX instructions, or are they completely different? Do we have a way to emit VEX instructions?

I'm not too familiar with x86 but I'd like to pick this up with some help.

cc: @abrown

Wasmtime GitHub notifications bot (Jul 17 2022 at 13:29):

afonso360 labeled issue #4462:

:wave: Hey,

It looks like we don't have any lowering for the fma instruction when used with SIMD types.

.clif Test Case

function %fma_f32x4(f32x4, f32x4, f32x4) -> f32x4 {
block0(v0: f32x4, v1: f32x4, v2: f32x4):
    v3 = fma v0, v1, v2
    return v3
}

Steps to Reproduce

cargo run -- compile --target x86_64 ./the-above.clif

Expected Results

A successful compilation.

Actual Results

We don't have this implemented

thread 'main' panicked at 'internal error: entered unreachable code: implemented in ISLE: inst = `v3 = fma.f32x4 v0, v1, v2`, type = `Some(types::F32X4)`', cranelift\codegen\src\isa\x64\lower.rs:808:9
n

Versions and Environment

Cranelift version or commit: main
Operating system: Windows
Architecture: x86_64

Extra Info

In #4460 I tried to implement this using the vfmadd231ps instruction, but I had issues implementing emit for it and I'd like some help.

It looks like that instruction is only available in VEX encoding (or EVEX for avx512 but I don't have a machine with that). Is our EvexInstruction encoder suitable for emitting VEX instructions, or are they completely different? Do we have a way to emit VEX instructions?

I'm not too familiar with x86 but I'd like to pick this up with some help.

cc: @abrown

Wasmtime GitHub notifications bot (Jul 17 2022 at 13:40):

afonso360 edited issue #4462:

:wave: Hey,

It looks like we don't have any lowering for the fma instruction when used with SIMD types.

.clif Test Case

function %fma_f32x4(f32x4, f32x4, f32x4) -> f32x4 {
block0(v0: f32x4, v1: f32x4, v2: f32x4):
    v3 = fma v0, v1, v2
    return v3
}

Steps to Reproduce

cargo run -- compile --target x86_64 ./the-above.clif

Expected Results

A successful compilation.

Actual Results

We don't have this implemented

thread 'main' panicked at 'internal error: entered unreachable code: implemented in ISLE: inst = `v3 = fma.f32x4 v0, v1, v2`, type = `Some(types::F32X4)`', cranelift\codegen\src\isa\x64\lower.rs:808:9
n

Versions and Environment

Cranelift version or commit: main
Operating system: Windows
Architecture: x86_64

Extra Info

In #4460 I tried to implement this using the vfmadd231ps instruction, but I had issues encoding the instruction and I'd like some help.

It looks like that instruction is only available in VEX encoding (or EVEX for avx512 but I don't have a machine with that). Is our EvexInstruction encoder suitable for emitting VEX instructions, or are they completely different? Do we have a way to emit VEX instructions?

I'm not too familiar with x86 but I'd like to pick this up with some help.

cc: @abrown

Wasmtime GitHub notifications bot (Jul 18 2022 at 17:08):

abrown commented on issue #4462:

Hm, so here are some thoughts:

• ideally we would have implementations of this for every possible machine, e.g., your libcall implementation in #4460, an AVX implementation for machines that have that, an AVX512 implementation for machines that have that, etc. That is a lot of work, however, so I would propose we only do two of those here: a) the libcall implementation for very old machines and b) either an AVX or AVX512 implementation for newer machines
• To decide "b," it helps to understand that (apologies if you know all this) AVX instructions use a VEX encoding, AVX512 instructions use an EVEX encoding, and that currently in Cranelift we have only really implemented EVEX. Adding a VEX encoding implementation would be nice, but it's a bit of work.
• That particular instruction is available for 128-bit vectors (XMMs) in both the VEX encoding under the FMA flag and in the EVEX encoding under the AVX512VL and AVX512F flags. Those AVX512 flags are usually only available on server class CPUs but the FMA flag is likely available on most recent x86 machines. So, there seems to be a decent justification for implementing VEX encodings: if we add it, we can lower fma to a single instruction on more machines than the AVX512. (On the flip side, you could implement vfmadd231ps with the EVEX encoding but then not be able to run it locally.)
• To implement VEX encodings, we would need to fill in codegen/src/isa/x64/encoding/vex.rs in much the same fashion as I did for evex.rs. In this case, however, we could follow the AVX instruction format guide in section 2.3 of Intel's Software Developer's Manual, Volume 2. With that in place, we would probably need think through how to emit AVX instructions; e.g., something like Avx512Opcode but perhaps we want to be able to decide on the VEX/EVEX encoding at a later time (?). Finally, we would need a has_fma flag in codegen/meta/src/isa/x86.rs and to plumb that through in a few places.

Hopefully that information helps. Ping me on Zulip if you want to have a more "live" discussion.

Wasmtime GitHub notifications bot (Jul 25 2022 at 22:01):

cfallin closed issue #4462:

:wave: Hey,

It looks like we don't have any lowering for the fma instruction when used with SIMD types.

.clif Test Case

function %fma_f32x4(f32x4, f32x4, f32x4) -> f32x4 {
block0(v0: f32x4, v1: f32x4, v2: f32x4):
    v3 = fma v0, v1, v2
    return v3
}

Steps to Reproduce

cargo run -- compile --target x86_64 ./the-above.clif

Expected Results

A successful compilation.

Actual Results

We don't have this implemented

thread 'main' panicked at 'internal error: entered unreachable code: implemented in ISLE: inst = `v3 = fma.f32x4 v0, v1, v2`, type = `Some(types::F32X4)`', cranelift\codegen\src\isa\x64\lower.rs:808:9
n

Versions and Environment

Cranelift version or commit: main
Operating system: Windows
Architecture: x86_64

Extra Info

In #4460 I tried to implement this using the vfmadd231ps instruction, but I had issues encoding the instruction and I'd like some help.

It looks like that instruction is only available in VEX encoding (or EVEX for avx512 but I don't have a machine with that). Is our EvexInstruction encoder suitable for emitting VEX instructions, or are they completely different? Do we have a way to emit VEX instructions?

I'm not too familiar with x86 but I'd like to pick this up with some help.

cc: @abrown

Last updated: Dec 23 2024 at 12:05 UTC