0f1b5f115a
ptxas is a proprietary compiler from Nvidia that can compile PTX to machine code (SASS). It has a lot of diagnostics to catch errors in PTX, which can be used to verify PTX output from llc. Set -DPXTAS_EXECUTABLE=/path/to/ptxas CMake option to enable it. If this option is not set, then ptxas is substituted to true which effectively disables all ptxas RUN lines. LLVM_PTXAS_EXECUTABLE environment variable takes precedence over the CMake option, and allows to override ptxas executable that is used for LIT without complete re-configuration. Differential Revision: https://reviews.llvm.org/D121727
36 lines
1.3 KiB
LLVM
36 lines
1.3 KiB
LLVM
; RUN: llc < %s -march=nvptx64 -mcpu=sm_20 -fp-contract=fast | FileCheck %s --check-prefix=FAST
|
|
; RUN: llc < %s -march=nvptx64 -mcpu=sm_30 | FileCheck %s --check-prefix=DEFAULT
|
|
; RUN: %if ptxas %{ llc < %s -march=nvptx64 -mcpu=sm_20 -fp-contract=fast | %ptxas-verify %}
|
|
; RUN: %if ptxas %{ llc < %s -march=nvptx64 -mcpu=sm_30 | %ptxas-verify %if !ptxas-11.0 %{-arch=sm_30%} %}
|
|
|
|
target triple = "nvptx64-unknown-cuda"
|
|
|
|
;; Make sure we are generating proper instruction sequences for fused ops
|
|
;; If fusion is allowed, we try to form fma.rn at the PTX level, and emit
|
|
;; add.f32 otherwise. Without an explicit rounding mode on add.f32, ptxas
|
|
;; is free to fuse with a multiply if it is able. If fusion is not allowed,
|
|
;; we do not form fma.rn at the PTX level and explicitly generate add.rn
|
|
;; for all adds to prevent ptxas from fusion the ops.
|
|
|
|
;; FAST-LABEL: @t0
|
|
;; DEFAULT-LABEL: @t0
|
|
define float @t0(float %a, float %b, float %c) {
|
|
;; FAST: fma.rn.f32
|
|
;; DEFAULT: mul.rn.f32
|
|
;; DEFAULT: add.rn.f32
|
|
%v0 = fmul float %a, %b
|
|
%v1 = fadd float %v0, %c
|
|
ret float %v1
|
|
}
|
|
|
|
;; FAST-LABEL: @t1
|
|
;; DEFAULT-LABEL: @t1
|
|
define float @t1(float %a, float %b) {
|
|
;; We cannot form an fma here, but make sure we explicitly emit add.rn.f32
|
|
;; to prevent ptxas from fusing this with anything else.
|
|
;; FAST: add.f32
|
|
;; DEFAULT: add.rn.f32
|
|
%v1 = fadd float %a, %b
|
|
ret float %v1
|
|
}
|