schedule_injective of arm_cpu should consider dtype itemsize

[sunwayforever]

test code:

import tvm
from tvm import relay

x = relay.var("x", shape=(1, 1024), dtype="int8")
y = relay.add(x, x)
func = relay.Function([x], y)
mod = tvm.IRModule.from_expr(func)

with tvm.transform.PassContext(opt_level=3):
    graph, lib, params = relay.build(
        mod,
        target="llvm --device=arm_cpu -mtriple=armv7a-linux-gnueabihf -mattr=+neon",
        params=None,
    )

lib.save("/tmp/a.o")

arm-linux-gnueabihf-objdump -d /tmp/a.o will produce output like:

000002a8 <tvmgen_default_fused_add_compute_>:
2a8: e3a02000 mov r2, #0
2ac: e0813002 add r3, r1, r2
2b0: f4e3083f vld1.32 {d16[0]}, [r3 :32]
2b4: e0803002 add r3, r0, r2
2b8: e2822004 add r2, r2, #4
2bc: e3520b01 cmp r2, #1024 ; 0x400
2c0: f3c80a30 vmovl.u8 q8, d16
2c4: f2d10530 vshl.s16 d16, d16, #1
2c8: f3f2012 vuzp.8 d16, d17
2cc: f4c3083f vst1.32 {d16[0]}, [r3 :32]
2d0: 1afffff5 bne 2ac <tvmgen_default_fused_add_compute_+0x4>
2d4: e12fff1e bx lr

when the PR is applied, it will change to:

000002a8 <tvmgen_default_fused_add_compute_>:
2a8: e3a02000 mov r2, #0
2ac: e0813002 add r3, r1, r2
2b0: f4630aef vld1.64 {d16-d17}, [r3 :128]
2b4: e0803002 add r3, r0, r2
2b8: e2822010 add r2, r2, #16
2bc: f2c90570 vshl.s8 q8, q8, #1
2c0: e3520b01 cmp r2, #1024 ; 0x400
2c4: f4430aef vst1.64 {d16-d17}, [r3 :128]
2c8: 1afffff7 bne 2ac <tvmgen_default_fused_add_compute_+0x4>
2cc: e12fff1e bx lr

because neon support up to 16x8-bit operations per instruction

[u99127]

Can a testcase be added for this ?

[sunwayforever]

Can a testcase be added for this ?

I'd like to, but don't know how to write the testcase...since the PR targets a sub-optimal issue, and I didn't find testcases related with schedule_injective and arm_cpu for reference.

[AndrewZhaoLuo]

Can you fill out the PR description with what you did and why.

Also a link or comment in the code explaining what is done would be helpful for reading.

[AndrewZhaoLuo]

LGTM, thanks for the documentation.

You will need to rebase on main in order to fix a flaky test.