BLAS: Convert alpha & beta to more appropriate types.

[maleadt]

Default boolean values for alpha and beta cause messy conversions when applied to e.g. Float16:

┌ @ /home/tim/Julia/pkg/GemmKernels/src/blas.jl:71 within `#9`
│ %6097  = Core.getfield(%6096, :alpha)::Bool
│┌ @ bool.jl:182 within `*` @ bool.jl:180
││┌ @ number.jl:221 within `copysign`
│││┌ @ floatfuncs.jl:17 within `signbit`
││││ %6098  = Base.bitcast(Base.Int16, %6063)::Int16
││││ @ floatfuncs.jl:17 within `signbit` @ int.jl:139
││││┌ @ promotion.jl:450 within `<`
│││││┌ @ promotion.jl:381 within `promote`
││││││┌ @ promotion.jl:358 within `_promote`
│││││││┌ @ number.jl:7 within `convert`
││││││││┌ @ boot.jl:784 within `Int64`
│││││││││┌ @ boot.jl:702 within `toInt64`
││││││││││ %6099  = Core.sext_int(Core.Int64, %6098)::Int64
│││││└└└└└
│││││ @ promotion.jl:450 within `<` @ int.jl:83
│││││ %6100  = Base.slt_int(%6099, 0)::Bool
│││└└
│││┌ @ operators.jl:269 within `!=`
││││┌ @ promotion.jl:499 within `==`
│││││ %6101  = (false === %6100)::Bool
││││└
││││┌ @ bool.jl:35 within `!`
│││││ %6102  = Base.not_int(%6101)::Bool
│││└└
│││┌ @ essentials.jl:575 within `ifelse`
││││ %6103  = Core.ifelse(%6102, Float16(-0.0), Float16(0.0))::Float16
││└└
││┌ @ essentials.jl:575 within `ifelse`
│││ %6104  = Core.ifelse(%6097, %6063, %6103)::Float16
│└└

vs simply

┌ @ /home/tim/Julia/pkg/GemmKernels/src/blas.jl:71 within `#13`
│ %6097  = Core.getfield(%6096, :alpha)::Float16
│┌ @ float.jl:410 within `*`
││ %6098  = Base.mul_float(%6063, %6097)::Float16
│└

LLVM IR:

julia> @code_llvm Float16(1) * true
;  @ bool.jl:182 within `*`
define half @"julia_*_146"(half %0, i8 zeroext %1) #0 {
top:
;  @ bool.jl:182 within `*` @ bool.jl:180
; ┌ @ number.jl:221 within `copysign`
; │┌ @ essentials.jl:575 within `ifelse`
    %2 = call half @llvm.copysign.f16(half 0xH0000, half %0)
; └└
; ┌ @ essentials.jl:575 within `ifelse`
   %3 = and i8 %1, 1
   %.not = icmp eq i8 %3, 0
   %4 = select i1 %.not, half %2, half %0
; └
;  @ bool.jl:182 within `*`
  ret half %4
}

Avoiding this reduces register pressure by 15% or so.

[maleadt]

Benchmark results for commit 4ffa934 (comparing to 1222d80):

ID	before	after	change
["BLAS", "Float16'*Float16'=Float16 (4096×4096×4096, alpha)"]	7.389 ms ± 795.688 μs	6.021 ms ± 554.176 μs	20.3% ✅
["BLAS", "Float16'*Float16'=Float32 (4096×4096×4096, alpha)"]	6.438 ms ± 690.212 μs	6.128 ms ± 690.534 μs	6.8% ✅
["BLAS", "Float16'*Float16=Float16 (4096×4096×4096, alpha)"]	7.706 ms ± 22.187 μs	5.761 ms ± 35.014 μs	25.5% ✅
["BLAS", "Float16*Float16'=Float16 (4096×4096×4096, alpha)"]	7.104 ms ± 569.471 μs	5.807 ms ± 20.311 μs	-6.8% ❌
["BLAS", "Float16*Float16'=Float32 (4096×4096×4096, alpha)"]	6.689 ms ± 1.446 ms	6.356 ms ± 1.377 ms	6.3% ✅
["BLAS", "Float16*Float16=Float16 (4096×4096×4096, alpha)"]	5.419 ms ± 13.667 μs	4.184 ms ± 12.511 μs	22.7% ✅
["BLAS", "Float16*Float16=Float16 (4096×4096×4096, alpha, beta)"]	5.569 ms ± 24.554 μs	4.403 ms ± 17.014 μs	21.2% ✅
["BLAS", "Float16*Float16=Float16 (4096×4096×4096, beta)"]	3.865 ms ± 18.291 μs	3.311 ms ± 13.259 μs	13.8% ✅
["BLAS", "Float16*Float16=Float32 (4096×4096×4096, beta)"]	5.409 ms ± 8.565 μs	5.077 ms ± 15.586 μs	6.3% ✅

[codecov]

Codecov Report

Patch coverage: 100.00% and project coverage change: +0.26 🎉

Comparison is base (1222d80) 29.97% compared to head (4ffa934) 30.23%.

Additional details and impacted files

@@            Coverage Diff             @@
##           master     #129      +/-   ##
==========================================
+ Coverage   29.97%   30.23%   +0.26%     
==========================================
  Files          11       11              
  Lines         794      797       +3     
==========================================
+ Hits          238      241       +3     
  Misses        556      556              

Impacted Files	Coverage Δ
src/blas.jl	84.84% <100.00%> (+1.51%)	⬆️

☔ View full report in Codecov by Sentry.
📢 Do you have feedback about the report comment? Let us know in this issue.