Deprecate some old static solvers for new ASICs

[atamazov]

• [x] (1) Some old static solvers should be allowed only for the following targets:

  • gfx900/906 (Vega10/20, MI50/60, Radeon VII etc)
  • gfx908/90a (MI100, MI200)
  • gfx103X (Navi 2X)
  • ℹ️ Note that gfx110X (Navi 3X) is NOT included
  • Env.var MIOPEN_DEBUG_ENABLE_DEPRECATED_SOLVERS can be used to override, i.e. enable old static solvers for all targets.
  • ✔️ Implemented in #2081

• (2) The list of solvers:

  • ConvAsm1x1U
  • ConvAsm1x1UV2
  • ConvAsm3x3U
  • ConvAsm5x10u2v2b1
  • ConvAsm5x10u2v2f1
  • ConvAsm7x7c3h224w224k64u2v2p3q3f1
  • ConvAsmBwdWrW1x1
  • ConvAsmBwdWrW3x3
  • ConvCkIgemmFwdV6r1DlopsNchw
  • ConvHipImplicitGemmBwdDataV1R1
  • ConvHipImplicitGemmBwdDataV1R1Xdlops
  • ConvHipImplicitGemmBwdDataV4R1
  • ConvHipImplicitGemmBwdDataV4R1Xdlops
  • ConvHipImplicitGemmForwardV4R4Xdlops
  • ConvHipImplicitGemmForwardV4R4Xdlops_Padded_Gemm
  • ConvHipImplicitGemmForwardV4R5Xdlops
  • ConvHipImplicitGemmV4R1Fwd
  • ConvHipImplicitGemmV4R4Fwd
  • ConvHipImplicitGemmV4R4WrW
  • ConvHipImplicitGemmWrwV4R4Xdlops
  • ConvHipImplicitGemmWrwV4R4Xdlops_Padded_Gemm
  • ConvOclBwdWrW1x1
  • ConvOclBwdWrW2<N>
  • ConvOclBwdWrW2NonTunable
  • ConvOclBwdWrW53
  • ConvOclDirectFwd
  • ConvOclDirectFwd11x11
  • ConvOclDirectFwd1x1
  • ConvOclDirectFwdGen

• [ ] (3) The solvers listed at (2) are to be removed when support for the targets listed at (1) is discontinued.

Some clarifications: https://github.com/ROCmSoftwarePlatform/MIOpen/pull/2072#discussion_r1159102468

---

[Attribution] @junliume @johnny-keker https://github.com/ROCmSoftwarePlatform/MIOpen/labels/specification https://github.com/ROCmSoftwarePlatform/MIOpen/labels/urgency_high

• Proposed assignees: @JehandadKhan @atamazov

[atamazov]

[attribution] @JehandadKhan I think that this is https://github.com/ROCm/MIOpen/labels/urgency_low https://github.com/ROCm/MIOpen/labels/value_low. All that's left is to remove the solvers (which may only happen in the future)

[atamazov]

@junliume @JehandadKhan @Kirpich30000 As shown at https://github.com/ROCm/MIOpen/issues/2492#issuecomment-1911121477, ConvOclBwdWrW2<*> deliver good performance (especially with grouped configs). Shall we consider excluding these from the list of deprecated solvers?

FYI quote from https://github.com/ROCm/MIOpen/issues/2492#issuecomment-1919952275:

Maybe it is worth enabling ConvOclBwdWrW2<*> for the new GPUs to avoid inadequate performance of WrW convolutions, especially when n_groups > 1 (we can even enable them only for group convolutions).