19a3d7b5db
When we work on the triton kernel with tensor descriptor created on the host side, there is a error message `amdgpu_user_sgpr_count smaller than than implied by enabled user SGPRs`. After some debugging, we find the `USER_SGPR_COUNT` is not updated with GFX125 and this patch updates it for USER_SGPR_COUNT from https://llvm.org/docs/AMDGPUUsage.html#amdgpu-amdhsa-compute-pgm-rsrc2-gfx6-gfx12-table. On GFX125, COMPUTE_PGM_RSRC2::USER_SGPR_COUNT is 6 bits wide. The MC helper S_00B84C_USER_SGPR only masks to 5 bits; when the true user SGPR count is 32 or more, the masked value wraps (e.g. 32 -> 0). `AMDGPUAsmPrinter` then emits a .amdhsa_user_sgpr_count with 0, that disagrees with the implied count from enabled user SGPRs (including kernarg preload), and finally assembling llc output with `llvm-mc` fails in `AMDGPUAsmParser` --------- Co-authored-by: Shilei Tian <i@tianshilei.me>
129 lines
4.1 KiB
C++
129 lines
4.1 KiB
C++
//===--- SIProgramInfo.h ----------------------------------------*- C++ -*-===//
|
|
//
|
|
// Part of the LLVM Project, under the Apache License v2.0 with LLVM Exceptions.
|
|
// See https://llvm.org/LICENSE.txt for license information.
|
|
// SPDX-License-Identifier: Apache-2.0 WITH LLVM-exception
|
|
//
|
|
//===----------------------------------------------------------------------===//
|
|
//
|
|
/// \file
|
|
/// Defines struct to track resource usage and hardware flags for kernels and
|
|
/// entry functions.
|
|
///
|
|
//
|
|
//===----------------------------------------------------------------------===//
|
|
|
|
#ifndef LLVM_LIB_TARGET_AMDGPU_SIPROGRAMINFO_H
|
|
#define LLVM_LIB_TARGET_AMDGPU_SIPROGRAMINFO_H
|
|
|
|
#include "llvm/IR/CallingConv.h"
|
|
#include "llvm/Support/Compiler.h"
|
|
#include <cstdint>
|
|
#include <optional>
|
|
|
|
namespace llvm {
|
|
|
|
class GCNSubtarget;
|
|
class MCContext;
|
|
class MCExpr;
|
|
class MachineFunction;
|
|
|
|
/// Track resource usage for kernels / entry functions.
|
|
struct LLVM_EXTERNAL_VISIBILITY SIProgramInfo {
|
|
std::optional<uint64_t> CodeSizeInBytes;
|
|
|
|
// Fields set in PGM_RSRC1 pm4 packet.
|
|
const MCExpr *VGPRBlocks = nullptr;
|
|
const MCExpr *SGPRBlocks = nullptr;
|
|
uint32_t Priority = 0;
|
|
uint32_t FloatMode = 0;
|
|
uint32_t Priv = 0;
|
|
uint32_t DX10Clamp = 0;
|
|
uint32_t DebugMode = 0;
|
|
uint32_t IEEEMode = 0;
|
|
uint32_t WgpMode = 0; // GFX10+
|
|
uint32_t MemOrdered = 0; // GFX10+
|
|
uint32_t FwdProgress = 0; // GFX10+
|
|
uint32_t RrWgMode = 0; // GFX12+
|
|
const MCExpr *ScratchSize = nullptr;
|
|
|
|
// State used to calculate fields set in PGM_RSRC2 pm4 packet.
|
|
uint32_t LDSBlocks = 0;
|
|
const MCExpr *ScratchBlocks = nullptr;
|
|
|
|
// Fields set in PGM_RSRC2 pm4 packet
|
|
const MCExpr *ScratchEnable = nullptr;
|
|
uint32_t UserSGPR = 0;
|
|
uint32_t TrapHandlerEnable = 0;
|
|
uint32_t TGIdXEnable = 0;
|
|
uint32_t TGIdYEnable = 0;
|
|
uint32_t TGIdZEnable = 0;
|
|
uint32_t TGSizeEnable = 0;
|
|
uint32_t TIdIGCompCount = 0;
|
|
uint32_t EXCPEnMSB = 0;
|
|
uint32_t LdsSize = 0;
|
|
uint32_t EXCPEnable = 0;
|
|
|
|
const MCExpr *ComputePGMRSrc3 = nullptr;
|
|
|
|
const MCExpr *NumVGPR = nullptr;
|
|
const MCExpr *NumArchVGPR = nullptr;
|
|
const MCExpr *NumAccVGPR = nullptr;
|
|
const MCExpr *AccumOffset = nullptr;
|
|
uint32_t TgSplit = 0;
|
|
const MCExpr *NumSGPR = nullptr;
|
|
unsigned SGPRSpill = 0;
|
|
unsigned VGPRSpill = 0;
|
|
uint32_t LDSSize = 0;
|
|
const MCExpr *FlatUsed = nullptr;
|
|
|
|
// Number of SGPRs that meets number of waves per execution unit request.
|
|
const MCExpr *NumSGPRsForWavesPerEU = nullptr;
|
|
|
|
// Number of VGPRs that meets number of waves per execution unit request.
|
|
const MCExpr *NumVGPRsForWavesPerEU = nullptr;
|
|
|
|
// Number of named barriers used by the kernel.
|
|
const MCExpr *NamedBarCnt = nullptr;
|
|
|
|
// Final occupancy.
|
|
const MCExpr *Occupancy = nullptr;
|
|
|
|
// Whether there is recursion, dynamic allocas, indirect calls or some other
|
|
// reason there may be statically unknown stack usage.
|
|
const MCExpr *DynamicCallStack = nullptr;
|
|
|
|
// Bonus information for debugging.
|
|
const MCExpr *VCCUsed = nullptr;
|
|
|
|
SIProgramInfo() = default;
|
|
|
|
// The constructor sets the values for each member as shown in the struct.
|
|
// However, setting the MCExpr members to their zero value equivalent
|
|
// happens in reset together with (duplicated) value re-set for the
|
|
// non-MCExpr members.
|
|
void reset(const MachineFunction &MF);
|
|
|
|
// Get function code size and cache the value.
|
|
// If \p IsLowerBound is set it returns a minimal code size which is safe
|
|
// to address.
|
|
uint64_t getFunctionCodeSize(const MachineFunction &MF,
|
|
bool IsLowerBound = false);
|
|
|
|
/// Compute the value of the ComputePGMRsrc1 register.
|
|
const MCExpr *getComputePGMRSrc1(const GCNSubtarget &ST,
|
|
MCContext &Ctx) const;
|
|
const MCExpr *getPGMRSrc1(CallingConv::ID CC, const GCNSubtarget &ST,
|
|
MCContext &Ctx) const;
|
|
|
|
/// Compute the value of the ComputePGMRsrc2 register.
|
|
const MCExpr *getComputePGMRSrc2(const GCNSubtarget &ST,
|
|
MCContext &Ctx) const;
|
|
const MCExpr *getPGMRSrc2(CallingConv::ID CC, const GCNSubtarget &ST,
|
|
MCContext &Ctx) const;
|
|
};
|
|
|
|
} // namespace llvm
|
|
|
|
#endif // LLVM_LIB_TARGET_AMDGPU_SIPROGRAMINFO_H
|