#include "SIDefines.h"
#include "SIMachineFunctionInfo.h"
#include "SIRegisterInfo.h"
+#include "llvm/CodeGen/MachineFrameInfo.h"
#include "llvm/MC/MCContext.h"
#include "llvm/MC/MCSectionELF.h"
#include "llvm/MC/MCStreamer.h"
// precision, and leaves single precision to flush all and does not report
// CL_FP_DENORM for CL_DEVICE_SINGLE_FP_CONFIG. Mesa's OpenCL currently reports
// CL_FP_DENORM for both.
-static uint32_t getFPMode(const MachineFunction &) {
+//
+// FIXME: It seems some instructions do not support single precision denormals
+// regardless of the mode (exp_*_f32, rcp_*_f32, rsq_*_f32, rsq_*f32, sqrt_f32,
+// and sin_f32, cos_f32 on most parts).
+
+// We want to use these instructions, and using fp32 denormals also causes
+// instructions to run at the double precision rate for the device so it's
+// probably best to just report no single precision denormals.
+static uint32_t getFPMode(const MachineFunction &F) {
+ const AMDGPUSubtarget& ST = F.getTarget().getSubtarget<AMDGPUSubtarget>();
+ // TODO: Is there any real use for the flush in only / flush out only modes?
+
+ uint32_t FP32Denormals =
+ ST.hasFP32Denormals() ? FP_DENORM_FLUSH_NONE : FP_DENORM_FLUSH_IN_FLUSH_OUT;
+
+ uint32_t FP64Denormals =
+ ST.hasFP64Denormals() ? FP_DENORM_FLUSH_NONE : FP_DENORM_FLUSH_IN_FLUSH_OUT;
+
return FP_ROUND_MODE_SP(FP_ROUND_ROUND_TO_NEAREST) |
FP_ROUND_MODE_DP(FP_ROUND_ROUND_TO_NEAREST) |
- FP_DENORM_MODE_SP(FP_DENORM_FLUSH_NONE) |
- FP_DENORM_MODE_DP(FP_DENORM_FLUSH_NONE);
+ FP_DENORM_MODE_SP(FP32Denormals) |
+ FP_DENORM_MODE_DP(FP64Denormals);
}
static AsmPrinter *createAMDGPUAsmPrinterPass(TargetMachine &tm,
DisasmEnabled = TM.getSubtarget<AMDGPUSubtarget>().dumpCode();
}
+void AMDGPUAsmPrinter::EmitEndOfAsmFile(Module &M) {
+
+ // This label is used to mark the end of the .text section.
+ const TargetLoweringObjectFile &TLOF = getObjFileLowering();
+ OutStreamer.SwitchSection(TLOF.getTextSection());
+ MCSymbol *EndOfTextLabel =
+ OutContext.GetOrCreateSymbol(StringRef(END_OF_TEXT_LABEL_NAME));
+ OutStreamer.EmitLabel(EndOfTextLabel);
+}
+
bool AMDGPUAsmPrinter::runOnMachineFunction(MachineFunction &MF) {
+
+ // The starting address of all shader programs must be 256 bytes aligned.
+ MF.setAlignment(8);
+
SetupMachineFunction(MF);
- OutStreamer.emitRawComment(Twine('@') + MF.getName() + Twine(':'));
+ EmitFunctionHeader();
MCContext &Context = getObjFileLowering().getContext();
const MCSectionELF *ConfigSection = Context.getELFSection(".AMDGPU.config",
false);
OutStreamer.emitRawComment(" IeeeMode: " + Twine(KernelInfo.IEEEMode),
false);
+ OutStreamer.emitRawComment(" ScratchSize: " + Twine(KernelInfo.ScratchSize),
+ false);
} else {
R600MachineFunctionInfo *MFI = MF.getInfo<R600MachineFunctionInfo>();
OutStreamer.emitRawComment(
void AMDGPUAsmPrinter::EmitProgramInfoR600(const MachineFunction &MF) {
unsigned MaxGPR = 0;
bool killPixel = false;
- const R600RegisterInfo *RI
- = static_cast<const R600RegisterInfo*>(TM.getRegisterInfo());
+ const R600RegisterInfo *RI = static_cast<const R600RegisterInfo *>(
+ TM.getSubtargetImpl()->getRegisterInfo());
const R600MachineFunctionInfo *MFI = MF.getInfo<R600MachineFunctionInfo>();
const AMDGPUSubtarget &STM = TM.getSubtarget<AMDGPUSubtarget>();
void AMDGPUAsmPrinter::getSIProgramInfo(SIProgramInfo &ProgInfo,
const MachineFunction &MF) const {
+ const AMDGPUSubtarget &STM = TM.getSubtarget<AMDGPUSubtarget>();
+ const SIMachineFunctionInfo *MFI = MF.getInfo<SIMachineFunctionInfo>();
uint64_t CodeSize = 0;
unsigned MaxSGPR = 0;
unsigned MaxVGPR = 0;
bool VCCUsed = false;
- const SIRegisterInfo *RI
- = static_cast<const SIRegisterInfo*>(TM.getRegisterInfo());
+ bool FlatUsed = false;
+ const SIRegisterInfo *RI = static_cast<const SIRegisterInfo *>(
+ TM.getSubtargetImpl()->getRegisterInfo());
for (const MachineBasicBlock &MBB : MF) {
for (const MachineInstr &MI : MBB) {
reg == AMDGPU::VCC_HI) {
VCCUsed = true;
continue;
+ } else if (reg == AMDGPU::FLAT_SCR ||
+ reg == AMDGPU::FLAT_SCR_LO ||
+ reg == AMDGPU::FLAT_SCR_HI) {
+ FlatUsed = true;
+ continue;
}
switch (reg) {
if (VCCUsed)
MaxSGPR += 2;
- ProgInfo.NumVGPR = MaxVGPR;
- ProgInfo.NumSGPR = MaxSGPR;
+ if (FlatUsed)
+ MaxSGPR += 2;
+
+ // We found the maximum register index. They start at 0, so add one to get the
+ // number of registers.
+ ProgInfo.NumVGPR = MaxVGPR + 1;
+ ProgInfo.NumSGPR = MaxSGPR + 1;
+ ProgInfo.VGPRBlocks = (ProgInfo.NumVGPR - 1) / 4;
+ ProgInfo.SGPRBlocks = (ProgInfo.NumSGPR - 1) / 8;
// Set the value to initialize FP_ROUND and FP_DENORM parts of the mode
// register.
ProgInfo.FloatMode = getFPMode(MF);
// Do not clamp NAN to 0.
ProgInfo.DX10Clamp = 0;
- ProgInfo.CodeLen = CodeSize;
-}
+ const MachineFrameInfo *FrameInfo = MF.getFrameInfo();
+ ProgInfo.ScratchSize = FrameInfo->estimateStackSize(MF);
-void AMDGPUAsmPrinter::EmitProgramInfoSI(const MachineFunction &MF,
- const SIProgramInfo &KernelInfo) {
- const AMDGPUSubtarget &STM = TM.getSubtarget<AMDGPUSubtarget>();
- const SIMachineFunctionInfo *MFI = MF.getInfo<SIMachineFunctionInfo>();
-
- unsigned RsrcReg;
- switch (MFI->getShaderType()) {
- default: // Fall through
- case ShaderType::COMPUTE: RsrcReg = R_00B848_COMPUTE_PGM_RSRC1; break;
- case ShaderType::GEOMETRY: RsrcReg = R_00B228_SPI_SHADER_PGM_RSRC1_GS; break;
- case ShaderType::PIXEL: RsrcReg = R_00B028_SPI_SHADER_PGM_RSRC1_PS; break;
- case ShaderType::VERTEX: RsrcReg = R_00B128_SPI_SHADER_PGM_RSRC1_VS; break;
- }
+ ProgInfo.FlatUsed = FlatUsed;
+ ProgInfo.VCCUsed = VCCUsed;
+ ProgInfo.CodeLen = CodeSize;
unsigned LDSAlignShift;
if (STM.getGeneration() < AMDGPUSubtarget::SEA_ISLANDS) {
LDSAlignShift = 9;
}
- unsigned LDSBlocks =
- RoundUpToAlignment(MFI->LDSSize, 1 << LDSAlignShift) >> LDSAlignShift;
+ unsigned LDSSpillSize = MFI->LDSWaveSpillSize *
+ MFI->getMaximumWorkGroupSize(MF);
+
+ ProgInfo.LDSSize = MFI->LDSSize + LDSSpillSize;
+ ProgInfo.LDSBlocks =
+ RoundUpToAlignment(ProgInfo.LDSSize, 1 << LDSAlignShift) >> LDSAlignShift;
+
+ // Scratch is allocated in 256 dword blocks.
+ unsigned ScratchAlignShift = 10;
+ // We need to program the hardware with the amount of scratch memory that
+ // is used by the entire wave. ProgInfo.ScratchSize is the amount of
+ // scratch memory used per thread.
+ ProgInfo.ScratchBlocks =
+ RoundUpToAlignment(ProgInfo.ScratchSize * STM.getWavefrontSize(),
+ 1 << ScratchAlignShift) >> ScratchAlignShift;
+
+ ProgInfo.ComputePGMRSrc1 =
+ S_00B848_VGPRS(ProgInfo.VGPRBlocks) |
+ S_00B848_SGPRS(ProgInfo.SGPRBlocks) |
+ S_00B848_PRIORITY(ProgInfo.Priority) |
+ S_00B848_FLOAT_MODE(ProgInfo.FloatMode) |
+ S_00B848_PRIV(ProgInfo.Priv) |
+ S_00B848_DX10_CLAMP(ProgInfo.DX10Clamp) |
+ S_00B848_IEEE_MODE(ProgInfo.DebugMode) |
+ S_00B848_IEEE_MODE(ProgInfo.IEEEMode);
+
+ ProgInfo.ComputePGMRSrc2 =
+ S_00B84C_SCRATCH_EN(ProgInfo.ScratchBlocks > 0) |
+ S_00B84C_USER_SGPR(MFI->NumUserSGPRs) |
+ S_00B84C_TGID_X_EN(1) |
+ S_00B84C_TGID_Y_EN(1) |
+ S_00B84C_TGID_Z_EN(1) |
+ S_00B84C_TG_SIZE_EN(1) |
+ S_00B84C_TIDIG_COMP_CNT(2) |
+ S_00B84C_LDS_SIZE(ProgInfo.LDSBlocks);
+}
+
+static unsigned getRsrcReg(unsigned ShaderType) {
+ switch (ShaderType) {
+ default: // Fall through
+ case ShaderType::COMPUTE: return R_00B848_COMPUTE_PGM_RSRC1;
+ case ShaderType::GEOMETRY: return R_00B228_SPI_SHADER_PGM_RSRC1_GS;
+ case ShaderType::PIXEL: return R_00B028_SPI_SHADER_PGM_RSRC1_PS;
+ case ShaderType::VERTEX: return R_00B128_SPI_SHADER_PGM_RSRC1_VS;
+ }
+}
+
+void AMDGPUAsmPrinter::EmitProgramInfoSI(const MachineFunction &MF,
+ const SIProgramInfo &KernelInfo) {
+ const SIMachineFunctionInfo *MFI = MF.getInfo<SIMachineFunctionInfo>();
+ unsigned RsrcReg = getRsrcReg(MFI->getShaderType());
if (MFI->getShaderType() == ShaderType::COMPUTE) {
OutStreamer.EmitIntValue(R_00B848_COMPUTE_PGM_RSRC1, 4);
- const uint32_t ComputePGMRSrc1 =
- S_00B848_VGPRS(KernelInfo.NumVGPR / 4) |
- S_00B848_SGPRS(KernelInfo.NumSGPR / 8) |
- S_00B848_PRIORITY(KernelInfo.Priority) |
- S_00B848_FLOAT_MODE(KernelInfo.FloatMode) |
- S_00B848_PRIV(KernelInfo.Priv) |
- S_00B848_DX10_CLAMP(KernelInfo.DX10Clamp) |
- S_00B848_IEEE_MODE(KernelInfo.DebugMode) |
- S_00B848_IEEE_MODE(KernelInfo.IEEEMode);
-
- OutStreamer.EmitIntValue(ComputePGMRSrc1, 4);
+ OutStreamer.EmitIntValue(KernelInfo.ComputePGMRSrc1, 4);
OutStreamer.EmitIntValue(R_00B84C_COMPUTE_PGM_RSRC2, 4);
- OutStreamer.EmitIntValue(S_00B84C_LDS_SIZE(LDSBlocks), 4);
+ OutStreamer.EmitIntValue(KernelInfo.ComputePGMRSrc2, 4);
+
+ OutStreamer.EmitIntValue(R_00B860_COMPUTE_TMPRING_SIZE, 4);
+ OutStreamer.EmitIntValue(S_00B860_WAVESIZE(KernelInfo.ScratchBlocks), 4);
+
+ // TODO: Should probably note flat usage somewhere. SC emits a "FlatPtr32 =
+ // 0" comment but I don't see a corresponding field in the register spec.
} else {
OutStreamer.EmitIntValue(RsrcReg, 4);
- OutStreamer.EmitIntValue(S_00B028_VGPRS(KernelInfo.NumVGPR / 4) |
- S_00B028_SGPRS(KernelInfo.NumSGPR / 8), 4);
+ OutStreamer.EmitIntValue(S_00B028_VGPRS(KernelInfo.VGPRBlocks) |
+ S_00B028_SGPRS(KernelInfo.SGPRBlocks), 4);
}
if (MFI->getShaderType() == ShaderType::PIXEL) {
OutStreamer.EmitIntValue(R_00B02C_SPI_SHADER_PGM_RSRC2_PS, 4);
- OutStreamer.EmitIntValue(S_00B02C_EXTRA_LDS_SIZE(LDSBlocks), 4);
+ OutStreamer.EmitIntValue(S_00B02C_EXTRA_LDS_SIZE(KernelInfo.LDSBlocks), 4);
OutStreamer.EmitIntValue(R_0286CC_SPI_PS_INPUT_ENA, 4);
OutStreamer.EmitIntValue(MFI->PSInputAddr, 4);
}
projects / oota-llvm.git / blobdiff