R600/SI: Add FP mode bits to binary.

[arsenm]

Summary

The default rounding mode to initialize the mode register needs
to be reported to the runtime. Fill in other bits a kernel
may be interested in setting for future use.

[arsenm]

Default to flushing single precision denormals. It seems some important single precision instructions we would want to use do not support denormals regardless of the mode. It would be more work to properly support functions that would want to use them. It's also what the AMD OpenCL does and it kind of makes sense to match that behavior

[arsenm]

Ping

[arsenm]

r211791

lib/Target/R600/AMDGPUAsmPrinter.h

 private:
   struct SIProgramInfo {
     SIProgramInfo() :
-      CodeLen(0),
+      NumVGPR(0),
       NumSGPR(0),
-      NumVGPR(0) {}
+      Priority(0),
+      FloatMode(0),
+      Priv(0),
+      DX10Clamp(0),
+      DebugMode(0),
+      IEEEMode(0),
+      CodeLen(0) {}
 
+    // Fields set in PGM_RSRC1 pm4 packet.
+    uint32_t NumVGPR;
+    uint32_t NumSGPR;
+    uint32_t Priority;
+    uint32_t FloatMode;
+    uint32_t Priv;
+    uint32_t DX10Clamp;
+    uint32_t DebugMode;
+    uint32_t IEEEMode;
+
+    // Bonus information for debugging.
     uint64_t CodeLen;
-    unsigned NumSGPR;
-    unsigned NumVGPR;
   };
 
   void getSIProgramInfo(SIProgramInfo &Out, MachineFunction &MF) const;

lib/Target/R600/AMDGPUAsmPrinter.cpp

 
 using namespace llvm;
 
+// TODO: This should get the default rounding mode from the kernel. We just set
+// the default here, but this could change if the OpenCL rounding mode pragmas
+// are used.
+//
+// The denormal mode here should match what is reported by the OpenCL runtime
+// for the CL_FP_DENORM bit from CL_DEVICE_{HALF|SINGLE|DOUBLE}_FP_CONFIG, but
+// can also be override to flush with the -cl-denorms-are-zero compiler flag.
+//
+// AMD OpenCL only sets flush none and reports CL_FP_DENORM for double
+// 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.
+//
+// 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 be able to use these so it's
+// probably best to just report no single precision denormals.
+static uint32_t getFPMode(MachineFunction &) {
+  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_IN_FLUSH_OUT) |
+         FP_DENORM_MODE_DP(FP_DENORM_FLUSH_NONE);
+}
 
 static AsmPrinter *createAMDGPUAsmPrinterPass(TargetMachine &tm,
                                               MCStreamer &Streamer) {

                                  false);
       OutStreamer.emitRawComment(" NumVgprs: " + Twine(KernelInfo.NumVGPR),
                                  false);
+      OutStreamer.emitRawComment(" FloatMode: " + Twine(KernelInfo.FloatMode),
+                                 false);
+      OutStreamer.emitRawComment(" IeeeMode: " + Twine(KernelInfo.IEEEMode),
+                                 false);
     } else {
       R600MachineFunctionInfo *MFI = MF.getInfo<R600MachineFunctionInfo>();
       OutStreamer.emitRawComment(

   if (VCCUsed)
     MaxSGPR += 2;
 
-  ProgInfo.CodeLen = CodeSize;
-  ProgInfo.NumSGPR = MaxSGPR;
   ProgInfo.NumVGPR = MaxVGPR;
+  ProgInfo.NumSGPR = MaxSGPR;
+
+  // Set the value to initialize FP_ROUND and FP_DENORM parts of the mode
+  // register.
+  ProgInfo.FloatMode = getFPMode(MF);
+
+  // XXX: Not quite sure what this does, but sc seems to unset this.
+  ProgInfo.IEEEMode = 0;
+
+  // Do not clamp NAN to 0.
+  ProgInfo.DX10Clamp = 0;
+
+  ProgInfo.CodeLen = CodeSize;
 }
 
 void AMDGPUAsmPrinter::EmitProgramInfoSI(MachineFunction &MF,
                                          const SIProgramInfo &KernelInfo) {
   const AMDGPUSubtarget &STM = TM.getSubtarget<AMDGPUSubtarget>();
-
   SIMachineFunctionInfo *MFI = MF.getInfo<SIMachineFunctionInfo>();
+
   unsigned RsrcReg;
   switch (MFI->ShaderType) {
   default: // Fall through

   case ShaderType::VERTEX:   RsrcReg = R_00B128_SPI_SHADER_PGM_RSRC1_VS; break;
   }
 
-  OutStreamer.EmitIntValue(RsrcReg, 4);
-  OutStreamer.EmitIntValue(S_00B028_VGPRS(KernelInfo.NumVGPR / 4) |
-                           S_00B028_SGPRS(KernelInfo.NumSGPR / 8), 4);
-
   unsigned LDSAlignShift;
   if (STM.getGeneration() < AMDGPUSubtarget::SEA_ISLANDS) {
-    // LDS is allocated in 64 dword blocks
+    // LDS is allocated in 64 dword blocks.
     LDSAlignShift = 8;
   } else {
-    // LDS is allocated in 128 dword blocks
+    // LDS is allocated in 128 dword blocks.
     LDSAlignShift = 9;
   }
+
   unsigned LDSBlocks =
-          RoundUpToAlignment(MFI->LDSSize, 1 << LDSAlignShift) >> LDSAlignShift;
 
   if (MFI->ShaderType == 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(R_00B84C_COMPUTE_PGM_RSRC2, 4);
     OutStreamer.EmitIntValue(S_00B84C_LDS_SIZE(LDSBlocks), 4);
+  } else {
+    OutStreamer.EmitIntValue(RsrcReg, 4);
+    OutStreamer.EmitIntValue(S_00B028_VGPRS(KernelInfo.NumVGPR / 4) |
+                             S_00B028_SGPRS(KernelInfo.NumSGPR / 8), 4);
   }
+
   if (MFI->ShaderType == ShaderType::PIXEL) {
     OutStreamer.EmitIntValue(R_00B02C_SPI_SHADER_PGM_RSRC2_PS, 4);
     OutStreamer.EmitIntValue(S_00B02C_EXTRA_LDS_SIZE(LDSBlocks), 4);

lib/Target/R600/SIDefines.h

 #define   S_00B84C_LDS_SIZE(x)                                        (((x) & 0x1FF) << 15)
 #define R_0286CC_SPI_PS_INPUT_ENA                                       0x0286CC
 
+
+#define R_00B848_COMPUTE_PGM_RSRC1                                      0x00B848
+#define   S_00B848_VGPRS(x)                                           (((x) & 0x3F) << 0)
+#define   G_00B848_VGPRS(x)                                           (((x) >> 0) & 0x3F)
+#define   C_00B848_VGPRS                                              0xFFFFFFC0
+#define   S_00B848_SGPRS(x)                                           (((x) & 0x0F) << 6)
+#define   G_00B848_SGPRS(x)                                           (((x) >> 6) & 0x0F)
+#define   C_00B848_SGPRS                                              0xFFFFFC3F
+#define   S_00B848_PRIORITY(x)                                        (((x) & 0x03) << 10)
+#define   G_00B848_PRIORITY(x)                                        (((x) >> 10) & 0x03)
+#define   C_00B848_PRIORITY                                           0xFFFFF3FF
+#define   S_00B848_FLOAT_MODE(x)                                      (((x) & 0xFF) << 12)
+#define   G_00B848_FLOAT_MODE(x)                                      (((x) >> 12) & 0xFF)
+#define   C_00B848_FLOAT_MODE                                         0xFFF00FFF
+#define   S_00B848_PRIV(x)                                            (((x) & 0x1) << 20)
+#define   G_00B848_PRIV(x)                                            (((x) >> 20) & 0x1)
+#define   C_00B848_PRIV                                               0xFFEFFFFF
+#define   S_00B848_DX10_CLAMP(x)                                      (((x) & 0x1) << 21)
+#define   G_00B848_DX10_CLAMP(x)                                      (((x) >> 21) & 0x1)
+#define   C_00B848_DX10_CLAMP                                         0xFFDFFFFF
+#define   S_00B848_DEBUG_MODE(x)                                      (((x) & 0x1) << 22)
+#define   G_00B848_DEBUG_MODE(x)                                      (((x) >> 22) & 0x1)
+#define   C_00B848_DEBUG_MODE                                         0xFFBFFFFF
+#define   S_00B848_IEEE_MODE(x)                                       (((x) & 0x1) << 23)
+#define   G_00B848_IEEE_MODE(x)                                       (((x) >> 23) & 0x1)
+#define   C_00B848_IEEE_MODE                                          0xFF7FFFFF
+
+
+// Helpers for setting FLOAT_MODE
+#define FP_ROUND_ROUND_TO_NEAREST 0
+#define FP_ROUND_ROUND_TO_INF 1
+#define FP_ROUND_ROUND_TO_NEGINF 2
+#define FP_ROUND_ROUND_TO_ZERO 3
+
+// Bits 3:0 control rounding mode. 1:0 control single precision, 3:2 double
+// precision.
+#define FP_ROUND_MODE_SP(x) ((x) & 0x3)
+#define FP_ROUND_MODE_DP(x) (((x) & 0x3) << 2)
+
+#define FP_DENORM_FLUSH_IN_FLUSH_OUT 0
+#define FP_DENORM_FLUSH_OUT 1
+#define FP_DENORM_FLUSH_IN 2
+#define FP_DENORM_FLUSH_NONE 3
+
+
+// Bits 7:4 control denormal handling. 5:4 control single precision, 6:7 double
+// precision.
+#define FP_DENORM_MODE_SP(x) (((x) & 0x3) << 4)
+#define FP_DENORM_MODE_DP(x) (((x) & 0x3) << 6)
+
 #endif // SIDEFINES_H_

test/CodeGen/R600/default-fp-mode.ll

+; RUN: llc -march=r600 -mcpu=SI < %s | FileCheck -check-prefix=SI %s
+
+; SI-LABEL: @test_kernel
+; SI: FloatMode: 192
+; SI: IeeeMode: 0
+define void @test_kernel(float addrspace(1)* %out0, double addrspace(1)* %out1) nounwind {
+  store float 0.0, float addrspace(1)* %out0
+  store double 0.0, double addrspace(1)* %out1
+  ret void
+}