229 | AMD 250 manual

25112 Rev. 3.06 September 2005

Software Optimization Guide for AMD64 Processors

instructions that are not dependent on previous or presently executing operations so that the processor can mask the execution latency by keeping itself busy, as illustrated below:

Instruction	0	2	4	6	8	10	12	14	16	18
MOVQ	xxxxxx
MOVQ	xxxxxx
MOVQ		xxxxxx
MOVQ		xxxxxx
PSWAPD		xxxxxx
PSWAPD		xxxxxx
PSWAPD			xxxxxx
PSWAPD			xxxxxx
PFMUL		xxxxxxxxxxxxxxxxxx
PFMUL			xxxxxxxxxxxxxxxxxx
PFMUL			xxxxxxxxxxxxxxxxxx
PFMUL				xxxxxxxxxxxxxxxxxx
PFMUL				xxxxxxxxxxxxxxxxxx
PFMUL					xxxxxxxxxxxxxxxxxx
PFMUL					xxxxxxxxxxxxxxxxxx
PFMUL						xxxxxxxxxxxxxxxxxx
PFPNACC					xxxxxxxxxxxxxxxxxxx
PFPNACC						xxxxxxxxxxxxxxxxxxx
PFPNACC						xxxxxxxxxxxxxxxxxxx
PFPNACC							xxxxxxxxxxxxxxxxxxx

Multiplying four complex single-precision numbers only takes 17 cycles as opposed to 14 cycles to multiply one complex single-precision number. The floating-point pipes are kept busy by feeding new instructions into the floating-point pipeline each cycle. In the arrangement above, 24 floating-point operations are performed in 17 cycles, achieving more than a 3.5x increase in performance.

The last optimization in both implementations is the use of the MOVNTQ and MOVNTPS instructions, nontemporal writes to memory that stream data to main memory. These instructions increase throughput to memory and make more efficient use of the bandwidth provided by the processor and memory controller. Nontemporal writes, such as MOVNTQ, MOVNTPS, and MOVNTDQ, should only be used on data that is not going to be accessed again in the near future.

Chapter 9

Optimizing with SIMD Instructions

229

Image 245

AMD 250 manual 229

Contents

Software Optimization Guide For AMD64 Processors 2001 2005 Advanced Micro Devices, Inc. All rights reserved Contents General 64-Bit Optimizations Chapter Cache and Memory Optimizations Chapter Integer Optimizations 159 Chapter X87 Floating-Point Optimizations 237 Viii Appendix B Implementation of Write-Combining Index Software Optimization Guide for AMD64 Processors Tables Tables Xii Tables Figures Xiv Revision History Xvi Revision History Getting Started Quickly Intended Audience Using This Guide Numbering Systems Special InformationTypographic Notation Providing Feedback Internal Instruction Formats Important New TermsPrimitive Operations Mrom Types of InstructionsInstructions, Macro-ops and Micro-ops Guideline Key OptimizationsKey Optimizations by Rank Optimizations by Rank C++ Source-Level Optimizations C++ Source-Level Optimizations Optimization Declarations of Floating-Point ValuesApplication Rationale Using Arrays and Pointers Matrix Example Instead, use the equivalent array notation Additional Considerations Related Information Unrolling Small Loops Expression Order in Compound Branch Conditions Chapter C++ Source-Level Optimizations Long Logical Expressions in If Statements Arrange Boolean Operands for Quick Expression Evaluation If *p == y && strlenp Dynamic Memory Allocation Consideration Listing 3. Avoid Unnecessary Store-to-Load Dependencies Application Examples Matching Store and Load Size Listing 6. Preferred C++ Source-Level Optimizations Switch and Noncontiguous Case Expressions Example Related Information Arranging Cases by Probability of Occurrence Use of Function Prototypes Use of const Type Qualifier Rationale and Examples Generic Loop Hoisting Listing Chapter C++ Source-Level Optimizations Local Static Functions Explicit Parallelism in Code Listing 11. Preferred Extracting Common Subexpressions Listing 15. Example 2 Preferred Sorting and Padding C and C++ Structures Sorting and Padding C and C++ Structures C++ Source-Level Optimizations Sorting Local Variables Related Information Replacing Integer Division with Multiplication Listing 16. Avoid Frequently Dereferenced Pointer Arguments Listing 17. Preferred Array Indices Rational 23 32-Bit Integral Data Types Sign of Integer Operands Listing 20. Example 2 Avoid Accelerating Floating-Point Division and Square Root Examples Fast Floating-Point-to-Integer Conversion Listing 23. Slow Branches Dependent on Integer Comparisons Are Fast Speeding Up Branches Based on Comparisons Between Floats Comparisons against Positive Constant Comparisons among Two Floats Improving Performance in Linux Libraries Software Optimization Guide for AMD64 Processors General 64-Bit Optimizations 64-Bit Registers and Integer Arithmetic This code performs 64-bit addition using 32-bit registers ESP+8ESP+4 = multiplicand Background 64-Bit Arithmetic and Large-Integer Multiplication G1 = c3 E1 + f1 + g0 = c2 D1 + e0 + f0 = c1 D0 = c0 XMM END Text Segment 128-Bit Media Instructions and Floating-Point Operations 32-Bit Legacy GPRs and Small Unsigned Integers Chapter General 64-Bit Optimizations General 64-Bit Optimizations Instruction-Decoding Optimizations DirectPath Instructions Load-Execute Integer Instructions Optimization Load-Execute Instructions Movss xmm0, floatvar1 mulss xmm0, floatvar2 Application Branch Targets in Program Hot Spots 32/64-Bit vs -Bit Forms of the LEA Instruction Take Advantage of x86 and AMD64 Complex Addressing Modes Cmpb %al,0x68e35%r10,%r13 Short Instruction Encodings Partial-Register Reads and Writes Avoid Functions That Do not Allocate Local Variables Using Leave for Function EpiloguesFunctions That Allocate Local Variables Traditional function epilogue looks like this Alternatives to Shld Instruction Lea reg1, reg1*8+reg2 10 8-Bit Sign-Extended Immediate Values 11 8-Bit Sign-Extended Displacements NOP Code Padding with Operand-Size Override Instruction-Decoding Optimizations Cache and Memory Optimizations Examples-Store-to-Load-Forwarding Stalls Memory-Size MismatchesBit Avoid Avoid Preferred If Stores Are Close to the Load Preferred If the Contents of MM0 are No Longer NeededExamples-Large-to-small Mismatches Preferred If the Stores and Loads are Close Together, Option Natural Alignment of Data Objects Cache-Coherent Nonuniform Memory Access ccNUMA CPU0 OS Implications Dual-Core AMD Opteron Processor Configuration Multiprocessor Considerations Store-to-Load Forwarding Pitfalls-True Dependencies Store-to-Load Forwarding RestrictionsNarrow-to-Wide Store-Buffer Data-Forwarding Restriction 100 101 Wide-to-Narrow Store-Buffer Data-Forwarding RestrictionMisaligned Store-Buffer Data-Forwarding Restriction One Supported Store-to-Load Forwarding Case High-Byte Store-Buffer Data-Forwarding RestrictionStore-to-Load Forwarding-False Dependencies 102 103 Summary of Store-to-Load-Forwarding Pitfalls to Avoid 104 Prefetch InstructionsPrefetching versus Preloading Hardware Prefetching Unit-Stride AccessPREFETCH/W versus PREFETCHNTA/T0/T1/T2 105 106 Prefetchw versus PrefetchWrite-Combining Usage 107 Multiple Prefetches 108 Determining Prefetch Distance Processor-Limited Code Memory-Limited Code Cache and Memory Optimizations Definitions 111 Prefetch at Least 64 Bytes Away from Surrounding Stores Streaming-Store/Non-Temporal Instructions 113 Write-combining Fields Used to Address the Multibank L1 Data Cache How to Know If a Bank Conflict ExistsL1 Data Cache Bank Conflicts 115 Placing Code and Data in the Same 64-Byte Cache Line 117 Sorting and Padding C and C++ Structures 118 119 120 Memory CopyCopying Small Data Structures 121 Extend Arguments to 32 Bits Before Pushing onto Stack Stack ConsiderationsOptimized Stack Usage 122 123 Cache Issues when Writing Instruction Bytes to Memory 124 Interleave Loads and Stores 125 This Chapter 126 Density of Branches Align 127 128 Two-Byte Near-Return RET Instruction 129 Signed Integer ABS Function x = labsx Branches That Depend on Random DataUnsigned Integer min Function z = x y ? x y 130 131 Conditional Write 132 Pairing Call and Return 133 Recursive Functions 134 135 Nonzero Code-Segment Base Values 136 Replacing Branches with ComputationMuxing Constructs 137 SSE Solution PreferredMMX Solution Avoid Example 1 C Code Sample Code Translated into AMD64 CodeExample 1 3DNow! Code Example 2 C Code Example 3 3DNow! Code Example 3 C CodeExample 4 C Code Example 4 3DNow! Code 140 Example 5 C CodeExample 5 3DNow! Code 141 Loop Instruction 142 Far Control-Transfer Instructions 143 Chapter Scheduling Optimizations 144 Instruction Scheduling by Latency 145 Loop UnrollingLoop Unrolling Example Complete Loop Unrolling Complete Loop UnrollingPartial Loop Unrolling Example Partial Loop Unrolling 147 Fadd 148 Deriving the Loop Control for Partially Unrolled Loops 149 Inline Functions 150 Additional Recommendations 151 Address-Generation InterlocksAddress-Generation Interlocks 152 153 Movzx and Movsx 154 Pointer Arithmetic in Loops 155 156 157 Pushing Memory Data Directly onto the Stack 158 159 Chapter Integer Optimizations Multiplication by Reciprocal Division Utility Replacing Division with MultiplicationSigned Division Utility Unsigned Division Utility Algorithm Divisors 1 = d 231, Odd d Unsigned Division by Multiplication of ConstantAlgorithm Divisors 231 = d 161 Simpler Code for Restricted Dividend Signed Division by Multiplication of ConstantAlgorithm Divisors 2 = d Signed Division by Signed Division by -2n Signed Division by 2nRemainder of Signed Division by 2 or Remainder of Signed Division by 2n or -2n 164 Alternative Code for Multiplying by a Constant 165 166 Latency of Repeated String Instructions Repeated String InstructionsGuidelines for Repeated String Instructions 167 168 Use the Largest Possible Operand Size 169 Using XOR to Clear Integer RegistersAcceptable Bit Addition Efficient 64-Bit Integer Arithmetic in 32-Bit ModeBit Subtraction Bit Negation Bit Multiplication Bit Right ShiftBit Unsigned Division 171 172 173 Bit Signed Division 174 Bit Unsigned Remainder Computation 175 176 Bit Signed Remainder Computation 177 178 179 180 Integer Version 181 Efficient Binary-to-ASCII Decimal ConversionBinary-to-ASCII Decimal Conversion Retaining Leading Zeros 182 183 Binary-to-ASCII Decimal Conversion Suppressing Leading Zeros 184 185 186 Unsigned Integer Division 187 188 189 Signed Integer DivisionExample Code 190 191 192 Optimizing Integer Division 193 Optimizing with Simd Instructions 194 195 Ensure All Packed Floating-Point Data are Aligned 196 Rationale-Single Precision 197 Rational-Double Precision 198 Use MOVLPx/MOVHPx Instructions for Unaligned Data Access 199 Use Movapd and Movaps Instead of Movupd and Movups Loop type Description 200 201 202 Double-Precision 32 ⋅ 32 Matrix Multiplication 203 204 205 206 207 208 Passing Data between MMX and 3DNow! Instructions 209 Storing Floating-Point Data in MMX Registers 210 Emms and Femms Usage XMM Text Segment 211 212 213 214 215 Single PrecisionDouble Precision 216 Clearing MMX and XMM Registers with XOR Instructions 217 218 219 220 221 Code below Puts the Floating Point Sign Mask 222 223 224 225 226 Pfpnacc 227 228 229 230 Listing 27 ⋅ 4 Matrix Multiplication SSE 231 XMM3 232 Listing 28 ⋅ 4 Matrix Multiplication 3DNow! Technology 233 234 235 236 237 X87 Floating-Point Optimizations 238 Using Multiplication Rather Than Division 239 Achieving Two Floating-Point Operations per Clock Cycle 240 241 242 Align and Pack DirectPath x87 Instructions 243 244 Floating-Point Compare Instructions 245 Using the Fxch Instruction Rather Than FST/FLD Pairs 246 Floating-Point Subexpression Elimination 247 Accumulating Precision-Sensitive Quantities in x87 Registers 248 Avoiding Extended-Precision Data 249 Key Microarchitecture Features 251 Processor Block DiagramSuperscalar Processor L1 Instruction Cache AMD Athlon 64 and AMD Opteron Processors Block Diagram 253 L1 Instruction Cache Specifications by ProcessorBranch-Prediction Table Fetch-Decode Unit L1 Instruction TLB SpecificationsInstruction Control Unit Translation-Lookaside Buffer 10 L1 Data Cache L1 Data TLB SpecificationsL2 TLB Specifications Integer Execution Unit L1 Data Cache Specifications by ProcessorInteger Scheduler 257 Floating-Point Scheduler Floating-Point Unit Floating-Point Execution UnitLoad-Store Unit 259 16 L2 Cache HyperTransport Technology Interface Buses for AMD Athlon 64 and AMD Opteron ProcessorIntegrated Memory Controller 261 HyperTransport Technology Software Optimization Guide for AMD64 Processors 263 Write-Combining Definitions and Abbreviations 264 Programming DetailsWrite-combining Operations 265 Write-Combining Completion Events 266 Sending Write-Buffer Data to the System 267 Optimizations 268 269 Appendix C Instruction Latencies Example Instruction Entry Understanding Instruction EntriesParts of the Instruction Entry 270 271 Interpreting Placeholders 272 Interpreting Latencies Integer Instructions Integer Instructions273 AAA ADD reg16/32/64, mem16/32/64 274 Bswap EAX/RAX/R8 275 276 277 Mem16/32/64 CMOVNP/CMOVPO reg16/32/64, reg16/32/64 278 CMP reg16/32/64, mreg16/32/64 279 280 281 282 283 JA/JNBE disp16/32 284 Lahf 285 286 Mfence 287 288 NOP Xchg EAX, EAX 289 290 291 292 Rdmsr 293Rdpmc Rdtsc Sahf 294 SBB reg16/32/64, mem16/32/64 295 296 297 298 STC 299STD STI Syscall 300Sysenter Sysexit 301 Xchg AX/EAX/RAX, DI/EDI/RDI/R15 302Xchg AX/EAX/RAX, SI/ESI/RSI/R14 303 MMX Technology InstructionsMMX Technology Instructions Fmul 304 305 306 307 X87 Floating-Point InstructionsX87 Floating-Point Instructions Fadd Fcompp 308Fadd Fcos Fdecstp FADD/FMUL Fstore Finit 309Fincstp 310 311 312 Fxch 313Fxtract FYL2X 3DNow! Technology Instructions 3DNow! Technology Instructions314 Femms 315 DNow! Technology Instructions 316 3DNow! Technology Extensions3DNow! Technology Extensions 317 SSE InstructionsSSE Instructions 318 319 320 321 Mem64 Prefetchnta mem8 0Fh 18h Mm-000-xxx DirectPath 322 Sfence 323 324 325 326 SSE2 InstructionsSSE2 Instructions Fmul Fstore 3270FH 328 Maskmovdqu 329 Fadd Fmul Fstore 330 331 332 Fadd Fmul 333 334 335 336 337 338 Fmul Punpckhdq 339Fmul Punpckhbw 340 341 342 SSE3 InstructionsSSE3 Instructions 343 344 345 Fast-Write Optimizations 346 Fast-Write Optimizations for Graphics-Engine Programming 347 Cacheable-Memory Command Structure 348 349 Fast-Write Optimizations for Video-Memory Copies 350 351 Memory Optimizations 352 Northbridge Command Flow 353 Optimizations for Texture-Map Copies to AGP MemoryOptimizations for Vertex-Geometry Copies to AGP Memory 354 355 Types of XMM-Register DataTypes of SSE and SSE2 Instructions 356 Half-Register Operations 357 Zeroing Out an XMM RegisterClearing XMM Registers 358 359 Reuse of Dead Registers 360 Moving Data Between XMM Registers and GPRs 361 Saving and Restoring Registers of Unknown Format 362 SSE and SSE2 Copy Loops 363 Explicit Load Instructions 364 Data ConversionConverting Scalar Values Converting Directly from Memory Converting Vector Values365 INT GPR FPS 366 367 Numerics 368 Index