kernel_mode_neon.rst 5.6 KB

123456789101112131415161718192021222324252627282930313233343536373839404142434445464748495051525354555657585960616263646566676869707172737475767778798081828384858687888990919293949596979899100101102103104105106107108109110111112113114115116117118119120121122123124
  1. ================
  2. Kernel mode NEON
  3. ================
  4. TL;DR summary
  5. -------------
  6. * Use only NEON instructions, or VFP instructions that don't rely on support
  7. code
  8. * Isolate your NEON code in a separate compilation unit, and compile it with
  9. '-march=armv7-a -mfpu=neon -mfloat-abi=softfp'
  10. * Put kernel_neon_begin() and kernel_neon_end() calls around the calls into your
  11. NEON code
  12. * Don't sleep in your NEON code, and be aware that it will be executed with
  13. preemption disabled
  14. Introduction
  15. ------------
  16. It is possible to use NEON instructions (and in some cases, VFP instructions) in
  17. code that runs in kernel mode. However, for performance reasons, the NEON/VFP
  18. register file is not preserved and restored at every context switch or taken
  19. exception like the normal register file is, so some manual intervention is
  20. required. Furthermore, special care is required for code that may sleep [i.e.,
  21. may call schedule()], as NEON or VFP instructions will be executed in a
  22. non-preemptible section for reasons outlined below.
  23. Lazy preserve and restore
  24. -------------------------
  25. The NEON/VFP register file is managed using lazy preserve (on UP systems) and
  26. lazy restore (on both SMP and UP systems). This means that the register file is
  27. kept 'live', and is only preserved and restored when multiple tasks are
  28. contending for the NEON/VFP unit (or, in the SMP case, when a task migrates to
  29. another core). Lazy restore is implemented by disabling the NEON/VFP unit after
  30. every context switch, resulting in a trap when subsequently a NEON/VFP
  31. instruction is issued, allowing the kernel to step in and perform the restore if
  32. necessary.
  33. Any use of the NEON/VFP unit in kernel mode should not interfere with this, so
  34. it is required to do an 'eager' preserve of the NEON/VFP register file, and
  35. enable the NEON/VFP unit explicitly so no exceptions are generated on first
  36. subsequent use. This is handled by the function kernel_neon_begin(), which
  37. should be called before any kernel mode NEON or VFP instructions are issued.
  38. Likewise, the NEON/VFP unit should be disabled again after use to make sure user
  39. mode will hit the lazy restore trap upon next use. This is handled by the
  40. function kernel_neon_end().
  41. Interruptions in kernel mode
  42. ----------------------------
  43. For reasons of performance and simplicity, it was decided that there shall be no
  44. preserve/restore mechanism for the kernel mode NEON/VFP register contents. This
  45. implies that interruptions of a kernel mode NEON section can only be allowed if
  46. they are guaranteed not to touch the NEON/VFP registers. For this reason, the
  47. following rules and restrictions apply in the kernel:
  48. * NEON/VFP code is not allowed in interrupt context;
  49. * NEON/VFP code is not allowed to sleep;
  50. * NEON/VFP code is executed with preemption disabled.
  51. If latency is a concern, it is possible to put back to back calls to
  52. kernel_neon_end() and kernel_neon_begin() in places in your code where none of
  53. the NEON registers are live. (Additional calls to kernel_neon_begin() should be
  54. reasonably cheap if no context switch occurred in the meantime)
  55. VFP and support code
  56. --------------------
  57. Earlier versions of VFP (prior to version 3) rely on software support for things
  58. like IEEE-754 compliant underflow handling etc. When the VFP unit needs such
  59. software assistance, it signals the kernel by raising an undefined instruction
  60. exception. The kernel responds by inspecting the VFP control registers and the
  61. current instruction and arguments, and emulates the instruction in software.
  62. Such software assistance is currently not implemented for VFP instructions
  63. executed in kernel mode. If such a condition is encountered, the kernel will
  64. fail and generate an OOPS.
  65. Separating NEON code from ordinary code
  66. ---------------------------------------
  67. The compiler is not aware of the special significance of kernel_neon_begin() and
  68. kernel_neon_end(), i.e., that it is only allowed to issue NEON/VFP instructions
  69. between calls to these respective functions. Furthermore, GCC may generate NEON
  70. instructions of its own at -O3 level if -mfpu=neon is selected, and even if the
  71. kernel is currently compiled at -O2, future changes may result in NEON/VFP
  72. instructions appearing in unexpected places if no special care is taken.
  73. Therefore, the recommended and only supported way of using NEON/VFP in the
  74. kernel is by adhering to the following rules:
  75. * isolate the NEON code in a separate compilation unit and compile it with
  76. '-march=armv7-a -mfpu=neon -mfloat-abi=softfp';
  77. * issue the calls to kernel_neon_begin(), kernel_neon_end() as well as the calls
  78. into the unit containing the NEON code from a compilation unit which is *not*
  79. built with the GCC flag '-mfpu=neon' set.
  80. As the kernel is compiled with '-msoft-float', the above will guarantee that
  81. both NEON and VFP instructions will only ever appear in designated compilation
  82. units at any optimization level.
  83. NEON assembler
  84. --------------
  85. NEON assembler is supported with no additional caveats as long as the rules
  86. above are followed.
  87. NEON code generated by GCC
  88. --------------------------
  89. The GCC option -ftree-vectorize (implied by -O3) tries to exploit implicit
  90. parallelism, and generates NEON code from ordinary C source code. This is fully
  91. supported as long as the rules above are followed.
  92. NEON intrinsics
  93. ---------------
  94. NEON intrinsics are also supported. However, as code using NEON intrinsics
  95. relies on the GCC header <arm_neon.h>, (which #includes <stdint.h>), you should
  96. observe the following in addition to the rules above:
  97. * Compile the unit containing the NEON intrinsics with '-ffreestanding' so GCC
  98. uses its builtin version of <stdint.h> (this is a C99 header which the kernel
  99. does not supply);
  100. * Include <arm_neon.h> last, or at least after <linux/types.h>