1 =============================
2 User Guide for NVPTX Back-end
3 =============================
13 To support GPU programming, the NVPTX back-end supports a subset of LLVM IR
14 along with a defined set of conventions used to represent GPU programming
15 concepts. This document provides an overview of the general usage of the back-
16 end, including a description of the conventions used and the set of accepted
21 This document assumes a basic familiarity with CUDA and the PTX
22 assembly language. Information about the CUDA Driver API and the PTX assembly
23 language can be found in the `CUDA documentation
24 <http://docs.nvidia.com/cuda/index.html>`_.
31 Marking Functions as Kernels
32 ----------------------------
34 In PTX, there are two types of functions: *device functions*, which are only
35 callable by device code, and *kernel functions*, which are callable by host
36 code. By default, the back-end will emit device functions. Metadata is used to
37 declare a function as a kernel function. This metadata is attached to the
38 ``nvvm.annotations`` named metadata object, and has the following format:
42 !0 = !{<function-ref>, metadata !"kernel", i32 1}
44 The first parameter is a reference to the kernel function. The following
45 example shows a kernel function calling a device function in LLVM IR. The
46 function ``@my_kernel`` is callable from host code, but ``@my_fmad`` is not.
50 define float @my_fmad(float %x, float %y, float %z) {
51 %mul = fmul float %x, %y
52 %add = fadd float %mul, %z
56 define void @my_kernel(float* %ptr) {
57 %val = load float, float* %ptr
58 %ret = call float @my_fmad(float %val, float %val, float %val)
59 store float %ret, float* %ptr
63 !nvvm.annotations = !{!1}
64 !1 = !{void (float*)* @my_kernel, !"kernel", i32 1}
66 When compiled, the PTX kernel functions are callable by host-side code.
74 The NVPTX back-end uses the following address space mapping:
76 ============= ======================
77 Address Space Memory Space
78 ============= ======================
85 ============= ======================
87 Every global variable and pointer type is assigned to one of these address
88 spaces, with 0 being the default address space. Intrinsics are provided which
89 can be used to convert pointers between the generic and non-generic address
92 As an example, the following IR will define an array ``@g`` that resides in
97 @g = internal addrspace(1) global [4 x i32] [ i32 0, i32 1, i32 2, i32 3 ]
99 LLVM IR functions can read and write to this array, and host-side code can
100 copy data to it by name with the CUDA Driver API.
102 Note that since address space 0 is the generic space, it is illegal to have
103 global variables in address space 0. Address space 0 is the default address
104 space in LLVM, so the ``addrspace(N)`` annotation is *required* for global
111 The NVPTX target uses the module triple to select between 32/64-bit code
112 generation and the driver-compiler interface to use. The triple architecture
113 can be one of ``nvptx`` (32-bit PTX) or ``nvptx64`` (64-bit PTX). The
114 operating system should be one of ``cuda`` or ``nvcl``, which determines the
115 interface used by the generated code to communicate with the driver. Most
116 users will want to use ``cuda`` as the operating system, which makes the
117 generated PTX compatible with the CUDA Driver API.
119 Example: 32-bit PTX for CUDA Driver API: ``nvptx-nvidia-cuda``
121 Example: 64-bit PTX for CUDA Driver API: ``nvptx64-nvidia-cuda``
125 .. _nvptx_intrinsics:
130 Address Space Conversion
131 ------------------------
133 '``llvm.nvvm.ptr.*.to.gen``' Intrinsics
134 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
139 These are overloaded intrinsics. You can use these on any pointer types.
143 declare i8* @llvm.nvvm.ptr.global.to.gen.p0i8.p1i8(i8 addrspace(1)*)
144 declare i8* @llvm.nvvm.ptr.shared.to.gen.p0i8.p3i8(i8 addrspace(3)*)
145 declare i8* @llvm.nvvm.ptr.constant.to.gen.p0i8.p4i8(i8 addrspace(4)*)
146 declare i8* @llvm.nvvm.ptr.local.to.gen.p0i8.p5i8(i8 addrspace(5)*)
151 The '``llvm.nvvm.ptr.*.to.gen``' intrinsics convert a pointer in a non-generic
152 address space to a generic address space pointer.
157 These intrinsics modify the pointer value to be a valid generic address space
161 '``llvm.nvvm.ptr.gen.to.*``' Intrinsics
162 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
167 These are overloaded intrinsics. You can use these on any pointer types.
171 declare i8 addrspace(1)* @llvm.nvvm.ptr.gen.to.global.p1i8.p0i8(i8*)
172 declare i8 addrspace(3)* @llvm.nvvm.ptr.gen.to.shared.p3i8.p0i8(i8*)
173 declare i8 addrspace(4)* @llvm.nvvm.ptr.gen.to.constant.p4i8.p0i8(i8*)
174 declare i8 addrspace(5)* @llvm.nvvm.ptr.gen.to.local.p5i8.p0i8(i8*)
179 The '``llvm.nvvm.ptr.gen.to.*``' intrinsics convert a pointer in the generic
180 address space to a pointer in the target address space. Note that these
181 intrinsics are only useful if the address space of the target address space of
182 the pointer is known. It is not legal to use address space conversion
183 intrinsics to convert a pointer from one non-generic address space to another
184 non-generic address space.
189 These intrinsics modify the pointer value to be a valid pointer in the target
190 non-generic address space.
193 Reading PTX Special Registers
194 -----------------------------
196 '``llvm.nvvm.read.ptx.sreg.*``'
197 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
204 declare i32 @llvm.nvvm.read.ptx.sreg.tid.x()
205 declare i32 @llvm.nvvm.read.ptx.sreg.tid.y()
206 declare i32 @llvm.nvvm.read.ptx.sreg.tid.z()
207 declare i32 @llvm.nvvm.read.ptx.sreg.ntid.x()
208 declare i32 @llvm.nvvm.read.ptx.sreg.ntid.y()
209 declare i32 @llvm.nvvm.read.ptx.sreg.ntid.z()
210 declare i32 @llvm.nvvm.read.ptx.sreg.ctaid.x()
211 declare i32 @llvm.nvvm.read.ptx.sreg.ctaid.y()
212 declare i32 @llvm.nvvm.read.ptx.sreg.ctaid.z()
213 declare i32 @llvm.nvvm.read.ptx.sreg.nctaid.x()
214 declare i32 @llvm.nvvm.read.ptx.sreg.nctaid.y()
215 declare i32 @llvm.nvvm.read.ptx.sreg.nctaid.z()
216 declare i32 @llvm.nvvm.read.ptx.sreg.warpsize()
221 The '``@llvm.nvvm.read.ptx.sreg.*``' intrinsics provide access to the PTX
222 special registers, in particular the kernel launch bounds. These registers
223 map in the following way to CUDA builtins:
225 ============ =====================================
226 CUDA Builtin PTX Special Register Intrinsic
227 ============ =====================================
228 ``threadId`` ``@llvm.nvvm.read.ptx.sreg.tid.*``
229 ``blockIdx`` ``@llvm.nvvm.read.ptx.sreg.ctaid.*``
230 ``blockDim`` ``@llvm.nvvm.read.ptx.sreg.ntid.*``
231 ``gridDim`` ``@llvm.nvvm.read.ptx.sreg.nctaid.*``
232 ============ =====================================
238 '``llvm.nvvm.barrier0``'
239 ^^^^^^^^^^^^^^^^^^^^^^^^^^^
246 declare void @llvm.nvvm.barrier0()
251 The '``@llvm.nvvm.barrier0()``' intrinsic emits a PTX ``bar.sync 0``
252 instruction, equivalent to the ``__syncthreads()`` call in CUDA.
258 For the full set of NVPTX intrinsics, please see the
259 ``include/llvm/IR/IntrinsicsNVVM.td`` file in the LLVM source tree.
264 Linking with Libdevice
265 ======================
267 The CUDA Toolkit comes with an LLVM bitcode library called ``libdevice`` that
268 implements many common mathematical functions. This library can be used as a
269 high-performance math library for any compilers using the LLVM NVPTX target.
270 The library can be found under ``nvvm/libdevice/`` in the CUDA Toolkit and
271 there is a separate version for each compute architecture.
273 For a list of all math functions implemented in libdevice, see
274 `libdevice Users Guide <http://docs.nvidia.com/cuda/libdevice-users-guide/index.html>`_.
276 To accommodate various math-related compiler flags that can affect code
277 generation of libdevice code, the library code depends on a special LLVM IR
278 pass (``NVVMReflect``) to handle conditional compilation within LLVM IR. This
279 pass looks for calls to the ``@__nvvm_reflect`` function and replaces them
280 with constants based on the defined reflection parameters. Such conditional
281 code often follows a pattern:
285 float my_function(float a) {
286 if (__nvvm_reflect("FASTMATH"))
287 return my_function_fast(a);
289 return my_function_precise(a);
292 The default value for all unspecified reflection parameters is zero.
294 The ``NVVMReflect`` pass should be executed early in the optimization
295 pipeline, immediately after the link stage. The ``internalize`` pass is also
296 recommended to remove unused math functions from the resulting PTX. For an
297 input IR module ``module.bc``, the following compilation flow is recommended:
299 1. Save list of external functions in ``module.bc``
300 2. Link ``module.bc`` with ``libdevice.compute_XX.YY.bc``
301 3. Internalize all functions not in list from (1)
302 4. Eliminate all unused internal functions
303 5. Run ``NVVMReflect`` pass
304 6. Run standard optimization pipeline
308 ``linkonce`` and ``linkonce_odr`` linkage types are not suitable for the
309 libdevice functions. It is possible to link two IR modules that have been
310 linked against libdevice using different reflection variables.
312 Since the ``NVVMReflect`` pass replaces conditionals with constants, it will
313 often leave behind dead code of the form:
319 br i1 true, label %foo, label %bar
326 Therefore, it is recommended that ``NVVMReflect`` is executed early in the
327 optimization pipeline before dead-code elimination.
329 The NVPTX TargetMachine knows how to schedule ``NVVMReflect`` at the beginning
330 of your pass manager; just use the following code when setting up your pass
335 std::unique_ptr<TargetMachine> TM = ...;
336 PassManagerBuilder PMBuilder(...);
338 TM->adjustPassManager(PMBuilder);
340 Reflection Parameters
341 ---------------------
343 The libdevice library currently uses the following reflection parameters to
344 control code generation:
346 ==================== ======================================================
348 ==================== ======================================================
349 ``__CUDA_FTZ=[0,1]`` Use optimized code paths that flush subnormals to zero
350 ==================== ======================================================
352 The value of this flag is determined by the "nvvm-reflect-ftz" module flag.
353 The following sets the ftz flag to 1.
357 !llvm.module.flag = !{!0}
358 !0 = !{i32 4, !"nvvm-reflect-ftz", i32 1}
360 (``i32 4`` indicates that the value set here overrides the value in another
361 module we link with. See the `LangRef <LangRef.html#module-flags-metadata>`
367 The most common way to execute PTX assembly on a GPU device is to use the CUDA
368 Driver API. This API is a low-level interface to the GPU driver and allows for
369 JIT compilation of PTX code to native GPU machine code.
371 Initializing the Driver API:
378 // Initialize the driver API
380 // Get a handle to the first compute device
381 cuDeviceGet(&device, 0);
382 // Create a compute device context
383 cuCtxCreate(&context, 0, device);
385 JIT compiling a PTX string to a device binary:
392 // JIT compile a null-terminated PTX string
393 cuModuleLoadData(&module, (void*)PTXString);
395 // Get a handle to the "myfunction" kernel function
396 cuModuleGetFunction(&function, module, "myfunction");
398 For full examples of executing PTX assembly, please see the `CUDA Samples
399 <https://developer.nvidia.com/cuda-downloads>`_ distribution.
405 ptxas complains of undefined function: __nvvm_reflect
406 -----------------------------------------------------
408 When linking with libdevice, the ``NVVMReflect`` pass must be used. See
409 :ref:`libdevice` for more information.
412 Tutorial: A Simple Compute Kernel
413 =================================
415 To start, let us take a look at a simple compute kernel written directly in
416 LLVM IR. The kernel implements vector addition, where each thread computes one
417 element of the output vector C from the input vectors A and B. To make this
418 easier, we also assume that only a single CTA (thread block) will be launched,
419 and that it will be one dimensional.
427 target datalayout = "e-p:64:64:64-i1:8:8-i8:8:8-i16:16:16-i32:32:32-i64:64:64-f32:32:32-f64:64:64-v16:16:16-v32:32:32-v64:64:64-v128:128:128-n16:32:64"
428 target triple = "nvptx64-nvidia-cuda"
430 ; Intrinsic to read X component of thread ID
431 declare i32 @llvm.nvvm.read.ptx.sreg.tid.x() readnone nounwind
433 define void @kernel(float addrspace(1)* %A,
434 float addrspace(1)* %B,
435 float addrspace(1)* %C) {
438 %id = tail call i32 @llvm.nvvm.read.ptx.sreg.tid.x() readnone nounwind
440 ; Compute pointers into A, B, and C
441 %ptrA = getelementptr float, float addrspace(1)* %A, i32 %id
442 %ptrB = getelementptr float, float addrspace(1)* %B, i32 %id
443 %ptrC = getelementptr float, float addrspace(1)* %C, i32 %id
446 %valA = load float, float addrspace(1)* %ptrA, align 4
447 %valB = load float, float addrspace(1)* %ptrB, align 4
450 %valC = fadd float %valA, %valB
453 store float %valC, float addrspace(1)* %ptrC, align 4
458 !nvvm.annotations = !{!0}
459 !0 = !{void (float addrspace(1)*,
461 float addrspace(1)*)* @kernel, !"kernel", i32 1}
464 We can use the LLVM ``llc`` tool to directly run the NVPTX code generator:
468 # llc -mcpu=sm_20 kernel.ll -o kernel.ptx
473 If you want to generate 32-bit code, change ``p:64:64:64`` to ``p:32:32:32``
474 in the module data layout string and use ``nvptx-nvidia-cuda`` as the
478 The output we get from ``llc`` (as of LLVM 3.4):
483 // Generated by LLVM NVPTX Back-End
492 .visible .entry kernel(
493 .param .u64 kernel_param_0,
494 .param .u64 kernel_param_1,
495 .param .u64 kernel_param_2
503 ld.param.u64 %rl1, [kernel_param_0];
505 mul.wide.s32 %rl2, %r1, 4;
506 add.s64 %rl3, %rl1, %rl2;
507 ld.param.u64 %rl4, [kernel_param_1];
508 add.s64 %rl5, %rl4, %rl2;
509 ld.param.u64 %rl6, [kernel_param_2];
510 add.s64 %rl7, %rl6, %rl2;
511 ld.global.f32 %f1, [%rl3];
512 ld.global.f32 %f2, [%rl5];
513 add.f32 %f3, %f1, %f2;
514 st.global.f32 [%rl7], %f3;
519 Dissecting the Kernel
520 ---------------------
522 Now let us dissect the LLVM IR that makes up this kernel.
527 The data layout string determines the size in bits of common data types, their
528 ABI alignment, and their storage size. For NVPTX, you should use one of the
535 target datalayout = "e-p:32:32:32-i1:8:8-i8:8:8-i16:16:16-i32:32:32-i64:64:64-f32:32:32-f64:64:64-v16:16:16-v32:32:32-v64:64:64-v128:128:128-n16:32:64"
541 target datalayout = "e-p:64:64:64-i1:8:8-i8:8:8-i16:16:16-i32:32:32-i64:64:64-f32:32:32-f64:64:64-v16:16:16-v32:32:32-v64:64:64-v128:128:128-n16:32:64"
547 In this example, we use the ``@llvm.nvvm.read.ptx.sreg.tid.x`` intrinsic to
548 read the X component of the current thread's ID, which corresponds to a read
549 of register ``%tid.x`` in PTX. The NVPTX back-end supports a large set of
550 intrinsics. A short list is shown below; please see
551 ``include/llvm/IR/IntrinsicsNVVM.td`` for the full list.
554 ================================================ ====================
555 Intrinsic CUDA Equivalent
556 ================================================ ====================
557 ``i32 @llvm.nvvm.read.ptx.sreg.tid.{x,y,z}`` threadIdx.{x,y,z}
558 ``i32 @llvm.nvvm.read.ptx.sreg.ctaid.{x,y,z}`` blockIdx.{x,y,z}
559 ``i32 @llvm.nvvm.read.ptx.sreg.ntid.{x,y,z}`` blockDim.{x,y,z}
560 ``i32 @llvm.nvvm.read.ptx.sreg.nctaid.{x,y,z}`` gridDim.{x,y,z}
561 ``void @llvm.nvvm.barrier0()`` __syncthreads()
562 ================================================ ====================
568 You may have noticed that all of the pointer types in the LLVM IR example had
569 an explicit address space specifier. What is address space 1? NVIDIA GPU
570 devices (generally) have four types of memory:
572 - Global: Large, off-chip memory
573 - Shared: Small, on-chip memory shared among all threads in a CTA
574 - Local: Per-thread, private memory
575 - Constant: Read-only memory shared across all threads
577 These different types of memory are represented in LLVM IR as address spaces.
578 There is also a fifth address space used by the NVPTX code generator that
579 corresponds to the "generic" address space. This address space can represent
580 addresses in any other address space (with a few exceptions). This allows
581 users to write IR functions that can load/store memory using the same
582 instructions. Intrinsics are provided to convert pointers between the generic
583 and non-generic address spaces.
585 See :ref:`address_spaces` and :ref:`nvptx_intrinsics` for more information.
591 In PTX, a function can be either a `kernel` function (callable from the host
592 program), or a `device` function (callable only from GPU code). You can think
593 of `kernel` functions as entry-points in the GPU program. To mark an LLVM IR
594 function as a `kernel` function, we make use of special LLVM metadata. The
595 NVPTX back-end will look for a named metadata node called
596 ``nvvm.annotations``. This named metadata must contain a list of metadata that
597 describe the IR. For our purposes, we need to declare a metadata node that
598 assigns the "kernel" attribute to the LLVM IR function that should be emitted
599 as a PTX `kernel` function. These metadata nodes take the form:
603 !{<function ref>, metadata !"kernel", i32 1}
605 For the previous example, we have:
609 !nvvm.annotations = !{!0}
610 !0 = !{void (float addrspace(1)*,
612 float addrspace(1)*)* @kernel, !"kernel", i32 1}
614 Here, we have a single metadata declaration in ``nvvm.annotations``. This
615 metadata annotates our ``@kernel`` function with the ``kernel`` attribute.
621 Generating PTX from LLVM IR is all well and good, but how do we execute it on
622 a real GPU device? The CUDA Driver API provides a convenient mechanism for
623 loading and JIT compiling PTX to a native GPU device, and launching a kernel.
624 The API is similar to OpenCL. A simple example showing how to load and
625 execute our vector addition code is shown below. Note that for brevity this
626 code does not perform much error checking!
630 You can also use the ``ptxas`` tool provided by the CUDA Toolkit to offline
631 compile PTX to machine code (SASS) for a specific GPU architecture. Such
632 binaries can be loaded by the CUDA Driver API in the same way as PTX. This
633 can be useful for reducing startup time by precompiling the PTX kernels.
644 void checkCudaErrors(CUresult err) {
645 assert(err == CUDA_SUCCESS);
648 /// main - Program entry point
649 int main(int argc, char **argv) {
657 // CUDA initialization
658 checkCudaErrors(cuInit(0));
659 checkCudaErrors(cuDeviceGetCount(&devCount));
660 checkCudaErrors(cuDeviceGet(&device, 0));
663 checkCudaErrors(cuDeviceGetName(name, 128, device));
664 std::cout << "Using CUDA Device [0]: " << name << "\n";
666 int devMajor, devMinor;
667 checkCudaErrors(cuDeviceComputeCapability(&devMajor, &devMinor, device));
668 std::cout << "Device Compute Capability: "
669 << devMajor << "." << devMinor << "\n";
671 std::cerr << "ERROR: Device 0 is not SM 2.0 or greater\n";
675 std::ifstream t("kernel.ptx");
677 std::cerr << "kernel.ptx not found\n";
680 std::string str((std::istreambuf_iterator<char>(t)),
681 std::istreambuf_iterator<char>());
683 // Create driver context
684 checkCudaErrors(cuCtxCreate(&context, 0, device));
686 // Create module for object
687 checkCudaErrors(cuModuleLoadDataEx(&cudaModule, str.c_str(), 0, 0, 0));
689 // Get kernel function
690 checkCudaErrors(cuModuleGetFunction(&function, cudaModule, "kernel"));
693 CUdeviceptr devBufferA;
694 CUdeviceptr devBufferB;
695 CUdeviceptr devBufferC;
697 checkCudaErrors(cuMemAlloc(&devBufferA, sizeof(float)*16));
698 checkCudaErrors(cuMemAlloc(&devBufferB, sizeof(float)*16));
699 checkCudaErrors(cuMemAlloc(&devBufferC, sizeof(float)*16));
701 float* hostA = new float[16];
702 float* hostB = new float[16];
703 float* hostC = new float[16];
706 for (unsigned i = 0; i != 16; ++i) {
708 hostB[i] = (float)(2*i);
712 checkCudaErrors(cuMemcpyHtoD(devBufferA, &hostA[0], sizeof(float)*16));
713 checkCudaErrors(cuMemcpyHtoD(devBufferB, &hostB[0], sizeof(float)*16));
716 unsigned blockSizeX = 16;
717 unsigned blockSizeY = 1;
718 unsigned blockSizeZ = 1;
719 unsigned gridSizeX = 1;
720 unsigned gridSizeY = 1;
721 unsigned gridSizeZ = 1;
724 void *KernelParams[] = { &devBufferA, &devBufferB, &devBufferC };
726 std::cout << "Launching kernel\n";
729 checkCudaErrors(cuLaunchKernel(function, gridSizeX, gridSizeY, gridSizeZ,
730 blockSizeX, blockSizeY, blockSizeZ,
731 0, NULL, KernelParams, NULL));
733 // Retrieve device data
734 checkCudaErrors(cuMemcpyDtoH(&hostC[0], devBufferC, sizeof(float)*16));
737 std::cout << "Results:\n";
738 for (unsigned i = 0; i != 16; ++i) {
739 std::cout << hostA[i] << " + " << hostB[i] << " = " << hostC[i] << "\n";
743 // Clean up after ourselves
749 checkCudaErrors(cuMemFree(devBufferA));
750 checkCudaErrors(cuMemFree(devBufferB));
751 checkCudaErrors(cuMemFree(devBufferC));
752 checkCudaErrors(cuModuleUnload(cudaModule));
753 checkCudaErrors(cuCtxDestroy(context));
759 You will need to link with the CUDA driver and specify the path to cuda.h.
763 # clang++ sample.cpp -o sample -O2 -g -I/usr/local/cuda-5.5/include -lcuda
765 We don't need to specify a path to ``libcuda.so`` since this is installed in a
766 system location by the driver, not the CUDA toolkit.
768 If everything goes as planned, you should see the following output when
769 running the compiled program:
773 Using CUDA Device [0]: GeForce GTX 680
774 Device Compute Capability: 3.0
796 You will likely see a different device identifier based on your hardware
799 Tutorial: Linking with Libdevice
800 ================================
802 In this tutorial, we show a simple example of linking LLVM IR with the
803 libdevice library. We will use the same kernel as the previous tutorial,
804 except that we will compute ``C = pow(A, B)`` instead of ``C = A + B``.
805 Libdevice provides an ``__nv_powf`` function that we will use.
809 target datalayout = "e-p:64:64:64-i1:8:8-i8:8:8-i16:16:16-i32:32:32-i64:64:64-f32:32:32-f64:64:64-v16:16:16-v32:32:32-v64:64:64-v128:128:128-n16:32:64"
810 target triple = "nvptx64-nvidia-cuda"
812 ; Intrinsic to read X component of thread ID
813 declare i32 @llvm.nvvm.read.ptx.sreg.tid.x() readnone nounwind
815 declare float @__nv_powf(float, float)
817 define void @kernel(float addrspace(1)* %A,
818 float addrspace(1)* %B,
819 float addrspace(1)* %C) {
822 %id = tail call i32 @llvm.nvvm.read.ptx.sreg.tid.x() readnone nounwind
824 ; Compute pointers into A, B, and C
825 %ptrA = getelementptr float, float addrspace(1)* %A, i32 %id
826 %ptrB = getelementptr float, float addrspace(1)* %B, i32 %id
827 %ptrC = getelementptr float, float addrspace(1)* %C, i32 %id
830 %valA = load float, float addrspace(1)* %ptrA, align 4
831 %valB = load float, float addrspace(1)* %ptrB, align 4
833 ; Compute C = pow(A, B)
834 %valC = call float @__nv_powf(float %valA, float %valB)
837 store float %valC, float addrspace(1)* %ptrC, align 4
842 !nvvm.annotations = !{!0}
843 !0 = !{void (float addrspace(1)*,
845 float addrspace(1)*)* @kernel, !"kernel", i32 1}
848 To compile this kernel, we perform the following steps:
850 1. Link with libdevice
851 2. Internalize all but the public kernel function
852 3. Run ``NVVMReflect`` and set ``__CUDA_FTZ`` to 0
853 4. Optimize the linked module
854 5. Codegen the module
857 These steps can be performed by the LLVM ``llvm-link``, ``opt``, and ``llc``
858 tools. In a complete compiler, these steps can also be performed entirely
859 programmatically by setting up an appropriate pass configuration (see
864 # llvm-link t2.bc libdevice.compute_20.10.bc -o t2.linked.bc
865 # opt -internalize -internalize-public-api-list=kernel -nvvm-reflect-list=__CUDA_FTZ=0 -nvvm-reflect -O3 t2.linked.bc -o t2.opt.bc
866 # llc -mcpu=sm_20 t2.opt.bc -o t2.ptx
870 The ``-nvvm-reflect-list=_CUDA_FTZ=0`` is not strictly required, as any
871 undefined variables will default to zero. It is shown here for evaluation
875 This gives us the following PTX (excerpt):
880 // Generated by LLVM NVPTX Back-End
889 .visible .entry kernel(
890 .param .u64 kernel_param_0,
891 .param .u64 kernel_param_1,
892 .param .u64 kernel_param_2
901 ld.param.u64 %rl2, [kernel_param_0];
903 ld.param.u64 %rl3, [kernel_param_1];
904 mul.wide.s32 %rl4, %r3, 4;
905 add.s64 %rl5, %rl2, %rl4;
906 ld.param.u64 %rl6, [kernel_param_2];
907 add.s64 %rl7, %rl3, %rl4;
908 add.s64 %rl1, %rl6, %rl4;
909 ld.global.f32 %f1, [%rl5];
910 ld.global.f32 %f2, [%rl7];
911 setp.eq.f32 %p1, %f1, 0f3F800000;
912 setp.eq.f32 %p2, %f2, 0f00000000;
913 or.pred %p3, %p1, %p2;
917 mov.f32 %f110, 0f3F800000;
918 st.global.f32 [%rl1], %f110;
920 BB0_2: // %__nv_isnanf.exit.i
922 setp.gtu.f32 %p4, %f4, 0f7F800000;
924 // %bb.3: // %__nv_isnanf.exit5.i
926 setp.le.f32 %p5, %f5, 0f7F800000;
928 BB0_4: // %.critedge1.i
929 add.f32 %f110, %f1, %f2;
930 st.global.f32 [%rl1], %f110;
932 BB0_5: // %__nv_isinff.exit.i
936 BB0_26: // %__nv_truncf.exit.i.i.i.i.i
937 mul.f32 %f90, %f107, 0f3FB8AA3B;
938 cvt.rzi.f32.f32 %f91, %f90;
939 mov.f32 %f92, 0fBF317200;
940 fma.rn.f32 %f93, %f91, %f92, %f107;
941 mov.f32 %f94, 0fB5BFBE8E;
942 fma.rn.f32 %f95, %f91, %f94, %f93;
943 mul.f32 %f89, %f95, 0f3FB8AA3B;
945 ex2.approx.ftz.f32 %f88,%f89;
947 add.f32 %f96, %f91, 0f00000000;
948 ex2.approx.f32 %f97, %f96;
949 mul.f32 %f98, %f88, %f97;
950 setp.lt.f32 %p15, %f107, 0fC2D20000;
951 selp.f32 %f99, 0f00000000, %f98, %p15;
952 setp.gt.f32 %p16, %f107, 0f42D20000;
953 selp.f32 %f110, 0f7F800000, %f99, %p16;
954 setp.eq.f32 %p17, %f110, 0f7F800000;
957 fma.rn.f32 %f110, %f110, %f108, %f110;
958 BB0_28: // %__internal_accurate_powf.exit.i
959 setp.lt.f32 %p18, %f1, 0f00000000;
960 setp.eq.f32 %p19, %f3, 0f3F800000;
961 and.pred %p20, %p18, %p19;
966 xor.b32 %r10, %r9, -2147483648;
968 BB0_30: // %__nv_powf.exit
969 st.global.f32 [%rl1], %f110;