-
Notifications
You must be signed in to change notification settings - Fork 2
Home
HSA PRM Specification revision: 1.0 Final (2015-03-06)
Test suite revision: 2015-03-27 (commit 0ecca4a05179f219c700069bbb18f7f78fc91730)
HSA PRM Conformance test suite is used to validate an implementation of Heterogeneous System Architecture Intermediate Language (HSAIL) virtual machine and language as described in "HSA Programmer's Reference Manual". The focus is on verification of ISA produced by HSAIL Finalizer and functionality of kernels on the agent. HSAIL Linking extension is used to finalize test programs while HSA Core Runtime is used to set up and invoke test kernels and to provide certain functionality for test code on the host (for some tests).
A test consists of scenario which describes test data, HSAIL program(s) in BRIG format, agent(s) code, dispatches and validation. The test is deemed passed if all scenario steps (including validation) are successful. Many basic tests for HSAIL instruction contain just one program with one kernel, one or several input and output buffers (used to store test data and test results), one dispatch and validation of output buffer. An example of such test is a test for add_u32 HSAIL instruction: one kernel, two input buffers and one output buffer.
Note that each test actually expands into several testcases, one for each valid combination of parameters. These parameters are described in Test details. For example, a test for the abs instruction consists of several tests for f32 and f64 types.
HSA PRM Conformance test suite uses CMake. The following dependencies need to be made available before the build:
HSAIL-Tools can be obtained from HSA github repository. HSAIL-Tools-PATH needs to be set to point to it. HSA Runtime and Finalizer extension includes. These can be obtained from HSA github repository as well. HSA-Runtime-Inc-PATH needs to be set to point to directory with hsa.h. HSA-Runtime-Ext-PATH needs to be set to point to directory with hsa_ext_finalize.h. cmake_windows.sh and cmake_linux.sh scripts contain examples of command lines that can be used to run CMake on Windows and Linux.
The test suite can be invoked as "hc" program that runs on the host.
To run the suite you should have a dynamic HSA RT libraries with corresponding bitness in the path (e.g. PATH variable for Windows, LD_LIBRARY_PATH for Linux).
A simple scenario of running all the tests on Linux 64-bit may look like:
export LD_LIBRARY_PATH="$HSA_RT_LIB:$LD_LIBRARY_PATH" bin/lnx64a/hc -tests prm/ -exclude amdhsa.exclude -runner hrunner -verbose -testlog results_linux64.log where $HSA_RT_LIB points to the HSA RT libraries, e.g. /compiler/dist/Obsidian/dist/linux/debug.hsa_foundation/lib/x86_64.
For convenience the package also contains Linux and Windows batch files, which take path to the HSA Runtime binaries and run the whole suite.
The following example demonstrates how to run these scripts on Linux. The results grouped by sub-suites will be displayed during the run along with overall statistics, per test results can be found in results_linux64.log:
$ ./run.sh /compiler/dist/Obsidian/dist/linux/debug.hsa_foundation/
prm/arithmetic/intfp/abs Passed: 56 Failed: 0 Error: 0 NA: 0 Total: 56 prm/arithmetic/intfp/add Passed: 255 Failed: 0 Error: 0 NA: 0 Total: 255 ... prm/special/misc/laneid Passed: 24 Failed: 0 Error: 0 NA: 0 Total: 24
Passed: 13729 Failed: 0 Error: 0 NA: 0 Total: 13729 Using -tests it is possible to specify a smaller set of tests, while -exclude allows to exclude tests (e.g. failing due to issues with the current HSA RT implementation).
The "hc" supports the following command line options:
-tests TestSet: prefix of test to run, e.g. -tests prm/ to run all the tests; -exclude File: file containing a list of tests to be excluded from testing; -verbose: enables detailed test output in a log file; -testlog File: name for a log file, the default name is test.log; -runner Runner: a mode of test grouping. May be either hrunner (default) or simple. By default tests are grouped by category. simple runner may be specified to avoid tests grouping. See option -testloglevel which also affects grouping. -testloglevel: test grouping depth, the default is 4. See also -runner option; -dump: dump HSAIL and BRIG test sources for each test under corresponding folder (prm/...); -results: path to folder which will contain dumped test sources (prm/...), the default is the current folder. Interpreting results
Every test has a unique identifier: test name. A test name consists of several words in lower case separated by slash (/). An example of test name is prm/special/dispatchpacket/dim/basic/kernel_1_200x1x1_64x1x1_ND. The following scheme for the words is used (starting from the beginning of test name):
special: Chapter in PRM specification dispatchpacket: Section in PRM specification dim: Instruction or group of instructions in PRM specification basic: Name of test/scenario kernel_1_200x1x1_64x1x1_ND: Name/identifier of testcase within test. For dim/basic test, it is the testcase for dim instruction in kernel body with dispatch dimension 1, grid sizes 200x1x1, workgroup sizes 64x1x1 and no control directives. The naming scheme for testcases depends on test. For many tests, it contains text representation of parameters described in Test details, separated by underscore (_). For some tests, simple counter is used. Refer to documentation below for the list of parameters.
A prefix of test names identifies test category. For example, prm/special/dispatchpacket/ are tests for dispatch packet operations. These categories may be used to select which tests to run (see option -tests).
All: kernel, function Kernel: just kernel Grid Geometry: geometry of a dispatch
###Consists of: Number of dimensions Grid sizes for x, y, z. Unused dimensions have value 1. Workgroup sizes for x, y, z. Unused dimensions have value 1. Used sets: All: various geometries, includes samples of partial workgroups/workitems, corner cases DimensionSet: representative geometry for each dimension OneGroupSet: one group, one dimension geometry DefaultPlusNGroupSet: set of geometry with 1 WorkItem, OneGroupSet, and 1 dim & N groups Boundary32Set: samples of geometry with sizes > 2^32 Boundary24Set: samples of geometry with sizes > 2^24 DegenerateSet: samples of degenerated geometry when a dimension is used, but has size 1 BarrierSet: samples of one-dimensional geometries for barrier testing FBarrierSet: special samples of geometries for fbarrier testing ImageSet: special samples of geometries for testing operations with images and samples ImageRdSet: special samples of geometries for testing rdimage operation Control Directives: control directives in BRIG modules
Location: location of control directives (function, kernel or module). List of enabled control directives. A test shall use all possible combinations of enabled control directives. For example, if two directives are enabled - requiredgridsize and requiredworkgroupsize, the test shall use the following combinations: [] (no control directives) [ requiredgridsize ] [ requiredworkgroupsize ] [ requiredgridsize, requiredworkgroupsize ] Used sets: DimensionSet: enabled directive is requireddim (directives affecting dimensions) GeometrySet: enabled directives are requireddim, requiredgridsize, requiredworkgroupsize (directives affecting grid geometry) GridSizeSet: enabled directives are requireddim, requiredgridsize (directives affecting grid size) DegenerateSet: enabled directives are requiredgridsize, requiredworkgroupsize (directives affecting computation of operations that can be simplified for degenerate set) Segment: memory segment
Atomic: flat, global, group HasFlatAddress: global, group, private (segments which may be accessed via flat address) MemFence: global, group Variable: global, readonly, kernarg, group, private, spill, arg All: all standard segments Linkage: variable linkage
All: program, module, linkage, arg Operand Kind: kind of instruction operand
All: register, immediate, WAVESIZE Dst Type: type of destination operand
All: all supported types Src Type: type of source operands
Used sets: All: all supported types Ftz: flush to zero modifier
All: all legal values Rounding: rounding modifier
All: all supported values Packed Controls: controls for processing packed operands
All: all supported values Compare Operation: operation used by cmp instruction
Used sets: All: all operations Atomic Operation: operation used by atomic instructions
Used sets: All: all operations Align: alignment
Used sets: All: all legal values Const: constant memory access
All: all legal values Equiv: equivalence class
All: all legal values Memory Order
All: rlx, screl, scacq, scar SignalAll: rlx, screl, scacq, scar SignalWait: rlx, screl, scacq MemFence: screl, scacq, scar Memory Scope
All: wi, wv, wg, cmp, sys Global: wv, wg, cmp, sys Group: wv, wg Width: width modifier
UpToWavesizeAndAll: 1, 2, 4, 8, 16, 32, WAVESIZE, All All: all legal values Test Data: data for testing arithmetic and memory operations
Standard: Includes the following values: regular values; boundary values (e.g. -128 and 127 for s8); special values (NaNs, infinities, subnormals, etc); values producing regular, boundary or special values as a result. Does not include the following values: values which signal exceptions; values which result in an undefined behavior. Exceptions Mask: a number that specifies a list of exceptions
All: all legal values Image Geometry: geometries associated with images [chapter 7.1.3]
All: all legal values Image Channel Order: channel orders associated with images [chapter 7.1.4.1]
All: all legal values Image Channel Type: channel types associated with images [chapter 7.1.4.2]
All: all legal values Image Array: image array size
All: 1, 2, 10 Image Property: names of image properties [chapter 7.5.2]
All: all legal values Sampler Coord: sampler coordinate normalization mode [chapter 7.1.8]
All: all legal values Sampler Filter: sampler filter mode [chapter 7.1.8]
All: all legal values Sampler Addressing: sampler addressing mode [chapter 7.1.8]
All: all legal values Sampler Property: names of sampler properties [chapter 7.5.2]
All: all legal values Test details
Segment: Variable Grid Geometry: OneGroupSet Code Location: Kernel arithmetic: Arithmetic operations [chapter 5]
intfp: Integer/floating point arithmetic operations [chapters 5.2 and 5.11]
Instructions: abs, add, borrow, carry, div, max, min, mul, mulhi, neg, rem, sub, ceil, floor, fma, fract, rint, sqrt, trunc Operand Kind: All Dst Type: All Ftz: All Rounding: All Packed controls: All Grid Geometry: OneGroupSet Code Location: Kernel Test Data: Standard intopt: Integer optimization operations [chapter 5.3]
Instructions: mad Operand Kind: All Dst Type: All Grid Geometry: OneGroupSet Code Location: Kernel Test Data: Standard 24int: 24-bit integer optimization operations [chapter 5.4]
Instructions: mad24, mad24hi, mul24, mul24hi Operand Kind: All Dst Type: All Grid Geometry: OneGroupSet Code Location: Kernel Test Data: Standard intshift: Integer shift operations [chapter 5.5]
Instructions: shl, shr Operand Kind: All Dst Type: All Grid Geometry: OneGroupSet Code Location: Kernel Test Data: Standard indbit: Individual bit operations [chapter 5.6]
Instructions: and, or, xor, not, popcount Operand Kind: All Dst Type: All Src Type: All Grid Geometry: OneGroupSet Code Location: Kernel Test Data: Standard bitstr: Bit string operations [chapter 5.7]
Instructions: bitextract, bitinsert, bitmask, bitrev, bitselect, firstbit, lastbit Operand Kind: All Dst Type: All Src Type: All Grid Geometry: OneGroupSet Code Location: Kernel Test Data: Standard copymove: Copy and move operations [chapter 5.8]
Instructions: combine, expand, mov Operand Kind: All Dst Type: All Src Type: All Grid Geometry: OneGroupSet Code Location: Kernel Test Data: Standard packed: Packet data operations [chapter 5.9]
Instructions: shuffle, unpacklo, unpackhi, pack, unpack Operand Kind: All Dst Type: All Src Type: All Grid Geometry: OneGroupSet Code Location: Kernel Test Data: Standard bitcmov: Bit conditional move operation [chapter 5.10]
Instructions: cmov Operand Kind: All Dst Type: All Grid Geometry: OneGroupSet Code Location: Kernel Test Data: Standard fpbit: Floating-point bit operations [chapter 5.12]
Instructions: class, copysign Operand Kind: All Dst Type: All Src Type: All Packed controls: All Grid Geometry: OneGroupSet Code Location: Kernel Test Data: Standard nativefp: Native floating-point operations [chapter 5.13]
Instructions: nsin, ncos, nlog2, nexp2, nsqrt, nrsqrt, nrcp, nfma Operand Kind: All Dst Type: All Grid Geometry: OneGroupSet Code Location: Kernel Test Data: Standard multimedia: Multimedia operations [chapter 5.14]
Instructions: bitalign, bytealign, lerp, packcvt, unpackcvt, sad, sadhi Operand Kind: All Dst Type: All Src Type: All Grid Geometry: OneGroupSet Code Location: Kernel Test Data: Standard compare: Compare operation [chapter 5.17]
Instructions: cmp Operand Kind: All Dst Type: All Src Type: All Compare Operation: All Ftz: All Grid Geometry: OneGroupSet Code Location: Kernel Test Data: Standard conversion: Conversion operation [chapter 5.18]
Instructions: cvt Operand Kind: All Dst Type: All Src Type: All Rounding: All Ftz: All Grid Geometry: OneGroupSet Code Location: Kernel Test Data: Standard address: Address operations [chapters 5 and 11]
null: Verify result of address operation for null address obtained with nullptr [chapter 11.4]
Instructions: stof, ftos, segmentp Segment: HasFlatAddress identity: Verify converted address accesses same location as address before conversion [chapter 5.16]
Instructions: stof, ftos Segment: HasFlatAddress segmentStore: use segment store/flat load and not vice versa nonull: use nonull in the instruction variable: Verify result of segmentp operation for flat address pointing into a variable [chapter 5.15]
Instructions: segmentp Segment: HasFlatAddress nonull: use nonull in the instruction lda/alignment: Verify result of lda operation is divisible by alignment [chapter 5.8]
Instructions: lda Segment: HasFlatAddress nonull: use nonull in the instruction memory: Memory operations [chapter 6]
memory/ordinary: Ordinary memory operations [chapters 6.3 and 6.4]
Instructions: ld, st Operand Kind: All Dst Type: All Segment: All Align: All Const: All Equiv: All Width: All Grid Geometry: OneGroupSet Code Location: Kernel Test Data: Standard memory/atomic: Atomic and atomicnoret memory operations [chapter 6.6 and 6.7]
Instructions: atomic, atomicnoret Operand Kind: All Dst Type: All Atomic operation: All (except for ld) Segment: Atomic Memory Order: All (except for rlx) Memory Scope: All (except for system) Equiv: 0 Grid Geometry: OneGroupSet Code Location: Kernel Test Data: Standard More tests are to be implemented.
memory/signal: Notification operations [chapter 6.8]
Instructions: signal, signalnoret Memory Order: SignalAll, SignalWait memfence: Memory fence operation [chapter 6.9]
Instructions: memfence Grid Geometry: OneGroupSet Segment: Memfence Memory Order: Memfence Memory Scope: Global, Group Image operations [chapter 7]
initializer/image: Image Creation and Image Handles [chapter 7.1.7]
Tests are to be implemented.
initializer/sampler: Sampler Creation and Sampler Handles [chapter 7.1.8]
Grid Geometry: OneGroupSet Code Location: All Sampler Coord: All Sampler Filter: All Sampler Addressing: All image_rd: Read Image Instruction [chapter 7.2]
Instructions: rdimage Grid Geometry: ImageRdSet Code Location: All Image Geometry: All Image Channel Order: All Image Channel Type: All Image Array: All Sampler Coord: All Sampler Filter: All Sampler Addressing: All Equiv: 0 image_ld: Load Image Instruction [chapter 7.3]
Instructions: ldimage Grid Geometry: ImageSet Code Location: All Image Geometry: All Image Channel Order: All Image Channel Type: All Image Array: All Equiv: 0 image_st: Store Image Instruction [chapter 7.4]
Instructions: stimage Grid Geometry: ImageSet Code Location: All Image Geometry: All Image Channel Order: All Image Channel Type: All Image Array: All Equiv: 0 image_query: Query Image Instruction [chapter 7.5]
Instructions: queryimage Grid Geometry: ImageSet Code Location: All Image Geometry: All Image Channel Order: All Image Channel Type: All Image Array: All Image Property: All image_query_sampler: Query Sampler Instruction [chapter 7.5]
Instructions: querysampler Grid Geometry: ImageSet Code Location: All Sampler Coord: All Sampler Filter: All Sampler Addressing: All Sampler Property: All imagefence: Image Fence Instruction [chapter 7.5]
Tests are to be implemented.
basic/br: Basic unconditional jump, verify expected result by setting value of HSAIL register [chapter 8]
Instructions: br Code Location: All basic: Basic conditional jump, switch conditional jump [chapter 8]
Instructions: cbr, sbr Grid Geometry: DefaultPlusNGroupSet Width: UpToWavesizeAndAll Operand Kind: All nested: Nested control flow (if-then-else) [chapter 8]
Instructions: cbr, br Grid Geometry: DefaultPlusNGroupSet Width: UpToWavesizeAndAll Operand Kind: All sand, sor: Short-circuit control flow [chapter 8]
Instructions: cbr Grid Geometry: DefaultPlusNGroupSet Width: UpToWavesizeAndAll Operand Kind: All Parallel Synchronization and communication operations [chapter 9]
barrier: Barrier operations [chapter 9.1]
Instructions: barrier Grid Geometry: BarrierSet Segment: Atomic Memory Scope: All Memory Order: All Tests for wavebarrier are to be implemented.
Fine-grain barrier operations [chapter 9.2]
Instructions: initfbar, joinfbar, waitfbar, arrivefbar, leavefbar, releasefbar, ldf Grid Geometry: FBarrierSet crosslane: Cross-lane operations [chapter 9.4]
Instructions: activelanecount, activelaneid, activelanemask Grid Geometry: OneGroupSet Code Location: All Tests for activelanepermute are to be implemented.
Function operations [chapter 10]
Direct call operation [chapter 10.6]
arguments: Verify passing argument/returning result of given type [chapter 10.2]
Types: all BRIG types More tests are to be implemented.
special: Special operations [chapter 11]
dispatchpacket: Dispatch packet operations [chapter 11.1]
basic: Verify result of dispatch packet operation
Instructions: currentworkgroupsize, dim, gridgroups, gridsize, workgroupid, workgroupsize, workitemabsid, workitemflatabsid, workitemflatid, workitemid Code Location: All Grid Geometry: DimensionSet for dim, All for others Control Directives: DimensionSet for dim, GeometrySet for others Type: u32/u64 for gridsize, workitemflatabsid, workitemflatid, u32 for others boundary32: Verify result of dispatch packet operation for workitems with workitemflatabsid around 2^32 boundary
Instructions: gridsize, workitemflatabsid, workitemflatid Code Location: All Grid Geometry: Boundary32Set Control Directives: GridSizeSet Type: u32/u64 boundary24: Verify result of dispatch packet operation for possible mul24 finalizer optimizations around 2^24 boundary
Instructions: workitemabsid, workitemflatabsid, workitemflatid Code Location: All Grid Geometry: Boundary24Set Control Directives: Boundary24Set Type: u32/u64 for workitemflatabsid, workitemflatid, u32 for other instructions degenerate: Verify result of dispatch packet operation for used dimension that dispatched with size 1
Instructions: currentworkgroupsize, gridgroups, gridsize, workgroupid, workgroupsize, workitemabsid, workitemflatid, workitemid Code Location: All Grid Geometry: DegenerateSet Control Directives: DegenerateSet Type: u32/u64 for workitemflatabsid, workitemflatid, u32 for other instructions packetid/basic: Compare result of packetid operation with value on the host
Instructions: packetid Code Location: All Grid Geometry: Boundary32Set Control Directives: None packetcompletionsig/basic: Compare result of packetcompletionsig operation with value on the host
Instructions: packetcompletionsig Code Location: All Grid Geometry: Boundary32Set Control Directives: None Exception operations [chapter 11.2]
Grid Geometry: OneGroupSet Code Location: All or Kernel Exceptions mask: All (where applicable) usermodequeue: User mode queue operations [chapter 11.3]
basic: Verify result of queueid, queueptr, ldk operation for a dispatch of a kernel
Instructions: queueid, queueptr, ldk Code Location: All basicindex: Verify result of queue index operation on a user mode queue
Instructions: ldqueuereadindex, ldqueuewriteindex, addqueuewriteindex, casqueuewriteindex, stqueuereadindex, stqueuewriteindex Code Location: All User mode queue type: separate, customized for each test misc: Miscellaneous operations [chapter 11.4]
kernargbaseptr/identity: Verify accessing memory at kernargbaseptr address
Code Location: All kernargbaseptr/alignment: Verify alignment of kernargbaseptr depending on alignment of kernel arguments
Code Location: All nop: Verify kernel with nop instruction
Code Location: All cuid/lessmax: Verify that result of cuid operation is always less than maxcuid
Code Locations: All Grid Geometry: All cuid/identity: Verify that result of cuid operation is same across workgroup
Code Locations: All Grid Geometry: All maxcuid/identity: Verify that result of maxcuid operation is same across grid
Code Locations: All Grid Geometry: All clock/monotonic: Verify that result of clock operation increases monotonically
Grid Geometry: OneGroupSet Code Location: All or Kernel waveid/lessmax: Verify that result of waveid operation is always less than maxwaveid
Code Locations: All Grid Geometry: All waveid/identity: Verify that result of waveid operation is same across wavefront
Code Locations: All Grid Geometry: All maxwaveid/identity: Verify that result of maxwaveid operation is same across grid
Code Locations: All Grid Geometry: All laneid/lessmax: Verify that result of laneid operation is always less than wavesize
Code Locations: All Grid Geometry: All laneid/sequence: Verify that result of laneid operation corresponds to work-items assigned to lanes in work-item flattened absolute ID order
Code Locations: All Grid Geometry: All Exceptions: [chapter 12]
Feature not implemented/not testable.
Grid Geometry: OneGroupSet Code Location: All or Kernel Exceptions mask: All (where applicable) Version: [chapter 14]
Tests to be implemented.
Grid Geometry: OneGroupSet Code Location: All Segment: Variable Linkage: All
Configuration/tests to be implemented.
Grid Geometry: All Code Location: Kernel Test suite internals
Information to be added when the suite is finalized.