[HIPIFY][#1490][fix] Populate the list of supported device atomic functions

+ Update hipify-perl accordingly
+ Add atomics.cu test on all atomics


[ROCm/hip commit: b5f17e8ff0]

projects/hip/bin/hipify-perl

@@ -1646,13 +1646,13 @@ sub transformKernelLaunch {
 }
 
 sub transformHostFunctions {
-    my $m = 0;
+    my $k = 0;
     foreach $func (
         "hipMemcpyToSymbol",
         "hipMemcpyToSymbolAsync"
     )
     {
-        $m += s/(?<!\/\/ CHECK: )($func)\s*\(\s*([^,]+)\s*,/$func\(HIP_SYMBOL\($2\),/g
+        $k += s/(?<!\/\/ CHECK: )($func)\s*\(\s*([^,]+)\s*,/$func\(HIP_SYMBOL\($2\),/g
     }
     foreach $func (
         "hipGetSymbolAddress",
@@ -1661,25 +1661,25 @@ sub transformHostFunctions {
         "hipMemcpyFromSymbolAsync"
     )
     {
-        $m += s/(?<!\/\/ CHECK: )($func)\s*\(\s*([^,]+)\s*,\s*([^,\)]+)\s*(,\s*|\))\s*/$func\($2, HIP_SYMBOL\($3\)$4/g;
+        $k += s/(?<!\/\/ CHECK: )($func)\s*\(\s*([^,]+)\s*,\s*([^,\)]+)\s*(,\s*|\))\s*/$func\($2, HIP_SYMBOL\($3\)$4/g;
     }
     foreach $func (
         "hipFuncSetCacheConfig"
     )
     {
-        $m += s/(?<!\/\/ CHECK: )($func)\s*\(\s*([^,]+)\s*,/$func\(reinterpret_cast<const void*>\($2\),/g
+        $k += s/(?<!\/\/ CHECK: )($func)\s*\(\s*([^,]+)\s*,/$func\(reinterpret_cast<const void*>\($2\),/g
     }
     foreach $func (
         "hipFuncGetAttributes"
     )
     {
-        $m += s/(?<!\/\/ CHECK: )($func)\s*\(\s*([^,]+)\s*,\s*([^,\)]+)\s*(,\s*|\))\s*/$func\($2, reinterpret_cast<const void*>\($3\)$4/g;
+        $k += s/(?<!\/\/ CHECK: )($func)\s*\(\s*([^,]+)\s*,\s*([^,\)]+)\s*(,\s*|\))\s*/$func\($2, reinterpret_cast<const void*>\($3\)$4/g;
     }
-    return $m;
+    return $k;
 }
 
 sub countSupportedDeviceFunctions {
-    my $m = 0;
+    my $k = 0;
     foreach $func (
         "__brev",
         "__brevll",
@@ -1935,6 +1935,17 @@ sub countSupportedDeviceFunctions {
         "atanf",
         "atanh",
         "atanhf",
+        "atomicAdd",
+        "atomicAnd",
+        "atomicCAS",
+        "atomicDec",
+        "atomicExch",
+        "atomicInc",
+        "atomicMax",
+        "atomicMin",
+        "atomicOr",
+        "atomicSub",
+        "atomicXor",
         "cbrt",
         "cbrtf",
         "ceil",
@@ -2134,15 +2145,15 @@ sub countSupportedDeviceFunctions {
         my $mt_namespace = m/(\w+)::($func)\s*\(\s*.*\s*\)/g;
         my $mt = m/($func)\s*\(\s*.*\s*\)/g;
         if ($mt && !$mt_namespace) {
-            $m += $mt;
+            $k += $mt;
         }
     }
-    return $m;
+    return $k;
 }
 
 sub warnUnsupportedDeviceFunctions {
     my $line_num = shift;
-    my $m = 0;
+    my $k = 0;
     foreach $func (
         "_Pow_int",
         "__brkpt",
@@ -2250,7 +2261,6 @@ sub warnUnsupportedDeviceFunctions {
         "__vsubus4",
         "_fdsign",
         "_ldsign",
-        "atomicAdd",
         "float2int",
         "float_as_int",
         "float_as_uint",
@@ -2276,11 +2286,11 @@ sub warnUnsupportedDeviceFunctions {
         my $mt_namespace = m/(\w+)::($func)\s*\(\s*.*\s*\)/g;
         my $mt = m/($func)\s*\(\s*.*\s*\)/g;
         if ($mt && !$mt_namespace) {
-            $m += $mt;
+            $k += $mt;
             print STDERR "  warning: $fileName:$line_num: unsupported device function \"$func\": $_\n";
         }
     }
-    return $m;
+    return $k;
 }
 
 # Count of transforms in all files

projects/hip/hipify-clang/src/CUDA2HIP_Device_functions.cpp

@@ -599,9 +599,21 @@ const std::map<llvm::StringRef, hipCounter> CUDA_DEVICE_FUNC_MAP{
   {"h2exp10",           {"", "", CONV_DEVICE_FUNC, API_RUNTIME}},
   {"h2cos",             {"", "", CONV_DEVICE_FUNC, API_RUNTIME}},
   {"h2sin",             {"", "", CONV_DEVICE_FUNC, API_RUNTIME}},
-  {"__shfl_sync",     {"", "", CONV_DEVICE_FUNC, API_RUNTIME, UNSUPPORTED}},
+  {"__shfl_sync",       {"", "", CONV_DEVICE_FUNC, API_RUNTIME, UNSUPPORTED}},
-  {"__shfl_up_sync",  {"", "", CONV_DEVICE_FUNC, API_RUNTIME, UNSUPPORTED}},
+  {"__shfl_up_sync",    {"", "", CONV_DEVICE_FUNC, API_RUNTIME, UNSUPPORTED}},
-  {"__shfl_down_sync",{"", "", CONV_DEVICE_FUNC, API_RUNTIME, UNSUPPORTED}},
+  {"__shfl_down_sync",  {"", "", CONV_DEVICE_FUNC, API_RUNTIME, UNSUPPORTED}},
-  {"__shfl_xor_sync", {"", "", CONV_DEVICE_FUNC, API_RUNTIME, UNSUPPORTED}},
+  {"__shfl_xor_sync",   {"", "", CONV_DEVICE_FUNC, API_RUNTIME, UNSUPPORTED}},
-  {"atomicAdd",       {"", "", CONV_DEVICE_FUNC, API_RUNTIME, UNSUPPORTED}},
+  // atomic functions
+  {"atomicAdd",         {"", "", CONV_DEVICE_FUNC, API_RUNTIME}},
+  {"atomicSub",         {"", "", CONV_DEVICE_FUNC, API_RUNTIME}},
+  {"atomicExch",        {"", "", CONV_DEVICE_FUNC, API_RUNTIME}},
+  {"atomicMin",         {"", "", CONV_DEVICE_FUNC, API_RUNTIME}},
+  {"atomicMax",         {"", "", CONV_DEVICE_FUNC, API_RUNTIME}},
+  {"atomicInc",         {"", "", CONV_DEVICE_FUNC, API_RUNTIME}},
+  {"atomicDec",         {"", "", CONV_DEVICE_FUNC, API_RUNTIME}},
+  {"atomicDec",         {"", "", CONV_DEVICE_FUNC, API_RUNTIME}},
+  {"atomicAnd",         {"", "", CONV_DEVICE_FUNC, API_RUNTIME}},
+  {"atomicOr",          {"", "", CONV_DEVICE_FUNC, API_RUNTIME}},
+  {"atomicXor",         {"", "", CONV_DEVICE_FUNC, API_RUNTIME}},
+  {"atomicCAS",         {"", "", CONV_DEVICE_FUNC, API_RUNTIME}},
 };

projects/hip/tests/hipify-clang/unit_tests/device/atomics.cu

@@ -0,0 +1,284 @@
+// RUN: %run_test hipify "%s" "%t" %hipify_args %clang_args "-Xclang" "-fcuda-allow-variadic-functions"
+
+/*
+Copyright (c) 2015-present Advanced Micro Devices, Inc. All rights reserved.
+Permission is hereby granted, free of charge, to any person obtaining a copy
+of this software and associated documentation files (the "Software"), to deal
+in the Software without restriction, including without limitation the rights
+to use, copy, modify, merge, publish, distribute, sublicense, and/or sell
+copies of the Software, and to permit persons to whom the Software is
+furnished to do so, subject to the following conditions:
+The above copyright notice and this permission notice shall be included in
+all copies or substantial portions of the Software.
+THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
+IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
+FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT.  IN NO EVENT SHALL THE
+AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
+LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM,
+OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN
+THE SOFTWARE.
+*/
+
+// CHECK: #include "hip/hip_runtime.h"
+#include "cuda_runtime.h"
+
+#include <algorithm>
+#include <stdlib.h>
+#include <stdio.h>
+#include <string.h>
+#include <math.h>
+#include <type_traits>
+
+#define EXIT_WAIVED 2
+
+const char* sampleName = "hipSimpleAtomicsTest";
+
+using namespace std;
+// Auto-Verification Code
+bool testResult = true;
+
+bool computeGoldBitwise(...) {
+    return true;
+}
+
+template<typename T, typename enable_if<is_integral<T>{}>::type* = nullptr>
+bool computeGoldBitwise(T* gpuData, int len) {
+    T val = 0xff;
+    for (int i = 0; i < len; ++i) {
+        // 9th element should be 1
+        val &= (2 * i + 7);
+    }
+    if (val != gpuData[8]) {
+        printf("atomicAnd failed\n");
+        return false;
+    }
+
+    val = 0;
+    for (int i = 0; i < len; ++i) {
+        // 10th element should be 0xff
+        val |= (1 << i);
+    }
+    if (val != gpuData[9]) {
+        printf("atomicOr failed\n");
+        return false;
+    }
+
+    val = 0xff;
+    for (int i = 0; i < len; ++i) {
+        // 11th element should be 0xff
+        val ^= i;
+    }
+    if (val != gpuData[10]) {
+        printf("atomicXor failed\n");
+        return false;
+    }
+
+    return true;
+}
+
+template<typename T>
+bool computeGold(T* gpuData, int len) {
+    T val = 0;
+    for (int i = 0; i < len; ++i) {
+        val += 10;
+    }
+    if (val != gpuData[0]) {
+        printf("atomicAdd failed\n");
+        return false;
+    }
+
+    val = 0;
+    for (int i = 0; i < len; ++i) {
+        val -= 10;
+    }
+    if (val != gpuData[1]) {
+        printf("atomicSub failed\n");
+        return false;
+    }
+
+    bool found = false;
+    for (T i = 0; i < len; ++i) {
+        // third element should be a member of [0, len)
+        if (i == gpuData[2]) {
+            found = true;
+            break;
+        }
+    }
+    if (!found) {
+        printf("atomicExch failed\n");
+        return false;
+    }
+
+    val = -(1 << 8);
+    for (T i = 0; i < len; ++i) {
+        // fourth element should be len-1
+        val = max(val, i);
+    }
+    if (val != gpuData[3]) {
+        printf("atomicMax failed\n");
+        return false;
+    }
+
+    val = 1 << 8;
+    for (T i = 0; i < len; ++i) {
+        val = min(val, i);
+    }
+    if (val != gpuData[4]) {
+        printf("atomicMin failed\n");
+        return false;
+    }
+
+    int limit = 17;
+    val = 0;
+    for (int i = 0; i < len; ++i) {
+        val = (val >= limit) ? 0 : val + 1;
+    }
+    if (val != gpuData[5]) {
+        printf("atomicInc failed\n");
+        return false;
+    }
+
+    limit = 137;
+    val = 0;
+    for (int i = 0; i < len; ++i) {
+        val = ((val == 0) || (val > limit)) ? limit : val - 1;
+    }
+    if (val != gpuData[6]) {
+        printf("atomicDec failed\n");
+        return false;
+    }
+
+    found = false;
+    for (T i = 0; i < len; ++i) {
+        // eighth element should be a member of [0, len)
+        if (i == gpuData[7]) {
+            found = true;
+            break;
+        }
+    }
+    if (!found) {
+        printf("atomicCAS failed\n");
+        return false;
+    }
+
+    return computeGoldBitwise(gpuData, len);
+}
+
+__device__
+void testKernelExch(...) {}
+
+template<typename T, typename enable_if<!is_same<T, double>{}>::type* = nullptr>
+__device__
+void testKernelExch(T* g_odata) {
+    // access thread id
+    const T tid = blockDim.x * blockIdx.x + threadIdx.x;
+    // Atomic exchange
+    atomicExch(&g_odata[2], tid);
+}
+
+__device__
+void testKernelSub(...) {}
+
+template<
+    typename T,
+    typename enable_if<
+        is_same<T, int>{} || is_same<T, unsigned int>{}>::type* = nullptr>
+__device__
+void testKernelSub(T* g_odata) {
+    // Atomic subtraction (final should be 0)
+    atomicSub(&g_odata[1], 10);
+}
+
+__device__
+void testKernelIntegral(...) {}
+
+template<typename T, typename enable_if<is_integral<T>{}>::type* = nullptr>
+__device__
+void testKernelIntegral(T* g_odata) {
+    // access thread id
+    const T tid = blockDim.x * blockIdx.x + threadIdx.x;
+    // Atomic maximum
+    atomicMax(&g_odata[3], tid);
+    // Atomic minimum
+    atomicMin(&g_odata[4], tid);
+    // Atomic increment (modulo 17+1)
+    atomicInc((unsigned int*)&g_odata[5], 17);
+    // Atomic decrement
+    atomicDec((unsigned int*)&g_odata[6], 137);
+    // Atomic compare-and-swap
+    atomicCAS(&g_odata[7], tid - 1, tid);
+    // Atomic AND
+    atomicAnd(&g_odata[8], 2 * tid + 7);
+    // Atomic OR
+    atomicOr(&g_odata[9], 1 << tid);
+    // Atomic XOR
+    atomicXor(&g_odata[10], tid);
+    testKernelSub(g_odata);
+}
+
+template<typename T>
+__global__ void testKernel(T* g_odata) {
+    // Atomic addition
+    atomicAdd(&g_odata[0], 10);
+    testKernelIntegral(g_odata);
+    testKernelExch(g_odata);
+}
+
+template<typename T>
+void runTest() {
+    // CHECK: hipDeviceProp_t deviceProp;
+    cudaDeviceProp deviceProp;
+    deviceProp.major = 0;
+    deviceProp.minor = 0;
+    int dev = 0;
+    // CHECK: hipGetDeviceProperties(&deviceProp, dev);
+    cudaGetDeviceProperties(&deviceProp, dev);
+    // Statistics about the GPU device
+    printf(
+        "> GPU device has %d Multi-Processors, "
+        "SM %d.%d compute capabilities\n\n",
+        deviceProp.multiProcessorCount, deviceProp.major, deviceProp.minor);
+    unsigned int numThreads = 256;
+    unsigned int numBlocks = 64;
+    unsigned int numData = 11;
+    unsigned int memSize = sizeof(T) * numData;
+
+    // Allocate mem for the result on host side
+    T* hOData = (T*)malloc(memSize);
+    // Initialize the memory
+    for (unsigned int i = 0; i < numData; i++) hOData[i] = 0;
+    // To make the AND and XOR tests generate something other than 0...
+    hOData[8] = hOData[10] = 0xff;
+    // Allocate device memory for result
+    T* dOData;
+    // CHECK: hipMalloc((void**)&dOData, memSize);
+    cudaMalloc((void**)&dOData, memSize);
+    // Copy host memory to device to initialize to zero
+    // CHECK: hipMemcpy(dOData, hOData, memSize, hipMemcpyHostToDevice);
+    cudaMemcpy(dOData, hOData, memSize, cudaMemcpyHostToDevice);
+    // Execute the kernel
+    // CHECK: hipLaunchKernelGGL((testKernel), dim3(numBlocks), dim3(numThreads), 0, 0, dOData);
+    testKernel<<<numBlocks, numThreads>>>(dOData);
+    // Copy result from device to host
+    // CHECK: hipMemcpy(hOData, dOData, memSize, hipMemcpyDeviceToHost);
+    cudaMemcpy(hOData, dOData, memSize, cudaMemcpyDeviceToHost);
+    // Compute reference solution
+    testResult = computeGold(hOData, numThreads * numBlocks);
+    // Cleanup memory
+    free(hOData);
+    // CHECK: hipFree(dOData);
+    cudaFree(dOData);
+}
+
+int main(int argc, char** argv) {
+    printf("%s starting...\n", sampleName);
+    runTest<int>();
+    runTest<unsigned int>();
+    runTest<unsigned long long>();
+    runTest<float>();
+    runTest<double>();
+    // CHECK: hipDeviceReset();
+    cudaDeviceReset();
+    printf("%s completed, returned %s\n", sampleName, testResult ? "OK" : "ERROR!");
+    exit(testResult ? EXIT_SUCCESS : EXIT_FAILURE);
+}