#include "function_list.h"
#include "test_functions.h"
#include "utility.h"
#include <cstring>
static int BuildKernel(const char *name, int vectorSize, cl_kernel *k,
cl_program *p, bool relaxedMode)
{
const char *c[] = { "__kernel void math_kernel",
sizeNames[vectorSize],
"( __global float",
sizeNames[vectorSize],
"* out, __global float",
sizeNames[vectorSize],
"* in1, __global float",
sizeNames[vectorSize],
"* in2, __global float",
sizeNames[vectorSize],
"* in3 )\n"
"{\n"
" size_t i = get_global_id(0);\n"
" out[i] = ",
name,
"( in1[i], in2[i], in3[i] );\n"
"}\n" };
const char *c3[] = {
"__kernel void math_kernel",
sizeNames[vectorSize],
"( __global float* out, __global float* in, __global float* in2, "
"__global float* in3)\n"
"{\n"
" size_t i = get_global_id(0);\n"
" if( i + 1 < get_global_size(0) )\n"
" {\n"
" float3 f0 = vload3( 0, in + 3 * i );\n"
" float3 f1 = vload3( 0, in2 + 3 * i );\n"
" float3 f2 = vload3( 0, in3 + 3 * i );\n"
" f0 = ",
name,
"( f0, f1, f2 );\n"
" vstore3( f0, 0, out + 3*i );\n"
" }\n"
" else\n"
" {\n"
" size_t parity = i & 1; // Figure out how many elements are "
"left over after BUFFER_SIZE % (3*sizeof(float)). Assume power of two "
"buffer size \n"
" float3 f0;\n"
" float3 f1;\n"
" float3 f2;\n"
" switch( parity )\n"
" {\n"
" case 1:\n"
" f0 = (float3)( in[3*i], NAN, NAN ); \n"
" f1 = (float3)( in2[3*i], NAN, NAN ); \n"
" f2 = (float3)( in3[3*i], NAN, NAN ); \n"
" break;\n"
" case 0:\n"
" f0 = (float3)( in[3*i], in[3*i+1], NAN ); \n"
" f1 = (float3)( in2[3*i], in2[3*i+1], NAN ); \n"
" f2 = (float3)( in3[3*i], in3[3*i+1], NAN ); \n"
" break;\n"
" }\n"
" f0 = ",
name,
"( f0, f1, f2 );\n"
" switch( parity )\n"
" {\n"
" case 0:\n"
" out[3*i+1] = f0.y; \n"
" // fall through\n"
" case 1:\n"
" out[3*i] = f0.x; \n"
" break;\n"
" }\n"
" }\n"
"}\n"
};
const char **kern = c;
size_t kernSize = sizeof(c) / sizeof(c[0]);
if (sizeValues[vectorSize] == 3)
{
kern = c3;
kernSize = sizeof(c3) / sizeof(c3[0]);
}
char testName[32];
snprintf(testName, sizeof(testName) - 1, "math_kernel%s",
sizeNames[vectorSize]);
return MakeKernel(kern, (cl_uint)kernSize, testName, k, p, relaxedMode);
}
typedef struct BuildKernelInfo
{
cl_uint offset;
cl_kernel *kernels;
cl_program *programs;
const char *nameInCode;
bool relaxedMode;
} BuildKernelInfo;
static cl_int BuildKernelFn(cl_uint job_id, cl_uint thread_id UNUSED, void *p)
{
BuildKernelInfo *info = (BuildKernelInfo *)p;
cl_uint i = info->offset + job_id;
return BuildKernel(info->nameInCode, i, info->kernels + i,
info->programs + i, info->relaxedMode);
}
int TestFunc_mad_Float(const Func *f, MTdata d, bool relaxedMode)
{
uint64_t i;
uint32_t j, k;
int error;
logFunctionInfo(f->name, sizeof(cl_float), relaxedMode);
cl_program programs[VECTOR_SIZE_COUNT];
cl_kernel kernels[VECTOR_SIZE_COUNT];
float maxError = 0.0f;
float maxErrorVal = 0.0f;
float maxErrorVal2 = 0.0f;
float maxErrorVal3 = 0.0f;
size_t bufferSize = (gWimpyMode) ? gWimpyBufferSize : BUFFER_SIZE;
uint64_t step = getTestStep(sizeof(float), bufferSize);
{
BuildKernelInfo build_info = { gMinVectorSizeIndex, kernels, programs,
f->nameInCode, relaxedMode };
if ((error = ThreadPool_Do(BuildKernelFn,
gMaxVectorSizeIndex - gMinVectorSizeIndex,
&build_info)))
return error;
}
for (i = 0; i < (1ULL << 32); i += step)
{
cl_uint *p = (cl_uint *)gIn;
cl_uint *p2 = (cl_uint *)gIn2;
cl_uint *p3 = (cl_uint *)gIn3;
for (j = 0; j < bufferSize / sizeof(float); j++)
{
p[j] = genrand_int32(d);
p2[j] = genrand_int32(d);
p3[j] = genrand_int32(d);
}
if ((error = clEnqueueWriteBuffer(gQueue, gInBuffer, CL_FALSE, 0,
bufferSize, gIn, 0, NULL, NULL)))
{
vlog_error("\n*** Error %d in clEnqueueWriteBuffer ***\n", error);
return error;
}
if ((error = clEnqueueWriteBuffer(gQueue, gInBuffer2, CL_FALSE, 0,
bufferSize, gIn2, 0, NULL, NULL)))
{
vlog_error("\n*** Error %d in clEnqueueWriteBuffer2 ***\n", error);
return error;
}
if ((error = clEnqueueWriteBuffer(gQueue, gInBuffer3, CL_FALSE, 0,
bufferSize, gIn3, 0, NULL, NULL)))
{
vlog_error("\n*** Error %d in clEnqueueWriteBuffer3 ***\n", error);
return error;
}
for (j = gMinVectorSizeIndex; j < gMaxVectorSizeIndex; j++)
{
uint32_t pattern = 0xffffdead;
memset_pattern4(gOut[j], &pattern, bufferSize);
if ((error =
clEnqueueWriteBuffer(gQueue, gOutBuffer[j], CL_FALSE, 0,
bufferSize, gOut[j], 0, NULL, NULL)))
{
vlog_error("\n*** Error %d in clEnqueueWriteBuffer2(%d) ***\n",
error, j);
goto exit;
}
}
for (j = gMinVectorSizeIndex; j < gMaxVectorSizeIndex; j++)
{
size_t vectorSize = sizeof(cl_float) * sizeValues[j];
size_t localCount = (bufferSize + vectorSize - 1)
/ vectorSize;
if ((error = clSetKernelArg(kernels[j], 0, sizeof(gOutBuffer[j]),
&gOutBuffer[j])))
{
LogBuildError(programs[j]);
goto exit;
}
if ((error = clSetKernelArg(kernels[j], 1, sizeof(gInBuffer),
&gInBuffer)))
{
LogBuildError(programs[j]);
goto exit;
}
if ((error = clSetKernelArg(kernels[j], 2, sizeof(gInBuffer2),
&gInBuffer2)))
{
LogBuildError(programs[j]);
goto exit;
}
if ((error = clSetKernelArg(kernels[j], 3, sizeof(gInBuffer3),
&gInBuffer3)))
{
LogBuildError(programs[j]);
goto exit;
}
if ((error =
clEnqueueNDRangeKernel(gQueue, kernels[j], 1, NULL,
&localCount, NULL, 0, NULL, NULL)))
{
vlog_error("FAILED -- could not execute kernel\n");
goto exit;
}
}
if ((error = clFlush(gQueue))) vlog("clFlush failed\n");
float *r = (float *)gOut_Ref;
float *s = (float *)gIn;
float *s2 = (float *)gIn2;
float *s3 = (float *)gIn3;
for (j = 0; j < bufferSize / sizeof(float); j++)
r[j] = (float)f->func.f_fff(s[j], s2[j], s3[j]);
for (j = gMinVectorSizeIndex; j < gMaxVectorSizeIndex; j++)
{
if ((error =
clEnqueueReadBuffer(gQueue, gOutBuffer[j], CL_TRUE, 0,
bufferSize, gOut[j], 0, NULL, NULL)))
{
vlog_error("ReadArray failed %d\n", error);
goto exit;
}
}
if (gSkipCorrectnessTesting) break;
if (0 == (i & 0x0fffffff))
{
vlog(".");
fflush(stdout);
}
}
if (!gSkipCorrectnessTesting)
{
if (gWimpyMode)
vlog("Wimp pass");
else
vlog("passed");
vlog("\t%8.2f @ {%a, %a, %a}", maxError, maxErrorVal, maxErrorVal2,
maxErrorVal3);
}
vlog("\n");
exit:
for (k = gMinVectorSizeIndex; k < gMaxVectorSizeIndex; k++)
{
clReleaseKernel(kernels[k]);
clReleaseProgram(programs[k]);
}
return error;
}