[Beignet] [VERSION 2, PATCH 2/2] Tested nineteen built-in functions
Homer Hsing
homer.xing at intel.com
Sun May 12 23:58:17 PDT 2013
Tested hypot, sinpi, log1p, logb, tanpi, sinh, cosh, tanh,
asinh, acosh, atanh, remainder, rint, fdim, fract, native_divide,
pown, rootn, ldexp.
Signed-off-by: Homer Hsing <homer.xing at intel.com>
---
kernels/compiler_math.cl | 11 ++++++++
kernels/compiler_math_2op.cl | 17 ++++++++++++
kernels/compiler_math_3op.cl | 10 +++++++
utests/CMakeLists.txt | 2 ++
utests/compiler_math.cpp | 18 ++++++++++++-
utests/compiler_math_2op.cpp | 63 ++++++++++++++++++++++++++++++++++++++++++++
utests/compiler_math_3op.cpp | 62 +++++++++++++++++++++++++++++++++++++++++++
7 files changed, 182 insertions(+), 1 deletion(-)
create mode 100644 kernels/compiler_math_2op.cl
create mode 100644 kernels/compiler_math_3op.cl
create mode 100644 utests/compiler_math_2op.cpp
create mode 100644 utests/compiler_math_3op.cpp
diff --git a/kernels/compiler_math.cl b/kernels/compiler_math.cl
index 494cba1..f325595 100644
--- a/kernels/compiler_math.cl
+++ b/kernels/compiler_math.cl
@@ -15,6 +15,17 @@ __kernel void compiler_math(__global float *dst, __global float *src) {
case 10: dst[i] = exp2(x); break;
case 11: dst[i] = exp10(x); break;
case 12: dst[i] = expm1(x); break;
+ case 13: dst[i] = log1p(x); break;
+ case 14: dst[i] = logb(x); break;
+ case 15: dst[i] = sinpi(x); break;
+ case 16: dst[i] = tanpi(x); break;
+ case 17: dst[i] = rint(x); break;
+ case 18: dst[i] = sinh(x); break;
+ case 19: dst[i] = cosh(x); break;
+ case 20: dst[i] = tanh(x); break;
+ case 21: dst[i] = asinh(x); break;
+ case 22: dst[i] = acosh(x); break;
+ case 23: dst[i] = atanh(x); break;
default: dst[i] = 1.f; break;
};
}
diff --git a/kernels/compiler_math_2op.cl b/kernels/compiler_math_2op.cl
new file mode 100644
index 0000000..3a3eac3
--- /dev/null
+++ b/kernels/compiler_math_2op.cl
@@ -0,0 +1,17 @@
+kernel void compiler_math_2op(global float *dst, global float *src1, global float *src2) {
+ int i = get_global_id(0);
+ const float x = src1[i], y = src2[i];
+ float z;
+ switch (i) {
+ case 0: dst[i] = native_divide(x, y); break;
+ case 1: dst[i] = fdim(x, y); break;
+ case 2: dst[i] = fract(x, &z); break;
+ case 3: dst[i] = hypot(x, y); break;
+ case 4: dst[i] = ldexp(x, y); break;
+ case 5: dst[i] = pown(x, (int)y); break;
+ case 6: dst[i] = remainder(x, y); break;
+ case 7: dst[i] = rootn(x, (int)(y+1)); break;
+ default: dst[i] = 1.f; break;
+ };
+}
+
diff --git a/kernels/compiler_math_3op.cl b/kernels/compiler_math_3op.cl
new file mode 100644
index 0000000..e1d7c37
--- /dev/null
+++ b/kernels/compiler_math_3op.cl
@@ -0,0 +1,10 @@
+kernel void compiler_math_3op(global float *dst, global float *src1, global float *src2, global float *src3) {
+ int i = get_global_id(0);
+ const float x = src1[i], y = src2[i], z = src3[i];
+ switch (i) {
+ case 0: dst[i] = mad(x, y, z); break;
+ case 1: dst[i] = fma(x, y, z); break;
+ default: dst[i] = 1.f; break;
+ };
+}
+
diff --git a/utests/CMakeLists.txt b/utests/CMakeLists.txt
index 20d5456..d7eb959 100644
--- a/utests/CMakeLists.txt
+++ b/utests/CMakeLists.txt
@@ -65,6 +65,8 @@ set (utests_sources
compiler_write_only_shorts.cpp
compiler_switch.cpp
compiler_math.cpp
+ compiler_math_2op.cpp
+ compiler_math_3op.cpp
compiler_insn_selection_min.cpp
compiler_insn_selection_max.cpp
compiler_insn_selection_masked_min_max.cpp
diff --git a/utests/compiler_math.cpp b/utests/compiler_math.cpp
index 879f649..8799b8e 100644
--- a/utests/compiler_math.cpp
+++ b/utests/compiler_math.cpp
@@ -20,6 +20,17 @@ static void cpu_compiler_math(float *dst, float *src, int i)
case 10: dst[i] = powf(2, x); break;
case 11: dst[i] = powf(10, x); break;
case 12: dst[i] = expf(x) - 1; break;
+ case 13: dst[i] = logf(x + 1); break;
+ case 14: dst[i] = floorf(log2f(x)); break;
+ case 15: dst[i] = sinf(PI * x); break;
+ case 16: dst[i] = tanf(PI * x); break;
+ case 17: dst[i] = 2 * roundf(x / 2); break;
+ case 18: dst[i] = sinhf(x); break;
+ case 19: dst[i] = coshf(x); break;
+ case 20: dst[i] = tanhf(x); break;
+ case 21: dst[i] = asinhf(x); break;
+ case 22: dst[i] = acoshf(x); break;
+ case 23: dst[i] = atanhf(x); break;
default: dst[i] = 1.f; break;
};
}
@@ -51,7 +62,12 @@ static void compiler_math(void)
for (int i = 0; i < 16; ++i) {
const float cpu = cpu_dst[i];
const float gpu = ((float*)buf_data[0])[i];
- OCL_ASSERT(fabs(gpu-cpu)/std::max(fabs(cpu), fabs(gpu)) < 1e-4f);
+ if (isinf(cpu))
+ OCL_ASSERT(isinf(gpu));
+ else if (isnan(cpu))
+ OCL_ASSERT(isnan(gpu));
+ else
+ OCL_ASSERT(fabs(gpu-cpu) <= 1e-4f * std::max(fabs(cpu), fabs(gpu)));
}
OCL_UNMAP_BUFFER(0);
OCL_UNMAP_BUFFER(1);
diff --git a/utests/compiler_math_2op.cpp b/utests/compiler_math_2op.cpp
new file mode 100644
index 0000000..4a5c18c
--- /dev/null
+++ b/utests/compiler_math_2op.cpp
@@ -0,0 +1,63 @@
+#include "utest_helper.hpp"
+#include <cmath>
+#include <algorithm>
+
+static void cpu_compiler_math(float *dst, float *src1, float *src2, int i)
+{
+ const float x = src1[i], y = src2[i];
+ switch (i) {
+ case 0: dst[i] = x / y; break;
+ case 1: dst[i] = x > y ? x - y : 0; break;
+ case 2: dst[i] = fminf(x - floorf(x), 0x1.FFFFFep-1F); break;
+ case 3: dst[i] = sqrtf(x*x + y*y); break;
+ case 4: dst[i] = x * powf(2, (int)y); break;
+ case 5: dst[i] = powf(x, (int)y); break;
+ case 6: dst[i] = x - roundf(x/y)*y; break;
+ case 7: dst[i] = powf(x, 1.f/(int)(y+1)); break;
+ default: dst[i] = 1.f; break;
+ };
+}
+
+static void compiler_math_2op(void)
+{
+ const size_t n = 32;
+ float cpu_dst[32], cpu_src1[32], cpu_src2[32];
+
+ // Setup kernel and buffers
+ OCL_CREATE_KERNEL("compiler_math_2op");
+ OCL_CREATE_BUFFER(buf[0], 0, n * sizeof(float), NULL);
+ OCL_CREATE_BUFFER(buf[1], 0, n * sizeof(float), NULL);
+ OCL_CREATE_BUFFER(buf[2], 0, n * sizeof(float), NULL);
+ OCL_SET_ARG(0, sizeof(cl_mem), &buf[0]);
+ OCL_SET_ARG(1, sizeof(cl_mem), &buf[1]);
+ OCL_SET_ARG(2, sizeof(cl_mem), &buf[2]);
+ globals[0] = 16;
+ locals[0] = 16;
+
+ OCL_MAP_BUFFER(1);
+ OCL_MAP_BUFFER(2);
+ for (uint32_t i = 0; i < 32; ++i) {
+ cpu_src1[i] = ((float*)buf_data[1])[i] = .1f * (rand() & 15);
+ cpu_src2[i] = ((float*)buf_data[2])[i] = .1f * (rand() & 15);
+ }
+ OCL_UNMAP_BUFFER(1);
+ OCL_UNMAP_BUFFER(2);
+ OCL_NDRANGE(1);
+
+ for (int i = 0; i < 16; ++i)
+ cpu_compiler_math(cpu_dst, cpu_src1, cpu_src2, i);
+ OCL_MAP_BUFFER(0);
+ for (int i = 0; i < 16; ++i) {
+ const float cpu = cpu_dst[i];
+ const float gpu = ((float*)buf_data[0])[i];
+ if (isinf(cpu))
+ OCL_ASSERT(isinf(gpu));
+ else if (isnan(cpu))
+ OCL_ASSERT(isnan(gpu));
+ else
+ OCL_ASSERT(fabs(gpu-cpu) <= 1e-4f * std::max(fabs(cpu), fabs(gpu)));
+ }
+ OCL_UNMAP_BUFFER(0);
+}
+
+MAKE_UTEST_FROM_FUNCTION(compiler_math_2op)
diff --git a/utests/compiler_math_3op.cpp b/utests/compiler_math_3op.cpp
new file mode 100644
index 0000000..145b275
--- /dev/null
+++ b/utests/compiler_math_3op.cpp
@@ -0,0 +1,62 @@
+#include "utest_helper.hpp"
+#include <cmath>
+#include <algorithm>
+
+static void cpu_compiler_math(float *dst, float *src1, float *src2, float *src3, int i)
+{
+ const float x = src1[i], y = src2[i], z = src3[i];
+ switch (i) {
+ case 0: dst[i] = x * y + z; break;
+ case 1: dst[i] = x * y + z; break;
+ default: dst[i] = 1.f; break;
+ };
+}
+
+static void compiler_math_3op(void)
+{
+ const size_t n = 32;
+ float cpu_dst[32], cpu_src1[32], cpu_src2[32], cpu_src3[32];
+
+ // Setup kernel and buffers
+ OCL_CREATE_KERNEL("compiler_math_3op");
+ OCL_CREATE_BUFFER(buf[0], 0, n * sizeof(float), NULL);
+ OCL_CREATE_BUFFER(buf[1], 0, n * sizeof(float), NULL);
+ OCL_CREATE_BUFFER(buf[2], 0, n * sizeof(float), NULL);
+ OCL_CREATE_BUFFER(buf[3], 0, n * sizeof(float), NULL);
+ OCL_SET_ARG(0, sizeof(cl_mem), &buf[0]);
+ OCL_SET_ARG(1, sizeof(cl_mem), &buf[1]);
+ OCL_SET_ARG(2, sizeof(cl_mem), &buf[2]);
+ OCL_SET_ARG(3, sizeof(cl_mem), &buf[3]);
+ globals[0] = 16;
+ locals[0] = 16;
+
+ OCL_MAP_BUFFER(1);
+ OCL_MAP_BUFFER(2);
+ OCL_MAP_BUFFER(3);
+ for (uint32_t i = 0; i < 32; ++i) {
+ cpu_src1[i] = ((float*)buf_data[1])[i] = .1f * (rand() & 15);
+ cpu_src2[i] = ((float*)buf_data[2])[i] = .1f * (rand() & 15);
+ cpu_src3[i] = ((float*)buf_data[3])[i] = .1f * (rand() & 15);
+ }
+ OCL_UNMAP_BUFFER(1);
+ OCL_UNMAP_BUFFER(2);
+ OCL_UNMAP_BUFFER(3);
+ OCL_NDRANGE(1);
+
+ for (int i = 0; i < 16; ++i)
+ cpu_compiler_math(cpu_dst, cpu_src1, cpu_src2, cpu_src3, i);
+ OCL_MAP_BUFFER(0);
+ for (int i = 0; i < 16; ++i) {
+ const float cpu = cpu_dst[i];
+ const float gpu = ((float*)buf_data[0])[i];
+ if (isinf(cpu))
+ OCL_ASSERT(isinf(gpu));
+ else if (isnan(cpu))
+ OCL_ASSERT(isnan(gpu));
+ else
+ OCL_ASSERT(fabs(gpu-cpu) <= 1e-4f * std::max(fabs(cpu), fabs(gpu)));
+ }
+ OCL_UNMAP_BUFFER(0);
+}
+
+MAKE_UTEST_FROM_FUNCTION(compiler_math_3op)
--
1.8.1.2
More information about the Beignet
mailing list