From cbd73dde3dd2da790bb663385a229ce22965c43c Mon Sep 17 00:00:00 2001 From: Julius Ikkala Date: Thu, 21 Aug 2025 01:13:52 +0300 Subject: Fix nextafter() (#8195) Fixes #8185. The previous implementation is incorrect and basically only works in the `x = 0` case. `delta` was the smallest possible positive value representable as a float, but that's below the rounding error of addition with almost all reasonably sized floats. This fixed implementation is based on bit twiddling instead. I've checked the float case against the C++ `nextafterf` with both a -inf -> inf and inf -> -inf sweep, in addition to the test included in this PR. --- tests/bugs/gh-8185.slang | 61 ++++++++++++++++++++++++++++++++++++++++++++++++ 1 file changed, 61 insertions(+) create mode 100644 tests/bugs/gh-8185.slang (limited to 'tests/bugs') diff --git a/tests/bugs/gh-8185.slang b/tests/bugs/gh-8185.slang new file mode 100644 index 000000000..7a9a32b51 --- /dev/null +++ b/tests/bugs/gh-8185.slang @@ -0,0 +1,61 @@ +//TEST(compute):COMPARE_COMPUTE(filecheck-buffer=CHECK): -cpu -compute -Xslang -DSKIP_HALF_PRECISION +//TEST(compute):COMPARE_COMPUTE(filecheck-buffer=CHECK): -dx12 -profile cs_6_2 -use-dxil -compute +//TEST(compute, vulkan):COMPARE_COMPUTE(filecheck-buffer=CHECK): -vk -emit-spirv-directly -compute +//TEST(compute):COMPARE_COMPUTE(filecheck-buffer=CHECK): -cuda -compute -Xslang -DSKIP_HALF_PRECISION +//TEST_DISABLED(compute):COMPARE_COMPUTE(filecheck-buffer=CHECK): -wgpu -compute + +//TEST_INPUT: ubuffer(data=[0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0], stride=4):out,name=outputBuffer +RWStructuredBuffer outputBuffer; + +void write_double(double f, int i) +{ + uint64_t v = bit_cast(f); + outputBuffer[i] = uint32_t(v); + outputBuffer[i+1] = uint32_t(v >> 32); +} + +[numthreads(1, 1, 1)] +void computeMain(int3 dispatchThreadID: SV_DispatchThreadID) +{ + outputBuffer[0] = bit_cast(nextafter(-0.0f, 1.0f)); // CHECK: 1 + outputBuffer[1] = bit_cast(nextafter(0.0f, 1.0f)); // CHECK-NEXT: 1 + outputBuffer[2] = bit_cast(nextafter(-0.0f, -1.0f)); // CHECK-NEXT: 80000001 + outputBuffer[3] = bit_cast(nextafter(0.0f, -1.0f)); // CHECK-NEXT: 80000001 + outputBuffer[4] = bit_cast(nextafter(1000.0f, 2000.0f)); // CHECK-NEXT: 447A0001 + outputBuffer[5] = bit_cast(nextafter(-1000.0f, -1.0f)); // CHECK-NEXT: C479FFFF + +#ifdef SKIP_HALF_PRECISION + outputBuffer[6] = 1; + outputBuffer[7] = 1; + outputBuffer[8] = 0x8001; + outputBuffer[9] = 0x8001; + outputBuffer[10] = 0x63D1; + outputBuffer[11] = 0xE3CF; +#else + outputBuffer[6] = bit_cast(nextafter(-0.0h, 1.0h)); // CHECK-NEXT: 1 + outputBuffer[7] = bit_cast(nextafter(0.0h, 1.0h)); // CHECK-NEXT: 1 + outputBuffer[8] = bit_cast(nextafter(-0.0h, -1.0h)); // CHECK-NEXT: 8001 + outputBuffer[9] = bit_cast(nextafter(0.0h, -1.0h)); // CHECK-NEXT: 8001 + outputBuffer[10] = bit_cast(nextafter(1000.0h, 2000.0h)); // CHECK-NEXT: 63D1 + outputBuffer[11] = bit_cast(nextafter(-1000.0h, -1.0h)); // CHECK-NEXT: E3CF +#endif + + // CHECK-NEXT: 1 + // CHECK-NEXT: 0 + write_double(nextafter(-0.0l, 1.0l), 12); + // CHECK-NEXT: 1 + // CHECK-NEXT: 0 + write_double(nextafter(0.0l, 1.0l), 14); + // CHECK-NEXT: 1 + // CHECK-NEXT: 80000000 + write_double(nextafter(-0.0l, -1.0l), 16); + // CHECK-NEXT: 1 + // CHECK-NEXT: 80000000 + write_double(nextafter(0.0l, -1.0l), 18); + // CHECK-NEXT: 1 + // CHECK-NEXT: 408F4000 + write_double(nextafter(1000.0l, 2000.0l), 20); + // CHECK-NEXT: FFFFFFFF + // CHECK-NEXT: C08F3FFF + write_double(nextafter(-1000.0l, -1.0l), 22); +} -- cgit v1.2.3