The greedy rewriter is used in many different flows and it has a lot of
convenience (work list management, debugging actions, tracing, etc). But
it combines two kinds of greedy behavior 1) how ops are matched, 2)
folding wherever it can.
These are independent forms of greedy and leads to inefficiency. E.g.,
cases where one need to create different phases in lowering and is
required to applying patterns in specific order split across different
passes. Using the driver one ends up needlessly retrying folding/having
multiple rounds of folding attempts, where one final run would have
sufficed.
Of course folks can locally avoid this behavior by just building their
own, but this is also a common requested feature that folks keep on
working around locally in suboptimal ways.
For downstream users, there should be no behavioral change. Updating
from the deprecated should just be a find and replace (e.g., `find ./
-type f -exec sed -i
's|applyPatternsAndFoldGreedily|applyPatternsGreedily|g' {} \;` variety)
as the API arguments hasn't changed between the two.
The patch introduces the required changes to update the pass declarations and definitions to use the new autogenerated files and allow dropping the old infrastructure.
Reviewed By: mehdi_amini, rriddle
Differential Review: https://reviews.llvm.org/D132838
The patch introduces the required changes to update the pass declarations and definitions to use the new autogenerated files and allow dropping the old infrastructure.
Reviewed By: mehdi_amini, rriddle
Differential Review: https://reviews.llvm.org/D132838
This is correct for all values, i.e. the same as promoting the division to fp32 in the NVPTX backend. But it is faster (~10% in average, sometimes more) because:
- it performs less Newton iterations
- it avoids the slow path for e.g. denormals
- it allows reuse of the reciprocal for multiple divisions by the same divisor
Test program:
```
#include <stdio.h>
#include "cuda_fp16.h"
// This is a variant of CUDA's own __hdiv which is fast than hdiv_promote below
// and doesn't suffer from the perf cliff of div.rn.fp32 with 'special' values.
__device__ half hdiv_newton(half a, half b) {
float fa = __half2float(a);
float fb = __half2float(b);
float rcp;
asm("{rcp.approx.ftz.f32 %0, %1;\n}" : "=f"(rcp) : "f"(fb));
float result = fa * rcp;
auto exponent = reinterpret_cast<const unsigned&>(result) & 0x7f800000;
if (exponent != 0 && exponent != 0x7f800000) {
float err = __fmaf_rn(-fb, result, fa);
result = __fmaf_rn(rcp, err, result);
}
return __float2half(result);
}
// Surprisingly, this is faster than CUDA's own __hdiv.
__device__ half hdiv_promote(half a, half b) {
return __float2half(__half2float(a) / __half2float(b));
}
// This is an approximation that is accurate up to 1 ulp.
__device__ half hdiv_approx(half a, half b) {
float fa = __half2float(a);
float fb = __half2float(b);
float result;
asm("{div.approx.ftz.f32 %0, %1, %2;\n}" : "=f"(result) : "f"(fa), "f"(fb));
return __float2half(result);
}
__global__ void CheckCorrectness() {
int i = threadIdx.x + blockIdx.x * blockDim.x;
half x = reinterpret_cast<const half&>(i);
for (int j = 0; j < 65536; ++j) {
half y = reinterpret_cast<const half&>(j);
half d1 = hdiv_newton(x, y);
half d2 = hdiv_promote(x, y);
auto s1 = reinterpret_cast<const short&>(d1);
auto s2 = reinterpret_cast<const short&>(d2);
if (s1 != s2) {
printf("%f (%u) / %f (%u), got %f (%hu), expected: %f (%hu)\n",
__half2float(x), i, __half2float(y), j, __half2float(d1), s1,
__half2float(d2), s2);
//__trap();
}
}
}
__device__ half dst;
__global__ void ProfileBuiltin(half x) {
#pragma unroll 1
for (int i = 0; i < 10000000; ++i) {
x = x / x;
}
dst = x;
}
__global__ void ProfilePromote(half x) {
#pragma unroll 1
for (int i = 0; i < 10000000; ++i) {
x = hdiv_promote(x, x);
}
dst = x;
}
__global__ void ProfileNewton(half x) {
#pragma unroll 1
for (int i = 0; i < 10000000; ++i) {
x = hdiv_newton(x, x);
}
dst = x;
}
__global__ void ProfileApprox(half x) {
#pragma unroll 1
for (int i = 0; i < 10000000; ++i) {
x = hdiv_approx(x, x);
}
dst = x;
}
int main() {
CheckCorrectness<<<256, 256>>>();
half one = __float2half(1.0f);
ProfileBuiltin<<<1, 1>>>(one); // 1.001s
ProfilePromote<<<1, 1>>>(one); // 0.560s
ProfileNewton<<<1, 1>>>(one); // 0.508s
ProfileApprox<<<1, 1>>>(one); // 0.304s
auto status = cudaDeviceSynchronize();
printf("%s\n", cudaGetErrorString(status));
}
```
Reviewed By: herhut
Differential Revision: https://reviews.llvm.org/D126158
