Errata

This line contains two mistakes:

"bool bSSEInstructions = (CpuInfo[3] >> 24 && 0x1)"

First, the integer array is name CPUInfo and not CpuInfo.

Secondly, logical and operator "&&" with constant 0x1 (true) does not have any effect on the result of the expression. I guess it should be a bit-wise and operation instead.

Note from the Author or Editor:
Should read:

bool bSSEInstructions = (CPUInfo[3] >> 24 & 0x1);

Originally submitted on: http://blogs.msdn.com/b/nativeconcurrency/archive/2012/10/11/c-amp-book-now-available.aspx#10361456

Jo Blow
20 Oct 2012 8:34 PM

Just started reading the book. I'd like to point out that parallelization of delayed recurrence relationships is actually possible contrary to what it says on p10. One can partition the array into pieces. If some pieces depend other pieces it is still possible, in some cases, to parallel them. The example gives a[k] = a[k-1] + b[k]. We can assume that when k-1 is outside the partition it has a value of zero. This, then will throw off each value in the array by some constant. We can add it back in after the fact quite easily. Essentially it is a boundary value problem. The problem is that we have to potentially loop back over the entire array(possibly multiple times) and this may defeat the speed up in the first place. It will depend on the specific case. (there are other "tricks" that could potentially be used too... the point here, is only that it is possible).

Ade: This should read:

For example, this loop is not parallelizable in its current form:

I was puzzled by the "acc = r * s;" single CPU code on page 36 in this function:

void NBodySimpleInteractionEngine::BodyBodyInteraction(const ParticleCpu* const pParticlesIn,
ParticleCpu& particleOut, int numParticles) const
{
float_3 pos(particleOut.pos);
float_3 vel(particleOut.vel);
float_3 acc(0.0f);

std::for_each(pParticlesIn, pParticlesIn + numParticles, [=, &acc](const ParticleCpu& p)
{
const float_3 r = p.pos - pos;

float distSqr = SqrLength(r) + m_softeningSquared;
float invDist = 1.0f / sqrt(distSqr);
float invDistCube = invDist * invDist * invDist;
float s = m_particleMass * invDistCube;

acc = r * s;
});

vel += acc * m_deltaTime;
vel *= m_dampingFactor;
pos += vel * m_deltaTime;

particleOut.pos = pos;
particleOut.vel = vel;
}

because the final value of acc depended ONLY on the last call of the lambda. But the sum of all the accelerations caused by each point should be the final value of acc.

And in fact, if I look at the AMP version, I find the code I expected (acc += r * s;)

//--------------------------------------------------------------------------------------
// Calculate the acceleration (force * mass) change for a pair of particles.
//--------------------------------------------------------------------------------------

void BodyBodyInteraction(float_3& acc, const float_3 particlePosition,
const float_3 otherParticlePosition,
float softeningSquared, float particleMass) restrict(amp)
{
float_3 r = otherParticlePosition - particlePosition;

float distSqr = SqrLength(r) + softeningSquared;
float invDist = concurrency::fast_math::rsqrt(distSqr);
float invDistCube = invDist * invDist * invDist;

float s = particleMass * invDistCube;

acc += r * s;
}


So it looks like a minor bug to be fixed, and Amit agrees.
http://social.msdn.microsoft.com/Forums/en-US/parallelcppnative/thread/657296d8-0322-4a7e-b453-c6c12f4a5553

Note from the Author or Editor:
The line:

acc = r * s;

Should read:

acc += r * s;

"i += TS" and "k < TS" should almost certainly be
"i += TileSize" and "k < TileSize"

Note from the Author or Editor:
Should read:

for (int i = 0; i < W; i += TileSize)
{
tile_static float sA[TileSize][TileSize];
tile_static float sB[TileSize][TileSize];
sA[row][col] = a(tidx.global[0], col + i);
sB[row][col] = b(row + i, tidx.global[1]);
for (int k = 0; k < TileSize; k++)
sum += sA[row][k] * sB[k][col];
}

Currently the TDR feature is not supported correctly in the NVIDIA and AMD drivers. This is tracked in a issue on CodePlex.

http://ampbook.codeplex.com/workitem/33361

While the code and text in the book is correct it will not work correctly with the current drivers. No accelerator_view_removed is thrown.