Skip to content

stop rejected mutation steps early and fix the AVX-512 build #987

New issue

Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.

Sign up for GitHub

By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.

Already on GitHub? Sign in to your account

Open
MZoxx wants to merge 1 commit into
base: develop
Choose a base branch
from
Open

stop rejected mutation steps early and fix the AVX-512 build #987

Show file tree
Hide file tree
Changes from all commits
Commits
Show all changes
1 commit
Select commit
File filter

Filter by extension

Filter by extension
Conversations
Failed to load comments.
Loading
Jump to
Jump to file
Failed to load files.
Loading
Diff view
Diff view
2 changes: 1 addition & 1 deletion cmake/CompilerSetup.cmake
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -190,7 +190,7 @@ if(IS_MSVC)
endif()
elseif(IS_CLANG OR IS_GCC)
if(ENABLE_AVX512)
set(CPU_INSTRUCTION_FLAGS "-mavx -mavx2 -mavx512f -mavx512cd -mavx512vl -mavx512bw -mavx512dq" CACHE INTERNAL "CPU instruction set flags" FORCE)
set(CPU_INSTRUCTION_FLAGS "-mavx -mavx2 -mavx512f -mavx512cd -mavx512vl -mavx512bw -mavx512dq -mavx512vbmi" CACHE INTERNAL "CPU instruction set flags" FORCE)
message(STATUS "GCC/Clang: Enabling AVX-512 and AVX/AVX2")
else()
set(CPU_INSTRUCTION_FLAGS "-mavx -mavx2" CACHE INTERNAL "CPU instruction set flags" FORCE)
Expand Down
36 changes: 24 additions & 12 deletions src/mining/score_bpp9000.h
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -5,6 +5,10 @@

// bpp9000 scorer: recurrent ternary ANN (trits {0,1,2}, 2=UNKNOWN); only per-neuron LUTs mutate.
// Bit-exact with test/score_bpp9000_reference.h.
#if defined(__AVX512F__) && (defined(__AVX512VBMI__) || defined(_MSC_VER))
#define BPP9000_AVX512 1
#endif

namespace score_engine
{

Expand Down Expand Up @@ -99,7 +103,7 @@ struct ScoreBpp9000
// Stride of the interleaved wiring tuples {self, nbr0, nbr1, nbr2} (see neuronWiring below).
static constexpr unsigned int WIRING_STRIDE = 4;

#if defined(__AVX512F__)
#if defined(BPP9000_AVX512)
// AVX-512 window-batched stat
static constexpr unsigned long long SIMD_LANES = 64;
static constexpr unsigned long long SIMD_WINDOWS = 128;
Expand Down Expand Up @@ -145,7 +149,7 @@ struct ScoreBpp9000
unsigned long long updatedNeuronIndices[maxNumberOfNeurons];
unsigned long long numberOfUpdatedNeurons;

#if !defined(__AVX512F__) // AVX2 fallback (score_common.h requires AVX2 or AVX-512)
#if !defined(BPP9000_AVX512) // AVX2 fallback (score_common.h requires AVX2 or AVX-512)
// AVX2 fallback state: SIMD_LANES windows/batch, one trit per byte-lane, [neuron][lane] layout.
static constexpr unsigned long long SIMD_LANES = 32;
alignas(64) unsigned char simdCur[maxNumberOfNeurons * SIMD_LANES];
Expand Down Expand Up @@ -264,7 +268,7 @@ struct ScoreBpp9000
neuronWiring[k * WIRING_STRIDE + 3] = neighborIndices[n * numberOfNeighbors + 2];
}

#if defined(__AVX512F__)
#if defined(BPP9000_AVX512)
// Cone of influence: closure of {signal, output} under neighbour edges. Skipping updates outside
// the cone is bit-exact.
bool inCone[maxNumberOfNeurons];
Expand Down Expand Up @@ -404,13 +408,13 @@ struct ScoreBpp9000
}

// Sliding-window self-clocked score via the window-batched SIMD kernel (AVX-512 or AVX2, bit-exact).
unsigned int score()
unsigned int score(unsigned int failureCap = 0xFFFFFFFFu)
{
// PROFILE_NAMED_SCOPE("bpp9000:score");
return scoreSIMD();
return scoreSIMD(failureCap);
}

#if defined(__AVX512F__)
#if defined(BPP9000_AVX512)
// Ttry to minimal the L1-L3 cache as much as possible
inline void tickNeuron(const unsigned char* cur, unsigned char* nxt,
const unsigned char* lut, unsigned long long k)
Expand Down Expand Up @@ -455,7 +459,7 @@ struct ScoreBpp9000

// Window-batched score: SIMD_WINDOWS windows/batch, two per byte-lane (lo bits[0:3], hi bits[4:7]),
// halves independent. INFINITE_ERROR iff any window times out, else the failure count.
unsigned int scoreSIMD()
unsigned int scoreSIMD(unsigned int failureCap)
{
unsigned int numberOfFailures = 0;

Expand Down Expand Up @@ -831,13 +835,17 @@ struct ScoreBpp9000
numberOfFailures++;
}
}
if (numberOfFailures > failureCap)
{
return numberOfFailures;
}
}

return numberOfFailures;
}
#endif // defined(__AVX512F__)
#endif // defined(BPP9000_AVX512)

#if !defined(__AVX512F__)
#if !defined(BPP9000_AVX512)
void processTickSIMD(const unsigned char* cur, unsigned char* nxt)
{
// PROFILE_NAMED_SCOPE("bpp9000:processTick");
Expand Down Expand Up @@ -867,7 +875,7 @@ struct ScoreBpp9000
}

// AVX2 window-batched score
unsigned int scoreSIMD()
unsigned int scoreSIMD(unsigned int failureCap)
{
unsigned int numberOfFailures = 0;
const unsigned int sigIdx = signalNeuronIndex;
Expand Down Expand Up @@ -950,11 +958,15 @@ struct ScoreBpp9000
++numberOfFailures;
}
}
if (numberOfFailures > failureCap)
{
return numberOfFailures;
}
}

return numberOfFailures;
}
#endif // !defined(__AVX512F__)
#endif // !defined(BPP9000_AVX512)

// Flip one LUT entry of one non-input neuron (bit 0 picks the new trit, the rest picks the line).
void mutate(unsigned long long mutationSeed)
Expand Down Expand Up @@ -1034,7 +1046,7 @@ struct ScoreBpp9000
mutate(initValue.mutationSeed[s * MAX_LUT_ENTRIES_PER_STEP + i]);
}

unsigned int r = score();
const unsigned int r = (s < K) ? score() : score(cur);

bool accept = false;
if (s < K)
Expand Down
Loading