borisbat

MoltenVK argument-buffers compute store-loss repro — SPIR-V kernels

Nine compute kernels from the dasLLAMA test suite (daslang-generated SPIR-V, disassembled with spirv-dis; spirv-as rebuilds the binaries). All nine pass spirv-val (default and --target-env vulkan1.2). On Apple M1 Max, macOS 26.5.2, MoltenVK 1.4.2 (brew), with Metal argument buffers enabled (MVK_CONFIG_USE_METAL_ARGUMENT_BUFFERS=1, the Apple Silicon default) the seven kernels marked FAIL dispatch without any error but none of their device stores land; with MVK_CONFIG_USE_METAL_ARGUMENT_BUFFERS=0 all nine produce correct results. Fences signal and buffer copies in the same submission run in both cases.

borisbat

    // this will split the work between all available threads
    // if chunk_count is -1, default number of chunks will be computed
    void JobQue::parallel_for ( int from, int to, const JobChunk & chunk, int chunk_count, int step )
    {
        if ( from >= to )
            return;
        // so the idea here is that if say 1 thread is busy
        // we don't pay 2x time. smaller chunks are more evenly distributed
        if ( chunk_count==-1 )
            chunk_count = mThreadCount * 4;

borisbat

mat3 cotangent_frame( vec3 N, vec3 p, vec2 uv )
{
    // get edge vectors of the pixel triangle
    vec3 dp1 = dFdx( p );
    vec3 dp2 = dFdy( p );
    vec2 duv1 = dFdx( uv );
    vec2 duv2 = dFdy( uv );

    // solve the linear system
    vec3 dp2perp = cross( dp2, N );

borisbat

       ///////////////////////////////////////////
        // parse library_visual_scenes (nodes) here
        
        void importDaeNodeArray ( const pugi::xpath_node_set & nodes, vector<NodePtr> & outNodes, const WeakNodePtr & parent )
        {
			using namespace pugi;
            if ( nodes.size()==0 )
                return;
            outNodes.resize(nodes.size());
            for ( unsigned int ni=0; ni<nodes.size(); ni++ )

borisbat

  vec3 finalColor = vec3( 0.0 );
	float steps = 0.0;
	
	int initialSteps = max ( 4, g_Steps / 3 );
	int maxEstSteps = 32;
	
	float kFactor = 2.0 / g_Steps;
	
	// do initial steps
	for ( int i=0; i<initialSteps; ++i )

borisbat

  vec3 finalColor = vec3( 0.0 );
	float steps = 0.0;
	
	int initialSteps = max ( 4, g_Steps / 3 );
	int maxEstSteps = 32;
	float factor = max ( 1.0, length(g_light_color.rgb)/length(vec3(1.0)) );
	float tolerance = 1.0 / ( (g_Steps * 4.0 + initialSteps) * factor );
	
	// do initial steps
	for ( int i=0; i<initialSteps; ++i )

borisbat

#include <iostream>
#include <vector>
#include <functional>
#include <initializer_list>

#define USE_EMPLACE     1
#define USE_UREF        1

using namespace std;

borisbat

class Report
{
public:
    Report() : i_am_dead(false) { cout << "  Report();" << endl; }
    Report(const Report & ) { cout << "  Report(const Report&);" << endl; };
    Report(Report && h) { cout << "  Report(&&);" << endl; h.i_am_dead = true; }
    Report & operator = (const Report&) { cout << "  Report::=(const Report&);" << endl; return *this; };
    ~Report() { cout << "  ~Report() i_am_dead=" << i_am_dead << endl; }
    bool i_am_dead;
};

borisbat

class Report
{
public:
    Report() { cout << "Report();" << endl; }
    Report(const Report & ) { cout << "Report(const & Report);" << endl; };
    Report(Report &&) { cout << "Report(&&);" << endl; }
    Report & operator = (const Report &) { cout << "Report::=(const Report &);" << endl; return *this; };
};

template <typename Tc, typename T>

borisbat

#include <iostream>
#include <vector>
#include <functional>
#include <initializer_list>
#include <string>

using namespace std;

template <typename TT>
class RArray : protected vector<TT>