Mja $dZddlZdZddZdZy)u Replacement for `_tiled_matmul_src` and `_TiledKernel` in `vgpu_tiled.py`. Bug: shader exposed `M_dim`/`N_dim`/`K_dim` as `uniform int` but the host never assigned them, so they defaulted to 0, every bound check failed, the inner loop ran zero iterations, and C was never written — leaving uninitialised float32 garbage in the readback. Fix: bake M, K, N into the source as #define so the compiler sees constants and the runtime-undefined-uniform trap is gone. Also zero the output buffer once at allocation as a belt-and-braces defence. Nc|d}|d}|d}|d}|d}|d} |d} ||z| | zz} ||z| | zz} d|d |d |d |d |d | d| d| d| d|d|d|dS)NTS_KTSMTSNWPTMWPTNRTSMRTSNuT#version 430 core // Tile geometry for this kernel — compile-time. #define TS_K z #define TSM z #define TSN z #define WPTM z #define WPTN z #define RTSM z #define RTSN z #define LPTA z #define LPTB u // Shader-shape constants — BAKE, do not use uniforms. // Prevents the "uniform defaults to 0" trap that you just hit. #define MDIM z #define NDIM z #define KDIM u layout(local_size_x = RTSM, local_size_y = RTSN) in; layout(std430, binding = 0) readonly buffer A_buf { float A[]; }; layout(std430, binding = 1) readonly buffer B_buf { float B[]; }; layout(std430, binding = 2) writeonly buffer C_buf { float C[]; }; // A is stored transposed so the inner-loop read pattern // `Asub[k][tidm + wm*RTSM]` is bank-coalesced on Intel/AMD after the // first tile. shared float Asub[TS_K][TSM]; // The +2 column pad prevents 32-way bank conflicts on Intel. shared float Bsub[TSN][TS_K + 2]; void main() { uint tidm = gl_LocalInvocationID.x; // 0..RTSM-1 uint tidn = gl_LocalInvocationID.y; // 0..RTSN-1 uint offM = TSM * gl_WorkGroupID.x; uint offN = TSN * gl_WorkGroupID.y; float acc[WPTM][WPTN]; for (uint i = 0u; i < WPTM; ++i) for (uint j = 0u; j < WPTN; ++j) acc[i][j] = 0.0; uint nTiles = (KDIM + TS_K - 1u) / TS_K; for (uint t = 0u; t < nTiles; ++t) { uint kBase = t * TS_K; // ─── Load A tile into shared (transposed) ───────────────────── for (uint la = 0u; la < LPTA; ++la) { uint lin = la * (RTSM * RTSN) + tidn * RTSM + tidm; uint row = lin % TSM; uint col = lin / TSM; uint gRow = offM + row; uint gK = kBase + col; Asub[col][row] = (gRow < MDIM && gK < KDIM) ? A[gRow * KDIM + gK] : 0.0; } // ─── Load B tile into shared (column-padded) ────────────────── for (uint lb = 0u; lb < LPTB; ++lb) { uint lin = lb * (RTSM * RTSN) + tidn * RTSM + tidm; uint row = lin % TSN; uint col = lin / TSN; uint gCol = offN + row; uint gK = kBase + col; Bsub[row][col] = (gCol < NDIM && gK < KDIM) ? B[gK * NDIM + gCol] : 0.0; } barrier(); // Inner loop within tile. // Last (possibly partial) tile: bound K so we never read past KDIM. uint tileK = min(TS_K, KDIM - kBase); for (uint k = 0u; k < tileK; ++k) { float Areg; for (uint wm = 0u; wm < WPTM; ++wm) { Areg = Asub[k][tidm + wm * RTSM]; for (uint wn = 0u; wn < WPTN; ++wn) { float Breg = Bsub[tidn + wn * RTSN][k]; acc[wm][wn] += Areg * Breg; } } } barrier(); } for (uint wm = 0u; wm < WPTM; ++wm) { uint gRow = offM + tidm + wm * RTSM; if (gRow >= MDIM) continue; for (uint wn = 0u; wn < WPTN; ++wn) { uint gCol = offN + tidn + wn * RTSN; if (gCol >= NDIM) continue; // Row-major write matches `np.frombuffer().reshape(M, N)`. C[gRow * NDIM + gCol] = acc[wm][wn]; } } } ) MKNgeorrrrrr r LPTALPTBs 0/home/per/Documents/VGPU/tiled_matmul_fix_bak.py_tiled_matmul_srcrs v;Dc%jE # v;Ds6{ v;Ds6{ $JD4K (D $JD4K (DvuuvvvvvvsssV!f fc tvr"ttjfdt}t|}jj | jj zdzjj zdz jj zdz td|dzdz |dz td|dzdz |dz G   f dd }td d d d d d d|dd|d|S)Nc t|z S)N)abs)xks rz'_compile_tiled_matmul..sc!a%jr)key)reserverrc > eZdZdZ f dZ fdZy)+_compile_tiled_matmul.._TiledKernel progABCgxgyr r rtile_kclearedc |_c|_|_|_c|_|_c|_|_|_ |_ d|_ y)NFr!) selfA_bufB_bufC_bufr r rr&r'rr"s r__init__z4_compile_tiled_matmul.._TiledKernel.__init__sTDI%*E5 "DFDFDF!2 DGTW%&1 "DFDFDFDK DLrc R|jj|j|jj|j|jj d|jj d|j j d|j j|j|jd jj|j j}tj|tjj!}|j"s^t%|dk7j't)|j*dz }|dkrt-ddddd|z d zd d d |_|S) Nrr)dtypeg{Gz?z[VGPU] tiled matmul rz: C buffer is dz.1fuT% zeros — shader produced no output. Likely a bug in the shader, not in this call.T)r#writetobytesr$bind_to_storage_bufferr%r"runr&r'ctxfinishreadnp frombufferfloat32copyr)floatsummaxsize RuntimeError) r+A_arrB_arrbufoutnz_ratior r r self_outers r__call__z4_compile_tiled_matmul.._TiledKernel.__call__s? FFLL ) FFLL ) FF ) )! , FF ) )! , FF ) )! , IIMM$''477A . NN ! ! # &&++-C--2::6;;=C<<!#(!12S15EEd?&.qc1#QqchJ+C01%&'' $ JrN)__name__ __module__ __qualname__ __slots__r/rJ) r,r-r.r r rr&r'rr"rIs r _TiledKernelr s9  ! !  rrOu#[VGPU] ✅ compiled tiled matmul M=z K=z N=z k=z @ workgroup ×u × r r ) TILE_GEOMETRIESminkeysrr8compute_shaderbufferrAprint)rIr r rrrsrcrOr,r-r.r&r'r"s````` @@@@@@r_compile_tiled_matmulrXs\ $$&,@ A ! C Aq!S )C >> ( ( -D NN ! !!a%!) ! 4E NN ! !!a%!) ! 4E NN ! !!a%!) ! 4E QSZ!#E 2 3B QSZ!#E 2 3B%%%N  cQCs1#T!%Brd$s6{m2c&k] DE >rc dddd}||_y)z=Adds `cache.verify_matmul(M, K, N, k=16)`. Use after install.gMbP?)atolc L|j |js tdytjj ||j tj}tjj ||j tj}||z}|j|||j||} ttjtj| |z } | |k} td|d|d|d|d| dd | rd nd z| S) z;Run a small ground-truth check before trusting a benchmark.u4[verify] CPU backend — skipped (nothing to verify)T)rz[verify] matmul(,z, k=z) err=z.2ez u✓u✗ ABOVE TOLERANCE) r8 gpu_availablerVr;randomrandnastyper=matmulreshaper?rAr) r+r r rrrZXWrefrGerroks r verify_matmulz%install_verify..verify_matmuls 88 4#5#5 H I IIOOAq ! ( ( 4 IIOOAq ! ( ( 4!ekk!Q!k$,,Q2BFF266#),-.Dj  1QCq4s'#c"E"8: ; rN)@riri)rh)clsrhs rinstall_verifyrls D &Cr)rj)__doc__numpyr;rrXrlr rrros!  md8|&r