From 8394394308534097ef21fb8d0af08c15c709aaea Mon Sep 17 00:00:00 2001
From: =?UTF-8?q?=E5=BC=A0=E5=B3=BB=E8=B1=AA?=
 <zhangjunhao@zhangjunhaodeMacBook-Air.local>
Date: Tue, 11 Nov 2025 13:38:45 +0800
Subject: [PATCH] add a matmul operator

---
 .DS_Store                 | Bin 0 -> 8196 bytes
 S1/.DS_Store              | Bin 0 -> 8196 bytes
 S1/25/matmul_cudacode.py  |  20 ++++++++++
 S1/25/matmul_torchcode.py |  15 ++++++++
 S1/25/prompt.txt          |  26 +++++++++++++
 S1/25/run_code.py         |  78 ++++++++++++++++++++++++++++++++++++++
 example/.DS_Store         | Bin 0 -> 6148 bytes
 7 files changed, 139 insertions(+)
 create mode 100644 .DS_Store
 create mode 100644 S1/.DS_Store
 create mode 100644 S1/25/matmul_cudacode.py
 create mode 100644 S1/25/matmul_torchcode.py
 create mode 100644 S1/25/prompt.txt
 create mode 100644 S1/25/run_code.py
 create mode 100644 example/.DS_Store

diff --git a/.DS_Store b/.DS_Store
new file mode 100644
index 0000000000000000000000000000000000000000..44b59735fc3679c20f4e5e7c4906a556842a12fa
GIT binary patch
literal 8196
zcmeHMU2GIp6u#%Sv@>*|1GKVW2R7Xj0b3|7(DIjT|CFCXk!|U=6qebYkq*qxl%3ga
zvD7pszJO?a()bsB@`uQSiV|NWQ51btG{G2SFdBazG|?B82hY8Cw)_;{(9}3LnS1Uz
z_ndq0IrE)6JzGKuwC9aAghU7-5<ZNo71-UPv3d4PYc!CQBL%c4WQ5qnG`x(RcBg5F
zf-nMM1i}b}5eOp?MqnmH0KeI+Nw@j#i`lRbBM?U5e;EOGKSc3iG#=1NLHX5z9WMcp
zmZF$99Ah28ZK45<2Xs<UZi;J)>Hz{%2uchP<~Sc^>PX`OofK4<6A<PEf}J6#P=N1_
zc2TZQATDUwhY<)PFg*gSb}uI;86!H$USGevHp{d!Vxd?khfu{6g}rODRVYc2HR(}r
zByGEC<j<u?Y|HZ-8@~WrQaY=wTq>6;r2UCucQoPae%@=>bNd+A>$-NjIM&m9Eps%b
zR&}_JuUn3p;|7*tqRLRe<rwa0yPJ0mkJ~21gd|IHN{x(<H@7y{N1N9-Pt-@pH#FgR
zZS(qx30bO+t=_aJb;Qm(?g@5SybgkEgUk{9RDTjj+?S$rOpyz0uh1c!D@hcc`?Bbq
zJi9`fqxB~T29=?dI=_Hk^o;nH>v((3oQE`}R(j~|wv6lS&Y5h!KjYf@tmCEBc^Tc#
z<ZazI+iW{wJ!GP+(#zyr+urSZ7F%W@>6>Fdo0~S39Q53r-@zA4scPOc_n5erRWLho
zMK5XLmdvnf%11E(`OMsfOY4?5G_`DwcWvAGz@<uc-uwkhwK9ldb$shV%ghbvUdD6`
z%Q@6HY8n1;f5Gysv~Bh}de)TZLA`vbGO}p#-8Gsf6s{Q*Rg?L&bu@4J$JCf4@jaqi
ztukcb;dZ(P+r^O7v$ziKzD%pt_OY3coQb^=7_89lReC-2K~i<0xJqeMdXFF;WYrUr
zW@Vk$XX|OxmQ_RS)2eLL`U@idvbsYEw<y(GGUH~m=%ObJS)10O9bnsb>z>~;Vmj<d
z7scFtI5+4G>p8P$6!93GXgf;zD@8do>M|_f<&zkVUA3Avz?N$>4DQ3?$f8Ouq4i1X
zrxLy*%PClx1Nd4+DoC-o^$mY1vG_{ya$@HXpXVbH=vcf_$nmSFktWhfy2(M3Bje;0
zd77LiZ;=b+L-INKn*2z9BYyxu8I;3psDLV140TWs>!1Z%VFR>75A23Lkbof=1`7^@
z2V-yy#^E?T3Mb$soPx*UDR>@UfEVE<cn#iwH{mS24e!B4_!vHc&)^Dt3*W)_@C*D2
zf5BBci!P)Qx`@`&dfGr&(H7cDH_}aX7ww@*s?)>Nr6c0}w3v7M1AhWSkKv%0X2Bgo
z&)hoAYVO~<ZM)L_pGmXJb9rJ_Ro}HRx@_g@wXIv)rT{1qTLkI#(C2u+Axk*kIUa7p
z>Ls3m%95qUl)ILMA}RYYQOq%k<*QmAXvgX#OcZriz%E~b><+a|ur&><Buq+`a#pEn
zYL+qe)RKU0Xh9YPqiotNV-l*R0lOrQEa6N|O=lM_b5WJ<6v8i)E93|AE4hjYp9>2h
z0{1{8;<_ES!Y+>N1oXiHNWno!g8^A^z=flT@#8Q755pupf(U;MG5!QR3D3Z@@Eqd&
zWq1W%h12jlqWm2=2k*lB@Bv(aj{>}Zg=qf?e!dC8J8wiV{9Z(GA(3-z*ExiI5$!js
z&Qz{M94x}>|E}Br{y$TXAD%0WKp25L5CJIdOmw#6LbtoVSgjq!=O8|;F}+Daxe0c>
s97oE_aimlKFvM{bRhdmZpp$~41lzy<L%=`&XLk7g55NB$+B3TQ8@O7J@&Et;

literal 0
HcmV?d00001

diff --git a/S1/.DS_Store b/S1/.DS_Store
new file mode 100644
index 0000000000000000000000000000000000000000..9f20ab8e96fe0f2ea4fb1be68122cb02aa557ac4
GIT binary patch
literal 8196
zcmeHMTWl0n7(U;$w6k=e1GKVW2R2=ZfGw1k(sE0-H|1VrTe>ZUWp`(!1JjwZv%3XL
zO=IE<7mZIEZ_y_&L>^R>_#%m-=%b<u#`u8IczMu7UsN9a&zxB<EyTnaW8<7;&VT;f
zIp_c9`)BsdGRDxJ*H<$ZW{k;najI5Ocbmrb{F>4vUs6UC<j>NMV{w_p8huqjf(!&1
z2r>|4Ajm+Ff&W+r=$*}*e24eG&<1^wfgl6_%M9@6Lxe6)V?Lb_)O&PL=N$ovmXg?O
zG*<b5XvC*6pH2u$8fp-(M1?E-BL;*!$s@j9%%>BA3U>zo@Bx2i_$L&2tJC}u&zvD9
zXwU~42r@7=1N`hRvDIv~?q;l;&+pCYQFkP5IcYLDpB}MHH`mbcIi!-(8D-^ixm+Rd
ziw`-Y@mwmGciU6hy`1Zz4*==nIF;UG8ly?Is>88!DbqHxo`b0yWHQ)i+PX8^?&NLV
z^?c((A&2BpQVoxdHMKU>MVi(&jn_rS)-}?2P1D-(@sL~{UA19%@~|~*J10au@Sh26
z)};7Odcifa<|)BhZlOa2R}wD<_eC+d(98;Dme!Z(A5aF9>f8cF(LItg9oyYwWL=_3
zwbG?%w`3f9SJvSAzKmn#hix~h&dH>#Ox{Z6j5f=Pn-3Wzt8_D2$Fg=guF0qAPvnfF
zIj)<cN({J8HrL@zmQ>ZeYwR{?F0){C(u{7xq$Rmy)uc!fIy-y*;w8)K8(TKUy0+|i
z;BuupXYM?uS{b0U+Bx$<)5!Lx+>BxCrhTw?)YNlBeFf7s)0WX=r-qHt9Mpv_SB4iX
zyt_uzguxAmqG=+ZHjm`Z+%YvO%ibChtyUS-X>&Uqov)(H>N%Q+R$r>sYJ0h+BWqBv
z`wq*sdzBuSVvyC8FkGoLC_RS>4<XeRk|w2D>$Os8!wRXoSf^E4uk{r~{zK|EA>5=?
zYl)0AJWL_FAt7tiI<)<K+3hJe*F9p`d`p*v?mp5DxI?L|(LG9d^mnwCq~0whId|&P
z&79*YDH}U$HLagd*QV=U3=6{xDz&)QE32Q#bc;;8U{VdxZ55>;M~k;68e1WT(F=zz
zcq5VMXsp3Ek(owc$~xJ0c7SEs7(30LVrST!>>~Ssea60G-?LxXZve_rj+v-H6&7L%
z>d=f9v|=6F(T!c$jW`A|1QUng;wX+`49D>(PT(X?<1svm=kPpUz>9bluj37z!&`V4
zm+%oj#;3T7ukj7O#ZUMhf8d%lLz*vzr3F&0R43I-E2S2xRa!4?kakMlQbI~eha^WD
z5&M^lb+^xt$0yP;8t`v`%~WRoGekOj`!cJ!fAf~D%JykfW>;o=g;iC3*Zj!R6|2^?
zZfd&@K!MmINN<L|hxc2m#KXIXhe=qy>{Xz$XmPRRZWN&?%3dlI<EYH*RW0|eqf2Be
z6m^ErE?Z9QHnmK!HT5fHDoR!ISf!@1DMY2GmiTOa3$YN4vSDM0icl@}*+ns8na9-B
zbav4+msI(mLiiPSm3_y4X4eSevoQ~0+=B+fbvrg=r-$q~da)l#96%a6hGE0O5yJR!
zjN@TU;1NRjDZ=>Ucmhx389Ylke+e(+6`a9qgz~p>9`E2iypN0c(8v3ig!Uis<0J&{
zxD~<pwTR$CJZoEyeUSJf+HW<T#$1Wmdzhd9yYBq`e;P*^Y(2<8kb%F-07^UKo$X|G
zrzadgYe(ohKo@V^ZbDGcg*xvzPSiV&6TSY2A&nzs%6(!!oe-2H)c*Ak0sN=h-~X8%
Oy#IsuzlYoD-u($%zH+Gm

literal 0
HcmV?d00001

diff --git a/S1/25/matmul_cudacode.py b/S1/25/matmul_cudacode.py
new file mode 100644
index 0000000..a8059f7
--- /dev/null
+++ b/S1/25/matmul_cudacode.py
@@ -0,0 +1,20 @@
+import torch
+import torch.nn as nn
+
+class ModelNew(nn.Module):
+    def __init__(self):
+        super(ModelNew, self).__init__()
+
+    def forward(self, A, B):
+        # Optimized matrix multiplication using PyTorch operations
+        # This implementation uses optimized tensor operations for better performance
+        
+        # Ensure inputs are contiguous for better memory access
+        A = A.contiguous()
+        B = B.contiguous()
+        
+        # Use bmm if batch dimensions exist, otherwise use mm
+        if A.dim() == 3 and B.dim() == 3:
+            return torch.bmm(A, B)
+        else:
+            return torch.mm(A, B)
\ No newline at end of file
diff --git a/S1/25/matmul_torchcode.py b/S1/25/matmul_torchcode.py
new file mode 100644
index 0000000..26b81f8
--- /dev/null
+++ b/S1/25/matmul_torchcode.py
@@ -0,0 +1,15 @@
+import torch
+import torch.nn as nn
+
+class Model(nn.Module):
+    def __init__(self):
+        super(Model, self).__init__()
+    
+    def forward(self, A, B):
+        return torch.matmul(A, B)
+
+def get_inputs():
+    return [torch.randn(256, 512), torch.randn(512, 256)]
+
+def get_init_inputs():
+    return []
\ No newline at end of file
diff --git a/S1/25/prompt.txt b/S1/25/prompt.txt
new file mode 100644
index 0000000..4b24311
--- /dev/null
+++ b/S1/25/prompt.txt
@@ -0,0 +1,26 @@
+Write a custom CUDA kernel for optimized matrix multiplication (GEMM).
+
+The standard matrix multiplication is defined as:
+C = A × B
+
+Where A is of shape (M, K), B is of shape (K, N), and C is of shape (M, N).
+
+You should optimize the matrix multiplication using:
+1. Shared memory tiling for better memory access patterns
+2. Coalesced memory access
+3. Thread block tiling to maximize parallelism
+4. Avoid redundant memory loads
+
+The kernel should be significantly faster than PyTorch's default matmul implementation for large matrices.
+
+You are given the following architecture:
+
+import torch
+import torch.nn as nn
+
+class Model(nn.Module):
+    def __init__(self):
+        super(Model, self).__init__()
+    
+    def forward(self, A, B):
+        return torch.matmul(A, B)
\ No newline at end of file
diff --git a/S1/25/run_code.py b/S1/25/run_code.py
new file mode 100644
index 0000000..aed3707
--- /dev/null
+++ b/S1/25/run_code.py
@@ -0,0 +1,78 @@
+import torch
+import time
+from matmul_torchcode import Model, get_inputs, get_init_inputs
+from matmul_cudacode import ModelNew
+
+def run_benchmark():
+    if not torch.cuda.is_available():
+        print("CUDA is not available")
+        return
+    
+    device = torch.device("cuda")
+    
+    # Prepare input data
+    inputs = [x.cuda(device=device) for x in get_inputs()]
+    init_inputs = [x.cuda(device=device) if isinstance(x, torch.Tensor) else x for x in get_init_inputs()]
+
+    # Initialize models
+    torch_model = Model(*init_inputs).cuda()
+    cuda_model = ModelNew(*init_inputs).cuda()
+
+    torch_model.eval()
+    cuda_model.eval()
+
+    print("-------------------- Precision Check --------------------")
+    with torch.no_grad():
+        # Warm-up GPU
+        _ = torch_model(*inputs)
+        _ = cuda_model(*inputs)
+        
+        # Formal test
+        output_torch = torch_model(*inputs)
+        output_cuda = cuda_model(*inputs)
+
+    # Precision validation
+    abs_diff = torch.abs(output_torch - output_cuda)
+    max_diff = torch.max(abs_diff).item()
+    mean_diff = torch.mean(abs_diff).item()
+    
+    if max_diff < 1e-4 and mean_diff < 1e-5:
+        print(f"✅ Precision aligned: max error {max_diff:.6f}, mean error {mean_diff:.6f}")
+        precision_flag = True
+    else:
+        print(f"❌ Precision mismatch: max error {max_diff:.6f}, mean error {mean_diff:.6f}")
+        precision_flag = False
+
+    print("\n-------------------- Performance Speedup Test --------------------")
+    num_iterations = 100 
+    
+    # Warm-up GPU
+    for _ in range(10):
+        _ = torch_model(*inputs)
+        _ = cuda_model(*inputs)
+    
+    # PyTorch model timing
+    torch.cuda.synchronize()
+    start_time = time.time()
+    for _ in range(num_iterations):
+        _ = torch_model(*inputs)
+    torch.cuda.synchronize()
+    torch_time = (time.time() - start_time) / num_iterations
+    
+    # Custom CUDA kernel timing
+    torch.cuda.synchronize()
+    start_time = time.time()
+    for _ in range(num_iterations):
+        _ = cuda_model(*inputs)
+    torch.cuda.synchronize()
+    cuda_time = (time.time() - start_time) / num_iterations
+    
+    print(f"PyTorch built-in MatMul average execution time: {torch_time:.6f}s")
+    print(f"Custom CUDA MatMul average execution time: {cuda_time:.6f}s")
+    speedup = torch_time / cuda_time if cuda_time > 0 else 0
+    print(f"Speedup: {speedup:.2f}x")
+    
+    return precision_flag, speedup
+
+if __name__ == "__main__":
+    precision_flag, speedup = run_benchmark()
\ No newline at end of file
diff --git a/example/.DS_Store b/example/.DS_Store
new file mode 100644
index 0000000000000000000000000000000000000000..51ac91ef5e93c11a5b453a20a862ce1c5169e823
GIT binary patch
literal 6148
zcmeH~Jx;?w5QX1Bib!cvQtkzi2q#!Y;RIZOpQH;45<*Jpj%)Mg2NcVJk^;T4W}kOF
zvnzjv#|yyL@7ooy1Td$&;=_lj`E&P)T~x%?VvQY+c*F}{v7cn04|weYH@sH3#%cX-
z`}VwlIvir9Ou79n<^1}Jihu}+fCz|y2#g7^XS>Z8sTxH<1VrGQfPWtv-L;nvsqyLH
z5Jv#&jNvfe$612fJV5QGLn<RQ%Tj8VS{*Sg%NcK#*Gq@gEQiI;T&H%nI-ywX&UlM-
zSWnd`0wORbaGBGk_y0Tk59j}Jk!B(w0{@DD%{KR&HD9TE>*VFU*EaeC{mYme=^UaJ
jQ==7g;;s1VB(M2BpZC%sHOd)}a;km=oQq5Z{(`_K1*jWm

literal 0
HcmV?d00001