Skip to content

fix(compute): reuse dst GPU memory instead of allocating per call (#84) #85

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

Merged
dndungu merged 2 commits into from
Apr 10, 2026
+126 −32
Merged

fix(compute): reuse dst GPU memory instead of allocating per call (#84) #85

Show file tree
Hide file tree
Changes from all commits
Commits
Show all changes
2 commits
Select commit Hold shift + click to select a range
e2232b6
fix(compute): reuse dst GPU memory instead of allocating per call (#84)
dndungu Apr 10, 2026
88f7394
style(compute): gofmt gpu_engine.go
dndungu Apr 10, 2026
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
32 changes: 23 additions & 9 deletions compute/gpu_engine.go
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -978,13 +978,16 @@ func (e *GPUEngine[T]) MatMul(ctx context.Context, a, b *tensor.TensorNumeric[T]
return nil, err
}

// Allocate device output.
devCTotal, err := e.pool.Alloc(e.deviceID, outputBytes)
if err != nil {
e.oomFallbackCount.Add(1)
e.logger.Warn("MatMul: GPU output alloc failed, falling back to CPU", "error", err.Error())
// Reuse dst's existing GPU memory when possible (#84).
devCTotal, reusedC := tryReuseDstPtr[T](batchSize*cMatSize, dst)
if !reusedC {
devCTotal, err = e.pool.Alloc(e.deviceID, outputBytes)
if err != nil {
e.oomFallbackCount.Add(1)
e.logger.Warn("MatMul: GPU output alloc failed, falling back to CPU", "error", err.Error())

return e.cpu.MatMul(ctx, a, b, dst...)
return e.cpu.MatMul(ctx, a, b, dst...)
}
}

// Use strided batched GEMM when available for float32 with batch > 1.
Expand Down Expand Up @@ -1013,9 +1016,14 @@ func (e *GPUEngine[T]) MatMul(ctx context.Context, a, b *tensor.TensorNumeric[T]
if err := batched.SgemmStridedBatched(m, n, k, 1.0,
devA, strideA, devB, strideBVal, 0.0,
devCTotal, strideC, batchSize); err != nil {
e.pool.Free(e.deviceID, devCTotal, outputBytes)
if !reusedC {
e.pool.Free(e.deviceID, devCTotal, outputBytes)
}
return nil, fmt.Errorf("MatMul: batched GEMM: %w", err)
}
if reusedC {
return finishReusedDst[T](dst[0], outShape), nil
}
return makeGPUResult[T](e, outShape, devCTotal, batchSize*cMatSize, dst...)
}
}
Expand Down Expand Up @@ -1052,12 +1060,17 @@ func (e *GPUEngine[T]) MatMul(ctx context.Context, a, b *tensor.TensorNumeric[T]
}

if blasErr != nil {
e.pool.Free(e.deviceID, devCTotal, outputBytes)
if !reusedC {
e.pool.Free(e.deviceID, devCTotal, outputBytes)
}

return nil, fmt.Errorf("MatMul: BLAS batch %d: %w", batch, blasErr)
}
}

if reusedC {
return finishReusedDst[T](dst[0], outShape), nil
}
return makeGPUResult[T](e, outShape, devCTotal, batchSize*cMatSize, dst...)
}

Expand Down Expand Up @@ -1256,7 +1269,8 @@ func (e *GPUEngine[T]) matMulQ4(ctx context.Context, qs *tensor.Q4Storage, a, b
// matMulQ4BWeight handles MatMul where B has Q4 storage (virtual-transposed weight).
// B's shape after virtual transpose is [K, N], but the Q4 data is laid out as [N, K].
// We compute C[M, N] = A[M, K] * dequant(B)^T by reformulating as:
// C_temp[N, M] = gemm_q4(B_q4[N, K], A^T[K, M])
//
// C_temp[N, M] = gemm_q4(B_q4[N, K], A^T[K, M])
//
// For GEMV (M=1), A^T[K,1] is just A's data as a column, and C_temp[N,1]
// can be reshaped to [1, N] without a physical transpose.
Expand Down
25 changes: 20 additions & 5 deletions compute/gpu_engine_memory.go
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -132,9 +132,14 @@ func (e *GPUEngine[T]) Transpose(ctx context.Context, a *tensor.TensorNumeric[T]
}

byteSize := total * f32Size
devOut, err := e.pool.Alloc(e.deviceID, byteSize)
if err != nil {
return e.cpu.Transpose(ctx, a, axes, dst...)

// Reuse dst's existing GPU memory when possible (#84).
devOut, reused := tryReuseDstPtr[T](total, dst)
if !reused {
devOut, err = e.pool.Alloc(e.deviceID, byteSize)
if err != nil {
return e.cpu.Transpose(ctx, a, axes, dst...)
}
}

// Fast path: 2D transpose.
Expand All @@ -145,9 +150,14 @@ func (e *GPUEngine[T]) Transpose(ctx context.Context, a *tensor.TensorNumeric[T]
"cols", fmt.Sprintf("%d", shape[1]))
}
if err := e.kernels.Transpose2D(devIn, devOut, shape[0], shape[1], e.stream); err != nil {
e.pool.Free(e.deviceID, devOut, byteSize)
if !reused {
e.pool.Free(e.deviceID, devOut, byteSize)
}
return nil, err
}
if reused {
return finishReusedDst[T](dst[0], outShape), nil
}
return makeGPUResult[T](e, outShape, devOut, total, dst...)
}

Expand Down Expand Up @@ -175,10 +185,15 @@ func (e *GPUEngine[T]) Transpose(ctx context.Context, a *tensor.TensorNumeric[T]
}

if err := e.kernels.TransposeND(devIn, devOut, inStrides32, outStrides32, perm32, rank, total, e.stream); err != nil {
e.pool.Free(e.deviceID, devOut, byteSize)
if !reused {
e.pool.Free(e.deviceID, devOut, byteSize)
}
return nil, err
}

if reused {
return finishReusedDst[T](dst[0], outShape), nil
}
return makeGPUResult[T](e, outShape, devOut, total, dst...)
}

Expand Down
101 changes: 83 additions & 18 deletions compute/gpu_kernels.go
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -115,6 +115,35 @@ func getDevicePtr[T tensor.Numeric](e *GPUEngine[T], t *tensor.TensorNumeric[T])
return devPtr, cleanup, nil
}

// tryReuseDstPtr checks whether dst[0] already has a GPUStorage with at least
// neededElems capacity. If so, it returns the existing device pointer so the
// caller can write kernel output directly into it, avoiding a pool.Alloc and
// the resulting GC-pressure from orphaned GPUStorage objects. See ztensor#84.
func tryReuseDstPtr[T tensor.Numeric](neededElems int, dst []*tensor.TensorNumeric[T]) (unsafe.Pointer, bool) {
if len(dst) == 0 || dst[0] == nil {
return nil, false
}
gs, ok := dst[0].GetStorage().(*tensor.GPUStorage[T])
if !ok || gs.Len() < neededElems {
return nil, false
}
return gs.Ptr(), true
}

// finishReusedDst updates dst's shape and strides in place after a kernel has
// written into dst's existing device memory. No new GPUStorage is created.
func finishReusedDst[T tensor.Numeric](dst *tensor.TensorNumeric[T], shape []int) *tensor.TensorNumeric[T] {
strides := make([]int, len(shape))
stride := 1
for i := len(shape) - 1; i >= 0; i-- {
strides[i] = stride
stride *= shape[i]
}
dst.SetShape(shape)
dst.SetStrides(strides)
return dst
}

// makeGPUResult creates a tensor with pool-backed GPUStorage wrapping the given
// device pointer. When the tensor is freed, the pointer is returned to the pool
// for reuse instead of calling cudaFree.
Expand Down Expand Up @@ -522,17 +551,26 @@ func gpuBinaryOp[T tensor.Numeric](

byteSize := n * f32Size

devC, err := e.pool.Alloc(e.deviceID, byteSize)
if err != nil {
return nil, err
// Reuse dst's existing GPU memory when possible (#84).
devC, reused := tryReuseDstPtr[T](n, dst)
if !reused {
devC, err = e.pool.Alloc(e.deviceID, byteSize)
if err != nil {
return nil, err
}
}

if err := kernelFn(devA, devB, devC, n, e.stream); err != nil {
e.pool.Free(e.deviceID, devC, byteSize)
if !reused {
e.pool.Free(e.deviceID, devC, byteSize)
}

return nil, err
}

if reused {
return finishReusedDst[T](dst[0], a.Shape()), nil
}
return makeGPUResult[T](e, a.Shape(), devC, n, dst...)
}

Expand All @@ -559,17 +597,26 @@ func gpuUnaryOp[T tensor.Numeric](

byteSize := n * f32Size

devC, err := e.pool.Alloc(e.deviceID, byteSize)
if err != nil {
return nil, err
// Reuse dst's existing GPU memory when possible (#84).
devC, reused := tryReuseDstPtr[T](n, dst)
if !reused {
devC, err = e.pool.Alloc(e.deviceID, byteSize)
if err != nil {
return nil, err
}
}

if err := kernelFn(devA, devC, n, e.stream); err != nil {
e.pool.Free(e.deviceID, devC, byteSize)
if !reused {
e.pool.Free(e.deviceID, devC, byteSize)
}

return nil, err
}

if reused {
return finishReusedDst[T](dst[0], a.Shape()), nil
}
return makeGPUResult[T](e, a.Shape(), devC, n, dst...)
}

Expand Down Expand Up @@ -597,17 +644,26 @@ func gpuScalarOp[T tensor.Numeric](

byteSize := n * f32Size

devC, err := e.pool.Alloc(e.deviceID, byteSize)
if err != nil {
return nil, err
// Reuse dst's existing GPU memory when possible (#84).
devC, reused := tryReuseDstPtr[T](n, dst)
if !reused {
devC, err = e.pool.Alloc(e.deviceID, byteSize)
if err != nil {
return nil, err
}
}

if err := kernelFn(devA, scalar, devC, n, e.stream); err != nil {
e.pool.Free(e.deviceID, devC, byteSize)
if !reused {
e.pool.Free(e.deviceID, devC, byteSize)
}

return nil, err
}

if reused {
return finishReusedDst[T](dst[0], a.Shape()), nil
}
return makeGPUResult[T](e, a.Shape(), devC, n, dst...)
}

Expand Down Expand Up @@ -957,20 +1013,29 @@ func (e *GPUEngine[T]) gpuSum(ctx context.Context, a *tensor.TensorNumeric[T], a

outByteSize := numStripes * f32Size

devOut, err := e.pool.Alloc(e.deviceID, outByteSize)
if err != nil {
e.oomFallbackCount.Add(1)
e.logger.Warn("Sum: GPU output alloc failed, falling back to CPU", "error", err.Error())
// Reuse dst's existing GPU memory when possible (#84).
devOut, reused := tryReuseDstPtr[T](numStripes, dst)
if !reused {
devOut, err = e.pool.Alloc(e.deviceID, outByteSize)
if err != nil {
e.oomFallbackCount.Add(1)
e.logger.Warn("Sum: GPU output alloc failed, falling back to CPU", "error", err.Error())

return e.cpu.Sum(ctx, a, axis, keepDims, dst...)
return e.cpu.Sum(ctx, a, axis, keepDims, dst...)
}
}

if err := e.kernels.SumAxis(devIn, devOut, outer, inner, axisSize, e.stream); err != nil {
e.pool.Free(e.deviceID, devOut, outByteSize)
if !reused {
e.pool.Free(e.deviceID, devOut, outByteSize)
}

return nil, err
}

if reused {
return finishReusedDst[T](dst[0], newShape), nil
}
return makeGPUResult[T](e, newShape, devOut, numStripes, dst...)
}

Expand Down
Loading