Programmatic Dependent Launch (PDL) for more performance on newer NVIDIA GPUs (Hopper+) (#22522)

* Adds initial PDL setup.

* Adds PDL barriers based on simple heuristic: place "sync" before first input pointer access, and "launch" after last write, e.g. to tensors like dst.

* Further optimization pass of the first half of kernels

* Optimized PDL barriers for the second batch of kernels

* Further refinements after rebase.

* Moves pdl logic to separate function, removes some whitespace

* Strips post-hoc PDL logic

* Adds stream capture PDL setup. Enrolls quantize_q8_1 to leverage pdl to
overlap execution with previous kernels

* Enrolls mul_mat_vec_q, rms_norm_f32 and k_bin_bcast (partly) into PDL

* Enrolls mmvf, rope, set-rows and topk kernels for gpt-oss into PDL

* Introduce ggml_cuda_kernel_launch, to abstract away cudaLaunchKernelEx,
to enable hip/musa compatibility

* Enrolls cpy_scalar_contiguous, k_get_rows_float and rms_norm_f32

* Enrolls flash_attn_combine_results

* Fix: Drops needless and broken check of CUDA arch for PDL. PDL either
works or is without effect.

* Enrolls flash-attention kernels to pdl

* Fix: inlines ggml_cuda_kernel_launch, and uses perfect forwarding for
kernels args. This fixes PDL.

* Perf: Enrolls k_bin_bcast variadic template invocation into PDL, via
and template alias and template expansion

* Enrolls all remaining kernels for qwen3-coder-next into PDL

* Remove all PDL LC calls to create a baseline

* Added LC according to internal guidance and tested kernel performance.

* Enrols missing qwen3-5 kernels passively into PDL.

* Kernel optimizations (LC signals) for qwen3.5

* Enrolls ssm-scan kernels into PDL

* Adds GGML_CUDA_PDL command line option to toggle PDL.

* Fix: Ada and lower compilation by guarding PDL calls correctly

* Cleanup: Removes commented out GGML_CUDA_PDL_LC

* Cleanup: Removes experimental comments

* Adds 90-virtual to build script so that Hopper GPUs can leverage PDL.

* Adds stricter checks to enable PDL, adds env-check to disable it, and removes now superfluous compile option to enable PDL.

* Fix: Correct PDL en/disablement based on device-side arch check. Host
side check is UB. Required moving from macros to inlined functions

* Fix: default-disable PDL. Enable by setting GGML_CUDA_ENABLE_PDL=1

* Enable PDL by default for Hopper+ devices

* Enrolls softcap_f32 and two flash_attn kernels into PDL.

* Improves flash attn PDL barrier placement

* Fix: Perf regression on ada; excludes ada and below from PDL launches

* Improves some sync barrier placements

* Drops superfluous constructor

* Adds #endif guard comments

* Reverts experimental change to top-k-moe.cu, which moved expensive allocations
in front of the PDL barrier. It did not have a meaningful impact.

* Exchanges GGML_CUDA_DISABLE_PDL with GGML_CUDA_PDL. IFF GGML_CUDA_PDL=0
PDL is disabled

* Revert "Drops superfluous constructor". Adds const to remaining
arguments

This reverts commit 12b1d250da0089ae02a9bb71bbb3fd6d70f6f2f1.

* Cleanup: Removes and fixes some comments and whitespace

* Clarifies comment of sync-barrier position

* Relocates and refactors PDL launch functions and accessories

* Adds error checking to the regular kernel launch path

* Drops "auto" in favor of "ggml_cuda_kernel_params"

* Adds "const" to ggml_cuda_kernel_launch_params

* [Whitespace] Adds final newline to common.cuh to make editorconfig CI job happy

ggml/src/ggml-cuda/cpy.cu

@@ -16,6 +16,7 @@ static __global__ void cpy_scalar(const char * cx, char * cdst, const int64_t ne
                                   const int64_t ne00, const int64_t ne01, const int64_t ne02, const int64_t nb00, const int64_t nb01, const int64_t nb02,
                                   const int64_t nb03, const int64_t ne10, const int64_t ne11, const int64_t ne12, const int64_t nb10, const int64_t nb11,
                                   const int64_t nb12, const int64_t nb13) {
+    ggml_cuda_pdl_lc();
     const int64_t i = (int64_t)blockDim.x*blockIdx.x + threadIdx.x;
 
     if (i >= ne) {
@@ -36,6 +37,7 @@ static __global__ void cpy_scalar(const char * cx, char * cdst, const int64_t ne
     const int64_t i10 = i - i13*ne10*ne11*ne12 - i12*ne10*ne11 - i11*ne10;
     const int64_t dst_offset = i10*nb10 + i11*nb11 + i12*nb12 + i13 * nb13;
 
+    ggml_cuda_pdl_sync();
     cpy_1(cx + x_offset, cdst + dst_offset);
 }
 
@@ -59,6 +61,7 @@ static __global__ void cpy_scalar_transpose(const char * cx, char * cdst, const
     __shared__ float tile[2][CUDA_CPY_TILE_DIM_2D][CUDA_CPY_TILE_DIM_2D+1];
     int cur_tile_buf = 0;
 
+    ggml_cuda_pdl_sync();
 #pragma unroll
     for (int i = 0; i < CUDA_CPY_BLOCK_NM; ++i) {
 
@@ -142,6 +145,7 @@ static __global__ void cpy_f32_q(const char * cx, char * cdst, const int64_t ne,
     const int64_t i10 = i - i13*ne10*ne11*ne12 - i12*ne10*ne11 - i11*ne10;
     const int64_t dst_offset = (i10/qk)*nb10 + i11*nb11 + i12*nb12 + i13*nb13;
 
+    ggml_cuda_pdl_sync();
     cpy_blck(cx + x_offset, cdst + dst_offset);
 }
 
@@ -168,6 +172,7 @@ static __global__ void cpy_q_f32(const char * cx, char * cdst, const int64_t ne,
     const int64_t i10 = i - i13*ne10*ne11*ne12 - i12*ne10*ne11 - i11*ne10;
     const int64_t dst_offset = i10*nb10 + i11*nb11 + i12*nb12 + i13*nb13;
 
+    ggml_cuda_pdl_sync();
     cpy_blck(cx + x_offset, cdst + dst_offset);
 }
 
@@ -182,6 +187,7 @@ static __global__ void cpy_scalar_contiguous(const char * cx, char * cdst, const
     const src_t * x = (const src_t *) cx;
     dst_t *     dst = (dst_t *) cdst;
 
+    ggml_cuda_pdl_sync();
     dst[i] = ggml_cuda_cast<dst_t>(x[i]);
 }
 
@@ -192,8 +198,8 @@ cudaStream_t stream) {
 
     const int64_t num_blocks = (ne + CUDA_CPY_BLOCK_SIZE - 1) / CUDA_CPY_BLOCK_SIZE;
     GGML_ASSERT(num_blocks < UINT_MAX);
-    cpy_scalar_contiguous<src_t, dst_t><<<num_blocks, CUDA_CPY_BLOCK_SIZE, 0, stream>>>
-        (cx, cdst, ne);
+    const ggml_cuda_kernel_launch_params launch_params = ggml_cuda_kernel_launch_params((dim3)num_blocks, CUDA_CPY_BLOCK_SIZE, 0, stream);
+    ggml_cuda_kernel_launch(cpy_scalar_contiguous<src_t, dst_t>, launch_params, cx, cdst, ne);
 }
 
 template<typename src_t, typename dst_t, bool transposed = false>
@@ -223,13 +229,15 @@ static void ggml_cpy_scalar_cuda(
         GGML_ASSERT(grid_z < USHRT_MAX);
         dim3 dimGrid(grid_x, grid_y, grid_z);
         dim3 dimBlock(CUDA_CPY_TILE_DIM_2D, CUDA_CPY_BLOCK_ROWS, 1);
-        cpy_scalar_transpose<dst_t><<<dimGrid, dimBlock, 0, stream>>>
-            (cx, cdst, ne, ne00n, ne01n, ne02n, nb00, nb01, nb02, nb03, ne10, ne11, ne12, nb10, nb11, nb12, nb13);
+        const ggml_cuda_kernel_launch_params launch_params = ggml_cuda_kernel_launch_params(dimGrid, dimBlock, 0, stream);
+        ggml_cuda_kernel_launch(cpy_scalar_transpose<dst_t>, launch_params,
+            cx, cdst, ne, ne00n, ne01n, ne02n, nb00, nb01, nb02, nb03, ne10, ne11, ne12, nb10, nb11, nb12, nb13);
     } else {
         const int64_t num_blocks = (ne + CUDA_CPY_BLOCK_SIZE - 1) / CUDA_CPY_BLOCK_SIZE;
         GGML_ASSERT(num_blocks < UINT_MAX);
-        cpy_scalar<cpy_1_scalar<src_t, dst_t>><<<num_blocks, CUDA_CPY_BLOCK_SIZE, 0, stream>>>
-            (cx, cdst, ne, ne00, ne01, ne02, nb00, nb01, nb02, nb03, ne10, ne11, ne12, nb10, nb11, nb12, nb13);
+        const ggml_cuda_kernel_launch_params launch_params = ggml_cuda_kernel_launch_params((dim3)num_blocks, CUDA_CPY_BLOCK_SIZE, 0, stream);
+        ggml_cuda_kernel_launch(cpy_scalar<cpy_1_scalar<src_t, dst_t>>, launch_params,
+            cx, cdst, ne, ne00, ne01, ne02, nb00, nb01, nb02, nb03, ne10, ne11, ne12, nb10, nb11, nb12, nb13);
     }
 }