From 84d9ce0faddad002cd26a5a3fcd29f71fb426a1f Mon Sep 17 00:00:00 2001 From: lhez Date: Mon, 2 Nov 2026 03:59:05 -0800 Subject: [PATCH] opencl: add topk_mask (#31441) --- ggml/src/ggml-opencl/CMakeLists.txt | 1 + ggml/src/ggml-opencl/ggml-opencl.cpp | 146 +++++++++++++++++++++ ggml/src/ggml-opencl/kernels/top_k_mask.cl | 129 ++++++++++++++++++ 3 files changed, 276 insertions(+) create mode 100644 ggml/src/ggml-opencl/kernels/top_k_mask.cl diff --git a/ggml/src/ggml-opencl/CMakeLists.txt b/ggml/src/ggml-opencl/CMakeLists.txt index f905370a3..79fc37b60 100644 --- a/ggml/src/ggml-opencl/CMakeLists.txt +++ b/ggml/src/ggml-opencl/CMakeLists.txt @@ -67,6 +67,7 @@ set(GGML_OPENCL_KERNELS argmax argsort top_k + top_k_mask lightning_indexer tri fill diff --git a/ggml/src/ggml-opencl/ggml-opencl.cpp b/ggml/src/ggml-opencl/ggml-opencl.cpp index 6cfd0c0f9..97956bed1 100644 --- a/ggml/src/ggml-opencl/ggml-opencl.cpp +++ b/ggml/src/ggml-opencl/ggml-opencl.cpp @@ -873,6 +873,7 @@ struct ggml_backend_opencl_context { ggml_cl_adreno_xmem_attn_state adreno_xmem_attn; #endif cl_kernel kernel_top_k_f32, kernel_top_k_merge_f32; + cl_kernel kernel_top_k_mask[2], kernel_top_k_mask_fill[2]; cl_kernel kernel_lightning_indexer; cl_kernel kernel_get_rows_f32, kernel_get_rows_f16, kernel_get_rows_i32, kernel_get_rows_q4_0; cl_kernel kernel_set_rows_f32_i64, kernel_set_rows_f32_i32, kernel_set_rows_f16_i64, kernel_set_rows_f16_i32; @@ -2298,6 +2299,27 @@ static void load_cl_kernels(ggml_backend_opencl_context *backend_ctx) { GGML_LOG_CONT("."); } + // top_k_mask + { +#ifdef GGML_OPENCL_EMBED_KERNELS + const std::string kernel_src { + #include "top_k_mask.cl.h" + }; +#else + const std::string kernel_src = read_file("top_k_mask.cl"); +#endif + // Scores can contain NaN and -inf, so do not use the default fast-math options. + const std::string opts = "-cl-std=" + opencl_c_std + + (backend_ctx->adreno_use_large_buffer ? " -qcom-enable-large-buffer" : ""); + for (int i = 0; i < 2; ++i) { + cl_program prog = build_program_from_source(backend_ctx, kernel_src.c_str(), opts + " -DMASK_F16=" + std::to_string(i == 0)); + CL_CHECK((backend_ctx->kernel_top_k_mask[i] = clCreateKernel(prog, "kernel_top_k_mask", &err), err)); + CL_CHECK((backend_ctx->kernel_top_k_mask_fill[i] = clCreateKernel(prog, "kernel_top_k_mask_fill", &err), err)); + CL_CHECK(clReleaseProgram(prog)); + } + GGML_LOG_CONT("."); + } + // get_rows { #ifdef GGML_OPENCL_EMBED_KERNELS @@ -11800,6 +11822,14 @@ static bool ggml_opencl_supports_op(ggml_backend_dev_t dev, const struct ggml_te op->src[2]->nb[0] == sizeof(float) && op->src[3]->nb[0] == sizeof(ggml_fp16_t); case GGML_OP_TOP_K: return op->src[0]->type == GGML_TYPE_F32 && op->type == GGML_TYPE_I32; + case GGML_OP_TOP_K_MASK: + return (op->type == GGML_TYPE_F16 || op->type == GGML_TYPE_F32) && + op->src[0]->type == op->type && + op->src[0]->nb[0] == ggml_type_size(op->type) && op->nb[0] == ggml_type_size(op->type) && + op->src[1]->type == GGML_TYPE_I32 && + (!op->src[2] || op->src[2]->type == GGML_TYPE_F32) && + (!op->src[3] || op->src[3]->type == GGML_TYPE_I32) && + (!op->src[4] || op->src[4]->type == GGML_TYPE_I32); case GGML_OP_ARGSORT: { load_cl_kernels_argsort(backend_ctx); @@ -36968,6 +36998,119 @@ static void ggml_cl_top_k(ggml_backend_t backend, const ggml_tensor * src0, cons } } +static void ggml_cl_top_k_mask(ggml_backend_t backend, const ggml_tensor * src0, const ggml_tensor * src1, ggml_tensor * dst) { + ggml_backend_opencl_context * backend_ctx = (ggml_backend_opencl_context *) backend->context; + const ggml_tensor * score = dst->src[2]; + const ggml_tensor * blocks = dst->src[3]; + const ggml_tensor * extra = dst->src[4]; + const int type = dst->type == GGML_TYPE_F16 ? 0 : 1; + cl_kernel kernel = backend_ctx->kernel_top_k_mask[type]; + + const ggml_tensor_extra_cl * extram = (const ggml_tensor_extra_cl *) src0->extra; + const ggml_tensor_extra_cl * extrak = (const ggml_tensor_extra_cl *) src1->extra; + const ggml_tensor_extra_cl * extras = score ? (const ggml_tensor_extra_cl *) score->extra : nullptr; + const ggml_tensor_extra_cl * extrab = blocks ? (const ggml_tensor_extra_cl *) blocks->extra : nullptr; + const ggml_tensor_extra_cl * extrae = extra ? (const ggml_tensor_extra_cl *) extra->extra : nullptr; + const ggml_tensor_extra_cl * extrad = (const ggml_tensor_extra_cl *) dst->extra; + + const cl_mem buffer_s = extras ? extras->data_device : nullptr; + const cl_mem buffer_b = extrab ? extrab->data_device : nullptr; + const cl_mem buffer_e = extrae ? extrae->data_device : nullptr; + + const cl_ulong offsetm = extram->offset + src0->view_offs; + const cl_ulong offsetk = extrak->offset + src1->view_offs; + const cl_ulong offsets = extras ? extras->offset + score->view_offs : 0; + const cl_ulong offsetb = extrab ? extrab->offset + blocks->view_offs : 0; + const cl_ulong offsete = extrae ? extrae->offset + extra->view_offs : 0; + const cl_ulong offsetd = extrad->offset + dst->view_offs; + + // single launch: fill == true, two launches: fill = false + // X2 benefits from a parallel fill for long single rows; mobile Adreno favors one launch. + const cl_int fill = !(backend_ctx->adreno_gen == ADRENO_GPU_GEN::X2E && + ggml_nrows(dst) == 1 && dst->ne[0] >= 65536 && ggml_is_contiguous(dst)); + + if (!fill) { + GGML_ASSERT(ggml_is_contiguous(dst)); + + cl_kernel kernel_fill = backend_ctx->kernel_top_k_mask_fill[type]; + + const cl_ulong n = ggml_nelements(dst); + const size_t nv = 16/ggml_type_size(dst->type); + const size_t nth = MIN((size_t) 256, backend_ctx->get_kernel_workgroup_size(kernel_fill)); + + CL_CHECK(clSetKernelArg(kernel_fill, 0, sizeof(cl_mem), &extrad->data_device)); + CL_CHECK(clSetKernelArg(kernel_fill, 1, sizeof(cl_ulong), &offsetd)); + CL_CHECK(clSetKernelArg(kernel_fill, 2, sizeof(cl_ulong), &n)); + + size_t global[] = {((n + nv - 1)/nv + nth - 1)/nth*nth}; + size_t local[] = {nth}; + backend_ctx->enqueue_ndrange_kernel(kernel_fill, 1, global, local, dst); + } + + GGML_TENSOR_LOCALS(cl_ulong, nem, src0, ne); + GGML_TENSOR_LOCALS(cl_ulong, nbm, src0, nb); + GGML_TENSOR_LOCALS(cl_ulong, nek, src1, ne); + GGML_TENSOR_LOCALS(cl_ulong, nbk, src1, nb); + GGML_TENSOR_LOCALS(cl_ulong, nes, score, ne); + GGML_TENSOR_LOCALS(cl_ulong, nbs, score, nb); + GGML_TENSOR_LOCALS(cl_ulong, neb, blocks, ne); + GGML_TENSOR_LOCALS(cl_ulong, nbb, blocks, nb); + GGML_TENSOR_LOCALS(cl_ulong, nee, extra, ne); + GGML_TENSOR_LOCALS(cl_ulong, nbe, extra, nb); + GGML_TENSOR_LOCALS(cl_ulong, ned, dst, ne); + GGML_TENSOR_LOCALS(cl_ulong, nbd, dst, nb); + + CL_CHECK(clSetKernelArg(kernel, 0, sizeof(cl_mem), &extram->data_device)); + CL_CHECK(clSetKernelArg(kernel, 1, sizeof(cl_ulong), &offsetm)); + CL_CHECK(clSetKernelArg(kernel, 2, sizeof(cl_mem), &extrak->data_device)); + CL_CHECK(clSetKernelArg(kernel, 3, sizeof(cl_ulong), &offsetk)); + CL_CHECK(clSetKernelArg(kernel, 4, sizeof(cl_mem), &buffer_s)); + CL_CHECK(clSetKernelArg(kernel, 5, sizeof(cl_ulong), &offsets)); + CL_CHECK(clSetKernelArg(kernel, 6, sizeof(cl_mem), &buffer_b)); + CL_CHECK(clSetKernelArg(kernel, 7, sizeof(cl_ulong), &offsetb)); + CL_CHECK(clSetKernelArg(kernel, 8, sizeof(cl_mem), &buffer_e)); + CL_CHECK(clSetKernelArg(kernel, 9, sizeof(cl_ulong), &offsete)); + CL_CHECK(clSetKernelArg(kernel, 10, sizeof(cl_mem), &extrad->data_device)); + CL_CHECK(clSetKernelArg(kernel, 11, sizeof(cl_ulong), &offsetd)); + CL_CHECK(clSetKernelArg(kernel, 12, sizeof(cl_ulong), &nem0)); + CL_CHECK(clSetKernelArg(kernel, 13, sizeof(cl_ulong), &nem1)); + CL_CHECK(clSetKernelArg(kernel, 14, sizeof(cl_ulong), &nem2)); + CL_CHECK(clSetKernelArg(kernel, 15, sizeof(cl_ulong), &nem3)); + CL_CHECK(clSetKernelArg(kernel, 16, sizeof(cl_ulong), &nbm1)); + CL_CHECK(clSetKernelArg(kernel, 17, sizeof(cl_ulong), &nbm2)); + CL_CHECK(clSetKernelArg(kernel, 18, sizeof(cl_ulong), &nbm3)); + CL_CHECK(clSetKernelArg(kernel, 19, sizeof(cl_ulong), &nek0)); + CL_CHECK(clSetKernelArg(kernel, 20, sizeof(cl_ulong), &nbk0)); + CL_CHECK(clSetKernelArg(kernel, 21, sizeof(cl_ulong), &nbk1)); + CL_CHECK(clSetKernelArg(kernel, 22, sizeof(cl_ulong), &nbk2)); + CL_CHECK(clSetKernelArg(kernel, 23, sizeof(cl_ulong), &nbk3)); + CL_CHECK(clSetKernelArg(kernel, 24, sizeof(cl_ulong), &nes0)); + CL_CHECK(clSetKernelArg(kernel, 25, sizeof(cl_ulong), &nbs0)); + CL_CHECK(clSetKernelArg(kernel, 26, sizeof(cl_ulong), &nbs1)); + CL_CHECK(clSetKernelArg(kernel, 27, sizeof(cl_ulong), &nbs2)); + CL_CHECK(clSetKernelArg(kernel, 28, sizeof(cl_ulong), &nbs3)); + CL_CHECK(clSetKernelArg(kernel, 29, sizeof(cl_ulong), &neb0)); + CL_CHECK(clSetKernelArg(kernel, 30, sizeof(cl_ulong), &neb1)); + CL_CHECK(clSetKernelArg(kernel, 31, sizeof(cl_ulong), &nbb0)); + CL_CHECK(clSetKernelArg(kernel, 32, sizeof(cl_ulong), &nbb1)); + CL_CHECK(clSetKernelArg(kernel, 33, sizeof(cl_ulong), &nee0)); + CL_CHECK(clSetKernelArg(kernel, 34, sizeof(cl_ulong), &nbe0)); + CL_CHECK(clSetKernelArg(kernel, 35, sizeof(cl_ulong), &nbe1)); + CL_CHECK(clSetKernelArg(kernel, 36, sizeof(cl_ulong), &nbe2)); + CL_CHECK(clSetKernelArg(kernel, 37, sizeof(cl_ulong), &nbe3)); + CL_CHECK(clSetKernelArg(kernel, 38, sizeof(cl_ulong), &ned1)); + CL_CHECK(clSetKernelArg(kernel, 39, sizeof(cl_ulong), &ned2)); + CL_CHECK(clSetKernelArg(kernel, 40, sizeof(cl_ulong), &nbd1)); + CL_CHECK(clSetKernelArg(kernel, 41, sizeof(cl_ulong), &nbd2)); + CL_CHECK(clSetKernelArg(kernel, 42, sizeof(cl_ulong), &nbd3)); + CL_CHECK(clSetKernelArg(kernel, 43, sizeof(cl_int), &fill)); + + const size_t nth = MIN((size_t) 512, backend_ctx->get_kernel_workgroup_size(kernel)); + size_t global[] = {nth, (size_t) ggml_nrows(dst)}; + size_t local[] = {nth, 1}; + backend_ctx->enqueue_ndrange_kernel(kernel, 2, global, local, dst); +} + static void ggml_cl_argsort(ggml_backend_t backend, const ggml_tensor * src0, const ggml_tensor * src1, ggml_tensor * dst) { GGML_ASSERT(src0); GGML_ASSERT(src0->extra); @@ -37893,6 +38036,9 @@ bool ggml_cl_compute_forward(ggml_backend_t backend, struct ggml_tensor * tensor case GGML_OP_TOP_K: func = ggml_cl_top_k; break; + case GGML_OP_TOP_K_MASK: + func = ggml_cl_top_k_mask; + break; case GGML_OP_ARGSORT: if (!any_on_device) { return false; diff --git a/ggml/src/ggml-opencl/kernels/top_k_mask.cl b/ggml/src/ggml-opencl/kernels/top_k_mask.cl new file mode 100644 index 000000000..00fc0fb13 --- /dev/null +++ b/ggml/src/ggml-opencl/kernels/top_k_mask.cl @@ -0,0 +1,129 @@ +// Integer storage preserves the mask bits without FP16 conversion. +#if MASK_F16 +#define MASK_T ushort +#define MASK_V ushort8 +#define MASK_NV 8 +#define MASK_NINF 0xfc00 +#define MASK_STORE vstore8 +#else +#define MASK_T uint +#define MASK_V uint4 +#define MASK_NV 4 +#define MASK_NINF 0xff800000u +#define MASK_STORE vstore4 +#endif + +kernel void kernel_top_k_mask_fill( + global char * dst, + ulong offsetd, + ulong n +) { + global MASK_T * out = (global MASK_T *) (dst + offsetd); + const ulong c = get_global_id(0)*MASK_NV; + if (c + MASK_NV <= n) { + MASK_STORE((MASK_V) (MASK_NINF), 0, out + c); + } else { + for (ulong j = c; j < n; ++j) { + out[j] = MASK_NINF; + } + } +} + +kernel void kernel_top_k_mask( + global const char * mask, + ulong offsetm, + global const char * top_k, + ulong offsetk, + global const char * score, + ulong offsets, + global const char * blocks, + ulong offsetb, + global const char * extra, + ulong offsete, + global char * dst, + ulong offsetd, + ulong nem0, + ulong nem1, + ulong nem2, + ulong nem3, + ulong nbm1, + ulong nbm2, + ulong nbm3, + ulong nek0, + ulong nbk0, + ulong nbk1, + ulong nbk2, + ulong nbk3, + ulong nes0, + ulong nbs0, + ulong nbs1, + ulong nbs2, + ulong nbs3, + ulong neb0, + ulong neb1, + ulong nbb0, + ulong nbb1, + ulong nee0, + ulong nbe0, + ulong nbe1, + ulong nbe2, + ulong nbe3, + ulong ned1, + ulong ned2, + ulong nbd1, + ulong nbd2, + ulong nbd3, + int fill +) { + const ulong row = get_group_id(1); + const ulong i1 = row % ned1; + const ulong i2 = (row / ned1) % ned2; + const ulong i3 = row / (ned1*ned2); + const ulong tid = get_local_id(0); + const ulong nth = get_local_size(0); + const ulong n_pick = blocks ? neb1 : score ? nes0 : nem0; + + global const MASK_T * mask_row = (global const MASK_T *) (mask + offsetm + + (i1 % nem1)*nbm1 + (i2 % nem2)*nbm2 + (i3 % nem3)*nbm3); + global MASK_T * dst_row = (global MASK_T *) (dst + offsetd + i1*nbd1 + i2*nbd2 + i3*nbd3); + global const char * top_k_row = top_k + offsetk + i1*nbk1 + i2*nbk2 + i3*nbk3; + + if (fill) { + for (ulong c = tid; c < nem0/MASK_NV; c += nth) { + MASK_STORE((MASK_V) (MASK_NINF), c, dst_row); + } + for (ulong c = nem0/MASK_NV*MASK_NV + tid; c < nem0; c += nth) { + dst_row[c] = MASK_NINF; + } + barrier(CLK_GLOBAL_MEM_FENCE); + } + + // Repeated cells copy the same mask value, as in the CUDA and Metal kernels. + for (ulong j = tid; j < nek0 + nee0; j += nth) { + if (j >= nek0) { + const int c = *(global const int *) (extra + offsete + (j - nek0)*nbe0 + i1*nbe1 + i2*nbe2 + i3*nbe3); + if (c >= 0 && (ulong) c < nem0) { + dst_row[c] = mask_row[c]; + } + continue; + } + + const int p = *(global const int *) (top_k_row + j*nbk0); + if (p < 0 || (ulong) p >= n_pick) { + continue; + } + if (score && !(*(global const float *) (score + offsets + p*nbs0 + i1*nbs1 + i2*nbs2 + i3*nbs3) > -INFINITY)) { + continue; + } + if (blocks) { + for (ulong k = 0; k < neb0; ++k) { + const int c = *(global const int *) (blocks + offsetb + k*nbb0 + p*nbb1); + if (c >= 0 && (ulong) c < nem0) { + dst_row[c] = mask_row[c]; + } + } + } else if ((ulong) p < nem0) { + dst_row[p] = mask_row[p]; + } + } +}