sycl : fixed sycl CLAMP to actually support strided view (#31784)

Modify `ggml_sycl_op_clamp` to use the unary helper function
`ggml_sycl_detail::ggml_sycl_op_unary` to handle non-contigous or
strided views. The fix uses a pattern that matches other stride-safe
unary ops, and is a simpler execution path.

Deleted dead template `unary_op_clamp_kernel`, the flat `clamp` kernel
launcher from element_wise.cpp, and deleted (now) unused
`SYCL_CLAMP_BLOCK_SIZE` from presets.hpp.
This commit is contained in:
Michael Taylor authored and GitHub committed 2026-11-11 13:48:20 +08:00
1 parent aa75ff3239
commit f3b407ca1b
2 files changed
+5 -30

No files matched your search

+5 -29
View File
@@ -361,13 +361,6 @@ static void unary_op_sqr_kernel(const T * x, T * dst, const int k, const sycl::n
}
}
template<typename T>
static void unary_op_clamp_kernel(const T * x, T * dst, const int k, const sycl::nd_item<1> &item_ct1, float min_val, float max_val) {
SYCL_GLOBAL_ID_LOOP(k, item_ct1) {
dst[i] = op_clamp(x[i], min_val, max_val);
}
}
template<typename T>
static void unary_op_ceil_kernel(const T * x, T * dst, const int k, const sycl::nd_item<1> &item_ct1) {
SYCL_GLOBAL_ID_LOOP(k, item_ct1) {
@@ -375,14 +368,6 @@ static void unary_op_ceil_kernel(const T * x, T * dst, const int k, const sycl::
}
}
template<typename T>
static void clamp(const T * x, T * dst, const float min, const float max, const int k,
const sycl::nd_item<1> &item_ct1) {
SYCL_GLOBAL_ID_LOOP(k, item_ct1) {
dst[i] = x[i] < static_cast<T>(min) ? static_cast<T>(min) : (x[i] > static_cast<T>(max) ? static_cast<T>(max) : x[i]);
}
}
template<typename T, typename F>
static void unary_gated_op_flat_kernel(const T * x, const T * g, T * dst, const uint64_t k, const sycl::nd_item<1> & item_ct1, F func) {
SYCL_GLOBAL_ID_LOOP(k, item_ct1) {
@@ -929,20 +914,11 @@ static inline void ggml_sycl_op_sqr(ggml_backend_sycl_context & ctx, ggml_tensor
}
static inline void ggml_sycl_op_clamp(ggml_backend_sycl_context & ctx, ggml_tensor * dst) {
float min_val;
float max_val;
memcpy(&min_val, dst->op_params, sizeof(float));
memcpy(&max_val, (float *) dst->op_params + 1, sizeof(float));
ggml_sycl_detail::dispatch_ggml_sycl_op_unary(ctx, dst,
[](const auto* src, auto* dst_ptr, int k_elements, queue_ptr stream, float min_arg, float max_arg) {
const int num_blocks = ceil_div(k_elements, SYCL_CLAMP_BLOCK_SIZE);
stream->parallel_for(
sycl::nd_range<1>(sycl::range<1>(num_blocks) * sycl::range<1>(SYCL_CLAMP_BLOCK_SIZE),
sycl::range<1>(SYCL_CLAMP_BLOCK_SIZE)),
[=](sycl::nd_item<1> item_ct1) [[sycl::reqd_sub_group_size(WARP_SIZE)]] {
clamp(src, dst_ptr, min_arg, max_arg, k_elements, item_ct1);
});
}, min_val, max_val);
const float min_val = ggml_get_op_params_f32(dst, 0);
const float max_val = ggml_get_op_params_f32(dst, 1);
ggml_sycl_detail::ggml_sycl_op_unary(ctx, dst, [min_val, max_val](auto x) {
return op_clamp(x, min_val, max_val);
});
}
static inline void ggml_sycl_op_xielu(ggml_backend_sycl_context & ctx, ggml_tensor * dst) {
-1
View File
@@ -33,7 +33,6 @@
#define SYCL_SET_BLOCK_SIZE 256
#define SYCL_CPY_BLOCK_SIZE 32
#define SYCL_SCALE_BLOCK_SIZE 256
#define SYCL_CLAMP_BLOCK_SIZE 256
#define SYCL_ROPE_BLOCK_SIZE 256
#define SYCL_ALIBI_BLOCK_SIZE 32
#define SYCL_DIAG_MASK_INF_BLOCK_SIZE 32