cactus-react-native 1.10.3 → 1.12.0

package/ios/cactus.xcframework/ios-arm64/cactus.framework/Headers/graph.h

@@ -122,13 +122,14 @@ enum class Activation {
 enum class OpType {
     INPUT, PRECISION_CAST,
     ADD, ADD_CLIPPED, SUBTRACT, MULTIPLY, DIVIDE,
+    ABS, POW, FLATTEN, VIEW,
     MATMUL, TRANSPOSE, RESHAPE, SLICE, GATHER, EMBEDDING,
     BILINEAR_INTERPOLATION,
     SUM, MEAN, VARIANCE, MIN, MAX,
     RMS_NORM, ROPE, ROPE_GPTJ, SOFTMAX, ATTENTION, ATTENTION_INT8_HYBRID, REL_POS_BIAS, CONV1D_CAUSAL, CONV1D_K3, CONV1D_K7S3, CONV1D, CONV1D_SAME_DEPTHWISE_K9, CONV1D_POINTWISE, CONV2D_K3S2P1, CONV2D_DEPTHWISE_K3S2P1, CONV2D_POINTWISE_1X1, GLU, BATCHNORM,
     SCALAR_ADD, SCALAR_SUBTRACT, SCALAR_MULTIPLY, SCALAR_DIVIDE, SCALAR_EXP, SCALAR_SQRT, SCALAR_COS, SCALAR_SIN, SCALAR_LOG,
     RELU, SILU, GELU, GELU_ERF, SIGMOID, TANH,
-    SAMPLE, CONCAT,
+    SAMPLE, CONCAT, CAT,
     SCATTER_TOPK,
     TOPK, LAYERNORM, GROUPNORM,
     MOE_LAYER,
@@ -136,7 +137,17 @@ enum class OpType {
     PERSISTENT,
     QUANTIZE_ACTIVATIONS,
     LSTM_CELL,
-    STFT
+    GATED_DELTANET_DECODE,
+    GATED_DELTANET_PREFILL,
+    STFT,
+    ALTUP_PREDICT,
+    ALTUP_CORRECT,
+    GAUSSIAN_TOPK,
+    MAXPOOL1D,
+    BILSTM_SEQUENCE,
+    LEAKY_RELU,
+    CONV2D_K3S1P1,
+    STATS_POOL
 };
 
 struct PrecisionTraits {
@@ -315,6 +326,7 @@ struct OpParams {
     size_t window_size = 0;
     bool is_causal = true;  
     bool attention_mask_is_additive = false;
+    float logit_cap = 0.0f;
     std::vector<size_t> new_shape;
     std::vector<size_t> permutation;
     Precision output_precision = Precision::INT8;
@@ -350,6 +362,10 @@ struct OpParams {
     size_t num_kv_heads = 0;
     size_t head_dim = 0;
     size_t num_fft_bins = 0;
+    size_t chunk_size = 0;
+    size_t num_altup_inputs = 0;
+    size_t v_head_dim = 0;
+    size_t kernel_size = 0;
 };
 
 struct GraphNode {
@@ -362,6 +378,28 @@ struct GraphNode {
     GraphNode(size_t node_id, OpType type);
 };
 
+using nodes_vector = std::vector<std::unique_ptr<GraphNode>>;
+using node_index_map_t = std::unordered_map<size_t, size_t>;
+
+inline const BufferDesc& get_input(const GraphNode& node, size_t idx,
+                                   const nodes_vector& nodes,
+                                   const node_index_map_t& node_index_map) {
+    return nodes[node_index_map.at(node.input_ids[idx])]->output_buffer;
+}
+
+struct AxisDims {
+    size_t outer, axis_size, inner;
+    static AxisDims from_shape(const std::vector<size_t>& shape, size_t axis) {
+        AxisDims d;
+        d.outer = 1;
+        for (size_t i = 0; i < axis; i++) d.outer *= shape[i];
+        d.axis_size = shape[axis];
+        d.inner = 1;
+        for (size_t i = axis + 1; i < shape.size(); i++) d.inner *= shape[i];
+        return d;
+    }
+};
+
 template<typename T>
 void dispatch_binary_op(OpType op, const T* lhs, const T* rhs, T* output, size_t count);
 
@@ -383,6 +421,14 @@ void compute_groupnorm_node(GraphNode& node, const std::vector<std::unique_ptr<G
 void compute_persistent_node(GraphNode& node, const std::vector<std::unique_ptr<GraphNode>>& nodes, const std::unordered_map<size_t, size_t>& node_index_map);
 void compute_index_node(GraphNode& node, const std::vector<std::unique_ptr<GraphNode>>& nodes, const std::unordered_map<size_t, size_t>& node_index_map);
 void compute_lstm_cell_node(GraphNode& node, const std::vector<std::unique_ptr<GraphNode>>& nodes, const std::unordered_map<size_t, size_t>& node_index_map);
+void compute_gated_deltanet_decode_node(GraphNode& node, const std::vector<std::unique_ptr<GraphNode>>& nodes, const std::unordered_map<size_t, size_t>& node_index_map);
+void compute_gated_deltanet_prefill_node(GraphNode& node, const std::vector<std::unique_ptr<GraphNode>>& nodes, const std::unordered_map<size_t, size_t>& node_index_map);
+void compute_altup_predict_node(GraphNode& node, const std::vector<std::unique_ptr<GraphNode>>& nodes, const std::unordered_map<size_t, size_t>& node_index_map);
+void compute_altup_correct_node(GraphNode& node, const std::vector<std::unique_ptr<GraphNode>>& nodes, const std::unordered_map<size_t, size_t>& node_index_map);
+void compute_maxpool1d_node(GraphNode& node, const std::vector<std::unique_ptr<GraphNode>>& nodes, const std::unordered_map<size_t, size_t>& node_index_map);
+void compute_bilstm_sequence_node(GraphNode& node, const std::vector<std::unique_ptr<GraphNode>>& nodes, const std::unordered_map<size_t, size_t>& node_index_map);
+void compute_conv2d_k3s1p1_node(GraphNode& node, const std::vector<std::unique_ptr<GraphNode>>& nodes, const std::unordered_map<size_t, size_t>& node_index_map);
+void compute_stats_pool_node(GraphNode& node, const std::vector<std::unique_ptr<GraphNode>>& nodes, const std::unordered_map<size_t, size_t>& node_index_map);
 
 void shrink_thread_local_buffers();
 class BufferPool {
@@ -437,7 +483,6 @@ public:
     size_t multiply(size_t input1, size_t input2);
     size_t divide(size_t input1, size_t input2);
     
-    
     size_t scalar_add(size_t input, float value);
     size_t scalar_subtract(size_t input, float value);
     size_t scalar_multiply(size_t input, float value);
@@ -455,6 +500,11 @@ public:
     size_t sigmoid(size_t input);
     size_t tanh(size_t input);
     size_t glu(size_t input, int axis = -1);
+
+    size_t abs(size_t input);
+    size_t pow(size_t input, float exponent);
+    size_t view(size_t input, const std::vector<size_t>& new_shape);
+    size_t flatten(size_t input, int start_dim = 0, int end_dim = -1);
     
     size_t matmul(size_t input1, size_t input2, bool pretransposed_rhs = false, ComputeBackend backend = ComputeBackend::CPU);
     size_t transpose(size_t input, ComputeBackend backend = ComputeBackend::CPU);
@@ -497,7 +547,9 @@ public:
                      size_t num_experts_per_tok,
                      bool normalize_routing,
                      float epsilon,
-                     float routed_scaling_factor);
+                     float routed_scaling_factor,
+                     Activation activation = Activation::SILU,
+                     size_t per_expert_scale = 0);
     size_t moe_layer(size_t hidden,
                      size_t routing_probs,
                      size_t topk_indices,
@@ -518,13 +570,15 @@ public:
     size_t attention(size_t query, size_t key, size_t value, float scale, size_t position_offset, size_t window_size, ComputeBackend backend = ComputeBackend::CPU);
     size_t attention_masked(size_t query, size_t key, size_t value, size_t mask, float scale,
                             bool is_causal = true, ComputeBackend backend = ComputeBackend::CPU,
-                            bool additive_mask = false, size_t position_offset = 0, size_t window_size = 0);
+                            bool additive_mask = false, size_t position_offset = 0, size_t window_size = 0,
+                            float logit_cap = 0.0f);
     size_t rel_pos_bias(size_t query, size_t relative_key, float scale);
 
     size_t attention_int8_hybrid(size_t query, size_t key_new, size_t value_new, float scale, size_t position_offset,
                                  const int8_t* cached_keys, const int8_t* cached_values,
                                  const float* k_scales, const float* v_scales,
-                                 size_t cache_len, size_t num_kv_heads, size_t head_dim, size_t window_size = 0);
+                                 size_t cache_len, size_t num_kv_heads, size_t head_dim,
+                                 size_t window_size = 0, size_t v_head_dim = 0);
 
     size_t conv1d_causal(size_t input, size_t weight, size_t kernel_size, size_t dilation = 1);
     size_t conv1d_k3(size_t input, size_t weight, size_t stride);
@@ -543,12 +597,30 @@ public:
     size_t conv2d_pointwise_1x1(size_t input, size_t weight, size_t bias);
 
     size_t lstm_cell(size_t input, size_t h_prev, size_t c_prev, size_t weight_ih, size_t weight_hh, size_t bias_ih, size_t bias_hh);
+    size_t gated_deltanet_decode(size_t query, size_t key, size_t value, size_t gate_log, size_t beta,
+                                 size_t initial_state, float scale = 0.0f);
+    size_t gated_deltanet_prefill(size_t query, size_t key, size_t value, size_t gate_log, size_t beta,
+                                  size_t initial_state, size_t chunk_size = 64, float scale = 0.0f);
     size_t stft(size_t input, size_t weight, size_t stride, size_t num_fft_bins);
 
+    size_t altup_predict(size_t coefs, const size_t* streams, size_t num_streams);
+    size_t altup_correct(size_t coefs, size_t innovation, const size_t* predictions, size_t num_predictions);
+
+    size_t gaussian_topk(size_t input, float ppf);
+
+    size_t maxpool1d(size_t input, size_t kernel_size, size_t stride);
+    size_t leaky_relu(size_t input, float negative_slope = 0.01f);
+    size_t bilstm_sequence(size_t input, size_t w_ih_fwd, size_t w_hh_fwd, size_t b_ih_fwd, size_t b_hh_fwd,
+                           size_t w_ih_bwd, size_t w_hh_bwd, size_t b_ih_bwd, size_t b_hh_bwd);
+    size_t conv2d_k3s1p1(size_t input, size_t weight);
+    size_t conv2d_k3s1p1(size_t input, size_t weight, size_t bias);
+    size_t stats_pool(size_t input);
+
     size_t sample(size_t logits, float temperature = 0.6f, float top_p = 0.95f, size_t top_k = 20,
                   const std::unordered_map<uint32_t, float>& logit_bias = {});
     
     size_t concat(size_t input1, size_t input2, int axis = 0);
+    size_t cat(const std::vector<size_t>& inputs, int axis);
     size_t scatter_topk(size_t indices, size_t values, size_t num_classes);
     
     void set_input(size_t node_id, const void* data, Precision precision);
@@ -653,4 +725,4 @@ namespace GraphFile {
     };
 }
 
-#endif
+#endif

package/ios/cactus.xcframework/ios-arm64/cactus.framework/Headers/kernel.h

@@ -11,7 +11,9 @@ enum class ScalarOpType {
     SUBTRACT,
     MULTIPLY,
     DIVIDE,
+    ABS,
     EXP,
+    POW,
     SQRT,
     COS,
     SIN,
@@ -54,6 +56,14 @@ void cactus_matmul_int8(const int8_t* A, const float* A_scales,
                         const int8_t* B, const __fp16* B_scales,
                         __fp16* C, size_t M, size_t K, size_t N, size_t group_size);
 
+void cactus_gemv_int8_i8mm(const int8_t* A, float A_scale,
+                            const int8_t* B, const __fp16* B_scales,
+                            __fp16* C, size_t K, size_t N, size_t group_size);
+
+void cactus_gemm_int8_i8mm(const int8_t* A, const float* A_scales,
+                            const int8_t* B, const __fp16* B_scales,
+                            __fp16* C, size_t M, size_t K, size_t N, size_t group_size);
+
 void cactus_gemv_int4(const int8_t* A, float A_scale,
                       const int8_t* B_packed, const __fp16* B_scales,
                       __fp16* C, size_t K, size_t N, size_t group_size);
@@ -97,6 +107,9 @@ void cactus_max_axis_f16(const __fp16* input, __fp16* output, size_t outer_size,
 void cactus_rms_norm_f16(const __fp16* input, const __fp16* weight, __fp16* output,
                           size_t batch_size, size_t dims, float eps);
 
+void cactus_layer_norm_f16(const __fp16* input, const __fp16* weight, const __fp16* bias,
+                            __fp16* output, size_t batch_size, size_t dims, float eps);
+
 void cactus_rope_f16(const __fp16* input, __fp16* output, size_t batch_size, size_t seq_len,
                       size_t num_heads, size_t head_dim, size_t start_pos, float theta);
 
@@ -108,6 +121,8 @@ void cactus_softmax_f16(const __fp16* input, __fp16* output, size_t batch_size,
 
 void cactus_relu_f16(const __fp16* input, __fp16* output, size_t num_elements);
 
+void cactus_leaky_relu_f16(const __fp16* input, __fp16* output, size_t num_elements, float negative_slope);
+
 void cactus_silu_f16(const __fp16* input, __fp16* output, size_t num_elements);
 
 void cactus_gelu_f16(const __fp16* input, __fp16* output, size_t num_elements);
@@ -163,21 +178,54 @@ void cactus_batchnorm_f32(
 void cactus_attention_f16(const __fp16* queries, const __fp16* keys, const __fp16* values, __fp16* output,
                           size_t batch_size, size_t seq_len, size_t kv_seq_len, size_t num_q_heads, size_t num_kv_heads,
                           size_t head_dim, float scale, const __fp16* mask, size_t position_offset = 0, size_t window_size = 0,
-                          bool is_causal = true, bool mask_is_additive = false, bool mask_per_head = false);
+                          bool is_causal = true, bool mask_is_additive = false, bool mask_per_head = false,
+                          size_t v_head_dim = 0, float logit_cap = 0.0f);
 
 void cactus_attention_hybrid_int8_fp16(
-    const __fp16* queries,        
-    const int8_t* keys_cached, 
-    const int8_t* values_cached, 
+    const __fp16* queries,
+    const int8_t* keys_cached,
+    const int8_t* values_cached,
     const float* k_scales,
-    const float* v_scales, 
-    const __fp16* keys_new, 
-    const __fp16* values_new, 
+    const float* v_scales,
+    const __fp16* keys_new,
+    const __fp16* values_new,
     __fp16* output,
     size_t batch_size, size_t seq_len, size_t cache_len, size_t new_len,
     size_t num_q_heads, size_t num_kv_heads, size_t head_dim,
     float scale, size_t position_offset = 0, bool is_causal = true, size_t window_size = 0,
-    size_t group_size = KV_QUANT_GROUP_SIZE);
+    size_t group_size = KV_QUANT_GROUP_SIZE, size_t v_head_dim = 0);
+
+void cactus_gated_deltanet_decode_f16(
+    const __fp16* q_data,
+    const __fp16* k_data,
+    const __fp16* v_data,
+    const __fp16* g_data,
+    const __fp16* b_data,
+    const __fp16* s_data,
+    __fp16* out,
+    size_t B,
+    size_t Hq,
+    size_t Hv,
+    size_t K,
+    size_t V,
+    float scale);
+
+void cactus_gated_deltanet_prefill_f16(
+    const __fp16* q_data,
+    const __fp16* k_data,
+    const __fp16* v_data,
+    const __fp16* g_data,
+    const __fp16* b_data,
+    const __fp16* s_data,
+    __fp16* out,
+    size_t B,
+    size_t T,
+    size_t Hq,
+    size_t Hv,
+    size_t K,
+    size_t V,
+    size_t requested_chunk_size,
+    float scale);
 
 void cactus_conv1d_causal_depthwise_f16(
     const __fp16* input,
@@ -244,6 +292,18 @@ void cactus_conv1d_same_depthwise_f16_k9(
     size_t C
 );
 
+void cactus_conv2d_f16_k3s1p1_nchw(
+    const __fp16* input,
+    const __fp16* weight,
+    const __fp16* bias,
+    __fp16* output,
+    size_t N,
+    size_t C_in,
+    size_t H,
+    size_t W,
+    size_t C_out
+);
+
 void cactus_conv2d_f16_k3s2p1_nchw(
     const __fp16* input,
     const __fp16* weight,
@@ -305,6 +365,8 @@ void cactus_sample_f16(const __fp16* logits, uint32_t* output, size_t vocab_size
 void cactus_concat_f16(const __fp16* input1, const __fp16* input2, __fp16* output,
                        const size_t* shape1, const size_t* shape2, const size_t* output_shape,
                        size_t ndims, int axis);
+void cactus_cat_f16(const __fp16** inputs, __fp16* output, const size_t** input_shapes,
+                      const size_t* output_shape, size_t num_inputs, size_t rank, int axis);
 
 void cactus_int8_to_fp32(const int8_t* src, float* dst, size_t count, float scale = 1.0f);
 void cactus_fp32_to_int8(const float* src, int8_t* dst, size_t count, float scale = 1.0f);
@@ -328,6 +390,30 @@ inline size_t kv_scales_count(size_t seq_len, size_t kv_heads, size_t head_dim,
 
 void cactus_unpack_int4_to_int8(const uint8_t* packed, int8_t* unpacked, size_t unpacked_count);
 
+void cactus_gaussian_topk_f16(
+    const __fp16* input,
+    __fp16* output,
+    size_t rows,
+    size_t cols,
+    float ppf);
+
+void cactus_altup_predict_f16(
+    const __fp16* coefs,
+    const __fp16* const* streams,
+    __fp16* output,
+    size_t n,
+    size_t seq_len,
+    size_t hidden_dim);
+
+void cactus_altup_correct_f16(
+    const __fp16* coefs,
+    const __fp16* innovation,
+    const __fp16* const* predictions,
+    __fp16* output,
+    size_t n,
+    size_t seq_len,
+    size_t hidden_dim);
+
 void cactus_lstm_cell_f16(
     const __fp16* x_input,
     const __fp16* h_prev,
@@ -343,4 +429,31 @@ void cactus_lstm_cell_f16(
     size_t hidden_size
 );
 
-#endif
+void cactus_bilstm_sequence_f16(
+    const __fp16* input,
+    const __fp16* weight_ih_fwd,
+    const __fp16* weight_hh_fwd,
+    const __fp16* bias_ih_fwd,
+    const __fp16* bias_hh_fwd,
+    const __fp16* weight_ih_bwd,
+    const __fp16* weight_hh_bwd,
+    const __fp16* bias_ih_bwd,
+    const __fp16* bias_hh_bwd,
+    __fp16* output,
+    size_t batch_size,
+    size_t seq_len,
+    size_t input_size,
+    size_t hidden_size
+);
+
+void cactus_maxpool1d_f16(
+    const __fp16* input,
+    __fp16* output,
+    size_t batch_size,
+    size_t channels,
+    size_t input_length,
+    size_t kernel_size,
+    size_t stride
+);
+
+#endif

package/ios/cactus.xcframework/ios-arm64/cactus.framework/Headers/kernel_utils.h

@@ -9,6 +9,8 @@
 #if defined(__ANDROID__)
 #include <sys/auxv.h>
 #include <asm/hwcap.h>
+#include <sched.h>
+#include <fstream>
 #endif
 #include <algorithm>
 #include <cmath>
@@ -44,6 +46,29 @@ inline void stream_store_f16x8(__fp16* dst, float16x8_t val) {
 #endif
 }
 
+inline bool cpu_has_i8mm() {
+#if defined(__aarch64__)
+    static std::once_flag once;
+    static bool has = false;
+
+    std::call_once(once, []() {
+#if defined(__APPLE__)
+    has = true;
+#elif defined(__ANDROID__)
+    unsigned long hwcap2 = getauxval(AT_HWCAP2);
+    #ifndef HWCAP2_I8MM
+    #define HWCAP2_I8MM (1 << 13)
+    #endif
+    has = (hwcap2 & HWCAP2_I8MM) != 0;
+#endif
+    });
+
+    return has;
+#else
+    return false;
+#endif
+}
+
 inline bool cpu_has_sme2() {
 #if defined(__aarch64__)
 	static std::once_flag once;
@@ -130,6 +155,33 @@ inline float32x4_t fast_tanh_f32x4(float32x4_t x) {
     return result;
 }
 
+constexpr size_t SIMD_F16_WIDTH = 8;
+
+inline size_t simd_align(size_t count, size_t width = SIMD_F16_WIDTH) {
+    return (count / width) * width;
+}
+
+inline void f16x8_split_f32(float16x8_t v, float32x4_t& lo, float32x4_t& hi) {
+    lo = vcvt_f32_f16(vget_low_f16(v));
+    hi = vcvt_f32_f16(vget_high_f16(v));
+}
+
+inline float16x8_t f32_merge_f16(float32x4_t lo, float32x4_t hi) {
+    return vcombine_f16(vcvt_f16_f32(lo), vcvt_f16_f32(hi));
+}
+
+inline float32x4_t fast_sigmoid_f32x4(float32x4_t x) {
+    const float32x4_t one = vdupq_n_f32(1.0f);
+    return vdivq_f32(one, vaddq_f32(one, fast_exp_f32x4(vnegq_f32(x))));
+}
+
+template<typename F32x4Op>
+inline float16x8_t apply_f32_op_on_f16x8(float16x8_t v, F32x4Op op) {
+    float32x4_t lo, hi;
+    f16x8_split_f32(v, lo, hi);
+    return f32_merge_f16(op(lo), op(hi));
+}
+
 inline void unpack_int4_as_int8x16x2(const uint8_t* ptr, int8x16_t& high_decoded, int8x16_t& low_decoded) {
     int8x16_t packed = vreinterpretq_s8_u8(vld1q_u8(ptr));
     high_decoded = vshrq_n_s8(packed, 4);
@@ -138,6 +190,80 @@ inline void unpack_int4_as_int8x16x2(const uint8_t* ptr, int8x16_t& high_decoded
 
 namespace CactusThreading {
 
+#if defined(__ANDROID__)
+    struct CoreTopology {
+        std::vector<int> performance_cores;  
+        std::vector<int> all_cores;
+
+        static CoreTopology& get() {
+            static CoreTopology topo = detect();
+            return topo;
+        }
+
+    private:
+        static int read_sysfs_int(const char* path) {
+            std::ifstream f(path);
+            if (!f.is_open()) return -1;
+            int val = -1;
+            f >> val;
+            return val;
+        }
+
+        static CoreTopology detect() {
+            CoreTopology topo;
+            constexpr int MAX_CPUS = 16;
+            std::vector<std::pair<int, int>> core_caps; 
+
+            for (int i = 0; i < MAX_CPUS; ++i) {
+                char path[128];
+
+                snprintf(path, sizeof(path),
+                         "/sys/devices/system/cpu/cpu%d/cpu_capacity", i);
+                int cap = read_sysfs_int(path);
+                if (cap > 0) {
+                    core_caps.push_back({i, cap});
+                    topo.all_cores.push_back(i);
+                    continue;
+                }
+
+                snprintf(path, sizeof(path),
+                         "/sys/devices/system/cpu/cpu%d/cpufreq/cpuinfo_max_freq", i);
+                int freq = read_sysfs_int(path);
+                if (freq > 0) {
+                    core_caps.push_back({i, freq});
+                    topo.all_cores.push_back(i);
+                }
+            }
+
+            if (core_caps.empty()) return topo;
+
+            int max_cap = 0;
+            for (auto& [id, cap] : core_caps) {
+                max_cap = std::max(max_cap, cap);
+            }
+
+            int threshold = static_cast<int>(max_cap * 0.70);
+            for (auto& [id, cap] : core_caps) {
+                if (cap >= threshold) {
+                    topo.performance_cores.push_back(id);
+                }
+            }
+
+            return topo;
+        }
+    };
+
+    inline bool pin_current_thread_to_cores(const std::vector<int>& cores) {
+        if (cores.empty()) return false;
+        cpu_set_t mask;
+        CPU_ZERO(&mask);
+        for (int core : cores) {
+            CPU_SET(core, &mask);
+        }
+        return sched_setaffinity(0, sizeof(mask), &mask) == 0;
+    }
+#endif
+
     class ThreadPool {
     private:
         static constexpr size_t MAX_WORKERS = 16;
@@ -184,9 +310,25 @@ namespace CactusThreading {
             : stop(false), pending_tasks(0) {
             num_workers_ = std::min(num_threads, MAX_WORKERS);
             if (num_workers_ == 0) num_workers_ = 1;
+
+#if defined(__ANDROID__)
+            auto& topo = CoreTopology::get();
+            if (!topo.performance_cores.empty()) {
+                num_workers_ = std::min(num_workers_, topo.performance_cores.size());
+            }
+#endif
+
             workers.reserve(num_workers_);
             for (size_t i = 0; i < num_workers_; ++i) {
-                workers.emplace_back(&ThreadPool::worker_thread, this);
+                workers.emplace_back([this]() {
+#if defined(__ANDROID__)
+                    auto& perf = CoreTopology::get().performance_cores;
+                    if (!perf.empty()) {
+                        pin_current_thread_to_cores(perf);
+                    }
+#endif
+                    worker_thread();
+                });
             }
         }
 
@@ -498,5 +640,52 @@ namespace CactusThreading {
 
 }
 
+template<typename SimdOp, typename ScalarOp>
+void elementwise_op_f16(const __fp16* input, __fp16* output, size_t num_elements,
+                        bool use_streaming, CactusThreading::ParallelConfig config,
+                        SimdOp simd_op, ScalarOp scalar_op, size_t unroll = 4) {
+    CactusThreading::parallel_for(num_elements, config,
+        [&](size_t start, size_t end) {
+            const size_t n = end - start;
+            const size_t vec_end = start + simd_align(n);
+
+            if (use_streaming && unroll >= 4) {
+                const size_t unrolled_end = start + simd_align(n, SIMD_F16_WIDTH * 4);
+                for (size_t i = start; i < unrolled_end; i += SIMD_F16_WIDTH * 4) {
+                    __builtin_prefetch(&input[i + 256], 0, 0);
+                    float16x8_t v0 = simd_op(vld1q_f16(&input[i]));
+                    float16x8_t v1 = simd_op(vld1q_f16(&input[i + 8]));
+                    float16x8_t v2 = simd_op(vld1q_f16(&input[i + 16]));
+                    float16x8_t v3 = simd_op(vld1q_f16(&input[i + 24]));
+                    stream_store_f16x8(&output[i], v0);
+                    stream_store_f16x8(&output[i + 8], v1);
+                    stream_store_f16x8(&output[i + 16], v2);
+                    stream_store_f16x8(&output[i + 24], v3);
+                }
+                for (size_t i = unrolled_end; i < vec_end; i += SIMD_F16_WIDTH) {
+                    stream_store_f16x8(&output[i], simd_op(vld1q_f16(&input[i])));
+                }
+            } else if (use_streaming && unroll >= 2) {
+                const size_t unrolled_end = start + simd_align(n, SIMD_F16_WIDTH * 2);
+                for (size_t i = start; i < unrolled_end; i += SIMD_F16_WIDTH * 2) {
+                    __builtin_prefetch(&input[i + 128], 0, 0);
+                    float16x8_t v0 = simd_op(vld1q_f16(&input[i]));
+                    float16x8_t v1 = simd_op(vld1q_f16(&input[i + 8]));
+                    stream_store_f16x8(&output[i], v0);
+                    stream_store_f16x8(&output[i + 8], v1);
+                }
+                for (size_t i = unrolled_end; i < vec_end; i += SIMD_F16_WIDTH) {
+                    stream_store_f16x8(&output[i], simd_op(vld1q_f16(&input[i])));
+                }
+            } else {
+                for (size_t i = start; i < vec_end; i += SIMD_F16_WIDTH) {
+                    vst1q_f16(&output[i], simd_op(vld1q_f16(&input[i])));
+                }
+            }
+            for (size_t i = vec_end; i < end; ++i) {
+                output[i] = scalar_op(input[i]);
+            }
+        });
+}
 
-#endif // KERNEL_UTILS_H 
+#endif // KERNEL_UTILS_H