ggml-hexagon: add HTP unary ops for ABS and LOG

Add HVX-accelerated implementations for GGML_OP_LOG and
GGML_UNARY_OP_ABS on the HTP backend.

- Register HTP_OP_UNARY_ABS and HTP_OP_UNARY_LOG in op_remap_to_htp()
- Add ABS and LOG to ggml_backend_hexagon_device_supports_op()
- Implement hvx_abs_f32_aa() in hvx-arith.h using hvx_vec_abs_f32()
- Implement hvx_log_f32_aa() in hvx-log.h using hvx_vec_log_f32()
- Add abs_f32() and log_f32() row-wise dispatch in unary-ops.c
- Define tiled and non-tiled task functions via DEFINE_UNARY_TASK and
  DEFINE_UNARY_TILED_TASK macros
- Route HTP_OP_UNARY_ABS and HTP_OP_UNARY_LOG through execute_op()
  in main.c

Author: cqderek

ggml/src/ggml-hexagon/ggml-hexagon.cpp

@@ -3453,6 +3453,7 @@ static htp_op_code op_remap_to_htp(const ggml_tensor * t) {
         case GGML_OP_CLAMP:           return HTP_OP_CLAMP;
         case GGML_OP_SQR:             return HTP_OP_SQR;
         case GGML_OP_SQRT:            return HTP_OP_SQRT;
+        case GGML_OP_LOG:             return HTP_OP_UNARY_LOG;
         case GGML_OP_SOFT_MAX:        return HTP_OP_SOFTMAX;
         case GGML_OP_SSM_CONV:        return HTP_OP_SSM_CONV;
         case GGML_OP_GATED_DELTA_NET: return HTP_OP_GATED_DELTA_NET;
@@ -3476,6 +3477,7 @@ static htp_op_code op_remap_to_htp(const ggml_tensor * t) {
                 case GGML_UNARY_OP_EXP:        return HTP_OP_UNARY_EXP;
                 case GGML_UNARY_OP_SOFTPLUS:   return HTP_OP_UNARY_SOFTPLUS;
                 case GGML_UNARY_OP_TANH:       return HTP_OP_UNARY_TANH;
+                case GGML_UNARY_OP_ABS:        return HTP_OP_UNARY_ABS;
             default:
                 break;
             }
@@ -4112,6 +4114,7 @@ static bool ggml_backend_hexagon_device_supports_op(ggml_backend_dev_t dev, cons
 
         case GGML_OP_SQR:
         case GGML_OP_SQRT:
+        case GGML_OP_LOG:
             supp = ggml_hexagon_supported_unary(sess, op);
             break;
 
@@ -4130,6 +4133,7 @@ static bool ggml_backend_hexagon_device_supports_op(ggml_backend_dev_t dev, cons
                 case GGML_UNARY_OP_SIGMOID:
                 case GGML_UNARY_OP_SOFTPLUS:
                 case GGML_UNARY_OP_TANH:
+                case GGML_UNARY_OP_ABS:
                 case GGML_UNARY_OP_SILU:
                 case GGML_UNARY_OP_GELU:
                 case GGML_UNARY_OP_GELU_QUICK:

ggml/src/ggml-hexagon/htp/htp-ops.h

@@ -70,6 +70,8 @@ enum htp_op_code {
     HTP_OP_UNARY_NEG,
     HTP_OP_UNARY_SOFTPLUS,
     HTP_OP_UNARY_TANH,
+    HTP_OP_UNARY_ABS,
+    HTP_OP_UNARY_LOG,
     HTP_OP_GLU_SWIGLU,
     HTP_OP_GLU_SWIGLU_OAI,
     HTP_OP_GLU_GEGLU,

ggml/src/ggml-hexagon/htp/hvx-arith.h

@@ -358,6 +358,34 @@ static inline void hvx_clamp_scalar_f32(uint8_t * restrict dst, const uint8_t *
     }
 }
 
+//
+// Abs
+//
+
+static inline void hvx_abs_f32_aa(uint8_t * restrict dst, const uint8_t * restrict src, uint32_t n) {
+    assert((unsigned long) dst % 128 == 0);
+    assert((unsigned long) src % 128 == 0);
+
+    HVX_Vector * restrict vdst = (HVX_Vector *) dst;
+    HVX_Vector * restrict vsrc = (HVX_Vector *) src;
+
+    const uint32_t elem_size = sizeof(float);
+    const uint32_t epv       = 128 / elem_size;
+    const uint32_t nvec      = n / epv;
+    const uint32_t nloe      = n % epv;
+
+    uint32_t i = 0;
+
+    _Pragma("unroll(4)")
+    for (; i < nvec; i++) {
+        vdst[i] = hvx_vec_abs_f32(vsrc[i]);
+    }
+    if (nloe) {
+        HVX_Vector v = hvx_vec_abs_f32(vsrc[i]);
+        hvx_vec_store_a((void *) &vdst[i], nloe * elem_size, v);
+    }
+}
+
 //
 // Square
 //

ggml/src/ggml-hexagon/htp/hvx-log.h

@@ -62,4 +62,28 @@ static inline HVX_Vector hvx_vec_log_f32(HVX_Vector x) {
     return hvx_vec_add_f32_f32(term_e, res);
 }
 
+static inline void hvx_log_f32_aa(uint8_t * restrict dst, const uint8_t * restrict src, uint32_t n) {
+    assert((unsigned long) dst % 128 == 0);
+    assert((unsigned long) src % 128 == 0);
+
+    HVX_Vector * restrict vdst = (HVX_Vector *) dst;
+    HVX_Vector * restrict vsrc = (HVX_Vector *) src;
+
+    const uint32_t elem_size = sizeof(float);
+    const uint32_t epv       = 128 / elem_size;
+    const uint32_t nvec      = n / epv;
+    const uint32_t nloe      = n % epv;
+
+    uint32_t i = 0;
+
+    _Pragma("unroll(4)")
+    for (; i < nvec; i++) {
+        vdst[i] = hvx_vec_log_f32(vsrc[i]);
+    }
+    if (nloe) {
+        HVX_Vector v = hvx_vec_log_f32(vsrc[i]);
+        hvx_vec_store_a((void *) &vdst[i], nloe * elem_size, v);
+    }
+}
+
 #endif /* HVX_LOG_H */

ggml/src/ggml-hexagon/htp/main.c

@@ -728,6 +728,8 @@ static int execute_op(struct htp_ops_context * octx) {
         case HTP_OP_UNARY_NEG:
         case HTP_OP_UNARY_EXP:
         case HTP_OP_UNARY_TANH:
+        case HTP_OP_UNARY_ABS:
+        case HTP_OP_UNARY_LOG:
         case HTP_OP_L2_NORM:
             return op_unary(octx);
 

ggml/src/ggml-hexagon/htp/unary-ops.c

@@ -443,6 +443,34 @@ static void tanh_f32(const float * restrict src,
     }
 }
 
+static void abs_f32(const float * restrict src,
+                    float * restrict dst,
+                    const uint32_t num_rows,
+                    const struct htp_unary_context * uctx) {
+    htp_unary_op_preamble;
+
+    for (uint32_t ir = 0; ir < num_rows; ir++) {
+        const uint8_t * restrict src_local = (const uint8_t *)src + (ir * src0_row_size_aligned);
+        uint8_t * restrict dst_local       = (uint8_t *)dst + (ir * dst_row_size_aligned);
+
+        hvx_abs_f32_aa(dst_local, src_local, ne0);
+    }
+}
+
+static void log_f32(const float * restrict src,
+                    float * restrict dst,
+                    const uint32_t num_rows,
+                    const struct htp_unary_context * uctx) {
+    htp_unary_op_preamble;
+
+    for (uint32_t ir = 0; ir < num_rows; ir++) {
+        const uint8_t * restrict src_local = (const uint8_t *)src + (ir * src0_row_size_aligned);
+        uint8_t * restrict dst_local       = (uint8_t *)dst + (ir * dst_row_size_aligned);
+
+        hvx_log_f32_aa(dst_local, src_local, ne0);
+    }
+}
+
 #define DEFINE_UNARY_TASK(NAME, IS_RMS_NORM_MUL, IS_TRI, CORE_EXPR)                                                 \
 static void unary_task_f32_##NAME(unsigned int nth, unsigned int ith, void * data) {                                \
     const struct htp_unary_context * uctx = (const struct htp_unary_context *) data;                                \
@@ -603,6 +631,8 @@ DEFINE_UNARY_TASK(unary_silu,     false, false, silu_f32(src0_vtcm, dst_vtcm, bl
 DEFINE_UNARY_TASK(unary_gelu,     false, false, gelu_f32(src0_vtcm, dst_vtcm, block_size, uctx))
 DEFINE_UNARY_TASK(unary_softplus, false, false, softplus_f32(src0_vtcm, dst_vtcm, block_size, uctx))
 DEFINE_UNARY_TASK(unary_tanh,     false, false, tanh_f32(src0_vtcm, dst_vtcm, block_size, uctx))
+DEFINE_UNARY_TASK(unary_abs,      false, false, abs_f32(src0_vtcm, dst_vtcm, block_size, uctx))
+DEFINE_UNARY_TASK(unary_log,      false, false, log_f32(src0_vtcm, dst_vtcm, block_size, uctx))
 DEFINE_UNARY_TASK(l2_norm,        false, false, l2_norm_f32(src0_vtcm, dst_vtcm, block_size, uctx))
 DEFINE_UNARY_TASK(tri,            false, true,  tri_f32(src0_vtcm, dst_vtcm, block_size, ir, uctx))
 
@@ -850,6 +880,8 @@ DEFINE_UNARY_TILED_TASK(unary_silu,     false, tile_silu_f32(dst_vtcm, src_vtcm,
 DEFINE_UNARY_TILED_TASK(unary_gelu,     false, tile_gelu_f32(dst_vtcm, src_vtcm, tw))
 DEFINE_UNARY_TILED_TASK(unary_softplus, false, tile_unary_softplus_f32(dst_vtcm, src_vtcm, tw))
 DEFINE_UNARY_TILED_TASK(unary_tanh,     false, hvx_tanh_f32_aa(dst_vtcm, src_vtcm, tw))
+DEFINE_UNARY_TILED_TASK(unary_abs,      false, hvx_abs_f32_aa(dst_vtcm, src_vtcm, tw))
+DEFINE_UNARY_TILED_TASK(unary_log,      false, hvx_log_f32_aa(dst_vtcm, src_vtcm, tw))
 DEFINE_UNARY_TILED_TASK(tri,            true,  tri_apply_tile_f32(src_vtcm, dst_vtcm, tw, col, i01, ne0, tri_ttype))
 
 static int execute_op_unary_f32(struct htp_ops_context * octx) {
@@ -875,6 +907,8 @@ static int execute_op_unary_f32(struct htp_ops_context * octx) {
         case HTP_OP_UNARY_GELU:      op_type = "gelu-f32";         break;
         case HTP_OP_UNARY_SOFTPLUS:  op_type = "softplus-f32";     break;
         case HTP_OP_UNARY_TANH:      op_type = "tanh-f32";         break;
+        case HTP_OP_UNARY_ABS:       op_type = "abs-f32";          break;
+        case HTP_OP_UNARY_LOG:       op_type = "log-f32";          break;
         case HTP_OP_L2_NORM:         op_type = "l2norm-f32";       break;
         case HTP_OP_TRI:             op_type = "tri-f32";          break;
 
@@ -973,6 +1007,8 @@ static int execute_op_unary_f32(struct htp_ops_context * octx) {
                 case HTP_OP_UNARY_GELU:      task_func = unary_task_f32_tiled_unary_gelu;     break;
                 case HTP_OP_UNARY_SOFTPLUS:  task_func = unary_task_f32_tiled_unary_softplus; break;
                 case HTP_OP_UNARY_TANH:      task_func = unary_task_f32_tiled_unary_tanh;     break;
+                case HTP_OP_UNARY_ABS:       task_func = unary_task_f32_tiled_unary_abs;      break;
+                case HTP_OP_UNARY_LOG:       task_func = unary_task_f32_tiled_unary_log;      break;
                 case HTP_OP_TRI:             task_func = unary_task_f32_tiled_tri;            break;
                 default:                     break;
             }
@@ -992,6 +1028,8 @@ static int execute_op_unary_f32(struct htp_ops_context * octx) {
                 case HTP_OP_UNARY_GELU:      task_func = unary_task_f32_unary_gelu;           break;
                 case HTP_OP_UNARY_SOFTPLUS:  task_func = unary_task_f32_unary_softplus;       break;
                 case HTP_OP_UNARY_TANH:      task_func = unary_task_f32_unary_tanh;           break;
+                case HTP_OP_UNARY_ABS:       task_func = unary_task_f32_unary_abs;            break;
+                case HTP_OP_UNARY_LOG:       task_func = unary_task_f32_unary_log;            break;
                 case HTP_OP_L2_NORM:         task_func = unary_task_f32_l2_norm;              break;
                 case HTP_OP_TRI:             task_func = unary_task_f32_tri;                  break;
                 default:                     break;

ggml/src/ggml-hexagon/htp/unary-ops.h

@@ -55,6 +55,8 @@ static inline bool htp_op_is_unary(uint32_t opcode) {
         case HTP_OP_UNARY_GELU:
         case HTP_OP_UNARY_SOFTPLUS:
         case HTP_OP_UNARY_TANH:
+        case HTP_OP_UNARY_ABS:
+        case HTP_OP_UNARY_LOG:
         case HTP_OP_L2_NORM:
         case HTP_OP_TRI:
             return true;

tests/test-backend-ops.cpp

@@ -2469,8 +2469,13 @@ struct test_set_rows : public test_case {
     // See dicussion here: https://github.com/ggml-org/llama.cpp/pull/23760#issuecomment-4566312209
     double max_nmse_err(ggml_backend_t backend) override {
         ggml_backend_reg_t reg = ggml_backend_dev_backend_reg(ggml_backend_get_device(backend));
-        if (type_dst == GGML_TYPE_Q8_0 && strcmp(ggml_backend_reg_name(reg), "WebGPU") == 0) {
-            return std::max(test_case::max_nmse_err(backend), 2e-7);
+        if (type_dst == GGML_TYPE_Q8_0) {
+            if (strcmp(ggml_backend_reg_name(reg), "WebGPU") == 0) {
+                return std::max(test_case::max_nmse_err(backend), 2e-7);
+            }
+            if (strcmp(ggml_backend_reg_name(reg), "HTP") == 0) {
+                return std::max(test_case::max_nmse_err(backend), 5e-6);
+            }
         }
         return test_case::max_nmse_err(backend);
     }
@@ -4120,9 +4125,10 @@ struct test_ssm_scan : public test_case {
     const int64_t n_seqs;
     const bool    xbc_overlap;
     const int64_t K;
+    const bool    weak_decay;
 
     std::string vars() override {
-        return VARS_TO_STR9(type, d_state, head_dim, n_head, n_group, n_seq_tokens, n_seqs, xbc_overlap, K);
+        return VARS_TO_STR10(type, d_state, head_dim, n_head, n_group, n_seq_tokens, n_seqs, xbc_overlap, K, weak_decay);
     }
 
     test_ssm_scan(ggml_type type = GGML_TYPE_F32,
@@ -4133,8 +4139,9 @@ struct test_ssm_scan : public test_case {
             int64_t n_seq_tokens = 32,
             int64_t n_seqs = 32,
             bool xbc_overlap = false,
-            int64_t K = 1)
-        : type(type), d_state(d_state), head_dim(head_dim), n_head(n_head), n_group(n_group), n_seq_tokens(n_seq_tokens), n_seqs(n_seqs), xbc_overlap(xbc_overlap), K(K) {}
+            int64_t K = 1,
+            bool weak_decay = false)
+        : type(type), d_state(d_state), head_dim(head_dim), n_head(n_head), n_group(n_group), n_seq_tokens(n_seq_tokens), n_seqs(n_seqs), xbc_overlap(xbc_overlap), K(K), weak_decay(weak_decay) {}
 
     double max_nmse_err() override {
         // SSD path (head_dim > 1) uses FP16 intermediates (M matrix, X_dt); Mamba-1 is pure FP32.
@@ -4187,7 +4194,7 @@ struct test_ssm_scan : public test_case {
                 continue;
             } else if (t->ne[1] == n_head && t->ne[2] == 1) {
                 // A {1 or d_state, n_head}: negative decay (2-D tensor, ne[2]==1 distinguishes from 3-D/4-D tensors)
-                init_tensor_uniform(t, -1.0f, -0.5f);
+                init_tensor_uniform(t, weak_decay ? -0.02f : -1.0f, weak_decay ? -0.005f : -0.5f);
             } else {
                 init_tensor_uniform(t);
             }
@@ -9111,6 +9118,10 @@ static std::vector<std::unique_ptr<test_case>> make_test_cases_eval() {
     test_cases.emplace_back(new test_ssm_scan(GGML_TYPE_F32, 128, 64, 16, 2, 4, 2, false, /*K=*/4)); // Mamba-2 rollback snapshots
     test_cases.emplace_back(new test_ssm_scan(GGML_TYPE_F32, 128, 64, 16, 2, 8, 2, false, /*K=*/3)); // Mamba-2 rollback overflow
     test_cases.emplace_back(new test_ssm_scan_rollback(GGML_TYPE_F32, 128, 64, 16, 2, 8, 2, /*K=*/3)); // rollback snapshots match prefix states
+    test_cases.emplace_back(new test_ssm_scan(GGML_TYPE_F32, 128, 64, 16, 2, 64, 4)); // Metal SSD one chunk MMA only, no seq tail
+    test_cases.emplace_back(new test_ssm_scan(GGML_TYPE_F32, 128, 64, 16, 2, 65, 2)); // SSD one chunk + 1-token sequential tail
+    test_cases.emplace_back(new test_ssm_scan(GGML_TYPE_F32, 128, 64, 16, 2, 128, 2)); // SSD multi-chunk, no tail (exercises the chunk-to-chunk state handoff)
+    test_cases.emplace_back(new test_ssm_scan(GGML_TYPE_F32, 128, 64, 16, 2, 128, 2, false, /*K=*/1, /*weak_decay=*/true)); // SSD multi-chunk, carried state not numerically negligible
 
     test_cases.emplace_back(new test_rwkv_wkv6(GGML_TYPE_F32, 32, 64, 1, 1));
     test_cases.emplace_back(new test_rwkv_wkv6(GGML_TYPE_F32, 32, 64, 32, 1));
@@ -10573,6 +10584,101 @@ static std::vector<std::unique_ptr<test_case>> make_test_cases_from_file(const c
     return test_cases;
 }
 
+// ---- FA vec (Q,NE): forced-config numerical slice (Metal only) ----
+using set_fa_vec_override_t   = void (*)(int, int);
+using clear_fa_vec_override_t = void (*)(void);
+
+// NL = 32/NE must divide both dk/4 and dv/4.
+static std::vector<int> fa_vec_legal_ne(int dk, int dv) {
+    std::vector<int> r;
+    for (int ne : {1, 2, 4}) {
+        const int nl = 32 / ne;
+        if ((dk/4) % nl == 0 && (dv/4) % nl == 0) {
+            r.push_back(ne);
+        }
+    }
+    return r;
+}
+
+static bool op_names_filter_selects(const char * op_names_filter, const char * op_name) {
+    if (!op_names_filter) {
+        return true;
+    }
+    std::string_view filter(op_names_filter);
+    while (!filter.empty()) {
+        auto comma_pos = filter.find_first_of(',');
+        const auto lparen_pos = filter.find_first_of('(');
+        std::string_view entry;
+        if (lparen_pos < comma_pos) {
+            const auto rparen_pos = filter.find_first_of(')');
+            comma_pos = filter.find_first_of(',', rparen_pos);
+            entry = filter.substr(0, lparen_pos);
+        } else {
+            entry = filter.substr(0, comma_pos);
+        }
+        if (entry == op_name) {
+            return true;
+        }
+        filter = comma_pos != std::string_view::npos ? filter.substr(comma_pos + 1) : "";
+    }
+    return false;
+}
+
+// Covers padded rows, sinks, kvpad, multi-SIMDgroup reduction, quantized K/V, and MLA views.
+// The override is backend-global, so this runs after all parallel workers have joined.
+static bool run_fa_vec_slice(ggml_backend_t backend, ggml_backend_t backend_cpu, const char * op_names_filter) {
+    if (!op_names_filter_selects(op_names_filter, "FLASH_ATTN_EXT")) {
+        return true;
+    }
+
+    auto * reg = ggml_backend_dev_backend_reg(ggml_backend_get_device(backend));
+
+    auto set_ov   = (set_fa_vec_override_t)   ggml_backend_reg_get_proc_address(reg, "ggml_backend_metal_tuning_set_fa_vec_override");
+    auto clear_ov = (clear_fa_vec_override_t) ggml_backend_reg_get_proc_address(reg, "ggml_backend_metal_tuning_clear_fa_vec_override");
+    if (!set_ov || !clear_ov) {
+        return true;  // not the Metal backend: nothing to force
+    }
+
+    struct shape_t { int dk, dv; };
+    const shape_t   shapes[] = { { 128, 128 }, { 576, 512 } };  // mainstream head size + MLA shared K/V view
+    const int       ne01_pts[] = { 1, 3 };                      // decode, and padded rows for Q=2 and Q=4
+    const int       ne11_pts[] = { 512, 4097 };                 // nsg=1, and nsg>=2 together with kvpad
+    const ggml_type types[]    = { GGML_TYPE_F16, GGML_TYPE_Q4_0 };
+
+    int n_run = 0, n_fail = 0;
+    for (auto s : shapes) {
+        for (int ne : fa_vec_legal_ne(s.dk, s.dv)) {
+            for (int Q : { 1, 2, 4 }) {
+                for (ggml_type type_kv : types) {
+                    for (bool sinks : { false, true }) {
+                        for (int ne01 : ne01_pts) {
+                            for (int ne11 : ne11_pts) {
+                                set_ov(Q, ne);
+                                test_flash_attn_ext tc(s.dk, s.dv, /*nh=*/4, { 1, 1 }, /*kv=*/ne11, /*nb=*/ne01,
+                                                       /*mask=*/true, sinks, 0.0f, 0.0f, GGML_PREC_F32,
+                                                       type_kv, type_kv);
+                                auto st = tc.eval(backend, backend_cpu, "FLASH_ATTN_EXT", nullptr);
+                                clear_ov();
+
+                                if (st == test_status_t::FAIL) {
+                                    printf("  FAIL fa_vec slice: dk=%d dv=%d Q=%d ne=%d type=%s ne01=%d ne11=%d sinks=%d\n",
+                                           s.dk, s.dv, Q, ne, ggml_type_name(type_kv), ne01, ne11, (int) sinks);
+                                    n_fail++;
+                                }
+                                n_run++;
+                            }
+                        }
+                    }
+                }
+            }
+        }
+    }
+
+    printf("  fa_vec (Q,NE) slice: %d cases run, %d failed\n", n_run, n_fail);
+
+    return n_fail == 0;
+}
+
 static bool test_backend(ggml_backend_t backend, ggml_backend_dev_t dev, test_mode mode, const char * op_names_filter, const char * params_filter,
                          printer * output_printer, const char * test_file_path, int parallel_workers) {
     auto filter_test_cases = [](std::vector<std::unique_ptr<test_case>> & test_cases, const char * params_filter) {
@@ -10710,7 +10816,9 @@ static bool test_backend(ggml_backend_t backend, ggml_backend_dev_t dev, test_mo
         output_printer->print_summary(test_summary_info(n_ok, tests_run, false));
         output_printer->print_failed_tests(failed_tests);
 
-        return n_ok == tests_run;
+        const bool slice_ok = run_fa_vec_slice(backend, backend_cpu.get(), op_names_filter);
+
+        return n_ok == tests_run && slice_ok;
     }
 
     if (mode == MODE_GRAD) {