snapshot: debug ggml-hexagon swiglu-oai

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

@@ -3297,7 +3297,7 @@ static bool ggml_backend_hexagon_device_supports_op(ggml_backend_dev_t dev, cons
             break;
 
         case GGML_OP_GLU:
-            if ((ggml_get_glu_op(op) == GGML_GLU_OP_SWIGLU) /* || (ggml_get_glu_op(op) == GGML_GLU_OP_SWIGLU_OAI) */) {
+            if ((ggml_get_glu_op(op) == GGML_GLU_OP_SWIGLU) || (ggml_get_glu_op(op) == GGML_GLU_OP_SWIGLU_OAI) ) {
                 supp = ggml_hexagon_supported_activations(sess, op);
             }
             break;

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

@@ -229,9 +229,29 @@ static void glu_swiglu_oai_fp32_per_thread(const struct htp_tensor * src0,
         }
 
         // x (src0_spad_data) = std::min(src0_p[k], limit);
-        hvx_min_scalar_f32((const uint8_t *) src0, limit, src0_spad_data, nc);
+        //hvx_min_scalar_f32((const uint8_t *) src0, limit, src0_spad_data, nc);
-        // y1 (src1_spad_data) = std::clamp(src1_p[k], -limit, limit);
+        // // y1 (src1_spad_data) = std::clamp(src1_p[k], -limit, limit);
-        hvx_clamp_scalar_f32((const uint8_t *) src1, limit, limit, src1_spad_data, nc);
+        // hvx_clamp_scalar_f32((const uint8_t *) src1, limit, limit, src1_spad_data, nc);
+
+        // do manual limit
+        for (int i = 0; i < nc; i++) {
+            if (src0[i] > limit) {
+                ((float *) src0_spad_data)[i] = limit;
+            } else {
+                ((float *) src0_spad_data)[i] = src0[i];
+            }
+        }
+
+        for (int i = 0; i < nc; i++) {
+            if (src1[i] > limit) {
+                ((float *) src1_spad_data)[i] = limit;
+            } else if (src1[i] < -limit) {
+                ((float *) src1_spad_data)[i] = -limit;
+            } else {
+                ((float *) src1_spad_data)[i] = src1[i];
+            }
+        }
+
         // y (src1_spad_data)  = y1 + 1.f
         hvx_add_scalar_f32(src1_spad_data, 1.0, src1_spad_data, nc);
         // x1 (dst_spad_data) = alpha * (x)

ggml/src/ggml-hexagon/htp/hvx-utils.c

@@ -886,14 +886,34 @@ void hvx_min_scalar_f32(const uint8_t * restrict src, const float val, uint8_t *
 
     HVX_Vector vec_min = Q6_V_vsplat_R(val);
 
-    HVX_Vector * restrict vec_in  = (HVX_Vector *) src;
+    int unaligned_input_addr = 0;
-    HVX_Vector * restrict vec_out = (HVX_Vector *) dst;
+    int unaligned_output_addr = 0;
 
-    #pragma unroll(4)
+    if(htp_is_aligned((void *) src, VLEN) == 0) {
-    for (int i = 0; i < num_elems_whole; i += VLEN_FP32) {
+        unaligned_input_addr = 1;
-        vec_min    = Q6_Vsf_vmin_VsfVsf(vec_min, *vec_in++);
-        *vec_out++ = Q6_Vsf_equals_Vqf32(vec_min);
     }
+    if(htp_is_aligned((void *) dst, VLEN) == 0) {
+        unaligned_output_addr = 1;
+    }
+
+
+    if(unaligned_input_addr == 0 && unaligned_output_addr == 0){
+
+        HVX_Vector * restrict vec_in  = (HVX_Vector *) src;
+        HVX_Vector * restrict vec_out = (HVX_Vector *) dst;
+
+        #pragma unroll(4)
+        for (int i = 0; i < num_elems_whole; i += VLEN_FP32) {
+            vec_min    = Q6_Vsf_vmin_VsfVsf(vec_min, *vec_in++);
+            *vec_out++ = Q6_Vsf_equals_Vqf32(vec_min);
+        }
+    }
+    else if(unaligned_output_addr == 0){
+
+    }else if()
+
+
+
 
     if (left_over > 0) {
         const float * srcf = (const float *) src + num_elems_whole;

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

@@ -806,6 +806,7 @@ static void htp_packet_callback(dspqueue_t queue, int error, void * context) {
                 break;
 
             case HTP_OP_GLU_SWIGLU:
+            case HTP_OP_GLU_SWIGLU_OAI:
             case HTP_OP_SOFTMAX:
                 if ((n_bufs != 2) && (n_bufs != 3)) {
                     FARF(ERROR, "Bad act-req buffer list");