CANN: Add 310P operator support check (llama/12962)

ggml/src/ggml-cann/aclnn_ops.cpp

@@ -625,6 +625,10 @@ static void ggml_cann_avg_pool2d(ggml_backend_cann_context& ctx,
     bool count_include_pad = true;
     int64_t divisor_override = 0;
     int8_t cube_math_type = 0;
+#ifdef ASCEND_310P
+    cube_math_type = 1;
+#endif
+
     GGML_CANN_CALL_ACLNN_OP(AvgPool2d, acl_src, kernel_size, strides, paddings_avg,
                     ceil_mode, count_include_pad, divisor_override,
                     cube_math_type, acl_dst);
@@ -2590,6 +2594,10 @@ void ggml_cann_conv_transpose_1d(ggml_backend_cann_context& ctx, ggml_tensor* ds
     int64_t groups = 1;
     int8_t cubeMathType = 0;
 
+#ifdef ASCEND_310P
+    cubeMathType = 1;
+#endif
+
     GGML_CANN_CALL_ACLNN_OP(Convolution, acl_input, acl_weight, nullptr, stride,
         padding, dilation, transposed, padding, groups, acl_dst, cubeMathType);
 

ggml/src/ggml-cann/ggml-cann.cpp

@@ -2022,6 +2022,10 @@ static bool ggml_backend_cann_supports_op(ggml_backend_dev_t dev,
                     return true;
                 case GGML_TYPE_Q8_0:
                 case GGML_TYPE_Q4_0:
+#ifdef ASCEND_310P
+                    // Q4 && Q8 per group is not suppor on 310p device
+                    return false;
+#endif
                     // only support contiguous for quantized types.
                     return ggml_is_contiguous(op->src[0]) &&
                             ggml_is_contiguous(op->src[1]);
@@ -2107,6 +2111,12 @@ static bool ggml_backend_cann_supports_op(ggml_backend_dev_t dev,
         }
         case GGML_OP_POOL_2D: {
             const int32_t * opts = (const int32_t *) op->op_params;
+#ifdef ASCEND_310P
+            enum ggml_op_pool opt = static_cast<ggml_op_pool>(opts[0]);
+            if(opt == GGML_OP_POOL_MAX){
+                return false;
+            }
+#endif
             const int       k0   = opts[1];
             const int       k1   = opts[2];
             const int       p0   = opts[5];